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, 2015 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 * Arc buffers may have an associated eviction callback function. 103 * This function will be invoked prior to removing the buffer (e.g. 104 * in arc_do_user_evicts()). Note however that the data associated 105 * with the buffer may be evicted prior to the callback. The callback 106 * must be made with *no locks held* (to prevent deadlock). Additionally, 107 * the users of callbacks must ensure that their private data is 108 * protected from simultaneous callbacks from arc_clear_callback() 109 * and arc_do_user_evicts(). 110 * 111 * Note that the majority of the performance stats are manipulated 112 * with atomic operations. 113 * 114 * The L2ARC uses the l2ad_mtx on each vdev for the following: 115 * 116 * - L2ARC buflist creation 117 * - L2ARC buflist eviction 118 * - L2ARC write completion, which walks L2ARC buflists 119 * - ARC header destruction, as it removes from L2ARC buflists 120 * - ARC header release, as it removes from L2ARC buflists 121 */ 122 123 #include <sys/spa.h> 124 #include <sys/zio.h> 125 #include <sys/zio_compress.h> 126 #include <sys/zfs_context.h> 127 #include <sys/arc.h> 128 #include <sys/refcount.h> 129 #include <sys/vdev.h> 130 #include <sys/vdev_impl.h> 131 #include <sys/dsl_pool.h> 132 #include <sys/multilist.h> 133 #ifdef _KERNEL 134 #include <sys/vmsystm.h> 135 #include <vm/anon.h> 136 #include <sys/fs/swapnode.h> 137 #include <sys/dnlc.h> 138 #endif 139 #include <sys/callb.h> 140 #include <sys/kstat.h> 141 #include <zfs_fletcher.h> 142 #include <sys/zfs_ioctl.h> 143 144 #ifndef _KERNEL 145 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 146 boolean_t arc_watch = B_FALSE; 147 int arc_procfd; 148 #endif 149 150 static kmutex_t arc_reclaim_lock; 151 static kcondvar_t arc_reclaim_thread_cv; 152 static boolean_t arc_reclaim_thread_exit; 153 static kcondvar_t arc_reclaim_waiters_cv; 154 155 static kmutex_t arc_user_evicts_lock; 156 static kcondvar_t arc_user_evicts_cv; 157 static boolean_t arc_user_evicts_thread_exit; 158 159 uint_t arc_reduce_dnlc_percent = 3; 160 161 /* 162 * The number of headers to evict in arc_evict_state_impl() before 163 * dropping the sublist lock and evicting from another sublist. A lower 164 * value means we're more likely to evict the "correct" header (i.e. the 165 * oldest header in the arc state), but comes with higher overhead 166 * (i.e. more invocations of arc_evict_state_impl()). 167 */ 168 int zfs_arc_evict_batch_limit = 10; 169 170 /* 171 * The number of sublists used for each of the arc state lists. If this 172 * is not set to a suitable value by the user, it will be configured to 173 * the number of CPUs on the system in arc_init(). 174 */ 175 int zfs_arc_num_sublists_per_state = 0; 176 177 /* number of seconds before growing cache again */ 178 static int arc_grow_retry = 60; 179 180 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */ 181 int zfs_arc_overflow_shift = 8; 182 183 /* shift of arc_c for calculating both min and max arc_p */ 184 static int arc_p_min_shift = 4; 185 186 /* log2(fraction of arc to reclaim) */ 187 static int arc_shrink_shift = 7; 188 189 /* 190 * log2(fraction of ARC which must be free to allow growing). 191 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 192 * when reading a new block into the ARC, we will evict an equal-sized block 193 * from the ARC. 194 * 195 * This must be less than arc_shrink_shift, so that when we shrink the ARC, 196 * we will still not allow it to grow. 197 */ 198 int arc_no_grow_shift = 5; 199 200 201 /* 202 * minimum lifespan of a prefetch block in clock ticks 203 * (initialized in arc_init()) 204 */ 205 static int arc_min_prefetch_lifespan; 206 207 /* 208 * If this percent of memory is free, don't throttle. 209 */ 210 int arc_lotsfree_percent = 10; 211 212 static int arc_dead; 213 214 /* 215 * The arc has filled available memory and has now warmed up. 216 */ 217 static boolean_t arc_warm; 218 219 /* 220 * These tunables are for performance analysis. 221 */ 222 uint64_t zfs_arc_max; 223 uint64_t zfs_arc_min; 224 uint64_t zfs_arc_meta_limit = 0; 225 uint64_t zfs_arc_meta_min = 0; 226 int zfs_arc_grow_retry = 0; 227 int zfs_arc_shrink_shift = 0; 228 int zfs_arc_p_min_shift = 0; 229 int zfs_disable_dup_eviction = 0; 230 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 231 232 /* 233 * Note that buffers can be in one of 6 states: 234 * ARC_anon - anonymous (discussed below) 235 * ARC_mru - recently used, currently cached 236 * ARC_mru_ghost - recentely used, no longer in cache 237 * ARC_mfu - frequently used, currently cached 238 * ARC_mfu_ghost - frequently used, no longer in cache 239 * ARC_l2c_only - exists in L2ARC but not other states 240 * When there are no active references to the buffer, they are 241 * are linked onto a list in one of these arc states. These are 242 * the only buffers that can be evicted or deleted. Within each 243 * state there are multiple lists, one for meta-data and one for 244 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 245 * etc.) is tracked separately so that it can be managed more 246 * explicitly: favored over data, limited explicitly. 247 * 248 * Anonymous buffers are buffers that are not associated with 249 * a DVA. These are buffers that hold dirty block copies 250 * before they are written to stable storage. By definition, 251 * they are "ref'd" and are considered part of arc_mru 252 * that cannot be freed. Generally, they will aquire a DVA 253 * as they are written and migrate onto the arc_mru list. 254 * 255 * The ARC_l2c_only state is for buffers that are in the second 256 * level ARC but no longer in any of the ARC_m* lists. The second 257 * level ARC itself may also contain buffers that are in any of 258 * the ARC_m* states - meaning that a buffer can exist in two 259 * places. The reason for the ARC_l2c_only state is to keep the 260 * buffer header in the hash table, so that reads that hit the 261 * second level ARC benefit from these fast lookups. 262 */ 263 264 typedef struct arc_state { 265 /* 266 * list of evictable buffers 267 */ 268 multilist_t arcs_list[ARC_BUFC_NUMTYPES]; 269 /* 270 * total amount of evictable data in this state 271 */ 272 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; 273 /* 274 * total amount of data in this state; this includes: evictable, 275 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 276 */ 277 refcount_t arcs_size; 278 } arc_state_t; 279 280 /* The 6 states: */ 281 static arc_state_t ARC_anon; 282 static arc_state_t ARC_mru; 283 static arc_state_t ARC_mru_ghost; 284 static arc_state_t ARC_mfu; 285 static arc_state_t ARC_mfu_ghost; 286 static arc_state_t ARC_l2c_only; 287 288 typedef struct arc_stats { 289 kstat_named_t arcstat_hits; 290 kstat_named_t arcstat_misses; 291 kstat_named_t arcstat_demand_data_hits; 292 kstat_named_t arcstat_demand_data_misses; 293 kstat_named_t arcstat_demand_metadata_hits; 294 kstat_named_t arcstat_demand_metadata_misses; 295 kstat_named_t arcstat_prefetch_data_hits; 296 kstat_named_t arcstat_prefetch_data_misses; 297 kstat_named_t arcstat_prefetch_metadata_hits; 298 kstat_named_t arcstat_prefetch_metadata_misses; 299 kstat_named_t arcstat_mru_hits; 300 kstat_named_t arcstat_mru_ghost_hits; 301 kstat_named_t arcstat_mfu_hits; 302 kstat_named_t arcstat_mfu_ghost_hits; 303 kstat_named_t arcstat_deleted; 304 /* 305 * Number of buffers that could not be evicted because the hash lock 306 * was held by another thread. The lock may not necessarily be held 307 * by something using the same buffer, since hash locks are shared 308 * by multiple buffers. 309 */ 310 kstat_named_t arcstat_mutex_miss; 311 /* 312 * Number of buffers skipped because they have I/O in progress, are 313 * indrect prefetch buffers that have not lived long enough, or are 314 * not from the spa we're trying to evict from. 315 */ 316 kstat_named_t arcstat_evict_skip; 317 /* 318 * Number of times arc_evict_state() was unable to evict enough 319 * buffers to reach it's target amount. 320 */ 321 kstat_named_t arcstat_evict_not_enough; 322 kstat_named_t arcstat_evict_l2_cached; 323 kstat_named_t arcstat_evict_l2_eligible; 324 kstat_named_t arcstat_evict_l2_ineligible; 325 kstat_named_t arcstat_evict_l2_skip; 326 kstat_named_t arcstat_hash_elements; 327 kstat_named_t arcstat_hash_elements_max; 328 kstat_named_t arcstat_hash_collisions; 329 kstat_named_t arcstat_hash_chains; 330 kstat_named_t arcstat_hash_chain_max; 331 kstat_named_t arcstat_p; 332 kstat_named_t arcstat_c; 333 kstat_named_t arcstat_c_min; 334 kstat_named_t arcstat_c_max; 335 kstat_named_t arcstat_size; 336 /* 337 * Number of bytes consumed by internal ARC structures necessary 338 * for tracking purposes; these structures are not actually 339 * backed by ARC buffers. This includes arc_buf_hdr_t structures 340 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 341 * caches), and arc_buf_t structures (allocated via arc_buf_t 342 * cache). 343 */ 344 kstat_named_t arcstat_hdr_size; 345 /* 346 * Number of bytes consumed by ARC buffers of type equal to 347 * ARC_BUFC_DATA. This is generally consumed by buffers backing 348 * on disk user data (e.g. plain file contents). 349 */ 350 kstat_named_t arcstat_data_size; 351 /* 352 * Number of bytes consumed by ARC buffers of type equal to 353 * ARC_BUFC_METADATA. This is generally consumed by buffers 354 * backing on disk data that is used for internal ZFS 355 * structures (e.g. ZAP, dnode, indirect blocks, etc). 356 */ 357 kstat_named_t arcstat_metadata_size; 358 /* 359 * Number of bytes consumed by various buffers and structures 360 * not actually backed with ARC buffers. This includes bonus 361 * buffers (allocated directly via zio_buf_* functions), 362 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 363 * cache), and dnode_t structures (allocated via dnode_t cache). 364 */ 365 kstat_named_t arcstat_other_size; 366 /* 367 * Total number of bytes consumed by ARC buffers residing in the 368 * arc_anon state. This includes *all* buffers in the arc_anon 369 * state; e.g. data, metadata, evictable, and unevictable buffers 370 * are all included in this value. 371 */ 372 kstat_named_t arcstat_anon_size; 373 /* 374 * Number of bytes consumed by ARC buffers that meet the 375 * following criteria: backing buffers of type ARC_BUFC_DATA, 376 * residing in the arc_anon state, and are eligible for eviction 377 * (e.g. have no outstanding holds on the buffer). 378 */ 379 kstat_named_t arcstat_anon_evictable_data; 380 /* 381 * Number of bytes consumed by ARC buffers that meet the 382 * following criteria: backing buffers of type ARC_BUFC_METADATA, 383 * residing in the arc_anon state, and are eligible for eviction 384 * (e.g. have no outstanding holds on the buffer). 385 */ 386 kstat_named_t arcstat_anon_evictable_metadata; 387 /* 388 * Total number of bytes consumed by ARC buffers residing in the 389 * arc_mru state. This includes *all* buffers in the arc_mru 390 * state; e.g. data, metadata, evictable, and unevictable buffers 391 * are all included in this value. 392 */ 393 kstat_named_t arcstat_mru_size; 394 /* 395 * Number of bytes consumed by ARC buffers that meet the 396 * following criteria: backing buffers of type ARC_BUFC_DATA, 397 * residing in the arc_mru state, and are eligible for eviction 398 * (e.g. have no outstanding holds on the buffer). 399 */ 400 kstat_named_t arcstat_mru_evictable_data; 401 /* 402 * Number of bytes consumed by ARC buffers that meet the 403 * following criteria: backing buffers of type ARC_BUFC_METADATA, 404 * residing in the arc_mru state, and are eligible for eviction 405 * (e.g. have no outstanding holds on the buffer). 406 */ 407 kstat_named_t arcstat_mru_evictable_metadata; 408 /* 409 * Total number of bytes that *would have been* consumed by ARC 410 * buffers in the arc_mru_ghost state. The key thing to note 411 * here, is the fact that this size doesn't actually indicate 412 * RAM consumption. The ghost lists only consist of headers and 413 * don't actually have ARC buffers linked off of these headers. 414 * Thus, *if* the headers had associated ARC buffers, these 415 * buffers *would have* consumed this number of bytes. 416 */ 417 kstat_named_t arcstat_mru_ghost_size; 418 /* 419 * Number of bytes that *would have been* consumed by ARC 420 * buffers that are eligible for eviction, of type 421 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 422 */ 423 kstat_named_t arcstat_mru_ghost_evictable_data; 424 /* 425 * Number of bytes that *would have been* consumed by ARC 426 * buffers that are eligible for eviction, of type 427 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 428 */ 429 kstat_named_t arcstat_mru_ghost_evictable_metadata; 430 /* 431 * Total number of bytes consumed by ARC buffers residing in the 432 * arc_mfu state. This includes *all* buffers in the arc_mfu 433 * state; e.g. data, metadata, evictable, and unevictable buffers 434 * are all included in this value. 435 */ 436 kstat_named_t arcstat_mfu_size; 437 /* 438 * Number of bytes consumed by ARC buffers that are eligible for 439 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 440 * state. 441 */ 442 kstat_named_t arcstat_mfu_evictable_data; 443 /* 444 * Number of bytes consumed by ARC buffers that are eligible for 445 * eviction, of type ARC_BUFC_METADATA, and reside in the 446 * arc_mfu state. 447 */ 448 kstat_named_t arcstat_mfu_evictable_metadata; 449 /* 450 * Total number of bytes that *would have been* consumed by ARC 451 * buffers in the arc_mfu_ghost state. See the comment above 452 * arcstat_mru_ghost_size for more details. 453 */ 454 kstat_named_t arcstat_mfu_ghost_size; 455 /* 456 * Number of bytes that *would have been* consumed by ARC 457 * buffers that are eligible for eviction, of type 458 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 459 */ 460 kstat_named_t arcstat_mfu_ghost_evictable_data; 461 /* 462 * Number of bytes that *would have been* consumed by ARC 463 * buffers that are eligible for eviction, of type 464 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 465 */ 466 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 467 kstat_named_t arcstat_l2_hits; 468 kstat_named_t arcstat_l2_misses; 469 kstat_named_t arcstat_l2_feeds; 470 kstat_named_t arcstat_l2_rw_clash; 471 kstat_named_t arcstat_l2_read_bytes; 472 kstat_named_t arcstat_l2_write_bytes; 473 kstat_named_t arcstat_l2_writes_sent; 474 kstat_named_t arcstat_l2_writes_done; 475 kstat_named_t arcstat_l2_writes_error; 476 kstat_named_t arcstat_l2_writes_lock_retry; 477 kstat_named_t arcstat_l2_evict_lock_retry; 478 kstat_named_t arcstat_l2_evict_reading; 479 kstat_named_t arcstat_l2_evict_l1cached; 480 kstat_named_t arcstat_l2_free_on_write; 481 kstat_named_t arcstat_l2_cdata_free_on_write; 482 kstat_named_t arcstat_l2_abort_lowmem; 483 kstat_named_t arcstat_l2_cksum_bad; 484 kstat_named_t arcstat_l2_io_error; 485 kstat_named_t arcstat_l2_size; 486 kstat_named_t arcstat_l2_asize; 487 kstat_named_t arcstat_l2_hdr_size; 488 kstat_named_t arcstat_l2_compress_successes; 489 kstat_named_t arcstat_l2_compress_zeros; 490 kstat_named_t arcstat_l2_compress_failures; 491 kstat_named_t arcstat_memory_throttle_count; 492 kstat_named_t arcstat_duplicate_buffers; 493 kstat_named_t arcstat_duplicate_buffers_size; 494 kstat_named_t arcstat_duplicate_reads; 495 kstat_named_t arcstat_meta_used; 496 kstat_named_t arcstat_meta_limit; 497 kstat_named_t arcstat_meta_max; 498 kstat_named_t arcstat_meta_min; 499 kstat_named_t arcstat_sync_wait_for_async; 500 kstat_named_t arcstat_demand_hit_predictive_prefetch; 501 } arc_stats_t; 502 503 static arc_stats_t arc_stats = { 504 { "hits", KSTAT_DATA_UINT64 }, 505 { "misses", KSTAT_DATA_UINT64 }, 506 { "demand_data_hits", KSTAT_DATA_UINT64 }, 507 { "demand_data_misses", KSTAT_DATA_UINT64 }, 508 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 509 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 510 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 511 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 512 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 513 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 514 { "mru_hits", KSTAT_DATA_UINT64 }, 515 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 516 { "mfu_hits", KSTAT_DATA_UINT64 }, 517 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 518 { "deleted", KSTAT_DATA_UINT64 }, 519 { "mutex_miss", KSTAT_DATA_UINT64 }, 520 { "evict_skip", KSTAT_DATA_UINT64 }, 521 { "evict_not_enough", KSTAT_DATA_UINT64 }, 522 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 523 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 524 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 525 { "evict_l2_skip", KSTAT_DATA_UINT64 }, 526 { "hash_elements", KSTAT_DATA_UINT64 }, 527 { "hash_elements_max", KSTAT_DATA_UINT64 }, 528 { "hash_collisions", KSTAT_DATA_UINT64 }, 529 { "hash_chains", KSTAT_DATA_UINT64 }, 530 { "hash_chain_max", KSTAT_DATA_UINT64 }, 531 { "p", KSTAT_DATA_UINT64 }, 532 { "c", KSTAT_DATA_UINT64 }, 533 { "c_min", KSTAT_DATA_UINT64 }, 534 { "c_max", KSTAT_DATA_UINT64 }, 535 { "size", KSTAT_DATA_UINT64 }, 536 { "hdr_size", KSTAT_DATA_UINT64 }, 537 { "data_size", KSTAT_DATA_UINT64 }, 538 { "metadata_size", KSTAT_DATA_UINT64 }, 539 { "other_size", KSTAT_DATA_UINT64 }, 540 { "anon_size", KSTAT_DATA_UINT64 }, 541 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 542 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 543 { "mru_size", KSTAT_DATA_UINT64 }, 544 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 545 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 546 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 547 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 548 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 549 { "mfu_size", KSTAT_DATA_UINT64 }, 550 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 551 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 552 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 553 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 554 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 555 { "l2_hits", KSTAT_DATA_UINT64 }, 556 { "l2_misses", KSTAT_DATA_UINT64 }, 557 { "l2_feeds", KSTAT_DATA_UINT64 }, 558 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 559 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 560 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 561 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 562 { "l2_writes_done", KSTAT_DATA_UINT64 }, 563 { "l2_writes_error", KSTAT_DATA_UINT64 }, 564 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 565 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 566 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 567 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 568 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 569 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 570 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 571 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 572 { "l2_io_error", KSTAT_DATA_UINT64 }, 573 { "l2_size", KSTAT_DATA_UINT64 }, 574 { "l2_asize", KSTAT_DATA_UINT64 }, 575 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 576 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 577 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 578 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 579 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 580 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 581 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 582 { "duplicate_reads", KSTAT_DATA_UINT64 }, 583 { "arc_meta_used", KSTAT_DATA_UINT64 }, 584 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 585 { "arc_meta_max", KSTAT_DATA_UINT64 }, 586 { "arc_meta_min", KSTAT_DATA_UINT64 }, 587 { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 588 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 589 }; 590 591 #define ARCSTAT(stat) (arc_stats.stat.value.ui64) 592 593 #define ARCSTAT_INCR(stat, val) \ 594 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 595 596 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 597 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 598 599 #define ARCSTAT_MAX(stat, val) { \ 600 uint64_t m; \ 601 while ((val) > (m = arc_stats.stat.value.ui64) && \ 602 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 603 continue; \ 604 } 605 606 #define ARCSTAT_MAXSTAT(stat) \ 607 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 608 609 /* 610 * We define a macro to allow ARC hits/misses to be easily broken down by 611 * two separate conditions, giving a total of four different subtypes for 612 * each of hits and misses (so eight statistics total). 613 */ 614 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 615 if (cond1) { \ 616 if (cond2) { \ 617 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 618 } else { \ 619 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 620 } \ 621 } else { \ 622 if (cond2) { \ 623 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 624 } else { \ 625 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 626 } \ 627 } 628 629 kstat_t *arc_ksp; 630 static arc_state_t *arc_anon; 631 static arc_state_t *arc_mru; 632 static arc_state_t *arc_mru_ghost; 633 static arc_state_t *arc_mfu; 634 static arc_state_t *arc_mfu_ghost; 635 static arc_state_t *arc_l2c_only; 636 637 /* 638 * There are several ARC variables that are critical to export as kstats -- 639 * but we don't want to have to grovel around in the kstat whenever we wish to 640 * manipulate them. For these variables, we therefore define them to be in 641 * terms of the statistic variable. This assures that we are not introducing 642 * the possibility of inconsistency by having shadow copies of the variables, 643 * while still allowing the code to be readable. 644 */ 645 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 646 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 647 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 648 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 649 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 650 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 651 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 652 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 653 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 654 655 #define L2ARC_IS_VALID_COMPRESS(_c_) \ 656 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 657 658 static int arc_no_grow; /* Don't try to grow cache size */ 659 static uint64_t arc_tempreserve; 660 static uint64_t arc_loaned_bytes; 661 662 typedef struct arc_callback arc_callback_t; 663 664 struct arc_callback { 665 void *acb_private; 666 arc_done_func_t *acb_done; 667 arc_buf_t *acb_buf; 668 zio_t *acb_zio_dummy; 669 arc_callback_t *acb_next; 670 }; 671 672 typedef struct arc_write_callback arc_write_callback_t; 673 674 struct arc_write_callback { 675 void *awcb_private; 676 arc_done_func_t *awcb_ready; 677 arc_done_func_t *awcb_physdone; 678 arc_done_func_t *awcb_done; 679 arc_buf_t *awcb_buf; 680 }; 681 682 /* 683 * ARC buffers are separated into multiple structs as a memory saving measure: 684 * - Common fields struct, always defined, and embedded within it: 685 * - L2-only fields, always allocated but undefined when not in L2ARC 686 * - L1-only fields, only allocated when in L1ARC 687 * 688 * Buffer in L1 Buffer only in L2 689 * +------------------------+ +------------------------+ 690 * | arc_buf_hdr_t | | arc_buf_hdr_t | 691 * | | | | 692 * | | | | 693 * | | | | 694 * +------------------------+ +------------------------+ 695 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 696 * | (undefined if L1-only) | | | 697 * +------------------------+ +------------------------+ 698 * | l1arc_buf_hdr_t | 699 * | | 700 * | | 701 * | | 702 * | | 703 * +------------------------+ 704 * 705 * Because it's possible for the L2ARC to become extremely large, we can wind 706 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 707 * is minimized by only allocating the fields necessary for an L1-cached buffer 708 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 709 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 710 * words in pointers. arc_hdr_realloc() is used to switch a header between 711 * these two allocation states. 712 */ 713 typedef struct l1arc_buf_hdr { 714 kmutex_t b_freeze_lock; 715 #ifdef ZFS_DEBUG 716 /* 717 * used for debugging wtih kmem_flags - by allocating and freeing 718 * b_thawed when the buffer is thawed, we get a record of the stack 719 * trace that thawed it. 720 */ 721 void *b_thawed; 722 #endif 723 724 arc_buf_t *b_buf; 725 uint32_t b_datacnt; 726 /* for waiting on writes to complete */ 727 kcondvar_t b_cv; 728 729 /* protected by arc state mutex */ 730 arc_state_t *b_state; 731 multilist_node_t b_arc_node; 732 733 /* updated atomically */ 734 clock_t b_arc_access; 735 736 /* self protecting */ 737 refcount_t b_refcnt; 738 739 arc_callback_t *b_acb; 740 /* temporary buffer holder for in-flight compressed data */ 741 void *b_tmp_cdata; 742 } l1arc_buf_hdr_t; 743 744 typedef struct l2arc_dev l2arc_dev_t; 745 746 typedef struct l2arc_buf_hdr { 747 /* protected by arc_buf_hdr mutex */ 748 l2arc_dev_t *b_dev; /* L2ARC device */ 749 uint64_t b_daddr; /* disk address, offset byte */ 750 /* real alloc'd buffer size depending on b_compress applied */ 751 int32_t b_asize; 752 uint8_t b_compress; 753 754 list_node_t b_l2node; 755 } l2arc_buf_hdr_t; 756 757 struct arc_buf_hdr { 758 /* protected by hash lock */ 759 dva_t b_dva; 760 uint64_t b_birth; 761 /* 762 * Even though this checksum is only set/verified when a buffer is in 763 * the L1 cache, it needs to be in the set of common fields because it 764 * must be preserved from the time before a buffer is written out to 765 * L2ARC until after it is read back in. 766 */ 767 zio_cksum_t *b_freeze_cksum; 768 769 arc_buf_hdr_t *b_hash_next; 770 arc_flags_t b_flags; 771 772 /* immutable */ 773 int32_t b_size; 774 uint64_t b_spa; 775 776 /* L2ARC fields. Undefined when not in L2ARC. */ 777 l2arc_buf_hdr_t b_l2hdr; 778 /* L1ARC fields. Undefined when in l2arc_only state */ 779 l1arc_buf_hdr_t b_l1hdr; 780 }; 781 782 static arc_buf_t *arc_eviction_list; 783 static arc_buf_hdr_t arc_eviction_hdr; 784 785 #define GHOST_STATE(state) \ 786 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 787 (state) == arc_l2c_only) 788 789 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 790 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 791 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 792 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 793 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 794 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 795 796 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 797 #define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS) 798 #define HDR_L2_READING(hdr) \ 799 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 800 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 801 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 802 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 803 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 804 805 #define HDR_ISTYPE_METADATA(hdr) \ 806 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 807 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 808 809 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 810 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 811 812 /* 813 * Other sizes 814 */ 815 816 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 817 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 818 819 /* 820 * Hash table routines 821 */ 822 823 #define HT_LOCK_PAD 64 824 825 struct ht_lock { 826 kmutex_t ht_lock; 827 #ifdef _KERNEL 828 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 829 #endif 830 }; 831 832 #define BUF_LOCKS 256 833 typedef struct buf_hash_table { 834 uint64_t ht_mask; 835 arc_buf_hdr_t **ht_table; 836 struct ht_lock ht_locks[BUF_LOCKS]; 837 } buf_hash_table_t; 838 839 static buf_hash_table_t buf_hash_table; 840 841 #define BUF_HASH_INDEX(spa, dva, birth) \ 842 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 843 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 844 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 845 #define HDR_LOCK(hdr) \ 846 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 847 848 uint64_t zfs_crc64_table[256]; 849 850 /* 851 * Level 2 ARC 852 */ 853 854 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 855 #define L2ARC_HEADROOM 2 /* num of writes */ 856 /* 857 * If we discover during ARC scan any buffers to be compressed, we boost 858 * our headroom for the next scanning cycle by this percentage multiple. 859 */ 860 #define L2ARC_HEADROOM_BOOST 200 861 #define L2ARC_FEED_SECS 1 /* caching interval secs */ 862 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 863 864 /* 865 * Used to distinguish headers that are being process by 866 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk 867 * address. This can happen when the header is added to the l2arc's list 868 * of buffers to write in the first stage of l2arc_write_buffers(), but 869 * has not yet been written out which happens in the second stage of 870 * l2arc_write_buffers(). 871 */ 872 #define L2ARC_ADDR_UNSET ((uint64_t)(-1)) 873 874 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 875 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 876 877 /* L2ARC Performance Tunables */ 878 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 879 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 880 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 881 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 882 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 883 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 884 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 885 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 886 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 887 888 /* 889 * L2ARC Internals 890 */ 891 struct l2arc_dev { 892 vdev_t *l2ad_vdev; /* vdev */ 893 spa_t *l2ad_spa; /* spa */ 894 uint64_t l2ad_hand; /* next write location */ 895 uint64_t l2ad_start; /* first addr on device */ 896 uint64_t l2ad_end; /* last addr on device */ 897 boolean_t l2ad_first; /* first sweep through */ 898 boolean_t l2ad_writing; /* currently writing */ 899 kmutex_t l2ad_mtx; /* lock for buffer list */ 900 list_t l2ad_buflist; /* buffer list */ 901 list_node_t l2ad_node; /* device list node */ 902 refcount_t l2ad_alloc; /* allocated bytes */ 903 }; 904 905 static list_t L2ARC_dev_list; /* device list */ 906 static list_t *l2arc_dev_list; /* device list pointer */ 907 static kmutex_t l2arc_dev_mtx; /* device list mutex */ 908 static l2arc_dev_t *l2arc_dev_last; /* last device used */ 909 static list_t L2ARC_free_on_write; /* free after write buf list */ 910 static list_t *l2arc_free_on_write; /* free after write list ptr */ 911 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 912 static uint64_t l2arc_ndev; /* number of devices */ 913 914 typedef struct l2arc_read_callback { 915 arc_buf_t *l2rcb_buf; /* read buffer */ 916 spa_t *l2rcb_spa; /* spa */ 917 blkptr_t l2rcb_bp; /* original blkptr */ 918 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 919 int l2rcb_flags; /* original flags */ 920 enum zio_compress l2rcb_compress; /* applied compress */ 921 } l2arc_read_callback_t; 922 923 typedef struct l2arc_write_callback { 924 l2arc_dev_t *l2wcb_dev; /* device info */ 925 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 926 } l2arc_write_callback_t; 927 928 typedef struct l2arc_data_free { 929 /* protected by l2arc_free_on_write_mtx */ 930 void *l2df_data; 931 size_t l2df_size; 932 void (*l2df_func)(void *, size_t); 933 list_node_t l2df_list_node; 934 } l2arc_data_free_t; 935 936 static kmutex_t l2arc_feed_thr_lock; 937 static kcondvar_t l2arc_feed_thr_cv; 938 static uint8_t l2arc_thread_exit; 939 940 static void arc_get_data_buf(arc_buf_t *); 941 static void arc_access(arc_buf_hdr_t *, kmutex_t *); 942 static boolean_t arc_is_overflowing(); 943 static void arc_buf_watch(arc_buf_t *); 944 945 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 946 static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 947 948 static boolean_t l2arc_write_eligible(uint64_t, uint64_t, arc_buf_hdr_t *); 949 static void l2arc_read_done(zio_t *); 950 951 static boolean_t l2arc_compress_buf(arc_buf_hdr_t *); 952 static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 953 static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 954 955 static uint64_t 956 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 957 { 958 uint8_t *vdva = (uint8_t *)dva; 959 uint64_t crc = -1ULL; 960 int i; 961 962 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 963 964 for (i = 0; i < sizeof (dva_t); i++) 965 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 966 967 crc ^= (spa>>8) ^ birth; 968 969 return (crc); 970 } 971 972 #define BUF_EMPTY(buf) \ 973 ((buf)->b_dva.dva_word[0] == 0 && \ 974 (buf)->b_dva.dva_word[1] == 0) 975 976 #define BUF_EQUAL(spa, dva, birth, buf) \ 977 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 978 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 979 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 980 981 static void 982 buf_discard_identity(arc_buf_hdr_t *hdr) 983 { 984 hdr->b_dva.dva_word[0] = 0; 985 hdr->b_dva.dva_word[1] = 0; 986 hdr->b_birth = 0; 987 } 988 989 static arc_buf_hdr_t * 990 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 991 { 992 const dva_t *dva = BP_IDENTITY(bp); 993 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 994 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 995 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 996 arc_buf_hdr_t *hdr; 997 998 mutex_enter(hash_lock); 999 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1000 hdr = hdr->b_hash_next) { 1001 if (BUF_EQUAL(spa, dva, birth, hdr)) { 1002 *lockp = hash_lock; 1003 return (hdr); 1004 } 1005 } 1006 mutex_exit(hash_lock); 1007 *lockp = NULL; 1008 return (NULL); 1009 } 1010 1011 /* 1012 * Insert an entry into the hash table. If there is already an element 1013 * equal to elem in the hash table, then the already existing element 1014 * will be returned and the new element will not be inserted. 1015 * Otherwise returns NULL. 1016 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1017 */ 1018 static arc_buf_hdr_t * 1019 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1020 { 1021 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1022 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1023 arc_buf_hdr_t *fhdr; 1024 uint32_t i; 1025 1026 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1027 ASSERT(hdr->b_birth != 0); 1028 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1029 1030 if (lockp != NULL) { 1031 *lockp = hash_lock; 1032 mutex_enter(hash_lock); 1033 } else { 1034 ASSERT(MUTEX_HELD(hash_lock)); 1035 } 1036 1037 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1038 fhdr = fhdr->b_hash_next, i++) { 1039 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1040 return (fhdr); 1041 } 1042 1043 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1044 buf_hash_table.ht_table[idx] = hdr; 1045 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 1046 1047 /* collect some hash table performance data */ 1048 if (i > 0) { 1049 ARCSTAT_BUMP(arcstat_hash_collisions); 1050 if (i == 1) 1051 ARCSTAT_BUMP(arcstat_hash_chains); 1052 1053 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1054 } 1055 1056 ARCSTAT_BUMP(arcstat_hash_elements); 1057 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1058 1059 return (NULL); 1060 } 1061 1062 static void 1063 buf_hash_remove(arc_buf_hdr_t *hdr) 1064 { 1065 arc_buf_hdr_t *fhdr, **hdrp; 1066 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1067 1068 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1069 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1070 1071 hdrp = &buf_hash_table.ht_table[idx]; 1072 while ((fhdr = *hdrp) != hdr) { 1073 ASSERT(fhdr != NULL); 1074 hdrp = &fhdr->b_hash_next; 1075 } 1076 *hdrp = hdr->b_hash_next; 1077 hdr->b_hash_next = NULL; 1078 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1079 1080 /* collect some hash table performance data */ 1081 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1082 1083 if (buf_hash_table.ht_table[idx] && 1084 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1085 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1086 } 1087 1088 /* 1089 * Global data structures and functions for the buf kmem cache. 1090 */ 1091 static kmem_cache_t *hdr_full_cache; 1092 static kmem_cache_t *hdr_l2only_cache; 1093 static kmem_cache_t *buf_cache; 1094 1095 static void 1096 buf_fini(void) 1097 { 1098 int i; 1099 1100 kmem_free(buf_hash_table.ht_table, 1101 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1102 for (i = 0; i < BUF_LOCKS; i++) 1103 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1104 kmem_cache_destroy(hdr_full_cache); 1105 kmem_cache_destroy(hdr_l2only_cache); 1106 kmem_cache_destroy(buf_cache); 1107 } 1108 1109 /* 1110 * Constructor callback - called when the cache is empty 1111 * and a new buf is requested. 1112 */ 1113 /* ARGSUSED */ 1114 static int 1115 hdr_full_cons(void *vbuf, void *unused, int kmflag) 1116 { 1117 arc_buf_hdr_t *hdr = vbuf; 1118 1119 bzero(hdr, HDR_FULL_SIZE); 1120 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1121 refcount_create(&hdr->b_l1hdr.b_refcnt); 1122 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1123 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1124 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1125 1126 return (0); 1127 } 1128 1129 /* ARGSUSED */ 1130 static int 1131 hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1132 { 1133 arc_buf_hdr_t *hdr = vbuf; 1134 1135 bzero(hdr, HDR_L2ONLY_SIZE); 1136 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1137 1138 return (0); 1139 } 1140 1141 /* ARGSUSED */ 1142 static int 1143 buf_cons(void *vbuf, void *unused, int kmflag) 1144 { 1145 arc_buf_t *buf = vbuf; 1146 1147 bzero(buf, sizeof (arc_buf_t)); 1148 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1149 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1150 1151 return (0); 1152 } 1153 1154 /* 1155 * Destructor callback - called when a cached buf is 1156 * no longer required. 1157 */ 1158 /* ARGSUSED */ 1159 static void 1160 hdr_full_dest(void *vbuf, void *unused) 1161 { 1162 arc_buf_hdr_t *hdr = vbuf; 1163 1164 ASSERT(BUF_EMPTY(hdr)); 1165 cv_destroy(&hdr->b_l1hdr.b_cv); 1166 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1167 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1168 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1169 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1170 } 1171 1172 /* ARGSUSED */ 1173 static void 1174 hdr_l2only_dest(void *vbuf, void *unused) 1175 { 1176 arc_buf_hdr_t *hdr = vbuf; 1177 1178 ASSERT(BUF_EMPTY(hdr)); 1179 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1180 } 1181 1182 /* ARGSUSED */ 1183 static void 1184 buf_dest(void *vbuf, void *unused) 1185 { 1186 arc_buf_t *buf = vbuf; 1187 1188 mutex_destroy(&buf->b_evict_lock); 1189 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1190 } 1191 1192 /* 1193 * Reclaim callback -- invoked when memory is low. 1194 */ 1195 /* ARGSUSED */ 1196 static void 1197 hdr_recl(void *unused) 1198 { 1199 dprintf("hdr_recl called\n"); 1200 /* 1201 * umem calls the reclaim func when we destroy the buf cache, 1202 * which is after we do arc_fini(). 1203 */ 1204 if (!arc_dead) 1205 cv_signal(&arc_reclaim_thread_cv); 1206 } 1207 1208 static void 1209 buf_init(void) 1210 { 1211 uint64_t *ct; 1212 uint64_t hsize = 1ULL << 12; 1213 int i, j; 1214 1215 /* 1216 * The hash table is big enough to fill all of physical memory 1217 * with an average block size of zfs_arc_average_blocksize (default 8K). 1218 * By default, the table will take up 1219 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1220 */ 1221 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE) 1222 hsize <<= 1; 1223 retry: 1224 buf_hash_table.ht_mask = hsize - 1; 1225 buf_hash_table.ht_table = 1226 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1227 if (buf_hash_table.ht_table == NULL) { 1228 ASSERT(hsize > (1ULL << 8)); 1229 hsize >>= 1; 1230 goto retry; 1231 } 1232 1233 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1234 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1235 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1236 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1237 NULL, NULL, 0); 1238 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1239 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1240 1241 for (i = 0; i < 256; i++) 1242 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1243 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1244 1245 for (i = 0; i < BUF_LOCKS; i++) { 1246 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1247 NULL, MUTEX_DEFAULT, NULL); 1248 } 1249 } 1250 1251 /* 1252 * Transition between the two allocation states for the arc_buf_hdr struct. 1253 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 1254 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 1255 * version is used when a cache buffer is only in the L2ARC in order to reduce 1256 * memory usage. 1257 */ 1258 static arc_buf_hdr_t * 1259 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 1260 { 1261 ASSERT(HDR_HAS_L2HDR(hdr)); 1262 1263 arc_buf_hdr_t *nhdr; 1264 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1265 1266 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 1267 (old == hdr_l2only_cache && new == hdr_full_cache)); 1268 1269 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 1270 1271 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 1272 buf_hash_remove(hdr); 1273 1274 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 1275 1276 if (new == hdr_full_cache) { 1277 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1278 /* 1279 * arc_access and arc_change_state need to be aware that a 1280 * header has just come out of L2ARC, so we set its state to 1281 * l2c_only even though it's about to change. 1282 */ 1283 nhdr->b_l1hdr.b_state = arc_l2c_only; 1284 1285 /* Verify previous threads set to NULL before freeing */ 1286 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1287 } else { 1288 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1289 ASSERT0(hdr->b_l1hdr.b_datacnt); 1290 1291 /* 1292 * If we've reached here, We must have been called from 1293 * arc_evict_hdr(), as such we should have already been 1294 * removed from any ghost list we were previously on 1295 * (which protects us from racing with arc_evict_state), 1296 * thus no locking is needed during this check. 1297 */ 1298 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1299 1300 /* 1301 * A buffer must not be moved into the arc_l2c_only 1302 * state if it's not finished being written out to the 1303 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field 1304 * might try to be accessed, even though it was removed. 1305 */ 1306 VERIFY(!HDR_L2_WRITING(hdr)); 1307 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1308 1309 #ifdef ZFS_DEBUG 1310 if (hdr->b_l1hdr.b_thawed != NULL) { 1311 kmem_free(hdr->b_l1hdr.b_thawed, 1); 1312 hdr->b_l1hdr.b_thawed = NULL; 1313 } 1314 #endif 1315 1316 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR; 1317 } 1318 /* 1319 * The header has been reallocated so we need to re-insert it into any 1320 * lists it was on. 1321 */ 1322 (void) buf_hash_insert(nhdr, NULL); 1323 1324 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 1325 1326 mutex_enter(&dev->l2ad_mtx); 1327 1328 /* 1329 * We must place the realloc'ed header back into the list at 1330 * the same spot. Otherwise, if it's placed earlier in the list, 1331 * l2arc_write_buffers() could find it during the function's 1332 * write phase, and try to write it out to the l2arc. 1333 */ 1334 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 1335 list_remove(&dev->l2ad_buflist, hdr); 1336 1337 mutex_exit(&dev->l2ad_mtx); 1338 1339 /* 1340 * Since we're using the pointer address as the tag when 1341 * incrementing and decrementing the l2ad_alloc refcount, we 1342 * must remove the old pointer (that we're about to destroy) and 1343 * add the new pointer to the refcount. Otherwise we'd remove 1344 * the wrong pointer address when calling arc_hdr_destroy() later. 1345 */ 1346 1347 (void) refcount_remove_many(&dev->l2ad_alloc, 1348 hdr->b_l2hdr.b_asize, hdr); 1349 1350 (void) refcount_add_many(&dev->l2ad_alloc, 1351 nhdr->b_l2hdr.b_asize, nhdr); 1352 1353 buf_discard_identity(hdr); 1354 hdr->b_freeze_cksum = NULL; 1355 kmem_cache_free(old, hdr); 1356 1357 return (nhdr); 1358 } 1359 1360 1361 #define ARC_MINTIME (hz>>4) /* 62 ms */ 1362 1363 static void 1364 arc_cksum_verify(arc_buf_t *buf) 1365 { 1366 zio_cksum_t zc; 1367 1368 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1369 return; 1370 1371 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1372 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) { 1373 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1374 return; 1375 } 1376 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc); 1377 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1378 panic("buffer modified while frozen!"); 1379 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1380 } 1381 1382 static int 1383 arc_cksum_equal(arc_buf_t *buf) 1384 { 1385 zio_cksum_t zc; 1386 int equal; 1387 1388 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1389 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc); 1390 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1391 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1392 1393 return (equal); 1394 } 1395 1396 static void 1397 arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1398 { 1399 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1400 return; 1401 1402 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1403 if (buf->b_hdr->b_freeze_cksum != NULL) { 1404 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1405 return; 1406 } 1407 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1408 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1409 NULL, buf->b_hdr->b_freeze_cksum); 1410 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1411 arc_buf_watch(buf); 1412 } 1413 1414 #ifndef _KERNEL 1415 typedef struct procctl { 1416 long cmd; 1417 prwatch_t prwatch; 1418 } procctl_t; 1419 #endif 1420 1421 /* ARGSUSED */ 1422 static void 1423 arc_buf_unwatch(arc_buf_t *buf) 1424 { 1425 #ifndef _KERNEL 1426 if (arc_watch) { 1427 int result; 1428 procctl_t ctl; 1429 ctl.cmd = PCWATCH; 1430 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1431 ctl.prwatch.pr_size = 0; 1432 ctl.prwatch.pr_wflags = 0; 1433 result = write(arc_procfd, &ctl, sizeof (ctl)); 1434 ASSERT3U(result, ==, sizeof (ctl)); 1435 } 1436 #endif 1437 } 1438 1439 /* ARGSUSED */ 1440 static void 1441 arc_buf_watch(arc_buf_t *buf) 1442 { 1443 #ifndef _KERNEL 1444 if (arc_watch) { 1445 int result; 1446 procctl_t ctl; 1447 ctl.cmd = PCWATCH; 1448 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1449 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1450 ctl.prwatch.pr_wflags = WA_WRITE; 1451 result = write(arc_procfd, &ctl, sizeof (ctl)); 1452 ASSERT3U(result, ==, sizeof (ctl)); 1453 } 1454 #endif 1455 } 1456 1457 static arc_buf_contents_t 1458 arc_buf_type(arc_buf_hdr_t *hdr) 1459 { 1460 if (HDR_ISTYPE_METADATA(hdr)) { 1461 return (ARC_BUFC_METADATA); 1462 } else { 1463 return (ARC_BUFC_DATA); 1464 } 1465 } 1466 1467 static uint32_t 1468 arc_bufc_to_flags(arc_buf_contents_t type) 1469 { 1470 switch (type) { 1471 case ARC_BUFC_DATA: 1472 /* metadata field is 0 if buffer contains normal data */ 1473 return (0); 1474 case ARC_BUFC_METADATA: 1475 return (ARC_FLAG_BUFC_METADATA); 1476 default: 1477 break; 1478 } 1479 panic("undefined ARC buffer type!"); 1480 return ((uint32_t)-1); 1481 } 1482 1483 void 1484 arc_buf_thaw(arc_buf_t *buf) 1485 { 1486 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1487 if (buf->b_hdr->b_l1hdr.b_state != arc_anon) 1488 panic("modifying non-anon buffer!"); 1489 if (HDR_IO_IN_PROGRESS(buf->b_hdr)) 1490 panic("modifying buffer while i/o in progress!"); 1491 arc_cksum_verify(buf); 1492 } 1493 1494 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1495 if (buf->b_hdr->b_freeze_cksum != NULL) { 1496 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1497 buf->b_hdr->b_freeze_cksum = NULL; 1498 } 1499 1500 #ifdef ZFS_DEBUG 1501 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1502 if (buf->b_hdr->b_l1hdr.b_thawed != NULL) 1503 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1); 1504 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1505 } 1506 #endif 1507 1508 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1509 1510 arc_buf_unwatch(buf); 1511 } 1512 1513 void 1514 arc_buf_freeze(arc_buf_t *buf) 1515 { 1516 kmutex_t *hash_lock; 1517 1518 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1519 return; 1520 1521 hash_lock = HDR_LOCK(buf->b_hdr); 1522 mutex_enter(hash_lock); 1523 1524 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1525 buf->b_hdr->b_l1hdr.b_state == arc_anon); 1526 arc_cksum_compute(buf, B_FALSE); 1527 mutex_exit(hash_lock); 1528 1529 } 1530 1531 static void 1532 add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1533 { 1534 ASSERT(HDR_HAS_L1HDR(hdr)); 1535 ASSERT(MUTEX_HELD(hash_lock)); 1536 arc_state_t *state = hdr->b_l1hdr.b_state; 1537 1538 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 1539 (state != arc_anon)) { 1540 /* We don't use the L2-only state list. */ 1541 if (state != arc_l2c_only) { 1542 arc_buf_contents_t type = arc_buf_type(hdr); 1543 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt; 1544 multilist_t *list = &state->arcs_list[type]; 1545 uint64_t *size = &state->arcs_lsize[type]; 1546 1547 multilist_remove(list, hdr); 1548 1549 if (GHOST_STATE(state)) { 1550 ASSERT0(hdr->b_l1hdr.b_datacnt); 1551 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 1552 delta = hdr->b_size; 1553 } 1554 ASSERT(delta > 0); 1555 ASSERT3U(*size, >=, delta); 1556 atomic_add_64(size, -delta); 1557 } 1558 /* remove the prefetch flag if we get a reference */ 1559 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1560 } 1561 } 1562 1563 static int 1564 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1565 { 1566 int cnt; 1567 arc_state_t *state = hdr->b_l1hdr.b_state; 1568 1569 ASSERT(HDR_HAS_L1HDR(hdr)); 1570 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1571 ASSERT(!GHOST_STATE(state)); 1572 1573 /* 1574 * arc_l2c_only counts as a ghost state so we don't need to explicitly 1575 * check to prevent usage of the arc_l2c_only list. 1576 */ 1577 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 1578 (state != arc_anon)) { 1579 arc_buf_contents_t type = arc_buf_type(hdr); 1580 multilist_t *list = &state->arcs_list[type]; 1581 uint64_t *size = &state->arcs_lsize[type]; 1582 1583 multilist_insert(list, hdr); 1584 1585 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 1586 atomic_add_64(size, hdr->b_size * 1587 hdr->b_l1hdr.b_datacnt); 1588 } 1589 return (cnt); 1590 } 1591 1592 /* 1593 * Move the supplied buffer to the indicated state. The hash lock 1594 * for the buffer must be held by the caller. 1595 */ 1596 static void 1597 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1598 kmutex_t *hash_lock) 1599 { 1600 arc_state_t *old_state; 1601 int64_t refcnt; 1602 uint32_t datacnt; 1603 uint64_t from_delta, to_delta; 1604 arc_buf_contents_t buftype = arc_buf_type(hdr); 1605 1606 /* 1607 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 1608 * in arc_read() when bringing a buffer out of the L2ARC. However, the 1609 * L1 hdr doesn't always exist when we change state to arc_anon before 1610 * destroying a header, in which case reallocating to add the L1 hdr is 1611 * pointless. 1612 */ 1613 if (HDR_HAS_L1HDR(hdr)) { 1614 old_state = hdr->b_l1hdr.b_state; 1615 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 1616 datacnt = hdr->b_l1hdr.b_datacnt; 1617 } else { 1618 old_state = arc_l2c_only; 1619 refcnt = 0; 1620 datacnt = 0; 1621 } 1622 1623 ASSERT(MUTEX_HELD(hash_lock)); 1624 ASSERT3P(new_state, !=, old_state); 1625 ASSERT(refcnt == 0 || datacnt > 0); 1626 ASSERT(!GHOST_STATE(new_state) || datacnt == 0); 1627 ASSERT(old_state != arc_anon || datacnt <= 1); 1628 1629 from_delta = to_delta = datacnt * hdr->b_size; 1630 1631 /* 1632 * If this buffer is evictable, transfer it from the 1633 * old state list to the new state list. 1634 */ 1635 if (refcnt == 0) { 1636 if (old_state != arc_anon && old_state != arc_l2c_only) { 1637 uint64_t *size = &old_state->arcs_lsize[buftype]; 1638 1639 ASSERT(HDR_HAS_L1HDR(hdr)); 1640 multilist_remove(&old_state->arcs_list[buftype], hdr); 1641 1642 /* 1643 * If prefetching out of the ghost cache, 1644 * we will have a non-zero datacnt. 1645 */ 1646 if (GHOST_STATE(old_state) && datacnt == 0) { 1647 /* ghost elements have a ghost size */ 1648 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1649 from_delta = hdr->b_size; 1650 } 1651 ASSERT3U(*size, >=, from_delta); 1652 atomic_add_64(size, -from_delta); 1653 } 1654 if (new_state != arc_anon && new_state != arc_l2c_only) { 1655 uint64_t *size = &new_state->arcs_lsize[buftype]; 1656 1657 /* 1658 * An L1 header always exists here, since if we're 1659 * moving to some L1-cached state (i.e. not l2c_only or 1660 * anonymous), we realloc the header to add an L1hdr 1661 * beforehand. 1662 */ 1663 ASSERT(HDR_HAS_L1HDR(hdr)); 1664 multilist_insert(&new_state->arcs_list[buftype], hdr); 1665 1666 /* ghost elements have a ghost size */ 1667 if (GHOST_STATE(new_state)) { 1668 ASSERT0(datacnt); 1669 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1670 to_delta = hdr->b_size; 1671 } 1672 atomic_add_64(size, to_delta); 1673 } 1674 } 1675 1676 ASSERT(!BUF_EMPTY(hdr)); 1677 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1678 buf_hash_remove(hdr); 1679 1680 /* adjust state sizes (ignore arc_l2c_only) */ 1681 1682 if (to_delta && new_state != arc_l2c_only) { 1683 ASSERT(HDR_HAS_L1HDR(hdr)); 1684 if (GHOST_STATE(new_state)) { 1685 ASSERT0(datacnt); 1686 1687 /* 1688 * We moving a header to a ghost state, we first 1689 * remove all arc buffers. Thus, we'll have a 1690 * datacnt of zero, and no arc buffer to use for 1691 * the reference. As a result, we use the arc 1692 * header pointer for the reference. 1693 */ 1694 (void) refcount_add_many(&new_state->arcs_size, 1695 hdr->b_size, hdr); 1696 } else { 1697 ASSERT3U(datacnt, !=, 0); 1698 1699 /* 1700 * Each individual buffer holds a unique reference, 1701 * thus we must remove each of these references one 1702 * at a time. 1703 */ 1704 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1705 buf = buf->b_next) { 1706 (void) refcount_add_many(&new_state->arcs_size, 1707 hdr->b_size, buf); 1708 } 1709 } 1710 } 1711 1712 if (from_delta && old_state != arc_l2c_only) { 1713 ASSERT(HDR_HAS_L1HDR(hdr)); 1714 if (GHOST_STATE(old_state)) { 1715 /* 1716 * When moving a header off of a ghost state, 1717 * there's the possibility for datacnt to be 1718 * non-zero. This is because we first add the 1719 * arc buffer to the header prior to changing 1720 * the header's state. Since we used the header 1721 * for the reference when putting the header on 1722 * the ghost state, we must balance that and use 1723 * the header when removing off the ghost state 1724 * (even though datacnt is non zero). 1725 */ 1726 1727 IMPLY(datacnt == 0, new_state == arc_anon || 1728 new_state == arc_l2c_only); 1729 1730 (void) refcount_remove_many(&old_state->arcs_size, 1731 hdr->b_size, hdr); 1732 } else { 1733 ASSERT3P(datacnt, !=, 0); 1734 1735 /* 1736 * Each individual buffer holds a unique reference, 1737 * thus we must remove each of these references one 1738 * at a time. 1739 */ 1740 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1741 buf = buf->b_next) { 1742 (void) refcount_remove_many( 1743 &old_state->arcs_size, hdr->b_size, buf); 1744 } 1745 } 1746 } 1747 1748 if (HDR_HAS_L1HDR(hdr)) 1749 hdr->b_l1hdr.b_state = new_state; 1750 1751 /* 1752 * L2 headers should never be on the L2 state list since they don't 1753 * have L1 headers allocated. 1754 */ 1755 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 1756 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 1757 } 1758 1759 void 1760 arc_space_consume(uint64_t space, arc_space_type_t type) 1761 { 1762 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1763 1764 switch (type) { 1765 case ARC_SPACE_DATA: 1766 ARCSTAT_INCR(arcstat_data_size, space); 1767 break; 1768 case ARC_SPACE_META: 1769 ARCSTAT_INCR(arcstat_metadata_size, space); 1770 break; 1771 case ARC_SPACE_OTHER: 1772 ARCSTAT_INCR(arcstat_other_size, space); 1773 break; 1774 case ARC_SPACE_HDRS: 1775 ARCSTAT_INCR(arcstat_hdr_size, space); 1776 break; 1777 case ARC_SPACE_L2HDRS: 1778 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1779 break; 1780 } 1781 1782 if (type != ARC_SPACE_DATA) 1783 ARCSTAT_INCR(arcstat_meta_used, space); 1784 1785 atomic_add_64(&arc_size, space); 1786 } 1787 1788 void 1789 arc_space_return(uint64_t space, arc_space_type_t type) 1790 { 1791 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1792 1793 switch (type) { 1794 case ARC_SPACE_DATA: 1795 ARCSTAT_INCR(arcstat_data_size, -space); 1796 break; 1797 case ARC_SPACE_META: 1798 ARCSTAT_INCR(arcstat_metadata_size, -space); 1799 break; 1800 case ARC_SPACE_OTHER: 1801 ARCSTAT_INCR(arcstat_other_size, -space); 1802 break; 1803 case ARC_SPACE_HDRS: 1804 ARCSTAT_INCR(arcstat_hdr_size, -space); 1805 break; 1806 case ARC_SPACE_L2HDRS: 1807 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1808 break; 1809 } 1810 1811 if (type != ARC_SPACE_DATA) { 1812 ASSERT(arc_meta_used >= space); 1813 if (arc_meta_max < arc_meta_used) 1814 arc_meta_max = arc_meta_used; 1815 ARCSTAT_INCR(arcstat_meta_used, -space); 1816 } 1817 1818 ASSERT(arc_size >= space); 1819 atomic_add_64(&arc_size, -space); 1820 } 1821 1822 arc_buf_t * 1823 arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 1824 { 1825 arc_buf_hdr_t *hdr; 1826 arc_buf_t *buf; 1827 1828 ASSERT3U(size, >, 0); 1829 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 1830 ASSERT(BUF_EMPTY(hdr)); 1831 ASSERT3P(hdr->b_freeze_cksum, ==, NULL); 1832 hdr->b_size = size; 1833 hdr->b_spa = spa_load_guid(spa); 1834 1835 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1836 buf->b_hdr = hdr; 1837 buf->b_data = NULL; 1838 buf->b_efunc = NULL; 1839 buf->b_private = NULL; 1840 buf->b_next = NULL; 1841 1842 hdr->b_flags = arc_bufc_to_flags(type); 1843 hdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1844 1845 hdr->b_l1hdr.b_buf = buf; 1846 hdr->b_l1hdr.b_state = arc_anon; 1847 hdr->b_l1hdr.b_arc_access = 0; 1848 hdr->b_l1hdr.b_datacnt = 1; 1849 hdr->b_l1hdr.b_tmp_cdata = NULL; 1850 1851 arc_get_data_buf(buf); 1852 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 1853 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 1854 1855 return (buf); 1856 } 1857 1858 static char *arc_onloan_tag = "onloan"; 1859 1860 /* 1861 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1862 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1863 * buffers must be returned to the arc before they can be used by the DMU or 1864 * freed. 1865 */ 1866 arc_buf_t * 1867 arc_loan_buf(spa_t *spa, int size) 1868 { 1869 arc_buf_t *buf; 1870 1871 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1872 1873 atomic_add_64(&arc_loaned_bytes, size); 1874 return (buf); 1875 } 1876 1877 /* 1878 * Return a loaned arc buffer to the arc. 1879 */ 1880 void 1881 arc_return_buf(arc_buf_t *buf, void *tag) 1882 { 1883 arc_buf_hdr_t *hdr = buf->b_hdr; 1884 1885 ASSERT(buf->b_data != NULL); 1886 ASSERT(HDR_HAS_L1HDR(hdr)); 1887 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 1888 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 1889 1890 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1891 } 1892 1893 /* Detach an arc_buf from a dbuf (tag) */ 1894 void 1895 arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1896 { 1897 arc_buf_hdr_t *hdr = buf->b_hdr; 1898 1899 ASSERT(buf->b_data != NULL); 1900 ASSERT(HDR_HAS_L1HDR(hdr)); 1901 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 1902 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 1903 buf->b_efunc = NULL; 1904 buf->b_private = NULL; 1905 1906 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1907 } 1908 1909 static arc_buf_t * 1910 arc_buf_clone(arc_buf_t *from) 1911 { 1912 arc_buf_t *buf; 1913 arc_buf_hdr_t *hdr = from->b_hdr; 1914 uint64_t size = hdr->b_size; 1915 1916 ASSERT(HDR_HAS_L1HDR(hdr)); 1917 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 1918 1919 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1920 buf->b_hdr = hdr; 1921 buf->b_data = NULL; 1922 buf->b_efunc = NULL; 1923 buf->b_private = NULL; 1924 buf->b_next = hdr->b_l1hdr.b_buf; 1925 hdr->b_l1hdr.b_buf = buf; 1926 arc_get_data_buf(buf); 1927 bcopy(from->b_data, buf->b_data, size); 1928 1929 /* 1930 * This buffer already exists in the arc so create a duplicate 1931 * copy for the caller. If the buffer is associated with user data 1932 * then track the size and number of duplicates. These stats will be 1933 * updated as duplicate buffers are created and destroyed. 1934 */ 1935 if (HDR_ISTYPE_DATA(hdr)) { 1936 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1937 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1938 } 1939 hdr->b_l1hdr.b_datacnt += 1; 1940 return (buf); 1941 } 1942 1943 void 1944 arc_buf_add_ref(arc_buf_t *buf, void* tag) 1945 { 1946 arc_buf_hdr_t *hdr; 1947 kmutex_t *hash_lock; 1948 1949 /* 1950 * Check to see if this buffer is evicted. Callers 1951 * must verify b_data != NULL to know if the add_ref 1952 * was successful. 1953 */ 1954 mutex_enter(&buf->b_evict_lock); 1955 if (buf->b_data == NULL) { 1956 mutex_exit(&buf->b_evict_lock); 1957 return; 1958 } 1959 hash_lock = HDR_LOCK(buf->b_hdr); 1960 mutex_enter(hash_lock); 1961 hdr = buf->b_hdr; 1962 ASSERT(HDR_HAS_L1HDR(hdr)); 1963 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1964 mutex_exit(&buf->b_evict_lock); 1965 1966 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 1967 hdr->b_l1hdr.b_state == arc_mfu); 1968 1969 add_reference(hdr, hash_lock, tag); 1970 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1971 arc_access(hdr, hash_lock); 1972 mutex_exit(hash_lock); 1973 ARCSTAT_BUMP(arcstat_hits); 1974 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 1975 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 1976 data, metadata, hits); 1977 } 1978 1979 static void 1980 arc_buf_free_on_write(void *data, size_t size, 1981 void (*free_func)(void *, size_t)) 1982 { 1983 l2arc_data_free_t *df; 1984 1985 df = kmem_alloc(sizeof (*df), KM_SLEEP); 1986 df->l2df_data = data; 1987 df->l2df_size = size; 1988 df->l2df_func = free_func; 1989 mutex_enter(&l2arc_free_on_write_mtx); 1990 list_insert_head(l2arc_free_on_write, df); 1991 mutex_exit(&l2arc_free_on_write_mtx); 1992 } 1993 1994 /* 1995 * Free the arc data buffer. If it is an l2arc write in progress, 1996 * the buffer is placed on l2arc_free_on_write to be freed later. 1997 */ 1998 static void 1999 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 2000 { 2001 arc_buf_hdr_t *hdr = buf->b_hdr; 2002 2003 if (HDR_L2_WRITING(hdr)) { 2004 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 2005 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2006 } else { 2007 free_func(buf->b_data, hdr->b_size); 2008 } 2009 } 2010 2011 static void 2012 arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 2013 { 2014 ASSERT(HDR_HAS_L2HDR(hdr)); 2015 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx)); 2016 2017 /* 2018 * The b_tmp_cdata field is linked off of the b_l1hdr, so if 2019 * that doesn't exist, the header is in the arc_l2c_only state, 2020 * and there isn't anything to free (it's already been freed). 2021 */ 2022 if (!HDR_HAS_L1HDR(hdr)) 2023 return; 2024 2025 /* 2026 * The header isn't being written to the l2arc device, thus it 2027 * shouldn't have a b_tmp_cdata to free. 2028 */ 2029 if (!HDR_L2_WRITING(hdr)) { 2030 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2031 return; 2032 } 2033 2034 /* 2035 * The header does not have compression enabled. This can be due 2036 * to the buffer not being compressible, or because we're 2037 * freeing the buffer before the second phase of 2038 * l2arc_write_buffer() has started (which does the compression 2039 * step). In either case, b_tmp_cdata does not point to a 2040 * separately compressed buffer, so there's nothing to free (it 2041 * points to the same buffer as the arc_buf_t's b_data field). 2042 */ 2043 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_OFF) { 2044 hdr->b_l1hdr.b_tmp_cdata = NULL; 2045 return; 2046 } 2047 2048 /* 2049 * There's nothing to free since the buffer was all zero's and 2050 * compressed to a zero length buffer. 2051 */ 2052 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_EMPTY) { 2053 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2054 return; 2055 } 2056 2057 ASSERT(L2ARC_IS_VALID_COMPRESS(hdr->b_l2hdr.b_compress)); 2058 2059 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, 2060 hdr->b_size, zio_data_buf_free); 2061 2062 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 2063 hdr->b_l1hdr.b_tmp_cdata = NULL; 2064 } 2065 2066 /* 2067 * Free up buf->b_data and if 'remove' is set, then pull the 2068 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 2069 */ 2070 static void 2071 arc_buf_destroy(arc_buf_t *buf, boolean_t remove) 2072 { 2073 arc_buf_t **bufp; 2074 2075 /* free up data associated with the buf */ 2076 if (buf->b_data != NULL) { 2077 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 2078 uint64_t size = buf->b_hdr->b_size; 2079 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 2080 2081 arc_cksum_verify(buf); 2082 arc_buf_unwatch(buf); 2083 2084 if (type == ARC_BUFC_METADATA) { 2085 arc_buf_data_free(buf, zio_buf_free); 2086 arc_space_return(size, ARC_SPACE_META); 2087 } else { 2088 ASSERT(type == ARC_BUFC_DATA); 2089 arc_buf_data_free(buf, zio_data_buf_free); 2090 arc_space_return(size, ARC_SPACE_DATA); 2091 } 2092 2093 /* protected by hash lock, if in the hash table */ 2094 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) { 2095 uint64_t *cnt = &state->arcs_lsize[type]; 2096 2097 ASSERT(refcount_is_zero( 2098 &buf->b_hdr->b_l1hdr.b_refcnt)); 2099 ASSERT(state != arc_anon && state != arc_l2c_only); 2100 2101 ASSERT3U(*cnt, >=, size); 2102 atomic_add_64(cnt, -size); 2103 } 2104 2105 (void) refcount_remove_many(&state->arcs_size, size, buf); 2106 buf->b_data = NULL; 2107 2108 /* 2109 * If we're destroying a duplicate buffer make sure 2110 * that the appropriate statistics are updated. 2111 */ 2112 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 && 2113 HDR_ISTYPE_DATA(buf->b_hdr)) { 2114 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 2115 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 2116 } 2117 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0); 2118 buf->b_hdr->b_l1hdr.b_datacnt -= 1; 2119 } 2120 2121 /* only remove the buf if requested */ 2122 if (!remove) 2123 return; 2124 2125 /* remove the buf from the hdr list */ 2126 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf; 2127 bufp = &(*bufp)->b_next) 2128 continue; 2129 *bufp = buf->b_next; 2130 buf->b_next = NULL; 2131 2132 ASSERT(buf->b_efunc == NULL); 2133 2134 /* clean up the buf */ 2135 buf->b_hdr = NULL; 2136 kmem_cache_free(buf_cache, buf); 2137 } 2138 2139 static void 2140 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 2141 { 2142 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2143 l2arc_dev_t *dev = l2hdr->b_dev; 2144 2145 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 2146 ASSERT(HDR_HAS_L2HDR(hdr)); 2147 2148 list_remove(&dev->l2ad_buflist, hdr); 2149 2150 /* 2151 * We don't want to leak the b_tmp_cdata buffer that was 2152 * allocated in l2arc_write_buffers() 2153 */ 2154 arc_buf_l2_cdata_free(hdr); 2155 2156 /* 2157 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then 2158 * this header is being processed by l2arc_write_buffers() (i.e. 2159 * it's in the first stage of l2arc_write_buffers()). 2160 * Re-affirming that truth here, just to serve as a reminder. If 2161 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or 2162 * may not have its HDR_L2_WRITING flag set. (the write may have 2163 * completed, in which case HDR_L2_WRITING will be false and the 2164 * b_daddr field will point to the address of the buffer on disk). 2165 */ 2166 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr)); 2167 2168 /* 2169 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with 2170 * l2arc_write_buffers(). Since we've just removed this header 2171 * from the l2arc buffer list, this header will never reach the 2172 * second stage of l2arc_write_buffers(), which increments the 2173 * accounting stats for this header. Thus, we must be careful 2174 * not to decrement them for this header either. 2175 */ 2176 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) { 2177 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 2178 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 2179 2180 vdev_space_update(dev->l2ad_vdev, 2181 -l2hdr->b_asize, 0, 0); 2182 2183 (void) refcount_remove_many(&dev->l2ad_alloc, 2184 l2hdr->b_asize, hdr); 2185 } 2186 2187 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 2188 } 2189 2190 static void 2191 arc_hdr_destroy(arc_buf_hdr_t *hdr) 2192 { 2193 if (HDR_HAS_L1HDR(hdr)) { 2194 ASSERT(hdr->b_l1hdr.b_buf == NULL || 2195 hdr->b_l1hdr.b_datacnt > 0); 2196 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2197 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2198 } 2199 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2200 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 2201 2202 if (HDR_HAS_L2HDR(hdr)) { 2203 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2204 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 2205 2206 if (!buflist_held) 2207 mutex_enter(&dev->l2ad_mtx); 2208 2209 /* 2210 * Even though we checked this conditional above, we 2211 * need to check this again now that we have the 2212 * l2ad_mtx. This is because we could be racing with 2213 * another thread calling l2arc_evict() which might have 2214 * destroyed this header's L2 portion as we were waiting 2215 * to acquire the l2ad_mtx. If that happens, we don't 2216 * want to re-destroy the header's L2 portion. 2217 */ 2218 if (HDR_HAS_L2HDR(hdr)) 2219 arc_hdr_l2hdr_destroy(hdr); 2220 2221 if (!buflist_held) 2222 mutex_exit(&dev->l2ad_mtx); 2223 } 2224 2225 if (!BUF_EMPTY(hdr)) 2226 buf_discard_identity(hdr); 2227 2228 if (hdr->b_freeze_cksum != NULL) { 2229 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 2230 hdr->b_freeze_cksum = NULL; 2231 } 2232 2233 if (HDR_HAS_L1HDR(hdr)) { 2234 while (hdr->b_l1hdr.b_buf) { 2235 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2236 2237 if (buf->b_efunc != NULL) { 2238 mutex_enter(&arc_user_evicts_lock); 2239 mutex_enter(&buf->b_evict_lock); 2240 ASSERT(buf->b_hdr != NULL); 2241 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE); 2242 hdr->b_l1hdr.b_buf = buf->b_next; 2243 buf->b_hdr = &arc_eviction_hdr; 2244 buf->b_next = arc_eviction_list; 2245 arc_eviction_list = buf; 2246 mutex_exit(&buf->b_evict_lock); 2247 cv_signal(&arc_user_evicts_cv); 2248 mutex_exit(&arc_user_evicts_lock); 2249 } else { 2250 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE); 2251 } 2252 } 2253 #ifdef ZFS_DEBUG 2254 if (hdr->b_l1hdr.b_thawed != NULL) { 2255 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2256 hdr->b_l1hdr.b_thawed = NULL; 2257 } 2258 #endif 2259 } 2260 2261 ASSERT3P(hdr->b_hash_next, ==, NULL); 2262 if (HDR_HAS_L1HDR(hdr)) { 2263 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2264 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 2265 kmem_cache_free(hdr_full_cache, hdr); 2266 } else { 2267 kmem_cache_free(hdr_l2only_cache, hdr); 2268 } 2269 } 2270 2271 void 2272 arc_buf_free(arc_buf_t *buf, void *tag) 2273 { 2274 arc_buf_hdr_t *hdr = buf->b_hdr; 2275 int hashed = hdr->b_l1hdr.b_state != arc_anon; 2276 2277 ASSERT(buf->b_efunc == NULL); 2278 ASSERT(buf->b_data != NULL); 2279 2280 if (hashed) { 2281 kmutex_t *hash_lock = HDR_LOCK(hdr); 2282 2283 mutex_enter(hash_lock); 2284 hdr = buf->b_hdr; 2285 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2286 2287 (void) remove_reference(hdr, hash_lock, tag); 2288 if (hdr->b_l1hdr.b_datacnt > 1) { 2289 arc_buf_destroy(buf, TRUE); 2290 } else { 2291 ASSERT(buf == hdr->b_l1hdr.b_buf); 2292 ASSERT(buf->b_efunc == NULL); 2293 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2294 } 2295 mutex_exit(hash_lock); 2296 } else if (HDR_IO_IN_PROGRESS(hdr)) { 2297 int destroy_hdr; 2298 /* 2299 * We are in the middle of an async write. Don't destroy 2300 * this buffer unless the write completes before we finish 2301 * decrementing the reference count. 2302 */ 2303 mutex_enter(&arc_user_evicts_lock); 2304 (void) remove_reference(hdr, NULL, tag); 2305 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2306 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 2307 mutex_exit(&arc_user_evicts_lock); 2308 if (destroy_hdr) 2309 arc_hdr_destroy(hdr); 2310 } else { 2311 if (remove_reference(hdr, NULL, tag) > 0) 2312 arc_buf_destroy(buf, TRUE); 2313 else 2314 arc_hdr_destroy(hdr); 2315 } 2316 } 2317 2318 boolean_t 2319 arc_buf_remove_ref(arc_buf_t *buf, void* tag) 2320 { 2321 arc_buf_hdr_t *hdr = buf->b_hdr; 2322 kmutex_t *hash_lock = HDR_LOCK(hdr); 2323 boolean_t no_callback = (buf->b_efunc == NULL); 2324 2325 if (hdr->b_l1hdr.b_state == arc_anon) { 2326 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 2327 arc_buf_free(buf, tag); 2328 return (no_callback); 2329 } 2330 2331 mutex_enter(hash_lock); 2332 hdr = buf->b_hdr; 2333 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 2334 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2335 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2336 ASSERT(buf->b_data != NULL); 2337 2338 (void) remove_reference(hdr, hash_lock, tag); 2339 if (hdr->b_l1hdr.b_datacnt > 1) { 2340 if (no_callback) 2341 arc_buf_destroy(buf, TRUE); 2342 } else if (no_callback) { 2343 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL); 2344 ASSERT(buf->b_efunc == NULL); 2345 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2346 } 2347 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 || 2348 refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2349 mutex_exit(hash_lock); 2350 return (no_callback); 2351 } 2352 2353 int32_t 2354 arc_buf_size(arc_buf_t *buf) 2355 { 2356 return (buf->b_hdr->b_size); 2357 } 2358 2359 /* 2360 * Called from the DMU to determine if the current buffer should be 2361 * evicted. In order to ensure proper locking, the eviction must be initiated 2362 * from the DMU. Return true if the buffer is associated with user data and 2363 * duplicate buffers still exist. 2364 */ 2365 boolean_t 2366 arc_buf_eviction_needed(arc_buf_t *buf) 2367 { 2368 arc_buf_hdr_t *hdr; 2369 boolean_t evict_needed = B_FALSE; 2370 2371 if (zfs_disable_dup_eviction) 2372 return (B_FALSE); 2373 2374 mutex_enter(&buf->b_evict_lock); 2375 hdr = buf->b_hdr; 2376 if (hdr == NULL) { 2377 /* 2378 * We are in arc_do_user_evicts(); let that function 2379 * perform the eviction. 2380 */ 2381 ASSERT(buf->b_data == NULL); 2382 mutex_exit(&buf->b_evict_lock); 2383 return (B_FALSE); 2384 } else if (buf->b_data == NULL) { 2385 /* 2386 * We have already been added to the arc eviction list; 2387 * recommend eviction. 2388 */ 2389 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2390 mutex_exit(&buf->b_evict_lock); 2391 return (B_TRUE); 2392 } 2393 2394 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr)) 2395 evict_needed = B_TRUE; 2396 2397 mutex_exit(&buf->b_evict_lock); 2398 return (evict_needed); 2399 } 2400 2401 /* 2402 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 2403 * state of the header is dependent on it's state prior to entering this 2404 * function. The following transitions are possible: 2405 * 2406 * - arc_mru -> arc_mru_ghost 2407 * - arc_mfu -> arc_mfu_ghost 2408 * - arc_mru_ghost -> arc_l2c_only 2409 * - arc_mru_ghost -> deleted 2410 * - arc_mfu_ghost -> arc_l2c_only 2411 * - arc_mfu_ghost -> deleted 2412 */ 2413 static int64_t 2414 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 2415 { 2416 arc_state_t *evicted_state, *state; 2417 int64_t bytes_evicted = 0; 2418 2419 ASSERT(MUTEX_HELD(hash_lock)); 2420 ASSERT(HDR_HAS_L1HDR(hdr)); 2421 2422 state = hdr->b_l1hdr.b_state; 2423 if (GHOST_STATE(state)) { 2424 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2425 ASSERT(hdr->b_l1hdr.b_buf == NULL); 2426 2427 /* 2428 * l2arc_write_buffers() relies on a header's L1 portion 2429 * (i.e. it's b_tmp_cdata field) during it's write phase. 2430 * Thus, we cannot push a header onto the arc_l2c_only 2431 * state (removing it's L1 piece) until the header is 2432 * done being written to the l2arc. 2433 */ 2434 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 2435 ARCSTAT_BUMP(arcstat_evict_l2_skip); 2436 return (bytes_evicted); 2437 } 2438 2439 ARCSTAT_BUMP(arcstat_deleted); 2440 bytes_evicted += hdr->b_size; 2441 2442 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2443 2444 if (HDR_HAS_L2HDR(hdr)) { 2445 /* 2446 * This buffer is cached on the 2nd Level ARC; 2447 * don't destroy the header. 2448 */ 2449 arc_change_state(arc_l2c_only, hdr, hash_lock); 2450 /* 2451 * dropping from L1+L2 cached to L2-only, 2452 * realloc to remove the L1 header. 2453 */ 2454 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 2455 hdr_l2only_cache); 2456 } else { 2457 arc_change_state(arc_anon, hdr, hash_lock); 2458 arc_hdr_destroy(hdr); 2459 } 2460 return (bytes_evicted); 2461 } 2462 2463 ASSERT(state == arc_mru || state == arc_mfu); 2464 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2465 2466 /* prefetch buffers have a minimum lifespan */ 2467 if (HDR_IO_IN_PROGRESS(hdr) || 2468 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 2469 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 2470 arc_min_prefetch_lifespan)) { 2471 ARCSTAT_BUMP(arcstat_evict_skip); 2472 return (bytes_evicted); 2473 } 2474 2475 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 2476 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0); 2477 while (hdr->b_l1hdr.b_buf) { 2478 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2479 if (!mutex_tryenter(&buf->b_evict_lock)) { 2480 ARCSTAT_BUMP(arcstat_mutex_miss); 2481 break; 2482 } 2483 if (buf->b_data != NULL) 2484 bytes_evicted += hdr->b_size; 2485 if (buf->b_efunc != NULL) { 2486 mutex_enter(&arc_user_evicts_lock); 2487 arc_buf_destroy(buf, FALSE); 2488 hdr->b_l1hdr.b_buf = buf->b_next; 2489 buf->b_hdr = &arc_eviction_hdr; 2490 buf->b_next = arc_eviction_list; 2491 arc_eviction_list = buf; 2492 cv_signal(&arc_user_evicts_cv); 2493 mutex_exit(&arc_user_evicts_lock); 2494 mutex_exit(&buf->b_evict_lock); 2495 } else { 2496 mutex_exit(&buf->b_evict_lock); 2497 arc_buf_destroy(buf, TRUE); 2498 } 2499 } 2500 2501 if (HDR_HAS_L2HDR(hdr)) { 2502 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size); 2503 } else { 2504 if (l2arc_write_eligible(hdr->b_spa, UINT64_MAX, hdr)) 2505 ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size); 2506 else 2507 ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size); 2508 } 2509 2510 if (hdr->b_l1hdr.b_datacnt == 0) { 2511 arc_change_state(evicted_state, hdr, hash_lock); 2512 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2513 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2514 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2515 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2516 } 2517 2518 return (bytes_evicted); 2519 } 2520 2521 static uint64_t 2522 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 2523 uint64_t spa, int64_t bytes) 2524 { 2525 multilist_sublist_t *mls; 2526 uint64_t bytes_evicted = 0; 2527 arc_buf_hdr_t *hdr; 2528 kmutex_t *hash_lock; 2529 int evict_count = 0; 2530 2531 ASSERT3P(marker, !=, NULL); 2532 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2533 2534 mls = multilist_sublist_lock(ml, idx); 2535 2536 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 2537 hdr = multilist_sublist_prev(mls, marker)) { 2538 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 2539 (evict_count >= zfs_arc_evict_batch_limit)) 2540 break; 2541 2542 /* 2543 * To keep our iteration location, move the marker 2544 * forward. Since we're not holding hdr's hash lock, we 2545 * must be very careful and not remove 'hdr' from the 2546 * sublist. Otherwise, other consumers might mistake the 2547 * 'hdr' as not being on a sublist when they call the 2548 * multilist_link_active() function (they all rely on 2549 * the hash lock protecting concurrent insertions and 2550 * removals). multilist_sublist_move_forward() was 2551 * specifically implemented to ensure this is the case 2552 * (only 'marker' will be removed and re-inserted). 2553 */ 2554 multilist_sublist_move_forward(mls, marker); 2555 2556 /* 2557 * The only case where the b_spa field should ever be 2558 * zero, is the marker headers inserted by 2559 * arc_evict_state(). It's possible for multiple threads 2560 * to be calling arc_evict_state() concurrently (e.g. 2561 * dsl_pool_close() and zio_inject_fault()), so we must 2562 * skip any markers we see from these other threads. 2563 */ 2564 if (hdr->b_spa == 0) 2565 continue; 2566 2567 /* we're only interested in evicting buffers of a certain spa */ 2568 if (spa != 0 && hdr->b_spa != spa) { 2569 ARCSTAT_BUMP(arcstat_evict_skip); 2570 continue; 2571 } 2572 2573 hash_lock = HDR_LOCK(hdr); 2574 2575 /* 2576 * We aren't calling this function from any code path 2577 * that would already be holding a hash lock, so we're 2578 * asserting on this assumption to be defensive in case 2579 * this ever changes. Without this check, it would be 2580 * possible to incorrectly increment arcstat_mutex_miss 2581 * below (e.g. if the code changed such that we called 2582 * this function with a hash lock held). 2583 */ 2584 ASSERT(!MUTEX_HELD(hash_lock)); 2585 2586 if (mutex_tryenter(hash_lock)) { 2587 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 2588 mutex_exit(hash_lock); 2589 2590 bytes_evicted += evicted; 2591 2592 /* 2593 * If evicted is zero, arc_evict_hdr() must have 2594 * decided to skip this header, don't increment 2595 * evict_count in this case. 2596 */ 2597 if (evicted != 0) 2598 evict_count++; 2599 2600 /* 2601 * If arc_size isn't overflowing, signal any 2602 * threads that might happen to be waiting. 2603 * 2604 * For each header evicted, we wake up a single 2605 * thread. If we used cv_broadcast, we could 2606 * wake up "too many" threads causing arc_size 2607 * to significantly overflow arc_c; since 2608 * arc_get_data_buf() doesn't check for overflow 2609 * when it's woken up (it doesn't because it's 2610 * possible for the ARC to be overflowing while 2611 * full of un-evictable buffers, and the 2612 * function should proceed in this case). 2613 * 2614 * If threads are left sleeping, due to not 2615 * using cv_broadcast, they will be woken up 2616 * just before arc_reclaim_thread() sleeps. 2617 */ 2618 mutex_enter(&arc_reclaim_lock); 2619 if (!arc_is_overflowing()) 2620 cv_signal(&arc_reclaim_waiters_cv); 2621 mutex_exit(&arc_reclaim_lock); 2622 } else { 2623 ARCSTAT_BUMP(arcstat_mutex_miss); 2624 } 2625 } 2626 2627 multilist_sublist_unlock(mls); 2628 2629 return (bytes_evicted); 2630 } 2631 2632 /* 2633 * Evict buffers from the given arc state, until we've removed the 2634 * specified number of bytes. Move the removed buffers to the 2635 * appropriate evict state. 2636 * 2637 * This function makes a "best effort". It skips over any buffers 2638 * it can't get a hash_lock on, and so, may not catch all candidates. 2639 * It may also return without evicting as much space as requested. 2640 * 2641 * If bytes is specified using the special value ARC_EVICT_ALL, this 2642 * will evict all available (i.e. unlocked and evictable) buffers from 2643 * the given arc state; which is used by arc_flush(). 2644 */ 2645 static uint64_t 2646 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 2647 arc_buf_contents_t type) 2648 { 2649 uint64_t total_evicted = 0; 2650 multilist_t *ml = &state->arcs_list[type]; 2651 int num_sublists; 2652 arc_buf_hdr_t **markers; 2653 2654 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2655 2656 num_sublists = multilist_get_num_sublists(ml); 2657 2658 /* 2659 * If we've tried to evict from each sublist, made some 2660 * progress, but still have not hit the target number of bytes 2661 * to evict, we want to keep trying. The markers allow us to 2662 * pick up where we left off for each individual sublist, rather 2663 * than starting from the tail each time. 2664 */ 2665 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 2666 for (int i = 0; i < num_sublists; i++) { 2667 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 2668 2669 /* 2670 * A b_spa of 0 is used to indicate that this header is 2671 * a marker. This fact is used in arc_adjust_type() and 2672 * arc_evict_state_impl(). 2673 */ 2674 markers[i]->b_spa = 0; 2675 2676 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2677 multilist_sublist_insert_tail(mls, markers[i]); 2678 multilist_sublist_unlock(mls); 2679 } 2680 2681 /* 2682 * While we haven't hit our target number of bytes to evict, or 2683 * we're evicting all available buffers. 2684 */ 2685 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 2686 /* 2687 * Start eviction using a randomly selected sublist, 2688 * this is to try and evenly balance eviction across all 2689 * sublists. Always starting at the same sublist 2690 * (e.g. index 0) would cause evictions to favor certain 2691 * sublists over others. 2692 */ 2693 int sublist_idx = multilist_get_random_index(ml); 2694 uint64_t scan_evicted = 0; 2695 2696 for (int i = 0; i < num_sublists; i++) { 2697 uint64_t bytes_remaining; 2698 uint64_t bytes_evicted; 2699 2700 if (bytes == ARC_EVICT_ALL) 2701 bytes_remaining = ARC_EVICT_ALL; 2702 else if (total_evicted < bytes) 2703 bytes_remaining = bytes - total_evicted; 2704 else 2705 break; 2706 2707 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 2708 markers[sublist_idx], spa, bytes_remaining); 2709 2710 scan_evicted += bytes_evicted; 2711 total_evicted += bytes_evicted; 2712 2713 /* we've reached the end, wrap to the beginning */ 2714 if (++sublist_idx >= num_sublists) 2715 sublist_idx = 0; 2716 } 2717 2718 /* 2719 * If we didn't evict anything during this scan, we have 2720 * no reason to believe we'll evict more during another 2721 * scan, so break the loop. 2722 */ 2723 if (scan_evicted == 0) { 2724 /* This isn't possible, let's make that obvious */ 2725 ASSERT3S(bytes, !=, 0); 2726 2727 /* 2728 * When bytes is ARC_EVICT_ALL, the only way to 2729 * break the loop is when scan_evicted is zero. 2730 * In that case, we actually have evicted enough, 2731 * so we don't want to increment the kstat. 2732 */ 2733 if (bytes != ARC_EVICT_ALL) { 2734 ASSERT3S(total_evicted, <, bytes); 2735 ARCSTAT_BUMP(arcstat_evict_not_enough); 2736 } 2737 2738 break; 2739 } 2740 } 2741 2742 for (int i = 0; i < num_sublists; i++) { 2743 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2744 multilist_sublist_remove(mls, markers[i]); 2745 multilist_sublist_unlock(mls); 2746 2747 kmem_cache_free(hdr_full_cache, markers[i]); 2748 } 2749 kmem_free(markers, sizeof (*markers) * num_sublists); 2750 2751 return (total_evicted); 2752 } 2753 2754 /* 2755 * Flush all "evictable" data of the given type from the arc state 2756 * specified. This will not evict any "active" buffers (i.e. referenced). 2757 * 2758 * When 'retry' is set to FALSE, the function will make a single pass 2759 * over the state and evict any buffers that it can. Since it doesn't 2760 * continually retry the eviction, it might end up leaving some buffers 2761 * in the ARC due to lock misses. 2762 * 2763 * When 'retry' is set to TRUE, the function will continually retry the 2764 * eviction until *all* evictable buffers have been removed from the 2765 * state. As a result, if concurrent insertions into the state are 2766 * allowed (e.g. if the ARC isn't shutting down), this function might 2767 * wind up in an infinite loop, continually trying to evict buffers. 2768 */ 2769 static uint64_t 2770 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 2771 boolean_t retry) 2772 { 2773 uint64_t evicted = 0; 2774 2775 while (state->arcs_lsize[type] != 0) { 2776 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 2777 2778 if (!retry) 2779 break; 2780 } 2781 2782 return (evicted); 2783 } 2784 2785 /* 2786 * Evict the specified number of bytes from the state specified, 2787 * restricting eviction to the spa and type given. This function 2788 * prevents us from trying to evict more from a state's list than 2789 * is "evictable", and to skip evicting altogether when passed a 2790 * negative value for "bytes". In contrast, arc_evict_state() will 2791 * evict everything it can, when passed a negative value for "bytes". 2792 */ 2793 static uint64_t 2794 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 2795 arc_buf_contents_t type) 2796 { 2797 int64_t delta; 2798 2799 if (bytes > 0 && state->arcs_lsize[type] > 0) { 2800 delta = MIN(state->arcs_lsize[type], bytes); 2801 return (arc_evict_state(state, spa, delta, type)); 2802 } 2803 2804 return (0); 2805 } 2806 2807 /* 2808 * Evict metadata buffers from the cache, such that arc_meta_used is 2809 * capped by the arc_meta_limit tunable. 2810 */ 2811 static uint64_t 2812 arc_adjust_meta(void) 2813 { 2814 uint64_t total_evicted = 0; 2815 int64_t target; 2816 2817 /* 2818 * If we're over the meta limit, we want to evict enough 2819 * metadata to get back under the meta limit. We don't want to 2820 * evict so much that we drop the MRU below arc_p, though. If 2821 * we're over the meta limit more than we're over arc_p, we 2822 * evict some from the MRU here, and some from the MFU below. 2823 */ 2824 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 2825 (int64_t)(refcount_count(&arc_anon->arcs_size) + 2826 refcount_count(&arc_mru->arcs_size) - arc_p)); 2827 2828 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 2829 2830 /* 2831 * Similar to the above, we want to evict enough bytes to get us 2832 * below the meta limit, but not so much as to drop us below the 2833 * space alloted to the MFU (which is defined as arc_c - arc_p). 2834 */ 2835 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 2836 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 2837 2838 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 2839 2840 return (total_evicted); 2841 } 2842 2843 /* 2844 * Return the type of the oldest buffer in the given arc state 2845 * 2846 * This function will select a random sublist of type ARC_BUFC_DATA and 2847 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 2848 * is compared, and the type which contains the "older" buffer will be 2849 * returned. 2850 */ 2851 static arc_buf_contents_t 2852 arc_adjust_type(arc_state_t *state) 2853 { 2854 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; 2855 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; 2856 int data_idx = multilist_get_random_index(data_ml); 2857 int meta_idx = multilist_get_random_index(meta_ml); 2858 multilist_sublist_t *data_mls; 2859 multilist_sublist_t *meta_mls; 2860 arc_buf_contents_t type; 2861 arc_buf_hdr_t *data_hdr; 2862 arc_buf_hdr_t *meta_hdr; 2863 2864 /* 2865 * We keep the sublist lock until we're finished, to prevent 2866 * the headers from being destroyed via arc_evict_state(). 2867 */ 2868 data_mls = multilist_sublist_lock(data_ml, data_idx); 2869 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 2870 2871 /* 2872 * These two loops are to ensure we skip any markers that 2873 * might be at the tail of the lists due to arc_evict_state(). 2874 */ 2875 2876 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 2877 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 2878 if (data_hdr->b_spa != 0) 2879 break; 2880 } 2881 2882 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 2883 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 2884 if (meta_hdr->b_spa != 0) 2885 break; 2886 } 2887 2888 if (data_hdr == NULL && meta_hdr == NULL) { 2889 type = ARC_BUFC_DATA; 2890 } else if (data_hdr == NULL) { 2891 ASSERT3P(meta_hdr, !=, NULL); 2892 type = ARC_BUFC_METADATA; 2893 } else if (meta_hdr == NULL) { 2894 ASSERT3P(data_hdr, !=, NULL); 2895 type = ARC_BUFC_DATA; 2896 } else { 2897 ASSERT3P(data_hdr, !=, NULL); 2898 ASSERT3P(meta_hdr, !=, NULL); 2899 2900 /* The headers can't be on the sublist without an L1 header */ 2901 ASSERT(HDR_HAS_L1HDR(data_hdr)); 2902 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 2903 2904 if (data_hdr->b_l1hdr.b_arc_access < 2905 meta_hdr->b_l1hdr.b_arc_access) { 2906 type = ARC_BUFC_DATA; 2907 } else { 2908 type = ARC_BUFC_METADATA; 2909 } 2910 } 2911 2912 multilist_sublist_unlock(meta_mls); 2913 multilist_sublist_unlock(data_mls); 2914 2915 return (type); 2916 } 2917 2918 /* 2919 * Evict buffers from the cache, such that arc_size is capped by arc_c. 2920 */ 2921 static uint64_t 2922 arc_adjust(void) 2923 { 2924 uint64_t total_evicted = 0; 2925 uint64_t bytes; 2926 int64_t target; 2927 2928 /* 2929 * If we're over arc_meta_limit, we want to correct that before 2930 * potentially evicting data buffers below. 2931 */ 2932 total_evicted += arc_adjust_meta(); 2933 2934 /* 2935 * Adjust MRU size 2936 * 2937 * If we're over the target cache size, we want to evict enough 2938 * from the list to get back to our target size. We don't want 2939 * to evict too much from the MRU, such that it drops below 2940 * arc_p. So, if we're over our target cache size more than 2941 * the MRU is over arc_p, we'll evict enough to get back to 2942 * arc_p here, and then evict more from the MFU below. 2943 */ 2944 target = MIN((int64_t)(arc_size - arc_c), 2945 (int64_t)(refcount_count(&arc_anon->arcs_size) + 2946 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 2947 2948 /* 2949 * If we're below arc_meta_min, always prefer to evict data. 2950 * Otherwise, try to satisfy the requested number of bytes to 2951 * evict from the type which contains older buffers; in an 2952 * effort to keep newer buffers in the cache regardless of their 2953 * type. If we cannot satisfy the number of bytes from this 2954 * type, spill over into the next type. 2955 */ 2956 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 2957 arc_meta_used > arc_meta_min) { 2958 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 2959 total_evicted += bytes; 2960 2961 /* 2962 * If we couldn't evict our target number of bytes from 2963 * metadata, we try to get the rest from data. 2964 */ 2965 target -= bytes; 2966 2967 total_evicted += 2968 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 2969 } else { 2970 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 2971 total_evicted += bytes; 2972 2973 /* 2974 * If we couldn't evict our target number of bytes from 2975 * data, we try to get the rest from metadata. 2976 */ 2977 target -= bytes; 2978 2979 total_evicted += 2980 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 2981 } 2982 2983 /* 2984 * Adjust MFU size 2985 * 2986 * Now that we've tried to evict enough from the MRU to get its 2987 * size back to arc_p, if we're still above the target cache 2988 * size, we evict the rest from the MFU. 2989 */ 2990 target = arc_size - arc_c; 2991 2992 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 2993 arc_meta_used > arc_meta_min) { 2994 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 2995 total_evicted += bytes; 2996 2997 /* 2998 * If we couldn't evict our target number of bytes from 2999 * metadata, we try to get the rest from data. 3000 */ 3001 target -= bytes; 3002 3003 total_evicted += 3004 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3005 } else { 3006 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3007 total_evicted += bytes; 3008 3009 /* 3010 * If we couldn't evict our target number of bytes from 3011 * data, we try to get the rest from data. 3012 */ 3013 target -= bytes; 3014 3015 total_evicted += 3016 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3017 } 3018 3019 /* 3020 * Adjust ghost lists 3021 * 3022 * In addition to the above, the ARC also defines target values 3023 * for the ghost lists. The sum of the mru list and mru ghost 3024 * list should never exceed the target size of the cache, and 3025 * the sum of the mru list, mfu list, mru ghost list, and mfu 3026 * ghost list should never exceed twice the target size of the 3027 * cache. The following logic enforces these limits on the ghost 3028 * caches, and evicts from them as needed. 3029 */ 3030 target = refcount_count(&arc_mru->arcs_size) + 3031 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3032 3033 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3034 total_evicted += bytes; 3035 3036 target -= bytes; 3037 3038 total_evicted += 3039 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 3040 3041 /* 3042 * We assume the sum of the mru list and mfu list is less than 3043 * or equal to arc_c (we enforced this above), which means we 3044 * can use the simpler of the two equations below: 3045 * 3046 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3047 * mru ghost + mfu ghost <= arc_c 3048 */ 3049 target = refcount_count(&arc_mru_ghost->arcs_size) + 3050 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3051 3052 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3053 total_evicted += bytes; 3054 3055 target -= bytes; 3056 3057 total_evicted += 3058 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3059 3060 return (total_evicted); 3061 } 3062 3063 static void 3064 arc_do_user_evicts(void) 3065 { 3066 mutex_enter(&arc_user_evicts_lock); 3067 while (arc_eviction_list != NULL) { 3068 arc_buf_t *buf = arc_eviction_list; 3069 arc_eviction_list = buf->b_next; 3070 mutex_enter(&buf->b_evict_lock); 3071 buf->b_hdr = NULL; 3072 mutex_exit(&buf->b_evict_lock); 3073 mutex_exit(&arc_user_evicts_lock); 3074 3075 if (buf->b_efunc != NULL) 3076 VERIFY0(buf->b_efunc(buf->b_private)); 3077 3078 buf->b_efunc = NULL; 3079 buf->b_private = NULL; 3080 kmem_cache_free(buf_cache, buf); 3081 mutex_enter(&arc_user_evicts_lock); 3082 } 3083 mutex_exit(&arc_user_evicts_lock); 3084 } 3085 3086 void 3087 arc_flush(spa_t *spa, boolean_t retry) 3088 { 3089 uint64_t guid = 0; 3090 3091 /* 3092 * If retry is TRUE, a spa must not be specified since we have 3093 * no good way to determine if all of a spa's buffers have been 3094 * evicted from an arc state. 3095 */ 3096 ASSERT(!retry || spa == 0); 3097 3098 if (spa != NULL) 3099 guid = spa_load_guid(spa); 3100 3101 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3102 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3103 3104 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3105 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3106 3107 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3108 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3109 3110 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3111 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3112 3113 arc_do_user_evicts(); 3114 ASSERT(spa || arc_eviction_list == NULL); 3115 } 3116 3117 void 3118 arc_shrink(int64_t to_free) 3119 { 3120 if (arc_c > arc_c_min) { 3121 3122 if (arc_c > arc_c_min + to_free) 3123 atomic_add_64(&arc_c, -to_free); 3124 else 3125 arc_c = arc_c_min; 3126 3127 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3128 if (arc_c > arc_size) 3129 arc_c = MAX(arc_size, arc_c_min); 3130 if (arc_p > arc_c) 3131 arc_p = (arc_c >> 1); 3132 ASSERT(arc_c >= arc_c_min); 3133 ASSERT((int64_t)arc_p >= 0); 3134 } 3135 3136 if (arc_size > arc_c) 3137 (void) arc_adjust(); 3138 } 3139 3140 typedef enum free_memory_reason_t { 3141 FMR_UNKNOWN, 3142 FMR_NEEDFREE, 3143 FMR_LOTSFREE, 3144 FMR_SWAPFS_MINFREE, 3145 FMR_PAGES_PP_MAXIMUM, 3146 FMR_HEAP_ARENA, 3147 FMR_ZIO_ARENA, 3148 } free_memory_reason_t; 3149 3150 int64_t last_free_memory; 3151 free_memory_reason_t last_free_reason; 3152 3153 /* 3154 * Additional reserve of pages for pp_reserve. 3155 */ 3156 int64_t arc_pages_pp_reserve = 64; 3157 3158 /* 3159 * Additional reserve of pages for swapfs. 3160 */ 3161 int64_t arc_swapfs_reserve = 64; 3162 3163 /* 3164 * Return the amount of memory that can be consumed before reclaim will be 3165 * needed. Positive if there is sufficient free memory, negative indicates 3166 * the amount of memory that needs to be freed up. 3167 */ 3168 static int64_t 3169 arc_available_memory(void) 3170 { 3171 int64_t lowest = INT64_MAX; 3172 int64_t n; 3173 free_memory_reason_t r = FMR_UNKNOWN; 3174 3175 #ifdef _KERNEL 3176 if (needfree > 0) { 3177 n = PAGESIZE * (-needfree); 3178 if (n < lowest) { 3179 lowest = n; 3180 r = FMR_NEEDFREE; 3181 } 3182 } 3183 3184 /* 3185 * check that we're out of range of the pageout scanner. It starts to 3186 * schedule paging if freemem is less than lotsfree and needfree. 3187 * lotsfree is the high-water mark for pageout, and needfree is the 3188 * number of needed free pages. We add extra pages here to make sure 3189 * the scanner doesn't start up while we're freeing memory. 3190 */ 3191 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 3192 if (n < lowest) { 3193 lowest = n; 3194 r = FMR_LOTSFREE; 3195 } 3196 3197 /* 3198 * check to make sure that swapfs has enough space so that anon 3199 * reservations can still succeed. anon_resvmem() checks that the 3200 * availrmem is greater than swapfs_minfree, and the number of reserved 3201 * swap pages. We also add a bit of extra here just to prevent 3202 * circumstances from getting really dire. 3203 */ 3204 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 3205 desfree - arc_swapfs_reserve); 3206 if (n < lowest) { 3207 lowest = n; 3208 r = FMR_SWAPFS_MINFREE; 3209 } 3210 3211 3212 /* 3213 * Check that we have enough availrmem that memory locking (e.g., via 3214 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3215 * stores the number of pages that cannot be locked; when availrmem 3216 * drops below pages_pp_maximum, page locking mechanisms such as 3217 * page_pp_lock() will fail.) 3218 */ 3219 n = PAGESIZE * (availrmem - pages_pp_maximum - 3220 arc_pages_pp_reserve); 3221 if (n < lowest) { 3222 lowest = n; 3223 r = FMR_PAGES_PP_MAXIMUM; 3224 } 3225 3226 #if defined(__i386) 3227 /* 3228 * If we're on an i386 platform, it's possible that we'll exhaust the 3229 * kernel heap space before we ever run out of available physical 3230 * memory. Most checks of the size of the heap_area compare against 3231 * tune.t_minarmem, which is the minimum available real memory that we 3232 * can have in the system. However, this is generally fixed at 25 pages 3233 * which is so low that it's useless. In this comparison, we seek to 3234 * calculate the total heap-size, and reclaim if more than 3/4ths of the 3235 * heap is allocated. (Or, in the calculation, if less than 1/4th is 3236 * free) 3237 */ 3238 n = vmem_size(heap_arena, VMEM_FREE) - 3239 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 3240 if (n < lowest) { 3241 lowest = n; 3242 r = FMR_HEAP_ARENA; 3243 } 3244 #endif 3245 3246 /* 3247 * If zio data pages are being allocated out of a separate heap segment, 3248 * then enforce that the size of available vmem for this arena remains 3249 * above about 1/16th free. 3250 * 3251 * Note: The 1/16th arena free requirement was put in place 3252 * to aggressively evict memory from the arc in order to avoid 3253 * memory fragmentation issues. 3254 */ 3255 if (zio_arena != NULL) { 3256 n = vmem_size(zio_arena, VMEM_FREE) - 3257 (vmem_size(zio_arena, VMEM_ALLOC) >> 4); 3258 if (n < lowest) { 3259 lowest = n; 3260 r = FMR_ZIO_ARENA; 3261 } 3262 } 3263 #else 3264 /* Every 100 calls, free a small amount */ 3265 if (spa_get_random(100) == 0) 3266 lowest = -1024; 3267 #endif 3268 3269 last_free_memory = lowest; 3270 last_free_reason = r; 3271 3272 return (lowest); 3273 } 3274 3275 3276 /* 3277 * Determine if the system is under memory pressure and is asking 3278 * to reclaim memory. A return value of TRUE indicates that the system 3279 * is under memory pressure and that the arc should adjust accordingly. 3280 */ 3281 static boolean_t 3282 arc_reclaim_needed(void) 3283 { 3284 return (arc_available_memory() < 0); 3285 } 3286 3287 static void 3288 arc_kmem_reap_now(void) 3289 { 3290 size_t i; 3291 kmem_cache_t *prev_cache = NULL; 3292 kmem_cache_t *prev_data_cache = NULL; 3293 extern kmem_cache_t *zio_buf_cache[]; 3294 extern kmem_cache_t *zio_data_buf_cache[]; 3295 extern kmem_cache_t *range_seg_cache; 3296 3297 #ifdef _KERNEL 3298 if (arc_meta_used >= arc_meta_limit) { 3299 /* 3300 * We are exceeding our meta-data cache limit. 3301 * Purge some DNLC entries to release holds on meta-data. 3302 */ 3303 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 3304 } 3305 #if defined(__i386) 3306 /* 3307 * Reclaim unused memory from all kmem caches. 3308 */ 3309 kmem_reap(); 3310 #endif 3311 #endif 3312 3313 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 3314 if (zio_buf_cache[i] != prev_cache) { 3315 prev_cache = zio_buf_cache[i]; 3316 kmem_cache_reap_now(zio_buf_cache[i]); 3317 } 3318 if (zio_data_buf_cache[i] != prev_data_cache) { 3319 prev_data_cache = zio_data_buf_cache[i]; 3320 kmem_cache_reap_now(zio_data_buf_cache[i]); 3321 } 3322 } 3323 kmem_cache_reap_now(buf_cache); 3324 kmem_cache_reap_now(hdr_full_cache); 3325 kmem_cache_reap_now(hdr_l2only_cache); 3326 kmem_cache_reap_now(range_seg_cache); 3327 3328 if (zio_arena != NULL) { 3329 /* 3330 * Ask the vmem arena to reclaim unused memory from its 3331 * quantum caches. 3332 */ 3333 vmem_qcache_reap(zio_arena); 3334 } 3335 } 3336 3337 /* 3338 * Threads can block in arc_get_data_buf() waiting for this thread to evict 3339 * enough data and signal them to proceed. When this happens, the threads in 3340 * arc_get_data_buf() are sleeping while holding the hash lock for their 3341 * particular arc header. Thus, we must be careful to never sleep on a 3342 * hash lock in this thread. This is to prevent the following deadlock: 3343 * 3344 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", 3345 * waiting for the reclaim thread to signal it. 3346 * 3347 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 3348 * fails, and goes to sleep forever. 3349 * 3350 * This possible deadlock is avoided by always acquiring a hash lock 3351 * using mutex_tryenter() from arc_reclaim_thread(). 3352 */ 3353 static void 3354 arc_reclaim_thread(void) 3355 { 3356 hrtime_t growtime = 0; 3357 callb_cpr_t cpr; 3358 3359 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 3360 3361 mutex_enter(&arc_reclaim_lock); 3362 while (!arc_reclaim_thread_exit) { 3363 int64_t free_memory = arc_available_memory(); 3364 uint64_t evicted = 0; 3365 3366 mutex_exit(&arc_reclaim_lock); 3367 3368 if (free_memory < 0) { 3369 3370 arc_no_grow = B_TRUE; 3371 arc_warm = B_TRUE; 3372 3373 /* 3374 * Wait at least zfs_grow_retry (default 60) seconds 3375 * before considering growing. 3376 */ 3377 growtime = gethrtime() + SEC2NSEC(arc_grow_retry); 3378 3379 arc_kmem_reap_now(); 3380 3381 /* 3382 * If we are still low on memory, shrink the ARC 3383 * so that we have arc_shrink_min free space. 3384 */ 3385 free_memory = arc_available_memory(); 3386 3387 int64_t to_free = 3388 (arc_c >> arc_shrink_shift) - free_memory; 3389 if (to_free > 0) { 3390 #ifdef _KERNEL 3391 to_free = MAX(to_free, ptob(needfree)); 3392 #endif 3393 arc_shrink(to_free); 3394 } 3395 } else if (free_memory < arc_c >> arc_no_grow_shift) { 3396 arc_no_grow = B_TRUE; 3397 } else if (gethrtime() >= growtime) { 3398 arc_no_grow = B_FALSE; 3399 } 3400 3401 evicted = arc_adjust(); 3402 3403 mutex_enter(&arc_reclaim_lock); 3404 3405 /* 3406 * If evicted is zero, we couldn't evict anything via 3407 * arc_adjust(). This could be due to hash lock 3408 * collisions, but more likely due to the majority of 3409 * arc buffers being unevictable. Therefore, even if 3410 * arc_size is above arc_c, another pass is unlikely to 3411 * be helpful and could potentially cause us to enter an 3412 * infinite loop. 3413 */ 3414 if (arc_size <= arc_c || evicted == 0) { 3415 /* 3416 * We're either no longer overflowing, or we 3417 * can't evict anything more, so we should wake 3418 * up any threads before we go to sleep. 3419 */ 3420 cv_broadcast(&arc_reclaim_waiters_cv); 3421 3422 /* 3423 * Block until signaled, or after one second (we 3424 * might need to perform arc_kmem_reap_now() 3425 * even if we aren't being signalled) 3426 */ 3427 CALLB_CPR_SAFE_BEGIN(&cpr); 3428 (void) cv_timedwait_hires(&arc_reclaim_thread_cv, 3429 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 3430 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 3431 } 3432 } 3433 3434 arc_reclaim_thread_exit = FALSE; 3435 cv_broadcast(&arc_reclaim_thread_cv); 3436 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 3437 thread_exit(); 3438 } 3439 3440 static void 3441 arc_user_evicts_thread(void) 3442 { 3443 callb_cpr_t cpr; 3444 3445 CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG); 3446 3447 mutex_enter(&arc_user_evicts_lock); 3448 while (!arc_user_evicts_thread_exit) { 3449 mutex_exit(&arc_user_evicts_lock); 3450 3451 arc_do_user_evicts(); 3452 3453 /* 3454 * This is necessary in order for the mdb ::arc dcmd to 3455 * show up to date information. Since the ::arc command 3456 * does not call the kstat's update function, without 3457 * this call, the command may show stale stats for the 3458 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 3459 * with this change, the data might be up to 1 second 3460 * out of date; but that should suffice. The arc_state_t 3461 * structures can be queried directly if more accurate 3462 * information is needed. 3463 */ 3464 if (arc_ksp != NULL) 3465 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 3466 3467 mutex_enter(&arc_user_evicts_lock); 3468 3469 /* 3470 * Block until signaled, or after one second (we need to 3471 * call the arc's kstat update function regularly). 3472 */ 3473 CALLB_CPR_SAFE_BEGIN(&cpr); 3474 (void) cv_timedwait(&arc_user_evicts_cv, 3475 &arc_user_evicts_lock, ddi_get_lbolt() + hz); 3476 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock); 3477 } 3478 3479 arc_user_evicts_thread_exit = FALSE; 3480 cv_broadcast(&arc_user_evicts_cv); 3481 CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */ 3482 thread_exit(); 3483 } 3484 3485 /* 3486 * Adapt arc info given the number of bytes we are trying to add and 3487 * the state that we are comming from. This function is only called 3488 * when we are adding new content to the cache. 3489 */ 3490 static void 3491 arc_adapt(int bytes, arc_state_t *state) 3492 { 3493 int mult; 3494 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 3495 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 3496 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 3497 3498 if (state == arc_l2c_only) 3499 return; 3500 3501 ASSERT(bytes > 0); 3502 /* 3503 * Adapt the target size of the MRU list: 3504 * - if we just hit in the MRU ghost list, then increase 3505 * the target size of the MRU list. 3506 * - if we just hit in the MFU ghost list, then increase 3507 * the target size of the MFU list by decreasing the 3508 * target size of the MRU list. 3509 */ 3510 if (state == arc_mru_ghost) { 3511 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 3512 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 3513 3514 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 3515 } else if (state == arc_mfu_ghost) { 3516 uint64_t delta; 3517 3518 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 3519 mult = MIN(mult, 10); 3520 3521 delta = MIN(bytes * mult, arc_p); 3522 arc_p = MAX(arc_p_min, arc_p - delta); 3523 } 3524 ASSERT((int64_t)arc_p >= 0); 3525 3526 if (arc_reclaim_needed()) { 3527 cv_signal(&arc_reclaim_thread_cv); 3528 return; 3529 } 3530 3531 if (arc_no_grow) 3532 return; 3533 3534 if (arc_c >= arc_c_max) 3535 return; 3536 3537 /* 3538 * If we're within (2 * maxblocksize) bytes of the target 3539 * cache size, increment the target cache size 3540 */ 3541 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 3542 atomic_add_64(&arc_c, (int64_t)bytes); 3543 if (arc_c > arc_c_max) 3544 arc_c = arc_c_max; 3545 else if (state == arc_anon) 3546 atomic_add_64(&arc_p, (int64_t)bytes); 3547 if (arc_p > arc_c) 3548 arc_p = arc_c; 3549 } 3550 ASSERT((int64_t)arc_p >= 0); 3551 } 3552 3553 /* 3554 * Check if arc_size has grown past our upper threshold, determined by 3555 * zfs_arc_overflow_shift. 3556 */ 3557 static boolean_t 3558 arc_is_overflowing(void) 3559 { 3560 /* Always allow at least one block of overflow */ 3561 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 3562 arc_c >> zfs_arc_overflow_shift); 3563 3564 return (arc_size >= arc_c + overflow); 3565 } 3566 3567 /* 3568 * The buffer, supplied as the first argument, needs a data block. If we 3569 * are hitting the hard limit for the cache size, we must sleep, waiting 3570 * for the eviction thread to catch up. If we're past the target size 3571 * but below the hard limit, we'll only signal the reclaim thread and 3572 * continue on. 3573 */ 3574 static void 3575 arc_get_data_buf(arc_buf_t *buf) 3576 { 3577 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 3578 uint64_t size = buf->b_hdr->b_size; 3579 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 3580 3581 arc_adapt(size, state); 3582 3583 /* 3584 * If arc_size is currently overflowing, and has grown past our 3585 * upper limit, we must be adding data faster than the evict 3586 * thread can evict. Thus, to ensure we don't compound the 3587 * problem by adding more data and forcing arc_size to grow even 3588 * further past it's target size, we halt and wait for the 3589 * eviction thread to catch up. 3590 * 3591 * It's also possible that the reclaim thread is unable to evict 3592 * enough buffers to get arc_size below the overflow limit (e.g. 3593 * due to buffers being un-evictable, or hash lock collisions). 3594 * In this case, we want to proceed regardless if we're 3595 * overflowing; thus we don't use a while loop here. 3596 */ 3597 if (arc_is_overflowing()) { 3598 mutex_enter(&arc_reclaim_lock); 3599 3600 /* 3601 * Now that we've acquired the lock, we may no longer be 3602 * over the overflow limit, lets check. 3603 * 3604 * We're ignoring the case of spurious wake ups. If that 3605 * were to happen, it'd let this thread consume an ARC 3606 * buffer before it should have (i.e. before we're under 3607 * the overflow limit and were signalled by the reclaim 3608 * thread). As long as that is a rare occurrence, it 3609 * shouldn't cause any harm. 3610 */ 3611 if (arc_is_overflowing()) { 3612 cv_signal(&arc_reclaim_thread_cv); 3613 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 3614 } 3615 3616 mutex_exit(&arc_reclaim_lock); 3617 } 3618 3619 if (type == ARC_BUFC_METADATA) { 3620 buf->b_data = zio_buf_alloc(size); 3621 arc_space_consume(size, ARC_SPACE_META); 3622 } else { 3623 ASSERT(type == ARC_BUFC_DATA); 3624 buf->b_data = zio_data_buf_alloc(size); 3625 arc_space_consume(size, ARC_SPACE_DATA); 3626 } 3627 3628 /* 3629 * Update the state size. Note that ghost states have a 3630 * "ghost size" and so don't need to be updated. 3631 */ 3632 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) { 3633 arc_buf_hdr_t *hdr = buf->b_hdr; 3634 arc_state_t *state = hdr->b_l1hdr.b_state; 3635 3636 (void) refcount_add_many(&state->arcs_size, size, buf); 3637 3638 /* 3639 * If this is reached via arc_read, the link is 3640 * protected by the hash lock. If reached via 3641 * arc_buf_alloc, the header should not be accessed by 3642 * any other thread. And, if reached via arc_read_done, 3643 * the hash lock will protect it if it's found in the 3644 * hash table; otherwise no other thread should be 3645 * trying to [add|remove]_reference it. 3646 */ 3647 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 3648 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3649 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type], 3650 size); 3651 } 3652 /* 3653 * If we are growing the cache, and we are adding anonymous 3654 * data, and we have outgrown arc_p, update arc_p 3655 */ 3656 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 3657 (refcount_count(&arc_anon->arcs_size) + 3658 refcount_count(&arc_mru->arcs_size) > arc_p)) 3659 arc_p = MIN(arc_c, arc_p + size); 3660 } 3661 } 3662 3663 /* 3664 * This routine is called whenever a buffer is accessed. 3665 * NOTE: the hash lock is dropped in this function. 3666 */ 3667 static void 3668 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3669 { 3670 clock_t now; 3671 3672 ASSERT(MUTEX_HELD(hash_lock)); 3673 ASSERT(HDR_HAS_L1HDR(hdr)); 3674 3675 if (hdr->b_l1hdr.b_state == arc_anon) { 3676 /* 3677 * This buffer is not in the cache, and does not 3678 * appear in our "ghost" list. Add the new buffer 3679 * to the MRU state. 3680 */ 3681 3682 ASSERT0(hdr->b_l1hdr.b_arc_access); 3683 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3684 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3685 arc_change_state(arc_mru, hdr, hash_lock); 3686 3687 } else if (hdr->b_l1hdr.b_state == arc_mru) { 3688 now = ddi_get_lbolt(); 3689 3690 /* 3691 * If this buffer is here because of a prefetch, then either: 3692 * - clear the flag if this is a "referencing" read 3693 * (any subsequent access will bump this into the MFU state). 3694 * or 3695 * - move the buffer to the head of the list if this is 3696 * another prefetch (to make it less likely to be evicted). 3697 */ 3698 if (HDR_PREFETCH(hdr)) { 3699 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 3700 /* link protected by hash lock */ 3701 ASSERT(multilist_link_active( 3702 &hdr->b_l1hdr.b_arc_node)); 3703 } else { 3704 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3705 ARCSTAT_BUMP(arcstat_mru_hits); 3706 } 3707 hdr->b_l1hdr.b_arc_access = now; 3708 return; 3709 } 3710 3711 /* 3712 * This buffer has been "accessed" only once so far, 3713 * but it is still in the cache. Move it to the MFU 3714 * state. 3715 */ 3716 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 3717 /* 3718 * More than 125ms have passed since we 3719 * instantiated this buffer. Move it to the 3720 * most frequently used state. 3721 */ 3722 hdr->b_l1hdr.b_arc_access = now; 3723 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3724 arc_change_state(arc_mfu, hdr, hash_lock); 3725 } 3726 ARCSTAT_BUMP(arcstat_mru_hits); 3727 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 3728 arc_state_t *new_state; 3729 /* 3730 * This buffer has been "accessed" recently, but 3731 * was evicted from the cache. Move it to the 3732 * MFU state. 3733 */ 3734 3735 if (HDR_PREFETCH(hdr)) { 3736 new_state = arc_mru; 3737 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 3738 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3739 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3740 } else { 3741 new_state = arc_mfu; 3742 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3743 } 3744 3745 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3746 arc_change_state(new_state, hdr, hash_lock); 3747 3748 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3749 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 3750 /* 3751 * This buffer has been accessed more than once and is 3752 * still in the cache. Keep it in the MFU state. 3753 * 3754 * NOTE: an add_reference() that occurred when we did 3755 * the arc_read() will have kicked this off the list. 3756 * If it was a prefetch, we will explicitly move it to 3757 * the head of the list now. 3758 */ 3759 if ((HDR_PREFETCH(hdr)) != 0) { 3760 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3761 /* link protected by hash_lock */ 3762 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3763 } 3764 ARCSTAT_BUMP(arcstat_mfu_hits); 3765 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3766 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 3767 arc_state_t *new_state = arc_mfu; 3768 /* 3769 * This buffer has been accessed more than once but has 3770 * been evicted from the cache. Move it back to the 3771 * MFU state. 3772 */ 3773 3774 if (HDR_PREFETCH(hdr)) { 3775 /* 3776 * This is a prefetch access... 3777 * move this block back to the MRU state. 3778 */ 3779 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3780 new_state = arc_mru; 3781 } 3782 3783 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3784 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3785 arc_change_state(new_state, hdr, hash_lock); 3786 3787 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3788 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 3789 /* 3790 * This buffer is on the 2nd Level ARC. 3791 */ 3792 3793 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3794 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3795 arc_change_state(arc_mfu, hdr, hash_lock); 3796 } else { 3797 ASSERT(!"invalid arc state"); 3798 } 3799 } 3800 3801 /* a generic arc_done_func_t which you can use */ 3802 /* ARGSUSED */ 3803 void 3804 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3805 { 3806 if (zio == NULL || zio->io_error == 0) 3807 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3808 VERIFY(arc_buf_remove_ref(buf, arg)); 3809 } 3810 3811 /* a generic arc_done_func_t */ 3812 void 3813 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3814 { 3815 arc_buf_t **bufp = arg; 3816 if (zio && zio->io_error) { 3817 VERIFY(arc_buf_remove_ref(buf, arg)); 3818 *bufp = NULL; 3819 } else { 3820 *bufp = buf; 3821 ASSERT(buf->b_data); 3822 } 3823 } 3824 3825 static void 3826 arc_read_done(zio_t *zio) 3827 { 3828 arc_buf_hdr_t *hdr; 3829 arc_buf_t *buf; 3830 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3831 kmutex_t *hash_lock = NULL; 3832 arc_callback_t *callback_list, *acb; 3833 int freeable = FALSE; 3834 3835 buf = zio->io_private; 3836 hdr = buf->b_hdr; 3837 3838 /* 3839 * The hdr was inserted into hash-table and removed from lists 3840 * prior to starting I/O. We should find this header, since 3841 * it's in the hash table, and it should be legit since it's 3842 * not possible to evict it during the I/O. The only possible 3843 * reason for it not to be found is if we were freed during the 3844 * read. 3845 */ 3846 if (HDR_IN_HASH_TABLE(hdr)) { 3847 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3848 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3849 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3850 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3851 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3852 3853 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3854 &hash_lock); 3855 3856 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3857 hash_lock == NULL) || 3858 (found == hdr && 3859 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3860 (found == hdr && HDR_L2_READING(hdr))); 3861 } 3862 3863 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 3864 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 3865 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 3866 3867 /* byteswap if necessary */ 3868 callback_list = hdr->b_l1hdr.b_acb; 3869 ASSERT(callback_list != NULL); 3870 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3871 dmu_object_byteswap_t bswap = 3872 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3873 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3874 byteswap_uint64_array : 3875 dmu_ot_byteswap[bswap].ob_func; 3876 func(buf->b_data, hdr->b_size); 3877 } 3878 3879 arc_cksum_compute(buf, B_FALSE); 3880 arc_buf_watch(buf); 3881 3882 if (hash_lock && zio->io_error == 0 && 3883 hdr->b_l1hdr.b_state == arc_anon) { 3884 /* 3885 * Only call arc_access on anonymous buffers. This is because 3886 * if we've issued an I/O for an evicted buffer, we've already 3887 * called arc_access (to prevent any simultaneous readers from 3888 * getting confused). 3889 */ 3890 arc_access(hdr, hash_lock); 3891 } 3892 3893 /* create copies of the data buffer for the callers */ 3894 abuf = buf; 3895 for (acb = callback_list; acb; acb = acb->acb_next) { 3896 if (acb->acb_done) { 3897 if (abuf == NULL) { 3898 ARCSTAT_BUMP(arcstat_duplicate_reads); 3899 abuf = arc_buf_clone(buf); 3900 } 3901 acb->acb_buf = abuf; 3902 abuf = NULL; 3903 } 3904 } 3905 hdr->b_l1hdr.b_acb = NULL; 3906 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3907 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3908 if (abuf == buf) { 3909 ASSERT(buf->b_efunc == NULL); 3910 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 3911 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3912 } 3913 3914 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 3915 callback_list != NULL); 3916 3917 if (zio->io_error != 0) { 3918 hdr->b_flags |= ARC_FLAG_IO_ERROR; 3919 if (hdr->b_l1hdr.b_state != arc_anon) 3920 arc_change_state(arc_anon, hdr, hash_lock); 3921 if (HDR_IN_HASH_TABLE(hdr)) 3922 buf_hash_remove(hdr); 3923 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 3924 } 3925 3926 /* 3927 * Broadcast before we drop the hash_lock to avoid the possibility 3928 * that the hdr (and hence the cv) might be freed before we get to 3929 * the cv_broadcast(). 3930 */ 3931 cv_broadcast(&hdr->b_l1hdr.b_cv); 3932 3933 if (hash_lock != NULL) { 3934 mutex_exit(hash_lock); 3935 } else { 3936 /* 3937 * This block was freed while we waited for the read to 3938 * complete. It has been removed from the hash table and 3939 * moved to the anonymous state (so that it won't show up 3940 * in the cache). 3941 */ 3942 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 3943 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 3944 } 3945 3946 /* execute each callback and free its structure */ 3947 while ((acb = callback_list) != NULL) { 3948 if (acb->acb_done) 3949 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3950 3951 if (acb->acb_zio_dummy != NULL) { 3952 acb->acb_zio_dummy->io_error = zio->io_error; 3953 zio_nowait(acb->acb_zio_dummy); 3954 } 3955 3956 callback_list = acb->acb_next; 3957 kmem_free(acb, sizeof (arc_callback_t)); 3958 } 3959 3960 if (freeable) 3961 arc_hdr_destroy(hdr); 3962 } 3963 3964 /* 3965 * "Read" the block at the specified DVA (in bp) via the 3966 * cache. If the block is found in the cache, invoke the provided 3967 * callback immediately and return. Note that the `zio' parameter 3968 * in the callback will be NULL in this case, since no IO was 3969 * required. If the block is not in the cache pass the read request 3970 * on to the spa with a substitute callback function, so that the 3971 * requested block will be added to the cache. 3972 * 3973 * If a read request arrives for a block that has a read in-progress, 3974 * either wait for the in-progress read to complete (and return the 3975 * results); or, if this is a read with a "done" func, add a record 3976 * to the read to invoke the "done" func when the read completes, 3977 * and return; or just return. 3978 * 3979 * arc_read_done() will invoke all the requested "done" functions 3980 * for readers of this block. 3981 */ 3982 int 3983 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3984 void *private, zio_priority_t priority, int zio_flags, 3985 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 3986 { 3987 arc_buf_hdr_t *hdr = NULL; 3988 arc_buf_t *buf = NULL; 3989 kmutex_t *hash_lock = NULL; 3990 zio_t *rzio; 3991 uint64_t guid = spa_load_guid(spa); 3992 3993 ASSERT(!BP_IS_EMBEDDED(bp) || 3994 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3995 3996 top: 3997 if (!BP_IS_EMBEDDED(bp)) { 3998 /* 3999 * Embedded BP's have no DVA and require no I/O to "read". 4000 * Create an anonymous arc buf to back it. 4001 */ 4002 hdr = buf_hash_find(guid, bp, &hash_lock); 4003 } 4004 4005 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) { 4006 4007 *arc_flags |= ARC_FLAG_CACHED; 4008 4009 if (HDR_IO_IN_PROGRESS(hdr)) { 4010 4011 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 4012 priority == ZIO_PRIORITY_SYNC_READ) { 4013 /* 4014 * This sync read must wait for an 4015 * in-progress async read (e.g. a predictive 4016 * prefetch). Async reads are queued 4017 * separately at the vdev_queue layer, so 4018 * this is a form of priority inversion. 4019 * Ideally, we would "inherit" the demand 4020 * i/o's priority by moving the i/o from 4021 * the async queue to the synchronous queue, 4022 * but there is currently no mechanism to do 4023 * so. Track this so that we can evaluate 4024 * the magnitude of this potential performance 4025 * problem. 4026 * 4027 * Note that if the prefetch i/o is already 4028 * active (has been issued to the device), 4029 * the prefetch improved performance, because 4030 * we issued it sooner than we would have 4031 * without the prefetch. 4032 */ 4033 DTRACE_PROBE1(arc__sync__wait__for__async, 4034 arc_buf_hdr_t *, hdr); 4035 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 4036 } 4037 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4038 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH; 4039 } 4040 4041 if (*arc_flags & ARC_FLAG_WAIT) { 4042 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 4043 mutex_exit(hash_lock); 4044 goto top; 4045 } 4046 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4047 4048 if (done) { 4049 arc_callback_t *acb = NULL; 4050 4051 acb = kmem_zalloc(sizeof (arc_callback_t), 4052 KM_SLEEP); 4053 acb->acb_done = done; 4054 acb->acb_private = private; 4055 if (pio != NULL) 4056 acb->acb_zio_dummy = zio_null(pio, 4057 spa, NULL, NULL, NULL, zio_flags); 4058 4059 ASSERT(acb->acb_done != NULL); 4060 acb->acb_next = hdr->b_l1hdr.b_acb; 4061 hdr->b_l1hdr.b_acb = acb; 4062 add_reference(hdr, hash_lock, private); 4063 mutex_exit(hash_lock); 4064 return (0); 4065 } 4066 mutex_exit(hash_lock); 4067 return (0); 4068 } 4069 4070 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4071 hdr->b_l1hdr.b_state == arc_mfu); 4072 4073 if (done) { 4074 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4075 /* 4076 * This is a demand read which does not have to 4077 * wait for i/o because we did a predictive 4078 * prefetch i/o for it, which has completed. 4079 */ 4080 DTRACE_PROBE1( 4081 arc__demand__hit__predictive__prefetch, 4082 arc_buf_hdr_t *, hdr); 4083 ARCSTAT_BUMP( 4084 arcstat_demand_hit_predictive_prefetch); 4085 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH; 4086 } 4087 add_reference(hdr, hash_lock, private); 4088 /* 4089 * If this block is already in use, create a new 4090 * copy of the data so that we will be guaranteed 4091 * that arc_release() will always succeed. 4092 */ 4093 buf = hdr->b_l1hdr.b_buf; 4094 ASSERT(buf); 4095 ASSERT(buf->b_data); 4096 if (HDR_BUF_AVAILABLE(hdr)) { 4097 ASSERT(buf->b_efunc == NULL); 4098 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4099 } else { 4100 buf = arc_buf_clone(buf); 4101 } 4102 4103 } else if (*arc_flags & ARC_FLAG_PREFETCH && 4104 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4105 hdr->b_flags |= ARC_FLAG_PREFETCH; 4106 } 4107 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4108 arc_access(hdr, hash_lock); 4109 if (*arc_flags & ARC_FLAG_L2CACHE) 4110 hdr->b_flags |= ARC_FLAG_L2CACHE; 4111 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4112 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4113 mutex_exit(hash_lock); 4114 ARCSTAT_BUMP(arcstat_hits); 4115 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4116 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4117 data, metadata, hits); 4118 4119 if (done) 4120 done(NULL, buf, private); 4121 } else { 4122 uint64_t size = BP_GET_LSIZE(bp); 4123 arc_callback_t *acb; 4124 vdev_t *vd = NULL; 4125 uint64_t addr = 0; 4126 boolean_t devw = B_FALSE; 4127 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 4128 int32_t b_asize = 0; 4129 4130 if (hdr == NULL) { 4131 /* this block is not in the cache */ 4132 arc_buf_hdr_t *exists = NULL; 4133 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4134 buf = arc_buf_alloc(spa, size, private, type); 4135 hdr = buf->b_hdr; 4136 if (!BP_IS_EMBEDDED(bp)) { 4137 hdr->b_dva = *BP_IDENTITY(bp); 4138 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 4139 exists = buf_hash_insert(hdr, &hash_lock); 4140 } 4141 if (exists != NULL) { 4142 /* somebody beat us to the hash insert */ 4143 mutex_exit(hash_lock); 4144 buf_discard_identity(hdr); 4145 (void) arc_buf_remove_ref(buf, private); 4146 goto top; /* restart the IO request */ 4147 } 4148 4149 /* 4150 * If there is a callback, we pass our reference to 4151 * it; otherwise we remove our reference. 4152 */ 4153 if (done == NULL) { 4154 (void) remove_reference(hdr, hash_lock, 4155 private); 4156 } 4157 if (*arc_flags & ARC_FLAG_PREFETCH) 4158 hdr->b_flags |= ARC_FLAG_PREFETCH; 4159 if (*arc_flags & ARC_FLAG_L2CACHE) 4160 hdr->b_flags |= ARC_FLAG_L2CACHE; 4161 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4162 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4163 if (BP_GET_LEVEL(bp) > 0) 4164 hdr->b_flags |= ARC_FLAG_INDIRECT; 4165 } else { 4166 /* 4167 * This block is in the ghost cache. If it was L2-only 4168 * (and thus didn't have an L1 hdr), we realloc the 4169 * header to add an L1 hdr. 4170 */ 4171 if (!HDR_HAS_L1HDR(hdr)) { 4172 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 4173 hdr_full_cache); 4174 } 4175 4176 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4177 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4178 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4179 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 4180 4181 /* 4182 * If there is a callback, we pass a reference to it. 4183 */ 4184 if (done != NULL) 4185 add_reference(hdr, hash_lock, private); 4186 if (*arc_flags & ARC_FLAG_PREFETCH) 4187 hdr->b_flags |= ARC_FLAG_PREFETCH; 4188 if (*arc_flags & ARC_FLAG_L2CACHE) 4189 hdr->b_flags |= ARC_FLAG_L2CACHE; 4190 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4191 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4192 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 4193 buf->b_hdr = hdr; 4194 buf->b_data = NULL; 4195 buf->b_efunc = NULL; 4196 buf->b_private = NULL; 4197 buf->b_next = NULL; 4198 hdr->b_l1hdr.b_buf = buf; 4199 ASSERT0(hdr->b_l1hdr.b_datacnt); 4200 hdr->b_l1hdr.b_datacnt = 1; 4201 arc_get_data_buf(buf); 4202 arc_access(hdr, hash_lock); 4203 } 4204 4205 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 4206 hdr->b_flags |= ARC_FLAG_PREDICTIVE_PREFETCH; 4207 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 4208 4209 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4210 acb->acb_done = done; 4211 acb->acb_private = private; 4212 4213 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4214 hdr->b_l1hdr.b_acb = acb; 4215 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4216 4217 if (HDR_HAS_L2HDR(hdr) && 4218 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 4219 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 4220 addr = hdr->b_l2hdr.b_daddr; 4221 b_compress = hdr->b_l2hdr.b_compress; 4222 b_asize = hdr->b_l2hdr.b_asize; 4223 /* 4224 * Lock out device removal. 4225 */ 4226 if (vdev_is_dead(vd) || 4227 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 4228 vd = NULL; 4229 } 4230 4231 if (hash_lock != NULL) 4232 mutex_exit(hash_lock); 4233 4234 /* 4235 * At this point, we have a level 1 cache miss. Try again in 4236 * L2ARC if possible. 4237 */ 4238 ASSERT3U(hdr->b_size, ==, size); 4239 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 4240 uint64_t, size, zbookmark_phys_t *, zb); 4241 ARCSTAT_BUMP(arcstat_misses); 4242 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4243 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4244 data, metadata, misses); 4245 4246 if (priority == ZIO_PRIORITY_ASYNC_READ) 4247 hdr->b_flags |= ARC_FLAG_PRIO_ASYNC_READ; 4248 else 4249 hdr->b_flags &= ~ARC_FLAG_PRIO_ASYNC_READ; 4250 4251 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 4252 /* 4253 * Read from the L2ARC if the following are true: 4254 * 1. The L2ARC vdev was previously cached. 4255 * 2. This buffer still has L2ARC metadata. 4256 * 3. This buffer isn't currently writing to the L2ARC. 4257 * 4. The L2ARC entry wasn't evicted, which may 4258 * also have invalidated the vdev. 4259 * 5. This isn't prefetch and l2arc_noprefetch is set. 4260 */ 4261 if (HDR_HAS_L2HDR(hdr) && 4262 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 4263 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 4264 l2arc_read_callback_t *cb; 4265 4266 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 4267 ARCSTAT_BUMP(arcstat_l2_hits); 4268 4269 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 4270 KM_SLEEP); 4271 cb->l2rcb_buf = buf; 4272 cb->l2rcb_spa = spa; 4273 cb->l2rcb_bp = *bp; 4274 cb->l2rcb_zb = *zb; 4275 cb->l2rcb_flags = zio_flags; 4276 cb->l2rcb_compress = b_compress; 4277 4278 ASSERT(addr >= VDEV_LABEL_START_SIZE && 4279 addr + size < vd->vdev_psize - 4280 VDEV_LABEL_END_SIZE); 4281 4282 /* 4283 * l2arc read. The SCL_L2ARC lock will be 4284 * released by l2arc_read_done(). 4285 * Issue a null zio if the underlying buffer 4286 * was squashed to zero size by compression. 4287 */ 4288 if (b_compress == ZIO_COMPRESS_EMPTY) { 4289 rzio = zio_null(pio, spa, vd, 4290 l2arc_read_done, cb, 4291 zio_flags | ZIO_FLAG_DONT_CACHE | 4292 ZIO_FLAG_CANFAIL | 4293 ZIO_FLAG_DONT_PROPAGATE | 4294 ZIO_FLAG_DONT_RETRY); 4295 } else { 4296 rzio = zio_read_phys(pio, vd, addr, 4297 b_asize, buf->b_data, 4298 ZIO_CHECKSUM_OFF, 4299 l2arc_read_done, cb, priority, 4300 zio_flags | ZIO_FLAG_DONT_CACHE | 4301 ZIO_FLAG_CANFAIL | 4302 ZIO_FLAG_DONT_PROPAGATE | 4303 ZIO_FLAG_DONT_RETRY, B_FALSE); 4304 } 4305 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 4306 zio_t *, rzio); 4307 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 4308 4309 if (*arc_flags & ARC_FLAG_NOWAIT) { 4310 zio_nowait(rzio); 4311 return (0); 4312 } 4313 4314 ASSERT(*arc_flags & ARC_FLAG_WAIT); 4315 if (zio_wait(rzio) == 0) 4316 return (0); 4317 4318 /* l2arc read error; goto zio_read() */ 4319 } else { 4320 DTRACE_PROBE1(l2arc__miss, 4321 arc_buf_hdr_t *, hdr); 4322 ARCSTAT_BUMP(arcstat_l2_misses); 4323 if (HDR_L2_WRITING(hdr)) 4324 ARCSTAT_BUMP(arcstat_l2_rw_clash); 4325 spa_config_exit(spa, SCL_L2ARC, vd); 4326 } 4327 } else { 4328 if (vd != NULL) 4329 spa_config_exit(spa, SCL_L2ARC, vd); 4330 if (l2arc_ndev != 0) { 4331 DTRACE_PROBE1(l2arc__miss, 4332 arc_buf_hdr_t *, hdr); 4333 ARCSTAT_BUMP(arcstat_l2_misses); 4334 } 4335 } 4336 4337 rzio = zio_read(pio, spa, bp, buf->b_data, size, 4338 arc_read_done, buf, priority, zio_flags, zb); 4339 4340 if (*arc_flags & ARC_FLAG_WAIT) 4341 return (zio_wait(rzio)); 4342 4343 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4344 zio_nowait(rzio); 4345 } 4346 return (0); 4347 } 4348 4349 void 4350 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 4351 { 4352 ASSERT(buf->b_hdr != NULL); 4353 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon); 4354 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) || 4355 func == NULL); 4356 ASSERT(buf->b_efunc == NULL); 4357 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 4358 4359 buf->b_efunc = func; 4360 buf->b_private = private; 4361 } 4362 4363 /* 4364 * Notify the arc that a block was freed, and thus will never be used again. 4365 */ 4366 void 4367 arc_freed(spa_t *spa, const blkptr_t *bp) 4368 { 4369 arc_buf_hdr_t *hdr; 4370 kmutex_t *hash_lock; 4371 uint64_t guid = spa_load_guid(spa); 4372 4373 ASSERT(!BP_IS_EMBEDDED(bp)); 4374 4375 hdr = buf_hash_find(guid, bp, &hash_lock); 4376 if (hdr == NULL) 4377 return; 4378 if (HDR_BUF_AVAILABLE(hdr)) { 4379 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 4380 add_reference(hdr, hash_lock, FTAG); 4381 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4382 mutex_exit(hash_lock); 4383 4384 arc_release(buf, FTAG); 4385 (void) arc_buf_remove_ref(buf, FTAG); 4386 } else { 4387 mutex_exit(hash_lock); 4388 } 4389 4390 } 4391 4392 /* 4393 * Clear the user eviction callback set by arc_set_callback(), first calling 4394 * it if it exists. Because the presence of a callback keeps an arc_buf cached 4395 * clearing the callback may result in the arc_buf being destroyed. However, 4396 * it will not result in the *last* arc_buf being destroyed, hence the data 4397 * will remain cached in the ARC. We make a copy of the arc buffer here so 4398 * that we can process the callback without holding any locks. 4399 * 4400 * It's possible that the callback is already in the process of being cleared 4401 * by another thread. In this case we can not clear the callback. 4402 * 4403 * Returns B_TRUE if the callback was successfully called and cleared. 4404 */ 4405 boolean_t 4406 arc_clear_callback(arc_buf_t *buf) 4407 { 4408 arc_buf_hdr_t *hdr; 4409 kmutex_t *hash_lock; 4410 arc_evict_func_t *efunc = buf->b_efunc; 4411 void *private = buf->b_private; 4412 4413 mutex_enter(&buf->b_evict_lock); 4414 hdr = buf->b_hdr; 4415 if (hdr == NULL) { 4416 /* 4417 * We are in arc_do_user_evicts(). 4418 */ 4419 ASSERT(buf->b_data == NULL); 4420 mutex_exit(&buf->b_evict_lock); 4421 return (B_FALSE); 4422 } else if (buf->b_data == NULL) { 4423 /* 4424 * We are on the eviction list; process this buffer now 4425 * but let arc_do_user_evicts() do the reaping. 4426 */ 4427 buf->b_efunc = NULL; 4428 mutex_exit(&buf->b_evict_lock); 4429 VERIFY0(efunc(private)); 4430 return (B_TRUE); 4431 } 4432 hash_lock = HDR_LOCK(hdr); 4433 mutex_enter(hash_lock); 4434 hdr = buf->b_hdr; 4435 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4436 4437 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <, 4438 hdr->b_l1hdr.b_datacnt); 4439 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4440 hdr->b_l1hdr.b_state == arc_mfu); 4441 4442 buf->b_efunc = NULL; 4443 buf->b_private = NULL; 4444 4445 if (hdr->b_l1hdr.b_datacnt > 1) { 4446 mutex_exit(&buf->b_evict_lock); 4447 arc_buf_destroy(buf, TRUE); 4448 } else { 4449 ASSERT(buf == hdr->b_l1hdr.b_buf); 4450 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4451 mutex_exit(&buf->b_evict_lock); 4452 } 4453 4454 mutex_exit(hash_lock); 4455 VERIFY0(efunc(private)); 4456 return (B_TRUE); 4457 } 4458 4459 /* 4460 * Release this buffer from the cache, making it an anonymous buffer. This 4461 * must be done after a read and prior to modifying the buffer contents. 4462 * If the buffer has more than one reference, we must make 4463 * a new hdr for the buffer. 4464 */ 4465 void 4466 arc_release(arc_buf_t *buf, void *tag) 4467 { 4468 arc_buf_hdr_t *hdr = buf->b_hdr; 4469 4470 /* 4471 * It would be nice to assert that if it's DMU metadata (level > 4472 * 0 || it's the dnode file), then it must be syncing context. 4473 * But we don't know that information at this level. 4474 */ 4475 4476 mutex_enter(&buf->b_evict_lock); 4477 4478 ASSERT(HDR_HAS_L1HDR(hdr)); 4479 4480 /* 4481 * We don't grab the hash lock prior to this check, because if 4482 * the buffer's header is in the arc_anon state, it won't be 4483 * linked into the hash table. 4484 */ 4485 if (hdr->b_l1hdr.b_state == arc_anon) { 4486 mutex_exit(&buf->b_evict_lock); 4487 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4488 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 4489 ASSERT(!HDR_HAS_L2HDR(hdr)); 4490 ASSERT(BUF_EMPTY(hdr)); 4491 4492 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1); 4493 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 4494 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4495 4496 ASSERT3P(buf->b_efunc, ==, NULL); 4497 ASSERT3P(buf->b_private, ==, NULL); 4498 4499 hdr->b_l1hdr.b_arc_access = 0; 4500 arc_buf_thaw(buf); 4501 4502 return; 4503 } 4504 4505 kmutex_t *hash_lock = HDR_LOCK(hdr); 4506 mutex_enter(hash_lock); 4507 4508 /* 4509 * This assignment is only valid as long as the hash_lock is 4510 * held, we must be careful not to reference state or the 4511 * b_state field after dropping the lock. 4512 */ 4513 arc_state_t *state = hdr->b_l1hdr.b_state; 4514 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4515 ASSERT3P(state, !=, arc_anon); 4516 4517 /* this buffer is not on any list */ 4518 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 4519 4520 if (HDR_HAS_L2HDR(hdr)) { 4521 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4522 4523 /* 4524 * We have to recheck this conditional again now that 4525 * we're holding the l2ad_mtx to prevent a race with 4526 * another thread which might be concurrently calling 4527 * l2arc_evict(). In that case, l2arc_evict() might have 4528 * destroyed the header's L2 portion as we were waiting 4529 * to acquire the l2ad_mtx. 4530 */ 4531 if (HDR_HAS_L2HDR(hdr)) 4532 arc_hdr_l2hdr_destroy(hdr); 4533 4534 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4535 } 4536 4537 /* 4538 * Do we have more than one buf? 4539 */ 4540 if (hdr->b_l1hdr.b_datacnt > 1) { 4541 arc_buf_hdr_t *nhdr; 4542 arc_buf_t **bufp; 4543 uint64_t blksz = hdr->b_size; 4544 uint64_t spa = hdr->b_spa; 4545 arc_buf_contents_t type = arc_buf_type(hdr); 4546 uint32_t flags = hdr->b_flags; 4547 4548 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 4549 /* 4550 * Pull the data off of this hdr and attach it to 4551 * a new anonymous hdr. 4552 */ 4553 (void) remove_reference(hdr, hash_lock, tag); 4554 bufp = &hdr->b_l1hdr.b_buf; 4555 while (*bufp != buf) 4556 bufp = &(*bufp)->b_next; 4557 *bufp = buf->b_next; 4558 buf->b_next = NULL; 4559 4560 ASSERT3P(state, !=, arc_l2c_only); 4561 4562 (void) refcount_remove_many( 4563 &state->arcs_size, hdr->b_size, buf); 4564 4565 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 4566 ASSERT3P(state, !=, arc_l2c_only); 4567 uint64_t *size = &state->arcs_lsize[type]; 4568 ASSERT3U(*size, >=, hdr->b_size); 4569 atomic_add_64(size, -hdr->b_size); 4570 } 4571 4572 /* 4573 * We're releasing a duplicate user data buffer, update 4574 * our statistics accordingly. 4575 */ 4576 if (HDR_ISTYPE_DATA(hdr)) { 4577 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 4578 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 4579 -hdr->b_size); 4580 } 4581 hdr->b_l1hdr.b_datacnt -= 1; 4582 arc_cksum_verify(buf); 4583 arc_buf_unwatch(buf); 4584 4585 mutex_exit(hash_lock); 4586 4587 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 4588 nhdr->b_size = blksz; 4589 nhdr->b_spa = spa; 4590 4591 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 4592 nhdr->b_flags |= arc_bufc_to_flags(type); 4593 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 4594 4595 nhdr->b_l1hdr.b_buf = buf; 4596 nhdr->b_l1hdr.b_datacnt = 1; 4597 nhdr->b_l1hdr.b_state = arc_anon; 4598 nhdr->b_l1hdr.b_arc_access = 0; 4599 nhdr->b_l1hdr.b_tmp_cdata = NULL; 4600 nhdr->b_freeze_cksum = NULL; 4601 4602 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 4603 buf->b_hdr = nhdr; 4604 mutex_exit(&buf->b_evict_lock); 4605 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf); 4606 } else { 4607 mutex_exit(&buf->b_evict_lock); 4608 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 4609 /* protected by hash lock, or hdr is on arc_anon */ 4610 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4611 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4612 arc_change_state(arc_anon, hdr, hash_lock); 4613 hdr->b_l1hdr.b_arc_access = 0; 4614 mutex_exit(hash_lock); 4615 4616 buf_discard_identity(hdr); 4617 arc_buf_thaw(buf); 4618 } 4619 buf->b_efunc = NULL; 4620 buf->b_private = NULL; 4621 } 4622 4623 int 4624 arc_released(arc_buf_t *buf) 4625 { 4626 int released; 4627 4628 mutex_enter(&buf->b_evict_lock); 4629 released = (buf->b_data != NULL && 4630 buf->b_hdr->b_l1hdr.b_state == arc_anon); 4631 mutex_exit(&buf->b_evict_lock); 4632 return (released); 4633 } 4634 4635 #ifdef ZFS_DEBUG 4636 int 4637 arc_referenced(arc_buf_t *buf) 4638 { 4639 int referenced; 4640 4641 mutex_enter(&buf->b_evict_lock); 4642 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 4643 mutex_exit(&buf->b_evict_lock); 4644 return (referenced); 4645 } 4646 #endif 4647 4648 static void 4649 arc_write_ready(zio_t *zio) 4650 { 4651 arc_write_callback_t *callback = zio->io_private; 4652 arc_buf_t *buf = callback->awcb_buf; 4653 arc_buf_hdr_t *hdr = buf->b_hdr; 4654 4655 ASSERT(HDR_HAS_L1HDR(hdr)); 4656 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 4657 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4658 callback->awcb_ready(zio, buf, callback->awcb_private); 4659 4660 /* 4661 * If the IO is already in progress, then this is a re-write 4662 * attempt, so we need to thaw and re-compute the cksum. 4663 * It is the responsibility of the callback to handle the 4664 * accounting for any re-write attempt. 4665 */ 4666 if (HDR_IO_IN_PROGRESS(hdr)) { 4667 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 4668 if (hdr->b_freeze_cksum != NULL) { 4669 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 4670 hdr->b_freeze_cksum = NULL; 4671 } 4672 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 4673 } 4674 arc_cksum_compute(buf, B_FALSE); 4675 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4676 } 4677 4678 /* 4679 * The SPA calls this callback for each physical write that happens on behalf 4680 * of a logical write. See the comment in dbuf_write_physdone() for details. 4681 */ 4682 static void 4683 arc_write_physdone(zio_t *zio) 4684 { 4685 arc_write_callback_t *cb = zio->io_private; 4686 if (cb->awcb_physdone != NULL) 4687 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 4688 } 4689 4690 static void 4691 arc_write_done(zio_t *zio) 4692 { 4693 arc_write_callback_t *callback = zio->io_private; 4694 arc_buf_t *buf = callback->awcb_buf; 4695 arc_buf_hdr_t *hdr = buf->b_hdr; 4696 4697 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4698 4699 if (zio->io_error == 0) { 4700 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 4701 buf_discard_identity(hdr); 4702 } else { 4703 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 4704 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 4705 } 4706 } else { 4707 ASSERT(BUF_EMPTY(hdr)); 4708 } 4709 4710 /* 4711 * If the block to be written was all-zero or compressed enough to be 4712 * embedded in the BP, no write was performed so there will be no 4713 * dva/birth/checksum. The buffer must therefore remain anonymous 4714 * (and uncached). 4715 */ 4716 if (!BUF_EMPTY(hdr)) { 4717 arc_buf_hdr_t *exists; 4718 kmutex_t *hash_lock; 4719 4720 ASSERT(zio->io_error == 0); 4721 4722 arc_cksum_verify(buf); 4723 4724 exists = buf_hash_insert(hdr, &hash_lock); 4725 if (exists != NULL) { 4726 /* 4727 * This can only happen if we overwrite for 4728 * sync-to-convergence, because we remove 4729 * buffers from the hash table when we arc_free(). 4730 */ 4731 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 4732 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4733 panic("bad overwrite, hdr=%p exists=%p", 4734 (void *)hdr, (void *)exists); 4735 ASSERT(refcount_is_zero( 4736 &exists->b_l1hdr.b_refcnt)); 4737 arc_change_state(arc_anon, exists, hash_lock); 4738 mutex_exit(hash_lock); 4739 arc_hdr_destroy(exists); 4740 exists = buf_hash_insert(hdr, &hash_lock); 4741 ASSERT3P(exists, ==, NULL); 4742 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 4743 /* nopwrite */ 4744 ASSERT(zio->io_prop.zp_nopwrite); 4745 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4746 panic("bad nopwrite, hdr=%p exists=%p", 4747 (void *)hdr, (void *)exists); 4748 } else { 4749 /* Dedup */ 4750 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4751 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 4752 ASSERT(BP_GET_DEDUP(zio->io_bp)); 4753 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 4754 } 4755 } 4756 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4757 /* if it's not anon, we are doing a scrub */ 4758 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 4759 arc_access(hdr, hash_lock); 4760 mutex_exit(hash_lock); 4761 } else { 4762 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4763 } 4764 4765 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4766 callback->awcb_done(zio, buf, callback->awcb_private); 4767 4768 kmem_free(callback, sizeof (arc_write_callback_t)); 4769 } 4770 4771 zio_t * 4772 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 4773 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 4774 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 4775 arc_done_func_t *done, void *private, zio_priority_t priority, 4776 int zio_flags, const zbookmark_phys_t *zb) 4777 { 4778 arc_buf_hdr_t *hdr = buf->b_hdr; 4779 arc_write_callback_t *callback; 4780 zio_t *zio; 4781 4782 ASSERT(ready != NULL); 4783 ASSERT(done != NULL); 4784 ASSERT(!HDR_IO_ERROR(hdr)); 4785 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4786 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4787 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4788 if (l2arc) 4789 hdr->b_flags |= ARC_FLAG_L2CACHE; 4790 if (l2arc_compress) 4791 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4792 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4793 callback->awcb_ready = ready; 4794 callback->awcb_physdone = physdone; 4795 callback->awcb_done = done; 4796 callback->awcb_private = private; 4797 callback->awcb_buf = buf; 4798 4799 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4800 arc_write_ready, arc_write_physdone, arc_write_done, callback, 4801 priority, zio_flags, zb); 4802 4803 return (zio); 4804 } 4805 4806 static int 4807 arc_memory_throttle(uint64_t reserve, uint64_t txg) 4808 { 4809 #ifdef _KERNEL 4810 uint64_t available_memory = ptob(freemem); 4811 static uint64_t page_load = 0; 4812 static uint64_t last_txg = 0; 4813 4814 #if defined(__i386) 4815 available_memory = 4816 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 4817 #endif 4818 4819 if (freemem > physmem * arc_lotsfree_percent / 100) 4820 return (0); 4821 4822 if (txg > last_txg) { 4823 last_txg = txg; 4824 page_load = 0; 4825 } 4826 /* 4827 * If we are in pageout, we know that memory is already tight, 4828 * the arc is already going to be evicting, so we just want to 4829 * continue to let page writes occur as quickly as possible. 4830 */ 4831 if (curproc == proc_pageout) { 4832 if (page_load > MAX(ptob(minfree), available_memory) / 4) 4833 return (SET_ERROR(ERESTART)); 4834 /* Note: reserve is inflated, so we deflate */ 4835 page_load += reserve / 8; 4836 return (0); 4837 } else if (page_load > 0 && arc_reclaim_needed()) { 4838 /* memory is low, delay before restarting */ 4839 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4840 return (SET_ERROR(EAGAIN)); 4841 } 4842 page_load = 0; 4843 #endif 4844 return (0); 4845 } 4846 4847 void 4848 arc_tempreserve_clear(uint64_t reserve) 4849 { 4850 atomic_add_64(&arc_tempreserve, -reserve); 4851 ASSERT((int64_t)arc_tempreserve >= 0); 4852 } 4853 4854 int 4855 arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4856 { 4857 int error; 4858 uint64_t anon_size; 4859 4860 if (reserve > arc_c/4 && !arc_no_grow) 4861 arc_c = MIN(arc_c_max, reserve * 4); 4862 if (reserve > arc_c) 4863 return (SET_ERROR(ENOMEM)); 4864 4865 /* 4866 * Don't count loaned bufs as in flight dirty data to prevent long 4867 * network delays from blocking transactions that are ready to be 4868 * assigned to a txg. 4869 */ 4870 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 4871 arc_loaned_bytes), 0); 4872 4873 /* 4874 * Writes will, almost always, require additional memory allocations 4875 * in order to compress/encrypt/etc the data. We therefore need to 4876 * make sure that there is sufficient available memory for this. 4877 */ 4878 error = arc_memory_throttle(reserve, txg); 4879 if (error != 0) 4880 return (error); 4881 4882 /* 4883 * Throttle writes when the amount of dirty data in the cache 4884 * gets too large. We try to keep the cache less than half full 4885 * of dirty blocks so that our sync times don't grow too large. 4886 * Note: if two requests come in concurrently, we might let them 4887 * both succeed, when one of them should fail. Not a huge deal. 4888 */ 4889 4890 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4891 anon_size > arc_c / 4) { 4892 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4893 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4894 arc_tempreserve>>10, 4895 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4896 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4897 reserve>>10, arc_c>>10); 4898 return (SET_ERROR(ERESTART)); 4899 } 4900 atomic_add_64(&arc_tempreserve, reserve); 4901 return (0); 4902 } 4903 4904 static void 4905 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 4906 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 4907 { 4908 size->value.ui64 = refcount_count(&state->arcs_size); 4909 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA]; 4910 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA]; 4911 } 4912 4913 static int 4914 arc_kstat_update(kstat_t *ksp, int rw) 4915 { 4916 arc_stats_t *as = ksp->ks_data; 4917 4918 if (rw == KSTAT_WRITE) { 4919 return (EACCES); 4920 } else { 4921 arc_kstat_update_state(arc_anon, 4922 &as->arcstat_anon_size, 4923 &as->arcstat_anon_evictable_data, 4924 &as->arcstat_anon_evictable_metadata); 4925 arc_kstat_update_state(arc_mru, 4926 &as->arcstat_mru_size, 4927 &as->arcstat_mru_evictable_data, 4928 &as->arcstat_mru_evictable_metadata); 4929 arc_kstat_update_state(arc_mru_ghost, 4930 &as->arcstat_mru_ghost_size, 4931 &as->arcstat_mru_ghost_evictable_data, 4932 &as->arcstat_mru_ghost_evictable_metadata); 4933 arc_kstat_update_state(arc_mfu, 4934 &as->arcstat_mfu_size, 4935 &as->arcstat_mfu_evictable_data, 4936 &as->arcstat_mfu_evictable_metadata); 4937 arc_kstat_update_state(arc_mfu_ghost, 4938 &as->arcstat_mfu_ghost_size, 4939 &as->arcstat_mfu_ghost_evictable_data, 4940 &as->arcstat_mfu_ghost_evictable_metadata); 4941 } 4942 4943 return (0); 4944 } 4945 4946 /* 4947 * This function *must* return indices evenly distributed between all 4948 * sublists of the multilist. This is needed due to how the ARC eviction 4949 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 4950 * distributed between all sublists and uses this assumption when 4951 * deciding which sublist to evict from and how much to evict from it. 4952 */ 4953 unsigned int 4954 arc_state_multilist_index_func(multilist_t *ml, void *obj) 4955 { 4956 arc_buf_hdr_t *hdr = obj; 4957 4958 /* 4959 * We rely on b_dva to generate evenly distributed index 4960 * numbers using buf_hash below. So, as an added precaution, 4961 * let's make sure we never add empty buffers to the arc lists. 4962 */ 4963 ASSERT(!BUF_EMPTY(hdr)); 4964 4965 /* 4966 * The assumption here, is the hash value for a given 4967 * arc_buf_hdr_t will remain constant throughout it's lifetime 4968 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 4969 * Thus, we don't need to store the header's sublist index 4970 * on insertion, as this index can be recalculated on removal. 4971 * 4972 * Also, the low order bits of the hash value are thought to be 4973 * distributed evenly. Otherwise, in the case that the multilist 4974 * has a power of two number of sublists, each sublists' usage 4975 * would not be evenly distributed. 4976 */ 4977 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 4978 multilist_get_num_sublists(ml)); 4979 } 4980 4981 void 4982 arc_init(void) 4983 { 4984 /* 4985 * allmem is "all memory that we could possibly use". 4986 */ 4987 #ifdef _KERNEL 4988 uint64_t allmem = ptob(physmem - swapfs_minfree); 4989 #else 4990 uint64_t allmem = (physmem * PAGESIZE) / 2; 4991 #endif 4992 4993 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 4994 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 4995 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 4996 4997 mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 4998 cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL); 4999 5000 /* Convert seconds to clock ticks */ 5001 arc_min_prefetch_lifespan = 1 * hz; 5002 5003 /* Start out with 1/8 of all memory */ 5004 arc_c = allmem / 8; 5005 5006 #ifdef _KERNEL 5007 /* 5008 * On architectures where the physical memory can be larger 5009 * than the addressable space (intel in 32-bit mode), we may 5010 * need to limit the cache to 1/8 of VM size. 5011 */ 5012 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 5013 #endif 5014 5015 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */ 5016 arc_c_min = MAX(allmem / 32, 64 << 20); 5017 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */ 5018 if (allmem >= 1 << 30) 5019 arc_c_max = allmem - (1 << 30); 5020 else 5021 arc_c_max = arc_c_min; 5022 arc_c_max = MAX(allmem * 3 / 4, arc_c_max); 5023 5024 /* 5025 * In userland, there's only the memory pressure that we artificially 5026 * create (see arc_available_memory()). Don't let arc_c get too 5027 * small, because it can cause transactions to be larger than 5028 * arc_c, causing arc_tempreserve_space() to fail. 5029 */ 5030 #ifndef _KERNEL 5031 arc_c_min = arc_c_max / 2; 5032 #endif 5033 5034 /* 5035 * Allow the tunables to override our calculations if they are 5036 * reasonable (ie. over 64MB) 5037 */ 5038 if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) 5039 arc_c_max = zfs_arc_max; 5040 if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max) 5041 arc_c_min = zfs_arc_min; 5042 5043 arc_c = arc_c_max; 5044 arc_p = (arc_c >> 1); 5045 5046 /* limit meta-data to 1/4 of the arc capacity */ 5047 arc_meta_limit = arc_c_max / 4; 5048 5049 /* Allow the tunable to override if it is reasonable */ 5050 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 5051 arc_meta_limit = zfs_arc_meta_limit; 5052 5053 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 5054 arc_c_min = arc_meta_limit / 2; 5055 5056 if (zfs_arc_meta_min > 0) { 5057 arc_meta_min = zfs_arc_meta_min; 5058 } else { 5059 arc_meta_min = arc_c_min / 2; 5060 } 5061 5062 if (zfs_arc_grow_retry > 0) 5063 arc_grow_retry = zfs_arc_grow_retry; 5064 5065 if (zfs_arc_shrink_shift > 0) 5066 arc_shrink_shift = zfs_arc_shrink_shift; 5067 5068 /* 5069 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 5070 */ 5071 if (arc_no_grow_shift >= arc_shrink_shift) 5072 arc_no_grow_shift = arc_shrink_shift - 1; 5073 5074 if (zfs_arc_p_min_shift > 0) 5075 arc_p_min_shift = zfs_arc_p_min_shift; 5076 5077 if (zfs_arc_num_sublists_per_state < 1) 5078 zfs_arc_num_sublists_per_state = MAX(boot_ncpus, 1); 5079 5080 /* if kmem_flags are set, lets try to use less memory */ 5081 if (kmem_debugging()) 5082 arc_c = arc_c / 2; 5083 if (arc_c < arc_c_min) 5084 arc_c = arc_c_min; 5085 5086 arc_anon = &ARC_anon; 5087 arc_mru = &ARC_mru; 5088 arc_mru_ghost = &ARC_mru_ghost; 5089 arc_mfu = &ARC_mfu; 5090 arc_mfu_ghost = &ARC_mfu_ghost; 5091 arc_l2c_only = &ARC_l2c_only; 5092 arc_size = 0; 5093 5094 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 5095 sizeof (arc_buf_hdr_t), 5096 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5097 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5098 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 5099 sizeof (arc_buf_hdr_t), 5100 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5101 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5102 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 5103 sizeof (arc_buf_hdr_t), 5104 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5105 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5106 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 5107 sizeof (arc_buf_hdr_t), 5108 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5109 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5110 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 5111 sizeof (arc_buf_hdr_t), 5112 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5113 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5114 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 5115 sizeof (arc_buf_hdr_t), 5116 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5117 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5118 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 5119 sizeof (arc_buf_hdr_t), 5120 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5121 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5122 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 5123 sizeof (arc_buf_hdr_t), 5124 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5125 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5126 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], 5127 sizeof (arc_buf_hdr_t), 5128 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5129 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5130 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], 5131 sizeof (arc_buf_hdr_t), 5132 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5133 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5134 5135 refcount_create(&arc_anon->arcs_size); 5136 refcount_create(&arc_mru->arcs_size); 5137 refcount_create(&arc_mru_ghost->arcs_size); 5138 refcount_create(&arc_mfu->arcs_size); 5139 refcount_create(&arc_mfu_ghost->arcs_size); 5140 refcount_create(&arc_l2c_only->arcs_size); 5141 5142 buf_init(); 5143 5144 arc_reclaim_thread_exit = FALSE; 5145 arc_user_evicts_thread_exit = FALSE; 5146 arc_eviction_list = NULL; 5147 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 5148 5149 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 5150 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 5151 5152 if (arc_ksp != NULL) { 5153 arc_ksp->ks_data = &arc_stats; 5154 arc_ksp->ks_update = arc_kstat_update; 5155 kstat_install(arc_ksp); 5156 } 5157 5158 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 5159 TS_RUN, minclsyspri); 5160 5161 (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0, 5162 TS_RUN, minclsyspri); 5163 5164 arc_dead = FALSE; 5165 arc_warm = B_FALSE; 5166 5167 /* 5168 * Calculate maximum amount of dirty data per pool. 5169 * 5170 * If it has been set by /etc/system, take that. 5171 * Otherwise, use a percentage of physical memory defined by 5172 * zfs_dirty_data_max_percent (default 10%) with a cap at 5173 * zfs_dirty_data_max_max (default 4GB). 5174 */ 5175 if (zfs_dirty_data_max == 0) { 5176 zfs_dirty_data_max = physmem * PAGESIZE * 5177 zfs_dirty_data_max_percent / 100; 5178 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 5179 zfs_dirty_data_max_max); 5180 } 5181 } 5182 5183 void 5184 arc_fini(void) 5185 { 5186 mutex_enter(&arc_reclaim_lock); 5187 arc_reclaim_thread_exit = TRUE; 5188 /* 5189 * The reclaim thread will set arc_reclaim_thread_exit back to 5190 * FALSE when it is finished exiting; we're waiting for that. 5191 */ 5192 while (arc_reclaim_thread_exit) { 5193 cv_signal(&arc_reclaim_thread_cv); 5194 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 5195 } 5196 mutex_exit(&arc_reclaim_lock); 5197 5198 mutex_enter(&arc_user_evicts_lock); 5199 arc_user_evicts_thread_exit = TRUE; 5200 /* 5201 * The user evicts thread will set arc_user_evicts_thread_exit 5202 * to FALSE when it is finished exiting; we're waiting for that. 5203 */ 5204 while (arc_user_evicts_thread_exit) { 5205 cv_signal(&arc_user_evicts_cv); 5206 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock); 5207 } 5208 mutex_exit(&arc_user_evicts_lock); 5209 5210 /* Use TRUE to ensure *all* buffers are evicted */ 5211 arc_flush(NULL, TRUE); 5212 5213 arc_dead = TRUE; 5214 5215 if (arc_ksp != NULL) { 5216 kstat_delete(arc_ksp); 5217 arc_ksp = NULL; 5218 } 5219 5220 mutex_destroy(&arc_reclaim_lock); 5221 cv_destroy(&arc_reclaim_thread_cv); 5222 cv_destroy(&arc_reclaim_waiters_cv); 5223 5224 mutex_destroy(&arc_user_evicts_lock); 5225 cv_destroy(&arc_user_evicts_cv); 5226 5227 refcount_destroy(&arc_anon->arcs_size); 5228 refcount_destroy(&arc_mru->arcs_size); 5229 refcount_destroy(&arc_mru_ghost->arcs_size); 5230 refcount_destroy(&arc_mfu->arcs_size); 5231 refcount_destroy(&arc_mfu_ghost->arcs_size); 5232 refcount_destroy(&arc_l2c_only->arcs_size); 5233 5234 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 5235 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 5236 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 5237 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 5238 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]); 5239 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 5240 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 5241 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 5242 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 5243 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]); 5244 5245 buf_fini(); 5246 5247 ASSERT0(arc_loaned_bytes); 5248 } 5249 5250 /* 5251 * Level 2 ARC 5252 * 5253 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 5254 * It uses dedicated storage devices to hold cached data, which are populated 5255 * using large infrequent writes. The main role of this cache is to boost 5256 * the performance of random read workloads. The intended L2ARC devices 5257 * include short-stroked disks, solid state disks, and other media with 5258 * substantially faster read latency than disk. 5259 * 5260 * +-----------------------+ 5261 * | ARC | 5262 * +-----------------------+ 5263 * | ^ ^ 5264 * | | | 5265 * l2arc_feed_thread() arc_read() 5266 * | | | 5267 * | l2arc read | 5268 * V | | 5269 * +---------------+ | 5270 * | L2ARC | | 5271 * +---------------+ | 5272 * | ^ | 5273 * l2arc_write() | | 5274 * | | | 5275 * V | | 5276 * +-------+ +-------+ 5277 * | vdev | | vdev | 5278 * | cache | | cache | 5279 * +-------+ +-------+ 5280 * +=========+ .-----. 5281 * : L2ARC : |-_____-| 5282 * : devices : | Disks | 5283 * +=========+ `-_____-' 5284 * 5285 * Read requests are satisfied from the following sources, in order: 5286 * 5287 * 1) ARC 5288 * 2) vdev cache of L2ARC devices 5289 * 3) L2ARC devices 5290 * 4) vdev cache of disks 5291 * 5) disks 5292 * 5293 * Some L2ARC device types exhibit extremely slow write performance. 5294 * To accommodate for this there are some significant differences between 5295 * the L2ARC and traditional cache design: 5296 * 5297 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 5298 * the ARC behave as usual, freeing buffers and placing headers on ghost 5299 * lists. The ARC does not send buffers to the L2ARC during eviction as 5300 * this would add inflated write latencies for all ARC memory pressure. 5301 * 5302 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 5303 * It does this by periodically scanning buffers from the eviction-end of 5304 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 5305 * not already there. It scans until a headroom of buffers is satisfied, 5306 * which itself is a buffer for ARC eviction. If a compressible buffer is 5307 * found during scanning and selected for writing to an L2ARC device, we 5308 * temporarily boost scanning headroom during the next scan cycle to make 5309 * sure we adapt to compression effects (which might significantly reduce 5310 * the data volume we write to L2ARC). The thread that does this is 5311 * l2arc_feed_thread(), illustrated below; example sizes are included to 5312 * provide a better sense of ratio than this diagram: 5313 * 5314 * head --> tail 5315 * +---------------------+----------+ 5316 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 5317 * +---------------------+----------+ | o L2ARC eligible 5318 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 5319 * +---------------------+----------+ | 5320 * 15.9 Gbytes ^ 32 Mbytes | 5321 * headroom | 5322 * l2arc_feed_thread() 5323 * | 5324 * l2arc write hand <--[oooo]--' 5325 * | 8 Mbyte 5326 * | write max 5327 * V 5328 * +==============================+ 5329 * L2ARC dev |####|#|###|###| |####| ... | 5330 * +==============================+ 5331 * 32 Gbytes 5332 * 5333 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 5334 * evicted, then the L2ARC has cached a buffer much sooner than it probably 5335 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 5336 * safe to say that this is an uncommon case, since buffers at the end of 5337 * the ARC lists have moved there due to inactivity. 5338 * 5339 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 5340 * then the L2ARC simply misses copying some buffers. This serves as a 5341 * pressure valve to prevent heavy read workloads from both stalling the ARC 5342 * with waits and clogging the L2ARC with writes. This also helps prevent 5343 * the potential for the L2ARC to churn if it attempts to cache content too 5344 * quickly, such as during backups of the entire pool. 5345 * 5346 * 5. After system boot and before the ARC has filled main memory, there are 5347 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 5348 * lists can remain mostly static. Instead of searching from tail of these 5349 * lists as pictured, the l2arc_feed_thread() will search from the list heads 5350 * for eligible buffers, greatly increasing its chance of finding them. 5351 * 5352 * The L2ARC device write speed is also boosted during this time so that 5353 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 5354 * there are no L2ARC reads, and no fear of degrading read performance 5355 * through increased writes. 5356 * 5357 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 5358 * the vdev queue can aggregate them into larger and fewer writes. Each 5359 * device is written to in a rotor fashion, sweeping writes through 5360 * available space then repeating. 5361 * 5362 * 7. The L2ARC does not store dirty content. It never needs to flush 5363 * write buffers back to disk based storage. 5364 * 5365 * 8. If an ARC buffer is written (and dirtied) which also exists in the 5366 * L2ARC, the now stale L2ARC buffer is immediately dropped. 5367 * 5368 * The performance of the L2ARC can be tweaked by a number of tunables, which 5369 * may be necessary for different workloads: 5370 * 5371 * l2arc_write_max max write bytes per interval 5372 * l2arc_write_boost extra write bytes during device warmup 5373 * l2arc_noprefetch skip caching prefetched buffers 5374 * l2arc_headroom number of max device writes to precache 5375 * l2arc_headroom_boost when we find compressed buffers during ARC 5376 * scanning, we multiply headroom by this 5377 * percentage factor for the next scan cycle, 5378 * since more compressed buffers are likely to 5379 * be present 5380 * l2arc_feed_secs seconds between L2ARC writing 5381 * 5382 * Tunables may be removed or added as future performance improvements are 5383 * integrated, and also may become zpool properties. 5384 * 5385 * There are three key functions that control how the L2ARC warms up: 5386 * 5387 * l2arc_write_eligible() check if a buffer is eligible to cache 5388 * l2arc_write_size() calculate how much to write 5389 * l2arc_write_interval() calculate sleep delay between writes 5390 * 5391 * These three functions determine what to write, how much, and how quickly 5392 * to send writes. 5393 */ 5394 5395 static boolean_t 5396 l2arc_write_eligible(uint64_t spa_guid, uint64_t sync_txg, arc_buf_hdr_t *hdr) 5397 { 5398 /* 5399 * A buffer is *not* eligible for the L2ARC if it: 5400 * 1. belongs to a different spa. 5401 * 2. is already cached on the L2ARC. 5402 * 3. has an I/O in progress (it may be an incomplete read). 5403 * 4. is flagged not eligible (zfs property). 5404 * 5. is part of the syncing txg (and thus subject to change). 5405 */ 5406 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) || 5407 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr) || 5408 hdr->b_birth >= sync_txg) 5409 return (B_FALSE); 5410 5411 return (B_TRUE); 5412 } 5413 5414 static uint64_t 5415 l2arc_write_size(void) 5416 { 5417 uint64_t size; 5418 5419 /* 5420 * Make sure our globals have meaningful values in case the user 5421 * altered them. 5422 */ 5423 size = l2arc_write_max; 5424 if (size == 0) { 5425 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 5426 "be greater than zero, resetting it to the default (%d)", 5427 L2ARC_WRITE_SIZE); 5428 size = l2arc_write_max = L2ARC_WRITE_SIZE; 5429 } 5430 5431 if (arc_warm == B_FALSE) 5432 size += l2arc_write_boost; 5433 5434 return (size); 5435 5436 } 5437 5438 static clock_t 5439 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 5440 { 5441 clock_t interval, next, now; 5442 5443 /* 5444 * If the ARC lists are busy, increase our write rate; if the 5445 * lists are stale, idle back. This is achieved by checking 5446 * how much we previously wrote - if it was more than half of 5447 * what we wanted, schedule the next write much sooner. 5448 */ 5449 if (l2arc_feed_again && wrote > (wanted / 2)) 5450 interval = (hz * l2arc_feed_min_ms) / 1000; 5451 else 5452 interval = hz * l2arc_feed_secs; 5453 5454 now = ddi_get_lbolt(); 5455 next = MAX(now, MIN(now + interval, began + interval)); 5456 5457 return (next); 5458 } 5459 5460 /* 5461 * Cycle through L2ARC devices. This is how L2ARC load balances. 5462 * If a device is returned, this also returns holding the spa config lock. 5463 */ 5464 static l2arc_dev_t * 5465 l2arc_dev_get_next(void) 5466 { 5467 l2arc_dev_t *first, *next = NULL; 5468 5469 /* 5470 * Lock out the removal of spas (spa_namespace_lock), then removal 5471 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 5472 * both locks will be dropped and a spa config lock held instead. 5473 */ 5474 mutex_enter(&spa_namespace_lock); 5475 mutex_enter(&l2arc_dev_mtx); 5476 5477 /* if there are no vdevs, there is nothing to do */ 5478 if (l2arc_ndev == 0) 5479 goto out; 5480 5481 first = NULL; 5482 next = l2arc_dev_last; 5483 do { 5484 /* loop around the list looking for a non-faulted vdev */ 5485 if (next == NULL) { 5486 next = list_head(l2arc_dev_list); 5487 } else { 5488 next = list_next(l2arc_dev_list, next); 5489 if (next == NULL) 5490 next = list_head(l2arc_dev_list); 5491 } 5492 5493 /* if we have come back to the start, bail out */ 5494 if (first == NULL) 5495 first = next; 5496 else if (next == first) 5497 break; 5498 5499 } while (vdev_is_dead(next->l2ad_vdev)); 5500 5501 /* if we were unable to find any usable vdevs, return NULL */ 5502 if (vdev_is_dead(next->l2ad_vdev)) 5503 next = NULL; 5504 5505 l2arc_dev_last = next; 5506 5507 out: 5508 mutex_exit(&l2arc_dev_mtx); 5509 5510 /* 5511 * Grab the config lock to prevent the 'next' device from being 5512 * removed while we are writing to it. 5513 */ 5514 if (next != NULL) 5515 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 5516 mutex_exit(&spa_namespace_lock); 5517 5518 return (next); 5519 } 5520 5521 /* 5522 * Free buffers that were tagged for destruction. 5523 */ 5524 static void 5525 l2arc_do_free_on_write() 5526 { 5527 list_t *buflist; 5528 l2arc_data_free_t *df, *df_prev; 5529 5530 mutex_enter(&l2arc_free_on_write_mtx); 5531 buflist = l2arc_free_on_write; 5532 5533 for (df = list_tail(buflist); df; df = df_prev) { 5534 df_prev = list_prev(buflist, df); 5535 ASSERT(df->l2df_data != NULL); 5536 ASSERT(df->l2df_func != NULL); 5537 df->l2df_func(df->l2df_data, df->l2df_size); 5538 list_remove(buflist, df); 5539 kmem_free(df, sizeof (l2arc_data_free_t)); 5540 } 5541 5542 mutex_exit(&l2arc_free_on_write_mtx); 5543 } 5544 5545 /* 5546 * A write to a cache device has completed. Update all headers to allow 5547 * reads from these buffers to begin. 5548 */ 5549 static void 5550 l2arc_write_done(zio_t *zio) 5551 { 5552 l2arc_write_callback_t *cb; 5553 l2arc_dev_t *dev; 5554 list_t *buflist; 5555 arc_buf_hdr_t *head, *hdr, *hdr_prev; 5556 kmutex_t *hash_lock; 5557 int64_t bytes_dropped = 0; 5558 5559 cb = zio->io_private; 5560 ASSERT(cb != NULL); 5561 dev = cb->l2wcb_dev; 5562 ASSERT(dev != NULL); 5563 head = cb->l2wcb_head; 5564 ASSERT(head != NULL); 5565 buflist = &dev->l2ad_buflist; 5566 ASSERT(buflist != NULL); 5567 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 5568 l2arc_write_callback_t *, cb); 5569 5570 if (zio->io_error != 0) 5571 ARCSTAT_BUMP(arcstat_l2_writes_error); 5572 5573 /* 5574 * All writes completed, or an error was hit. 5575 */ 5576 top: 5577 mutex_enter(&dev->l2ad_mtx); 5578 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 5579 hdr_prev = list_prev(buflist, hdr); 5580 5581 hash_lock = HDR_LOCK(hdr); 5582 5583 /* 5584 * We cannot use mutex_enter or else we can deadlock 5585 * with l2arc_write_buffers (due to swapping the order 5586 * the hash lock and l2ad_mtx are taken). 5587 */ 5588 if (!mutex_tryenter(hash_lock)) { 5589 /* 5590 * Missed the hash lock. We must retry so we 5591 * don't leave the ARC_FLAG_L2_WRITING bit set. 5592 */ 5593 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 5594 5595 /* 5596 * We don't want to rescan the headers we've 5597 * already marked as having been written out, so 5598 * we reinsert the head node so we can pick up 5599 * where we left off. 5600 */ 5601 list_remove(buflist, head); 5602 list_insert_after(buflist, hdr, head); 5603 5604 mutex_exit(&dev->l2ad_mtx); 5605 5606 /* 5607 * We wait for the hash lock to become available 5608 * to try and prevent busy waiting, and increase 5609 * the chance we'll be able to acquire the lock 5610 * the next time around. 5611 */ 5612 mutex_enter(hash_lock); 5613 mutex_exit(hash_lock); 5614 goto top; 5615 } 5616 5617 /* 5618 * We could not have been moved into the arc_l2c_only 5619 * state while in-flight due to our ARC_FLAG_L2_WRITING 5620 * bit being set. Let's just ensure that's being enforced. 5621 */ 5622 ASSERT(HDR_HAS_L1HDR(hdr)); 5623 5624 /* 5625 * We may have allocated a buffer for L2ARC compression, 5626 * we must release it to avoid leaking this data. 5627 */ 5628 l2arc_release_cdata_buf(hdr); 5629 5630 if (zio->io_error != 0) { 5631 /* 5632 * Error - drop L2ARC entry. 5633 */ 5634 list_remove(buflist, hdr); 5635 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 5636 5637 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 5638 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5639 5640 bytes_dropped += hdr->b_l2hdr.b_asize; 5641 (void) refcount_remove_many(&dev->l2ad_alloc, 5642 hdr->b_l2hdr.b_asize, hdr); 5643 } 5644 5645 /* 5646 * Allow ARC to begin reads and ghost list evictions to 5647 * this L2ARC entry. 5648 */ 5649 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5650 5651 mutex_exit(hash_lock); 5652 } 5653 5654 atomic_inc_64(&l2arc_writes_done); 5655 list_remove(buflist, head); 5656 ASSERT(!HDR_HAS_L1HDR(head)); 5657 kmem_cache_free(hdr_l2only_cache, head); 5658 mutex_exit(&dev->l2ad_mtx); 5659 5660 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 5661 5662 l2arc_do_free_on_write(); 5663 5664 kmem_free(cb, sizeof (l2arc_write_callback_t)); 5665 } 5666 5667 /* 5668 * A read to a cache device completed. Validate buffer contents before 5669 * handing over to the regular ARC routines. 5670 */ 5671 static void 5672 l2arc_read_done(zio_t *zio) 5673 { 5674 l2arc_read_callback_t *cb; 5675 arc_buf_hdr_t *hdr; 5676 arc_buf_t *buf; 5677 kmutex_t *hash_lock; 5678 int equal; 5679 5680 ASSERT(zio->io_vd != NULL); 5681 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 5682 5683 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 5684 5685 cb = zio->io_private; 5686 ASSERT(cb != NULL); 5687 buf = cb->l2rcb_buf; 5688 ASSERT(buf != NULL); 5689 5690 hash_lock = HDR_LOCK(buf->b_hdr); 5691 mutex_enter(hash_lock); 5692 hdr = buf->b_hdr; 5693 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5694 5695 /* 5696 * If the buffer was compressed, decompress it first. 5697 */ 5698 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 5699 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 5700 ASSERT(zio->io_data != NULL); 5701 ASSERT3U(zio->io_size, ==, hdr->b_size); 5702 ASSERT3U(BP_GET_LSIZE(&cb->l2rcb_bp), ==, hdr->b_size); 5703 5704 /* 5705 * Check this survived the L2ARC journey. 5706 */ 5707 equal = arc_cksum_equal(buf); 5708 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 5709 mutex_exit(hash_lock); 5710 zio->io_private = buf; 5711 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 5712 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 5713 arc_read_done(zio); 5714 } else { 5715 mutex_exit(hash_lock); 5716 /* 5717 * Buffer didn't survive caching. Increment stats and 5718 * reissue to the original storage device. 5719 */ 5720 if (zio->io_error != 0) { 5721 ARCSTAT_BUMP(arcstat_l2_io_error); 5722 } else { 5723 zio->io_error = SET_ERROR(EIO); 5724 } 5725 if (!equal) 5726 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 5727 5728 /* 5729 * If there's no waiter, issue an async i/o to the primary 5730 * storage now. If there *is* a waiter, the caller must 5731 * issue the i/o in a context where it's OK to block. 5732 */ 5733 if (zio->io_waiter == NULL) { 5734 zio_t *pio = zio_unique_parent(zio); 5735 5736 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 5737 5738 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 5739 buf->b_data, hdr->b_size, arc_read_done, buf, 5740 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 5741 } 5742 } 5743 5744 kmem_free(cb, sizeof (l2arc_read_callback_t)); 5745 } 5746 5747 /* 5748 * This is the list priority from which the L2ARC will search for pages to 5749 * cache. This is used within loops (0..3) to cycle through lists in the 5750 * desired order. This order can have a significant effect on cache 5751 * performance. 5752 * 5753 * Currently the metadata lists are hit first, MFU then MRU, followed by 5754 * the data lists. This function returns a locked list, and also returns 5755 * the lock pointer. 5756 */ 5757 static multilist_sublist_t * 5758 l2arc_sublist_lock(int list_num) 5759 { 5760 multilist_t *ml = NULL; 5761 unsigned int idx; 5762 5763 ASSERT(list_num >= 0 && list_num <= 3); 5764 5765 switch (list_num) { 5766 case 0: 5767 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 5768 break; 5769 case 1: 5770 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 5771 break; 5772 case 2: 5773 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 5774 break; 5775 case 3: 5776 ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; 5777 break; 5778 } 5779 5780 /* 5781 * Return a randomly-selected sublist. This is acceptable 5782 * because the caller feeds only a little bit of data for each 5783 * call (8MB). Subsequent calls will result in different 5784 * sublists being selected. 5785 */ 5786 idx = multilist_get_random_index(ml); 5787 return (multilist_sublist_lock(ml, idx)); 5788 } 5789 5790 /* 5791 * Evict buffers from the device write hand to the distance specified in 5792 * bytes. This distance may span populated buffers, it may span nothing. 5793 * This is clearing a region on the L2ARC device ready for writing. 5794 * If the 'all' boolean is set, every buffer is evicted. 5795 */ 5796 static void 5797 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 5798 { 5799 list_t *buflist; 5800 arc_buf_hdr_t *hdr, *hdr_prev; 5801 kmutex_t *hash_lock; 5802 uint64_t taddr; 5803 5804 buflist = &dev->l2ad_buflist; 5805 5806 if (!all && dev->l2ad_first) { 5807 /* 5808 * This is the first sweep through the device. There is 5809 * nothing to evict. 5810 */ 5811 return; 5812 } 5813 5814 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 5815 /* 5816 * When nearing the end of the device, evict to the end 5817 * before the device write hand jumps to the start. 5818 */ 5819 taddr = dev->l2ad_end; 5820 } else { 5821 taddr = dev->l2ad_hand + distance; 5822 } 5823 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 5824 uint64_t, taddr, boolean_t, all); 5825 5826 top: 5827 mutex_enter(&dev->l2ad_mtx); 5828 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 5829 hdr_prev = list_prev(buflist, hdr); 5830 5831 hash_lock = HDR_LOCK(hdr); 5832 5833 /* 5834 * We cannot use mutex_enter or else we can deadlock 5835 * with l2arc_write_buffers (due to swapping the order 5836 * the hash lock and l2ad_mtx are taken). 5837 */ 5838 if (!mutex_tryenter(hash_lock)) { 5839 /* 5840 * Missed the hash lock. Retry. 5841 */ 5842 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 5843 mutex_exit(&dev->l2ad_mtx); 5844 mutex_enter(hash_lock); 5845 mutex_exit(hash_lock); 5846 goto top; 5847 } 5848 5849 if (HDR_L2_WRITE_HEAD(hdr)) { 5850 /* 5851 * We hit a write head node. Leave it for 5852 * l2arc_write_done(). 5853 */ 5854 list_remove(buflist, hdr); 5855 mutex_exit(hash_lock); 5856 continue; 5857 } 5858 5859 if (!all && HDR_HAS_L2HDR(hdr) && 5860 (hdr->b_l2hdr.b_daddr > taddr || 5861 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 5862 /* 5863 * We've evicted to the target address, 5864 * or the end of the device. 5865 */ 5866 mutex_exit(hash_lock); 5867 break; 5868 } 5869 5870 ASSERT(HDR_HAS_L2HDR(hdr)); 5871 if (!HDR_HAS_L1HDR(hdr)) { 5872 ASSERT(!HDR_L2_READING(hdr)); 5873 /* 5874 * This doesn't exist in the ARC. Destroy. 5875 * arc_hdr_destroy() will call list_remove() 5876 * and decrement arcstat_l2_size. 5877 */ 5878 arc_change_state(arc_anon, hdr, hash_lock); 5879 arc_hdr_destroy(hdr); 5880 } else { 5881 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 5882 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 5883 /* 5884 * Invalidate issued or about to be issued 5885 * reads, since we may be about to write 5886 * over this location. 5887 */ 5888 if (HDR_L2_READING(hdr)) { 5889 ARCSTAT_BUMP(arcstat_l2_evict_reading); 5890 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 5891 } 5892 5893 /* Ensure this header has finished being written */ 5894 ASSERT(!HDR_L2_WRITING(hdr)); 5895 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 5896 5897 arc_hdr_l2hdr_destroy(hdr); 5898 } 5899 mutex_exit(hash_lock); 5900 } 5901 mutex_exit(&dev->l2ad_mtx); 5902 } 5903 5904 /* 5905 * Find and write ARC buffers to the L2ARC device. 5906 * 5907 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 5908 * for reading until they have completed writing. 5909 * The headroom_boost is an in-out parameter used to maintain headroom boost 5910 * state between calls to this function. 5911 * 5912 * Returns the number of bytes actually written (which may be smaller than 5913 * the delta by which the device hand has changed due to alignment). 5914 */ 5915 static uint64_t 5916 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5917 boolean_t *headroom_boost) 5918 { 5919 arc_buf_hdr_t *hdr, *hdr_prev, *head; 5920 uint64_t write_asize, write_sz, headroom, 5921 buf_compress_minsz; 5922 void *buf_data; 5923 boolean_t full; 5924 l2arc_write_callback_t *cb; 5925 zio_t *pio, *wzio; 5926 uint64_t guid = spa_load_guid(spa); 5927 uint64_t sync_txg = spa_syncing_txg(spa); 5928 const boolean_t do_headroom_boost = *headroom_boost; 5929 5930 ASSERT(dev->l2ad_vdev != NULL); 5931 5932 /* Lower the flag now, we might want to raise it again later. */ 5933 *headroom_boost = B_FALSE; 5934 5935 pio = NULL; 5936 write_sz = write_asize = 0; 5937 full = B_FALSE; 5938 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 5939 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 5940 head->b_flags |= ARC_FLAG_HAS_L2HDR; 5941 5942 /* 5943 * We will want to try to compress buffers that are at least 2x the 5944 * device sector size. 5945 */ 5946 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5947 5948 /* 5949 * Copy buffers for L2ARC writing. 5950 */ 5951 for (int try = 0; try <= 3; try++) { 5952 multilist_sublist_t *mls = l2arc_sublist_lock(try); 5953 uint64_t passed_sz = 0; 5954 5955 /* 5956 * L2ARC fast warmup. 5957 * 5958 * Until the ARC is warm and starts to evict, read from the 5959 * head of the ARC lists rather than the tail. 5960 */ 5961 if (arc_warm == B_FALSE) 5962 hdr = multilist_sublist_head(mls); 5963 else 5964 hdr = multilist_sublist_tail(mls); 5965 5966 headroom = target_sz * l2arc_headroom; 5967 if (do_headroom_boost) 5968 headroom = (headroom * l2arc_headroom_boost) / 100; 5969 5970 for (; hdr; hdr = hdr_prev) { 5971 kmutex_t *hash_lock; 5972 uint64_t buf_sz; 5973 uint64_t buf_a_sz; 5974 5975 if (arc_warm == B_FALSE) 5976 hdr_prev = multilist_sublist_next(mls, hdr); 5977 else 5978 hdr_prev = multilist_sublist_prev(mls, hdr); 5979 5980 hash_lock = HDR_LOCK(hdr); 5981 if (!mutex_tryenter(hash_lock)) { 5982 /* 5983 * Skip this buffer rather than waiting. 5984 */ 5985 continue; 5986 } 5987 5988 passed_sz += hdr->b_size; 5989 if (passed_sz > headroom) { 5990 /* 5991 * Searched too far. 5992 */ 5993 mutex_exit(hash_lock); 5994 break; 5995 } 5996 5997 if (!l2arc_write_eligible(guid, sync_txg, hdr)) { 5998 mutex_exit(hash_lock); 5999 continue; 6000 } 6001 6002 /* 6003 * Assume that the buffer is not going to be compressed 6004 * and could take more space on disk because of a larger 6005 * disk block size. 6006 */ 6007 buf_sz = hdr->b_size; 6008 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 6009 6010 if ((write_asize + buf_a_sz) > target_sz) { 6011 full = B_TRUE; 6012 mutex_exit(hash_lock); 6013 break; 6014 } 6015 6016 if (pio == NULL) { 6017 /* 6018 * Insert a dummy header on the buflist so 6019 * l2arc_write_done() can find where the 6020 * write buffers begin without searching. 6021 */ 6022 mutex_enter(&dev->l2ad_mtx); 6023 list_insert_head(&dev->l2ad_buflist, head); 6024 mutex_exit(&dev->l2ad_mtx); 6025 6026 cb = kmem_alloc( 6027 sizeof (l2arc_write_callback_t), KM_SLEEP); 6028 cb->l2wcb_dev = dev; 6029 cb->l2wcb_head = head; 6030 pio = zio_root(spa, l2arc_write_done, cb, 6031 ZIO_FLAG_CANFAIL); 6032 } 6033 6034 /* 6035 * Create and add a new L2ARC header. 6036 */ 6037 hdr->b_l2hdr.b_dev = dev; 6038 hdr->b_flags |= ARC_FLAG_L2_WRITING; 6039 /* 6040 * Temporarily stash the data buffer in b_tmp_cdata. 6041 * The subsequent write step will pick it up from 6042 * there. This is because can't access b_l1hdr.b_buf 6043 * without holding the hash_lock, which we in turn 6044 * can't access without holding the ARC list locks 6045 * (which we want to avoid during compression/writing). 6046 */ 6047 hdr->b_l2hdr.b_compress = ZIO_COMPRESS_OFF; 6048 hdr->b_l2hdr.b_asize = hdr->b_size; 6049 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data; 6050 6051 /* 6052 * Explicitly set the b_daddr field to a known 6053 * value which means "invalid address". This 6054 * enables us to differentiate which stage of 6055 * l2arc_write_buffers() the particular header 6056 * is in (e.g. this loop, or the one below). 6057 * ARC_FLAG_L2_WRITING is not enough to make 6058 * this distinction, and we need to know in 6059 * order to do proper l2arc vdev accounting in 6060 * arc_release() and arc_hdr_destroy(). 6061 * 6062 * Note, we can't use a new flag to distinguish 6063 * the two stages because we don't hold the 6064 * header's hash_lock below, in the second stage 6065 * of this function. Thus, we can't simply 6066 * change the b_flags field to denote that the 6067 * IO has been sent. We can change the b_daddr 6068 * field of the L2 portion, though, since we'll 6069 * be holding the l2ad_mtx; which is why we're 6070 * using it to denote the header's state change. 6071 */ 6072 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET; 6073 6074 hdr->b_flags |= ARC_FLAG_HAS_L2HDR; 6075 6076 mutex_enter(&dev->l2ad_mtx); 6077 list_insert_head(&dev->l2ad_buflist, hdr); 6078 mutex_exit(&dev->l2ad_mtx); 6079 6080 /* 6081 * Compute and store the buffer cksum before 6082 * writing. On debug the cksum is verified first. 6083 */ 6084 arc_cksum_verify(hdr->b_l1hdr.b_buf); 6085 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE); 6086 6087 mutex_exit(hash_lock); 6088 6089 write_sz += buf_sz; 6090 write_asize += buf_a_sz; 6091 } 6092 6093 multilist_sublist_unlock(mls); 6094 6095 if (full == B_TRUE) 6096 break; 6097 } 6098 6099 /* No buffers selected for writing? */ 6100 if (pio == NULL) { 6101 ASSERT0(write_sz); 6102 ASSERT(!HDR_HAS_L1HDR(head)); 6103 kmem_cache_free(hdr_l2only_cache, head); 6104 return (0); 6105 } 6106 6107 mutex_enter(&dev->l2ad_mtx); 6108 6109 /* 6110 * Note that elsewhere in this file arcstat_l2_asize 6111 * and the used space on l2ad_vdev are updated using b_asize, 6112 * which is not necessarily rounded up to the device block size. 6113 * Too keep accounting consistent we do the same here as well: 6114 * stats_size accumulates the sum of b_asize of the written buffers, 6115 * while write_asize accumulates the sum of b_asize rounded up 6116 * to the device block size. 6117 * The latter sum is used only to validate the corectness of the code. 6118 */ 6119 uint64_t stats_size = 0; 6120 write_asize = 0; 6121 6122 /* 6123 * Now start writing the buffers. We're starting at the write head 6124 * and work backwards, retracing the course of the buffer selector 6125 * loop above. 6126 */ 6127 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr; 6128 hdr = list_prev(&dev->l2ad_buflist, hdr)) { 6129 uint64_t buf_sz; 6130 6131 /* 6132 * We rely on the L1 portion of the header below, so 6133 * it's invalid for this header to have been evicted out 6134 * of the ghost cache, prior to being written out. The 6135 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 6136 */ 6137 ASSERT(HDR_HAS_L1HDR(hdr)); 6138 6139 /* 6140 * We shouldn't need to lock the buffer here, since we flagged 6141 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 6142 * take care to only access its L2 cache parameters. In 6143 * particular, hdr->l1hdr.b_buf may be invalid by now due to 6144 * ARC eviction. 6145 */ 6146 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 6147 6148 if ((HDR_L2COMPRESS(hdr)) && 6149 hdr->b_l2hdr.b_asize >= buf_compress_minsz) { 6150 if (l2arc_compress_buf(hdr)) { 6151 /* 6152 * If compression succeeded, enable headroom 6153 * boost on the next scan cycle. 6154 */ 6155 *headroom_boost = B_TRUE; 6156 } 6157 } 6158 6159 /* 6160 * Pick up the buffer data we had previously stashed away 6161 * (and now potentially also compressed). 6162 */ 6163 buf_data = hdr->b_l1hdr.b_tmp_cdata; 6164 buf_sz = hdr->b_l2hdr.b_asize; 6165 6166 /* 6167 * We need to do this regardless if buf_sz is zero or 6168 * not, otherwise, when this l2hdr is evicted we'll 6169 * remove a reference that was never added. 6170 */ 6171 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr); 6172 6173 /* Compression may have squashed the buffer to zero length. */ 6174 if (buf_sz != 0) { 6175 uint64_t buf_a_sz; 6176 6177 wzio = zio_write_phys(pio, dev->l2ad_vdev, 6178 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 6179 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 6180 ZIO_FLAG_CANFAIL, B_FALSE); 6181 6182 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 6183 zio_t *, wzio); 6184 (void) zio_nowait(wzio); 6185 6186 stats_size += buf_sz; 6187 6188 /* 6189 * Keep the clock hand suitably device-aligned. 6190 */ 6191 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 6192 write_asize += buf_a_sz; 6193 dev->l2ad_hand += buf_a_sz; 6194 } 6195 } 6196 6197 mutex_exit(&dev->l2ad_mtx); 6198 6199 ASSERT3U(write_asize, <=, target_sz); 6200 ARCSTAT_BUMP(arcstat_l2_writes_sent); 6201 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 6202 ARCSTAT_INCR(arcstat_l2_size, write_sz); 6203 ARCSTAT_INCR(arcstat_l2_asize, stats_size); 6204 vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0); 6205 6206 /* 6207 * Bump device hand to the device start if it is approaching the end. 6208 * l2arc_evict() will already have evicted ahead for this case. 6209 */ 6210 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 6211 dev->l2ad_hand = dev->l2ad_start; 6212 dev->l2ad_first = B_FALSE; 6213 } 6214 6215 dev->l2ad_writing = B_TRUE; 6216 (void) zio_wait(pio); 6217 dev->l2ad_writing = B_FALSE; 6218 6219 return (write_asize); 6220 } 6221 6222 /* 6223 * Compresses an L2ARC buffer. 6224 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its 6225 * size in l2hdr->b_asize. This routine tries to compress the data and 6226 * depending on the compression result there are three possible outcomes: 6227 * *) The buffer was incompressible. The original l2hdr contents were left 6228 * untouched and are ready for writing to an L2 device. 6229 * *) The buffer was all-zeros, so there is no need to write it to an L2 6230 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 6231 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 6232 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 6233 * data buffer which holds the compressed data to be written, and b_asize 6234 * tells us how much data there is. b_compress is set to the appropriate 6235 * compression algorithm. Once writing is done, invoke 6236 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 6237 * 6238 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 6239 * buffer was incompressible). 6240 */ 6241 static boolean_t 6242 l2arc_compress_buf(arc_buf_hdr_t *hdr) 6243 { 6244 void *cdata; 6245 size_t csize, len, rounded; 6246 ASSERT(HDR_HAS_L2HDR(hdr)); 6247 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 6248 6249 ASSERT(HDR_HAS_L1HDR(hdr)); 6250 ASSERT3S(l2hdr->b_compress, ==, ZIO_COMPRESS_OFF); 6251 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6252 6253 len = l2hdr->b_asize; 6254 cdata = zio_data_buf_alloc(len); 6255 ASSERT3P(cdata, !=, NULL); 6256 csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata, 6257 cdata, l2hdr->b_asize); 6258 6259 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); 6260 if (rounded > csize) { 6261 bzero((char *)cdata + csize, rounded - csize); 6262 csize = rounded; 6263 } 6264 6265 if (csize == 0) { 6266 /* zero block, indicate that there's nothing to write */ 6267 zio_data_buf_free(cdata, len); 6268 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 6269 l2hdr->b_asize = 0; 6270 hdr->b_l1hdr.b_tmp_cdata = NULL; 6271 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 6272 return (B_TRUE); 6273 } else if (csize > 0 && csize < len) { 6274 /* 6275 * Compression succeeded, we'll keep the cdata around for 6276 * writing and release it afterwards. 6277 */ 6278 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 6279 l2hdr->b_asize = csize; 6280 hdr->b_l1hdr.b_tmp_cdata = cdata; 6281 ARCSTAT_BUMP(arcstat_l2_compress_successes); 6282 return (B_TRUE); 6283 } else { 6284 /* 6285 * Compression failed, release the compressed buffer. 6286 * l2hdr will be left unmodified. 6287 */ 6288 zio_data_buf_free(cdata, len); 6289 ARCSTAT_BUMP(arcstat_l2_compress_failures); 6290 return (B_FALSE); 6291 } 6292 } 6293 6294 /* 6295 * Decompresses a zio read back from an l2arc device. On success, the 6296 * underlying zio's io_data buffer is overwritten by the uncompressed 6297 * version. On decompression error (corrupt compressed stream), the 6298 * zio->io_error value is set to signal an I/O error. 6299 * 6300 * Please note that the compressed data stream is not checksummed, so 6301 * if the underlying device is experiencing data corruption, we may feed 6302 * corrupt data to the decompressor, so the decompressor needs to be 6303 * able to handle this situation (LZ4 does). 6304 */ 6305 static void 6306 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 6307 { 6308 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 6309 6310 if (zio->io_error != 0) { 6311 /* 6312 * An io error has occured, just restore the original io 6313 * size in preparation for a main pool read. 6314 */ 6315 zio->io_orig_size = zio->io_size = hdr->b_size; 6316 return; 6317 } 6318 6319 if (c == ZIO_COMPRESS_EMPTY) { 6320 /* 6321 * An empty buffer results in a null zio, which means we 6322 * need to fill its io_data after we're done restoring the 6323 * buffer's contents. 6324 */ 6325 ASSERT(hdr->b_l1hdr.b_buf != NULL); 6326 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size); 6327 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data; 6328 } else { 6329 ASSERT(zio->io_data != NULL); 6330 /* 6331 * We copy the compressed data from the start of the arc buffer 6332 * (the zio_read will have pulled in only what we need, the 6333 * rest is garbage which we will overwrite at decompression) 6334 * and then decompress back to the ARC data buffer. This way we 6335 * can minimize copying by simply decompressing back over the 6336 * original compressed data (rather than decompressing to an 6337 * aux buffer and then copying back the uncompressed buffer, 6338 * which is likely to be much larger). 6339 */ 6340 uint64_t csize; 6341 void *cdata; 6342 6343 csize = zio->io_size; 6344 cdata = zio_data_buf_alloc(csize); 6345 bcopy(zio->io_data, cdata, csize); 6346 if (zio_decompress_data(c, cdata, zio->io_data, csize, 6347 hdr->b_size) != 0) 6348 zio->io_error = EIO; 6349 zio_data_buf_free(cdata, csize); 6350 } 6351 6352 /* Restore the expected uncompressed IO size. */ 6353 zio->io_orig_size = zio->io_size = hdr->b_size; 6354 } 6355 6356 /* 6357 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 6358 * This buffer serves as a temporary holder of compressed data while 6359 * the buffer entry is being written to an l2arc device. Once that is 6360 * done, we can dispose of it. 6361 */ 6362 static void 6363 l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 6364 { 6365 ASSERT(HDR_HAS_L2HDR(hdr)); 6366 enum zio_compress comp = hdr->b_l2hdr.b_compress; 6367 6368 ASSERT(HDR_HAS_L1HDR(hdr)); 6369 ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp)); 6370 6371 if (comp == ZIO_COMPRESS_OFF) { 6372 /* 6373 * In this case, b_tmp_cdata points to the same buffer 6374 * as the arc_buf_t's b_data field. We don't want to 6375 * free it, since the arc_buf_t will handle that. 6376 */ 6377 hdr->b_l1hdr.b_tmp_cdata = NULL; 6378 } else if (comp == ZIO_COMPRESS_EMPTY) { 6379 /* 6380 * In this case, b_tmp_cdata was compressed to an empty 6381 * buffer, thus there's nothing to free and b_tmp_cdata 6382 * should have been set to NULL in l2arc_write_buffers(). 6383 */ 6384 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 6385 } else { 6386 /* 6387 * If the data was compressed, then we've allocated a 6388 * temporary buffer for it, so now we need to release it. 6389 */ 6390 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6391 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, 6392 hdr->b_size); 6393 hdr->b_l1hdr.b_tmp_cdata = NULL; 6394 } 6395 6396 } 6397 6398 /* 6399 * This thread feeds the L2ARC at regular intervals. This is the beating 6400 * heart of the L2ARC. 6401 */ 6402 static void 6403 l2arc_feed_thread(void) 6404 { 6405 callb_cpr_t cpr; 6406 l2arc_dev_t *dev; 6407 spa_t *spa; 6408 uint64_t size, wrote; 6409 clock_t begin, next = ddi_get_lbolt(); 6410 boolean_t headroom_boost = B_FALSE; 6411 6412 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 6413 6414 mutex_enter(&l2arc_feed_thr_lock); 6415 6416 while (l2arc_thread_exit == 0) { 6417 CALLB_CPR_SAFE_BEGIN(&cpr); 6418 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 6419 next); 6420 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 6421 next = ddi_get_lbolt() + hz; 6422 6423 /* 6424 * Quick check for L2ARC devices. 6425 */ 6426 mutex_enter(&l2arc_dev_mtx); 6427 if (l2arc_ndev == 0) { 6428 mutex_exit(&l2arc_dev_mtx); 6429 continue; 6430 } 6431 mutex_exit(&l2arc_dev_mtx); 6432 begin = ddi_get_lbolt(); 6433 6434 /* 6435 * This selects the next l2arc device to write to, and in 6436 * doing so the next spa to feed from: dev->l2ad_spa. This 6437 * will return NULL if there are now no l2arc devices or if 6438 * they are all faulted. 6439 * 6440 * If a device is returned, its spa's config lock is also 6441 * held to prevent device removal. l2arc_dev_get_next() 6442 * will grab and release l2arc_dev_mtx. 6443 */ 6444 if ((dev = l2arc_dev_get_next()) == NULL) 6445 continue; 6446 6447 spa = dev->l2ad_spa; 6448 ASSERT(spa != NULL); 6449 6450 /* 6451 * If the pool is read-only then force the feed thread to 6452 * sleep a little longer. 6453 */ 6454 if (!spa_writeable(spa)) { 6455 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 6456 spa_config_exit(spa, SCL_L2ARC, dev); 6457 continue; 6458 } 6459 6460 /* 6461 * Avoid contributing to memory pressure. 6462 */ 6463 if (arc_reclaim_needed()) { 6464 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 6465 spa_config_exit(spa, SCL_L2ARC, dev); 6466 continue; 6467 } 6468 6469 ARCSTAT_BUMP(arcstat_l2_feeds); 6470 6471 size = l2arc_write_size(); 6472 6473 /* 6474 * Evict L2ARC buffers that will be overwritten. 6475 */ 6476 l2arc_evict(dev, size, B_FALSE); 6477 6478 /* 6479 * Write ARC buffers. 6480 */ 6481 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 6482 6483 /* 6484 * Calculate interval between writes. 6485 */ 6486 next = l2arc_write_interval(begin, size, wrote); 6487 spa_config_exit(spa, SCL_L2ARC, dev); 6488 } 6489 6490 l2arc_thread_exit = 0; 6491 cv_broadcast(&l2arc_feed_thr_cv); 6492 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 6493 thread_exit(); 6494 } 6495 6496 boolean_t 6497 l2arc_vdev_present(vdev_t *vd) 6498 { 6499 l2arc_dev_t *dev; 6500 6501 mutex_enter(&l2arc_dev_mtx); 6502 for (dev = list_head(l2arc_dev_list); dev != NULL; 6503 dev = list_next(l2arc_dev_list, dev)) { 6504 if (dev->l2ad_vdev == vd) 6505 break; 6506 } 6507 mutex_exit(&l2arc_dev_mtx); 6508 6509 return (dev != NULL); 6510 } 6511 6512 /* 6513 * Add a vdev for use by the L2ARC. By this point the spa has already 6514 * validated the vdev and opened it. 6515 */ 6516 void 6517 l2arc_add_vdev(spa_t *spa, vdev_t *vd) 6518 { 6519 l2arc_dev_t *adddev; 6520 6521 ASSERT(!l2arc_vdev_present(vd)); 6522 6523 /* 6524 * Create a new l2arc device entry. 6525 */ 6526 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 6527 adddev->l2ad_spa = spa; 6528 adddev->l2ad_vdev = vd; 6529 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 6530 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 6531 adddev->l2ad_hand = adddev->l2ad_start; 6532 adddev->l2ad_first = B_TRUE; 6533 adddev->l2ad_writing = B_FALSE; 6534 6535 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 6536 /* 6537 * This is a list of all ARC buffers that are still valid on the 6538 * device. 6539 */ 6540 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 6541 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 6542 6543 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 6544 refcount_create(&adddev->l2ad_alloc); 6545 6546 /* 6547 * Add device to global list 6548 */ 6549 mutex_enter(&l2arc_dev_mtx); 6550 list_insert_head(l2arc_dev_list, adddev); 6551 atomic_inc_64(&l2arc_ndev); 6552 mutex_exit(&l2arc_dev_mtx); 6553 } 6554 6555 /* 6556 * Remove a vdev from the L2ARC. 6557 */ 6558 void 6559 l2arc_remove_vdev(vdev_t *vd) 6560 { 6561 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 6562 6563 /* 6564 * Find the device by vdev 6565 */ 6566 mutex_enter(&l2arc_dev_mtx); 6567 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 6568 nextdev = list_next(l2arc_dev_list, dev); 6569 if (vd == dev->l2ad_vdev) { 6570 remdev = dev; 6571 break; 6572 } 6573 } 6574 ASSERT(remdev != NULL); 6575 6576 /* 6577 * Remove device from global list 6578 */ 6579 list_remove(l2arc_dev_list, remdev); 6580 l2arc_dev_last = NULL; /* may have been invalidated */ 6581 atomic_dec_64(&l2arc_ndev); 6582 mutex_exit(&l2arc_dev_mtx); 6583 6584 /* 6585 * Clear all buflists and ARC references. L2ARC device flush. 6586 */ 6587 l2arc_evict(remdev, 0, B_TRUE); 6588 list_destroy(&remdev->l2ad_buflist); 6589 mutex_destroy(&remdev->l2ad_mtx); 6590 refcount_destroy(&remdev->l2ad_alloc); 6591 kmem_free(remdev, sizeof (l2arc_dev_t)); 6592 } 6593 6594 void 6595 l2arc_init(void) 6596 { 6597 l2arc_thread_exit = 0; 6598 l2arc_ndev = 0; 6599 l2arc_writes_sent = 0; 6600 l2arc_writes_done = 0; 6601 6602 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 6603 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 6604 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 6605 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 6606 6607 l2arc_dev_list = &L2ARC_dev_list; 6608 l2arc_free_on_write = &L2ARC_free_on_write; 6609 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 6610 offsetof(l2arc_dev_t, l2ad_node)); 6611 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 6612 offsetof(l2arc_data_free_t, l2df_list_node)); 6613 } 6614 6615 void 6616 l2arc_fini(void) 6617 { 6618 /* 6619 * This is called from dmu_fini(), which is called from spa_fini(); 6620 * Because of this, we can assume that all l2arc devices have 6621 * already been removed when the pools themselves were removed. 6622 */ 6623 6624 l2arc_do_free_on_write(); 6625 6626 mutex_destroy(&l2arc_feed_thr_lock); 6627 cv_destroy(&l2arc_feed_thr_cv); 6628 mutex_destroy(&l2arc_dev_mtx); 6629 mutex_destroy(&l2arc_free_on_write_mtx); 6630 6631 list_destroy(l2arc_dev_list); 6632 list_destroy(l2arc_free_on_write); 6633 } 6634 6635 void 6636 l2arc_start(void) 6637 { 6638 if (!(spa_mode_global & FWRITE)) 6639 return; 6640 6641 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 6642 TS_RUN, minclsyspri); 6643 } 6644 6645 void 6646 l2arc_stop(void) 6647 { 6648 if (!(spa_mode_global & FWRITE)) 6649 return; 6650 6651 mutex_enter(&l2arc_feed_thr_lock); 6652 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 6653 l2arc_thread_exit = 1; 6654 while (l2arc_thread_exit != 0) 6655 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 6656 mutex_exit(&l2arc_feed_thr_lock); 6657 } 6658 6659 /* 6660 * dump arc cache to user mode for debugging purposes 6661 */ 6662 static void 6663 arc_dump_entry(arc_buf_hdr_t *entry, arc_info_t *outp) 6664 { 6665 outp->ai_dva = entry->b_dva; 6666 outp->ai_birth = entry->b_birth; 6667 outp->ai_flags = entry->b_flags; 6668 outp->ai_spa = entry->b_spa; 6669 outp->ai_size = entry->b_size; 6670 if (HDR_HAS_L1HDR(entry)) { 6671 arc_state_t *state = entry->b_l1hdr.b_state; 6672 if (state == arc_anon) 6673 outp->ai_state = AIS_ANON; 6674 else if (state == arc_mru) 6675 outp->ai_state = AIS_MRU; 6676 else if (state == arc_mru_ghost) 6677 outp->ai_state = AIS_MRU_GHOST; 6678 else if (state == arc_mfu) 6679 outp->ai_state = AIS_MFU; 6680 else if (state == arc_mfu_ghost) 6681 outp->ai_state = AIS_MFU_GHOST; 6682 else if (state == arc_l2c_only) 6683 outp->ai_state = AIS_L2C_ONLY; 6684 else 6685 outp->ai_state = AIS_UNKNOWN; 6686 } else { 6687 outp->ai_state = AIS_NO_L1HDR; 6688 } 6689 } 6690 6691 int 6692 arc_dump(int start_bucket, void *buf, size_t bufsize, size_t *returned_bytes) 6693 { 6694 int i; 6695 arc_info_t *outp = buf + sizeof(arc_info_hdr_t); 6696 arc_info_t *maxp = buf + bufsize; 6697 arc_info_hdr_t *aih = buf; 6698 size_t nbuckets = buf_hash_table.ht_mask + 1; 6699 size_t bph = nbuckets / BUF_LOCKS; /* buckets per hash */ 6700 kmutex_t *last_lock = NULL; 6701 6702 if (bufsize < sizeof(arc_info_hdr_t)) 6703 return (ENOMEM); 6704 6705 aih->aih_buckets = nbuckets; 6706 aih->aih_buf_locks = BUF_LOCKS; 6707 6708 ASSERT(start_bucket >= 0); 6709 ASSERT(start_bucket < nbuckets); 6710 6711 for (i = start_bucket; i < nbuckets; ++i) { 6712 kmutex_t *hash_lock; 6713 arc_buf_hdr_t *entry; 6714 arc_info_t *dryrun = outp; 6715 int bucket; 6716 6717 /* 6718 * transform index. We want to enumerate the buckets in an 6719 * order that allows us to keep the mutex as long as possible 6720 */ 6721 bucket = (i / bph) + (i % bph) * BUF_LOCKS; 6722 6723 hash_lock = BUF_HASH_LOCK(bucket); 6724 if (hash_lock != last_lock) { 6725 if (last_lock) 6726 mutex_exit(last_lock); 6727 mutex_enter(hash_lock); 6728 } 6729 last_lock = hash_lock; 6730 /* count entries to see if they will fit */ 6731 entry = buf_hash_table.ht_table[bucket]; 6732 while (entry != NULL) { 6733 ++dryrun; 6734 entry = entry->b_hash_next; 6735 } 6736 if (dryrun > maxp) { 6737 break; 6738 } 6739 /* actually copy entries */ 6740 entry = buf_hash_table.ht_table[bucket]; 6741 while (entry != NULL) { 6742 arc_dump_entry(entry, outp); 6743 ++outp; 6744 entry = entry->b_hash_next; 6745 } 6746 } 6747 if (last_lock) 6748 mutex_exit(last_lock); 6749 6750 *returned_bytes = (void *)outp - buf; 6751 aih->aih_entries = (*returned_bytes - sizeof(*aih)) / sizeof(*outp); 6752 6753 if (i <= buf_hash_table.ht_mask) 6754 aih->aih_next = i; 6755 else 6756 aih->aih_next = 0; 6757 6758 return (0); 6759 } 6760