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 2011 Nexenta Systems, Inc. All rights reserved. 24 * Copyright (c) 2013 by Delphix. All rights reserved. 25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. 26 */ 27 28 /* 29 * DVA-based Adjustable Replacement Cache 30 * 31 * While much of the theory of operation used here is 32 * based on the self-tuning, low overhead replacement cache 33 * presented by Megiddo and Modha at FAST 2003, there are some 34 * significant differences: 35 * 36 * 1. The Megiddo and Modha model assumes any page is evictable. 37 * Pages in its cache cannot be "locked" into memory. This makes 38 * the eviction algorithm simple: evict the last page in the list. 39 * This also make the performance characteristics easy to reason 40 * about. Our cache is not so simple. At any given moment, some 41 * subset of the blocks in the cache are un-evictable because we 42 * have handed out a reference to them. Blocks are only evictable 43 * when there are no external references active. This makes 44 * eviction far more problematic: we choose to evict the evictable 45 * blocks that are the "lowest" in the list. 46 * 47 * There are times when it is not possible to evict the requested 48 * space. In these circumstances we are unable to adjust the cache 49 * size. To prevent the cache growing unbounded at these times we 50 * implement a "cache throttle" that slows the flow of new data 51 * into the cache until we can make space available. 52 * 53 * 2. The Megiddo and Modha model assumes a fixed cache size. 54 * Pages are evicted when the cache is full and there is a cache 55 * miss. Our model has a variable sized cache. It grows with 56 * high use, but also tries to react to memory pressure from the 57 * operating system: decreasing its size when system memory is 58 * tight. 59 * 60 * 3. The Megiddo and Modha model assumes a fixed page size. All 61 * elements of the cache are therefor exactly the same size. So 62 * when adjusting the cache size following a cache miss, its simply 63 * a matter of choosing a single page to evict. In our model, we 64 * have variable sized cache blocks (rangeing from 512 bytes to 65 * 128K bytes). We therefor choose a set of blocks to evict to make 66 * space for a cache miss that approximates as closely as possible 67 * the space used by the new block. 68 * 69 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 70 * by N. Megiddo & D. Modha, FAST 2003 71 */ 72 73 /* 74 * The locking model: 75 * 76 * A new reference to a cache buffer can be obtained in two 77 * ways: 1) via a hash table lookup using the DVA as a key, 78 * or 2) via one of the ARC lists. The arc_read() interface 79 * uses method 1, while the internal arc algorithms for 80 * adjusting the cache use method 2. We therefor provide two 81 * types of locks: 1) the hash table lock array, and 2) the 82 * arc list locks. 83 * 84 * Buffers do not have their own mutexes, rather they rely on the 85 * hash table mutexes for the bulk of their protection (i.e. most 86 * fields in the arc_buf_hdr_t are protected by these mutexes). 87 * 88 * buf_hash_find() returns the appropriate mutex (held) when it 89 * locates the requested buffer in the hash table. It returns 90 * NULL for the mutex if the buffer was not in the table. 91 * 92 * buf_hash_remove() expects the appropriate hash mutex to be 93 * already held before it is invoked. 94 * 95 * Each arc state also has a mutex which is used to protect the 96 * buffer list associated with the state. When attempting to 97 * obtain a hash table lock while holding an arc list lock you 98 * must use: mutex_tryenter() to avoid deadlock. Also note that 99 * the active state mutex must be held before the ghost state mutex. 100 * 101 * Arc buffers may have an associated eviction callback function. 102 * This function will be invoked prior to removing the buffer (e.g. 103 * in arc_do_user_evicts()). Note however that the data associated 104 * with the buffer may be evicted prior to the callback. The callback 105 * must be made with *no locks held* (to prevent deadlock). Additionally, 106 * the users of callbacks must ensure that their private data is 107 * protected from simultaneous callbacks from arc_buf_evict() 108 * and arc_do_user_evicts(). 109 * 110 * Note that the majority of the performance stats are manipulated 111 * with atomic operations. 112 * 113 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 114 * 115 * - L2ARC buflist creation 116 * - L2ARC buflist eviction 117 * - L2ARC write completion, which walks L2ARC buflists 118 * - ARC header destruction, as it removes from L2ARC buflists 119 * - ARC header release, as it removes from L2ARC buflists 120 */ 121 122 #include <sys/spa.h> 123 #include <sys/zio.h> 124 #include <sys/zio_compress.h> 125 #include <sys/zfs_context.h> 126 #include <sys/arc.h> 127 #include <sys/refcount.h> 128 #include <sys/vdev.h> 129 #include <sys/vdev_impl.h> 130 #ifdef _KERNEL 131 #include <sys/vmsystm.h> 132 #include <vm/anon.h> 133 #include <sys/fs/swapnode.h> 134 #include <sys/dnlc.h> 135 #endif 136 #include <sys/callb.h> 137 #include <sys/kstat.h> 138 #include <zfs_fletcher.h> 139 140 #ifndef _KERNEL 141 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 142 boolean_t arc_watch = B_FALSE; 143 int arc_procfd; 144 #endif 145 146 static kmutex_t arc_reclaim_thr_lock; 147 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 148 static uint8_t arc_thread_exit; 149 150 extern int zfs_write_limit_shift; 151 extern uint64_t zfs_write_limit_max; 152 extern kmutex_t zfs_write_limit_lock; 153 154 #define ARC_REDUCE_DNLC_PERCENT 3 155 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 156 157 typedef enum arc_reclaim_strategy { 158 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 159 ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 160 } arc_reclaim_strategy_t; 161 162 /* number of seconds before growing cache again */ 163 static int arc_grow_retry = 60; 164 165 /* shift of arc_c for calculating both min and max arc_p */ 166 static int arc_p_min_shift = 4; 167 168 /* log2(fraction of arc to reclaim) */ 169 static int arc_shrink_shift = 5; 170 171 /* 172 * minimum lifespan of a prefetch block in clock ticks 173 * (initialized in arc_init()) 174 */ 175 static int arc_min_prefetch_lifespan; 176 177 static int arc_dead; 178 179 /* 180 * The arc has filled available memory and has now warmed up. 181 */ 182 static boolean_t arc_warm; 183 184 /* 185 * These tunables are for performance analysis. 186 */ 187 uint64_t zfs_arc_max; 188 uint64_t zfs_arc_min; 189 uint64_t zfs_arc_meta_limit = 0; 190 int zfs_arc_grow_retry = 0; 191 int zfs_arc_shrink_shift = 0; 192 int zfs_arc_p_min_shift = 0; 193 int zfs_disable_dup_eviction = 0; 194 195 /* 196 * Note that buffers can be in one of 6 states: 197 * ARC_anon - anonymous (discussed below) 198 * ARC_mru - recently used, currently cached 199 * ARC_mru_ghost - recentely used, no longer in cache 200 * ARC_mfu - frequently used, currently cached 201 * ARC_mfu_ghost - frequently used, no longer in cache 202 * ARC_l2c_only - exists in L2ARC but not other states 203 * When there are no active references to the buffer, they are 204 * are linked onto a list in one of these arc states. These are 205 * the only buffers that can be evicted or deleted. Within each 206 * state there are multiple lists, one for meta-data and one for 207 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 208 * etc.) is tracked separately so that it can be managed more 209 * explicitly: favored over data, limited explicitly. 210 * 211 * Anonymous buffers are buffers that are not associated with 212 * a DVA. These are buffers that hold dirty block copies 213 * before they are written to stable storage. By definition, 214 * they are "ref'd" and are considered part of arc_mru 215 * that cannot be freed. Generally, they will aquire a DVA 216 * as they are written and migrate onto the arc_mru list. 217 * 218 * The ARC_l2c_only state is for buffers that are in the second 219 * level ARC but no longer in any of the ARC_m* lists. The second 220 * level ARC itself may also contain buffers that are in any of 221 * the ARC_m* states - meaning that a buffer can exist in two 222 * places. The reason for the ARC_l2c_only state is to keep the 223 * buffer header in the hash table, so that reads that hit the 224 * second level ARC benefit from these fast lookups. 225 */ 226 227 typedef struct arc_state { 228 list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */ 229 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 230 uint64_t arcs_size; /* total amount of data in this state */ 231 kmutex_t arcs_mtx; 232 } arc_state_t; 233 234 /* The 6 states: */ 235 static arc_state_t ARC_anon; 236 static arc_state_t ARC_mru; 237 static arc_state_t ARC_mru_ghost; 238 static arc_state_t ARC_mfu; 239 static arc_state_t ARC_mfu_ghost; 240 static arc_state_t ARC_l2c_only; 241 242 typedef struct arc_stats { 243 kstat_named_t arcstat_hits; 244 kstat_named_t arcstat_misses; 245 kstat_named_t arcstat_demand_data_hits; 246 kstat_named_t arcstat_demand_data_misses; 247 kstat_named_t arcstat_demand_metadata_hits; 248 kstat_named_t arcstat_demand_metadata_misses; 249 kstat_named_t arcstat_prefetch_data_hits; 250 kstat_named_t arcstat_prefetch_data_misses; 251 kstat_named_t arcstat_prefetch_metadata_hits; 252 kstat_named_t arcstat_prefetch_metadata_misses; 253 kstat_named_t arcstat_mru_hits; 254 kstat_named_t arcstat_mru_ghost_hits; 255 kstat_named_t arcstat_mfu_hits; 256 kstat_named_t arcstat_mfu_ghost_hits; 257 kstat_named_t arcstat_deleted; 258 kstat_named_t arcstat_recycle_miss; 259 kstat_named_t arcstat_mutex_miss; 260 kstat_named_t arcstat_evict_skip; 261 kstat_named_t arcstat_evict_l2_cached; 262 kstat_named_t arcstat_evict_l2_eligible; 263 kstat_named_t arcstat_evict_l2_ineligible; 264 kstat_named_t arcstat_hash_elements; 265 kstat_named_t arcstat_hash_elements_max; 266 kstat_named_t arcstat_hash_collisions; 267 kstat_named_t arcstat_hash_chains; 268 kstat_named_t arcstat_hash_chain_max; 269 kstat_named_t arcstat_p; 270 kstat_named_t arcstat_c; 271 kstat_named_t arcstat_c_min; 272 kstat_named_t arcstat_c_max; 273 kstat_named_t arcstat_size; 274 kstat_named_t arcstat_hdr_size; 275 kstat_named_t arcstat_data_size; 276 kstat_named_t arcstat_other_size; 277 kstat_named_t arcstat_l2_hits; 278 kstat_named_t arcstat_l2_misses; 279 kstat_named_t arcstat_l2_feeds; 280 kstat_named_t arcstat_l2_rw_clash; 281 kstat_named_t arcstat_l2_read_bytes; 282 kstat_named_t arcstat_l2_write_bytes; 283 kstat_named_t arcstat_l2_writes_sent; 284 kstat_named_t arcstat_l2_writes_done; 285 kstat_named_t arcstat_l2_writes_error; 286 kstat_named_t arcstat_l2_writes_hdr_miss; 287 kstat_named_t arcstat_l2_evict_lock_retry; 288 kstat_named_t arcstat_l2_evict_reading; 289 kstat_named_t arcstat_l2_free_on_write; 290 kstat_named_t arcstat_l2_abort_lowmem; 291 kstat_named_t arcstat_l2_cksum_bad; 292 kstat_named_t arcstat_l2_io_error; 293 kstat_named_t arcstat_l2_size; 294 kstat_named_t arcstat_l2_asize; 295 kstat_named_t arcstat_l2_hdr_size; 296 kstat_named_t arcstat_l2_compress_successes; 297 kstat_named_t arcstat_l2_compress_zeros; 298 kstat_named_t arcstat_l2_compress_failures; 299 kstat_named_t arcstat_memory_throttle_count; 300 kstat_named_t arcstat_duplicate_buffers; 301 kstat_named_t arcstat_duplicate_buffers_size; 302 kstat_named_t arcstat_duplicate_reads; 303 kstat_named_t arcstat_meta_used; 304 kstat_named_t arcstat_meta_limit; 305 kstat_named_t arcstat_meta_max; 306 } arc_stats_t; 307 308 static arc_stats_t arc_stats = { 309 { "hits", KSTAT_DATA_UINT64 }, 310 { "misses", KSTAT_DATA_UINT64 }, 311 { "demand_data_hits", KSTAT_DATA_UINT64 }, 312 { "demand_data_misses", KSTAT_DATA_UINT64 }, 313 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 314 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 315 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 316 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 317 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 318 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 319 { "mru_hits", KSTAT_DATA_UINT64 }, 320 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 321 { "mfu_hits", KSTAT_DATA_UINT64 }, 322 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 323 { "deleted", KSTAT_DATA_UINT64 }, 324 { "recycle_miss", KSTAT_DATA_UINT64 }, 325 { "mutex_miss", KSTAT_DATA_UINT64 }, 326 { "evict_skip", KSTAT_DATA_UINT64 }, 327 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 328 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 329 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 330 { "hash_elements", KSTAT_DATA_UINT64 }, 331 { "hash_elements_max", KSTAT_DATA_UINT64 }, 332 { "hash_collisions", KSTAT_DATA_UINT64 }, 333 { "hash_chains", KSTAT_DATA_UINT64 }, 334 { "hash_chain_max", KSTAT_DATA_UINT64 }, 335 { "p", KSTAT_DATA_UINT64 }, 336 { "c", KSTAT_DATA_UINT64 }, 337 { "c_min", KSTAT_DATA_UINT64 }, 338 { "c_max", KSTAT_DATA_UINT64 }, 339 { "size", KSTAT_DATA_UINT64 }, 340 { "hdr_size", KSTAT_DATA_UINT64 }, 341 { "data_size", KSTAT_DATA_UINT64 }, 342 { "other_size", KSTAT_DATA_UINT64 }, 343 { "l2_hits", KSTAT_DATA_UINT64 }, 344 { "l2_misses", KSTAT_DATA_UINT64 }, 345 { "l2_feeds", KSTAT_DATA_UINT64 }, 346 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 347 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 348 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 349 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 350 { "l2_writes_done", KSTAT_DATA_UINT64 }, 351 { "l2_writes_error", KSTAT_DATA_UINT64 }, 352 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 353 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 354 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 355 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 356 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 357 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 358 { "l2_io_error", KSTAT_DATA_UINT64 }, 359 { "l2_size", KSTAT_DATA_UINT64 }, 360 { "l2_asize", KSTAT_DATA_UINT64 }, 361 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 362 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 363 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 364 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 365 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 366 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 367 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 368 { "duplicate_reads", KSTAT_DATA_UINT64 }, 369 { "arc_meta_used", KSTAT_DATA_UINT64 }, 370 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 371 { "arc_meta_max", KSTAT_DATA_UINT64 } 372 }; 373 374 #define ARCSTAT(stat) (arc_stats.stat.value.ui64) 375 376 #define ARCSTAT_INCR(stat, val) \ 377 atomic_add_64(&arc_stats.stat.value.ui64, (val)); 378 379 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 380 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 381 382 #define ARCSTAT_MAX(stat, val) { \ 383 uint64_t m; \ 384 while ((val) > (m = arc_stats.stat.value.ui64) && \ 385 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 386 continue; \ 387 } 388 389 #define ARCSTAT_MAXSTAT(stat) \ 390 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 391 392 /* 393 * We define a macro to allow ARC hits/misses to be easily broken down by 394 * two separate conditions, giving a total of four different subtypes for 395 * each of hits and misses (so eight statistics total). 396 */ 397 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 398 if (cond1) { \ 399 if (cond2) { \ 400 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 401 } else { \ 402 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 403 } \ 404 } else { \ 405 if (cond2) { \ 406 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 407 } else { \ 408 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 409 } \ 410 } 411 412 kstat_t *arc_ksp; 413 static arc_state_t *arc_anon; 414 static arc_state_t *arc_mru; 415 static arc_state_t *arc_mru_ghost; 416 static arc_state_t *arc_mfu; 417 static arc_state_t *arc_mfu_ghost; 418 static arc_state_t *arc_l2c_only; 419 420 /* 421 * There are several ARC variables that are critical to export as kstats -- 422 * but we don't want to have to grovel around in the kstat whenever we wish to 423 * manipulate them. For these variables, we therefore define them to be in 424 * terms of the statistic variable. This assures that we are not introducing 425 * the possibility of inconsistency by having shadow copies of the variables, 426 * while still allowing the code to be readable. 427 */ 428 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 429 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 430 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 431 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 432 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 433 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 434 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 435 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 436 437 #define L2ARC_IS_VALID_COMPRESS(_c_) \ 438 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 439 440 static int arc_no_grow; /* Don't try to grow cache size */ 441 static uint64_t arc_tempreserve; 442 static uint64_t arc_loaned_bytes; 443 444 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 445 446 typedef struct arc_callback arc_callback_t; 447 448 struct arc_callback { 449 void *acb_private; 450 arc_done_func_t *acb_done; 451 arc_buf_t *acb_buf; 452 zio_t *acb_zio_dummy; 453 arc_callback_t *acb_next; 454 }; 455 456 typedef struct arc_write_callback arc_write_callback_t; 457 458 struct arc_write_callback { 459 void *awcb_private; 460 arc_done_func_t *awcb_ready; 461 arc_done_func_t *awcb_done; 462 arc_buf_t *awcb_buf; 463 }; 464 465 struct arc_buf_hdr { 466 /* protected by hash lock */ 467 dva_t b_dva; 468 uint64_t b_birth; 469 uint64_t b_cksum0; 470 471 kmutex_t b_freeze_lock; 472 zio_cksum_t *b_freeze_cksum; 473 void *b_thawed; 474 475 arc_buf_hdr_t *b_hash_next; 476 arc_buf_t *b_buf; 477 uint32_t b_flags; 478 uint32_t b_datacnt; 479 480 arc_callback_t *b_acb; 481 kcondvar_t b_cv; 482 483 /* immutable */ 484 arc_buf_contents_t b_type; 485 uint64_t b_size; 486 uint64_t b_spa; 487 488 /* protected by arc state mutex */ 489 arc_state_t *b_state; 490 list_node_t b_arc_node; 491 492 /* updated atomically */ 493 clock_t b_arc_access; 494 495 /* self protecting */ 496 refcount_t b_refcnt; 497 498 l2arc_buf_hdr_t *b_l2hdr; 499 list_node_t b_l2node; 500 }; 501 502 static arc_buf_t *arc_eviction_list; 503 static kmutex_t arc_eviction_mtx; 504 static arc_buf_hdr_t arc_eviction_hdr; 505 static void arc_get_data_buf(arc_buf_t *buf); 506 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 507 static int arc_evict_needed(arc_buf_contents_t type); 508 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes); 509 static void arc_buf_watch(arc_buf_t *buf); 510 511 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); 512 513 #define GHOST_STATE(state) \ 514 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 515 (state) == arc_l2c_only) 516 517 /* 518 * Private ARC flags. These flags are private ARC only flags that will show up 519 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 520 * be passed in as arc_flags in things like arc_read. However, these flags 521 * should never be passed and should only be set by ARC code. When adding new 522 * public flags, make sure not to smash the private ones. 523 */ 524 525 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 526 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 527 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 528 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 529 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 530 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 531 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 532 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 533 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 534 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 535 536 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 537 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 538 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 539 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 540 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 541 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 542 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 543 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 544 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 545 (hdr)->b_l2hdr != NULL) 546 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 547 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 548 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 549 550 /* 551 * Other sizes 552 */ 553 554 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 555 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 556 557 /* 558 * Hash table routines 559 */ 560 561 #define HT_LOCK_PAD 64 562 563 struct ht_lock { 564 kmutex_t ht_lock; 565 #ifdef _KERNEL 566 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 567 #endif 568 }; 569 570 #define BUF_LOCKS 256 571 typedef struct buf_hash_table { 572 uint64_t ht_mask; 573 arc_buf_hdr_t **ht_table; 574 struct ht_lock ht_locks[BUF_LOCKS]; 575 } buf_hash_table_t; 576 577 static buf_hash_table_t buf_hash_table; 578 579 #define BUF_HASH_INDEX(spa, dva, birth) \ 580 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 581 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 582 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 583 #define HDR_LOCK(hdr) \ 584 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 585 586 uint64_t zfs_crc64_table[256]; 587 588 /* 589 * Level 2 ARC 590 */ 591 592 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 593 #define L2ARC_HEADROOM 2 /* num of writes */ 594 /* 595 * If we discover during ARC scan any buffers to be compressed, we boost 596 * our headroom for the next scanning cycle by this percentage multiple. 597 */ 598 #define L2ARC_HEADROOM_BOOST 200 599 #define L2ARC_FEED_SECS 1 /* caching interval secs */ 600 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 601 602 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 603 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 604 605 /* 606 * L2ARC Performance Tunables 607 */ 608 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 609 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 610 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 611 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 612 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 613 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 614 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 615 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 616 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 617 618 /* 619 * L2ARC Internals 620 */ 621 typedef struct l2arc_dev { 622 vdev_t *l2ad_vdev; /* vdev */ 623 spa_t *l2ad_spa; /* spa */ 624 uint64_t l2ad_hand; /* next write location */ 625 uint64_t l2ad_start; /* first addr on device */ 626 uint64_t l2ad_end; /* last addr on device */ 627 uint64_t l2ad_evict; /* last addr eviction reached */ 628 boolean_t l2ad_first; /* first sweep through */ 629 boolean_t l2ad_writing; /* currently writing */ 630 list_t *l2ad_buflist; /* buffer list */ 631 list_node_t l2ad_node; /* device list node */ 632 } l2arc_dev_t; 633 634 static list_t L2ARC_dev_list; /* device list */ 635 static list_t *l2arc_dev_list; /* device list pointer */ 636 static kmutex_t l2arc_dev_mtx; /* device list mutex */ 637 static l2arc_dev_t *l2arc_dev_last; /* last device used */ 638 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 639 static list_t L2ARC_free_on_write; /* free after write buf list */ 640 static list_t *l2arc_free_on_write; /* free after write list ptr */ 641 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 642 static uint64_t l2arc_ndev; /* number of devices */ 643 644 typedef struct l2arc_read_callback { 645 arc_buf_t *l2rcb_buf; /* read buffer */ 646 spa_t *l2rcb_spa; /* spa */ 647 blkptr_t l2rcb_bp; /* original blkptr */ 648 zbookmark_t l2rcb_zb; /* original bookmark */ 649 int l2rcb_flags; /* original flags */ 650 enum zio_compress l2rcb_compress; /* applied compress */ 651 } l2arc_read_callback_t; 652 653 typedef struct l2arc_write_callback { 654 l2arc_dev_t *l2wcb_dev; /* device info */ 655 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 656 } l2arc_write_callback_t; 657 658 struct l2arc_buf_hdr { 659 /* protected by arc_buf_hdr mutex */ 660 l2arc_dev_t *b_dev; /* L2ARC device */ 661 uint64_t b_daddr; /* disk address, offset byte */ 662 /* compression applied to buffer data */ 663 enum zio_compress b_compress; 664 /* real alloc'd buffer size depending on b_compress applied */ 665 int b_asize; 666 /* temporary buffer holder for in-flight compressed data */ 667 void *b_tmp_cdata; 668 }; 669 670 typedef struct l2arc_data_free { 671 /* protected by l2arc_free_on_write_mtx */ 672 void *l2df_data; 673 size_t l2df_size; 674 void (*l2df_func)(void *, size_t); 675 list_node_t l2df_list_node; 676 } l2arc_data_free_t; 677 678 static kmutex_t l2arc_feed_thr_lock; 679 static kcondvar_t l2arc_feed_thr_cv; 680 static uint8_t l2arc_thread_exit; 681 682 static void l2arc_read_done(zio_t *zio); 683 static void l2arc_hdr_stat_add(void); 684 static void l2arc_hdr_stat_remove(void); 685 686 static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr); 687 static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, 688 enum zio_compress c); 689 static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab); 690 691 static uint64_t 692 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 693 { 694 uint8_t *vdva = (uint8_t *)dva; 695 uint64_t crc = -1ULL; 696 int i; 697 698 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 699 700 for (i = 0; i < sizeof (dva_t); i++) 701 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 702 703 crc ^= (spa>>8) ^ birth; 704 705 return (crc); 706 } 707 708 #define BUF_EMPTY(buf) \ 709 ((buf)->b_dva.dva_word[0] == 0 && \ 710 (buf)->b_dva.dva_word[1] == 0 && \ 711 (buf)->b_birth == 0) 712 713 #define BUF_EQUAL(spa, dva, birth, buf) \ 714 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 715 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 716 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 717 718 static void 719 buf_discard_identity(arc_buf_hdr_t *hdr) 720 { 721 hdr->b_dva.dva_word[0] = 0; 722 hdr->b_dva.dva_word[1] = 0; 723 hdr->b_birth = 0; 724 hdr->b_cksum0 = 0; 725 } 726 727 static arc_buf_hdr_t * 728 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp) 729 { 730 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 731 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 732 arc_buf_hdr_t *buf; 733 734 mutex_enter(hash_lock); 735 for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 736 buf = buf->b_hash_next) { 737 if (BUF_EQUAL(spa, dva, birth, buf)) { 738 *lockp = hash_lock; 739 return (buf); 740 } 741 } 742 mutex_exit(hash_lock); 743 *lockp = NULL; 744 return (NULL); 745 } 746 747 /* 748 * Insert an entry into the hash table. If there is already an element 749 * equal to elem in the hash table, then the already existing element 750 * will be returned and the new element will not be inserted. 751 * Otherwise returns NULL. 752 */ 753 static arc_buf_hdr_t * 754 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 755 { 756 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 757 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 758 arc_buf_hdr_t *fbuf; 759 uint32_t i; 760 761 ASSERT(!HDR_IN_HASH_TABLE(buf)); 762 *lockp = hash_lock; 763 mutex_enter(hash_lock); 764 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 765 fbuf = fbuf->b_hash_next, i++) { 766 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 767 return (fbuf); 768 } 769 770 buf->b_hash_next = buf_hash_table.ht_table[idx]; 771 buf_hash_table.ht_table[idx] = buf; 772 buf->b_flags |= ARC_IN_HASH_TABLE; 773 774 /* collect some hash table performance data */ 775 if (i > 0) { 776 ARCSTAT_BUMP(arcstat_hash_collisions); 777 if (i == 1) 778 ARCSTAT_BUMP(arcstat_hash_chains); 779 780 ARCSTAT_MAX(arcstat_hash_chain_max, i); 781 } 782 783 ARCSTAT_BUMP(arcstat_hash_elements); 784 ARCSTAT_MAXSTAT(arcstat_hash_elements); 785 786 return (NULL); 787 } 788 789 static void 790 buf_hash_remove(arc_buf_hdr_t *buf) 791 { 792 arc_buf_hdr_t *fbuf, **bufp; 793 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 794 795 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 796 ASSERT(HDR_IN_HASH_TABLE(buf)); 797 798 bufp = &buf_hash_table.ht_table[idx]; 799 while ((fbuf = *bufp) != buf) { 800 ASSERT(fbuf != NULL); 801 bufp = &fbuf->b_hash_next; 802 } 803 *bufp = buf->b_hash_next; 804 buf->b_hash_next = NULL; 805 buf->b_flags &= ~ARC_IN_HASH_TABLE; 806 807 /* collect some hash table performance data */ 808 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 809 810 if (buf_hash_table.ht_table[idx] && 811 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 812 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 813 } 814 815 /* 816 * Global data structures and functions for the buf kmem cache. 817 */ 818 static kmem_cache_t *hdr_cache; 819 static kmem_cache_t *buf_cache; 820 821 static void 822 buf_fini(void) 823 { 824 int i; 825 826 kmem_free(buf_hash_table.ht_table, 827 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 828 for (i = 0; i < BUF_LOCKS; i++) 829 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 830 kmem_cache_destroy(hdr_cache); 831 kmem_cache_destroy(buf_cache); 832 } 833 834 /* 835 * Constructor callback - called when the cache is empty 836 * and a new buf is requested. 837 */ 838 /* ARGSUSED */ 839 static int 840 hdr_cons(void *vbuf, void *unused, int kmflag) 841 { 842 arc_buf_hdr_t *buf = vbuf; 843 844 bzero(buf, sizeof (arc_buf_hdr_t)); 845 refcount_create(&buf->b_refcnt); 846 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 847 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 848 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 849 850 return (0); 851 } 852 853 /* ARGSUSED */ 854 static int 855 buf_cons(void *vbuf, void *unused, int kmflag) 856 { 857 arc_buf_t *buf = vbuf; 858 859 bzero(buf, sizeof (arc_buf_t)); 860 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 861 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 862 863 return (0); 864 } 865 866 /* 867 * Destructor callback - called when a cached buf is 868 * no longer required. 869 */ 870 /* ARGSUSED */ 871 static void 872 hdr_dest(void *vbuf, void *unused) 873 { 874 arc_buf_hdr_t *buf = vbuf; 875 876 ASSERT(BUF_EMPTY(buf)); 877 refcount_destroy(&buf->b_refcnt); 878 cv_destroy(&buf->b_cv); 879 mutex_destroy(&buf->b_freeze_lock); 880 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 881 } 882 883 /* ARGSUSED */ 884 static void 885 buf_dest(void *vbuf, void *unused) 886 { 887 arc_buf_t *buf = vbuf; 888 889 mutex_destroy(&buf->b_evict_lock); 890 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 891 } 892 893 /* 894 * Reclaim callback -- invoked when memory is low. 895 */ 896 /* ARGSUSED */ 897 static void 898 hdr_recl(void *unused) 899 { 900 dprintf("hdr_recl called\n"); 901 /* 902 * umem calls the reclaim func when we destroy the buf cache, 903 * which is after we do arc_fini(). 904 */ 905 if (!arc_dead) 906 cv_signal(&arc_reclaim_thr_cv); 907 } 908 909 static void 910 buf_init(void) 911 { 912 uint64_t *ct; 913 uint64_t hsize = 1ULL << 12; 914 int i, j; 915 916 /* 917 * The hash table is big enough to fill all of physical memory 918 * with an average 64K block size. The table will take up 919 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers). 920 */ 921 while (hsize * 65536 < physmem * PAGESIZE) 922 hsize <<= 1; 923 retry: 924 buf_hash_table.ht_mask = hsize - 1; 925 buf_hash_table.ht_table = 926 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 927 if (buf_hash_table.ht_table == NULL) { 928 ASSERT(hsize > (1ULL << 8)); 929 hsize >>= 1; 930 goto retry; 931 } 932 933 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 934 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 935 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 936 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 937 938 for (i = 0; i < 256; i++) 939 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 940 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 941 942 for (i = 0; i < BUF_LOCKS; i++) { 943 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 944 NULL, MUTEX_DEFAULT, NULL); 945 } 946 } 947 948 #define ARC_MINTIME (hz>>4) /* 62 ms */ 949 950 static void 951 arc_cksum_verify(arc_buf_t *buf) 952 { 953 zio_cksum_t zc; 954 955 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 956 return; 957 958 mutex_enter(&buf->b_hdr->b_freeze_lock); 959 if (buf->b_hdr->b_freeze_cksum == NULL || 960 (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 961 mutex_exit(&buf->b_hdr->b_freeze_lock); 962 return; 963 } 964 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 965 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 966 panic("buffer modified while frozen!"); 967 mutex_exit(&buf->b_hdr->b_freeze_lock); 968 } 969 970 static int 971 arc_cksum_equal(arc_buf_t *buf) 972 { 973 zio_cksum_t zc; 974 int equal; 975 976 mutex_enter(&buf->b_hdr->b_freeze_lock); 977 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 978 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 979 mutex_exit(&buf->b_hdr->b_freeze_lock); 980 981 return (equal); 982 } 983 984 static void 985 arc_cksum_compute(arc_buf_t *buf, boolean_t force) 986 { 987 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 988 return; 989 990 mutex_enter(&buf->b_hdr->b_freeze_lock); 991 if (buf->b_hdr->b_freeze_cksum != NULL) { 992 mutex_exit(&buf->b_hdr->b_freeze_lock); 993 return; 994 } 995 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 996 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 997 buf->b_hdr->b_freeze_cksum); 998 mutex_exit(&buf->b_hdr->b_freeze_lock); 999 arc_buf_watch(buf); 1000 } 1001 1002 #ifndef _KERNEL 1003 typedef struct procctl { 1004 long cmd; 1005 prwatch_t prwatch; 1006 } procctl_t; 1007 #endif 1008 1009 /* ARGSUSED */ 1010 static void 1011 arc_buf_unwatch(arc_buf_t *buf) 1012 { 1013 #ifndef _KERNEL 1014 if (arc_watch) { 1015 int result; 1016 procctl_t ctl; 1017 ctl.cmd = PCWATCH; 1018 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1019 ctl.prwatch.pr_size = 0; 1020 ctl.prwatch.pr_wflags = 0; 1021 result = write(arc_procfd, &ctl, sizeof (ctl)); 1022 ASSERT3U(result, ==, sizeof (ctl)); 1023 } 1024 #endif 1025 } 1026 1027 /* ARGSUSED */ 1028 static void 1029 arc_buf_watch(arc_buf_t *buf) 1030 { 1031 #ifndef _KERNEL 1032 if (arc_watch) { 1033 int result; 1034 procctl_t ctl; 1035 ctl.cmd = PCWATCH; 1036 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1037 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1038 ctl.prwatch.pr_wflags = WA_WRITE; 1039 result = write(arc_procfd, &ctl, sizeof (ctl)); 1040 ASSERT3U(result, ==, sizeof (ctl)); 1041 } 1042 #endif 1043 } 1044 1045 void 1046 arc_buf_thaw(arc_buf_t *buf) 1047 { 1048 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1049 if (buf->b_hdr->b_state != arc_anon) 1050 panic("modifying non-anon buffer!"); 1051 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1052 panic("modifying buffer while i/o in progress!"); 1053 arc_cksum_verify(buf); 1054 } 1055 1056 mutex_enter(&buf->b_hdr->b_freeze_lock); 1057 if (buf->b_hdr->b_freeze_cksum != NULL) { 1058 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1059 buf->b_hdr->b_freeze_cksum = NULL; 1060 } 1061 1062 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1063 if (buf->b_hdr->b_thawed) 1064 kmem_free(buf->b_hdr->b_thawed, 1); 1065 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1066 } 1067 1068 mutex_exit(&buf->b_hdr->b_freeze_lock); 1069 1070 arc_buf_unwatch(buf); 1071 } 1072 1073 void 1074 arc_buf_freeze(arc_buf_t *buf) 1075 { 1076 kmutex_t *hash_lock; 1077 1078 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1079 return; 1080 1081 hash_lock = HDR_LOCK(buf->b_hdr); 1082 mutex_enter(hash_lock); 1083 1084 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1085 buf->b_hdr->b_state == arc_anon); 1086 arc_cksum_compute(buf, B_FALSE); 1087 mutex_exit(hash_lock); 1088 1089 } 1090 1091 static void 1092 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1093 { 1094 ASSERT(MUTEX_HELD(hash_lock)); 1095 1096 if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1097 (ab->b_state != arc_anon)) { 1098 uint64_t delta = ab->b_size * ab->b_datacnt; 1099 list_t *list = &ab->b_state->arcs_list[ab->b_type]; 1100 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1101 1102 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx)); 1103 mutex_enter(&ab->b_state->arcs_mtx); 1104 ASSERT(list_link_active(&ab->b_arc_node)); 1105 list_remove(list, ab); 1106 if (GHOST_STATE(ab->b_state)) { 1107 ASSERT0(ab->b_datacnt); 1108 ASSERT3P(ab->b_buf, ==, NULL); 1109 delta = ab->b_size; 1110 } 1111 ASSERT(delta > 0); 1112 ASSERT3U(*size, >=, delta); 1113 atomic_add_64(size, -delta); 1114 mutex_exit(&ab->b_state->arcs_mtx); 1115 /* remove the prefetch flag if we get a reference */ 1116 if (ab->b_flags & ARC_PREFETCH) 1117 ab->b_flags &= ~ARC_PREFETCH; 1118 } 1119 } 1120 1121 static int 1122 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1123 { 1124 int cnt; 1125 arc_state_t *state = ab->b_state; 1126 1127 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1128 ASSERT(!GHOST_STATE(state)); 1129 1130 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1131 (state != arc_anon)) { 1132 uint64_t *size = &state->arcs_lsize[ab->b_type]; 1133 1134 ASSERT(!MUTEX_HELD(&state->arcs_mtx)); 1135 mutex_enter(&state->arcs_mtx); 1136 ASSERT(!list_link_active(&ab->b_arc_node)); 1137 list_insert_head(&state->arcs_list[ab->b_type], ab); 1138 ASSERT(ab->b_datacnt > 0); 1139 atomic_add_64(size, ab->b_size * ab->b_datacnt); 1140 mutex_exit(&state->arcs_mtx); 1141 } 1142 return (cnt); 1143 } 1144 1145 /* 1146 * Move the supplied buffer to the indicated state. The mutex 1147 * for the buffer must be held by the caller. 1148 */ 1149 static void 1150 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1151 { 1152 arc_state_t *old_state = ab->b_state; 1153 int64_t refcnt = refcount_count(&ab->b_refcnt); 1154 uint64_t from_delta, to_delta; 1155 1156 ASSERT(MUTEX_HELD(hash_lock)); 1157 ASSERT(new_state != old_state); 1158 ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1159 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1160 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); 1161 1162 from_delta = to_delta = ab->b_datacnt * ab->b_size; 1163 1164 /* 1165 * If this buffer is evictable, transfer it from the 1166 * old state list to the new state list. 1167 */ 1168 if (refcnt == 0) { 1169 if (old_state != arc_anon) { 1170 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx); 1171 uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1172 1173 if (use_mutex) 1174 mutex_enter(&old_state->arcs_mtx); 1175 1176 ASSERT(list_link_active(&ab->b_arc_node)); 1177 list_remove(&old_state->arcs_list[ab->b_type], ab); 1178 1179 /* 1180 * If prefetching out of the ghost cache, 1181 * we will have a non-zero datacnt. 1182 */ 1183 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1184 /* ghost elements have a ghost size */ 1185 ASSERT(ab->b_buf == NULL); 1186 from_delta = ab->b_size; 1187 } 1188 ASSERT3U(*size, >=, from_delta); 1189 atomic_add_64(size, -from_delta); 1190 1191 if (use_mutex) 1192 mutex_exit(&old_state->arcs_mtx); 1193 } 1194 if (new_state != arc_anon) { 1195 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx); 1196 uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1197 1198 if (use_mutex) 1199 mutex_enter(&new_state->arcs_mtx); 1200 1201 list_insert_head(&new_state->arcs_list[ab->b_type], ab); 1202 1203 /* ghost elements have a ghost size */ 1204 if (GHOST_STATE(new_state)) { 1205 ASSERT(ab->b_datacnt == 0); 1206 ASSERT(ab->b_buf == NULL); 1207 to_delta = ab->b_size; 1208 } 1209 atomic_add_64(size, to_delta); 1210 1211 if (use_mutex) 1212 mutex_exit(&new_state->arcs_mtx); 1213 } 1214 } 1215 1216 ASSERT(!BUF_EMPTY(ab)); 1217 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) 1218 buf_hash_remove(ab); 1219 1220 /* adjust state sizes */ 1221 if (to_delta) 1222 atomic_add_64(&new_state->arcs_size, to_delta); 1223 if (from_delta) { 1224 ASSERT3U(old_state->arcs_size, >=, from_delta); 1225 atomic_add_64(&old_state->arcs_size, -from_delta); 1226 } 1227 ab->b_state = new_state; 1228 1229 /* adjust l2arc hdr stats */ 1230 if (new_state == arc_l2c_only) 1231 l2arc_hdr_stat_add(); 1232 else if (old_state == arc_l2c_only) 1233 l2arc_hdr_stat_remove(); 1234 } 1235 1236 void 1237 arc_space_consume(uint64_t space, arc_space_type_t type) 1238 { 1239 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1240 1241 switch (type) { 1242 case ARC_SPACE_DATA: 1243 ARCSTAT_INCR(arcstat_data_size, space); 1244 break; 1245 case ARC_SPACE_OTHER: 1246 ARCSTAT_INCR(arcstat_other_size, space); 1247 break; 1248 case ARC_SPACE_HDRS: 1249 ARCSTAT_INCR(arcstat_hdr_size, space); 1250 break; 1251 case ARC_SPACE_L2HDRS: 1252 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1253 break; 1254 } 1255 1256 ARCSTAT_INCR(arcstat_meta_used, space); 1257 atomic_add_64(&arc_size, space); 1258 } 1259 1260 void 1261 arc_space_return(uint64_t space, arc_space_type_t type) 1262 { 1263 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1264 1265 switch (type) { 1266 case ARC_SPACE_DATA: 1267 ARCSTAT_INCR(arcstat_data_size, -space); 1268 break; 1269 case ARC_SPACE_OTHER: 1270 ARCSTAT_INCR(arcstat_other_size, -space); 1271 break; 1272 case ARC_SPACE_HDRS: 1273 ARCSTAT_INCR(arcstat_hdr_size, -space); 1274 break; 1275 case ARC_SPACE_L2HDRS: 1276 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1277 break; 1278 } 1279 1280 ASSERT(arc_meta_used >= space); 1281 if (arc_meta_max < arc_meta_used) 1282 arc_meta_max = arc_meta_used; 1283 ARCSTAT_INCR(arcstat_meta_used, -space); 1284 ASSERT(arc_size >= space); 1285 atomic_add_64(&arc_size, -space); 1286 } 1287 1288 void * 1289 arc_data_buf_alloc(uint64_t size) 1290 { 1291 if (arc_evict_needed(ARC_BUFC_DATA)) 1292 cv_signal(&arc_reclaim_thr_cv); 1293 atomic_add_64(&arc_size, size); 1294 return (zio_data_buf_alloc(size)); 1295 } 1296 1297 void 1298 arc_data_buf_free(void *buf, uint64_t size) 1299 { 1300 zio_data_buf_free(buf, size); 1301 ASSERT(arc_size >= size); 1302 atomic_add_64(&arc_size, -size); 1303 } 1304 1305 arc_buf_t * 1306 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1307 { 1308 arc_buf_hdr_t *hdr; 1309 arc_buf_t *buf; 1310 1311 ASSERT3U(size, >, 0); 1312 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1313 ASSERT(BUF_EMPTY(hdr)); 1314 hdr->b_size = size; 1315 hdr->b_type = type; 1316 hdr->b_spa = spa_load_guid(spa); 1317 hdr->b_state = arc_anon; 1318 hdr->b_arc_access = 0; 1319 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1320 buf->b_hdr = hdr; 1321 buf->b_data = NULL; 1322 buf->b_efunc = NULL; 1323 buf->b_private = NULL; 1324 buf->b_next = NULL; 1325 hdr->b_buf = buf; 1326 arc_get_data_buf(buf); 1327 hdr->b_datacnt = 1; 1328 hdr->b_flags = 0; 1329 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1330 (void) refcount_add(&hdr->b_refcnt, tag); 1331 1332 return (buf); 1333 } 1334 1335 static char *arc_onloan_tag = "onloan"; 1336 1337 /* 1338 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1339 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1340 * buffers must be returned to the arc before they can be used by the DMU or 1341 * freed. 1342 */ 1343 arc_buf_t * 1344 arc_loan_buf(spa_t *spa, int size) 1345 { 1346 arc_buf_t *buf; 1347 1348 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1349 1350 atomic_add_64(&arc_loaned_bytes, size); 1351 return (buf); 1352 } 1353 1354 /* 1355 * Return a loaned arc buffer to the arc. 1356 */ 1357 void 1358 arc_return_buf(arc_buf_t *buf, void *tag) 1359 { 1360 arc_buf_hdr_t *hdr = buf->b_hdr; 1361 1362 ASSERT(buf->b_data != NULL); 1363 (void) refcount_add(&hdr->b_refcnt, tag); 1364 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1365 1366 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1367 } 1368 1369 /* Detach an arc_buf from a dbuf (tag) */ 1370 void 1371 arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1372 { 1373 arc_buf_hdr_t *hdr; 1374 1375 ASSERT(buf->b_data != NULL); 1376 hdr = buf->b_hdr; 1377 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1378 (void) refcount_remove(&hdr->b_refcnt, tag); 1379 buf->b_efunc = NULL; 1380 buf->b_private = NULL; 1381 1382 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1383 } 1384 1385 static arc_buf_t * 1386 arc_buf_clone(arc_buf_t *from) 1387 { 1388 arc_buf_t *buf; 1389 arc_buf_hdr_t *hdr = from->b_hdr; 1390 uint64_t size = hdr->b_size; 1391 1392 ASSERT(hdr->b_state != arc_anon); 1393 1394 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1395 buf->b_hdr = hdr; 1396 buf->b_data = NULL; 1397 buf->b_efunc = NULL; 1398 buf->b_private = NULL; 1399 buf->b_next = hdr->b_buf; 1400 hdr->b_buf = buf; 1401 arc_get_data_buf(buf); 1402 bcopy(from->b_data, buf->b_data, size); 1403 1404 /* 1405 * This buffer already exists in the arc so create a duplicate 1406 * copy for the caller. If the buffer is associated with user data 1407 * then track the size and number of duplicates. These stats will be 1408 * updated as duplicate buffers are created and destroyed. 1409 */ 1410 if (hdr->b_type == ARC_BUFC_DATA) { 1411 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1412 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1413 } 1414 hdr->b_datacnt += 1; 1415 return (buf); 1416 } 1417 1418 void 1419 arc_buf_add_ref(arc_buf_t *buf, void* tag) 1420 { 1421 arc_buf_hdr_t *hdr; 1422 kmutex_t *hash_lock; 1423 1424 /* 1425 * Check to see if this buffer is evicted. Callers 1426 * must verify b_data != NULL to know if the add_ref 1427 * was successful. 1428 */ 1429 mutex_enter(&buf->b_evict_lock); 1430 if (buf->b_data == NULL) { 1431 mutex_exit(&buf->b_evict_lock); 1432 return; 1433 } 1434 hash_lock = HDR_LOCK(buf->b_hdr); 1435 mutex_enter(hash_lock); 1436 hdr = buf->b_hdr; 1437 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1438 mutex_exit(&buf->b_evict_lock); 1439 1440 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1441 add_reference(hdr, hash_lock, tag); 1442 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1443 arc_access(hdr, hash_lock); 1444 mutex_exit(hash_lock); 1445 ARCSTAT_BUMP(arcstat_hits); 1446 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1447 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1448 data, metadata, hits); 1449 } 1450 1451 /* 1452 * Free the arc data buffer. If it is an l2arc write in progress, 1453 * the buffer is placed on l2arc_free_on_write to be freed later. 1454 */ 1455 static void 1456 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1457 { 1458 arc_buf_hdr_t *hdr = buf->b_hdr; 1459 1460 if (HDR_L2_WRITING(hdr)) { 1461 l2arc_data_free_t *df; 1462 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1463 df->l2df_data = buf->b_data; 1464 df->l2df_size = hdr->b_size; 1465 df->l2df_func = free_func; 1466 mutex_enter(&l2arc_free_on_write_mtx); 1467 list_insert_head(l2arc_free_on_write, df); 1468 mutex_exit(&l2arc_free_on_write_mtx); 1469 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1470 } else { 1471 free_func(buf->b_data, hdr->b_size); 1472 } 1473 } 1474 1475 static void 1476 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all) 1477 { 1478 arc_buf_t **bufp; 1479 1480 /* free up data associated with the buf */ 1481 if (buf->b_data) { 1482 arc_state_t *state = buf->b_hdr->b_state; 1483 uint64_t size = buf->b_hdr->b_size; 1484 arc_buf_contents_t type = buf->b_hdr->b_type; 1485 1486 arc_cksum_verify(buf); 1487 arc_buf_unwatch(buf); 1488 1489 if (!recycle) { 1490 if (type == ARC_BUFC_METADATA) { 1491 arc_buf_data_free(buf, zio_buf_free); 1492 arc_space_return(size, ARC_SPACE_DATA); 1493 } else { 1494 ASSERT(type == ARC_BUFC_DATA); 1495 arc_buf_data_free(buf, zio_data_buf_free); 1496 ARCSTAT_INCR(arcstat_data_size, -size); 1497 atomic_add_64(&arc_size, -size); 1498 } 1499 } 1500 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1501 uint64_t *cnt = &state->arcs_lsize[type]; 1502 1503 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1504 ASSERT(state != arc_anon); 1505 1506 ASSERT3U(*cnt, >=, size); 1507 atomic_add_64(cnt, -size); 1508 } 1509 ASSERT3U(state->arcs_size, >=, size); 1510 atomic_add_64(&state->arcs_size, -size); 1511 buf->b_data = NULL; 1512 1513 /* 1514 * If we're destroying a duplicate buffer make sure 1515 * that the appropriate statistics are updated. 1516 */ 1517 if (buf->b_hdr->b_datacnt > 1 && 1518 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1519 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1520 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1521 } 1522 ASSERT(buf->b_hdr->b_datacnt > 0); 1523 buf->b_hdr->b_datacnt -= 1; 1524 } 1525 1526 /* only remove the buf if requested */ 1527 if (!all) 1528 return; 1529 1530 /* remove the buf from the hdr list */ 1531 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1532 continue; 1533 *bufp = buf->b_next; 1534 buf->b_next = NULL; 1535 1536 ASSERT(buf->b_efunc == NULL); 1537 1538 /* clean up the buf */ 1539 buf->b_hdr = NULL; 1540 kmem_cache_free(buf_cache, buf); 1541 } 1542 1543 static void 1544 arc_hdr_destroy(arc_buf_hdr_t *hdr) 1545 { 1546 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1547 ASSERT3P(hdr->b_state, ==, arc_anon); 1548 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1549 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1550 1551 if (l2hdr != NULL) { 1552 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1553 /* 1554 * To prevent arc_free() and l2arc_evict() from 1555 * attempting to free the same buffer at the same time, 1556 * a FREE_IN_PROGRESS flag is given to arc_free() to 1557 * give it priority. l2arc_evict() can't destroy this 1558 * header while we are waiting on l2arc_buflist_mtx. 1559 * 1560 * The hdr may be removed from l2ad_buflist before we 1561 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1562 */ 1563 if (!buflist_held) { 1564 mutex_enter(&l2arc_buflist_mtx); 1565 l2hdr = hdr->b_l2hdr; 1566 } 1567 1568 if (l2hdr != NULL) { 1569 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1570 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1571 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1572 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1573 if (hdr->b_state == arc_l2c_only) 1574 l2arc_hdr_stat_remove(); 1575 hdr->b_l2hdr = NULL; 1576 } 1577 1578 if (!buflist_held) 1579 mutex_exit(&l2arc_buflist_mtx); 1580 } 1581 1582 if (!BUF_EMPTY(hdr)) { 1583 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1584 buf_discard_identity(hdr); 1585 } 1586 while (hdr->b_buf) { 1587 arc_buf_t *buf = hdr->b_buf; 1588 1589 if (buf->b_efunc) { 1590 mutex_enter(&arc_eviction_mtx); 1591 mutex_enter(&buf->b_evict_lock); 1592 ASSERT(buf->b_hdr != NULL); 1593 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1594 hdr->b_buf = buf->b_next; 1595 buf->b_hdr = &arc_eviction_hdr; 1596 buf->b_next = arc_eviction_list; 1597 arc_eviction_list = buf; 1598 mutex_exit(&buf->b_evict_lock); 1599 mutex_exit(&arc_eviction_mtx); 1600 } else { 1601 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1602 } 1603 } 1604 if (hdr->b_freeze_cksum != NULL) { 1605 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1606 hdr->b_freeze_cksum = NULL; 1607 } 1608 if (hdr->b_thawed) { 1609 kmem_free(hdr->b_thawed, 1); 1610 hdr->b_thawed = NULL; 1611 } 1612 1613 ASSERT(!list_link_active(&hdr->b_arc_node)); 1614 ASSERT3P(hdr->b_hash_next, ==, NULL); 1615 ASSERT3P(hdr->b_acb, ==, NULL); 1616 kmem_cache_free(hdr_cache, hdr); 1617 } 1618 1619 void 1620 arc_buf_free(arc_buf_t *buf, void *tag) 1621 { 1622 arc_buf_hdr_t *hdr = buf->b_hdr; 1623 int hashed = hdr->b_state != arc_anon; 1624 1625 ASSERT(buf->b_efunc == NULL); 1626 ASSERT(buf->b_data != NULL); 1627 1628 if (hashed) { 1629 kmutex_t *hash_lock = HDR_LOCK(hdr); 1630 1631 mutex_enter(hash_lock); 1632 hdr = buf->b_hdr; 1633 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1634 1635 (void) remove_reference(hdr, hash_lock, tag); 1636 if (hdr->b_datacnt > 1) { 1637 arc_buf_destroy(buf, FALSE, TRUE); 1638 } else { 1639 ASSERT(buf == hdr->b_buf); 1640 ASSERT(buf->b_efunc == NULL); 1641 hdr->b_flags |= ARC_BUF_AVAILABLE; 1642 } 1643 mutex_exit(hash_lock); 1644 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1645 int destroy_hdr; 1646 /* 1647 * We are in the middle of an async write. Don't destroy 1648 * this buffer unless the write completes before we finish 1649 * decrementing the reference count. 1650 */ 1651 mutex_enter(&arc_eviction_mtx); 1652 (void) remove_reference(hdr, NULL, tag); 1653 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1654 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1655 mutex_exit(&arc_eviction_mtx); 1656 if (destroy_hdr) 1657 arc_hdr_destroy(hdr); 1658 } else { 1659 if (remove_reference(hdr, NULL, tag) > 0) 1660 arc_buf_destroy(buf, FALSE, TRUE); 1661 else 1662 arc_hdr_destroy(hdr); 1663 } 1664 } 1665 1666 boolean_t 1667 arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1668 { 1669 arc_buf_hdr_t *hdr = buf->b_hdr; 1670 kmutex_t *hash_lock = HDR_LOCK(hdr); 1671 boolean_t no_callback = (buf->b_efunc == NULL); 1672 1673 if (hdr->b_state == arc_anon) { 1674 ASSERT(hdr->b_datacnt == 1); 1675 arc_buf_free(buf, tag); 1676 return (no_callback); 1677 } 1678 1679 mutex_enter(hash_lock); 1680 hdr = buf->b_hdr; 1681 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1682 ASSERT(hdr->b_state != arc_anon); 1683 ASSERT(buf->b_data != NULL); 1684 1685 (void) remove_reference(hdr, hash_lock, tag); 1686 if (hdr->b_datacnt > 1) { 1687 if (no_callback) 1688 arc_buf_destroy(buf, FALSE, TRUE); 1689 } else if (no_callback) { 1690 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1691 ASSERT(buf->b_efunc == NULL); 1692 hdr->b_flags |= ARC_BUF_AVAILABLE; 1693 } 1694 ASSERT(no_callback || hdr->b_datacnt > 1 || 1695 refcount_is_zero(&hdr->b_refcnt)); 1696 mutex_exit(hash_lock); 1697 return (no_callback); 1698 } 1699 1700 int 1701 arc_buf_size(arc_buf_t *buf) 1702 { 1703 return (buf->b_hdr->b_size); 1704 } 1705 1706 /* 1707 * Called from the DMU to determine if the current buffer should be 1708 * evicted. In order to ensure proper locking, the eviction must be initiated 1709 * from the DMU. Return true if the buffer is associated with user data and 1710 * duplicate buffers still exist. 1711 */ 1712 boolean_t 1713 arc_buf_eviction_needed(arc_buf_t *buf) 1714 { 1715 arc_buf_hdr_t *hdr; 1716 boolean_t evict_needed = B_FALSE; 1717 1718 if (zfs_disable_dup_eviction) 1719 return (B_FALSE); 1720 1721 mutex_enter(&buf->b_evict_lock); 1722 hdr = buf->b_hdr; 1723 if (hdr == NULL) { 1724 /* 1725 * We are in arc_do_user_evicts(); let that function 1726 * perform the eviction. 1727 */ 1728 ASSERT(buf->b_data == NULL); 1729 mutex_exit(&buf->b_evict_lock); 1730 return (B_FALSE); 1731 } else if (buf->b_data == NULL) { 1732 /* 1733 * We have already been added to the arc eviction list; 1734 * recommend eviction. 1735 */ 1736 ASSERT3P(hdr, ==, &arc_eviction_hdr); 1737 mutex_exit(&buf->b_evict_lock); 1738 return (B_TRUE); 1739 } 1740 1741 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 1742 evict_needed = B_TRUE; 1743 1744 mutex_exit(&buf->b_evict_lock); 1745 return (evict_needed); 1746 } 1747 1748 /* 1749 * Evict buffers from list until we've removed the specified number of 1750 * bytes. Move the removed buffers to the appropriate evict state. 1751 * If the recycle flag is set, then attempt to "recycle" a buffer: 1752 * - look for a buffer to evict that is `bytes' long. 1753 * - return the data block from this buffer rather than freeing it. 1754 * This flag is used by callers that are trying to make space for a 1755 * new buffer in a full arc cache. 1756 * 1757 * This function makes a "best effort". It skips over any buffers 1758 * it can't get a hash_lock on, and so may not catch all candidates. 1759 * It may also return without evicting as much space as requested. 1760 */ 1761 static void * 1762 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 1763 arc_buf_contents_t type) 1764 { 1765 arc_state_t *evicted_state; 1766 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1767 arc_buf_hdr_t *ab, *ab_prev = NULL; 1768 list_t *list = &state->arcs_list[type]; 1769 kmutex_t *hash_lock; 1770 boolean_t have_lock; 1771 void *stolen = NULL; 1772 1773 ASSERT(state == arc_mru || state == arc_mfu); 1774 1775 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 1776 1777 mutex_enter(&state->arcs_mtx); 1778 mutex_enter(&evicted_state->arcs_mtx); 1779 1780 for (ab = list_tail(list); ab; ab = ab_prev) { 1781 ab_prev = list_prev(list, ab); 1782 /* prefetch buffers have a minimum lifespan */ 1783 if (HDR_IO_IN_PROGRESS(ab) || 1784 (spa && ab->b_spa != spa) || 1785 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 1786 ddi_get_lbolt() - ab->b_arc_access < 1787 arc_min_prefetch_lifespan)) { 1788 skipped++; 1789 continue; 1790 } 1791 /* "lookahead" for better eviction candidate */ 1792 if (recycle && ab->b_size != bytes && 1793 ab_prev && ab_prev->b_size == bytes) 1794 continue; 1795 hash_lock = HDR_LOCK(ab); 1796 have_lock = MUTEX_HELD(hash_lock); 1797 if (have_lock || mutex_tryenter(hash_lock)) { 1798 ASSERT0(refcount_count(&ab->b_refcnt)); 1799 ASSERT(ab->b_datacnt > 0); 1800 while (ab->b_buf) { 1801 arc_buf_t *buf = ab->b_buf; 1802 if (!mutex_tryenter(&buf->b_evict_lock)) { 1803 missed += 1; 1804 break; 1805 } 1806 if (buf->b_data) { 1807 bytes_evicted += ab->b_size; 1808 if (recycle && ab->b_type == type && 1809 ab->b_size == bytes && 1810 !HDR_L2_WRITING(ab)) { 1811 stolen = buf->b_data; 1812 recycle = FALSE; 1813 } 1814 } 1815 if (buf->b_efunc) { 1816 mutex_enter(&arc_eviction_mtx); 1817 arc_buf_destroy(buf, 1818 buf->b_data == stolen, FALSE); 1819 ab->b_buf = buf->b_next; 1820 buf->b_hdr = &arc_eviction_hdr; 1821 buf->b_next = arc_eviction_list; 1822 arc_eviction_list = buf; 1823 mutex_exit(&arc_eviction_mtx); 1824 mutex_exit(&buf->b_evict_lock); 1825 } else { 1826 mutex_exit(&buf->b_evict_lock); 1827 arc_buf_destroy(buf, 1828 buf->b_data == stolen, TRUE); 1829 } 1830 } 1831 1832 if (ab->b_l2hdr) { 1833 ARCSTAT_INCR(arcstat_evict_l2_cached, 1834 ab->b_size); 1835 } else { 1836 if (l2arc_write_eligible(ab->b_spa, ab)) { 1837 ARCSTAT_INCR(arcstat_evict_l2_eligible, 1838 ab->b_size); 1839 } else { 1840 ARCSTAT_INCR( 1841 arcstat_evict_l2_ineligible, 1842 ab->b_size); 1843 } 1844 } 1845 1846 if (ab->b_datacnt == 0) { 1847 arc_change_state(evicted_state, ab, hash_lock); 1848 ASSERT(HDR_IN_HASH_TABLE(ab)); 1849 ab->b_flags |= ARC_IN_HASH_TABLE; 1850 ab->b_flags &= ~ARC_BUF_AVAILABLE; 1851 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 1852 } 1853 if (!have_lock) 1854 mutex_exit(hash_lock); 1855 if (bytes >= 0 && bytes_evicted >= bytes) 1856 break; 1857 } else { 1858 missed += 1; 1859 } 1860 } 1861 1862 mutex_exit(&evicted_state->arcs_mtx); 1863 mutex_exit(&state->arcs_mtx); 1864 1865 if (bytes_evicted < bytes) 1866 dprintf("only evicted %lld bytes from %x", 1867 (longlong_t)bytes_evicted, state); 1868 1869 if (skipped) 1870 ARCSTAT_INCR(arcstat_evict_skip, skipped); 1871 1872 if (missed) 1873 ARCSTAT_INCR(arcstat_mutex_miss, missed); 1874 1875 /* 1876 * We have just evicted some data into the ghost state, make 1877 * sure we also adjust the ghost state size if necessary. 1878 */ 1879 if (arc_no_grow && 1880 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) { 1881 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size + 1882 arc_mru_ghost->arcs_size - arc_c; 1883 1884 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) { 1885 int64_t todelete = 1886 MIN(arc_mru_ghost->arcs_lsize[type], mru_over); 1887 arc_evict_ghost(arc_mru_ghost, NULL, todelete); 1888 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) { 1889 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type], 1890 arc_mru_ghost->arcs_size + 1891 arc_mfu_ghost->arcs_size - arc_c); 1892 arc_evict_ghost(arc_mfu_ghost, NULL, todelete); 1893 } 1894 } 1895 1896 return (stolen); 1897 } 1898 1899 /* 1900 * Remove buffers from list until we've removed the specified number of 1901 * bytes. Destroy the buffers that are removed. 1902 */ 1903 static void 1904 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 1905 { 1906 arc_buf_hdr_t *ab, *ab_prev; 1907 arc_buf_hdr_t marker = { 0 }; 1908 list_t *list = &state->arcs_list[ARC_BUFC_DATA]; 1909 kmutex_t *hash_lock; 1910 uint64_t bytes_deleted = 0; 1911 uint64_t bufs_skipped = 0; 1912 1913 ASSERT(GHOST_STATE(state)); 1914 top: 1915 mutex_enter(&state->arcs_mtx); 1916 for (ab = list_tail(list); ab; ab = ab_prev) { 1917 ab_prev = list_prev(list, ab); 1918 if (spa && ab->b_spa != spa) 1919 continue; 1920 1921 /* ignore markers */ 1922 if (ab->b_spa == 0) 1923 continue; 1924 1925 hash_lock = HDR_LOCK(ab); 1926 /* caller may be trying to modify this buffer, skip it */ 1927 if (MUTEX_HELD(hash_lock)) 1928 continue; 1929 if (mutex_tryenter(hash_lock)) { 1930 ASSERT(!HDR_IO_IN_PROGRESS(ab)); 1931 ASSERT(ab->b_buf == NULL); 1932 ARCSTAT_BUMP(arcstat_deleted); 1933 bytes_deleted += ab->b_size; 1934 1935 if (ab->b_l2hdr != NULL) { 1936 /* 1937 * This buffer is cached on the 2nd Level ARC; 1938 * don't destroy the header. 1939 */ 1940 arc_change_state(arc_l2c_only, ab, hash_lock); 1941 mutex_exit(hash_lock); 1942 } else { 1943 arc_change_state(arc_anon, ab, hash_lock); 1944 mutex_exit(hash_lock); 1945 arc_hdr_destroy(ab); 1946 } 1947 1948 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 1949 if (bytes >= 0 && bytes_deleted >= bytes) 1950 break; 1951 } else if (bytes < 0) { 1952 /* 1953 * Insert a list marker and then wait for the 1954 * hash lock to become available. Once its 1955 * available, restart from where we left off. 1956 */ 1957 list_insert_after(list, ab, &marker); 1958 mutex_exit(&state->arcs_mtx); 1959 mutex_enter(hash_lock); 1960 mutex_exit(hash_lock); 1961 mutex_enter(&state->arcs_mtx); 1962 ab_prev = list_prev(list, &marker); 1963 list_remove(list, &marker); 1964 } else 1965 bufs_skipped += 1; 1966 } 1967 mutex_exit(&state->arcs_mtx); 1968 1969 if (list == &state->arcs_list[ARC_BUFC_DATA] && 1970 (bytes < 0 || bytes_deleted < bytes)) { 1971 list = &state->arcs_list[ARC_BUFC_METADATA]; 1972 goto top; 1973 } 1974 1975 if (bufs_skipped) { 1976 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 1977 ASSERT(bytes >= 0); 1978 } 1979 1980 if (bytes_deleted < bytes) 1981 dprintf("only deleted %lld bytes from %p", 1982 (longlong_t)bytes_deleted, state); 1983 } 1984 1985 static void 1986 arc_adjust(void) 1987 { 1988 int64_t adjustment, delta; 1989 1990 /* 1991 * Adjust MRU size 1992 */ 1993 1994 adjustment = MIN((int64_t)(arc_size - arc_c), 1995 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 1996 arc_p)); 1997 1998 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 1999 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2000 (void) arc_evict(arc_mru, NULL, delta, FALSE, ARC_BUFC_DATA); 2001 adjustment -= delta; 2002 } 2003 2004 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2005 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2006 (void) arc_evict(arc_mru, NULL, delta, FALSE, 2007 ARC_BUFC_METADATA); 2008 } 2009 2010 /* 2011 * Adjust MFU size 2012 */ 2013 2014 adjustment = arc_size - arc_c; 2015 2016 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2017 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2018 (void) arc_evict(arc_mfu, NULL, delta, FALSE, ARC_BUFC_DATA); 2019 adjustment -= delta; 2020 } 2021 2022 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2023 int64_t delta = MIN(adjustment, 2024 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2025 (void) arc_evict(arc_mfu, NULL, delta, FALSE, 2026 ARC_BUFC_METADATA); 2027 } 2028 2029 /* 2030 * Adjust ghost lists 2031 */ 2032 2033 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2034 2035 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2036 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2037 arc_evict_ghost(arc_mru_ghost, NULL, delta); 2038 } 2039 2040 adjustment = 2041 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2042 2043 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2044 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2045 arc_evict_ghost(arc_mfu_ghost, NULL, delta); 2046 } 2047 } 2048 2049 static void 2050 arc_do_user_evicts(void) 2051 { 2052 mutex_enter(&arc_eviction_mtx); 2053 while (arc_eviction_list != NULL) { 2054 arc_buf_t *buf = arc_eviction_list; 2055 arc_eviction_list = buf->b_next; 2056 mutex_enter(&buf->b_evict_lock); 2057 buf->b_hdr = NULL; 2058 mutex_exit(&buf->b_evict_lock); 2059 mutex_exit(&arc_eviction_mtx); 2060 2061 if (buf->b_efunc != NULL) 2062 VERIFY(buf->b_efunc(buf) == 0); 2063 2064 buf->b_efunc = NULL; 2065 buf->b_private = NULL; 2066 kmem_cache_free(buf_cache, buf); 2067 mutex_enter(&arc_eviction_mtx); 2068 } 2069 mutex_exit(&arc_eviction_mtx); 2070 } 2071 2072 /* 2073 * Flush all *evictable* data from the cache for the given spa. 2074 * NOTE: this will not touch "active" (i.e. referenced) data. 2075 */ 2076 void 2077 arc_flush(spa_t *spa) 2078 { 2079 uint64_t guid = 0; 2080 2081 if (spa) 2082 guid = spa_load_guid(spa); 2083 2084 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) { 2085 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2086 if (spa) 2087 break; 2088 } 2089 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) { 2090 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2091 if (spa) 2092 break; 2093 } 2094 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) { 2095 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2096 if (spa) 2097 break; 2098 } 2099 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) { 2100 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2101 if (spa) 2102 break; 2103 } 2104 2105 arc_evict_ghost(arc_mru_ghost, guid, -1); 2106 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2107 2108 mutex_enter(&arc_reclaim_thr_lock); 2109 arc_do_user_evicts(); 2110 mutex_exit(&arc_reclaim_thr_lock); 2111 ASSERT(spa || arc_eviction_list == NULL); 2112 } 2113 2114 void 2115 arc_shrink(void) 2116 { 2117 if (arc_c > arc_c_min) { 2118 uint64_t to_free; 2119 2120 #ifdef _KERNEL 2121 to_free = MAX(arc_c >> arc_shrink_shift, ptob(needfree)); 2122 #else 2123 to_free = arc_c >> arc_shrink_shift; 2124 #endif 2125 if (arc_c > arc_c_min + to_free) 2126 atomic_add_64(&arc_c, -to_free); 2127 else 2128 arc_c = arc_c_min; 2129 2130 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2131 if (arc_c > arc_size) 2132 arc_c = MAX(arc_size, arc_c_min); 2133 if (arc_p > arc_c) 2134 arc_p = (arc_c >> 1); 2135 ASSERT(arc_c >= arc_c_min); 2136 ASSERT((int64_t)arc_p >= 0); 2137 } 2138 2139 if (arc_size > arc_c) 2140 arc_adjust(); 2141 } 2142 2143 /* 2144 * Determine if the system is under memory pressure and is asking 2145 * to reclaim memory. A return value of 1 indicates that the system 2146 * is under memory pressure and that the arc should adjust accordingly. 2147 */ 2148 static int 2149 arc_reclaim_needed(void) 2150 { 2151 uint64_t extra; 2152 2153 #ifdef _KERNEL 2154 2155 if (needfree) 2156 return (1); 2157 2158 /* 2159 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2160 */ 2161 extra = desfree; 2162 2163 /* 2164 * check that we're out of range of the pageout scanner. It starts to 2165 * schedule paging if freemem is less than lotsfree and needfree. 2166 * lotsfree is the high-water mark for pageout, and needfree is the 2167 * number of needed free pages. We add extra pages here to make sure 2168 * the scanner doesn't start up while we're freeing memory. 2169 */ 2170 if (freemem < lotsfree + needfree + extra) 2171 return (1); 2172 2173 /* 2174 * check to make sure that swapfs has enough space so that anon 2175 * reservations can still succeed. anon_resvmem() checks that the 2176 * availrmem is greater than swapfs_minfree, and the number of reserved 2177 * swap pages. We also add a bit of extra here just to prevent 2178 * circumstances from getting really dire. 2179 */ 2180 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2181 return (1); 2182 2183 #if defined(__i386) 2184 /* 2185 * If we're on an i386 platform, it's possible that we'll exhaust the 2186 * kernel heap space before we ever run out of available physical 2187 * memory. Most checks of the size of the heap_area compare against 2188 * tune.t_minarmem, which is the minimum available real memory that we 2189 * can have in the system. However, this is generally fixed at 25 pages 2190 * which is so low that it's useless. In this comparison, we seek to 2191 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2192 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2193 * free) 2194 */ 2195 if (vmem_size(heap_arena, VMEM_FREE) < 2196 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) 2197 return (1); 2198 #endif 2199 2200 /* 2201 * If zio data pages are being allocated out of a separate heap segment, 2202 * then enforce that the size of available vmem for this arena remains 2203 * above about 1/16th free. 2204 * 2205 * Note: The 1/16th arena free requirement was put in place 2206 * to aggressively evict memory from the arc in order to avoid 2207 * memory fragmentation issues. 2208 */ 2209 if (zio_arena != NULL && 2210 vmem_size(zio_arena, VMEM_FREE) < 2211 (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2212 return (1); 2213 #else 2214 if (spa_get_random(100) == 0) 2215 return (1); 2216 #endif 2217 return (0); 2218 } 2219 2220 static void 2221 arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2222 { 2223 size_t i; 2224 kmem_cache_t *prev_cache = NULL; 2225 kmem_cache_t *prev_data_cache = NULL; 2226 extern kmem_cache_t *zio_buf_cache[]; 2227 extern kmem_cache_t *zio_data_buf_cache[]; 2228 2229 #ifdef _KERNEL 2230 if (arc_meta_used >= arc_meta_limit) { 2231 /* 2232 * We are exceeding our meta-data cache limit. 2233 * Purge some DNLC entries to release holds on meta-data. 2234 */ 2235 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2236 } 2237 #if defined(__i386) 2238 /* 2239 * Reclaim unused memory from all kmem caches. 2240 */ 2241 kmem_reap(); 2242 #endif 2243 #endif 2244 2245 /* 2246 * An aggressive reclamation will shrink the cache size as well as 2247 * reap free buffers from the arc kmem caches. 2248 */ 2249 if (strat == ARC_RECLAIM_AGGR) 2250 arc_shrink(); 2251 2252 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2253 if (zio_buf_cache[i] != prev_cache) { 2254 prev_cache = zio_buf_cache[i]; 2255 kmem_cache_reap_now(zio_buf_cache[i]); 2256 } 2257 if (zio_data_buf_cache[i] != prev_data_cache) { 2258 prev_data_cache = zio_data_buf_cache[i]; 2259 kmem_cache_reap_now(zio_data_buf_cache[i]); 2260 } 2261 } 2262 kmem_cache_reap_now(buf_cache); 2263 kmem_cache_reap_now(hdr_cache); 2264 2265 /* 2266 * Ask the vmem areana to reclaim unused memory from its 2267 * quantum caches. 2268 */ 2269 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 2270 vmem_qcache_reap(zio_arena); 2271 } 2272 2273 static void 2274 arc_reclaim_thread(void) 2275 { 2276 clock_t growtime = 0; 2277 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2278 callb_cpr_t cpr; 2279 2280 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2281 2282 mutex_enter(&arc_reclaim_thr_lock); 2283 while (arc_thread_exit == 0) { 2284 if (arc_reclaim_needed()) { 2285 2286 if (arc_no_grow) { 2287 if (last_reclaim == ARC_RECLAIM_CONS) { 2288 last_reclaim = ARC_RECLAIM_AGGR; 2289 } else { 2290 last_reclaim = ARC_RECLAIM_CONS; 2291 } 2292 } else { 2293 arc_no_grow = TRUE; 2294 last_reclaim = ARC_RECLAIM_AGGR; 2295 membar_producer(); 2296 } 2297 2298 /* reset the growth delay for every reclaim */ 2299 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2300 2301 arc_kmem_reap_now(last_reclaim); 2302 arc_warm = B_TRUE; 2303 2304 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2305 arc_no_grow = FALSE; 2306 } 2307 2308 arc_adjust(); 2309 2310 if (arc_eviction_list != NULL) 2311 arc_do_user_evicts(); 2312 2313 /* block until needed, or one second, whichever is shorter */ 2314 CALLB_CPR_SAFE_BEGIN(&cpr); 2315 (void) cv_timedwait(&arc_reclaim_thr_cv, 2316 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz)); 2317 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2318 } 2319 2320 arc_thread_exit = 0; 2321 cv_broadcast(&arc_reclaim_thr_cv); 2322 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2323 thread_exit(); 2324 } 2325 2326 /* 2327 * Adapt arc info given the number of bytes we are trying to add and 2328 * the state that we are comming from. This function is only called 2329 * when we are adding new content to the cache. 2330 */ 2331 static void 2332 arc_adapt(int bytes, arc_state_t *state) 2333 { 2334 int mult; 2335 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2336 2337 if (state == arc_l2c_only) 2338 return; 2339 2340 ASSERT(bytes > 0); 2341 /* 2342 * Adapt the target size of the MRU list: 2343 * - if we just hit in the MRU ghost list, then increase 2344 * the target size of the MRU list. 2345 * - if we just hit in the MFU ghost list, then increase 2346 * the target size of the MFU list by decreasing the 2347 * target size of the MRU list. 2348 */ 2349 if (state == arc_mru_ghost) { 2350 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2351 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2352 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2353 2354 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2355 } else if (state == arc_mfu_ghost) { 2356 uint64_t delta; 2357 2358 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2359 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2360 mult = MIN(mult, 10); 2361 2362 delta = MIN(bytes * mult, arc_p); 2363 arc_p = MAX(arc_p_min, arc_p - delta); 2364 } 2365 ASSERT((int64_t)arc_p >= 0); 2366 2367 if (arc_reclaim_needed()) { 2368 cv_signal(&arc_reclaim_thr_cv); 2369 return; 2370 } 2371 2372 if (arc_no_grow) 2373 return; 2374 2375 if (arc_c >= arc_c_max) 2376 return; 2377 2378 /* 2379 * If we're within (2 * maxblocksize) bytes of the target 2380 * cache size, increment the target cache size 2381 */ 2382 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2383 atomic_add_64(&arc_c, (int64_t)bytes); 2384 if (arc_c > arc_c_max) 2385 arc_c = arc_c_max; 2386 else if (state == arc_anon) 2387 atomic_add_64(&arc_p, (int64_t)bytes); 2388 if (arc_p > arc_c) 2389 arc_p = arc_c; 2390 } 2391 ASSERT((int64_t)arc_p >= 0); 2392 } 2393 2394 /* 2395 * Check if the cache has reached its limits and eviction is required 2396 * prior to insert. 2397 */ 2398 static int 2399 arc_evict_needed(arc_buf_contents_t type) 2400 { 2401 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2402 return (1); 2403 2404 if (arc_reclaim_needed()) 2405 return (1); 2406 2407 return (arc_size > arc_c); 2408 } 2409 2410 /* 2411 * The buffer, supplied as the first argument, needs a data block. 2412 * So, if we are at cache max, determine which cache should be victimized. 2413 * We have the following cases: 2414 * 2415 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2416 * In this situation if we're out of space, but the resident size of the MFU is 2417 * under the limit, victimize the MFU cache to satisfy this insertion request. 2418 * 2419 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2420 * Here, we've used up all of the available space for the MRU, so we need to 2421 * evict from our own cache instead. Evict from the set of resident MRU 2422 * entries. 2423 * 2424 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2425 * c minus p represents the MFU space in the cache, since p is the size of the 2426 * cache that is dedicated to the MRU. In this situation there's still space on 2427 * the MFU side, so the MRU side needs to be victimized. 2428 * 2429 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2430 * MFU's resident set is consuming more space than it has been allotted. In 2431 * this situation, we must victimize our own cache, the MFU, for this insertion. 2432 */ 2433 static void 2434 arc_get_data_buf(arc_buf_t *buf) 2435 { 2436 arc_state_t *state = buf->b_hdr->b_state; 2437 uint64_t size = buf->b_hdr->b_size; 2438 arc_buf_contents_t type = buf->b_hdr->b_type; 2439 2440 arc_adapt(size, state); 2441 2442 /* 2443 * We have not yet reached cache maximum size, 2444 * just allocate a new buffer. 2445 */ 2446 if (!arc_evict_needed(type)) { 2447 if (type == ARC_BUFC_METADATA) { 2448 buf->b_data = zio_buf_alloc(size); 2449 arc_space_consume(size, ARC_SPACE_DATA); 2450 } else { 2451 ASSERT(type == ARC_BUFC_DATA); 2452 buf->b_data = zio_data_buf_alloc(size); 2453 ARCSTAT_INCR(arcstat_data_size, size); 2454 atomic_add_64(&arc_size, size); 2455 } 2456 goto out; 2457 } 2458 2459 /* 2460 * If we are prefetching from the mfu ghost list, this buffer 2461 * will end up on the mru list; so steal space from there. 2462 */ 2463 if (state == arc_mfu_ghost) 2464 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2465 else if (state == arc_mru_ghost) 2466 state = arc_mru; 2467 2468 if (state == arc_mru || state == arc_anon) { 2469 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2470 state = (arc_mfu->arcs_lsize[type] >= size && 2471 arc_p > mru_used) ? arc_mfu : arc_mru; 2472 } else { 2473 /* MFU cases */ 2474 uint64_t mfu_space = arc_c - arc_p; 2475 state = (arc_mru->arcs_lsize[type] >= size && 2476 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2477 } 2478 if ((buf->b_data = arc_evict(state, NULL, size, TRUE, type)) == NULL) { 2479 if (type == ARC_BUFC_METADATA) { 2480 buf->b_data = zio_buf_alloc(size); 2481 arc_space_consume(size, ARC_SPACE_DATA); 2482 } else { 2483 ASSERT(type == ARC_BUFC_DATA); 2484 buf->b_data = zio_data_buf_alloc(size); 2485 ARCSTAT_INCR(arcstat_data_size, size); 2486 atomic_add_64(&arc_size, size); 2487 } 2488 ARCSTAT_BUMP(arcstat_recycle_miss); 2489 } 2490 ASSERT(buf->b_data != NULL); 2491 out: 2492 /* 2493 * Update the state size. Note that ghost states have a 2494 * "ghost size" and so don't need to be updated. 2495 */ 2496 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2497 arc_buf_hdr_t *hdr = buf->b_hdr; 2498 2499 atomic_add_64(&hdr->b_state->arcs_size, size); 2500 if (list_link_active(&hdr->b_arc_node)) { 2501 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2502 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2503 } 2504 /* 2505 * If we are growing the cache, and we are adding anonymous 2506 * data, and we have outgrown arc_p, update arc_p 2507 */ 2508 if (arc_size < arc_c && hdr->b_state == arc_anon && 2509 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2510 arc_p = MIN(arc_c, arc_p + size); 2511 } 2512 } 2513 2514 /* 2515 * This routine is called whenever a buffer is accessed. 2516 * NOTE: the hash lock is dropped in this function. 2517 */ 2518 static void 2519 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2520 { 2521 clock_t now; 2522 2523 ASSERT(MUTEX_HELD(hash_lock)); 2524 2525 if (buf->b_state == arc_anon) { 2526 /* 2527 * This buffer is not in the cache, and does not 2528 * appear in our "ghost" list. Add the new buffer 2529 * to the MRU state. 2530 */ 2531 2532 ASSERT(buf->b_arc_access == 0); 2533 buf->b_arc_access = ddi_get_lbolt(); 2534 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2535 arc_change_state(arc_mru, buf, hash_lock); 2536 2537 } else if (buf->b_state == arc_mru) { 2538 now = ddi_get_lbolt(); 2539 2540 /* 2541 * If this buffer is here because of a prefetch, then either: 2542 * - clear the flag if this is a "referencing" read 2543 * (any subsequent access will bump this into the MFU state). 2544 * or 2545 * - move the buffer to the head of the list if this is 2546 * another prefetch (to make it less likely to be evicted). 2547 */ 2548 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2549 if (refcount_count(&buf->b_refcnt) == 0) { 2550 ASSERT(list_link_active(&buf->b_arc_node)); 2551 } else { 2552 buf->b_flags &= ~ARC_PREFETCH; 2553 ARCSTAT_BUMP(arcstat_mru_hits); 2554 } 2555 buf->b_arc_access = now; 2556 return; 2557 } 2558 2559 /* 2560 * This buffer has been "accessed" only once so far, 2561 * but it is still in the cache. Move it to the MFU 2562 * state. 2563 */ 2564 if (now > buf->b_arc_access + ARC_MINTIME) { 2565 /* 2566 * More than 125ms have passed since we 2567 * instantiated this buffer. Move it to the 2568 * most frequently used state. 2569 */ 2570 buf->b_arc_access = now; 2571 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2572 arc_change_state(arc_mfu, buf, hash_lock); 2573 } 2574 ARCSTAT_BUMP(arcstat_mru_hits); 2575 } else if (buf->b_state == arc_mru_ghost) { 2576 arc_state_t *new_state; 2577 /* 2578 * This buffer has been "accessed" recently, but 2579 * was evicted from the cache. Move it to the 2580 * MFU state. 2581 */ 2582 2583 if (buf->b_flags & ARC_PREFETCH) { 2584 new_state = arc_mru; 2585 if (refcount_count(&buf->b_refcnt) > 0) 2586 buf->b_flags &= ~ARC_PREFETCH; 2587 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2588 } else { 2589 new_state = arc_mfu; 2590 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2591 } 2592 2593 buf->b_arc_access = ddi_get_lbolt(); 2594 arc_change_state(new_state, buf, hash_lock); 2595 2596 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2597 } else if (buf->b_state == arc_mfu) { 2598 /* 2599 * This buffer has been accessed more than once and is 2600 * still in the cache. Keep it in the MFU state. 2601 * 2602 * NOTE: an add_reference() that occurred when we did 2603 * the arc_read() will have kicked this off the list. 2604 * If it was a prefetch, we will explicitly move it to 2605 * the head of the list now. 2606 */ 2607 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2608 ASSERT(refcount_count(&buf->b_refcnt) == 0); 2609 ASSERT(list_link_active(&buf->b_arc_node)); 2610 } 2611 ARCSTAT_BUMP(arcstat_mfu_hits); 2612 buf->b_arc_access = ddi_get_lbolt(); 2613 } else if (buf->b_state == arc_mfu_ghost) { 2614 arc_state_t *new_state = arc_mfu; 2615 /* 2616 * This buffer has been accessed more than once but has 2617 * been evicted from the cache. Move it back to the 2618 * MFU state. 2619 */ 2620 2621 if (buf->b_flags & ARC_PREFETCH) { 2622 /* 2623 * This is a prefetch access... 2624 * move this block back to the MRU state. 2625 */ 2626 ASSERT0(refcount_count(&buf->b_refcnt)); 2627 new_state = arc_mru; 2628 } 2629 2630 buf->b_arc_access = ddi_get_lbolt(); 2631 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2632 arc_change_state(new_state, buf, hash_lock); 2633 2634 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 2635 } else if (buf->b_state == arc_l2c_only) { 2636 /* 2637 * This buffer is on the 2nd Level ARC. 2638 */ 2639 2640 buf->b_arc_access = ddi_get_lbolt(); 2641 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2642 arc_change_state(arc_mfu, buf, hash_lock); 2643 } else { 2644 ASSERT(!"invalid arc state"); 2645 } 2646 } 2647 2648 /* a generic arc_done_func_t which you can use */ 2649 /* ARGSUSED */ 2650 void 2651 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 2652 { 2653 if (zio == NULL || zio->io_error == 0) 2654 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 2655 VERIFY(arc_buf_remove_ref(buf, arg)); 2656 } 2657 2658 /* a generic arc_done_func_t */ 2659 void 2660 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 2661 { 2662 arc_buf_t **bufp = arg; 2663 if (zio && zio->io_error) { 2664 VERIFY(arc_buf_remove_ref(buf, arg)); 2665 *bufp = NULL; 2666 } else { 2667 *bufp = buf; 2668 ASSERT(buf->b_data); 2669 } 2670 } 2671 2672 static void 2673 arc_read_done(zio_t *zio) 2674 { 2675 arc_buf_hdr_t *hdr, *found; 2676 arc_buf_t *buf; 2677 arc_buf_t *abuf; /* buffer we're assigning to callback */ 2678 kmutex_t *hash_lock; 2679 arc_callback_t *callback_list, *acb; 2680 int freeable = FALSE; 2681 2682 buf = zio->io_private; 2683 hdr = buf->b_hdr; 2684 2685 /* 2686 * The hdr was inserted into hash-table and removed from lists 2687 * prior to starting I/O. We should find this header, since 2688 * it's in the hash table, and it should be legit since it's 2689 * not possible to evict it during the I/O. The only possible 2690 * reason for it not to be found is if we were freed during the 2691 * read. 2692 */ 2693 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth, 2694 &hash_lock); 2695 2696 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) || 2697 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 2698 (found == hdr && HDR_L2_READING(hdr))); 2699 2700 hdr->b_flags &= ~ARC_L2_EVICTED; 2701 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 2702 hdr->b_flags &= ~ARC_L2CACHE; 2703 2704 /* byteswap if necessary */ 2705 callback_list = hdr->b_acb; 2706 ASSERT(callback_list != NULL); 2707 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 2708 dmu_object_byteswap_t bswap = 2709 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 2710 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 2711 byteswap_uint64_array : 2712 dmu_ot_byteswap[bswap].ob_func; 2713 func(buf->b_data, hdr->b_size); 2714 } 2715 2716 arc_cksum_compute(buf, B_FALSE); 2717 arc_buf_watch(buf); 2718 2719 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 2720 /* 2721 * Only call arc_access on anonymous buffers. This is because 2722 * if we've issued an I/O for an evicted buffer, we've already 2723 * called arc_access (to prevent any simultaneous readers from 2724 * getting confused). 2725 */ 2726 arc_access(hdr, hash_lock); 2727 } 2728 2729 /* create copies of the data buffer for the callers */ 2730 abuf = buf; 2731 for (acb = callback_list; acb; acb = acb->acb_next) { 2732 if (acb->acb_done) { 2733 if (abuf == NULL) { 2734 ARCSTAT_BUMP(arcstat_duplicate_reads); 2735 abuf = arc_buf_clone(buf); 2736 } 2737 acb->acb_buf = abuf; 2738 abuf = NULL; 2739 } 2740 } 2741 hdr->b_acb = NULL; 2742 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 2743 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 2744 if (abuf == buf) { 2745 ASSERT(buf->b_efunc == NULL); 2746 ASSERT(hdr->b_datacnt == 1); 2747 hdr->b_flags |= ARC_BUF_AVAILABLE; 2748 } 2749 2750 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 2751 2752 if (zio->io_error != 0) { 2753 hdr->b_flags |= ARC_IO_ERROR; 2754 if (hdr->b_state != arc_anon) 2755 arc_change_state(arc_anon, hdr, hash_lock); 2756 if (HDR_IN_HASH_TABLE(hdr)) 2757 buf_hash_remove(hdr); 2758 freeable = refcount_is_zero(&hdr->b_refcnt); 2759 } 2760 2761 /* 2762 * Broadcast before we drop the hash_lock to avoid the possibility 2763 * that the hdr (and hence the cv) might be freed before we get to 2764 * the cv_broadcast(). 2765 */ 2766 cv_broadcast(&hdr->b_cv); 2767 2768 if (hash_lock) { 2769 mutex_exit(hash_lock); 2770 } else { 2771 /* 2772 * This block was freed while we waited for the read to 2773 * complete. It has been removed from the hash table and 2774 * moved to the anonymous state (so that it won't show up 2775 * in the cache). 2776 */ 2777 ASSERT3P(hdr->b_state, ==, arc_anon); 2778 freeable = refcount_is_zero(&hdr->b_refcnt); 2779 } 2780 2781 /* execute each callback and free its structure */ 2782 while ((acb = callback_list) != NULL) { 2783 if (acb->acb_done) 2784 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 2785 2786 if (acb->acb_zio_dummy != NULL) { 2787 acb->acb_zio_dummy->io_error = zio->io_error; 2788 zio_nowait(acb->acb_zio_dummy); 2789 } 2790 2791 callback_list = acb->acb_next; 2792 kmem_free(acb, sizeof (arc_callback_t)); 2793 } 2794 2795 if (freeable) 2796 arc_hdr_destroy(hdr); 2797 } 2798 2799 /* 2800 * "Read" the block at the specified DVA (in bp) via the 2801 * cache. If the block is found in the cache, invoke the provided 2802 * callback immediately and return. Note that the `zio' parameter 2803 * in the callback will be NULL in this case, since no IO was 2804 * required. If the block is not in the cache pass the read request 2805 * on to the spa with a substitute callback function, so that the 2806 * requested block will be added to the cache. 2807 * 2808 * If a read request arrives for a block that has a read in-progress, 2809 * either wait for the in-progress read to complete (and return the 2810 * results); or, if this is a read with a "done" func, add a record 2811 * to the read to invoke the "done" func when the read completes, 2812 * and return; or just return. 2813 * 2814 * arc_read_done() will invoke all the requested "done" functions 2815 * for readers of this block. 2816 */ 2817 int 2818 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 2819 void *private, int priority, int zio_flags, uint32_t *arc_flags, 2820 const zbookmark_t *zb) 2821 { 2822 arc_buf_hdr_t *hdr; 2823 arc_buf_t *buf = NULL; 2824 kmutex_t *hash_lock; 2825 zio_t *rzio; 2826 uint64_t guid = spa_load_guid(spa); 2827 2828 top: 2829 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp), 2830 &hash_lock); 2831 if (hdr && hdr->b_datacnt > 0) { 2832 2833 *arc_flags |= ARC_CACHED; 2834 2835 if (HDR_IO_IN_PROGRESS(hdr)) { 2836 2837 if (*arc_flags & ARC_WAIT) { 2838 cv_wait(&hdr->b_cv, hash_lock); 2839 mutex_exit(hash_lock); 2840 goto top; 2841 } 2842 ASSERT(*arc_flags & ARC_NOWAIT); 2843 2844 if (done) { 2845 arc_callback_t *acb = NULL; 2846 2847 acb = kmem_zalloc(sizeof (arc_callback_t), 2848 KM_SLEEP); 2849 acb->acb_done = done; 2850 acb->acb_private = private; 2851 if (pio != NULL) 2852 acb->acb_zio_dummy = zio_null(pio, 2853 spa, NULL, NULL, NULL, zio_flags); 2854 2855 ASSERT(acb->acb_done != NULL); 2856 acb->acb_next = hdr->b_acb; 2857 hdr->b_acb = acb; 2858 add_reference(hdr, hash_lock, private); 2859 mutex_exit(hash_lock); 2860 return (0); 2861 } 2862 mutex_exit(hash_lock); 2863 return (0); 2864 } 2865 2866 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 2867 2868 if (done) { 2869 add_reference(hdr, hash_lock, private); 2870 /* 2871 * If this block is already in use, create a new 2872 * copy of the data so that we will be guaranteed 2873 * that arc_release() will always succeed. 2874 */ 2875 buf = hdr->b_buf; 2876 ASSERT(buf); 2877 ASSERT(buf->b_data); 2878 if (HDR_BUF_AVAILABLE(hdr)) { 2879 ASSERT(buf->b_efunc == NULL); 2880 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 2881 } else { 2882 buf = arc_buf_clone(buf); 2883 } 2884 2885 } else if (*arc_flags & ARC_PREFETCH && 2886 refcount_count(&hdr->b_refcnt) == 0) { 2887 hdr->b_flags |= ARC_PREFETCH; 2888 } 2889 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 2890 arc_access(hdr, hash_lock); 2891 if (*arc_flags & ARC_L2CACHE) 2892 hdr->b_flags |= ARC_L2CACHE; 2893 if (*arc_flags & ARC_L2COMPRESS) 2894 hdr->b_flags |= ARC_L2COMPRESS; 2895 mutex_exit(hash_lock); 2896 ARCSTAT_BUMP(arcstat_hits); 2897 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 2898 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 2899 data, metadata, hits); 2900 2901 if (done) 2902 done(NULL, buf, private); 2903 } else { 2904 uint64_t size = BP_GET_LSIZE(bp); 2905 arc_callback_t *acb; 2906 vdev_t *vd = NULL; 2907 uint64_t addr = 0; 2908 boolean_t devw = B_FALSE; 2909 2910 if (hdr == NULL) { 2911 /* this block is not in the cache */ 2912 arc_buf_hdr_t *exists; 2913 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 2914 buf = arc_buf_alloc(spa, size, private, type); 2915 hdr = buf->b_hdr; 2916 hdr->b_dva = *BP_IDENTITY(bp); 2917 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 2918 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 2919 exists = buf_hash_insert(hdr, &hash_lock); 2920 if (exists) { 2921 /* somebody beat us to the hash insert */ 2922 mutex_exit(hash_lock); 2923 buf_discard_identity(hdr); 2924 (void) arc_buf_remove_ref(buf, private); 2925 goto top; /* restart the IO request */ 2926 } 2927 /* if this is a prefetch, we don't have a reference */ 2928 if (*arc_flags & ARC_PREFETCH) { 2929 (void) remove_reference(hdr, hash_lock, 2930 private); 2931 hdr->b_flags |= ARC_PREFETCH; 2932 } 2933 if (*arc_flags & ARC_L2CACHE) 2934 hdr->b_flags |= ARC_L2CACHE; 2935 if (*arc_flags & ARC_L2COMPRESS) 2936 hdr->b_flags |= ARC_L2COMPRESS; 2937 if (BP_GET_LEVEL(bp) > 0) 2938 hdr->b_flags |= ARC_INDIRECT; 2939 } else { 2940 /* this block is in the ghost cache */ 2941 ASSERT(GHOST_STATE(hdr->b_state)); 2942 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2943 ASSERT0(refcount_count(&hdr->b_refcnt)); 2944 ASSERT(hdr->b_buf == NULL); 2945 2946 /* if this is a prefetch, we don't have a reference */ 2947 if (*arc_flags & ARC_PREFETCH) 2948 hdr->b_flags |= ARC_PREFETCH; 2949 else 2950 add_reference(hdr, hash_lock, private); 2951 if (*arc_flags & ARC_L2CACHE) 2952 hdr->b_flags |= ARC_L2CACHE; 2953 if (*arc_flags & ARC_L2COMPRESS) 2954 hdr->b_flags |= ARC_L2COMPRESS; 2955 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2956 buf->b_hdr = hdr; 2957 buf->b_data = NULL; 2958 buf->b_efunc = NULL; 2959 buf->b_private = NULL; 2960 buf->b_next = NULL; 2961 hdr->b_buf = buf; 2962 ASSERT(hdr->b_datacnt == 0); 2963 hdr->b_datacnt = 1; 2964 arc_get_data_buf(buf); 2965 arc_access(hdr, hash_lock); 2966 } 2967 2968 ASSERT(!GHOST_STATE(hdr->b_state)); 2969 2970 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 2971 acb->acb_done = done; 2972 acb->acb_private = private; 2973 2974 ASSERT(hdr->b_acb == NULL); 2975 hdr->b_acb = acb; 2976 hdr->b_flags |= ARC_IO_IN_PROGRESS; 2977 2978 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL && 2979 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 2980 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 2981 addr = hdr->b_l2hdr->b_daddr; 2982 /* 2983 * Lock out device removal. 2984 */ 2985 if (vdev_is_dead(vd) || 2986 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 2987 vd = NULL; 2988 } 2989 2990 mutex_exit(hash_lock); 2991 2992 ASSERT3U(hdr->b_size, ==, size); 2993 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 2994 uint64_t, size, zbookmark_t *, zb); 2995 ARCSTAT_BUMP(arcstat_misses); 2996 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 2997 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 2998 data, metadata, misses); 2999 3000 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3001 /* 3002 * Read from the L2ARC if the following are true: 3003 * 1. The L2ARC vdev was previously cached. 3004 * 2. This buffer still has L2ARC metadata. 3005 * 3. This buffer isn't currently writing to the L2ARC. 3006 * 4. The L2ARC entry wasn't evicted, which may 3007 * also have invalidated the vdev. 3008 * 5. This isn't prefetch and l2arc_noprefetch is set. 3009 */ 3010 if (hdr->b_l2hdr != NULL && 3011 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3012 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3013 l2arc_read_callback_t *cb; 3014 3015 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3016 ARCSTAT_BUMP(arcstat_l2_hits); 3017 3018 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3019 KM_SLEEP); 3020 cb->l2rcb_buf = buf; 3021 cb->l2rcb_spa = spa; 3022 cb->l2rcb_bp = *bp; 3023 cb->l2rcb_zb = *zb; 3024 cb->l2rcb_flags = zio_flags; 3025 cb->l2rcb_compress = hdr->b_l2hdr->b_compress; 3026 3027 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3028 addr + size < vd->vdev_psize - 3029 VDEV_LABEL_END_SIZE); 3030 3031 /* 3032 * l2arc read. The SCL_L2ARC lock will be 3033 * released by l2arc_read_done(). 3034 * Issue a null zio if the underlying buffer 3035 * was squashed to zero size by compression. 3036 */ 3037 if (hdr->b_l2hdr->b_compress == 3038 ZIO_COMPRESS_EMPTY) { 3039 rzio = zio_null(pio, spa, vd, 3040 l2arc_read_done, cb, 3041 zio_flags | ZIO_FLAG_DONT_CACHE | 3042 ZIO_FLAG_CANFAIL | 3043 ZIO_FLAG_DONT_PROPAGATE | 3044 ZIO_FLAG_DONT_RETRY); 3045 } else { 3046 rzio = zio_read_phys(pio, vd, addr, 3047 hdr->b_l2hdr->b_asize, 3048 buf->b_data, ZIO_CHECKSUM_OFF, 3049 l2arc_read_done, cb, priority, 3050 zio_flags | ZIO_FLAG_DONT_CACHE | 3051 ZIO_FLAG_CANFAIL | 3052 ZIO_FLAG_DONT_PROPAGATE | 3053 ZIO_FLAG_DONT_RETRY, B_FALSE); 3054 } 3055 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3056 zio_t *, rzio); 3057 ARCSTAT_INCR(arcstat_l2_read_bytes, 3058 hdr->b_l2hdr->b_asize); 3059 3060 if (*arc_flags & ARC_NOWAIT) { 3061 zio_nowait(rzio); 3062 return (0); 3063 } 3064 3065 ASSERT(*arc_flags & ARC_WAIT); 3066 if (zio_wait(rzio) == 0) 3067 return (0); 3068 3069 /* l2arc read error; goto zio_read() */ 3070 } else { 3071 DTRACE_PROBE1(l2arc__miss, 3072 arc_buf_hdr_t *, hdr); 3073 ARCSTAT_BUMP(arcstat_l2_misses); 3074 if (HDR_L2_WRITING(hdr)) 3075 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3076 spa_config_exit(spa, SCL_L2ARC, vd); 3077 } 3078 } else { 3079 if (vd != NULL) 3080 spa_config_exit(spa, SCL_L2ARC, vd); 3081 if (l2arc_ndev != 0) { 3082 DTRACE_PROBE1(l2arc__miss, 3083 arc_buf_hdr_t *, hdr); 3084 ARCSTAT_BUMP(arcstat_l2_misses); 3085 } 3086 } 3087 3088 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3089 arc_read_done, buf, priority, zio_flags, zb); 3090 3091 if (*arc_flags & ARC_WAIT) 3092 return (zio_wait(rzio)); 3093 3094 ASSERT(*arc_flags & ARC_NOWAIT); 3095 zio_nowait(rzio); 3096 } 3097 return (0); 3098 } 3099 3100 void 3101 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3102 { 3103 ASSERT(buf->b_hdr != NULL); 3104 ASSERT(buf->b_hdr->b_state != arc_anon); 3105 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3106 ASSERT(buf->b_efunc == NULL); 3107 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3108 3109 buf->b_efunc = func; 3110 buf->b_private = private; 3111 } 3112 3113 /* 3114 * This is used by the DMU to let the ARC know that a buffer is 3115 * being evicted, so the ARC should clean up. If this arc buf 3116 * is not yet in the evicted state, it will be put there. 3117 */ 3118 int 3119 arc_buf_evict(arc_buf_t *buf) 3120 { 3121 arc_buf_hdr_t *hdr; 3122 kmutex_t *hash_lock; 3123 arc_buf_t **bufp; 3124 3125 mutex_enter(&buf->b_evict_lock); 3126 hdr = buf->b_hdr; 3127 if (hdr == NULL) { 3128 /* 3129 * We are in arc_do_user_evicts(). 3130 */ 3131 ASSERT(buf->b_data == NULL); 3132 mutex_exit(&buf->b_evict_lock); 3133 return (0); 3134 } else if (buf->b_data == NULL) { 3135 arc_buf_t copy = *buf; /* structure assignment */ 3136 /* 3137 * We are on the eviction list; process this buffer now 3138 * but let arc_do_user_evicts() do the reaping. 3139 */ 3140 buf->b_efunc = NULL; 3141 mutex_exit(&buf->b_evict_lock); 3142 VERIFY(copy.b_efunc(©) == 0); 3143 return (1); 3144 } 3145 hash_lock = HDR_LOCK(hdr); 3146 mutex_enter(hash_lock); 3147 hdr = buf->b_hdr; 3148 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3149 3150 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3151 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3152 3153 /* 3154 * Pull this buffer off of the hdr 3155 */ 3156 bufp = &hdr->b_buf; 3157 while (*bufp != buf) 3158 bufp = &(*bufp)->b_next; 3159 *bufp = buf->b_next; 3160 3161 ASSERT(buf->b_data != NULL); 3162 arc_buf_destroy(buf, FALSE, FALSE); 3163 3164 if (hdr->b_datacnt == 0) { 3165 arc_state_t *old_state = hdr->b_state; 3166 arc_state_t *evicted_state; 3167 3168 ASSERT(hdr->b_buf == NULL); 3169 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 3170 3171 evicted_state = 3172 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3173 3174 mutex_enter(&old_state->arcs_mtx); 3175 mutex_enter(&evicted_state->arcs_mtx); 3176 3177 arc_change_state(evicted_state, hdr, hash_lock); 3178 ASSERT(HDR_IN_HASH_TABLE(hdr)); 3179 hdr->b_flags |= ARC_IN_HASH_TABLE; 3180 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3181 3182 mutex_exit(&evicted_state->arcs_mtx); 3183 mutex_exit(&old_state->arcs_mtx); 3184 } 3185 mutex_exit(hash_lock); 3186 mutex_exit(&buf->b_evict_lock); 3187 3188 VERIFY(buf->b_efunc(buf) == 0); 3189 buf->b_efunc = NULL; 3190 buf->b_private = NULL; 3191 buf->b_hdr = NULL; 3192 buf->b_next = NULL; 3193 kmem_cache_free(buf_cache, buf); 3194 return (1); 3195 } 3196 3197 /* 3198 * Release this buffer from the cache. This must be done 3199 * after a read and prior to modifying the buffer contents. 3200 * If the buffer has more than one reference, we must make 3201 * a new hdr for the buffer. 3202 */ 3203 void 3204 arc_release(arc_buf_t *buf, void *tag) 3205 { 3206 arc_buf_hdr_t *hdr; 3207 kmutex_t *hash_lock = NULL; 3208 l2arc_buf_hdr_t *l2hdr; 3209 uint64_t buf_size; 3210 3211 /* 3212 * It would be nice to assert that if it's DMU metadata (level > 3213 * 0 || it's the dnode file), then it must be syncing context. 3214 * But we don't know that information at this level. 3215 */ 3216 3217 mutex_enter(&buf->b_evict_lock); 3218 hdr = buf->b_hdr; 3219 3220 /* this buffer is not on any list */ 3221 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3222 3223 if (hdr->b_state == arc_anon) { 3224 /* this buffer is already released */ 3225 ASSERT(buf->b_efunc == NULL); 3226 } else { 3227 hash_lock = HDR_LOCK(hdr); 3228 mutex_enter(hash_lock); 3229 hdr = buf->b_hdr; 3230 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3231 } 3232 3233 l2hdr = hdr->b_l2hdr; 3234 if (l2hdr) { 3235 mutex_enter(&l2arc_buflist_mtx); 3236 hdr->b_l2hdr = NULL; 3237 } 3238 buf_size = hdr->b_size; 3239 3240 /* 3241 * Do we have more than one buf? 3242 */ 3243 if (hdr->b_datacnt > 1) { 3244 arc_buf_hdr_t *nhdr; 3245 arc_buf_t **bufp; 3246 uint64_t blksz = hdr->b_size; 3247 uint64_t spa = hdr->b_spa; 3248 arc_buf_contents_t type = hdr->b_type; 3249 uint32_t flags = hdr->b_flags; 3250 3251 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3252 /* 3253 * Pull the data off of this hdr and attach it to 3254 * a new anonymous hdr. 3255 */ 3256 (void) remove_reference(hdr, hash_lock, tag); 3257 bufp = &hdr->b_buf; 3258 while (*bufp != buf) 3259 bufp = &(*bufp)->b_next; 3260 *bufp = buf->b_next; 3261 buf->b_next = NULL; 3262 3263 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3264 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3265 if (refcount_is_zero(&hdr->b_refcnt)) { 3266 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3267 ASSERT3U(*size, >=, hdr->b_size); 3268 atomic_add_64(size, -hdr->b_size); 3269 } 3270 3271 /* 3272 * We're releasing a duplicate user data buffer, update 3273 * our statistics accordingly. 3274 */ 3275 if (hdr->b_type == ARC_BUFC_DATA) { 3276 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3277 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3278 -hdr->b_size); 3279 } 3280 hdr->b_datacnt -= 1; 3281 arc_cksum_verify(buf); 3282 arc_buf_unwatch(buf); 3283 3284 mutex_exit(hash_lock); 3285 3286 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3287 nhdr->b_size = blksz; 3288 nhdr->b_spa = spa; 3289 nhdr->b_type = type; 3290 nhdr->b_buf = buf; 3291 nhdr->b_state = arc_anon; 3292 nhdr->b_arc_access = 0; 3293 nhdr->b_flags = flags & ARC_L2_WRITING; 3294 nhdr->b_l2hdr = NULL; 3295 nhdr->b_datacnt = 1; 3296 nhdr->b_freeze_cksum = NULL; 3297 (void) refcount_add(&nhdr->b_refcnt, tag); 3298 buf->b_hdr = nhdr; 3299 mutex_exit(&buf->b_evict_lock); 3300 atomic_add_64(&arc_anon->arcs_size, blksz); 3301 } else { 3302 mutex_exit(&buf->b_evict_lock); 3303 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3304 ASSERT(!list_link_active(&hdr->b_arc_node)); 3305 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3306 if (hdr->b_state != arc_anon) 3307 arc_change_state(arc_anon, hdr, hash_lock); 3308 hdr->b_arc_access = 0; 3309 if (hash_lock) 3310 mutex_exit(hash_lock); 3311 3312 buf_discard_identity(hdr); 3313 arc_buf_thaw(buf); 3314 } 3315 buf->b_efunc = NULL; 3316 buf->b_private = NULL; 3317 3318 if (l2hdr) { 3319 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3320 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3321 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3322 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3323 mutex_exit(&l2arc_buflist_mtx); 3324 } 3325 } 3326 3327 int 3328 arc_released(arc_buf_t *buf) 3329 { 3330 int released; 3331 3332 mutex_enter(&buf->b_evict_lock); 3333 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3334 mutex_exit(&buf->b_evict_lock); 3335 return (released); 3336 } 3337 3338 int 3339 arc_has_callback(arc_buf_t *buf) 3340 { 3341 int callback; 3342 3343 mutex_enter(&buf->b_evict_lock); 3344 callback = (buf->b_efunc != NULL); 3345 mutex_exit(&buf->b_evict_lock); 3346 return (callback); 3347 } 3348 3349 #ifdef ZFS_DEBUG 3350 int 3351 arc_referenced(arc_buf_t *buf) 3352 { 3353 int referenced; 3354 3355 mutex_enter(&buf->b_evict_lock); 3356 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3357 mutex_exit(&buf->b_evict_lock); 3358 return (referenced); 3359 } 3360 #endif 3361 3362 static void 3363 arc_write_ready(zio_t *zio) 3364 { 3365 arc_write_callback_t *callback = zio->io_private; 3366 arc_buf_t *buf = callback->awcb_buf; 3367 arc_buf_hdr_t *hdr = buf->b_hdr; 3368 3369 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3370 callback->awcb_ready(zio, buf, callback->awcb_private); 3371 3372 /* 3373 * If the IO is already in progress, then this is a re-write 3374 * attempt, so we need to thaw and re-compute the cksum. 3375 * It is the responsibility of the callback to handle the 3376 * accounting for any re-write attempt. 3377 */ 3378 if (HDR_IO_IN_PROGRESS(hdr)) { 3379 mutex_enter(&hdr->b_freeze_lock); 3380 if (hdr->b_freeze_cksum != NULL) { 3381 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3382 hdr->b_freeze_cksum = NULL; 3383 } 3384 mutex_exit(&hdr->b_freeze_lock); 3385 } 3386 arc_cksum_compute(buf, B_FALSE); 3387 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3388 } 3389 3390 static void 3391 arc_write_done(zio_t *zio) 3392 { 3393 arc_write_callback_t *callback = zio->io_private; 3394 arc_buf_t *buf = callback->awcb_buf; 3395 arc_buf_hdr_t *hdr = buf->b_hdr; 3396 3397 ASSERT(hdr->b_acb == NULL); 3398 3399 if (zio->io_error == 0) { 3400 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3401 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3402 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3403 } else { 3404 ASSERT(BUF_EMPTY(hdr)); 3405 } 3406 3407 /* 3408 * If the block to be written was all-zero, we may have 3409 * compressed it away. In this case no write was performed 3410 * so there will be no dva/birth/checksum. The buffer must 3411 * therefore remain anonymous (and uncached). 3412 */ 3413 if (!BUF_EMPTY(hdr)) { 3414 arc_buf_hdr_t *exists; 3415 kmutex_t *hash_lock; 3416 3417 ASSERT(zio->io_error == 0); 3418 3419 arc_cksum_verify(buf); 3420 3421 exists = buf_hash_insert(hdr, &hash_lock); 3422 if (exists) { 3423 /* 3424 * This can only happen if we overwrite for 3425 * sync-to-convergence, because we remove 3426 * buffers from the hash table when we arc_free(). 3427 */ 3428 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3429 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3430 panic("bad overwrite, hdr=%p exists=%p", 3431 (void *)hdr, (void *)exists); 3432 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3433 arc_change_state(arc_anon, exists, hash_lock); 3434 mutex_exit(hash_lock); 3435 arc_hdr_destroy(exists); 3436 exists = buf_hash_insert(hdr, &hash_lock); 3437 ASSERT3P(exists, ==, NULL); 3438 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3439 /* nopwrite */ 3440 ASSERT(zio->io_prop.zp_nopwrite); 3441 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3442 panic("bad nopwrite, hdr=%p exists=%p", 3443 (void *)hdr, (void *)exists); 3444 } else { 3445 /* Dedup */ 3446 ASSERT(hdr->b_datacnt == 1); 3447 ASSERT(hdr->b_state == arc_anon); 3448 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3449 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3450 } 3451 } 3452 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3453 /* if it's not anon, we are doing a scrub */ 3454 if (!exists && hdr->b_state == arc_anon) 3455 arc_access(hdr, hash_lock); 3456 mutex_exit(hash_lock); 3457 } else { 3458 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3459 } 3460 3461 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3462 callback->awcb_done(zio, buf, callback->awcb_private); 3463 3464 kmem_free(callback, sizeof (arc_write_callback_t)); 3465 } 3466 3467 zio_t * 3468 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3469 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3470 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *done, 3471 void *private, int priority, int zio_flags, const zbookmark_t *zb) 3472 { 3473 arc_buf_hdr_t *hdr = buf->b_hdr; 3474 arc_write_callback_t *callback; 3475 zio_t *zio; 3476 3477 ASSERT(ready != NULL); 3478 ASSERT(done != NULL); 3479 ASSERT(!HDR_IO_ERROR(hdr)); 3480 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3481 ASSERT(hdr->b_acb == NULL); 3482 if (l2arc) 3483 hdr->b_flags |= ARC_L2CACHE; 3484 if (l2arc_compress) 3485 hdr->b_flags |= ARC_L2COMPRESS; 3486 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3487 callback->awcb_ready = ready; 3488 callback->awcb_done = done; 3489 callback->awcb_private = private; 3490 callback->awcb_buf = buf; 3491 3492 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3493 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb); 3494 3495 return (zio); 3496 } 3497 3498 static int 3499 arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg) 3500 { 3501 #ifdef _KERNEL 3502 uint64_t available_memory = ptob(freemem); 3503 static uint64_t page_load = 0; 3504 static uint64_t last_txg = 0; 3505 3506 #if defined(__i386) 3507 available_memory = 3508 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 3509 #endif 3510 if (available_memory >= zfs_write_limit_max) 3511 return (0); 3512 3513 if (txg > last_txg) { 3514 last_txg = txg; 3515 page_load = 0; 3516 } 3517 /* 3518 * If we are in pageout, we know that memory is already tight, 3519 * the arc is already going to be evicting, so we just want to 3520 * continue to let page writes occur as quickly as possible. 3521 */ 3522 if (curproc == proc_pageout) { 3523 if (page_load > MAX(ptob(minfree), available_memory) / 4) 3524 return (SET_ERROR(ERESTART)); 3525 /* Note: reserve is inflated, so we deflate */ 3526 page_load += reserve / 8; 3527 return (0); 3528 } else if (page_load > 0 && arc_reclaim_needed()) { 3529 /* memory is low, delay before restarting */ 3530 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3531 return (SET_ERROR(EAGAIN)); 3532 } 3533 page_load = 0; 3534 3535 if (arc_size > arc_c_min) { 3536 uint64_t evictable_memory = 3537 arc_mru->arcs_lsize[ARC_BUFC_DATA] + 3538 arc_mru->arcs_lsize[ARC_BUFC_METADATA] + 3539 arc_mfu->arcs_lsize[ARC_BUFC_DATA] + 3540 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]; 3541 available_memory += MIN(evictable_memory, arc_size - arc_c_min); 3542 } 3543 3544 if (inflight_data > available_memory / 4) { 3545 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3546 return (SET_ERROR(ERESTART)); 3547 } 3548 #endif 3549 return (0); 3550 } 3551 3552 void 3553 arc_tempreserve_clear(uint64_t reserve) 3554 { 3555 atomic_add_64(&arc_tempreserve, -reserve); 3556 ASSERT((int64_t)arc_tempreserve >= 0); 3557 } 3558 3559 int 3560 arc_tempreserve_space(uint64_t reserve, uint64_t txg) 3561 { 3562 int error; 3563 uint64_t anon_size; 3564 3565 #ifdef ZFS_DEBUG 3566 /* 3567 * Once in a while, fail for no reason. Everything should cope. 3568 */ 3569 if (spa_get_random(10000) == 0) { 3570 dprintf("forcing random failure\n"); 3571 return (SET_ERROR(ERESTART)); 3572 } 3573 #endif 3574 if (reserve > arc_c/4 && !arc_no_grow) 3575 arc_c = MIN(arc_c_max, reserve * 4); 3576 if (reserve > arc_c) 3577 return (SET_ERROR(ENOMEM)); 3578 3579 /* 3580 * Don't count loaned bufs as in flight dirty data to prevent long 3581 * network delays from blocking transactions that are ready to be 3582 * assigned to a txg. 3583 */ 3584 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 3585 3586 /* 3587 * Writes will, almost always, require additional memory allocations 3588 * in order to compress/encrypt/etc the data. We therefor need to 3589 * make sure that there is sufficient available memory for this. 3590 */ 3591 if (error = arc_memory_throttle(reserve, anon_size, txg)) 3592 return (error); 3593 3594 /* 3595 * Throttle writes when the amount of dirty data in the cache 3596 * gets too large. We try to keep the cache less than half full 3597 * of dirty blocks so that our sync times don't grow too large. 3598 * Note: if two requests come in concurrently, we might let them 3599 * both succeed, when one of them should fail. Not a huge deal. 3600 */ 3601 3602 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 3603 anon_size > arc_c / 4) { 3604 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 3605 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 3606 arc_tempreserve>>10, 3607 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 3608 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 3609 reserve>>10, arc_c>>10); 3610 return (SET_ERROR(ERESTART)); 3611 } 3612 atomic_add_64(&arc_tempreserve, reserve); 3613 return (0); 3614 } 3615 3616 void 3617 arc_init(void) 3618 { 3619 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 3620 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 3621 3622 /* Convert seconds to clock ticks */ 3623 arc_min_prefetch_lifespan = 1 * hz; 3624 3625 /* Start out with 1/8 of all memory */ 3626 arc_c = physmem * PAGESIZE / 8; 3627 3628 #ifdef _KERNEL 3629 /* 3630 * On architectures where the physical memory can be larger 3631 * than the addressable space (intel in 32-bit mode), we may 3632 * need to limit the cache to 1/8 of VM size. 3633 */ 3634 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 3635 #endif 3636 3637 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */ 3638 arc_c_min = MAX(arc_c / 4, 64<<20); 3639 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */ 3640 if (arc_c * 8 >= 1<<30) 3641 arc_c_max = (arc_c * 8) - (1<<30); 3642 else 3643 arc_c_max = arc_c_min; 3644 arc_c_max = MAX(arc_c * 6, arc_c_max); 3645 3646 /* 3647 * Allow the tunables to override our calculations if they are 3648 * reasonable (ie. over 64MB) 3649 */ 3650 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE) 3651 arc_c_max = zfs_arc_max; 3652 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max) 3653 arc_c_min = zfs_arc_min; 3654 3655 arc_c = arc_c_max; 3656 arc_p = (arc_c >> 1); 3657 3658 /* limit meta-data to 1/4 of the arc capacity */ 3659 arc_meta_limit = arc_c_max / 4; 3660 3661 /* Allow the tunable to override if it is reasonable */ 3662 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 3663 arc_meta_limit = zfs_arc_meta_limit; 3664 3665 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 3666 arc_c_min = arc_meta_limit / 2; 3667 3668 if (zfs_arc_grow_retry > 0) 3669 arc_grow_retry = zfs_arc_grow_retry; 3670 3671 if (zfs_arc_shrink_shift > 0) 3672 arc_shrink_shift = zfs_arc_shrink_shift; 3673 3674 if (zfs_arc_p_min_shift > 0) 3675 arc_p_min_shift = zfs_arc_p_min_shift; 3676 3677 /* if kmem_flags are set, lets try to use less memory */ 3678 if (kmem_debugging()) 3679 arc_c = arc_c / 2; 3680 if (arc_c < arc_c_min) 3681 arc_c = arc_c_min; 3682 3683 arc_anon = &ARC_anon; 3684 arc_mru = &ARC_mru; 3685 arc_mru_ghost = &ARC_mru_ghost; 3686 arc_mfu = &ARC_mfu; 3687 arc_mfu_ghost = &ARC_mfu_ghost; 3688 arc_l2c_only = &ARC_l2c_only; 3689 arc_size = 0; 3690 3691 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); 3692 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); 3693 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); 3694 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); 3695 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); 3696 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); 3697 3698 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 3699 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3700 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 3701 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3702 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 3703 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3704 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 3705 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3706 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 3707 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3708 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 3709 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3710 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 3711 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3712 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 3713 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3714 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], 3715 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3716 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], 3717 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 3718 3719 buf_init(); 3720 3721 arc_thread_exit = 0; 3722 arc_eviction_list = NULL; 3723 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 3724 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 3725 3726 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 3727 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 3728 3729 if (arc_ksp != NULL) { 3730 arc_ksp->ks_data = &arc_stats; 3731 kstat_install(arc_ksp); 3732 } 3733 3734 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 3735 TS_RUN, minclsyspri); 3736 3737 arc_dead = FALSE; 3738 arc_warm = B_FALSE; 3739 3740 if (zfs_write_limit_max == 0) 3741 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift; 3742 else 3743 zfs_write_limit_shift = 0; 3744 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL); 3745 } 3746 3747 void 3748 arc_fini(void) 3749 { 3750 mutex_enter(&arc_reclaim_thr_lock); 3751 arc_thread_exit = 1; 3752 while (arc_thread_exit != 0) 3753 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 3754 mutex_exit(&arc_reclaim_thr_lock); 3755 3756 arc_flush(NULL); 3757 3758 arc_dead = TRUE; 3759 3760 if (arc_ksp != NULL) { 3761 kstat_delete(arc_ksp); 3762 arc_ksp = NULL; 3763 } 3764 3765 mutex_destroy(&arc_eviction_mtx); 3766 mutex_destroy(&arc_reclaim_thr_lock); 3767 cv_destroy(&arc_reclaim_thr_cv); 3768 3769 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 3770 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 3771 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 3772 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 3773 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 3774 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 3775 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 3776 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 3777 3778 mutex_destroy(&arc_anon->arcs_mtx); 3779 mutex_destroy(&arc_mru->arcs_mtx); 3780 mutex_destroy(&arc_mru_ghost->arcs_mtx); 3781 mutex_destroy(&arc_mfu->arcs_mtx); 3782 mutex_destroy(&arc_mfu_ghost->arcs_mtx); 3783 mutex_destroy(&arc_l2c_only->arcs_mtx); 3784 3785 mutex_destroy(&zfs_write_limit_lock); 3786 3787 buf_fini(); 3788 3789 ASSERT(arc_loaned_bytes == 0); 3790 } 3791 3792 /* 3793 * Level 2 ARC 3794 * 3795 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 3796 * It uses dedicated storage devices to hold cached data, which are populated 3797 * using large infrequent writes. The main role of this cache is to boost 3798 * the performance of random read workloads. The intended L2ARC devices 3799 * include short-stroked disks, solid state disks, and other media with 3800 * substantially faster read latency than disk. 3801 * 3802 * +-----------------------+ 3803 * | ARC | 3804 * +-----------------------+ 3805 * | ^ ^ 3806 * | | | 3807 * l2arc_feed_thread() arc_read() 3808 * | | | 3809 * | l2arc read | 3810 * V | | 3811 * +---------------+ | 3812 * | L2ARC | | 3813 * +---------------+ | 3814 * | ^ | 3815 * l2arc_write() | | 3816 * | | | 3817 * V | | 3818 * +-------+ +-------+ 3819 * | vdev | | vdev | 3820 * | cache | | cache | 3821 * +-------+ +-------+ 3822 * +=========+ .-----. 3823 * : L2ARC : |-_____-| 3824 * : devices : | Disks | 3825 * +=========+ `-_____-' 3826 * 3827 * Read requests are satisfied from the following sources, in order: 3828 * 3829 * 1) ARC 3830 * 2) vdev cache of L2ARC devices 3831 * 3) L2ARC devices 3832 * 4) vdev cache of disks 3833 * 5) disks 3834 * 3835 * Some L2ARC device types exhibit extremely slow write performance. 3836 * To accommodate for this there are some significant differences between 3837 * the L2ARC and traditional cache design: 3838 * 3839 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 3840 * the ARC behave as usual, freeing buffers and placing headers on ghost 3841 * lists. The ARC does not send buffers to the L2ARC during eviction as 3842 * this would add inflated write latencies for all ARC memory pressure. 3843 * 3844 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 3845 * It does this by periodically scanning buffers from the eviction-end of 3846 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 3847 * not already there. It scans until a headroom of buffers is satisfied, 3848 * which itself is a buffer for ARC eviction. If a compressible buffer is 3849 * found during scanning and selected for writing to an L2ARC device, we 3850 * temporarily boost scanning headroom during the next scan cycle to make 3851 * sure we adapt to compression effects (which might significantly reduce 3852 * the data volume we write to L2ARC). The thread that does this is 3853 * l2arc_feed_thread(), illustrated below; example sizes are included to 3854 * provide a better sense of ratio than this diagram: 3855 * 3856 * head --> tail 3857 * +---------------------+----------+ 3858 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 3859 * +---------------------+----------+ | o L2ARC eligible 3860 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 3861 * +---------------------+----------+ | 3862 * 15.9 Gbytes ^ 32 Mbytes | 3863 * headroom | 3864 * l2arc_feed_thread() 3865 * | 3866 * l2arc write hand <--[oooo]--' 3867 * | 8 Mbyte 3868 * | write max 3869 * V 3870 * +==============================+ 3871 * L2ARC dev |####|#|###|###| |####| ... | 3872 * +==============================+ 3873 * 32 Gbytes 3874 * 3875 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 3876 * evicted, then the L2ARC has cached a buffer much sooner than it probably 3877 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 3878 * safe to say that this is an uncommon case, since buffers at the end of 3879 * the ARC lists have moved there due to inactivity. 3880 * 3881 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 3882 * then the L2ARC simply misses copying some buffers. This serves as a 3883 * pressure valve to prevent heavy read workloads from both stalling the ARC 3884 * with waits and clogging the L2ARC with writes. This also helps prevent 3885 * the potential for the L2ARC to churn if it attempts to cache content too 3886 * quickly, such as during backups of the entire pool. 3887 * 3888 * 5. After system boot and before the ARC has filled main memory, there are 3889 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 3890 * lists can remain mostly static. Instead of searching from tail of these 3891 * lists as pictured, the l2arc_feed_thread() will search from the list heads 3892 * for eligible buffers, greatly increasing its chance of finding them. 3893 * 3894 * The L2ARC device write speed is also boosted during this time so that 3895 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 3896 * there are no L2ARC reads, and no fear of degrading read performance 3897 * through increased writes. 3898 * 3899 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 3900 * the vdev queue can aggregate them into larger and fewer writes. Each 3901 * device is written to in a rotor fashion, sweeping writes through 3902 * available space then repeating. 3903 * 3904 * 7. The L2ARC does not store dirty content. It never needs to flush 3905 * write buffers back to disk based storage. 3906 * 3907 * 8. If an ARC buffer is written (and dirtied) which also exists in the 3908 * L2ARC, the now stale L2ARC buffer is immediately dropped. 3909 * 3910 * The performance of the L2ARC can be tweaked by a number of tunables, which 3911 * may be necessary for different workloads: 3912 * 3913 * l2arc_write_max max write bytes per interval 3914 * l2arc_write_boost extra write bytes during device warmup 3915 * l2arc_noprefetch skip caching prefetched buffers 3916 * l2arc_headroom number of max device writes to precache 3917 * l2arc_headroom_boost when we find compressed buffers during ARC 3918 * scanning, we multiply headroom by this 3919 * percentage factor for the next scan cycle, 3920 * since more compressed buffers are likely to 3921 * be present 3922 * l2arc_feed_secs seconds between L2ARC writing 3923 * 3924 * Tunables may be removed or added as future performance improvements are 3925 * integrated, and also may become zpool properties. 3926 * 3927 * There are three key functions that control how the L2ARC warms up: 3928 * 3929 * l2arc_write_eligible() check if a buffer is eligible to cache 3930 * l2arc_write_size() calculate how much to write 3931 * l2arc_write_interval() calculate sleep delay between writes 3932 * 3933 * These three functions determine what to write, how much, and how quickly 3934 * to send writes. 3935 */ 3936 3937 static boolean_t 3938 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 3939 { 3940 /* 3941 * A buffer is *not* eligible for the L2ARC if it: 3942 * 1. belongs to a different spa. 3943 * 2. is already cached on the L2ARC. 3944 * 3. has an I/O in progress (it may be an incomplete read). 3945 * 4. is flagged not eligible (zfs property). 3946 */ 3947 if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL || 3948 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab)) 3949 return (B_FALSE); 3950 3951 return (B_TRUE); 3952 } 3953 3954 static uint64_t 3955 l2arc_write_size(void) 3956 { 3957 uint64_t size; 3958 3959 /* 3960 * Make sure our globals have meaningful values in case the user 3961 * altered them. 3962 */ 3963 size = l2arc_write_max; 3964 if (size == 0) { 3965 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 3966 "be greater than zero, resetting it to the default (%d)", 3967 L2ARC_WRITE_SIZE); 3968 size = l2arc_write_max = L2ARC_WRITE_SIZE; 3969 } 3970 3971 if (arc_warm == B_FALSE) 3972 size += l2arc_write_boost; 3973 3974 return (size); 3975 3976 } 3977 3978 static clock_t 3979 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 3980 { 3981 clock_t interval, next, now; 3982 3983 /* 3984 * If the ARC lists are busy, increase our write rate; if the 3985 * lists are stale, idle back. This is achieved by checking 3986 * how much we previously wrote - if it was more than half of 3987 * what we wanted, schedule the next write much sooner. 3988 */ 3989 if (l2arc_feed_again && wrote > (wanted / 2)) 3990 interval = (hz * l2arc_feed_min_ms) / 1000; 3991 else 3992 interval = hz * l2arc_feed_secs; 3993 3994 now = ddi_get_lbolt(); 3995 next = MAX(now, MIN(now + interval, began + interval)); 3996 3997 return (next); 3998 } 3999 4000 static void 4001 l2arc_hdr_stat_add(void) 4002 { 4003 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4004 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4005 } 4006 4007 static void 4008 l2arc_hdr_stat_remove(void) 4009 { 4010 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4011 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4012 } 4013 4014 /* 4015 * Cycle through L2ARC devices. This is how L2ARC load balances. 4016 * If a device is returned, this also returns holding the spa config lock. 4017 */ 4018 static l2arc_dev_t * 4019 l2arc_dev_get_next(void) 4020 { 4021 l2arc_dev_t *first, *next = NULL; 4022 4023 /* 4024 * Lock out the removal of spas (spa_namespace_lock), then removal 4025 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4026 * both locks will be dropped and a spa config lock held instead. 4027 */ 4028 mutex_enter(&spa_namespace_lock); 4029 mutex_enter(&l2arc_dev_mtx); 4030 4031 /* if there are no vdevs, there is nothing to do */ 4032 if (l2arc_ndev == 0) 4033 goto out; 4034 4035 first = NULL; 4036 next = l2arc_dev_last; 4037 do { 4038 /* loop around the list looking for a non-faulted vdev */ 4039 if (next == NULL) { 4040 next = list_head(l2arc_dev_list); 4041 } else { 4042 next = list_next(l2arc_dev_list, next); 4043 if (next == NULL) 4044 next = list_head(l2arc_dev_list); 4045 } 4046 4047 /* if we have come back to the start, bail out */ 4048 if (first == NULL) 4049 first = next; 4050 else if (next == first) 4051 break; 4052 4053 } while (vdev_is_dead(next->l2ad_vdev)); 4054 4055 /* if we were unable to find any usable vdevs, return NULL */ 4056 if (vdev_is_dead(next->l2ad_vdev)) 4057 next = NULL; 4058 4059 l2arc_dev_last = next; 4060 4061 out: 4062 mutex_exit(&l2arc_dev_mtx); 4063 4064 /* 4065 * Grab the config lock to prevent the 'next' device from being 4066 * removed while we are writing to it. 4067 */ 4068 if (next != NULL) 4069 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4070 mutex_exit(&spa_namespace_lock); 4071 4072 return (next); 4073 } 4074 4075 /* 4076 * Free buffers that were tagged for destruction. 4077 */ 4078 static void 4079 l2arc_do_free_on_write() 4080 { 4081 list_t *buflist; 4082 l2arc_data_free_t *df, *df_prev; 4083 4084 mutex_enter(&l2arc_free_on_write_mtx); 4085 buflist = l2arc_free_on_write; 4086 4087 for (df = list_tail(buflist); df; df = df_prev) { 4088 df_prev = list_prev(buflist, df); 4089 ASSERT(df->l2df_data != NULL); 4090 ASSERT(df->l2df_func != NULL); 4091 df->l2df_func(df->l2df_data, df->l2df_size); 4092 list_remove(buflist, df); 4093 kmem_free(df, sizeof (l2arc_data_free_t)); 4094 } 4095 4096 mutex_exit(&l2arc_free_on_write_mtx); 4097 } 4098 4099 /* 4100 * A write to a cache device has completed. Update all headers to allow 4101 * reads from these buffers to begin. 4102 */ 4103 static void 4104 l2arc_write_done(zio_t *zio) 4105 { 4106 l2arc_write_callback_t *cb; 4107 l2arc_dev_t *dev; 4108 list_t *buflist; 4109 arc_buf_hdr_t *head, *ab, *ab_prev; 4110 l2arc_buf_hdr_t *abl2; 4111 kmutex_t *hash_lock; 4112 4113 cb = zio->io_private; 4114 ASSERT(cb != NULL); 4115 dev = cb->l2wcb_dev; 4116 ASSERT(dev != NULL); 4117 head = cb->l2wcb_head; 4118 ASSERT(head != NULL); 4119 buflist = dev->l2ad_buflist; 4120 ASSERT(buflist != NULL); 4121 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4122 l2arc_write_callback_t *, cb); 4123 4124 if (zio->io_error != 0) 4125 ARCSTAT_BUMP(arcstat_l2_writes_error); 4126 4127 mutex_enter(&l2arc_buflist_mtx); 4128 4129 /* 4130 * All writes completed, or an error was hit. 4131 */ 4132 for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4133 ab_prev = list_prev(buflist, ab); 4134 4135 hash_lock = HDR_LOCK(ab); 4136 if (!mutex_tryenter(hash_lock)) { 4137 /* 4138 * This buffer misses out. It may be in a stage 4139 * of eviction. Its ARC_L2_WRITING flag will be 4140 * left set, denying reads to this buffer. 4141 */ 4142 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4143 continue; 4144 } 4145 4146 abl2 = ab->b_l2hdr; 4147 4148 /* 4149 * Release the temporary compressed buffer as soon as possible. 4150 */ 4151 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4152 l2arc_release_cdata_buf(ab); 4153 4154 if (zio->io_error != 0) { 4155 /* 4156 * Error - drop L2ARC entry. 4157 */ 4158 list_remove(buflist, ab); 4159 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4160 ab->b_l2hdr = NULL; 4161 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4162 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4163 } 4164 4165 /* 4166 * Allow ARC to begin reads to this L2ARC entry. 4167 */ 4168 ab->b_flags &= ~ARC_L2_WRITING; 4169 4170 mutex_exit(hash_lock); 4171 } 4172 4173 atomic_inc_64(&l2arc_writes_done); 4174 list_remove(buflist, head); 4175 kmem_cache_free(hdr_cache, head); 4176 mutex_exit(&l2arc_buflist_mtx); 4177 4178 l2arc_do_free_on_write(); 4179 4180 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4181 } 4182 4183 /* 4184 * A read to a cache device completed. Validate buffer contents before 4185 * handing over to the regular ARC routines. 4186 */ 4187 static void 4188 l2arc_read_done(zio_t *zio) 4189 { 4190 l2arc_read_callback_t *cb; 4191 arc_buf_hdr_t *hdr; 4192 arc_buf_t *buf; 4193 kmutex_t *hash_lock; 4194 int equal; 4195 4196 ASSERT(zio->io_vd != NULL); 4197 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4198 4199 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4200 4201 cb = zio->io_private; 4202 ASSERT(cb != NULL); 4203 buf = cb->l2rcb_buf; 4204 ASSERT(buf != NULL); 4205 4206 hash_lock = HDR_LOCK(buf->b_hdr); 4207 mutex_enter(hash_lock); 4208 hdr = buf->b_hdr; 4209 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4210 4211 /* 4212 * If the buffer was compressed, decompress it first. 4213 */ 4214 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4215 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4216 ASSERT(zio->io_data != NULL); 4217 4218 /* 4219 * Check this survived the L2ARC journey. 4220 */ 4221 equal = arc_cksum_equal(buf); 4222 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4223 mutex_exit(hash_lock); 4224 zio->io_private = buf; 4225 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4226 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4227 arc_read_done(zio); 4228 } else { 4229 mutex_exit(hash_lock); 4230 /* 4231 * Buffer didn't survive caching. Increment stats and 4232 * reissue to the original storage device. 4233 */ 4234 if (zio->io_error != 0) { 4235 ARCSTAT_BUMP(arcstat_l2_io_error); 4236 } else { 4237 zio->io_error = SET_ERROR(EIO); 4238 } 4239 if (!equal) 4240 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4241 4242 /* 4243 * If there's no waiter, issue an async i/o to the primary 4244 * storage now. If there *is* a waiter, the caller must 4245 * issue the i/o in a context where it's OK to block. 4246 */ 4247 if (zio->io_waiter == NULL) { 4248 zio_t *pio = zio_unique_parent(zio); 4249 4250 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4251 4252 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4253 buf->b_data, zio->io_size, arc_read_done, buf, 4254 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4255 } 4256 } 4257 4258 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4259 } 4260 4261 /* 4262 * This is the list priority from which the L2ARC will search for pages to 4263 * cache. This is used within loops (0..3) to cycle through lists in the 4264 * desired order. This order can have a significant effect on cache 4265 * performance. 4266 * 4267 * Currently the metadata lists are hit first, MFU then MRU, followed by 4268 * the data lists. This function returns a locked list, and also returns 4269 * the lock pointer. 4270 */ 4271 static list_t * 4272 l2arc_list_locked(int list_num, kmutex_t **lock) 4273 { 4274 list_t *list = NULL; 4275 4276 ASSERT(list_num >= 0 && list_num <= 3); 4277 4278 switch (list_num) { 4279 case 0: 4280 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 4281 *lock = &arc_mfu->arcs_mtx; 4282 break; 4283 case 1: 4284 list = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 4285 *lock = &arc_mru->arcs_mtx; 4286 break; 4287 case 2: 4288 list = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 4289 *lock = &arc_mfu->arcs_mtx; 4290 break; 4291 case 3: 4292 list = &arc_mru->arcs_list[ARC_BUFC_DATA]; 4293 *lock = &arc_mru->arcs_mtx; 4294 break; 4295 } 4296 4297 ASSERT(!(MUTEX_HELD(*lock))); 4298 mutex_enter(*lock); 4299 return (list); 4300 } 4301 4302 /* 4303 * Evict buffers from the device write hand to the distance specified in 4304 * bytes. This distance may span populated buffers, it may span nothing. 4305 * This is clearing a region on the L2ARC device ready for writing. 4306 * If the 'all' boolean is set, every buffer is evicted. 4307 */ 4308 static void 4309 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4310 { 4311 list_t *buflist; 4312 l2arc_buf_hdr_t *abl2; 4313 arc_buf_hdr_t *ab, *ab_prev; 4314 kmutex_t *hash_lock; 4315 uint64_t taddr; 4316 4317 buflist = dev->l2ad_buflist; 4318 4319 if (buflist == NULL) 4320 return; 4321 4322 if (!all && dev->l2ad_first) { 4323 /* 4324 * This is the first sweep through the device. There is 4325 * nothing to evict. 4326 */ 4327 return; 4328 } 4329 4330 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4331 /* 4332 * When nearing the end of the device, evict to the end 4333 * before the device write hand jumps to the start. 4334 */ 4335 taddr = dev->l2ad_end; 4336 } else { 4337 taddr = dev->l2ad_hand + distance; 4338 } 4339 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4340 uint64_t, taddr, boolean_t, all); 4341 4342 top: 4343 mutex_enter(&l2arc_buflist_mtx); 4344 for (ab = list_tail(buflist); ab; ab = ab_prev) { 4345 ab_prev = list_prev(buflist, ab); 4346 4347 hash_lock = HDR_LOCK(ab); 4348 if (!mutex_tryenter(hash_lock)) { 4349 /* 4350 * Missed the hash lock. Retry. 4351 */ 4352 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4353 mutex_exit(&l2arc_buflist_mtx); 4354 mutex_enter(hash_lock); 4355 mutex_exit(hash_lock); 4356 goto top; 4357 } 4358 4359 if (HDR_L2_WRITE_HEAD(ab)) { 4360 /* 4361 * We hit a write head node. Leave it for 4362 * l2arc_write_done(). 4363 */ 4364 list_remove(buflist, ab); 4365 mutex_exit(hash_lock); 4366 continue; 4367 } 4368 4369 if (!all && ab->b_l2hdr != NULL && 4370 (ab->b_l2hdr->b_daddr > taddr || 4371 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4372 /* 4373 * We've evicted to the target address, 4374 * or the end of the device. 4375 */ 4376 mutex_exit(hash_lock); 4377 break; 4378 } 4379 4380 if (HDR_FREE_IN_PROGRESS(ab)) { 4381 /* 4382 * Already on the path to destruction. 4383 */ 4384 mutex_exit(hash_lock); 4385 continue; 4386 } 4387 4388 if (ab->b_state == arc_l2c_only) { 4389 ASSERT(!HDR_L2_READING(ab)); 4390 /* 4391 * This doesn't exist in the ARC. Destroy. 4392 * arc_hdr_destroy() will call list_remove() 4393 * and decrement arcstat_l2_size. 4394 */ 4395 arc_change_state(arc_anon, ab, hash_lock); 4396 arc_hdr_destroy(ab); 4397 } else { 4398 /* 4399 * Invalidate issued or about to be issued 4400 * reads, since we may be about to write 4401 * over this location. 4402 */ 4403 if (HDR_L2_READING(ab)) { 4404 ARCSTAT_BUMP(arcstat_l2_evict_reading); 4405 ab->b_flags |= ARC_L2_EVICTED; 4406 } 4407 4408 /* 4409 * Tell ARC this no longer exists in L2ARC. 4410 */ 4411 if (ab->b_l2hdr != NULL) { 4412 abl2 = ab->b_l2hdr; 4413 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4414 ab->b_l2hdr = NULL; 4415 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4416 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4417 } 4418 list_remove(buflist, ab); 4419 4420 /* 4421 * This may have been leftover after a 4422 * failed write. 4423 */ 4424 ab->b_flags &= ~ARC_L2_WRITING; 4425 } 4426 mutex_exit(hash_lock); 4427 } 4428 mutex_exit(&l2arc_buflist_mtx); 4429 4430 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0); 4431 dev->l2ad_evict = taddr; 4432 } 4433 4434 /* 4435 * Find and write ARC buffers to the L2ARC device. 4436 * 4437 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4438 * for reading until they have completed writing. 4439 * The headroom_boost is an in-out parameter used to maintain headroom boost 4440 * state between calls to this function. 4441 * 4442 * Returns the number of bytes actually written (which may be smaller than 4443 * the delta by which the device hand has changed due to alignment). 4444 */ 4445 static uint64_t 4446 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 4447 boolean_t *headroom_boost) 4448 { 4449 arc_buf_hdr_t *ab, *ab_prev, *head; 4450 list_t *list; 4451 uint64_t write_asize, write_psize, write_sz, headroom, 4452 buf_compress_minsz; 4453 void *buf_data; 4454 kmutex_t *list_lock; 4455 boolean_t full; 4456 l2arc_write_callback_t *cb; 4457 zio_t *pio, *wzio; 4458 uint64_t guid = spa_load_guid(spa); 4459 const boolean_t do_headroom_boost = *headroom_boost; 4460 4461 ASSERT(dev->l2ad_vdev != NULL); 4462 4463 /* Lower the flag now, we might want to raise it again later. */ 4464 *headroom_boost = B_FALSE; 4465 4466 pio = NULL; 4467 write_sz = write_asize = write_psize = 0; 4468 full = B_FALSE; 4469 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 4470 head->b_flags |= ARC_L2_WRITE_HEAD; 4471 4472 /* 4473 * We will want to try to compress buffers that are at least 2x the 4474 * device sector size. 4475 */ 4476 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 4477 4478 /* 4479 * Copy buffers for L2ARC writing. 4480 */ 4481 mutex_enter(&l2arc_buflist_mtx); 4482 for (int try = 0; try <= 3; try++) { 4483 uint64_t passed_sz = 0; 4484 4485 list = l2arc_list_locked(try, &list_lock); 4486 4487 /* 4488 * L2ARC fast warmup. 4489 * 4490 * Until the ARC is warm and starts to evict, read from the 4491 * head of the ARC lists rather than the tail. 4492 */ 4493 if (arc_warm == B_FALSE) 4494 ab = list_head(list); 4495 else 4496 ab = list_tail(list); 4497 4498 headroom = target_sz * l2arc_headroom; 4499 if (do_headroom_boost) 4500 headroom = (headroom * l2arc_headroom_boost) / 100; 4501 4502 for (; ab; ab = ab_prev) { 4503 l2arc_buf_hdr_t *l2hdr; 4504 kmutex_t *hash_lock; 4505 uint64_t buf_sz; 4506 4507 if (arc_warm == B_FALSE) 4508 ab_prev = list_next(list, ab); 4509 else 4510 ab_prev = list_prev(list, ab); 4511 4512 hash_lock = HDR_LOCK(ab); 4513 if (!mutex_tryenter(hash_lock)) { 4514 /* 4515 * Skip this buffer rather than waiting. 4516 */ 4517 continue; 4518 } 4519 4520 passed_sz += ab->b_size; 4521 if (passed_sz > headroom) { 4522 /* 4523 * Searched too far. 4524 */ 4525 mutex_exit(hash_lock); 4526 break; 4527 } 4528 4529 if (!l2arc_write_eligible(guid, ab)) { 4530 mutex_exit(hash_lock); 4531 continue; 4532 } 4533 4534 if ((write_sz + ab->b_size) > target_sz) { 4535 full = B_TRUE; 4536 mutex_exit(hash_lock); 4537 break; 4538 } 4539 4540 if (pio == NULL) { 4541 /* 4542 * Insert a dummy header on the buflist so 4543 * l2arc_write_done() can find where the 4544 * write buffers begin without searching. 4545 */ 4546 list_insert_head(dev->l2ad_buflist, head); 4547 4548 cb = kmem_alloc( 4549 sizeof (l2arc_write_callback_t), KM_SLEEP); 4550 cb->l2wcb_dev = dev; 4551 cb->l2wcb_head = head; 4552 pio = zio_root(spa, l2arc_write_done, cb, 4553 ZIO_FLAG_CANFAIL); 4554 } 4555 4556 /* 4557 * Create and add a new L2ARC header. 4558 */ 4559 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 4560 l2hdr->b_dev = dev; 4561 ab->b_flags |= ARC_L2_WRITING; 4562 4563 /* 4564 * Temporarily stash the data buffer in b_tmp_cdata. 4565 * The subsequent write step will pick it up from 4566 * there. This is because can't access ab->b_buf 4567 * without holding the hash_lock, which we in turn 4568 * can't access without holding the ARC list locks 4569 * (which we want to avoid during compression/writing). 4570 */ 4571 l2hdr->b_compress = ZIO_COMPRESS_OFF; 4572 l2hdr->b_asize = ab->b_size; 4573 l2hdr->b_tmp_cdata = ab->b_buf->b_data; 4574 4575 buf_sz = ab->b_size; 4576 ab->b_l2hdr = l2hdr; 4577 4578 list_insert_head(dev->l2ad_buflist, ab); 4579 4580 /* 4581 * Compute and store the buffer cksum before 4582 * writing. On debug the cksum is verified first. 4583 */ 4584 arc_cksum_verify(ab->b_buf); 4585 arc_cksum_compute(ab->b_buf, B_TRUE); 4586 4587 mutex_exit(hash_lock); 4588 4589 write_sz += buf_sz; 4590 } 4591 4592 mutex_exit(list_lock); 4593 4594 if (full == B_TRUE) 4595 break; 4596 } 4597 4598 /* No buffers selected for writing? */ 4599 if (pio == NULL) { 4600 ASSERT0(write_sz); 4601 mutex_exit(&l2arc_buflist_mtx); 4602 kmem_cache_free(hdr_cache, head); 4603 return (0); 4604 } 4605 4606 /* 4607 * Now start writing the buffers. We're starting at the write head 4608 * and work backwards, retracing the course of the buffer selector 4609 * loop above. 4610 */ 4611 for (ab = list_prev(dev->l2ad_buflist, head); ab; 4612 ab = list_prev(dev->l2ad_buflist, ab)) { 4613 l2arc_buf_hdr_t *l2hdr; 4614 uint64_t buf_sz; 4615 4616 /* 4617 * We shouldn't need to lock the buffer here, since we flagged 4618 * it as ARC_L2_WRITING in the previous step, but we must take 4619 * care to only access its L2 cache parameters. In particular, 4620 * ab->b_buf may be invalid by now due to ARC eviction. 4621 */ 4622 l2hdr = ab->b_l2hdr; 4623 l2hdr->b_daddr = dev->l2ad_hand; 4624 4625 if ((ab->b_flags & ARC_L2COMPRESS) && 4626 l2hdr->b_asize >= buf_compress_minsz) { 4627 if (l2arc_compress_buf(l2hdr)) { 4628 /* 4629 * If compression succeeded, enable headroom 4630 * boost on the next scan cycle. 4631 */ 4632 *headroom_boost = B_TRUE; 4633 } 4634 } 4635 4636 /* 4637 * Pick up the buffer data we had previously stashed away 4638 * (and now potentially also compressed). 4639 */ 4640 buf_data = l2hdr->b_tmp_cdata; 4641 buf_sz = l2hdr->b_asize; 4642 4643 /* Compression may have squashed the buffer to zero length. */ 4644 if (buf_sz != 0) { 4645 uint64_t buf_p_sz; 4646 4647 wzio = zio_write_phys(pio, dev->l2ad_vdev, 4648 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 4649 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 4650 ZIO_FLAG_CANFAIL, B_FALSE); 4651 4652 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 4653 zio_t *, wzio); 4654 (void) zio_nowait(wzio); 4655 4656 write_asize += buf_sz; 4657 /* 4658 * Keep the clock hand suitably device-aligned. 4659 */ 4660 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 4661 write_psize += buf_p_sz; 4662 dev->l2ad_hand += buf_p_sz; 4663 } 4664 } 4665 4666 mutex_exit(&l2arc_buflist_mtx); 4667 4668 ASSERT3U(write_asize, <=, target_sz); 4669 ARCSTAT_BUMP(arcstat_l2_writes_sent); 4670 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 4671 ARCSTAT_INCR(arcstat_l2_size, write_sz); 4672 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 4673 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 4674 4675 /* 4676 * Bump device hand to the device start if it is approaching the end. 4677 * l2arc_evict() will already have evicted ahead for this case. 4678 */ 4679 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 4680 vdev_space_update(dev->l2ad_vdev, 4681 dev->l2ad_end - dev->l2ad_hand, 0, 0); 4682 dev->l2ad_hand = dev->l2ad_start; 4683 dev->l2ad_evict = dev->l2ad_start; 4684 dev->l2ad_first = B_FALSE; 4685 } 4686 4687 dev->l2ad_writing = B_TRUE; 4688 (void) zio_wait(pio); 4689 dev->l2ad_writing = B_FALSE; 4690 4691 return (write_asize); 4692 } 4693 4694 /* 4695 * Compresses an L2ARC buffer. 4696 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 4697 * size in l2hdr->b_asize. This routine tries to compress the data and 4698 * depending on the compression result there are three possible outcomes: 4699 * *) The buffer was incompressible. The original l2hdr contents were left 4700 * untouched and are ready for writing to an L2 device. 4701 * *) The buffer was all-zeros, so there is no need to write it to an L2 4702 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 4703 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 4704 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 4705 * data buffer which holds the compressed data to be written, and b_asize 4706 * tells us how much data there is. b_compress is set to the appropriate 4707 * compression algorithm. Once writing is done, invoke 4708 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 4709 * 4710 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 4711 * buffer was incompressible). 4712 */ 4713 static boolean_t 4714 l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 4715 { 4716 void *cdata; 4717 size_t csize, len; 4718 4719 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 4720 ASSERT(l2hdr->b_tmp_cdata != NULL); 4721 4722 len = l2hdr->b_asize; 4723 cdata = zio_data_buf_alloc(len); 4724 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 4725 cdata, l2hdr->b_asize); 4726 4727 if (csize == 0) { 4728 /* zero block, indicate that there's nothing to write */ 4729 zio_data_buf_free(cdata, len); 4730 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 4731 l2hdr->b_asize = 0; 4732 l2hdr->b_tmp_cdata = NULL; 4733 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 4734 return (B_TRUE); 4735 } else if (csize > 0 && csize < len) { 4736 /* 4737 * Compression succeeded, we'll keep the cdata around for 4738 * writing and release it afterwards. 4739 */ 4740 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 4741 l2hdr->b_asize = csize; 4742 l2hdr->b_tmp_cdata = cdata; 4743 ARCSTAT_BUMP(arcstat_l2_compress_successes); 4744 return (B_TRUE); 4745 } else { 4746 /* 4747 * Compression failed, release the compressed buffer. 4748 * l2hdr will be left unmodified. 4749 */ 4750 zio_data_buf_free(cdata, len); 4751 ARCSTAT_BUMP(arcstat_l2_compress_failures); 4752 return (B_FALSE); 4753 } 4754 } 4755 4756 /* 4757 * Decompresses a zio read back from an l2arc device. On success, the 4758 * underlying zio's io_data buffer is overwritten by the uncompressed 4759 * version. On decompression error (corrupt compressed stream), the 4760 * zio->io_error value is set to signal an I/O error. 4761 * 4762 * Please note that the compressed data stream is not checksummed, so 4763 * if the underlying device is experiencing data corruption, we may feed 4764 * corrupt data to the decompressor, so the decompressor needs to be 4765 * able to handle this situation (LZ4 does). 4766 */ 4767 static void 4768 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 4769 { 4770 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 4771 4772 if (zio->io_error != 0) { 4773 /* 4774 * An io error has occured, just restore the original io 4775 * size in preparation for a main pool read. 4776 */ 4777 zio->io_orig_size = zio->io_size = hdr->b_size; 4778 return; 4779 } 4780 4781 if (c == ZIO_COMPRESS_EMPTY) { 4782 /* 4783 * An empty buffer results in a null zio, which means we 4784 * need to fill its io_data after we're done restoring the 4785 * buffer's contents. 4786 */ 4787 ASSERT(hdr->b_buf != NULL); 4788 bzero(hdr->b_buf->b_data, hdr->b_size); 4789 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 4790 } else { 4791 ASSERT(zio->io_data != NULL); 4792 /* 4793 * We copy the compressed data from the start of the arc buffer 4794 * (the zio_read will have pulled in only what we need, the 4795 * rest is garbage which we will overwrite at decompression) 4796 * and then decompress back to the ARC data buffer. This way we 4797 * can minimize copying by simply decompressing back over the 4798 * original compressed data (rather than decompressing to an 4799 * aux buffer and then copying back the uncompressed buffer, 4800 * which is likely to be much larger). 4801 */ 4802 uint64_t csize; 4803 void *cdata; 4804 4805 csize = zio->io_size; 4806 cdata = zio_data_buf_alloc(csize); 4807 bcopy(zio->io_data, cdata, csize); 4808 if (zio_decompress_data(c, cdata, zio->io_data, csize, 4809 hdr->b_size) != 0) 4810 zio->io_error = EIO; 4811 zio_data_buf_free(cdata, csize); 4812 } 4813 4814 /* Restore the expected uncompressed IO size. */ 4815 zio->io_orig_size = zio->io_size = hdr->b_size; 4816 } 4817 4818 /* 4819 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 4820 * This buffer serves as a temporary holder of compressed data while 4821 * the buffer entry is being written to an l2arc device. Once that is 4822 * done, we can dispose of it. 4823 */ 4824 static void 4825 l2arc_release_cdata_buf(arc_buf_hdr_t *ab) 4826 { 4827 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 4828 4829 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) { 4830 /* 4831 * If the data was compressed, then we've allocated a 4832 * temporary buffer for it, so now we need to release it. 4833 */ 4834 ASSERT(l2hdr->b_tmp_cdata != NULL); 4835 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 4836 } 4837 l2hdr->b_tmp_cdata = NULL; 4838 } 4839 4840 /* 4841 * This thread feeds the L2ARC at regular intervals. This is the beating 4842 * heart of the L2ARC. 4843 */ 4844 static void 4845 l2arc_feed_thread(void) 4846 { 4847 callb_cpr_t cpr; 4848 l2arc_dev_t *dev; 4849 spa_t *spa; 4850 uint64_t size, wrote; 4851 clock_t begin, next = ddi_get_lbolt(); 4852 boolean_t headroom_boost = B_FALSE; 4853 4854 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 4855 4856 mutex_enter(&l2arc_feed_thr_lock); 4857 4858 while (l2arc_thread_exit == 0) { 4859 CALLB_CPR_SAFE_BEGIN(&cpr); 4860 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 4861 next); 4862 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 4863 next = ddi_get_lbolt() + hz; 4864 4865 /* 4866 * Quick check for L2ARC devices. 4867 */ 4868 mutex_enter(&l2arc_dev_mtx); 4869 if (l2arc_ndev == 0) { 4870 mutex_exit(&l2arc_dev_mtx); 4871 continue; 4872 } 4873 mutex_exit(&l2arc_dev_mtx); 4874 begin = ddi_get_lbolt(); 4875 4876 /* 4877 * This selects the next l2arc device to write to, and in 4878 * doing so the next spa to feed from: dev->l2ad_spa. This 4879 * will return NULL if there are now no l2arc devices or if 4880 * they are all faulted. 4881 * 4882 * If a device is returned, its spa's config lock is also 4883 * held to prevent device removal. l2arc_dev_get_next() 4884 * will grab and release l2arc_dev_mtx. 4885 */ 4886 if ((dev = l2arc_dev_get_next()) == NULL) 4887 continue; 4888 4889 spa = dev->l2ad_spa; 4890 ASSERT(spa != NULL); 4891 4892 /* 4893 * If the pool is read-only then force the feed thread to 4894 * sleep a little longer. 4895 */ 4896 if (!spa_writeable(spa)) { 4897 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 4898 spa_config_exit(spa, SCL_L2ARC, dev); 4899 continue; 4900 } 4901 4902 /* 4903 * Avoid contributing to memory pressure. 4904 */ 4905 if (arc_reclaim_needed()) { 4906 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 4907 spa_config_exit(spa, SCL_L2ARC, dev); 4908 continue; 4909 } 4910 4911 ARCSTAT_BUMP(arcstat_l2_feeds); 4912 4913 size = l2arc_write_size(); 4914 4915 /* 4916 * Evict L2ARC buffers that will be overwritten. 4917 */ 4918 l2arc_evict(dev, size, B_FALSE); 4919 4920 /* 4921 * Write ARC buffers. 4922 */ 4923 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 4924 4925 /* 4926 * Calculate interval between writes. 4927 */ 4928 next = l2arc_write_interval(begin, size, wrote); 4929 spa_config_exit(spa, SCL_L2ARC, dev); 4930 } 4931 4932 l2arc_thread_exit = 0; 4933 cv_broadcast(&l2arc_feed_thr_cv); 4934 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 4935 thread_exit(); 4936 } 4937 4938 boolean_t 4939 l2arc_vdev_present(vdev_t *vd) 4940 { 4941 l2arc_dev_t *dev; 4942 4943 mutex_enter(&l2arc_dev_mtx); 4944 for (dev = list_head(l2arc_dev_list); dev != NULL; 4945 dev = list_next(l2arc_dev_list, dev)) { 4946 if (dev->l2ad_vdev == vd) 4947 break; 4948 } 4949 mutex_exit(&l2arc_dev_mtx); 4950 4951 return (dev != NULL); 4952 } 4953 4954 /* 4955 * Add a vdev for use by the L2ARC. By this point the spa has already 4956 * validated the vdev and opened it. 4957 */ 4958 void 4959 l2arc_add_vdev(spa_t *spa, vdev_t *vd) 4960 { 4961 l2arc_dev_t *adddev; 4962 4963 ASSERT(!l2arc_vdev_present(vd)); 4964 4965 /* 4966 * Create a new l2arc device entry. 4967 */ 4968 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 4969 adddev->l2ad_spa = spa; 4970 adddev->l2ad_vdev = vd; 4971 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 4972 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 4973 adddev->l2ad_hand = adddev->l2ad_start; 4974 adddev->l2ad_evict = adddev->l2ad_start; 4975 adddev->l2ad_first = B_TRUE; 4976 adddev->l2ad_writing = B_FALSE; 4977 4978 /* 4979 * This is a list of all ARC buffers that are still valid on the 4980 * device. 4981 */ 4982 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 4983 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 4984 offsetof(arc_buf_hdr_t, b_l2node)); 4985 4986 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 4987 4988 /* 4989 * Add device to global list 4990 */ 4991 mutex_enter(&l2arc_dev_mtx); 4992 list_insert_head(l2arc_dev_list, adddev); 4993 atomic_inc_64(&l2arc_ndev); 4994 mutex_exit(&l2arc_dev_mtx); 4995 } 4996 4997 /* 4998 * Remove a vdev from the L2ARC. 4999 */ 5000 void 5001 l2arc_remove_vdev(vdev_t *vd) 5002 { 5003 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5004 5005 /* 5006 * Find the device by vdev 5007 */ 5008 mutex_enter(&l2arc_dev_mtx); 5009 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5010 nextdev = list_next(l2arc_dev_list, dev); 5011 if (vd == dev->l2ad_vdev) { 5012 remdev = dev; 5013 break; 5014 } 5015 } 5016 ASSERT(remdev != NULL); 5017 5018 /* 5019 * Remove device from global list 5020 */ 5021 list_remove(l2arc_dev_list, remdev); 5022 l2arc_dev_last = NULL; /* may have been invalidated */ 5023 atomic_dec_64(&l2arc_ndev); 5024 mutex_exit(&l2arc_dev_mtx); 5025 5026 /* 5027 * Clear all buflists and ARC references. L2ARC device flush. 5028 */ 5029 l2arc_evict(remdev, 0, B_TRUE); 5030 list_destroy(remdev->l2ad_buflist); 5031 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5032 kmem_free(remdev, sizeof (l2arc_dev_t)); 5033 } 5034 5035 void 5036 l2arc_init(void) 5037 { 5038 l2arc_thread_exit = 0; 5039 l2arc_ndev = 0; 5040 l2arc_writes_sent = 0; 5041 l2arc_writes_done = 0; 5042 5043 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5044 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5045 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5046 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5047 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5048 5049 l2arc_dev_list = &L2ARC_dev_list; 5050 l2arc_free_on_write = &L2ARC_free_on_write; 5051 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5052 offsetof(l2arc_dev_t, l2ad_node)); 5053 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5054 offsetof(l2arc_data_free_t, l2df_list_node)); 5055 } 5056 5057 void 5058 l2arc_fini(void) 5059 { 5060 /* 5061 * This is called from dmu_fini(), which is called from spa_fini(); 5062 * Because of this, we can assume that all l2arc devices have 5063 * already been removed when the pools themselves were removed. 5064 */ 5065 5066 l2arc_do_free_on_write(); 5067 5068 mutex_destroy(&l2arc_feed_thr_lock); 5069 cv_destroy(&l2arc_feed_thr_cv); 5070 mutex_destroy(&l2arc_dev_mtx); 5071 mutex_destroy(&l2arc_buflist_mtx); 5072 mutex_destroy(&l2arc_free_on_write_mtx); 5073 5074 list_destroy(l2arc_dev_list); 5075 list_destroy(l2arc_free_on_write); 5076 } 5077 5078 void 5079 l2arc_start(void) 5080 { 5081 if (!(spa_mode_global & FWRITE)) 5082 return; 5083 5084 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5085 TS_RUN, minclsyspri); 5086 } 5087 5088 void 5089 l2arc_stop(void) 5090 { 5091 if (!(spa_mode_global & FWRITE)) 5092 return; 5093 5094 mutex_enter(&l2arc_feed_thr_lock); 5095 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5096 l2arc_thread_exit = 1; 5097 while (l2arc_thread_exit != 0) 5098 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5099 mutex_exit(&l2arc_feed_thr_lock); 5100 } 5101