arc.c revision 286776
1157059Ssam/* 2157059Ssam * CDDL HEADER START 3157059Ssam * 4157059Ssam * The contents of this file are subject to the terms of the 5157059Ssam * Common Development and Distribution License (the "License"). 6157059Ssam * You may not use this file except in compliance with the License. 7157059Ssam * 8157059Ssam * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9157059Ssam * or http://www.opensolaris.org/os/licensing. 10157059Ssam * See the License for the specific language governing permissions 11157059Ssam * and limitations under the License. 12157059Ssam * 13157059Ssam * When distributing Covered Code, include this CDDL HEADER in each 14157059Ssam * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15157059Ssam * If applicable, add the following below this CDDL HEADER, with the 16157059Ssam * fields enclosed by brackets "[]" replaced with your own identifying 17157059Ssam * information: Portions Copyright [yyyy] [name of copyright owner] 18157059Ssam * 19157059Ssam * CDDL HEADER END 20157059Ssam */ 21157059Ssam/* 22157059Ssam * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23157059Ssam * Copyright (c) 2012, Joyent, Inc. All rights reserved. 24157059Ssam * Copyright (c) 2011, 2015 by Delphix. All rights reserved. 25157059Ssam * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 26157059Ssam * Copyright 2015 Nexenta Systems, Inc. All rights reserved. 27157059Ssam */ 28157059Ssam 29157059Ssam/* 30157059Ssam * DVA-based Adjustable Replacement Cache 31157059Ssam * 32157059Ssam * While much of the theory of operation used here is 33157059Ssam * based on the self-tuning, low overhead replacement cache 34157059Ssam * presented by Megiddo and Modha at FAST 2003, there are some 35157059Ssam * significant differences: 36157059Ssam * 37157059Ssam * 1. The Megiddo and Modha model assumes any page is evictable. 38157059Ssam * Pages in its cache cannot be "locked" into memory. This makes 39157059Ssam * the eviction algorithm simple: evict the last page in the list. 40157059Ssam * This also make the performance characteristics easy to reason 41157059Ssam * about. Our cache is not so simple. At any given moment, some 42157059Ssam * subset of the blocks in the cache are un-evictable because we 43157059Ssam * have handed out a reference to them. Blocks are only evictable 44157059Ssam * when there are no external references active. This makes 45157059Ssam * eviction far more problematic: we choose to evict the evictable 46157059Ssam * blocks that are the "lowest" in the list. 47157059Ssam * 48157059Ssam * There are times when it is not possible to evict the requested 49157059Ssam * space. In these circumstances we are unable to adjust the cache 50157059Ssam * size. To prevent the cache growing unbounded at these times we 51157059Ssam * implement a "cache throttle" that slows the flow of new data 52157059Ssam * into the cache until we can make space available. 53157059Ssam * 54157059Ssam * 2. The Megiddo and Modha model assumes a fixed cache size. 55157059Ssam * Pages are evicted when the cache is full and there is a cache 56157059Ssam * miss. Our model has a variable sized cache. It grows with 57157059Ssam * high use, but also tries to react to memory pressure from the 58157059Ssam * operating system: decreasing its size when system memory is 59157059Ssam * tight. 60157059Ssam * 61267992Shselasky * 3. The Megiddo and Modha model assumes a fixed page size. All 62267992Shselasky * elements of the cache are therefore exactly the same size. So 63157059Ssam * when adjusting the cache size following a cache miss, its simply 64157059Ssam * a matter of choosing a single page to evict. In our model, we 65157059Ssam * have variable sized cache blocks (rangeing from 512 bytes to 66157059Ssam * 128K bytes). We therefore choose a set of blocks to evict to make 67157059Ssam * space for a cache miss that approximates as closely as possible 68157059Ssam * the space used by the new block. 69157059Ssam * 70157059Ssam * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 71157059Ssam * by N. Megiddo & D. Modha, FAST 2003 72157059Ssam */ 73157059Ssam 74157059Ssam/* 75157059Ssam * The locking model: 76157059Ssam * 77157059Ssam * A new reference to a cache buffer can be obtained in two 78157059Ssam * ways: 1) via a hash table lookup using the DVA as a key, 79283291Sjkim * or 2) via one of the ARC lists. The arc_read() interface 80157059Ssam * uses method 1, while the internal arc algorithms for 81157059Ssam * adjusting the cache use method 2. We therefore provide two 82157059Ssam * types of locks: 1) the hash table lock array, and 2) the 83157059Ssam * arc list locks. 84158960Sphk * 85158960Sphk * Buffers do not have their own mutexes, rather they rely on the 86158960Sphk * hash table mutexes for the bulk of their protection (i.e. most 87158960Sphk * fields in the arc_buf_hdr_t are protected by these mutexes). 88158960Sphk * 89228631Savg * buf_hash_find() returns the appropriate mutex (held) when it 90228631Savg * locates the requested buffer in the hash table. It returns 91228631Savg * NULL for the mutex if the buffer was not in the table. 92228631Savg * 93228631Savg * buf_hash_remove() expects the appropriate hash mutex to be 94228631Savg * already held before it is invoked. 95228631Savg * 96228631Savg * Each arc state also has a mutex which is used to protect the 97228631Savg * buffer list associated with the state. When attempting to 98228631Savg * obtain a hash table lock while holding an arc list lock you 99157059Ssam * must use: mutex_tryenter() to avoid deadlock. Also note that 100158960Sphk * the active state mutex must be held before the ghost state mutex. 101157059Ssam * 102157059Ssam * Arc buffers may have an associated eviction callback function. 103157059Ssam * This function will be invoked prior to removing the buffer (e.g. 104157059Ssam * in arc_do_user_evicts()). Note however that the data associated 105157059Ssam * with the buffer may be evicted prior to the callback. The callback 106157059Ssam * must be made with *no locks held* (to prevent deadlock). Additionally, 107157059Ssam * the users of callbacks must ensure that their private data is 108157059Ssam * protected from simultaneous callbacks from arc_clear_callback() 109157059Ssam * and arc_do_user_evicts(). 110157059Ssam * 111157059Ssam * Note that the majority of the performance stats are manipulated 112157059Ssam * with atomic operations. 113157059Ssam * 114157059Ssam * The L2ARC uses the l2ad_mtx on each vdev for the following: 115157059Ssam * 116157059Ssam * - L2ARC buflist creation 117157059Ssam * - L2ARC buflist eviction 118157059Ssam * - L2ARC write completion, which walks L2ARC buflists 119157059Ssam * - ARC header destruction, as it removes from L2ARC buflists 120157059Ssam * - ARC header release, as it removes from L2ARC buflists 121157059Ssam */ 122157059Ssam 123157059Ssam#include <sys/spa.h> 124157059Ssam#include <sys/zio.h> 125157059Ssam#include <sys/zio_compress.h> 126157059Ssam#include <sys/zfs_context.h> 127157059Ssam#include <sys/arc.h> 128157059Ssam#include <sys/refcount.h> 129157059Ssam#include <sys/vdev.h> 130157059Ssam#include <sys/vdev_impl.h> 131157059Ssam#include <sys/dsl_pool.h> 132157059Ssam#include <sys/multilist.h> 133157059Ssam#ifdef _KERNEL 134157059Ssam#include <sys/dnlc.h> 135157059Ssam#endif 136157059Ssam#include <sys/callb.h> 137157059Ssam#include <sys/kstat.h> 138253604Savg#include <sys/trim_map.h> 139157059Ssam#include <zfs_fletcher.h> 140157059Ssam#include <sys/sdt.h> 141157059Ssam 142157059Ssam#include <vm/vm_pageout.h> 143157059Ssam#include <machine/vmparam.h> 144157059Ssam 145157059Ssam#ifdef illumos 146157059Ssam#ifndef _KERNEL 147157059Ssam/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 148157059Ssamboolean_t arc_watch = B_FALSE; 149157059Ssamint arc_procfd; 150157059Ssam#endif 151157059Ssam#endif /* illumos */ 152157059Ssam 153157059Ssamstatic kmutex_t arc_reclaim_lock; 154157059Ssamstatic kcondvar_t arc_reclaim_thread_cv; 155157059Ssamstatic boolean_t arc_reclaim_thread_exit; 156157059Ssamstatic kcondvar_t arc_reclaim_waiters_cv; 157157059Ssam 158157059Ssamstatic kmutex_t arc_user_evicts_lock; 159157059Ssamstatic kcondvar_t arc_user_evicts_cv; 160157059Ssamstatic boolean_t arc_user_evicts_thread_exit; 161157059Ssam 162157059Ssamuint_t arc_reduce_dnlc_percent = 3; 163157059Ssam 164157059Ssam/* 165157059Ssam * The number of headers to evict in arc_evict_state_impl() before 166157059Ssam * dropping the sublist lock and evicting from another sublist. A lower 167157059Ssam * value means we're more likely to evict the "correct" header (i.e. the 168157059Ssam * oldest header in the arc state), but comes with higher overhead 169157059Ssam * (i.e. more invocations of arc_evict_state_impl()). 170157059Ssam */ 171158960Sphkint zfs_arc_evict_batch_limit = 10; 172157059Ssam 173157059Ssam/* 174157059Ssam * The number of sublists used for each of the arc state lists. If this 175157059Ssam * is not set to a suitable value by the user, it will be configured to 176157059Ssam * the number of CPUs on the system in arc_init(). 177157059Ssam */ 178157059Ssamint zfs_arc_num_sublists_per_state = 0; 179157059Ssam 180157059Ssam/* number of seconds before growing cache again */ 181157059Ssamstatic int arc_grow_retry = 60; 182157059Ssam 183157059Ssam/* shift of arc_c for calculating overflow limit in arc_get_data_buf */ 184157059Ssamint zfs_arc_overflow_shift = 8; 185 186/* shift of arc_c for calculating both min and max arc_p */ 187static int arc_p_min_shift = 4; 188 189/* log2(fraction of arc to reclaim) */ 190static int arc_shrink_shift = 7; 191 192/* 193 * log2(fraction of ARC which must be free to allow growing). 194 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 195 * when reading a new block into the ARC, we will evict an equal-sized block 196 * from the ARC. 197 * 198 * This must be less than arc_shrink_shift, so that when we shrink the ARC, 199 * we will still not allow it to grow. 200 */ 201int arc_no_grow_shift = 5; 202 203 204/* 205 * minimum lifespan of a prefetch block in clock ticks 206 * (initialized in arc_init()) 207 */ 208static int arc_min_prefetch_lifespan; 209 210/* 211 * If this percent of memory is free, don't throttle. 212 */ 213int arc_lotsfree_percent = 10; 214 215static int arc_dead; 216extern int zfs_prefetch_disable; 217 218/* 219 * The arc has filled available memory and has now warmed up. 220 */ 221static boolean_t arc_warm; 222 223/* 224 * These tunables are for performance analysis. 225 */ 226uint64_t zfs_arc_max; 227uint64_t zfs_arc_min; 228uint64_t zfs_arc_meta_limit = 0; 229uint64_t zfs_arc_meta_min = 0; 230int zfs_arc_grow_retry = 0; 231int zfs_arc_shrink_shift = 0; 232int zfs_arc_p_min_shift = 0; 233int zfs_disable_dup_eviction = 0; 234uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 235u_int zfs_arc_free_target = 0; 236 237static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 238static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 239 240#ifdef _KERNEL 241static void 242arc_free_target_init(void *unused __unused) 243{ 244 245 zfs_arc_free_target = vm_pageout_wakeup_thresh; 246} 247SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 248 arc_free_target_init, NULL); 249 250TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 251TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 252TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 253SYSCTL_DECL(_vfs_zfs); 254SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 255 "Maximum ARC size"); 256SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 257 "Minimum ARC size"); 258SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 259 &zfs_arc_average_blocksize, 0, 260 "ARC average blocksize"); 261SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 262 &arc_shrink_shift, 0, 263 "log2(fraction of arc to reclaim)"); 264 265/* 266 * We don't have a tunable for arc_free_target due to the dependency on 267 * pagedaemon initialisation. 268 */ 269SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 270 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 271 sysctl_vfs_zfs_arc_free_target, "IU", 272 "Desired number of free pages below which ARC triggers reclaim"); 273 274static int 275sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 276{ 277 u_int val; 278 int err; 279 280 val = zfs_arc_free_target; 281 err = sysctl_handle_int(oidp, &val, 0, req); 282 if (err != 0 || req->newptr == NULL) 283 return (err); 284 285 if (val < minfree) 286 return (EINVAL); 287 if (val > vm_cnt.v_page_count) 288 return (EINVAL); 289 290 zfs_arc_free_target = val; 291 292 return (0); 293} 294 295/* 296 * Must be declared here, before the definition of corresponding kstat 297 * macro which uses the same names will confuse the compiler. 298 */ 299SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 300 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 301 sysctl_vfs_zfs_arc_meta_limit, "QU", 302 "ARC metadata limit"); 303#endif 304 305/* 306 * Note that buffers can be in one of 6 states: 307 * ARC_anon - anonymous (discussed below) 308 * ARC_mru - recently used, currently cached 309 * ARC_mru_ghost - recentely used, no longer in cache 310 * ARC_mfu - frequently used, currently cached 311 * ARC_mfu_ghost - frequently used, no longer in cache 312 * ARC_l2c_only - exists in L2ARC but not other states 313 * When there are no active references to the buffer, they are 314 * are linked onto a list in one of these arc states. These are 315 * the only buffers that can be evicted or deleted. Within each 316 * state there are multiple lists, one for meta-data and one for 317 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 318 * etc.) is tracked separately so that it can be managed more 319 * explicitly: favored over data, limited explicitly. 320 * 321 * Anonymous buffers are buffers that are not associated with 322 * a DVA. These are buffers that hold dirty block copies 323 * before they are written to stable storage. By definition, 324 * they are "ref'd" and are considered part of arc_mru 325 * that cannot be freed. Generally, they will aquire a DVA 326 * as they are written and migrate onto the arc_mru list. 327 * 328 * The ARC_l2c_only state is for buffers that are in the second 329 * level ARC but no longer in any of the ARC_m* lists. The second 330 * level ARC itself may also contain buffers that are in any of 331 * the ARC_m* states - meaning that a buffer can exist in two 332 * places. The reason for the ARC_l2c_only state is to keep the 333 * buffer header in the hash table, so that reads that hit the 334 * second level ARC benefit from these fast lookups. 335 */ 336 337typedef struct arc_state { 338 /* 339 * list of evictable buffers 340 */ 341 multilist_t arcs_list[ARC_BUFC_NUMTYPES]; 342 /* 343 * total amount of evictable data in this state 344 */ 345 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; 346 /* 347 * total amount of data in this state; this includes: evictable, 348 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 349 */ 350 refcount_t arcs_size; 351} arc_state_t; 352 353/* The 6 states: */ 354static arc_state_t ARC_anon; 355static arc_state_t ARC_mru; 356static arc_state_t ARC_mru_ghost; 357static arc_state_t ARC_mfu; 358static arc_state_t ARC_mfu_ghost; 359static arc_state_t ARC_l2c_only; 360 361typedef struct arc_stats { 362 kstat_named_t arcstat_hits; 363 kstat_named_t arcstat_misses; 364 kstat_named_t arcstat_demand_data_hits; 365 kstat_named_t arcstat_demand_data_misses; 366 kstat_named_t arcstat_demand_metadata_hits; 367 kstat_named_t arcstat_demand_metadata_misses; 368 kstat_named_t arcstat_prefetch_data_hits; 369 kstat_named_t arcstat_prefetch_data_misses; 370 kstat_named_t arcstat_prefetch_metadata_hits; 371 kstat_named_t arcstat_prefetch_metadata_misses; 372 kstat_named_t arcstat_mru_hits; 373 kstat_named_t arcstat_mru_ghost_hits; 374 kstat_named_t arcstat_mfu_hits; 375 kstat_named_t arcstat_mfu_ghost_hits; 376 kstat_named_t arcstat_allocated; 377 kstat_named_t arcstat_deleted; 378 /* 379 * Number of buffers that could not be evicted because the hash lock 380 * was held by another thread. The lock may not necessarily be held 381 * by something using the same buffer, since hash locks are shared 382 * by multiple buffers. 383 */ 384 kstat_named_t arcstat_mutex_miss; 385 /* 386 * Number of buffers skipped because they have I/O in progress, are 387 * indrect prefetch buffers that have not lived long enough, or are 388 * not from the spa we're trying to evict from. 389 */ 390 kstat_named_t arcstat_evict_skip; 391 /* 392 * Number of times arc_evict_state() was unable to evict enough 393 * buffers to reach it's target amount. 394 */ 395 kstat_named_t arcstat_evict_not_enough; 396 kstat_named_t arcstat_evict_l2_cached; 397 kstat_named_t arcstat_evict_l2_eligible; 398 kstat_named_t arcstat_evict_l2_ineligible; 399 kstat_named_t arcstat_evict_l2_skip; 400 kstat_named_t arcstat_hash_elements; 401 kstat_named_t arcstat_hash_elements_max; 402 kstat_named_t arcstat_hash_collisions; 403 kstat_named_t arcstat_hash_chains; 404 kstat_named_t arcstat_hash_chain_max; 405 kstat_named_t arcstat_p; 406 kstat_named_t arcstat_c; 407 kstat_named_t arcstat_c_min; 408 kstat_named_t arcstat_c_max; 409 kstat_named_t arcstat_size; 410 /* 411 * Number of bytes consumed by internal ARC structures necessary 412 * for tracking purposes; these structures are not actually 413 * backed by ARC buffers. This includes arc_buf_hdr_t structures 414 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 415 * caches), and arc_buf_t structures (allocated via arc_buf_t 416 * cache). 417 */ 418 kstat_named_t arcstat_hdr_size; 419 /* 420 * Number of bytes consumed by ARC buffers of type equal to 421 * ARC_BUFC_DATA. This is generally consumed by buffers backing 422 * on disk user data (e.g. plain file contents). 423 */ 424 kstat_named_t arcstat_data_size; 425 /* 426 * Number of bytes consumed by ARC buffers of type equal to 427 * ARC_BUFC_METADATA. This is generally consumed by buffers 428 * backing on disk data that is used for internal ZFS 429 * structures (e.g. ZAP, dnode, indirect blocks, etc). 430 */ 431 kstat_named_t arcstat_metadata_size; 432 /* 433 * Number of bytes consumed by various buffers and structures 434 * not actually backed with ARC buffers. This includes bonus 435 * buffers (allocated directly via zio_buf_* functions), 436 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 437 * cache), and dnode_t structures (allocated via dnode_t cache). 438 */ 439 kstat_named_t arcstat_other_size; 440 /* 441 * Total number of bytes consumed by ARC buffers residing in the 442 * arc_anon state. This includes *all* buffers in the arc_anon 443 * state; e.g. data, metadata, evictable, and unevictable buffers 444 * are all included in this value. 445 */ 446 kstat_named_t arcstat_anon_size; 447 /* 448 * Number of bytes consumed by ARC buffers that meet the 449 * following criteria: backing buffers of type ARC_BUFC_DATA, 450 * residing in the arc_anon state, and are eligible for eviction 451 * (e.g. have no outstanding holds on the buffer). 452 */ 453 kstat_named_t arcstat_anon_evictable_data; 454 /* 455 * Number of bytes consumed by ARC buffers that meet the 456 * following criteria: backing buffers of type ARC_BUFC_METADATA, 457 * residing in the arc_anon state, and are eligible for eviction 458 * (e.g. have no outstanding holds on the buffer). 459 */ 460 kstat_named_t arcstat_anon_evictable_metadata; 461 /* 462 * Total number of bytes consumed by ARC buffers residing in the 463 * arc_mru state. This includes *all* buffers in the arc_mru 464 * state; e.g. data, metadata, evictable, and unevictable buffers 465 * are all included in this value. 466 */ 467 kstat_named_t arcstat_mru_size; 468 /* 469 * Number of bytes consumed by ARC buffers that meet the 470 * following criteria: backing buffers of type ARC_BUFC_DATA, 471 * residing in the arc_mru state, and are eligible for eviction 472 * (e.g. have no outstanding holds on the buffer). 473 */ 474 kstat_named_t arcstat_mru_evictable_data; 475 /* 476 * Number of bytes consumed by ARC buffers that meet the 477 * following criteria: backing buffers of type ARC_BUFC_METADATA, 478 * residing in the arc_mru state, and are eligible for eviction 479 * (e.g. have no outstanding holds on the buffer). 480 */ 481 kstat_named_t arcstat_mru_evictable_metadata; 482 /* 483 * Total number of bytes that *would have been* consumed by ARC 484 * buffers in the arc_mru_ghost state. The key thing to note 485 * here, is the fact that this size doesn't actually indicate 486 * RAM consumption. The ghost lists only consist of headers and 487 * don't actually have ARC buffers linked off of these headers. 488 * Thus, *if* the headers had associated ARC buffers, these 489 * buffers *would have* consumed this number of bytes. 490 */ 491 kstat_named_t arcstat_mru_ghost_size; 492 /* 493 * Number of bytes that *would have been* consumed by ARC 494 * buffers that are eligible for eviction, of type 495 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 496 */ 497 kstat_named_t arcstat_mru_ghost_evictable_data; 498 /* 499 * Number of bytes that *would have been* consumed by ARC 500 * buffers that are eligible for eviction, of type 501 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 502 */ 503 kstat_named_t arcstat_mru_ghost_evictable_metadata; 504 /* 505 * Total number of bytes consumed by ARC buffers residing in the 506 * arc_mfu state. This includes *all* buffers in the arc_mfu 507 * state; e.g. data, metadata, evictable, and unevictable buffers 508 * are all included in this value. 509 */ 510 kstat_named_t arcstat_mfu_size; 511 /* 512 * Number of bytes consumed by ARC buffers that are eligible for 513 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 514 * state. 515 */ 516 kstat_named_t arcstat_mfu_evictable_data; 517 /* 518 * Number of bytes consumed by ARC buffers that are eligible for 519 * eviction, of type ARC_BUFC_METADATA, and reside in the 520 * arc_mfu state. 521 */ 522 kstat_named_t arcstat_mfu_evictable_metadata; 523 /* 524 * Total number of bytes that *would have been* consumed by ARC 525 * buffers in the arc_mfu_ghost state. See the comment above 526 * arcstat_mru_ghost_size for more details. 527 */ 528 kstat_named_t arcstat_mfu_ghost_size; 529 /* 530 * Number of bytes that *would have been* consumed by ARC 531 * buffers that are eligible for eviction, of type 532 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 533 */ 534 kstat_named_t arcstat_mfu_ghost_evictable_data; 535 /* 536 * Number of bytes that *would have been* consumed by ARC 537 * buffers that are eligible for eviction, of type 538 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 539 */ 540 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 541 kstat_named_t arcstat_l2_hits; 542 kstat_named_t arcstat_l2_misses; 543 kstat_named_t arcstat_l2_feeds; 544 kstat_named_t arcstat_l2_rw_clash; 545 kstat_named_t arcstat_l2_read_bytes; 546 kstat_named_t arcstat_l2_write_bytes; 547 kstat_named_t arcstat_l2_writes_sent; 548 kstat_named_t arcstat_l2_writes_done; 549 kstat_named_t arcstat_l2_writes_error; 550 kstat_named_t arcstat_l2_writes_lock_retry; 551 kstat_named_t arcstat_l2_evict_lock_retry; 552 kstat_named_t arcstat_l2_evict_reading; 553 kstat_named_t arcstat_l2_evict_l1cached; 554 kstat_named_t arcstat_l2_free_on_write; 555 kstat_named_t arcstat_l2_cdata_free_on_write; 556 kstat_named_t arcstat_l2_abort_lowmem; 557 kstat_named_t arcstat_l2_cksum_bad; 558 kstat_named_t arcstat_l2_io_error; 559 kstat_named_t arcstat_l2_size; 560 kstat_named_t arcstat_l2_asize; 561 kstat_named_t arcstat_l2_hdr_size; 562 kstat_named_t arcstat_l2_compress_successes; 563 kstat_named_t arcstat_l2_compress_zeros; 564 kstat_named_t arcstat_l2_compress_failures; 565 kstat_named_t arcstat_l2_write_trylock_fail; 566 kstat_named_t arcstat_l2_write_passed_headroom; 567 kstat_named_t arcstat_l2_write_spa_mismatch; 568 kstat_named_t arcstat_l2_write_in_l2; 569 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 570 kstat_named_t arcstat_l2_write_not_cacheable; 571 kstat_named_t arcstat_l2_write_full; 572 kstat_named_t arcstat_l2_write_buffer_iter; 573 kstat_named_t arcstat_l2_write_pios; 574 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 575 kstat_named_t arcstat_l2_write_buffer_list_iter; 576 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 577 kstat_named_t arcstat_memory_throttle_count; 578 kstat_named_t arcstat_duplicate_buffers; 579 kstat_named_t arcstat_duplicate_buffers_size; 580 kstat_named_t arcstat_duplicate_reads; 581 kstat_named_t arcstat_meta_used; 582 kstat_named_t arcstat_meta_limit; 583 kstat_named_t arcstat_meta_max; 584 kstat_named_t arcstat_meta_min; 585} arc_stats_t; 586 587static arc_stats_t arc_stats = { 588 { "hits", KSTAT_DATA_UINT64 }, 589 { "misses", KSTAT_DATA_UINT64 }, 590 { "demand_data_hits", KSTAT_DATA_UINT64 }, 591 { "demand_data_misses", KSTAT_DATA_UINT64 }, 592 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 593 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 594 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 595 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 596 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 597 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 598 { "mru_hits", KSTAT_DATA_UINT64 }, 599 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 600 { "mfu_hits", KSTAT_DATA_UINT64 }, 601 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 602 { "allocated", KSTAT_DATA_UINT64 }, 603 { "deleted", KSTAT_DATA_UINT64 }, 604 { "mutex_miss", KSTAT_DATA_UINT64 }, 605 { "evict_skip", KSTAT_DATA_UINT64 }, 606 { "evict_not_enough", KSTAT_DATA_UINT64 }, 607 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 608 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 609 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 610 { "evict_l2_skip", KSTAT_DATA_UINT64 }, 611 { "hash_elements", KSTAT_DATA_UINT64 }, 612 { "hash_elements_max", KSTAT_DATA_UINT64 }, 613 { "hash_collisions", KSTAT_DATA_UINT64 }, 614 { "hash_chains", KSTAT_DATA_UINT64 }, 615 { "hash_chain_max", KSTAT_DATA_UINT64 }, 616 { "p", KSTAT_DATA_UINT64 }, 617 { "c", KSTAT_DATA_UINT64 }, 618 { "c_min", KSTAT_DATA_UINT64 }, 619 { "c_max", KSTAT_DATA_UINT64 }, 620 { "size", KSTAT_DATA_UINT64 }, 621 { "hdr_size", KSTAT_DATA_UINT64 }, 622 { "data_size", KSTAT_DATA_UINT64 }, 623 { "metadata_size", KSTAT_DATA_UINT64 }, 624 { "other_size", KSTAT_DATA_UINT64 }, 625 { "anon_size", KSTAT_DATA_UINT64 }, 626 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 627 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 628 { "mru_size", KSTAT_DATA_UINT64 }, 629 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 630 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 631 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 632 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 633 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 634 { "mfu_size", KSTAT_DATA_UINT64 }, 635 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 636 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 637 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 638 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 639 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 640 { "l2_hits", KSTAT_DATA_UINT64 }, 641 { "l2_misses", KSTAT_DATA_UINT64 }, 642 { "l2_feeds", KSTAT_DATA_UINT64 }, 643 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 644 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 645 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 646 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 647 { "l2_writes_done", KSTAT_DATA_UINT64 }, 648 { "l2_writes_error", KSTAT_DATA_UINT64 }, 649 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 650 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 651 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 652 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 653 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 654 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 655 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 656 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 657 { "l2_io_error", KSTAT_DATA_UINT64 }, 658 { "l2_size", KSTAT_DATA_UINT64 }, 659 { "l2_asize", KSTAT_DATA_UINT64 }, 660 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 661 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 662 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 663 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 664 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 665 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 666 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 667 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 668 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 669 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 670 { "l2_write_full", KSTAT_DATA_UINT64 }, 671 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 672 { "l2_write_pios", KSTAT_DATA_UINT64 }, 673 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 674 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 675 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 676 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 677 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 678 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 679 { "duplicate_reads", KSTAT_DATA_UINT64 }, 680 { "arc_meta_used", KSTAT_DATA_UINT64 }, 681 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 682 { "arc_meta_max", KSTAT_DATA_UINT64 }, 683 { "arc_meta_min", KSTAT_DATA_UINT64 } 684}; 685 686#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 687 688#define ARCSTAT_INCR(stat, val) \ 689 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 690 691#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 692#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 693 694#define ARCSTAT_MAX(stat, val) { \ 695 uint64_t m; \ 696 while ((val) > (m = arc_stats.stat.value.ui64) && \ 697 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 698 continue; \ 699} 700 701#define ARCSTAT_MAXSTAT(stat) \ 702 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 703 704/* 705 * We define a macro to allow ARC hits/misses to be easily broken down by 706 * two separate conditions, giving a total of four different subtypes for 707 * each of hits and misses (so eight statistics total). 708 */ 709#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 710 if (cond1) { \ 711 if (cond2) { \ 712 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 713 } else { \ 714 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 715 } \ 716 } else { \ 717 if (cond2) { \ 718 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 719 } else { \ 720 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 721 } \ 722 } 723 724kstat_t *arc_ksp; 725static arc_state_t *arc_anon; 726static arc_state_t *arc_mru; 727static arc_state_t *arc_mru_ghost; 728static arc_state_t *arc_mfu; 729static arc_state_t *arc_mfu_ghost; 730static arc_state_t *arc_l2c_only; 731 732/* 733 * There are several ARC variables that are critical to export as kstats -- 734 * but we don't want to have to grovel around in the kstat whenever we wish to 735 * manipulate them. For these variables, we therefore define them to be in 736 * terms of the statistic variable. This assures that we are not introducing 737 * the possibility of inconsistency by having shadow copies of the variables, 738 * while still allowing the code to be readable. 739 */ 740#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 741#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 742#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 743#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 744#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 745#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 746#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 747#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 748#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 749 750#define L2ARC_IS_VALID_COMPRESS(_c_) \ 751 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 752 753static int arc_no_grow; /* Don't try to grow cache size */ 754static uint64_t arc_tempreserve; 755static uint64_t arc_loaned_bytes; 756 757typedef struct arc_callback arc_callback_t; 758 759struct arc_callback { 760 void *acb_private; 761 arc_done_func_t *acb_done; 762 arc_buf_t *acb_buf; 763 zio_t *acb_zio_dummy; 764 arc_callback_t *acb_next; 765}; 766 767typedef struct arc_write_callback arc_write_callback_t; 768 769struct arc_write_callback { 770 void *awcb_private; 771 arc_done_func_t *awcb_ready; 772 arc_done_func_t *awcb_physdone; 773 arc_done_func_t *awcb_done; 774 arc_buf_t *awcb_buf; 775}; 776 777/* 778 * ARC buffers are separated into multiple structs as a memory saving measure: 779 * - Common fields struct, always defined, and embedded within it: 780 * - L2-only fields, always allocated but undefined when not in L2ARC 781 * - L1-only fields, only allocated when in L1ARC 782 * 783 * Buffer in L1 Buffer only in L2 784 * +------------------------+ +------------------------+ 785 * | arc_buf_hdr_t | | arc_buf_hdr_t | 786 * | | | | 787 * | | | | 788 * | | | | 789 * +------------------------+ +------------------------+ 790 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 791 * | (undefined if L1-only) | | | 792 * +------------------------+ +------------------------+ 793 * | l1arc_buf_hdr_t | 794 * | | 795 * | | 796 * | | 797 * | | 798 * +------------------------+ 799 * 800 * Because it's possible for the L2ARC to become extremely large, we can wind 801 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 802 * is minimized by only allocating the fields necessary for an L1-cached buffer 803 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 804 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 805 * words in pointers. arc_hdr_realloc() is used to switch a header between 806 * these two allocation states. 807 */ 808typedef struct l1arc_buf_hdr { 809 kmutex_t b_freeze_lock; 810#ifdef ZFS_DEBUG 811 /* 812 * used for debugging wtih kmem_flags - by allocating and freeing 813 * b_thawed when the buffer is thawed, we get a record of the stack 814 * trace that thawed it. 815 */ 816 void *b_thawed; 817#endif 818 819 arc_buf_t *b_buf; 820 uint32_t b_datacnt; 821 /* for waiting on writes to complete */ 822 kcondvar_t b_cv; 823 824 /* protected by arc state mutex */ 825 arc_state_t *b_state; 826 multilist_node_t b_arc_node; 827 828 /* updated atomically */ 829 clock_t b_arc_access; 830 831 /* self protecting */ 832 refcount_t b_refcnt; 833 834 arc_callback_t *b_acb; 835 /* temporary buffer holder for in-flight compressed data */ 836 void *b_tmp_cdata; 837} l1arc_buf_hdr_t; 838 839typedef struct l2arc_dev l2arc_dev_t; 840 841typedef struct l2arc_buf_hdr { 842 /* protected by arc_buf_hdr mutex */ 843 l2arc_dev_t *b_dev; /* L2ARC device */ 844 uint64_t b_daddr; /* disk address, offset byte */ 845 /* real alloc'd buffer size depending on b_compress applied */ 846 int32_t b_asize; 847 848 list_node_t b_l2node; 849} l2arc_buf_hdr_t; 850 851struct arc_buf_hdr { 852 /* protected by hash lock */ 853 dva_t b_dva; 854 uint64_t b_birth; 855 /* 856 * Even though this checksum is only set/verified when a buffer is in 857 * the L1 cache, it needs to be in the set of common fields because it 858 * must be preserved from the time before a buffer is written out to 859 * L2ARC until after it is read back in. 860 */ 861 zio_cksum_t *b_freeze_cksum; 862 863 arc_buf_hdr_t *b_hash_next; 864 arc_flags_t b_flags; 865 866 /* immutable */ 867 int32_t b_size; 868 uint64_t b_spa; 869 870 /* L2ARC fields. Undefined when not in L2ARC. */ 871 l2arc_buf_hdr_t b_l2hdr; 872 /* L1ARC fields. Undefined when in l2arc_only state */ 873 l1arc_buf_hdr_t b_l1hdr; 874}; 875 876#ifdef _KERNEL 877static int 878sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 879{ 880 uint64_t val; 881 int err; 882 883 val = arc_meta_limit; 884 err = sysctl_handle_64(oidp, &val, 0, req); 885 if (err != 0 || req->newptr == NULL) 886 return (err); 887 888 if (val <= 0 || val > arc_c_max) 889 return (EINVAL); 890 891 arc_meta_limit = val; 892 return (0); 893} 894#endif 895 896static arc_buf_t *arc_eviction_list; 897static arc_buf_hdr_t arc_eviction_hdr; 898 899#define GHOST_STATE(state) \ 900 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 901 (state) == arc_l2c_only) 902 903#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 904#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 905#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 906#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 907#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 908#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 909 910#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 911#define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS) 912#define HDR_L2_READING(hdr) \ 913 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 914 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 915#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 916#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 917#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 918 919#define HDR_ISTYPE_METADATA(hdr) \ 920 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 921#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 922 923#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 924#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 925 926/* For storing compression mode in b_flags */ 927#define HDR_COMPRESS_OFFSET 24 928#define HDR_COMPRESS_NBITS 7 929 930#define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET(hdr->b_flags, \ 931 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS)) 932#define HDR_SET_COMPRESS(hdr, cmp) BF32_SET(hdr->b_flags, \ 933 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS, (cmp)) 934 935/* 936 * Other sizes 937 */ 938 939#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 940#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 941 942/* 943 * Hash table routines 944 */ 945 946#define HT_LOCK_PAD CACHE_LINE_SIZE 947 948struct ht_lock { 949 kmutex_t ht_lock; 950#ifdef _KERNEL 951 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 952#endif 953}; 954 955#define BUF_LOCKS 256 956typedef struct buf_hash_table { 957 uint64_t ht_mask; 958 arc_buf_hdr_t **ht_table; 959 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 960} buf_hash_table_t; 961 962static buf_hash_table_t buf_hash_table; 963 964#define BUF_HASH_INDEX(spa, dva, birth) \ 965 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 966#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 967#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 968#define HDR_LOCK(hdr) \ 969 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 970 971uint64_t zfs_crc64_table[256]; 972 973/* 974 * Level 2 ARC 975 */ 976 977#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 978#define L2ARC_HEADROOM 2 /* num of writes */ 979/* 980 * If we discover during ARC scan any buffers to be compressed, we boost 981 * our headroom for the next scanning cycle by this percentage multiple. 982 */ 983#define L2ARC_HEADROOM_BOOST 200 984#define L2ARC_FEED_SECS 1 /* caching interval secs */ 985#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 986 987/* 988 * Used to distinguish headers that are being process by 989 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk 990 * address. This can happen when the header is added to the l2arc's list 991 * of buffers to write in the first stage of l2arc_write_buffers(), but 992 * has not yet been written out which happens in the second stage of 993 * l2arc_write_buffers(). 994 */ 995#define L2ARC_ADDR_UNSET ((uint64_t)(-1)) 996 997#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 998#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 999 1000/* L2ARC Performance Tunables */ 1001uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1002uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1003uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1004uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1005uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1006uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1007boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1008boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1009boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1010 1011SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 1012 &l2arc_write_max, 0, "max write size"); 1013SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 1014 &l2arc_write_boost, 0, "extra write during warmup"); 1015SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 1016 &l2arc_headroom, 0, "number of dev writes"); 1017SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 1018 &l2arc_feed_secs, 0, "interval seconds"); 1019SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 1020 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 1021 1022SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 1023 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 1024SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 1025 &l2arc_feed_again, 0, "turbo warmup"); 1026SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 1027 &l2arc_norw, 0, "no reads during writes"); 1028 1029SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 1030 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state"); 1031SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 1032 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 1033SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 1034 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 1035 1036SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 1037 &ARC_mru.arcs_size.rc_count, 0, "size of mru state"); 1038SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 1039 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 1040SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 1041 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 1042 1043SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 1044 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state"); 1045SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 1046 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1047 "size of metadata in mru ghost state"); 1048SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 1049 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1050 "size of data in mru ghost state"); 1051 1052SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 1053 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state"); 1054SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 1055 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 1056SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 1057 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 1058 1059SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 1060 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state"); 1061SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 1062 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1063 "size of metadata in mfu ghost state"); 1064SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 1065 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1066 "size of data in mfu ghost state"); 1067 1068SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 1069 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state"); 1070 1071/* 1072 * L2ARC Internals 1073 */ 1074struct l2arc_dev { 1075 vdev_t *l2ad_vdev; /* vdev */ 1076 spa_t *l2ad_spa; /* spa */ 1077 uint64_t l2ad_hand; /* next write location */ 1078 uint64_t l2ad_start; /* first addr on device */ 1079 uint64_t l2ad_end; /* last addr on device */ 1080 boolean_t l2ad_first; /* first sweep through */ 1081 boolean_t l2ad_writing; /* currently writing */ 1082 kmutex_t l2ad_mtx; /* lock for buffer list */ 1083 list_t l2ad_buflist; /* buffer list */ 1084 list_node_t l2ad_node; /* device list node */ 1085 refcount_t l2ad_alloc; /* allocated bytes */ 1086}; 1087 1088static list_t L2ARC_dev_list; /* device list */ 1089static list_t *l2arc_dev_list; /* device list pointer */ 1090static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1091static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1092static list_t L2ARC_free_on_write; /* free after write buf list */ 1093static list_t *l2arc_free_on_write; /* free after write list ptr */ 1094static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1095static uint64_t l2arc_ndev; /* number of devices */ 1096 1097typedef struct l2arc_read_callback { 1098 arc_buf_t *l2rcb_buf; /* read buffer */ 1099 spa_t *l2rcb_spa; /* spa */ 1100 blkptr_t l2rcb_bp; /* original blkptr */ 1101 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1102 int l2rcb_flags; /* original flags */ 1103 enum zio_compress l2rcb_compress; /* applied compress */ 1104} l2arc_read_callback_t; 1105 1106typedef struct l2arc_write_callback { 1107 l2arc_dev_t *l2wcb_dev; /* device info */ 1108 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1109} l2arc_write_callback_t; 1110 1111typedef struct l2arc_data_free { 1112 /* protected by l2arc_free_on_write_mtx */ 1113 void *l2df_data; 1114 size_t l2df_size; 1115 void (*l2df_func)(void *, size_t); 1116 list_node_t l2df_list_node; 1117} l2arc_data_free_t; 1118 1119static kmutex_t l2arc_feed_thr_lock; 1120static kcondvar_t l2arc_feed_thr_cv; 1121static uint8_t l2arc_thread_exit; 1122 1123static void arc_get_data_buf(arc_buf_t *); 1124static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1125static boolean_t arc_is_overflowing(); 1126static void arc_buf_watch(arc_buf_t *); 1127 1128static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1129static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1130 1131static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1132static void l2arc_read_done(zio_t *); 1133 1134static boolean_t l2arc_compress_buf(arc_buf_hdr_t *); 1135static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 1136static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 1137 1138static uint64_t 1139buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1140{ 1141 uint8_t *vdva = (uint8_t *)dva; 1142 uint64_t crc = -1ULL; 1143 int i; 1144 1145 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1146 1147 for (i = 0; i < sizeof (dva_t); i++) 1148 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1149 1150 crc ^= (spa>>8) ^ birth; 1151 1152 return (crc); 1153} 1154 1155#define BUF_EMPTY(buf) \ 1156 ((buf)->b_dva.dva_word[0] == 0 && \ 1157 (buf)->b_dva.dva_word[1] == 0) 1158 1159#define BUF_EQUAL(spa, dva, birth, buf) \ 1160 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1161 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1162 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 1163 1164static void 1165buf_discard_identity(arc_buf_hdr_t *hdr) 1166{ 1167 hdr->b_dva.dva_word[0] = 0; 1168 hdr->b_dva.dva_word[1] = 0; 1169 hdr->b_birth = 0; 1170} 1171 1172static arc_buf_hdr_t * 1173buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1174{ 1175 const dva_t *dva = BP_IDENTITY(bp); 1176 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1177 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1178 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1179 arc_buf_hdr_t *hdr; 1180 1181 mutex_enter(hash_lock); 1182 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1183 hdr = hdr->b_hash_next) { 1184 if (BUF_EQUAL(spa, dva, birth, hdr)) { 1185 *lockp = hash_lock; 1186 return (hdr); 1187 } 1188 } 1189 mutex_exit(hash_lock); 1190 *lockp = NULL; 1191 return (NULL); 1192} 1193 1194/* 1195 * Insert an entry into the hash table. If there is already an element 1196 * equal to elem in the hash table, then the already existing element 1197 * will be returned and the new element will not be inserted. 1198 * Otherwise returns NULL. 1199 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1200 */ 1201static arc_buf_hdr_t * 1202buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1203{ 1204 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1205 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1206 arc_buf_hdr_t *fhdr; 1207 uint32_t i; 1208 1209 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1210 ASSERT(hdr->b_birth != 0); 1211 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1212 1213 if (lockp != NULL) { 1214 *lockp = hash_lock; 1215 mutex_enter(hash_lock); 1216 } else { 1217 ASSERT(MUTEX_HELD(hash_lock)); 1218 } 1219 1220 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1221 fhdr = fhdr->b_hash_next, i++) { 1222 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1223 return (fhdr); 1224 } 1225 1226 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1227 buf_hash_table.ht_table[idx] = hdr; 1228 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 1229 1230 /* collect some hash table performance data */ 1231 if (i > 0) { 1232 ARCSTAT_BUMP(arcstat_hash_collisions); 1233 if (i == 1) 1234 ARCSTAT_BUMP(arcstat_hash_chains); 1235 1236 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1237 } 1238 1239 ARCSTAT_BUMP(arcstat_hash_elements); 1240 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1241 1242 return (NULL); 1243} 1244 1245static void 1246buf_hash_remove(arc_buf_hdr_t *hdr) 1247{ 1248 arc_buf_hdr_t *fhdr, **hdrp; 1249 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1250 1251 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1252 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1253 1254 hdrp = &buf_hash_table.ht_table[idx]; 1255 while ((fhdr = *hdrp) != hdr) { 1256 ASSERT(fhdr != NULL); 1257 hdrp = &fhdr->b_hash_next; 1258 } 1259 *hdrp = hdr->b_hash_next; 1260 hdr->b_hash_next = NULL; 1261 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1262 1263 /* collect some hash table performance data */ 1264 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1265 1266 if (buf_hash_table.ht_table[idx] && 1267 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1268 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1269} 1270 1271/* 1272 * Global data structures and functions for the buf kmem cache. 1273 */ 1274static kmem_cache_t *hdr_full_cache; 1275static kmem_cache_t *hdr_l2only_cache; 1276static kmem_cache_t *buf_cache; 1277 1278static void 1279buf_fini(void) 1280{ 1281 int i; 1282 1283 kmem_free(buf_hash_table.ht_table, 1284 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1285 for (i = 0; i < BUF_LOCKS; i++) 1286 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1287 kmem_cache_destroy(hdr_full_cache); 1288 kmem_cache_destroy(hdr_l2only_cache); 1289 kmem_cache_destroy(buf_cache); 1290} 1291 1292/* 1293 * Constructor callback - called when the cache is empty 1294 * and a new buf is requested. 1295 */ 1296/* ARGSUSED */ 1297static int 1298hdr_full_cons(void *vbuf, void *unused, int kmflag) 1299{ 1300 arc_buf_hdr_t *hdr = vbuf; 1301 1302 bzero(hdr, HDR_FULL_SIZE); 1303 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1304 refcount_create(&hdr->b_l1hdr.b_refcnt); 1305 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1306 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1307 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1308 1309 return (0); 1310} 1311 1312/* ARGSUSED */ 1313static int 1314hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1315{ 1316 arc_buf_hdr_t *hdr = vbuf; 1317 1318 bzero(hdr, HDR_L2ONLY_SIZE); 1319 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1320 1321 return (0); 1322} 1323 1324/* ARGSUSED */ 1325static int 1326buf_cons(void *vbuf, void *unused, int kmflag) 1327{ 1328 arc_buf_t *buf = vbuf; 1329 1330 bzero(buf, sizeof (arc_buf_t)); 1331 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1332 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1333 1334 return (0); 1335} 1336 1337/* 1338 * Destructor callback - called when a cached buf is 1339 * no longer required. 1340 */ 1341/* ARGSUSED */ 1342static void 1343hdr_full_dest(void *vbuf, void *unused) 1344{ 1345 arc_buf_hdr_t *hdr = vbuf; 1346 1347 ASSERT(BUF_EMPTY(hdr)); 1348 cv_destroy(&hdr->b_l1hdr.b_cv); 1349 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1350 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1351 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1352 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1353} 1354 1355/* ARGSUSED */ 1356static void 1357hdr_l2only_dest(void *vbuf, void *unused) 1358{ 1359 arc_buf_hdr_t *hdr = vbuf; 1360 1361 ASSERT(BUF_EMPTY(hdr)); 1362 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1363} 1364 1365/* ARGSUSED */ 1366static void 1367buf_dest(void *vbuf, void *unused) 1368{ 1369 arc_buf_t *buf = vbuf; 1370 1371 mutex_destroy(&buf->b_evict_lock); 1372 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1373} 1374 1375/* 1376 * Reclaim callback -- invoked when memory is low. 1377 */ 1378/* ARGSUSED */ 1379static void 1380hdr_recl(void *unused) 1381{ 1382 dprintf("hdr_recl called\n"); 1383 /* 1384 * umem calls the reclaim func when we destroy the buf cache, 1385 * which is after we do arc_fini(). 1386 */ 1387 if (!arc_dead) 1388 cv_signal(&arc_reclaim_thread_cv); 1389} 1390 1391static void 1392buf_init(void) 1393{ 1394 uint64_t *ct; 1395 uint64_t hsize = 1ULL << 12; 1396 int i, j; 1397 1398 /* 1399 * The hash table is big enough to fill all of physical memory 1400 * with an average block size of zfs_arc_average_blocksize (default 8K). 1401 * By default, the table will take up 1402 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1403 */ 1404 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1405 hsize <<= 1; 1406retry: 1407 buf_hash_table.ht_mask = hsize - 1; 1408 buf_hash_table.ht_table = 1409 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1410 if (buf_hash_table.ht_table == NULL) { 1411 ASSERT(hsize > (1ULL << 8)); 1412 hsize >>= 1; 1413 goto retry; 1414 } 1415 1416 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1417 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1418 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1419 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1420 NULL, NULL, 0); 1421 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1422 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1423 1424 for (i = 0; i < 256; i++) 1425 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1426 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1427 1428 for (i = 0; i < BUF_LOCKS; i++) { 1429 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1430 NULL, MUTEX_DEFAULT, NULL); 1431 } 1432} 1433 1434/* 1435 * Transition between the two allocation states for the arc_buf_hdr struct. 1436 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 1437 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 1438 * version is used when a cache buffer is only in the L2ARC in order to reduce 1439 * memory usage. 1440 */ 1441static arc_buf_hdr_t * 1442arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 1443{ 1444 ASSERT(HDR_HAS_L2HDR(hdr)); 1445 1446 arc_buf_hdr_t *nhdr; 1447 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1448 1449 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 1450 (old == hdr_l2only_cache && new == hdr_full_cache)); 1451 1452 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 1453 1454 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 1455 buf_hash_remove(hdr); 1456 1457 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 1458 1459 if (new == hdr_full_cache) { 1460 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1461 /* 1462 * arc_access and arc_change_state need to be aware that a 1463 * header has just come out of L2ARC, so we set its state to 1464 * l2c_only even though it's about to change. 1465 */ 1466 nhdr->b_l1hdr.b_state = arc_l2c_only; 1467 1468 /* Verify previous threads set to NULL before freeing */ 1469 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1470 } else { 1471 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1472 ASSERT0(hdr->b_l1hdr.b_datacnt); 1473 1474 /* 1475 * If we've reached here, We must have been called from 1476 * arc_evict_hdr(), as such we should have already been 1477 * removed from any ghost list we were previously on 1478 * (which protects us from racing with arc_evict_state), 1479 * thus no locking is needed during this check. 1480 */ 1481 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1482 1483 /* 1484 * A buffer must not be moved into the arc_l2c_only 1485 * state if it's not finished being written out to the 1486 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field 1487 * might try to be accessed, even though it was removed. 1488 */ 1489 VERIFY(!HDR_L2_WRITING(hdr)); 1490 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1491 1492 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR; 1493 } 1494 /* 1495 * The header has been reallocated so we need to re-insert it into any 1496 * lists it was on. 1497 */ 1498 (void) buf_hash_insert(nhdr, NULL); 1499 1500 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 1501 1502 mutex_enter(&dev->l2ad_mtx); 1503 1504 /* 1505 * We must place the realloc'ed header back into the list at 1506 * the same spot. Otherwise, if it's placed earlier in the list, 1507 * l2arc_write_buffers() could find it during the function's 1508 * write phase, and try to write it out to the l2arc. 1509 */ 1510 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 1511 list_remove(&dev->l2ad_buflist, hdr); 1512 1513 mutex_exit(&dev->l2ad_mtx); 1514 1515 /* 1516 * Since we're using the pointer address as the tag when 1517 * incrementing and decrementing the l2ad_alloc refcount, we 1518 * must remove the old pointer (that we're about to destroy) and 1519 * add the new pointer to the refcount. Otherwise we'd remove 1520 * the wrong pointer address when calling arc_hdr_destroy() later. 1521 */ 1522 1523 (void) refcount_remove_many(&dev->l2ad_alloc, 1524 hdr->b_l2hdr.b_asize, hdr); 1525 1526 (void) refcount_add_many(&dev->l2ad_alloc, 1527 nhdr->b_l2hdr.b_asize, nhdr); 1528 1529 buf_discard_identity(hdr); 1530 hdr->b_freeze_cksum = NULL; 1531 kmem_cache_free(old, hdr); 1532 1533 return (nhdr); 1534} 1535 1536 1537#define ARC_MINTIME (hz>>4) /* 62 ms */ 1538 1539static void 1540arc_cksum_verify(arc_buf_t *buf) 1541{ 1542 zio_cksum_t zc; 1543 1544 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1545 return; 1546 1547 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1548 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) { 1549 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1550 return; 1551 } 1552 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1553 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1554 panic("buffer modified while frozen!"); 1555 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1556} 1557 1558static int 1559arc_cksum_equal(arc_buf_t *buf) 1560{ 1561 zio_cksum_t zc; 1562 int equal; 1563 1564 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1565 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1566 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1567 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1568 1569 return (equal); 1570} 1571 1572static void 1573arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1574{ 1575 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1576 return; 1577 1578 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1579 if (buf->b_hdr->b_freeze_cksum != NULL) { 1580 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1581 return; 1582 } 1583 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1584 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1585 buf->b_hdr->b_freeze_cksum); 1586 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1587#ifdef illumos 1588 arc_buf_watch(buf); 1589#endif 1590} 1591 1592#ifdef illumos 1593#ifndef _KERNEL 1594typedef struct procctl { 1595 long cmd; 1596 prwatch_t prwatch; 1597} procctl_t; 1598#endif 1599 1600/* ARGSUSED */ 1601static void 1602arc_buf_unwatch(arc_buf_t *buf) 1603{ 1604#ifndef _KERNEL 1605 if (arc_watch) { 1606 int result; 1607 procctl_t ctl; 1608 ctl.cmd = PCWATCH; 1609 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1610 ctl.prwatch.pr_size = 0; 1611 ctl.prwatch.pr_wflags = 0; 1612 result = write(arc_procfd, &ctl, sizeof (ctl)); 1613 ASSERT3U(result, ==, sizeof (ctl)); 1614 } 1615#endif 1616} 1617 1618/* ARGSUSED */ 1619static void 1620arc_buf_watch(arc_buf_t *buf) 1621{ 1622#ifndef _KERNEL 1623 if (arc_watch) { 1624 int result; 1625 procctl_t ctl; 1626 ctl.cmd = PCWATCH; 1627 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1628 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1629 ctl.prwatch.pr_wflags = WA_WRITE; 1630 result = write(arc_procfd, &ctl, sizeof (ctl)); 1631 ASSERT3U(result, ==, sizeof (ctl)); 1632 } 1633#endif 1634} 1635#endif /* illumos */ 1636 1637static arc_buf_contents_t 1638arc_buf_type(arc_buf_hdr_t *hdr) 1639{ 1640 if (HDR_ISTYPE_METADATA(hdr)) { 1641 return (ARC_BUFC_METADATA); 1642 } else { 1643 return (ARC_BUFC_DATA); 1644 } 1645} 1646 1647static uint32_t 1648arc_bufc_to_flags(arc_buf_contents_t type) 1649{ 1650 switch (type) { 1651 case ARC_BUFC_DATA: 1652 /* metadata field is 0 if buffer contains normal data */ 1653 return (0); 1654 case ARC_BUFC_METADATA: 1655 return (ARC_FLAG_BUFC_METADATA); 1656 default: 1657 break; 1658 } 1659 panic("undefined ARC buffer type!"); 1660 return ((uint32_t)-1); 1661} 1662 1663void 1664arc_buf_thaw(arc_buf_t *buf) 1665{ 1666 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1667 if (buf->b_hdr->b_l1hdr.b_state != arc_anon) 1668 panic("modifying non-anon buffer!"); 1669 if (HDR_IO_IN_PROGRESS(buf->b_hdr)) 1670 panic("modifying buffer while i/o in progress!"); 1671 arc_cksum_verify(buf); 1672 } 1673 1674 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1675 if (buf->b_hdr->b_freeze_cksum != NULL) { 1676 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1677 buf->b_hdr->b_freeze_cksum = NULL; 1678 } 1679 1680#ifdef ZFS_DEBUG 1681 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1682 if (buf->b_hdr->b_l1hdr.b_thawed != NULL) 1683 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1); 1684 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1685 } 1686#endif 1687 1688 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1689 1690#ifdef illumos 1691 arc_buf_unwatch(buf); 1692#endif 1693} 1694 1695void 1696arc_buf_freeze(arc_buf_t *buf) 1697{ 1698 kmutex_t *hash_lock; 1699 1700 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1701 return; 1702 1703 hash_lock = HDR_LOCK(buf->b_hdr); 1704 mutex_enter(hash_lock); 1705 1706 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1707 buf->b_hdr->b_l1hdr.b_state == arc_anon); 1708 arc_cksum_compute(buf, B_FALSE); 1709 mutex_exit(hash_lock); 1710 1711} 1712 1713static void 1714add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1715{ 1716 ASSERT(HDR_HAS_L1HDR(hdr)); 1717 ASSERT(MUTEX_HELD(hash_lock)); 1718 arc_state_t *state = hdr->b_l1hdr.b_state; 1719 1720 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 1721 (state != arc_anon)) { 1722 /* We don't use the L2-only state list. */ 1723 if (state != arc_l2c_only) { 1724 arc_buf_contents_t type = arc_buf_type(hdr); 1725 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt; 1726 multilist_t *list = &state->arcs_list[type]; 1727 uint64_t *size = &state->arcs_lsize[type]; 1728 1729 multilist_remove(list, hdr); 1730 1731 if (GHOST_STATE(state)) { 1732 ASSERT0(hdr->b_l1hdr.b_datacnt); 1733 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 1734 delta = hdr->b_size; 1735 } 1736 ASSERT(delta > 0); 1737 ASSERT3U(*size, >=, delta); 1738 atomic_add_64(size, -delta); 1739 } 1740 /* remove the prefetch flag if we get a reference */ 1741 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1742 } 1743} 1744 1745static int 1746remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1747{ 1748 int cnt; 1749 arc_state_t *state = hdr->b_l1hdr.b_state; 1750 1751 ASSERT(HDR_HAS_L1HDR(hdr)); 1752 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1753 ASSERT(!GHOST_STATE(state)); 1754 1755 /* 1756 * arc_l2c_only counts as a ghost state so we don't need to explicitly 1757 * check to prevent usage of the arc_l2c_only list. 1758 */ 1759 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 1760 (state != arc_anon)) { 1761 arc_buf_contents_t type = arc_buf_type(hdr); 1762 multilist_t *list = &state->arcs_list[type]; 1763 uint64_t *size = &state->arcs_lsize[type]; 1764 1765 multilist_insert(list, hdr); 1766 1767 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 1768 atomic_add_64(size, hdr->b_size * 1769 hdr->b_l1hdr.b_datacnt); 1770 } 1771 return (cnt); 1772} 1773 1774/* 1775 * Move the supplied buffer to the indicated state. The hash lock 1776 * for the buffer must be held by the caller. 1777 */ 1778static void 1779arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1780 kmutex_t *hash_lock) 1781{ 1782 arc_state_t *old_state; 1783 int64_t refcnt; 1784 uint32_t datacnt; 1785 uint64_t from_delta, to_delta; 1786 arc_buf_contents_t buftype = arc_buf_type(hdr); 1787 1788 /* 1789 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 1790 * in arc_read() when bringing a buffer out of the L2ARC. However, the 1791 * L1 hdr doesn't always exist when we change state to arc_anon before 1792 * destroying a header, in which case reallocating to add the L1 hdr is 1793 * pointless. 1794 */ 1795 if (HDR_HAS_L1HDR(hdr)) { 1796 old_state = hdr->b_l1hdr.b_state; 1797 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 1798 datacnt = hdr->b_l1hdr.b_datacnt; 1799 } else { 1800 old_state = arc_l2c_only; 1801 refcnt = 0; 1802 datacnt = 0; 1803 } 1804 1805 ASSERT(MUTEX_HELD(hash_lock)); 1806 ASSERT3P(new_state, !=, old_state); 1807 ASSERT(refcnt == 0 || datacnt > 0); 1808 ASSERT(!GHOST_STATE(new_state) || datacnt == 0); 1809 ASSERT(old_state != arc_anon || datacnt <= 1); 1810 1811 from_delta = to_delta = datacnt * hdr->b_size; 1812 1813 /* 1814 * If this buffer is evictable, transfer it from the 1815 * old state list to the new state list. 1816 */ 1817 if (refcnt == 0) { 1818 if (old_state != arc_anon && old_state != arc_l2c_only) { 1819 uint64_t *size = &old_state->arcs_lsize[buftype]; 1820 1821 ASSERT(HDR_HAS_L1HDR(hdr)); 1822 multilist_remove(&old_state->arcs_list[buftype], hdr); 1823 1824 /* 1825 * If prefetching out of the ghost cache, 1826 * we will have a non-zero datacnt. 1827 */ 1828 if (GHOST_STATE(old_state) && datacnt == 0) { 1829 /* ghost elements have a ghost size */ 1830 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1831 from_delta = hdr->b_size; 1832 } 1833 ASSERT3U(*size, >=, from_delta); 1834 atomic_add_64(size, -from_delta); 1835 } 1836 if (new_state != arc_anon && new_state != arc_l2c_only) { 1837 uint64_t *size = &new_state->arcs_lsize[buftype]; 1838 1839 /* 1840 * An L1 header always exists here, since if we're 1841 * moving to some L1-cached state (i.e. not l2c_only or 1842 * anonymous), we realloc the header to add an L1hdr 1843 * beforehand. 1844 */ 1845 ASSERT(HDR_HAS_L1HDR(hdr)); 1846 multilist_insert(&new_state->arcs_list[buftype], hdr); 1847 1848 /* ghost elements have a ghost size */ 1849 if (GHOST_STATE(new_state)) { 1850 ASSERT0(datacnt); 1851 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1852 to_delta = hdr->b_size; 1853 } 1854 atomic_add_64(size, to_delta); 1855 } 1856 } 1857 1858 ASSERT(!BUF_EMPTY(hdr)); 1859 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1860 buf_hash_remove(hdr); 1861 1862 /* adjust state sizes (ignore arc_l2c_only) */ 1863 1864 if (to_delta && new_state != arc_l2c_only) { 1865 ASSERT(HDR_HAS_L1HDR(hdr)); 1866 if (GHOST_STATE(new_state)) { 1867 ASSERT0(datacnt); 1868 1869 /* 1870 * We moving a header to a ghost state, we first 1871 * remove all arc buffers. Thus, we'll have a 1872 * datacnt of zero, and no arc buffer to use for 1873 * the reference. As a result, we use the arc 1874 * header pointer for the reference. 1875 */ 1876 (void) refcount_add_many(&new_state->arcs_size, 1877 hdr->b_size, hdr); 1878 } else { 1879 ASSERT3U(datacnt, !=, 0); 1880 1881 /* 1882 * Each individual buffer holds a unique reference, 1883 * thus we must remove each of these references one 1884 * at a time. 1885 */ 1886 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1887 buf = buf->b_next) { 1888 (void) refcount_add_many(&new_state->arcs_size, 1889 hdr->b_size, buf); 1890 } 1891 } 1892 } 1893 1894 if (from_delta && old_state != arc_l2c_only) { 1895 ASSERT(HDR_HAS_L1HDR(hdr)); 1896 if (GHOST_STATE(old_state)) { 1897 /* 1898 * When moving a header off of a ghost state, 1899 * there's the possibility for datacnt to be 1900 * non-zero. This is because we first add the 1901 * arc buffer to the header prior to changing 1902 * the header's state. Since we used the header 1903 * for the reference when putting the header on 1904 * the ghost state, we must balance that and use 1905 * the header when removing off the ghost state 1906 * (even though datacnt is non zero). 1907 */ 1908 1909 IMPLY(datacnt == 0, new_state == arc_anon || 1910 new_state == arc_l2c_only); 1911 1912 (void) refcount_remove_many(&old_state->arcs_size, 1913 hdr->b_size, hdr); 1914 } else { 1915 ASSERT3P(datacnt, !=, 0); 1916 1917 /* 1918 * Each individual buffer holds a unique reference, 1919 * thus we must remove each of these references one 1920 * at a time. 1921 */ 1922 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1923 buf = buf->b_next) { 1924 (void) refcount_remove_many( 1925 &old_state->arcs_size, hdr->b_size, buf); 1926 } 1927 } 1928 } 1929 1930 if (HDR_HAS_L1HDR(hdr)) 1931 hdr->b_l1hdr.b_state = new_state; 1932 1933 /* 1934 * L2 headers should never be on the L2 state list since they don't 1935 * have L1 headers allocated. 1936 */ 1937 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 1938 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 1939} 1940 1941void 1942arc_space_consume(uint64_t space, arc_space_type_t type) 1943{ 1944 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1945 1946 switch (type) { 1947 case ARC_SPACE_DATA: 1948 ARCSTAT_INCR(arcstat_data_size, space); 1949 break; 1950 case ARC_SPACE_META: 1951 ARCSTAT_INCR(arcstat_metadata_size, space); 1952 break; 1953 case ARC_SPACE_OTHER: 1954 ARCSTAT_INCR(arcstat_other_size, space); 1955 break; 1956 case ARC_SPACE_HDRS: 1957 ARCSTAT_INCR(arcstat_hdr_size, space); 1958 break; 1959 case ARC_SPACE_L2HDRS: 1960 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1961 break; 1962 } 1963 1964 if (type != ARC_SPACE_DATA) 1965 ARCSTAT_INCR(arcstat_meta_used, space); 1966 1967 atomic_add_64(&arc_size, space); 1968} 1969 1970void 1971arc_space_return(uint64_t space, arc_space_type_t type) 1972{ 1973 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1974 1975 switch (type) { 1976 case ARC_SPACE_DATA: 1977 ARCSTAT_INCR(arcstat_data_size, -space); 1978 break; 1979 case ARC_SPACE_META: 1980 ARCSTAT_INCR(arcstat_metadata_size, -space); 1981 break; 1982 case ARC_SPACE_OTHER: 1983 ARCSTAT_INCR(arcstat_other_size, -space); 1984 break; 1985 case ARC_SPACE_HDRS: 1986 ARCSTAT_INCR(arcstat_hdr_size, -space); 1987 break; 1988 case ARC_SPACE_L2HDRS: 1989 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1990 break; 1991 } 1992 1993 if (type != ARC_SPACE_DATA) { 1994 ASSERT(arc_meta_used >= space); 1995 if (arc_meta_max < arc_meta_used) 1996 arc_meta_max = arc_meta_used; 1997 ARCSTAT_INCR(arcstat_meta_used, -space); 1998 } 1999 2000 ASSERT(arc_size >= space); 2001 atomic_add_64(&arc_size, -space); 2002} 2003 2004arc_buf_t * 2005arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 2006{ 2007 arc_buf_hdr_t *hdr; 2008 arc_buf_t *buf; 2009 2010 ASSERT3U(size, >, 0); 2011 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2012 ASSERT(BUF_EMPTY(hdr)); 2013 ASSERT3P(hdr->b_freeze_cksum, ==, NULL); 2014 hdr->b_size = size; 2015 hdr->b_spa = spa_load_guid(spa); 2016 2017 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2018 buf->b_hdr = hdr; 2019 buf->b_data = NULL; 2020 buf->b_efunc = NULL; 2021 buf->b_private = NULL; 2022 buf->b_next = NULL; 2023 2024 hdr->b_flags = arc_bufc_to_flags(type); 2025 hdr->b_flags |= ARC_FLAG_HAS_L1HDR; 2026 2027 hdr->b_l1hdr.b_buf = buf; 2028 hdr->b_l1hdr.b_state = arc_anon; 2029 hdr->b_l1hdr.b_arc_access = 0; 2030 hdr->b_l1hdr.b_datacnt = 1; 2031 hdr->b_l1hdr.b_tmp_cdata = NULL; 2032 2033 arc_get_data_buf(buf); 2034 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2035 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2036 2037 return (buf); 2038} 2039 2040static char *arc_onloan_tag = "onloan"; 2041 2042/* 2043 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2044 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2045 * buffers must be returned to the arc before they can be used by the DMU or 2046 * freed. 2047 */ 2048arc_buf_t * 2049arc_loan_buf(spa_t *spa, int size) 2050{ 2051 arc_buf_t *buf; 2052 2053 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 2054 2055 atomic_add_64(&arc_loaned_bytes, size); 2056 return (buf); 2057} 2058 2059/* 2060 * Return a loaned arc buffer to the arc. 2061 */ 2062void 2063arc_return_buf(arc_buf_t *buf, void *tag) 2064{ 2065 arc_buf_hdr_t *hdr = buf->b_hdr; 2066 2067 ASSERT(buf->b_data != NULL); 2068 ASSERT(HDR_HAS_L1HDR(hdr)); 2069 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2070 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2071 2072 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 2073} 2074 2075/* Detach an arc_buf from a dbuf (tag) */ 2076void 2077arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2078{ 2079 arc_buf_hdr_t *hdr = buf->b_hdr; 2080 2081 ASSERT(buf->b_data != NULL); 2082 ASSERT(HDR_HAS_L1HDR(hdr)); 2083 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2084 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2085 buf->b_efunc = NULL; 2086 buf->b_private = NULL; 2087 2088 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 2089} 2090 2091static arc_buf_t * 2092arc_buf_clone(arc_buf_t *from) 2093{ 2094 arc_buf_t *buf; 2095 arc_buf_hdr_t *hdr = from->b_hdr; 2096 uint64_t size = hdr->b_size; 2097 2098 ASSERT(HDR_HAS_L1HDR(hdr)); 2099 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2100 2101 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2102 buf->b_hdr = hdr; 2103 buf->b_data = NULL; 2104 buf->b_efunc = NULL; 2105 buf->b_private = NULL; 2106 buf->b_next = hdr->b_l1hdr.b_buf; 2107 hdr->b_l1hdr.b_buf = buf; 2108 arc_get_data_buf(buf); 2109 bcopy(from->b_data, buf->b_data, size); 2110 2111 /* 2112 * This buffer already exists in the arc so create a duplicate 2113 * copy for the caller. If the buffer is associated with user data 2114 * then track the size and number of duplicates. These stats will be 2115 * updated as duplicate buffers are created and destroyed. 2116 */ 2117 if (HDR_ISTYPE_DATA(hdr)) { 2118 ARCSTAT_BUMP(arcstat_duplicate_buffers); 2119 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 2120 } 2121 hdr->b_l1hdr.b_datacnt += 1; 2122 return (buf); 2123} 2124 2125void 2126arc_buf_add_ref(arc_buf_t *buf, void* tag) 2127{ 2128 arc_buf_hdr_t *hdr; 2129 kmutex_t *hash_lock; 2130 2131 /* 2132 * Check to see if this buffer is evicted. Callers 2133 * must verify b_data != NULL to know if the add_ref 2134 * was successful. 2135 */ 2136 mutex_enter(&buf->b_evict_lock); 2137 if (buf->b_data == NULL) { 2138 mutex_exit(&buf->b_evict_lock); 2139 return; 2140 } 2141 hash_lock = HDR_LOCK(buf->b_hdr); 2142 mutex_enter(hash_lock); 2143 hdr = buf->b_hdr; 2144 ASSERT(HDR_HAS_L1HDR(hdr)); 2145 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2146 mutex_exit(&buf->b_evict_lock); 2147 2148 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 2149 hdr->b_l1hdr.b_state == arc_mfu); 2150 2151 add_reference(hdr, hash_lock, tag); 2152 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 2153 arc_access(hdr, hash_lock); 2154 mutex_exit(hash_lock); 2155 ARCSTAT_BUMP(arcstat_hits); 2156 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 2157 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 2158 data, metadata, hits); 2159} 2160 2161static void 2162arc_buf_free_on_write(void *data, size_t size, 2163 void (*free_func)(void *, size_t)) 2164{ 2165 l2arc_data_free_t *df; 2166 2167 df = kmem_alloc(sizeof (*df), KM_SLEEP); 2168 df->l2df_data = data; 2169 df->l2df_size = size; 2170 df->l2df_func = free_func; 2171 mutex_enter(&l2arc_free_on_write_mtx); 2172 list_insert_head(l2arc_free_on_write, df); 2173 mutex_exit(&l2arc_free_on_write_mtx); 2174} 2175 2176/* 2177 * Free the arc data buffer. If it is an l2arc write in progress, 2178 * the buffer is placed on l2arc_free_on_write to be freed later. 2179 */ 2180static void 2181arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 2182{ 2183 arc_buf_hdr_t *hdr = buf->b_hdr; 2184 2185 if (HDR_L2_WRITING(hdr)) { 2186 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 2187 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2188 } else { 2189 free_func(buf->b_data, hdr->b_size); 2190 } 2191} 2192 2193static void 2194arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 2195{ 2196 ASSERT(HDR_HAS_L2HDR(hdr)); 2197 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx)); 2198 2199 /* 2200 * The b_tmp_cdata field is linked off of the b_l1hdr, so if 2201 * that doesn't exist, the header is in the arc_l2c_only state, 2202 * and there isn't anything to free (it's already been freed). 2203 */ 2204 if (!HDR_HAS_L1HDR(hdr)) 2205 return; 2206 2207 /* 2208 * The header isn't being written to the l2arc device, thus it 2209 * shouldn't have a b_tmp_cdata to free. 2210 */ 2211 if (!HDR_L2_WRITING(hdr)) { 2212 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2213 return; 2214 } 2215 2216 /* 2217 * The header does not have compression enabled. This can be due 2218 * to the buffer not being compressible, or because we're 2219 * freeing the buffer before the second phase of 2220 * l2arc_write_buffer() has started (which does the compression 2221 * step). In either case, b_tmp_cdata does not point to a 2222 * separately compressed buffer, so there's nothing to free (it 2223 * points to the same buffer as the arc_buf_t's b_data field). 2224 */ 2225 if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) { 2226 hdr->b_l1hdr.b_tmp_cdata = NULL; 2227 return; 2228 } 2229 2230 /* 2231 * There's nothing to free since the buffer was all zero's and 2232 * compressed to a zero length buffer. 2233 */ 2234 if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_EMPTY) { 2235 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2236 return; 2237 } 2238 2239 ASSERT(L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr))); 2240 2241 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, 2242 hdr->b_size, zio_data_buf_free); 2243 2244 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 2245 hdr->b_l1hdr.b_tmp_cdata = NULL; 2246} 2247 2248/* 2249 * Free up buf->b_data and if 'remove' is set, then pull the 2250 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 2251 */ 2252static void 2253arc_buf_destroy(arc_buf_t *buf, boolean_t remove) 2254{ 2255 arc_buf_t **bufp; 2256 2257 /* free up data associated with the buf */ 2258 if (buf->b_data != NULL) { 2259 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 2260 uint64_t size = buf->b_hdr->b_size; 2261 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 2262 2263 arc_cksum_verify(buf); 2264#ifdef illumos 2265 arc_buf_unwatch(buf); 2266#endif 2267 2268 if (type == ARC_BUFC_METADATA) { 2269 arc_buf_data_free(buf, zio_buf_free); 2270 arc_space_return(size, ARC_SPACE_META); 2271 } else { 2272 ASSERT(type == ARC_BUFC_DATA); 2273 arc_buf_data_free(buf, zio_data_buf_free); 2274 arc_space_return(size, ARC_SPACE_DATA); 2275 } 2276 2277 /* protected by hash lock, if in the hash table */ 2278 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) { 2279 uint64_t *cnt = &state->arcs_lsize[type]; 2280 2281 ASSERT(refcount_is_zero( 2282 &buf->b_hdr->b_l1hdr.b_refcnt)); 2283 ASSERT(state != arc_anon && state != arc_l2c_only); 2284 2285 ASSERT3U(*cnt, >=, size); 2286 atomic_add_64(cnt, -size); 2287 } 2288 2289 (void) refcount_remove_many(&state->arcs_size, size, buf); 2290 buf->b_data = NULL; 2291 2292 /* 2293 * If we're destroying a duplicate buffer make sure 2294 * that the appropriate statistics are updated. 2295 */ 2296 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 && 2297 HDR_ISTYPE_DATA(buf->b_hdr)) { 2298 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 2299 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 2300 } 2301 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0); 2302 buf->b_hdr->b_l1hdr.b_datacnt -= 1; 2303 } 2304 2305 /* only remove the buf if requested */ 2306 if (!remove) 2307 return; 2308 2309 /* remove the buf from the hdr list */ 2310 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf; 2311 bufp = &(*bufp)->b_next) 2312 continue; 2313 *bufp = buf->b_next; 2314 buf->b_next = NULL; 2315 2316 ASSERT(buf->b_efunc == NULL); 2317 2318 /* clean up the buf */ 2319 buf->b_hdr = NULL; 2320 kmem_cache_free(buf_cache, buf); 2321} 2322 2323static void 2324arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 2325{ 2326 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2327 l2arc_dev_t *dev = l2hdr->b_dev; 2328 2329 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 2330 ASSERT(HDR_HAS_L2HDR(hdr)); 2331 2332 list_remove(&dev->l2ad_buflist, hdr); 2333 2334 /* 2335 * We don't want to leak the b_tmp_cdata buffer that was 2336 * allocated in l2arc_write_buffers() 2337 */ 2338 arc_buf_l2_cdata_free(hdr); 2339 2340 /* 2341 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then 2342 * this header is being processed by l2arc_write_buffers() (i.e. 2343 * it's in the first stage of l2arc_write_buffers()). 2344 * Re-affirming that truth here, just to serve as a reminder. If 2345 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or 2346 * may not have its HDR_L2_WRITING flag set. (the write may have 2347 * completed, in which case HDR_L2_WRITING will be false and the 2348 * b_daddr field will point to the address of the buffer on disk). 2349 */ 2350 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr)); 2351 2352 /* 2353 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with 2354 * l2arc_write_buffers(). Since we've just removed this header 2355 * from the l2arc buffer list, this header will never reach the 2356 * second stage of l2arc_write_buffers(), which increments the 2357 * accounting stats for this header. Thus, we must be careful 2358 * not to decrement them for this header either. 2359 */ 2360 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) { 2361 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 2362 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 2363 2364 vdev_space_update(dev->l2ad_vdev, 2365 -l2hdr->b_asize, 0, 0); 2366 2367 (void) refcount_remove_many(&dev->l2ad_alloc, 2368 l2hdr->b_asize, hdr); 2369 } 2370 2371 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 2372} 2373 2374static void 2375arc_hdr_destroy(arc_buf_hdr_t *hdr) 2376{ 2377 if (HDR_HAS_L1HDR(hdr)) { 2378 ASSERT(hdr->b_l1hdr.b_buf == NULL || 2379 hdr->b_l1hdr.b_datacnt > 0); 2380 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2381 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2382 } 2383 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2384 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 2385 2386 if (HDR_HAS_L2HDR(hdr)) { 2387 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2388 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 2389 2390 if (!buflist_held) 2391 mutex_enter(&dev->l2ad_mtx); 2392 2393 /* 2394 * Even though we checked this conditional above, we 2395 * need to check this again now that we have the 2396 * l2ad_mtx. This is because we could be racing with 2397 * another thread calling l2arc_evict() which might have 2398 * destroyed this header's L2 portion as we were waiting 2399 * to acquire the l2ad_mtx. If that happens, we don't 2400 * want to re-destroy the header's L2 portion. 2401 */ 2402 if (HDR_HAS_L2HDR(hdr)) { 2403 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET) 2404 trim_map_free(dev->l2ad_vdev, 2405 hdr->b_l2hdr.b_daddr, 2406 hdr->b_l2hdr.b_asize, 0); 2407 arc_hdr_l2hdr_destroy(hdr); 2408 } 2409 2410 if (!buflist_held) 2411 mutex_exit(&dev->l2ad_mtx); 2412 } 2413 2414 if (!BUF_EMPTY(hdr)) 2415 buf_discard_identity(hdr); 2416 2417 if (hdr->b_freeze_cksum != NULL) { 2418 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 2419 hdr->b_freeze_cksum = NULL; 2420 } 2421 2422 if (HDR_HAS_L1HDR(hdr)) { 2423 while (hdr->b_l1hdr.b_buf) { 2424 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2425 2426 if (buf->b_efunc != NULL) { 2427 mutex_enter(&arc_user_evicts_lock); 2428 mutex_enter(&buf->b_evict_lock); 2429 ASSERT(buf->b_hdr != NULL); 2430 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE); 2431 hdr->b_l1hdr.b_buf = buf->b_next; 2432 buf->b_hdr = &arc_eviction_hdr; 2433 buf->b_next = arc_eviction_list; 2434 arc_eviction_list = buf; 2435 mutex_exit(&buf->b_evict_lock); 2436 cv_signal(&arc_user_evicts_cv); 2437 mutex_exit(&arc_user_evicts_lock); 2438 } else { 2439 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE); 2440 } 2441 } 2442#ifdef ZFS_DEBUG 2443 if (hdr->b_l1hdr.b_thawed != NULL) { 2444 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2445 hdr->b_l1hdr.b_thawed = NULL; 2446 } 2447#endif 2448 } 2449 2450 ASSERT3P(hdr->b_hash_next, ==, NULL); 2451 if (HDR_HAS_L1HDR(hdr)) { 2452 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2453 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 2454 kmem_cache_free(hdr_full_cache, hdr); 2455 } else { 2456 kmem_cache_free(hdr_l2only_cache, hdr); 2457 } 2458} 2459 2460void 2461arc_buf_free(arc_buf_t *buf, void *tag) 2462{ 2463 arc_buf_hdr_t *hdr = buf->b_hdr; 2464 int hashed = hdr->b_l1hdr.b_state != arc_anon; 2465 2466 ASSERT(buf->b_efunc == NULL); 2467 ASSERT(buf->b_data != NULL); 2468 2469 if (hashed) { 2470 kmutex_t *hash_lock = HDR_LOCK(hdr); 2471 2472 mutex_enter(hash_lock); 2473 hdr = buf->b_hdr; 2474 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2475 2476 (void) remove_reference(hdr, hash_lock, tag); 2477 if (hdr->b_l1hdr.b_datacnt > 1) { 2478 arc_buf_destroy(buf, TRUE); 2479 } else { 2480 ASSERT(buf == hdr->b_l1hdr.b_buf); 2481 ASSERT(buf->b_efunc == NULL); 2482 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2483 } 2484 mutex_exit(hash_lock); 2485 } else if (HDR_IO_IN_PROGRESS(hdr)) { 2486 int destroy_hdr; 2487 /* 2488 * We are in the middle of an async write. Don't destroy 2489 * this buffer unless the write completes before we finish 2490 * decrementing the reference count. 2491 */ 2492 mutex_enter(&arc_user_evicts_lock); 2493 (void) remove_reference(hdr, NULL, tag); 2494 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2495 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 2496 mutex_exit(&arc_user_evicts_lock); 2497 if (destroy_hdr) 2498 arc_hdr_destroy(hdr); 2499 } else { 2500 if (remove_reference(hdr, NULL, tag) > 0) 2501 arc_buf_destroy(buf, TRUE); 2502 else 2503 arc_hdr_destroy(hdr); 2504 } 2505} 2506 2507boolean_t 2508arc_buf_remove_ref(arc_buf_t *buf, void* tag) 2509{ 2510 arc_buf_hdr_t *hdr = buf->b_hdr; 2511 kmutex_t *hash_lock = HDR_LOCK(hdr); 2512 boolean_t no_callback = (buf->b_efunc == NULL); 2513 2514 if (hdr->b_l1hdr.b_state == arc_anon) { 2515 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 2516 arc_buf_free(buf, tag); 2517 return (no_callback); 2518 } 2519 2520 mutex_enter(hash_lock); 2521 hdr = buf->b_hdr; 2522 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 2523 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2524 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2525 ASSERT(buf->b_data != NULL); 2526 2527 (void) remove_reference(hdr, hash_lock, tag); 2528 if (hdr->b_l1hdr.b_datacnt > 1) { 2529 if (no_callback) 2530 arc_buf_destroy(buf, TRUE); 2531 } else if (no_callback) { 2532 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL); 2533 ASSERT(buf->b_efunc == NULL); 2534 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2535 } 2536 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 || 2537 refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2538 mutex_exit(hash_lock); 2539 return (no_callback); 2540} 2541 2542int32_t 2543arc_buf_size(arc_buf_t *buf) 2544{ 2545 return (buf->b_hdr->b_size); 2546} 2547 2548/* 2549 * Called from the DMU to determine if the current buffer should be 2550 * evicted. In order to ensure proper locking, the eviction must be initiated 2551 * from the DMU. Return true if the buffer is associated with user data and 2552 * duplicate buffers still exist. 2553 */ 2554boolean_t 2555arc_buf_eviction_needed(arc_buf_t *buf) 2556{ 2557 arc_buf_hdr_t *hdr; 2558 boolean_t evict_needed = B_FALSE; 2559 2560 if (zfs_disable_dup_eviction) 2561 return (B_FALSE); 2562 2563 mutex_enter(&buf->b_evict_lock); 2564 hdr = buf->b_hdr; 2565 if (hdr == NULL) { 2566 /* 2567 * We are in arc_do_user_evicts(); let that function 2568 * perform the eviction. 2569 */ 2570 ASSERT(buf->b_data == NULL); 2571 mutex_exit(&buf->b_evict_lock); 2572 return (B_FALSE); 2573 } else if (buf->b_data == NULL) { 2574 /* 2575 * We have already been added to the arc eviction list; 2576 * recommend eviction. 2577 */ 2578 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2579 mutex_exit(&buf->b_evict_lock); 2580 return (B_TRUE); 2581 } 2582 2583 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr)) 2584 evict_needed = B_TRUE; 2585 2586 mutex_exit(&buf->b_evict_lock); 2587 return (evict_needed); 2588} 2589 2590/* 2591 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 2592 * state of the header is dependent on it's state prior to entering this 2593 * function. The following transitions are possible: 2594 * 2595 * - arc_mru -> arc_mru_ghost 2596 * - arc_mfu -> arc_mfu_ghost 2597 * - arc_mru_ghost -> arc_l2c_only 2598 * - arc_mru_ghost -> deleted 2599 * - arc_mfu_ghost -> arc_l2c_only 2600 * - arc_mfu_ghost -> deleted 2601 */ 2602static int64_t 2603arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 2604{ 2605 arc_state_t *evicted_state, *state; 2606 int64_t bytes_evicted = 0; 2607 2608 ASSERT(MUTEX_HELD(hash_lock)); 2609 ASSERT(HDR_HAS_L1HDR(hdr)); 2610 2611 state = hdr->b_l1hdr.b_state; 2612 if (GHOST_STATE(state)) { 2613 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2614 ASSERT(hdr->b_l1hdr.b_buf == NULL); 2615 2616 /* 2617 * l2arc_write_buffers() relies on a header's L1 portion 2618 * (i.e. it's b_tmp_cdata field) during it's write phase. 2619 * Thus, we cannot push a header onto the arc_l2c_only 2620 * state (removing it's L1 piece) until the header is 2621 * done being written to the l2arc. 2622 */ 2623 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 2624 ARCSTAT_BUMP(arcstat_evict_l2_skip); 2625 return (bytes_evicted); 2626 } 2627 2628 ARCSTAT_BUMP(arcstat_deleted); 2629 bytes_evicted += hdr->b_size; 2630 2631 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2632 2633 if (HDR_HAS_L2HDR(hdr)) { 2634 /* 2635 * This buffer is cached on the 2nd Level ARC; 2636 * don't destroy the header. 2637 */ 2638 arc_change_state(arc_l2c_only, hdr, hash_lock); 2639 /* 2640 * dropping from L1+L2 cached to L2-only, 2641 * realloc to remove the L1 header. 2642 */ 2643 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 2644 hdr_l2only_cache); 2645 } else { 2646 arc_change_state(arc_anon, hdr, hash_lock); 2647 arc_hdr_destroy(hdr); 2648 } 2649 return (bytes_evicted); 2650 } 2651 2652 ASSERT(state == arc_mru || state == arc_mfu); 2653 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2654 2655 /* prefetch buffers have a minimum lifespan */ 2656 if (HDR_IO_IN_PROGRESS(hdr) || 2657 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 2658 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 2659 arc_min_prefetch_lifespan)) { 2660 ARCSTAT_BUMP(arcstat_evict_skip); 2661 return (bytes_evicted); 2662 } 2663 2664 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 2665 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0); 2666 while (hdr->b_l1hdr.b_buf) { 2667 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2668 if (!mutex_tryenter(&buf->b_evict_lock)) { 2669 ARCSTAT_BUMP(arcstat_mutex_miss); 2670 break; 2671 } 2672 if (buf->b_data != NULL) 2673 bytes_evicted += hdr->b_size; 2674 if (buf->b_efunc != NULL) { 2675 mutex_enter(&arc_user_evicts_lock); 2676 arc_buf_destroy(buf, FALSE); 2677 hdr->b_l1hdr.b_buf = buf->b_next; 2678 buf->b_hdr = &arc_eviction_hdr; 2679 buf->b_next = arc_eviction_list; 2680 arc_eviction_list = buf; 2681 cv_signal(&arc_user_evicts_cv); 2682 mutex_exit(&arc_user_evicts_lock); 2683 mutex_exit(&buf->b_evict_lock); 2684 } else { 2685 mutex_exit(&buf->b_evict_lock); 2686 arc_buf_destroy(buf, TRUE); 2687 } 2688 } 2689 2690 if (HDR_HAS_L2HDR(hdr)) { 2691 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size); 2692 } else { 2693 if (l2arc_write_eligible(hdr->b_spa, hdr)) 2694 ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size); 2695 else 2696 ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size); 2697 } 2698 2699 if (hdr->b_l1hdr.b_datacnt == 0) { 2700 arc_change_state(evicted_state, hdr, hash_lock); 2701 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2702 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2703 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2704 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2705 } 2706 2707 return (bytes_evicted); 2708} 2709 2710static uint64_t 2711arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 2712 uint64_t spa, int64_t bytes) 2713{ 2714 multilist_sublist_t *mls; 2715 uint64_t bytes_evicted = 0; 2716 arc_buf_hdr_t *hdr; 2717 kmutex_t *hash_lock; 2718 int evict_count = 0; 2719 2720 ASSERT3P(marker, !=, NULL); 2721 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2722 2723 mls = multilist_sublist_lock(ml, idx); 2724 2725 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 2726 hdr = multilist_sublist_prev(mls, marker)) { 2727 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 2728 (evict_count >= zfs_arc_evict_batch_limit)) 2729 break; 2730 2731 /* 2732 * To keep our iteration location, move the marker 2733 * forward. Since we're not holding hdr's hash lock, we 2734 * must be very careful and not remove 'hdr' from the 2735 * sublist. Otherwise, other consumers might mistake the 2736 * 'hdr' as not being on a sublist when they call the 2737 * multilist_link_active() function (they all rely on 2738 * the hash lock protecting concurrent insertions and 2739 * removals). multilist_sublist_move_forward() was 2740 * specifically implemented to ensure this is the case 2741 * (only 'marker' will be removed and re-inserted). 2742 */ 2743 multilist_sublist_move_forward(mls, marker); 2744 2745 /* 2746 * The only case where the b_spa field should ever be 2747 * zero, is the marker headers inserted by 2748 * arc_evict_state(). It's possible for multiple threads 2749 * to be calling arc_evict_state() concurrently (e.g. 2750 * dsl_pool_close() and zio_inject_fault()), so we must 2751 * skip any markers we see from these other threads. 2752 */ 2753 if (hdr->b_spa == 0) 2754 continue; 2755 2756 /* we're only interested in evicting buffers of a certain spa */ 2757 if (spa != 0 && hdr->b_spa != spa) { 2758 ARCSTAT_BUMP(arcstat_evict_skip); 2759 continue; 2760 } 2761 2762 hash_lock = HDR_LOCK(hdr); 2763 2764 /* 2765 * We aren't calling this function from any code path 2766 * that would already be holding a hash lock, so we're 2767 * asserting on this assumption to be defensive in case 2768 * this ever changes. Without this check, it would be 2769 * possible to incorrectly increment arcstat_mutex_miss 2770 * below (e.g. if the code changed such that we called 2771 * this function with a hash lock held). 2772 */ 2773 ASSERT(!MUTEX_HELD(hash_lock)); 2774 2775 if (mutex_tryenter(hash_lock)) { 2776 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 2777 mutex_exit(hash_lock); 2778 2779 bytes_evicted += evicted; 2780 2781 /* 2782 * If evicted is zero, arc_evict_hdr() must have 2783 * decided to skip this header, don't increment 2784 * evict_count in this case. 2785 */ 2786 if (evicted != 0) 2787 evict_count++; 2788 2789 /* 2790 * If arc_size isn't overflowing, signal any 2791 * threads that might happen to be waiting. 2792 * 2793 * For each header evicted, we wake up a single 2794 * thread. If we used cv_broadcast, we could 2795 * wake up "too many" threads causing arc_size 2796 * to significantly overflow arc_c; since 2797 * arc_get_data_buf() doesn't check for overflow 2798 * when it's woken up (it doesn't because it's 2799 * possible for the ARC to be overflowing while 2800 * full of un-evictable buffers, and the 2801 * function should proceed in this case). 2802 * 2803 * If threads are left sleeping, due to not 2804 * using cv_broadcast, they will be woken up 2805 * just before arc_reclaim_thread() sleeps. 2806 */ 2807 mutex_enter(&arc_reclaim_lock); 2808 if (!arc_is_overflowing()) 2809 cv_signal(&arc_reclaim_waiters_cv); 2810 mutex_exit(&arc_reclaim_lock); 2811 } else { 2812 ARCSTAT_BUMP(arcstat_mutex_miss); 2813 } 2814 } 2815 2816 multilist_sublist_unlock(mls); 2817 2818 return (bytes_evicted); 2819} 2820 2821/* 2822 * Evict buffers from the given arc state, until we've removed the 2823 * specified number of bytes. Move the removed buffers to the 2824 * appropriate evict state. 2825 * 2826 * This function makes a "best effort". It skips over any buffers 2827 * it can't get a hash_lock on, and so, may not catch all candidates. 2828 * It may also return without evicting as much space as requested. 2829 * 2830 * If bytes is specified using the special value ARC_EVICT_ALL, this 2831 * will evict all available (i.e. unlocked and evictable) buffers from 2832 * the given arc state; which is used by arc_flush(). 2833 */ 2834static uint64_t 2835arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 2836 arc_buf_contents_t type) 2837{ 2838 uint64_t total_evicted = 0; 2839 multilist_t *ml = &state->arcs_list[type]; 2840 int num_sublists; 2841 arc_buf_hdr_t **markers; 2842 2843 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2844 2845 num_sublists = multilist_get_num_sublists(ml); 2846 2847 /* 2848 * If we've tried to evict from each sublist, made some 2849 * progress, but still have not hit the target number of bytes 2850 * to evict, we want to keep trying. The markers allow us to 2851 * pick up where we left off for each individual sublist, rather 2852 * than starting from the tail each time. 2853 */ 2854 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 2855 for (int i = 0; i < num_sublists; i++) { 2856 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 2857 2858 /* 2859 * A b_spa of 0 is used to indicate that this header is 2860 * a marker. This fact is used in arc_adjust_type() and 2861 * arc_evict_state_impl(). 2862 */ 2863 markers[i]->b_spa = 0; 2864 2865 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2866 multilist_sublist_insert_tail(mls, markers[i]); 2867 multilist_sublist_unlock(mls); 2868 } 2869 2870 /* 2871 * While we haven't hit our target number of bytes to evict, or 2872 * we're evicting all available buffers. 2873 */ 2874 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 2875 /* 2876 * Start eviction using a randomly selected sublist, 2877 * this is to try and evenly balance eviction across all 2878 * sublists. Always starting at the same sublist 2879 * (e.g. index 0) would cause evictions to favor certain 2880 * sublists over others. 2881 */ 2882 int sublist_idx = multilist_get_random_index(ml); 2883 uint64_t scan_evicted = 0; 2884 2885 for (int i = 0; i < num_sublists; i++) { 2886 uint64_t bytes_remaining; 2887 uint64_t bytes_evicted; 2888 2889 if (bytes == ARC_EVICT_ALL) 2890 bytes_remaining = ARC_EVICT_ALL; 2891 else if (total_evicted < bytes) 2892 bytes_remaining = bytes - total_evicted; 2893 else 2894 break; 2895 2896 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 2897 markers[sublist_idx], spa, bytes_remaining); 2898 2899 scan_evicted += bytes_evicted; 2900 total_evicted += bytes_evicted; 2901 2902 /* we've reached the end, wrap to the beginning */ 2903 if (++sublist_idx >= num_sublists) 2904 sublist_idx = 0; 2905 } 2906 2907 /* 2908 * If we didn't evict anything during this scan, we have 2909 * no reason to believe we'll evict more during another 2910 * scan, so break the loop. 2911 */ 2912 if (scan_evicted == 0) { 2913 /* This isn't possible, let's make that obvious */ 2914 ASSERT3S(bytes, !=, 0); 2915 2916 /* 2917 * When bytes is ARC_EVICT_ALL, the only way to 2918 * break the loop is when scan_evicted is zero. 2919 * In that case, we actually have evicted enough, 2920 * so we don't want to increment the kstat. 2921 */ 2922 if (bytes != ARC_EVICT_ALL) { 2923 ASSERT3S(total_evicted, <, bytes); 2924 ARCSTAT_BUMP(arcstat_evict_not_enough); 2925 } 2926 2927 break; 2928 } 2929 } 2930 2931 for (int i = 0; i < num_sublists; i++) { 2932 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2933 multilist_sublist_remove(mls, markers[i]); 2934 multilist_sublist_unlock(mls); 2935 2936 kmem_cache_free(hdr_full_cache, markers[i]); 2937 } 2938 kmem_free(markers, sizeof (*markers) * num_sublists); 2939 2940 return (total_evicted); 2941} 2942 2943/* 2944 * Flush all "evictable" data of the given type from the arc state 2945 * specified. This will not evict any "active" buffers (i.e. referenced). 2946 * 2947 * When 'retry' is set to FALSE, the function will make a single pass 2948 * over the state and evict any buffers that it can. Since it doesn't 2949 * continually retry the eviction, it might end up leaving some buffers 2950 * in the ARC due to lock misses. 2951 * 2952 * When 'retry' is set to TRUE, the function will continually retry the 2953 * eviction until *all* evictable buffers have been removed from the 2954 * state. As a result, if concurrent insertions into the state are 2955 * allowed (e.g. if the ARC isn't shutting down), this function might 2956 * wind up in an infinite loop, continually trying to evict buffers. 2957 */ 2958static uint64_t 2959arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 2960 boolean_t retry) 2961{ 2962 uint64_t evicted = 0; 2963 2964 while (state->arcs_lsize[type] != 0) { 2965 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 2966 2967 if (!retry) 2968 break; 2969 } 2970 2971 return (evicted); 2972} 2973 2974/* 2975 * Evict the specified number of bytes from the state specified, 2976 * restricting eviction to the spa and type given. This function 2977 * prevents us from trying to evict more from a state's list than 2978 * is "evictable", and to skip evicting altogether when passed a 2979 * negative value for "bytes". In contrast, arc_evict_state() will 2980 * evict everything it can, when passed a negative value for "bytes". 2981 */ 2982static uint64_t 2983arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 2984 arc_buf_contents_t type) 2985{ 2986 int64_t delta; 2987 2988 if (bytes > 0 && state->arcs_lsize[type] > 0) { 2989 delta = MIN(state->arcs_lsize[type], bytes); 2990 return (arc_evict_state(state, spa, delta, type)); 2991 } 2992 2993 return (0); 2994} 2995 2996/* 2997 * Evict metadata buffers from the cache, such that arc_meta_used is 2998 * capped by the arc_meta_limit tunable. 2999 */ 3000static uint64_t 3001arc_adjust_meta(void) 3002{ 3003 uint64_t total_evicted = 0; 3004 int64_t target; 3005 3006 /* 3007 * If we're over the meta limit, we want to evict enough 3008 * metadata to get back under the meta limit. We don't want to 3009 * evict so much that we drop the MRU below arc_p, though. If 3010 * we're over the meta limit more than we're over arc_p, we 3011 * evict some from the MRU here, and some from the MFU below. 3012 */ 3013 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3014 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3015 refcount_count(&arc_mru->arcs_size) - arc_p)); 3016 3017 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3018 3019 /* 3020 * Similar to the above, we want to evict enough bytes to get us 3021 * below the meta limit, but not so much as to drop us below the 3022 * space alloted to the MFU (which is defined as arc_c - arc_p). 3023 */ 3024 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3025 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3026 3027 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3028 3029 return (total_evicted); 3030} 3031 3032/* 3033 * Return the type of the oldest buffer in the given arc state 3034 * 3035 * This function will select a random sublist of type ARC_BUFC_DATA and 3036 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3037 * is compared, and the type which contains the "older" buffer will be 3038 * returned. 3039 */ 3040static arc_buf_contents_t 3041arc_adjust_type(arc_state_t *state) 3042{ 3043 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; 3044 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; 3045 int data_idx = multilist_get_random_index(data_ml); 3046 int meta_idx = multilist_get_random_index(meta_ml); 3047 multilist_sublist_t *data_mls; 3048 multilist_sublist_t *meta_mls; 3049 arc_buf_contents_t type; 3050 arc_buf_hdr_t *data_hdr; 3051 arc_buf_hdr_t *meta_hdr; 3052 3053 /* 3054 * We keep the sublist lock until we're finished, to prevent 3055 * the headers from being destroyed via arc_evict_state(). 3056 */ 3057 data_mls = multilist_sublist_lock(data_ml, data_idx); 3058 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3059 3060 /* 3061 * These two loops are to ensure we skip any markers that 3062 * might be at the tail of the lists due to arc_evict_state(). 3063 */ 3064 3065 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3066 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3067 if (data_hdr->b_spa != 0) 3068 break; 3069 } 3070 3071 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3072 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3073 if (meta_hdr->b_spa != 0) 3074 break; 3075 } 3076 3077 if (data_hdr == NULL && meta_hdr == NULL) { 3078 type = ARC_BUFC_DATA; 3079 } else if (data_hdr == NULL) { 3080 ASSERT3P(meta_hdr, !=, NULL); 3081 type = ARC_BUFC_METADATA; 3082 } else if (meta_hdr == NULL) { 3083 ASSERT3P(data_hdr, !=, NULL); 3084 type = ARC_BUFC_DATA; 3085 } else { 3086 ASSERT3P(data_hdr, !=, NULL); 3087 ASSERT3P(meta_hdr, !=, NULL); 3088 3089 /* The headers can't be on the sublist without an L1 header */ 3090 ASSERT(HDR_HAS_L1HDR(data_hdr)); 3091 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3092 3093 if (data_hdr->b_l1hdr.b_arc_access < 3094 meta_hdr->b_l1hdr.b_arc_access) { 3095 type = ARC_BUFC_DATA; 3096 } else { 3097 type = ARC_BUFC_METADATA; 3098 } 3099 } 3100 3101 multilist_sublist_unlock(meta_mls); 3102 multilist_sublist_unlock(data_mls); 3103 3104 return (type); 3105} 3106 3107/* 3108 * Evict buffers from the cache, such that arc_size is capped by arc_c. 3109 */ 3110static uint64_t 3111arc_adjust(void) 3112{ 3113 uint64_t total_evicted = 0; 3114 uint64_t bytes; 3115 int64_t target; 3116 3117 /* 3118 * If we're over arc_meta_limit, we want to correct that before 3119 * potentially evicting data buffers below. 3120 */ 3121 total_evicted += arc_adjust_meta(); 3122 3123 /* 3124 * Adjust MRU size 3125 * 3126 * If we're over the target cache size, we want to evict enough 3127 * from the list to get back to our target size. We don't want 3128 * to evict too much from the MRU, such that it drops below 3129 * arc_p. So, if we're over our target cache size more than 3130 * the MRU is over arc_p, we'll evict enough to get back to 3131 * arc_p here, and then evict more from the MFU below. 3132 */ 3133 target = MIN((int64_t)(arc_size - arc_c), 3134 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3135 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 3136 3137 /* 3138 * If we're below arc_meta_min, always prefer to evict data. 3139 * Otherwise, try to satisfy the requested number of bytes to 3140 * evict from the type which contains older buffers; in an 3141 * effort to keep newer buffers in the cache regardless of their 3142 * type. If we cannot satisfy the number of bytes from this 3143 * type, spill over into the next type. 3144 */ 3145 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3146 arc_meta_used > arc_meta_min) { 3147 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3148 total_evicted += bytes; 3149 3150 /* 3151 * If we couldn't evict our target number of bytes from 3152 * metadata, we try to get the rest from data. 3153 */ 3154 target -= bytes; 3155 3156 total_evicted += 3157 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3158 } else { 3159 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3160 total_evicted += bytes; 3161 3162 /* 3163 * If we couldn't evict our target number of bytes from 3164 * data, we try to get the rest from metadata. 3165 */ 3166 target -= bytes; 3167 3168 total_evicted += 3169 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3170 } 3171 3172 /* 3173 * Adjust MFU size 3174 * 3175 * Now that we've tried to evict enough from the MRU to get its 3176 * size back to arc_p, if we're still above the target cache 3177 * size, we evict the rest from the MFU. 3178 */ 3179 target = arc_size - arc_c; 3180 3181 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3182 arc_meta_used > arc_meta_min) { 3183 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3184 total_evicted += bytes; 3185 3186 /* 3187 * If we couldn't evict our target number of bytes from 3188 * metadata, we try to get the rest from data. 3189 */ 3190 target -= bytes; 3191 3192 total_evicted += 3193 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3194 } else { 3195 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3196 total_evicted += bytes; 3197 3198 /* 3199 * If we couldn't evict our target number of bytes from 3200 * data, we try to get the rest from data. 3201 */ 3202 target -= bytes; 3203 3204 total_evicted += 3205 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3206 } 3207 3208 /* 3209 * Adjust ghost lists 3210 * 3211 * In addition to the above, the ARC also defines target values 3212 * for the ghost lists. The sum of the mru list and mru ghost 3213 * list should never exceed the target size of the cache, and 3214 * the sum of the mru list, mfu list, mru ghost list, and mfu 3215 * ghost list should never exceed twice the target size of the 3216 * cache. The following logic enforces these limits on the ghost 3217 * caches, and evicts from them as needed. 3218 */ 3219 target = refcount_count(&arc_mru->arcs_size) + 3220 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3221 3222 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3223 total_evicted += bytes; 3224 3225 target -= bytes; 3226 3227 total_evicted += 3228 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 3229 3230 /* 3231 * We assume the sum of the mru list and mfu list is less than 3232 * or equal to arc_c (we enforced this above), which means we 3233 * can use the simpler of the two equations below: 3234 * 3235 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3236 * mru ghost + mfu ghost <= arc_c 3237 */ 3238 target = refcount_count(&arc_mru_ghost->arcs_size) + 3239 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3240 3241 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3242 total_evicted += bytes; 3243 3244 target -= bytes; 3245 3246 total_evicted += 3247 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3248 3249 return (total_evicted); 3250} 3251 3252static void 3253arc_do_user_evicts(void) 3254{ 3255 mutex_enter(&arc_user_evicts_lock); 3256 while (arc_eviction_list != NULL) { 3257 arc_buf_t *buf = arc_eviction_list; 3258 arc_eviction_list = buf->b_next; 3259 mutex_enter(&buf->b_evict_lock); 3260 buf->b_hdr = NULL; 3261 mutex_exit(&buf->b_evict_lock); 3262 mutex_exit(&arc_user_evicts_lock); 3263 3264 if (buf->b_efunc != NULL) 3265 VERIFY0(buf->b_efunc(buf->b_private)); 3266 3267 buf->b_efunc = NULL; 3268 buf->b_private = NULL; 3269 kmem_cache_free(buf_cache, buf); 3270 mutex_enter(&arc_user_evicts_lock); 3271 } 3272 mutex_exit(&arc_user_evicts_lock); 3273} 3274 3275void 3276arc_flush(spa_t *spa, boolean_t retry) 3277{ 3278 uint64_t guid = 0; 3279 3280 /* 3281 * If retry is TRUE, a spa must not be specified since we have 3282 * no good way to determine if all of a spa's buffers have been 3283 * evicted from an arc state. 3284 */ 3285 ASSERT(!retry || spa == 0); 3286 3287 if (spa != NULL) 3288 guid = spa_load_guid(spa); 3289 3290 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3291 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3292 3293 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3294 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3295 3296 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3297 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3298 3299 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3300 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3301 3302 arc_do_user_evicts(); 3303 ASSERT(spa || arc_eviction_list == NULL); 3304} 3305 3306void 3307arc_shrink(int64_t to_free) 3308{ 3309 if (arc_c > arc_c_min) { 3310 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 3311 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 3312 if (arc_c > arc_c_min + to_free) 3313 atomic_add_64(&arc_c, -to_free); 3314 else 3315 arc_c = arc_c_min; 3316 3317 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3318 if (arc_c > arc_size) 3319 arc_c = MAX(arc_size, arc_c_min); 3320 if (arc_p > arc_c) 3321 arc_p = (arc_c >> 1); 3322 3323 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 3324 arc_p); 3325 3326 ASSERT(arc_c >= arc_c_min); 3327 ASSERT((int64_t)arc_p >= 0); 3328 } 3329 3330 if (arc_size > arc_c) { 3331 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 3332 uint64_t, arc_c); 3333 (void) arc_adjust(); 3334 } 3335} 3336 3337static long needfree = 0; 3338 3339typedef enum free_memory_reason_t { 3340 FMR_UNKNOWN, 3341 FMR_NEEDFREE, 3342 FMR_LOTSFREE, 3343 FMR_SWAPFS_MINFREE, 3344 FMR_PAGES_PP_MAXIMUM, 3345 FMR_HEAP_ARENA, 3346 FMR_ZIO_ARENA, 3347 FMR_ZIO_FRAG, 3348} free_memory_reason_t; 3349 3350int64_t last_free_memory; 3351free_memory_reason_t last_free_reason; 3352 3353/* 3354 * Additional reserve of pages for pp_reserve. 3355 */ 3356int64_t arc_pages_pp_reserve = 64; 3357 3358/* 3359 * Additional reserve of pages for swapfs. 3360 */ 3361int64_t arc_swapfs_reserve = 64; 3362 3363/* 3364 * Return the amount of memory that can be consumed before reclaim will be 3365 * needed. Positive if there is sufficient free memory, negative indicates 3366 * the amount of memory that needs to be freed up. 3367 */ 3368static int64_t 3369arc_available_memory(void) 3370{ 3371 int64_t lowest = INT64_MAX; 3372 int64_t n; 3373 free_memory_reason_t r = FMR_UNKNOWN; 3374 3375#ifdef _KERNEL 3376 if (needfree > 0) { 3377 n = PAGESIZE * (-needfree); 3378 if (n < lowest) { 3379 lowest = n; 3380 r = FMR_NEEDFREE; 3381 } 3382 } 3383 3384 /* 3385 * Cooperate with pagedaemon when it's time for it to scan 3386 * and reclaim some pages. 3387 */ 3388 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 3389 if (n < lowest) { 3390 lowest = n; 3391 r = FMR_LOTSFREE; 3392 } 3393 3394#ifdef illumos 3395 /* 3396 * check that we're out of range of the pageout scanner. It starts to 3397 * schedule paging if freemem is less than lotsfree and needfree. 3398 * lotsfree is the high-water mark for pageout, and needfree is the 3399 * number of needed free pages. We add extra pages here to make sure 3400 * the scanner doesn't start up while we're freeing memory. 3401 */ 3402 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 3403 if (n < lowest) { 3404 lowest = n; 3405 r = FMR_LOTSFREE; 3406 } 3407 3408 /* 3409 * check to make sure that swapfs has enough space so that anon 3410 * reservations can still succeed. anon_resvmem() checks that the 3411 * availrmem is greater than swapfs_minfree, and the number of reserved 3412 * swap pages. We also add a bit of extra here just to prevent 3413 * circumstances from getting really dire. 3414 */ 3415 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 3416 desfree - arc_swapfs_reserve); 3417 if (n < lowest) { 3418 lowest = n; 3419 r = FMR_SWAPFS_MINFREE; 3420 } 3421 3422 3423 /* 3424 * Check that we have enough availrmem that memory locking (e.g., via 3425 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3426 * stores the number of pages that cannot be locked; when availrmem 3427 * drops below pages_pp_maximum, page locking mechanisms such as 3428 * page_pp_lock() will fail.) 3429 */ 3430 n = PAGESIZE * (availrmem - pages_pp_maximum - 3431 arc_pages_pp_reserve); 3432 if (n < lowest) { 3433 lowest = n; 3434 r = FMR_PAGES_PP_MAXIMUM; 3435 } 3436 3437#endif /* illumos */ 3438#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3439 /* 3440 * If we're on an i386 platform, it's possible that we'll exhaust the 3441 * kernel heap space before we ever run out of available physical 3442 * memory. Most checks of the size of the heap_area compare against 3443 * tune.t_minarmem, which is the minimum available real memory that we 3444 * can have in the system. However, this is generally fixed at 25 pages 3445 * which is so low that it's useless. In this comparison, we seek to 3446 * calculate the total heap-size, and reclaim if more than 3/4ths of the 3447 * heap is allocated. (Or, in the calculation, if less than 1/4th is 3448 * free) 3449 */ 3450 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 3451 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 3452 if (n < lowest) { 3453 lowest = n; 3454 r = FMR_HEAP_ARENA; 3455 } 3456#define zio_arena NULL 3457#else 3458#define zio_arena heap_arena 3459#endif 3460 3461 /* 3462 * If zio data pages are being allocated out of a separate heap segment, 3463 * then enforce that the size of available vmem for this arena remains 3464 * above about 1/16th free. 3465 * 3466 * Note: The 1/16th arena free requirement was put in place 3467 * to aggressively evict memory from the arc in order to avoid 3468 * memory fragmentation issues. 3469 */ 3470 if (zio_arena != NULL) { 3471 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 3472 (vmem_size(zio_arena, VMEM_ALLOC) >> 4); 3473 if (n < lowest) { 3474 lowest = n; 3475 r = FMR_ZIO_ARENA; 3476 } 3477 } 3478 3479 /* 3480 * Above limits know nothing about real level of KVA fragmentation. 3481 * Start aggressive reclamation if too little sequential KVA left. 3482 */ 3483 if (lowest > 0) { 3484 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ? 3485 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 3486 INT64_MAX; 3487 if (n < lowest) { 3488 lowest = n; 3489 r = FMR_ZIO_FRAG; 3490 } 3491 } 3492 3493#else /* _KERNEL */ 3494 /* Every 100 calls, free a small amount */ 3495 if (spa_get_random(100) == 0) 3496 lowest = -1024; 3497#endif /* _KERNEL */ 3498 3499 last_free_memory = lowest; 3500 last_free_reason = r; 3501 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 3502 return (lowest); 3503} 3504 3505 3506/* 3507 * Determine if the system is under memory pressure and is asking 3508 * to reclaim memory. A return value of TRUE indicates that the system 3509 * is under memory pressure and that the arc should adjust accordingly. 3510 */ 3511static boolean_t 3512arc_reclaim_needed(void) 3513{ 3514 return (arc_available_memory() < 0); 3515} 3516 3517extern kmem_cache_t *zio_buf_cache[]; 3518extern kmem_cache_t *zio_data_buf_cache[]; 3519extern kmem_cache_t *range_seg_cache; 3520 3521static __noinline void 3522arc_kmem_reap_now(void) 3523{ 3524 size_t i; 3525 kmem_cache_t *prev_cache = NULL; 3526 kmem_cache_t *prev_data_cache = NULL; 3527 3528 DTRACE_PROBE(arc__kmem_reap_start); 3529#ifdef _KERNEL 3530 if (arc_meta_used >= arc_meta_limit) { 3531 /* 3532 * We are exceeding our meta-data cache limit. 3533 * Purge some DNLC entries to release holds on meta-data. 3534 */ 3535 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 3536 } 3537#if defined(__i386) 3538 /* 3539 * Reclaim unused memory from all kmem caches. 3540 */ 3541 kmem_reap(); 3542#endif 3543#endif 3544 3545 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 3546 if (zio_buf_cache[i] != prev_cache) { 3547 prev_cache = zio_buf_cache[i]; 3548 kmem_cache_reap_now(zio_buf_cache[i]); 3549 } 3550 if (zio_data_buf_cache[i] != prev_data_cache) { 3551 prev_data_cache = zio_data_buf_cache[i]; 3552 kmem_cache_reap_now(zio_data_buf_cache[i]); 3553 } 3554 } 3555 kmem_cache_reap_now(buf_cache); 3556 kmem_cache_reap_now(hdr_full_cache); 3557 kmem_cache_reap_now(hdr_l2only_cache); 3558 kmem_cache_reap_now(range_seg_cache); 3559 3560#ifdef illumos 3561 if (zio_arena != NULL) { 3562 /* 3563 * Ask the vmem arena to reclaim unused memory from its 3564 * quantum caches. 3565 */ 3566 vmem_qcache_reap(zio_arena); 3567 } 3568#endif 3569 DTRACE_PROBE(arc__kmem_reap_end); 3570} 3571 3572/* 3573 * Threads can block in arc_get_data_buf() waiting for this thread to evict 3574 * enough data and signal them to proceed. When this happens, the threads in 3575 * arc_get_data_buf() are sleeping while holding the hash lock for their 3576 * particular arc header. Thus, we must be careful to never sleep on a 3577 * hash lock in this thread. This is to prevent the following deadlock: 3578 * 3579 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", 3580 * waiting for the reclaim thread to signal it. 3581 * 3582 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 3583 * fails, and goes to sleep forever. 3584 * 3585 * This possible deadlock is avoided by always acquiring a hash lock 3586 * using mutex_tryenter() from arc_reclaim_thread(). 3587 */ 3588static void 3589arc_reclaim_thread(void *dummy __unused) 3590{ 3591 clock_t growtime = 0; 3592 callb_cpr_t cpr; 3593 3594 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 3595 3596 mutex_enter(&arc_reclaim_lock); 3597 while (!arc_reclaim_thread_exit) { 3598 int64_t free_memory = arc_available_memory(); 3599 uint64_t evicted = 0; 3600 3601 mutex_exit(&arc_reclaim_lock); 3602 3603 if (free_memory < 0) { 3604 3605 arc_no_grow = B_TRUE; 3606 arc_warm = B_TRUE; 3607 3608 /* 3609 * Wait at least zfs_grow_retry (default 60) seconds 3610 * before considering growing. 3611 */ 3612 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 3613 3614 arc_kmem_reap_now(); 3615 3616 /* 3617 * If we are still low on memory, shrink the ARC 3618 * so that we have arc_shrink_min free space. 3619 */ 3620 free_memory = arc_available_memory(); 3621 3622 int64_t to_free = 3623 (arc_c >> arc_shrink_shift) - free_memory; 3624 if (to_free > 0) { 3625#ifdef _KERNEL 3626 to_free = MAX(to_free, ptob(needfree)); 3627#endif 3628 arc_shrink(to_free); 3629 } 3630 } else if (free_memory < arc_c >> arc_no_grow_shift) { 3631 arc_no_grow = B_TRUE; 3632 } else if (ddi_get_lbolt() >= growtime) { 3633 arc_no_grow = B_FALSE; 3634 } 3635 3636 evicted = arc_adjust(); 3637 3638 mutex_enter(&arc_reclaim_lock); 3639 3640 /* 3641 * If evicted is zero, we couldn't evict anything via 3642 * arc_adjust(). This could be due to hash lock 3643 * collisions, but more likely due to the majority of 3644 * arc buffers being unevictable. Therefore, even if 3645 * arc_size is above arc_c, another pass is unlikely to 3646 * be helpful and could potentially cause us to enter an 3647 * infinite loop. 3648 */ 3649 if (arc_size <= arc_c || evicted == 0) { 3650#ifdef _KERNEL 3651 needfree = 0; 3652#endif 3653 /* 3654 * We're either no longer overflowing, or we 3655 * can't evict anything more, so we should wake 3656 * up any threads before we go to sleep. 3657 */ 3658 cv_broadcast(&arc_reclaim_waiters_cv); 3659 3660 /* 3661 * Block until signaled, or after one second (we 3662 * might need to perform arc_kmem_reap_now() 3663 * even if we aren't being signalled) 3664 */ 3665 CALLB_CPR_SAFE_BEGIN(&cpr); 3666 (void) cv_timedwait(&arc_reclaim_thread_cv, 3667 &arc_reclaim_lock, hz); 3668 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 3669 } 3670 } 3671 3672 arc_reclaim_thread_exit = FALSE; 3673 cv_broadcast(&arc_reclaim_thread_cv); 3674 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 3675 thread_exit(); 3676} 3677 3678static void 3679arc_user_evicts_thread(void *dummy __unused) 3680{ 3681 callb_cpr_t cpr; 3682 3683 CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG); 3684 3685 mutex_enter(&arc_user_evicts_lock); 3686 while (!arc_user_evicts_thread_exit) { 3687 mutex_exit(&arc_user_evicts_lock); 3688 3689 arc_do_user_evicts(); 3690 3691 /* 3692 * This is necessary in order for the mdb ::arc dcmd to 3693 * show up to date information. Since the ::arc command 3694 * does not call the kstat's update function, without 3695 * this call, the command may show stale stats for the 3696 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 3697 * with this change, the data might be up to 1 second 3698 * out of date; but that should suffice. The arc_state_t 3699 * structures can be queried directly if more accurate 3700 * information is needed. 3701 */ 3702 if (arc_ksp != NULL) 3703 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 3704 3705 mutex_enter(&arc_user_evicts_lock); 3706 3707 /* 3708 * Block until signaled, or after one second (we need to 3709 * call the arc's kstat update function regularly). 3710 */ 3711 CALLB_CPR_SAFE_BEGIN(&cpr); 3712 (void) cv_timedwait(&arc_user_evicts_cv, 3713 &arc_user_evicts_lock, hz); 3714 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock); 3715 } 3716 3717 arc_user_evicts_thread_exit = FALSE; 3718 cv_broadcast(&arc_user_evicts_cv); 3719 CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */ 3720 thread_exit(); 3721} 3722 3723/* 3724 * Adapt arc info given the number of bytes we are trying to add and 3725 * the state that we are comming from. This function is only called 3726 * when we are adding new content to the cache. 3727 */ 3728static void 3729arc_adapt(int bytes, arc_state_t *state) 3730{ 3731 int mult; 3732 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 3733 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 3734 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 3735 3736 if (state == arc_l2c_only) 3737 return; 3738 3739 ASSERT(bytes > 0); 3740 /* 3741 * Adapt the target size of the MRU list: 3742 * - if we just hit in the MRU ghost list, then increase 3743 * the target size of the MRU list. 3744 * - if we just hit in the MFU ghost list, then increase 3745 * the target size of the MFU list by decreasing the 3746 * target size of the MRU list. 3747 */ 3748 if (state == arc_mru_ghost) { 3749 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 3750 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 3751 3752 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 3753 } else if (state == arc_mfu_ghost) { 3754 uint64_t delta; 3755 3756 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 3757 mult = MIN(mult, 10); 3758 3759 delta = MIN(bytes * mult, arc_p); 3760 arc_p = MAX(arc_p_min, arc_p - delta); 3761 } 3762 ASSERT((int64_t)arc_p >= 0); 3763 3764 if (arc_reclaim_needed()) { 3765 cv_signal(&arc_reclaim_thread_cv); 3766 return; 3767 } 3768 3769 if (arc_no_grow) 3770 return; 3771 3772 if (arc_c >= arc_c_max) 3773 return; 3774 3775 /* 3776 * If we're within (2 * maxblocksize) bytes of the target 3777 * cache size, increment the target cache size 3778 */ 3779 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 3780 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 3781 atomic_add_64(&arc_c, (int64_t)bytes); 3782 if (arc_c > arc_c_max) 3783 arc_c = arc_c_max; 3784 else if (state == arc_anon) 3785 atomic_add_64(&arc_p, (int64_t)bytes); 3786 if (arc_p > arc_c) 3787 arc_p = arc_c; 3788 } 3789 ASSERT((int64_t)arc_p >= 0); 3790} 3791 3792/* 3793 * Check if arc_size has grown past our upper threshold, determined by 3794 * zfs_arc_overflow_shift. 3795 */ 3796static boolean_t 3797arc_is_overflowing(void) 3798{ 3799 /* Always allow at least one block of overflow */ 3800 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 3801 arc_c >> zfs_arc_overflow_shift); 3802 3803 return (arc_size >= arc_c + overflow); 3804} 3805 3806/* 3807 * The buffer, supplied as the first argument, needs a data block. If we 3808 * are hitting the hard limit for the cache size, we must sleep, waiting 3809 * for the eviction thread to catch up. If we're past the target size 3810 * but below the hard limit, we'll only signal the reclaim thread and 3811 * continue on. 3812 */ 3813static void 3814arc_get_data_buf(arc_buf_t *buf) 3815{ 3816 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 3817 uint64_t size = buf->b_hdr->b_size; 3818 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 3819 3820 arc_adapt(size, state); 3821 3822 /* 3823 * If arc_size is currently overflowing, and has grown past our 3824 * upper limit, we must be adding data faster than the evict 3825 * thread can evict. Thus, to ensure we don't compound the 3826 * problem by adding more data and forcing arc_size to grow even 3827 * further past it's target size, we halt and wait for the 3828 * eviction thread to catch up. 3829 * 3830 * It's also possible that the reclaim thread is unable to evict 3831 * enough buffers to get arc_size below the overflow limit (e.g. 3832 * due to buffers being un-evictable, or hash lock collisions). 3833 * In this case, we want to proceed regardless if we're 3834 * overflowing; thus we don't use a while loop here. 3835 */ 3836 if (arc_is_overflowing()) { 3837 mutex_enter(&arc_reclaim_lock); 3838 3839 /* 3840 * Now that we've acquired the lock, we may no longer be 3841 * over the overflow limit, lets check. 3842 * 3843 * We're ignoring the case of spurious wake ups. If that 3844 * were to happen, it'd let this thread consume an ARC 3845 * buffer before it should have (i.e. before we're under 3846 * the overflow limit and were signalled by the reclaim 3847 * thread). As long as that is a rare occurrence, it 3848 * shouldn't cause any harm. 3849 */ 3850 if (arc_is_overflowing()) { 3851 cv_signal(&arc_reclaim_thread_cv); 3852 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 3853 } 3854 3855 mutex_exit(&arc_reclaim_lock); 3856 } 3857 3858 if (type == ARC_BUFC_METADATA) { 3859 buf->b_data = zio_buf_alloc(size); 3860 arc_space_consume(size, ARC_SPACE_META); 3861 } else { 3862 ASSERT(type == ARC_BUFC_DATA); 3863 buf->b_data = zio_data_buf_alloc(size); 3864 arc_space_consume(size, ARC_SPACE_DATA); 3865 } 3866 3867 /* 3868 * Update the state size. Note that ghost states have a 3869 * "ghost size" and so don't need to be updated. 3870 */ 3871 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) { 3872 arc_buf_hdr_t *hdr = buf->b_hdr; 3873 arc_state_t *state = hdr->b_l1hdr.b_state; 3874 3875 (void) refcount_add_many(&state->arcs_size, size, buf); 3876 3877 /* 3878 * If this is reached via arc_read, the link is 3879 * protected by the hash lock. If reached via 3880 * arc_buf_alloc, the header should not be accessed by 3881 * any other thread. And, if reached via arc_read_done, 3882 * the hash lock will protect it if it's found in the 3883 * hash table; otherwise no other thread should be 3884 * trying to [add|remove]_reference it. 3885 */ 3886 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 3887 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3888 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type], 3889 size); 3890 } 3891 /* 3892 * If we are growing the cache, and we are adding anonymous 3893 * data, and we have outgrown arc_p, update arc_p 3894 */ 3895 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 3896 (refcount_count(&arc_anon->arcs_size) + 3897 refcount_count(&arc_mru->arcs_size) > arc_p)) 3898 arc_p = MIN(arc_c, arc_p + size); 3899 } 3900 ARCSTAT_BUMP(arcstat_allocated); 3901} 3902 3903/* 3904 * This routine is called whenever a buffer is accessed. 3905 * NOTE: the hash lock is dropped in this function. 3906 */ 3907static void 3908arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3909{ 3910 clock_t now; 3911 3912 ASSERT(MUTEX_HELD(hash_lock)); 3913 ASSERT(HDR_HAS_L1HDR(hdr)); 3914 3915 if (hdr->b_l1hdr.b_state == arc_anon) { 3916 /* 3917 * This buffer is not in the cache, and does not 3918 * appear in our "ghost" list. Add the new buffer 3919 * to the MRU state. 3920 */ 3921 3922 ASSERT0(hdr->b_l1hdr.b_arc_access); 3923 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3924 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3925 arc_change_state(arc_mru, hdr, hash_lock); 3926 3927 } else if (hdr->b_l1hdr.b_state == arc_mru) { 3928 now = ddi_get_lbolt(); 3929 3930 /* 3931 * If this buffer is here because of a prefetch, then either: 3932 * - clear the flag if this is a "referencing" read 3933 * (any subsequent access will bump this into the MFU state). 3934 * or 3935 * - move the buffer to the head of the list if this is 3936 * another prefetch (to make it less likely to be evicted). 3937 */ 3938 if (HDR_PREFETCH(hdr)) { 3939 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 3940 /* link protected by hash lock */ 3941 ASSERT(multilist_link_active( 3942 &hdr->b_l1hdr.b_arc_node)); 3943 } else { 3944 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3945 ARCSTAT_BUMP(arcstat_mru_hits); 3946 } 3947 hdr->b_l1hdr.b_arc_access = now; 3948 return; 3949 } 3950 3951 /* 3952 * This buffer has been "accessed" only once so far, 3953 * but it is still in the cache. Move it to the MFU 3954 * state. 3955 */ 3956 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 3957 /* 3958 * More than 125ms have passed since we 3959 * instantiated this buffer. Move it to the 3960 * most frequently used state. 3961 */ 3962 hdr->b_l1hdr.b_arc_access = now; 3963 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3964 arc_change_state(arc_mfu, hdr, hash_lock); 3965 } 3966 ARCSTAT_BUMP(arcstat_mru_hits); 3967 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 3968 arc_state_t *new_state; 3969 /* 3970 * This buffer has been "accessed" recently, but 3971 * was evicted from the cache. Move it to the 3972 * MFU state. 3973 */ 3974 3975 if (HDR_PREFETCH(hdr)) { 3976 new_state = arc_mru; 3977 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 3978 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3979 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3980 } else { 3981 new_state = arc_mfu; 3982 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3983 } 3984 3985 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3986 arc_change_state(new_state, hdr, hash_lock); 3987 3988 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3989 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 3990 /* 3991 * This buffer has been accessed more than once and is 3992 * still in the cache. Keep it in the MFU state. 3993 * 3994 * NOTE: an add_reference() that occurred when we did 3995 * the arc_read() will have kicked this off the list. 3996 * If it was a prefetch, we will explicitly move it to 3997 * the head of the list now. 3998 */ 3999 if ((HDR_PREFETCH(hdr)) != 0) { 4000 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4001 /* link protected by hash_lock */ 4002 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4003 } 4004 ARCSTAT_BUMP(arcstat_mfu_hits); 4005 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4006 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 4007 arc_state_t *new_state = arc_mfu; 4008 /* 4009 * This buffer has been accessed more than once but has 4010 * been evicted from the cache. Move it back to the 4011 * MFU state. 4012 */ 4013 4014 if (HDR_PREFETCH(hdr)) { 4015 /* 4016 * This is a prefetch access... 4017 * move this block back to the MRU state. 4018 */ 4019 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4020 new_state = arc_mru; 4021 } 4022 4023 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4024 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4025 arc_change_state(new_state, hdr, hash_lock); 4026 4027 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 4028 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4029 /* 4030 * This buffer is on the 2nd Level ARC. 4031 */ 4032 4033 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4034 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4035 arc_change_state(arc_mfu, hdr, hash_lock); 4036 } else { 4037 ASSERT(!"invalid arc state"); 4038 } 4039} 4040 4041/* a generic arc_done_func_t which you can use */ 4042/* ARGSUSED */ 4043void 4044arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4045{ 4046 if (zio == NULL || zio->io_error == 0) 4047 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 4048 VERIFY(arc_buf_remove_ref(buf, arg)); 4049} 4050 4051/* a generic arc_done_func_t */ 4052void 4053arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4054{ 4055 arc_buf_t **bufp = arg; 4056 if (zio && zio->io_error) { 4057 VERIFY(arc_buf_remove_ref(buf, arg)); 4058 *bufp = NULL; 4059 } else { 4060 *bufp = buf; 4061 ASSERT(buf->b_data); 4062 } 4063} 4064 4065static void 4066arc_read_done(zio_t *zio) 4067{ 4068 arc_buf_hdr_t *hdr; 4069 arc_buf_t *buf; 4070 arc_buf_t *abuf; /* buffer we're assigning to callback */ 4071 kmutex_t *hash_lock = NULL; 4072 arc_callback_t *callback_list, *acb; 4073 int freeable = FALSE; 4074 4075 buf = zio->io_private; 4076 hdr = buf->b_hdr; 4077 4078 /* 4079 * The hdr was inserted into hash-table and removed from lists 4080 * prior to starting I/O. We should find this header, since 4081 * it's in the hash table, and it should be legit since it's 4082 * not possible to evict it during the I/O. The only possible 4083 * reason for it not to be found is if we were freed during the 4084 * read. 4085 */ 4086 if (HDR_IN_HASH_TABLE(hdr)) { 4087 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 4088 ASSERT3U(hdr->b_dva.dva_word[0], ==, 4089 BP_IDENTITY(zio->io_bp)->dva_word[0]); 4090 ASSERT3U(hdr->b_dva.dva_word[1], ==, 4091 BP_IDENTITY(zio->io_bp)->dva_word[1]); 4092 4093 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 4094 &hash_lock); 4095 4096 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 4097 hash_lock == NULL) || 4098 (found == hdr && 4099 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 4100 (found == hdr && HDR_L2_READING(hdr))); 4101 } 4102 4103 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 4104 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4105 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 4106 4107 /* byteswap if necessary */ 4108 callback_list = hdr->b_l1hdr.b_acb; 4109 ASSERT(callback_list != NULL); 4110 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 4111 dmu_object_byteswap_t bswap = 4112 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4113 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 4114 byteswap_uint64_array : 4115 dmu_ot_byteswap[bswap].ob_func; 4116 func(buf->b_data, hdr->b_size); 4117 } 4118 4119 arc_cksum_compute(buf, B_FALSE); 4120#ifdef illumos 4121 arc_buf_watch(buf); 4122#endif 4123 4124 if (hash_lock && zio->io_error == 0 && 4125 hdr->b_l1hdr.b_state == arc_anon) { 4126 /* 4127 * Only call arc_access on anonymous buffers. This is because 4128 * if we've issued an I/O for an evicted buffer, we've already 4129 * called arc_access (to prevent any simultaneous readers from 4130 * getting confused). 4131 */ 4132 arc_access(hdr, hash_lock); 4133 } 4134 4135 /* create copies of the data buffer for the callers */ 4136 abuf = buf; 4137 for (acb = callback_list; acb; acb = acb->acb_next) { 4138 if (acb->acb_done) { 4139 if (abuf == NULL) { 4140 ARCSTAT_BUMP(arcstat_duplicate_reads); 4141 abuf = arc_buf_clone(buf); 4142 } 4143 acb->acb_buf = abuf; 4144 abuf = NULL; 4145 } 4146 } 4147 hdr->b_l1hdr.b_acb = NULL; 4148 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4149 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 4150 if (abuf == buf) { 4151 ASSERT(buf->b_efunc == NULL); 4152 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4153 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4154 } 4155 4156 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 4157 callback_list != NULL); 4158 4159 if (zio->io_error != 0) { 4160 hdr->b_flags |= ARC_FLAG_IO_ERROR; 4161 if (hdr->b_l1hdr.b_state != arc_anon) 4162 arc_change_state(arc_anon, hdr, hash_lock); 4163 if (HDR_IN_HASH_TABLE(hdr)) 4164 buf_hash_remove(hdr); 4165 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4166 } 4167 4168 /* 4169 * Broadcast before we drop the hash_lock to avoid the possibility 4170 * that the hdr (and hence the cv) might be freed before we get to 4171 * the cv_broadcast(). 4172 */ 4173 cv_broadcast(&hdr->b_l1hdr.b_cv); 4174 4175 if (hash_lock != NULL) { 4176 mutex_exit(hash_lock); 4177 } else { 4178 /* 4179 * This block was freed while we waited for the read to 4180 * complete. It has been removed from the hash table and 4181 * moved to the anonymous state (so that it won't show up 4182 * in the cache). 4183 */ 4184 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 4185 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4186 } 4187 4188 /* execute each callback and free its structure */ 4189 while ((acb = callback_list) != NULL) { 4190 if (acb->acb_done) 4191 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4192 4193 if (acb->acb_zio_dummy != NULL) { 4194 acb->acb_zio_dummy->io_error = zio->io_error; 4195 zio_nowait(acb->acb_zio_dummy); 4196 } 4197 4198 callback_list = acb->acb_next; 4199 kmem_free(acb, sizeof (arc_callback_t)); 4200 } 4201 4202 if (freeable) 4203 arc_hdr_destroy(hdr); 4204} 4205 4206/* 4207 * "Read" the block at the specified DVA (in bp) via the 4208 * cache. If the block is found in the cache, invoke the provided 4209 * callback immediately and return. Note that the `zio' parameter 4210 * in the callback will be NULL in this case, since no IO was 4211 * required. If the block is not in the cache pass the read request 4212 * on to the spa with a substitute callback function, so that the 4213 * requested block will be added to the cache. 4214 * 4215 * If a read request arrives for a block that has a read in-progress, 4216 * either wait for the in-progress read to complete (and return the 4217 * results); or, if this is a read with a "done" func, add a record 4218 * to the read to invoke the "done" func when the read completes, 4219 * and return; or just return. 4220 * 4221 * arc_read_done() will invoke all the requested "done" functions 4222 * for readers of this block. 4223 */ 4224int 4225arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 4226 void *private, zio_priority_t priority, int zio_flags, 4227 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4228{ 4229 arc_buf_hdr_t *hdr = NULL; 4230 arc_buf_t *buf = NULL; 4231 kmutex_t *hash_lock = NULL; 4232 zio_t *rzio; 4233 uint64_t guid = spa_load_guid(spa); 4234 4235 ASSERT(!BP_IS_EMBEDDED(bp) || 4236 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 4237 4238top: 4239 if (!BP_IS_EMBEDDED(bp)) { 4240 /* 4241 * Embedded BP's have no DVA and require no I/O to "read". 4242 * Create an anonymous arc buf to back it. 4243 */ 4244 hdr = buf_hash_find(guid, bp, &hash_lock); 4245 } 4246 4247 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) { 4248 4249 *arc_flags |= ARC_FLAG_CACHED; 4250 4251 if (HDR_IO_IN_PROGRESS(hdr)) { 4252 4253 if (*arc_flags & ARC_FLAG_WAIT) { 4254 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 4255 mutex_exit(hash_lock); 4256 goto top; 4257 } 4258 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4259 4260 if (done) { 4261 arc_callback_t *acb = NULL; 4262 4263 acb = kmem_zalloc(sizeof (arc_callback_t), 4264 KM_SLEEP); 4265 acb->acb_done = done; 4266 acb->acb_private = private; 4267 if (pio != NULL) 4268 acb->acb_zio_dummy = zio_null(pio, 4269 spa, NULL, NULL, NULL, zio_flags); 4270 4271 ASSERT(acb->acb_done != NULL); 4272 acb->acb_next = hdr->b_l1hdr.b_acb; 4273 hdr->b_l1hdr.b_acb = acb; 4274 add_reference(hdr, hash_lock, private); 4275 mutex_exit(hash_lock); 4276 return (0); 4277 } 4278 mutex_exit(hash_lock); 4279 return (0); 4280 } 4281 4282 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4283 hdr->b_l1hdr.b_state == arc_mfu); 4284 4285 if (done) { 4286 add_reference(hdr, hash_lock, private); 4287 /* 4288 * If this block is already in use, create a new 4289 * copy of the data so that we will be guaranteed 4290 * that arc_release() will always succeed. 4291 */ 4292 buf = hdr->b_l1hdr.b_buf; 4293 ASSERT(buf); 4294 ASSERT(buf->b_data); 4295 if (HDR_BUF_AVAILABLE(hdr)) { 4296 ASSERT(buf->b_efunc == NULL); 4297 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4298 } else { 4299 buf = arc_buf_clone(buf); 4300 } 4301 4302 } else if (*arc_flags & ARC_FLAG_PREFETCH && 4303 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4304 hdr->b_flags |= ARC_FLAG_PREFETCH; 4305 } 4306 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4307 arc_access(hdr, hash_lock); 4308 if (*arc_flags & ARC_FLAG_L2CACHE) 4309 hdr->b_flags |= ARC_FLAG_L2CACHE; 4310 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4311 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4312 mutex_exit(hash_lock); 4313 ARCSTAT_BUMP(arcstat_hits); 4314 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4315 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4316 data, metadata, hits); 4317 4318 if (done) 4319 done(NULL, buf, private); 4320 } else { 4321 uint64_t size = BP_GET_LSIZE(bp); 4322 arc_callback_t *acb; 4323 vdev_t *vd = NULL; 4324 uint64_t addr = 0; 4325 boolean_t devw = B_FALSE; 4326 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 4327 int32_t b_asize = 0; 4328 4329 if (hdr == NULL) { 4330 /* this block is not in the cache */ 4331 arc_buf_hdr_t *exists = NULL; 4332 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4333 buf = arc_buf_alloc(spa, size, private, type); 4334 hdr = buf->b_hdr; 4335 if (!BP_IS_EMBEDDED(bp)) { 4336 hdr->b_dva = *BP_IDENTITY(bp); 4337 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 4338 exists = buf_hash_insert(hdr, &hash_lock); 4339 } 4340 if (exists != NULL) { 4341 /* somebody beat us to the hash insert */ 4342 mutex_exit(hash_lock); 4343 buf_discard_identity(hdr); 4344 (void) arc_buf_remove_ref(buf, private); 4345 goto top; /* restart the IO request */ 4346 } 4347 4348 /* if this is a prefetch, we don't have a reference */ 4349 if (*arc_flags & ARC_FLAG_PREFETCH) { 4350 (void) remove_reference(hdr, hash_lock, 4351 private); 4352 hdr->b_flags |= ARC_FLAG_PREFETCH; 4353 } 4354 if (*arc_flags & ARC_FLAG_L2CACHE) 4355 hdr->b_flags |= ARC_FLAG_L2CACHE; 4356 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4357 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4358 if (BP_GET_LEVEL(bp) > 0) 4359 hdr->b_flags |= ARC_FLAG_INDIRECT; 4360 } else { 4361 /* 4362 * This block is in the ghost cache. If it was L2-only 4363 * (and thus didn't have an L1 hdr), we realloc the 4364 * header to add an L1 hdr. 4365 */ 4366 if (!HDR_HAS_L1HDR(hdr)) { 4367 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 4368 hdr_full_cache); 4369 } 4370 4371 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4372 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4373 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4374 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 4375 4376 /* if this is a prefetch, we don't have a reference */ 4377 if (*arc_flags & ARC_FLAG_PREFETCH) 4378 hdr->b_flags |= ARC_FLAG_PREFETCH; 4379 else 4380 add_reference(hdr, hash_lock, private); 4381 if (*arc_flags & ARC_FLAG_L2CACHE) 4382 hdr->b_flags |= ARC_FLAG_L2CACHE; 4383 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4384 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4385 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 4386 buf->b_hdr = hdr; 4387 buf->b_data = NULL; 4388 buf->b_efunc = NULL; 4389 buf->b_private = NULL; 4390 buf->b_next = NULL; 4391 hdr->b_l1hdr.b_buf = buf; 4392 ASSERT0(hdr->b_l1hdr.b_datacnt); 4393 hdr->b_l1hdr.b_datacnt = 1; 4394 arc_get_data_buf(buf); 4395 arc_access(hdr, hash_lock); 4396 } 4397 4398 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 4399 4400 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4401 acb->acb_done = done; 4402 acb->acb_private = private; 4403 4404 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4405 hdr->b_l1hdr.b_acb = acb; 4406 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4407 4408 if (HDR_HAS_L2HDR(hdr) && 4409 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 4410 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 4411 addr = hdr->b_l2hdr.b_daddr; 4412 b_compress = HDR_GET_COMPRESS(hdr); 4413 b_asize = hdr->b_l2hdr.b_asize; 4414 /* 4415 * Lock out device removal. 4416 */ 4417 if (vdev_is_dead(vd) || 4418 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 4419 vd = NULL; 4420 } 4421 4422 if (hash_lock != NULL) 4423 mutex_exit(hash_lock); 4424 4425 /* 4426 * At this point, we have a level 1 cache miss. Try again in 4427 * L2ARC if possible. 4428 */ 4429 ASSERT3U(hdr->b_size, ==, size); 4430 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 4431 uint64_t, size, zbookmark_phys_t *, zb); 4432 ARCSTAT_BUMP(arcstat_misses); 4433 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4434 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4435 data, metadata, misses); 4436#ifdef _KERNEL 4437 curthread->td_ru.ru_inblock++; 4438#endif 4439 4440 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 4441 /* 4442 * Read from the L2ARC if the following are true: 4443 * 1. The L2ARC vdev was previously cached. 4444 * 2. This buffer still has L2ARC metadata. 4445 * 3. This buffer isn't currently writing to the L2ARC. 4446 * 4. The L2ARC entry wasn't evicted, which may 4447 * also have invalidated the vdev. 4448 * 5. This isn't prefetch and l2arc_noprefetch is set. 4449 */ 4450 if (HDR_HAS_L2HDR(hdr) && 4451 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 4452 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 4453 l2arc_read_callback_t *cb; 4454 4455 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 4456 ARCSTAT_BUMP(arcstat_l2_hits); 4457 4458 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 4459 KM_SLEEP); 4460 cb->l2rcb_buf = buf; 4461 cb->l2rcb_spa = spa; 4462 cb->l2rcb_bp = *bp; 4463 cb->l2rcb_zb = *zb; 4464 cb->l2rcb_flags = zio_flags; 4465 cb->l2rcb_compress = b_compress; 4466 4467 ASSERT(addr >= VDEV_LABEL_START_SIZE && 4468 addr + size < vd->vdev_psize - 4469 VDEV_LABEL_END_SIZE); 4470 4471 /* 4472 * l2arc read. The SCL_L2ARC lock will be 4473 * released by l2arc_read_done(). 4474 * Issue a null zio if the underlying buffer 4475 * was squashed to zero size by compression. 4476 */ 4477 if (b_compress == ZIO_COMPRESS_EMPTY) { 4478 rzio = zio_null(pio, spa, vd, 4479 l2arc_read_done, cb, 4480 zio_flags | ZIO_FLAG_DONT_CACHE | 4481 ZIO_FLAG_CANFAIL | 4482 ZIO_FLAG_DONT_PROPAGATE | 4483 ZIO_FLAG_DONT_RETRY); 4484 } else { 4485 rzio = zio_read_phys(pio, vd, addr, 4486 b_asize, buf->b_data, 4487 ZIO_CHECKSUM_OFF, 4488 l2arc_read_done, cb, priority, 4489 zio_flags | ZIO_FLAG_DONT_CACHE | 4490 ZIO_FLAG_CANFAIL | 4491 ZIO_FLAG_DONT_PROPAGATE | 4492 ZIO_FLAG_DONT_RETRY, B_FALSE); 4493 } 4494 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 4495 zio_t *, rzio); 4496 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 4497 4498 if (*arc_flags & ARC_FLAG_NOWAIT) { 4499 zio_nowait(rzio); 4500 return (0); 4501 } 4502 4503 ASSERT(*arc_flags & ARC_FLAG_WAIT); 4504 if (zio_wait(rzio) == 0) 4505 return (0); 4506 4507 /* l2arc read error; goto zio_read() */ 4508 } else { 4509 DTRACE_PROBE1(l2arc__miss, 4510 arc_buf_hdr_t *, hdr); 4511 ARCSTAT_BUMP(arcstat_l2_misses); 4512 if (HDR_L2_WRITING(hdr)) 4513 ARCSTAT_BUMP(arcstat_l2_rw_clash); 4514 spa_config_exit(spa, SCL_L2ARC, vd); 4515 } 4516 } else { 4517 if (vd != NULL) 4518 spa_config_exit(spa, SCL_L2ARC, vd); 4519 if (l2arc_ndev != 0) { 4520 DTRACE_PROBE1(l2arc__miss, 4521 arc_buf_hdr_t *, hdr); 4522 ARCSTAT_BUMP(arcstat_l2_misses); 4523 } 4524 } 4525 4526 rzio = zio_read(pio, spa, bp, buf->b_data, size, 4527 arc_read_done, buf, priority, zio_flags, zb); 4528 4529 if (*arc_flags & ARC_FLAG_WAIT) 4530 return (zio_wait(rzio)); 4531 4532 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4533 zio_nowait(rzio); 4534 } 4535 return (0); 4536} 4537 4538void 4539arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 4540{ 4541 ASSERT(buf->b_hdr != NULL); 4542 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon); 4543 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) || 4544 func == NULL); 4545 ASSERT(buf->b_efunc == NULL); 4546 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 4547 4548 buf->b_efunc = func; 4549 buf->b_private = private; 4550} 4551 4552/* 4553 * Notify the arc that a block was freed, and thus will never be used again. 4554 */ 4555void 4556arc_freed(spa_t *spa, const blkptr_t *bp) 4557{ 4558 arc_buf_hdr_t *hdr; 4559 kmutex_t *hash_lock; 4560 uint64_t guid = spa_load_guid(spa); 4561 4562 ASSERT(!BP_IS_EMBEDDED(bp)); 4563 4564 hdr = buf_hash_find(guid, bp, &hash_lock); 4565 if (hdr == NULL) 4566 return; 4567 if (HDR_BUF_AVAILABLE(hdr)) { 4568 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 4569 add_reference(hdr, hash_lock, FTAG); 4570 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4571 mutex_exit(hash_lock); 4572 4573 arc_release(buf, FTAG); 4574 (void) arc_buf_remove_ref(buf, FTAG); 4575 } else { 4576 mutex_exit(hash_lock); 4577 } 4578 4579} 4580 4581/* 4582 * Clear the user eviction callback set by arc_set_callback(), first calling 4583 * it if it exists. Because the presence of a callback keeps an arc_buf cached 4584 * clearing the callback may result in the arc_buf being destroyed. However, 4585 * it will not result in the *last* arc_buf being destroyed, hence the data 4586 * will remain cached in the ARC. We make a copy of the arc buffer here so 4587 * that we can process the callback without holding any locks. 4588 * 4589 * It's possible that the callback is already in the process of being cleared 4590 * by another thread. In this case we can not clear the callback. 4591 * 4592 * Returns B_TRUE if the callback was successfully called and cleared. 4593 */ 4594boolean_t 4595arc_clear_callback(arc_buf_t *buf) 4596{ 4597 arc_buf_hdr_t *hdr; 4598 kmutex_t *hash_lock; 4599 arc_evict_func_t *efunc = buf->b_efunc; 4600 void *private = buf->b_private; 4601 4602 mutex_enter(&buf->b_evict_lock); 4603 hdr = buf->b_hdr; 4604 if (hdr == NULL) { 4605 /* 4606 * We are in arc_do_user_evicts(). 4607 */ 4608 ASSERT(buf->b_data == NULL); 4609 mutex_exit(&buf->b_evict_lock); 4610 return (B_FALSE); 4611 } else if (buf->b_data == NULL) { 4612 /* 4613 * We are on the eviction list; process this buffer now 4614 * but let arc_do_user_evicts() do the reaping. 4615 */ 4616 buf->b_efunc = NULL; 4617 mutex_exit(&buf->b_evict_lock); 4618 VERIFY0(efunc(private)); 4619 return (B_TRUE); 4620 } 4621 hash_lock = HDR_LOCK(hdr); 4622 mutex_enter(hash_lock); 4623 hdr = buf->b_hdr; 4624 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4625 4626 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <, 4627 hdr->b_l1hdr.b_datacnt); 4628 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4629 hdr->b_l1hdr.b_state == arc_mfu); 4630 4631 buf->b_efunc = NULL; 4632 buf->b_private = NULL; 4633 4634 if (hdr->b_l1hdr.b_datacnt > 1) { 4635 mutex_exit(&buf->b_evict_lock); 4636 arc_buf_destroy(buf, TRUE); 4637 } else { 4638 ASSERT(buf == hdr->b_l1hdr.b_buf); 4639 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4640 mutex_exit(&buf->b_evict_lock); 4641 } 4642 4643 mutex_exit(hash_lock); 4644 VERIFY0(efunc(private)); 4645 return (B_TRUE); 4646} 4647 4648/* 4649 * Release this buffer from the cache, making it an anonymous buffer. This 4650 * must be done after a read and prior to modifying the buffer contents. 4651 * If the buffer has more than one reference, we must make 4652 * a new hdr for the buffer. 4653 */ 4654void 4655arc_release(arc_buf_t *buf, void *tag) 4656{ 4657 arc_buf_hdr_t *hdr = buf->b_hdr; 4658 4659 /* 4660 * It would be nice to assert that if it's DMU metadata (level > 4661 * 0 || it's the dnode file), then it must be syncing context. 4662 * But we don't know that information at this level. 4663 */ 4664 4665 mutex_enter(&buf->b_evict_lock); 4666 4667 ASSERT(HDR_HAS_L1HDR(hdr)); 4668 4669 /* 4670 * We don't grab the hash lock prior to this check, because if 4671 * the buffer's header is in the arc_anon state, it won't be 4672 * linked into the hash table. 4673 */ 4674 if (hdr->b_l1hdr.b_state == arc_anon) { 4675 mutex_exit(&buf->b_evict_lock); 4676 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4677 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 4678 ASSERT(!HDR_HAS_L2HDR(hdr)); 4679 ASSERT(BUF_EMPTY(hdr)); 4680 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1); 4681 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 4682 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4683 4684 ASSERT3P(buf->b_efunc, ==, NULL); 4685 ASSERT3P(buf->b_private, ==, NULL); 4686 4687 hdr->b_l1hdr.b_arc_access = 0; 4688 arc_buf_thaw(buf); 4689 4690 return; 4691 } 4692 4693 kmutex_t *hash_lock = HDR_LOCK(hdr); 4694 mutex_enter(hash_lock); 4695 4696 /* 4697 * This assignment is only valid as long as the hash_lock is 4698 * held, we must be careful not to reference state or the 4699 * b_state field after dropping the lock. 4700 */ 4701 arc_state_t *state = hdr->b_l1hdr.b_state; 4702 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4703 ASSERT3P(state, !=, arc_anon); 4704 4705 /* this buffer is not on any list */ 4706 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 4707 4708 if (HDR_HAS_L2HDR(hdr)) { 4709 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4710 4711 /* 4712 * We have to recheck this conditional again now that 4713 * we're holding the l2ad_mtx to prevent a race with 4714 * another thread which might be concurrently calling 4715 * l2arc_evict(). In that case, l2arc_evict() might have 4716 * destroyed the header's L2 portion as we were waiting 4717 * to acquire the l2ad_mtx. 4718 */ 4719 if (HDR_HAS_L2HDR(hdr)) { 4720 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET) 4721 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 4722 hdr->b_l2hdr.b_daddr, 4723 hdr->b_l2hdr.b_asize, 0); 4724 arc_hdr_l2hdr_destroy(hdr); 4725 } 4726 4727 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4728 } 4729 4730 /* 4731 * Do we have more than one buf? 4732 */ 4733 if (hdr->b_l1hdr.b_datacnt > 1) { 4734 arc_buf_hdr_t *nhdr; 4735 arc_buf_t **bufp; 4736 uint64_t blksz = hdr->b_size; 4737 uint64_t spa = hdr->b_spa; 4738 arc_buf_contents_t type = arc_buf_type(hdr); 4739 uint32_t flags = hdr->b_flags; 4740 4741 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 4742 /* 4743 * Pull the data off of this hdr and attach it to 4744 * a new anonymous hdr. 4745 */ 4746 (void) remove_reference(hdr, hash_lock, tag); 4747 bufp = &hdr->b_l1hdr.b_buf; 4748 while (*bufp != buf) 4749 bufp = &(*bufp)->b_next; 4750 *bufp = buf->b_next; 4751 buf->b_next = NULL; 4752 4753 ASSERT3P(state, !=, arc_l2c_only); 4754 4755 (void) refcount_remove_many( 4756 &state->arcs_size, hdr->b_size, buf); 4757 4758 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 4759 ASSERT3P(state, !=, arc_l2c_only); 4760 uint64_t *size = &state->arcs_lsize[type]; 4761 ASSERT3U(*size, >=, hdr->b_size); 4762 atomic_add_64(size, -hdr->b_size); 4763 } 4764 4765 /* 4766 * We're releasing a duplicate user data buffer, update 4767 * our statistics accordingly. 4768 */ 4769 if (HDR_ISTYPE_DATA(hdr)) { 4770 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 4771 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 4772 -hdr->b_size); 4773 } 4774 hdr->b_l1hdr.b_datacnt -= 1; 4775 arc_cksum_verify(buf); 4776#ifdef illumos 4777 arc_buf_unwatch(buf); 4778#endif 4779 4780 mutex_exit(hash_lock); 4781 4782 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 4783 nhdr->b_size = blksz; 4784 nhdr->b_spa = spa; 4785 4786 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 4787 nhdr->b_flags |= arc_bufc_to_flags(type); 4788 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 4789 4790 nhdr->b_l1hdr.b_buf = buf; 4791 nhdr->b_l1hdr.b_datacnt = 1; 4792 nhdr->b_l1hdr.b_state = arc_anon; 4793 nhdr->b_l1hdr.b_arc_access = 0; 4794 nhdr->b_l1hdr.b_tmp_cdata = NULL; 4795 nhdr->b_freeze_cksum = NULL; 4796 4797 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 4798 buf->b_hdr = nhdr; 4799 mutex_exit(&buf->b_evict_lock); 4800 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf); 4801 } else { 4802 mutex_exit(&buf->b_evict_lock); 4803 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 4804 /* protected by hash lock, or hdr is on arc_anon */ 4805 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4806 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4807 arc_change_state(arc_anon, hdr, hash_lock); 4808 hdr->b_l1hdr.b_arc_access = 0; 4809 mutex_exit(hash_lock); 4810 4811 buf_discard_identity(hdr); 4812 arc_buf_thaw(buf); 4813 } 4814 buf->b_efunc = NULL; 4815 buf->b_private = NULL; 4816} 4817 4818int 4819arc_released(arc_buf_t *buf) 4820{ 4821 int released; 4822 4823 mutex_enter(&buf->b_evict_lock); 4824 released = (buf->b_data != NULL && 4825 buf->b_hdr->b_l1hdr.b_state == arc_anon); 4826 mutex_exit(&buf->b_evict_lock); 4827 return (released); 4828} 4829 4830#ifdef ZFS_DEBUG 4831int 4832arc_referenced(arc_buf_t *buf) 4833{ 4834 int referenced; 4835 4836 mutex_enter(&buf->b_evict_lock); 4837 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 4838 mutex_exit(&buf->b_evict_lock); 4839 return (referenced); 4840} 4841#endif 4842 4843static void 4844arc_write_ready(zio_t *zio) 4845{ 4846 arc_write_callback_t *callback = zio->io_private; 4847 arc_buf_t *buf = callback->awcb_buf; 4848 arc_buf_hdr_t *hdr = buf->b_hdr; 4849 4850 ASSERT(HDR_HAS_L1HDR(hdr)); 4851 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 4852 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4853 callback->awcb_ready(zio, buf, callback->awcb_private); 4854 4855 /* 4856 * If the IO is already in progress, then this is a re-write 4857 * attempt, so we need to thaw and re-compute the cksum. 4858 * It is the responsibility of the callback to handle the 4859 * accounting for any re-write attempt. 4860 */ 4861 if (HDR_IO_IN_PROGRESS(hdr)) { 4862 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 4863 if (hdr->b_freeze_cksum != NULL) { 4864 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 4865 hdr->b_freeze_cksum = NULL; 4866 } 4867 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 4868 } 4869 arc_cksum_compute(buf, B_FALSE); 4870 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4871} 4872 4873/* 4874 * The SPA calls this callback for each physical write that happens on behalf 4875 * of a logical write. See the comment in dbuf_write_physdone() for details. 4876 */ 4877static void 4878arc_write_physdone(zio_t *zio) 4879{ 4880 arc_write_callback_t *cb = zio->io_private; 4881 if (cb->awcb_physdone != NULL) 4882 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 4883} 4884 4885static void 4886arc_write_done(zio_t *zio) 4887{ 4888 arc_write_callback_t *callback = zio->io_private; 4889 arc_buf_t *buf = callback->awcb_buf; 4890 arc_buf_hdr_t *hdr = buf->b_hdr; 4891 4892 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4893 4894 if (zio->io_error == 0) { 4895 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 4896 buf_discard_identity(hdr); 4897 } else { 4898 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 4899 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 4900 } 4901 } else { 4902 ASSERT(BUF_EMPTY(hdr)); 4903 } 4904 4905 /* 4906 * If the block to be written was all-zero or compressed enough to be 4907 * embedded in the BP, no write was performed so there will be no 4908 * dva/birth/checksum. The buffer must therefore remain anonymous 4909 * (and uncached). 4910 */ 4911 if (!BUF_EMPTY(hdr)) { 4912 arc_buf_hdr_t *exists; 4913 kmutex_t *hash_lock; 4914 4915 ASSERT(zio->io_error == 0); 4916 4917 arc_cksum_verify(buf); 4918 4919 exists = buf_hash_insert(hdr, &hash_lock); 4920 if (exists != NULL) { 4921 /* 4922 * This can only happen if we overwrite for 4923 * sync-to-convergence, because we remove 4924 * buffers from the hash table when we arc_free(). 4925 */ 4926 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 4927 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4928 panic("bad overwrite, hdr=%p exists=%p", 4929 (void *)hdr, (void *)exists); 4930 ASSERT(refcount_is_zero( 4931 &exists->b_l1hdr.b_refcnt)); 4932 arc_change_state(arc_anon, exists, hash_lock); 4933 mutex_exit(hash_lock); 4934 arc_hdr_destroy(exists); 4935 exists = buf_hash_insert(hdr, &hash_lock); 4936 ASSERT3P(exists, ==, NULL); 4937 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 4938 /* nopwrite */ 4939 ASSERT(zio->io_prop.zp_nopwrite); 4940 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4941 panic("bad nopwrite, hdr=%p exists=%p", 4942 (void *)hdr, (void *)exists); 4943 } else { 4944 /* Dedup */ 4945 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4946 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 4947 ASSERT(BP_GET_DEDUP(zio->io_bp)); 4948 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 4949 } 4950 } 4951 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4952 /* if it's not anon, we are doing a scrub */ 4953 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 4954 arc_access(hdr, hash_lock); 4955 mutex_exit(hash_lock); 4956 } else { 4957 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4958 } 4959 4960 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4961 callback->awcb_done(zio, buf, callback->awcb_private); 4962 4963 kmem_free(callback, sizeof (arc_write_callback_t)); 4964} 4965 4966zio_t * 4967arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 4968 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 4969 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 4970 arc_done_func_t *done, void *private, zio_priority_t priority, 4971 int zio_flags, const zbookmark_phys_t *zb) 4972{ 4973 arc_buf_hdr_t *hdr = buf->b_hdr; 4974 arc_write_callback_t *callback; 4975 zio_t *zio; 4976 4977 ASSERT(ready != NULL); 4978 ASSERT(done != NULL); 4979 ASSERT(!HDR_IO_ERROR(hdr)); 4980 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4981 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4982 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4983 if (l2arc) 4984 hdr->b_flags |= ARC_FLAG_L2CACHE; 4985 if (l2arc_compress) 4986 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4987 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4988 callback->awcb_ready = ready; 4989 callback->awcb_physdone = physdone; 4990 callback->awcb_done = done; 4991 callback->awcb_private = private; 4992 callback->awcb_buf = buf; 4993 4994 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4995 arc_write_ready, arc_write_physdone, arc_write_done, callback, 4996 priority, zio_flags, zb); 4997 4998 return (zio); 4999} 5000 5001static int 5002arc_memory_throttle(uint64_t reserve, uint64_t txg) 5003{ 5004#ifdef _KERNEL 5005 uint64_t available_memory = ptob(freemem); 5006 static uint64_t page_load = 0; 5007 static uint64_t last_txg = 0; 5008 5009#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 5010 available_memory = 5011 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 5012#endif 5013 5014 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 5015 return (0); 5016 5017 if (txg > last_txg) { 5018 last_txg = txg; 5019 page_load = 0; 5020 } 5021 /* 5022 * If we are in pageout, we know that memory is already tight, 5023 * the arc is already going to be evicting, so we just want to 5024 * continue to let page writes occur as quickly as possible. 5025 */ 5026 if (curproc == pageproc) { 5027 if (page_load > MAX(ptob(minfree), available_memory) / 4) 5028 return (SET_ERROR(ERESTART)); 5029 /* Note: reserve is inflated, so we deflate */ 5030 page_load += reserve / 8; 5031 return (0); 5032 } else if (page_load > 0 && arc_reclaim_needed()) { 5033 /* memory is low, delay before restarting */ 5034 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5035 return (SET_ERROR(EAGAIN)); 5036 } 5037 page_load = 0; 5038#endif 5039 return (0); 5040} 5041 5042void 5043arc_tempreserve_clear(uint64_t reserve) 5044{ 5045 atomic_add_64(&arc_tempreserve, -reserve); 5046 ASSERT((int64_t)arc_tempreserve >= 0); 5047} 5048 5049int 5050arc_tempreserve_space(uint64_t reserve, uint64_t txg) 5051{ 5052 int error; 5053 uint64_t anon_size; 5054 5055 if (reserve > arc_c/4 && !arc_no_grow) { 5056 arc_c = MIN(arc_c_max, reserve * 4); 5057 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 5058 } 5059 if (reserve > arc_c) 5060 return (SET_ERROR(ENOMEM)); 5061 5062 /* 5063 * Don't count loaned bufs as in flight dirty data to prevent long 5064 * network delays from blocking transactions that are ready to be 5065 * assigned to a txg. 5066 */ 5067 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 5068 arc_loaned_bytes), 0); 5069 5070 /* 5071 * Writes will, almost always, require additional memory allocations 5072 * in order to compress/encrypt/etc the data. We therefore need to 5073 * make sure that there is sufficient available memory for this. 5074 */ 5075 error = arc_memory_throttle(reserve, txg); 5076 if (error != 0) 5077 return (error); 5078 5079 /* 5080 * Throttle writes when the amount of dirty data in the cache 5081 * gets too large. We try to keep the cache less than half full 5082 * of dirty blocks so that our sync times don't grow too large. 5083 * Note: if two requests come in concurrently, we might let them 5084 * both succeed, when one of them should fail. Not a huge deal. 5085 */ 5086 5087 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 5088 anon_size > arc_c / 4) { 5089 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 5090 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5091 arc_tempreserve>>10, 5092 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 5093 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 5094 reserve>>10, arc_c>>10); 5095 return (SET_ERROR(ERESTART)); 5096 } 5097 atomic_add_64(&arc_tempreserve, reserve); 5098 return (0); 5099} 5100 5101static void 5102arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 5103 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 5104{ 5105 size->value.ui64 = refcount_count(&state->arcs_size); 5106 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA]; 5107 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA]; 5108} 5109 5110static int 5111arc_kstat_update(kstat_t *ksp, int rw) 5112{ 5113 arc_stats_t *as = ksp->ks_data; 5114 5115 if (rw == KSTAT_WRITE) { 5116 return (EACCES); 5117 } else { 5118 arc_kstat_update_state(arc_anon, 5119 &as->arcstat_anon_size, 5120 &as->arcstat_anon_evictable_data, 5121 &as->arcstat_anon_evictable_metadata); 5122 arc_kstat_update_state(arc_mru, 5123 &as->arcstat_mru_size, 5124 &as->arcstat_mru_evictable_data, 5125 &as->arcstat_mru_evictable_metadata); 5126 arc_kstat_update_state(arc_mru_ghost, 5127 &as->arcstat_mru_ghost_size, 5128 &as->arcstat_mru_ghost_evictable_data, 5129 &as->arcstat_mru_ghost_evictable_metadata); 5130 arc_kstat_update_state(arc_mfu, 5131 &as->arcstat_mfu_size, 5132 &as->arcstat_mfu_evictable_data, 5133 &as->arcstat_mfu_evictable_metadata); 5134 arc_kstat_update_state(arc_mfu_ghost, 5135 &as->arcstat_mfu_ghost_size, 5136 &as->arcstat_mfu_ghost_evictable_data, 5137 &as->arcstat_mfu_ghost_evictable_metadata); 5138 } 5139 5140 return (0); 5141} 5142 5143/* 5144 * This function *must* return indices evenly distributed between all 5145 * sublists of the multilist. This is needed due to how the ARC eviction 5146 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 5147 * distributed between all sublists and uses this assumption when 5148 * deciding which sublist to evict from and how much to evict from it. 5149 */ 5150unsigned int 5151arc_state_multilist_index_func(multilist_t *ml, void *obj) 5152{ 5153 arc_buf_hdr_t *hdr = obj; 5154 5155 /* 5156 * We rely on b_dva to generate evenly distributed index 5157 * numbers using buf_hash below. So, as an added precaution, 5158 * let's make sure we never add empty buffers to the arc lists. 5159 */ 5160 ASSERT(!BUF_EMPTY(hdr)); 5161 5162 /* 5163 * The assumption here, is the hash value for a given 5164 * arc_buf_hdr_t will remain constant throughout it's lifetime 5165 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 5166 * Thus, we don't need to store the header's sublist index 5167 * on insertion, as this index can be recalculated on removal. 5168 * 5169 * Also, the low order bits of the hash value are thought to be 5170 * distributed evenly. Otherwise, in the case that the multilist 5171 * has a power of two number of sublists, each sublists' usage 5172 * would not be evenly distributed. 5173 */ 5174 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 5175 multilist_get_num_sublists(ml)); 5176} 5177 5178#ifdef _KERNEL 5179static eventhandler_tag arc_event_lowmem = NULL; 5180 5181static void 5182arc_lowmem(void *arg __unused, int howto __unused) 5183{ 5184 5185 mutex_enter(&arc_reclaim_lock); 5186 /* XXX: Memory deficit should be passed as argument. */ 5187 needfree = btoc(arc_c >> arc_shrink_shift); 5188 DTRACE_PROBE(arc__needfree); 5189 cv_signal(&arc_reclaim_thread_cv); 5190 5191 /* 5192 * It is unsafe to block here in arbitrary threads, because we can come 5193 * here from ARC itself and may hold ARC locks and thus risk a deadlock 5194 * with ARC reclaim thread. 5195 */ 5196 if (curproc == pageproc) 5197 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 5198 mutex_exit(&arc_reclaim_lock); 5199} 5200#endif 5201 5202void 5203arc_init(void) 5204{ 5205 int i, prefetch_tunable_set = 0; 5206 5207 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 5208 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 5209 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 5210 5211 mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 5212 cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL); 5213 5214 /* Convert seconds to clock ticks */ 5215 arc_min_prefetch_lifespan = 1 * hz; 5216 5217 /* Start out with 1/8 of all memory */ 5218 arc_c = kmem_size() / 8; 5219 5220#ifdef illumos 5221#ifdef _KERNEL 5222 /* 5223 * On architectures where the physical memory can be larger 5224 * than the addressable space (intel in 32-bit mode), we may 5225 * need to limit the cache to 1/8 of VM size. 5226 */ 5227 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 5228#endif 5229#endif /* illumos */ 5230 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 5231 arc_c_min = MAX(arc_c / 4, 16 << 20); 5232 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 5233 if (arc_c * 8 >= 1 << 30) 5234 arc_c_max = (arc_c * 8) - (1 << 30); 5235 else 5236 arc_c_max = arc_c_min; 5237 arc_c_max = MAX(arc_c * 5, arc_c_max); 5238 5239#ifdef _KERNEL 5240 /* 5241 * Allow the tunables to override our calculations if they are 5242 * reasonable (ie. over 16MB) 5243 */ 5244 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 5245 arc_c_max = zfs_arc_max; 5246 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 5247 arc_c_min = zfs_arc_min; 5248#endif 5249 5250 arc_c = arc_c_max; 5251 arc_p = (arc_c >> 1); 5252 5253 /* limit meta-data to 1/4 of the arc capacity */ 5254 arc_meta_limit = arc_c_max / 4; 5255 5256 /* Allow the tunable to override if it is reasonable */ 5257 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 5258 arc_meta_limit = zfs_arc_meta_limit; 5259 5260 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 5261 arc_c_min = arc_meta_limit / 2; 5262 5263 if (zfs_arc_meta_min > 0) { 5264 arc_meta_min = zfs_arc_meta_min; 5265 } else { 5266 arc_meta_min = arc_c_min / 2; 5267 } 5268 5269 if (zfs_arc_grow_retry > 0) 5270 arc_grow_retry = zfs_arc_grow_retry; 5271 5272 if (zfs_arc_shrink_shift > 0) 5273 arc_shrink_shift = zfs_arc_shrink_shift; 5274 5275 /* 5276 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 5277 */ 5278 if (arc_no_grow_shift >= arc_shrink_shift) 5279 arc_no_grow_shift = arc_shrink_shift - 1; 5280 5281 if (zfs_arc_p_min_shift > 0) 5282 arc_p_min_shift = zfs_arc_p_min_shift; 5283 5284 if (zfs_arc_num_sublists_per_state < 1) 5285 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1); 5286 5287 /* if kmem_flags are set, lets try to use less memory */ 5288 if (kmem_debugging()) 5289 arc_c = arc_c / 2; 5290 if (arc_c < arc_c_min) 5291 arc_c = arc_c_min; 5292 5293 zfs_arc_min = arc_c_min; 5294 zfs_arc_max = arc_c_max; 5295 5296 arc_anon = &ARC_anon; 5297 arc_mru = &ARC_mru; 5298 arc_mru_ghost = &ARC_mru_ghost; 5299 arc_mfu = &ARC_mfu; 5300 arc_mfu_ghost = &ARC_mfu_ghost; 5301 arc_l2c_only = &ARC_l2c_only; 5302 arc_size = 0; 5303 5304 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 5305 sizeof (arc_buf_hdr_t), 5306 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5307 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5308 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 5309 sizeof (arc_buf_hdr_t), 5310 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5311 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5312 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 5313 sizeof (arc_buf_hdr_t), 5314 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5315 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5316 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 5317 sizeof (arc_buf_hdr_t), 5318 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5319 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5320 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 5321 sizeof (arc_buf_hdr_t), 5322 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5323 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5324 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 5325 sizeof (arc_buf_hdr_t), 5326 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5327 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5328 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 5329 sizeof (arc_buf_hdr_t), 5330 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5331 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5332 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 5333 sizeof (arc_buf_hdr_t), 5334 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5335 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5336 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], 5337 sizeof (arc_buf_hdr_t), 5338 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5339 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5340 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], 5341 sizeof (arc_buf_hdr_t), 5342 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5343 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5344 5345 refcount_create(&arc_anon->arcs_size); 5346 refcount_create(&arc_mru->arcs_size); 5347 refcount_create(&arc_mru_ghost->arcs_size); 5348 refcount_create(&arc_mfu->arcs_size); 5349 refcount_create(&arc_mfu_ghost->arcs_size); 5350 refcount_create(&arc_l2c_only->arcs_size); 5351 5352 buf_init(); 5353 5354 arc_reclaim_thread_exit = FALSE; 5355 arc_user_evicts_thread_exit = FALSE; 5356 arc_eviction_list = NULL; 5357 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 5358 5359 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 5360 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 5361 5362 if (arc_ksp != NULL) { 5363 arc_ksp->ks_data = &arc_stats; 5364 arc_ksp->ks_update = arc_kstat_update; 5365 kstat_install(arc_ksp); 5366 } 5367 5368 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 5369 TS_RUN, minclsyspri); 5370 5371#ifdef _KERNEL 5372 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 5373 EVENTHANDLER_PRI_FIRST); 5374#endif 5375 5376 (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0, 5377 TS_RUN, minclsyspri); 5378 5379 arc_dead = FALSE; 5380 arc_warm = B_FALSE; 5381 5382 /* 5383 * Calculate maximum amount of dirty data per pool. 5384 * 5385 * If it has been set by /etc/system, take that. 5386 * Otherwise, use a percentage of physical memory defined by 5387 * zfs_dirty_data_max_percent (default 10%) with a cap at 5388 * zfs_dirty_data_max_max (default 4GB). 5389 */ 5390 if (zfs_dirty_data_max == 0) { 5391 zfs_dirty_data_max = ptob(physmem) * 5392 zfs_dirty_data_max_percent / 100; 5393 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 5394 zfs_dirty_data_max_max); 5395 } 5396 5397#ifdef _KERNEL 5398 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 5399 prefetch_tunable_set = 1; 5400 5401#ifdef __i386__ 5402 if (prefetch_tunable_set == 0) { 5403 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 5404 "-- to enable,\n"); 5405 printf(" add \"vfs.zfs.prefetch_disable=0\" " 5406 "to /boot/loader.conf.\n"); 5407 zfs_prefetch_disable = 1; 5408 } 5409#else 5410 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 5411 prefetch_tunable_set == 0) { 5412 printf("ZFS NOTICE: Prefetch is disabled by default if less " 5413 "than 4GB of RAM is present;\n" 5414 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 5415 "to /boot/loader.conf.\n"); 5416 zfs_prefetch_disable = 1; 5417 } 5418#endif 5419 /* Warn about ZFS memory and address space requirements. */ 5420 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 5421 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 5422 "expect unstable behavior.\n"); 5423 } 5424 if (kmem_size() < 512 * (1 << 20)) { 5425 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 5426 "expect unstable behavior.\n"); 5427 printf(" Consider tuning vm.kmem_size and " 5428 "vm.kmem_size_max\n"); 5429 printf(" in /boot/loader.conf.\n"); 5430 } 5431#endif 5432} 5433 5434void 5435arc_fini(void) 5436{ 5437 mutex_enter(&arc_reclaim_lock); 5438 arc_reclaim_thread_exit = TRUE; 5439 /* 5440 * The reclaim thread will set arc_reclaim_thread_exit back to 5441 * FALSE when it is finished exiting; we're waiting for that. 5442 */ 5443 while (arc_reclaim_thread_exit) { 5444 cv_signal(&arc_reclaim_thread_cv); 5445 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 5446 } 5447 mutex_exit(&arc_reclaim_lock); 5448 5449 mutex_enter(&arc_user_evicts_lock); 5450 arc_user_evicts_thread_exit = TRUE; 5451 /* 5452 * The user evicts thread will set arc_user_evicts_thread_exit 5453 * to FALSE when it is finished exiting; we're waiting for that. 5454 */ 5455 while (arc_user_evicts_thread_exit) { 5456 cv_signal(&arc_user_evicts_cv); 5457 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock); 5458 } 5459 mutex_exit(&arc_user_evicts_lock); 5460 5461 /* Use TRUE to ensure *all* buffers are evicted */ 5462 arc_flush(NULL, TRUE); 5463 5464 arc_dead = TRUE; 5465 5466 if (arc_ksp != NULL) { 5467 kstat_delete(arc_ksp); 5468 arc_ksp = NULL; 5469 } 5470 5471 mutex_destroy(&arc_reclaim_lock); 5472 cv_destroy(&arc_reclaim_thread_cv); 5473 cv_destroy(&arc_reclaim_waiters_cv); 5474 5475 mutex_destroy(&arc_user_evicts_lock); 5476 cv_destroy(&arc_user_evicts_cv); 5477 5478 refcount_destroy(&arc_anon->arcs_size); 5479 refcount_destroy(&arc_mru->arcs_size); 5480 refcount_destroy(&arc_mru_ghost->arcs_size); 5481 refcount_destroy(&arc_mfu->arcs_size); 5482 refcount_destroy(&arc_mfu_ghost->arcs_size); 5483 refcount_destroy(&arc_l2c_only->arcs_size); 5484 5485 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 5486 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 5487 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 5488 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 5489 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 5490 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 5491 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 5492 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 5493 5494 buf_fini(); 5495 5496 ASSERT0(arc_loaned_bytes); 5497 5498#ifdef _KERNEL 5499 if (arc_event_lowmem != NULL) 5500 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 5501#endif 5502} 5503 5504/* 5505 * Level 2 ARC 5506 * 5507 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 5508 * It uses dedicated storage devices to hold cached data, which are populated 5509 * using large infrequent writes. The main role of this cache is to boost 5510 * the performance of random read workloads. The intended L2ARC devices 5511 * include short-stroked disks, solid state disks, and other media with 5512 * substantially faster read latency than disk. 5513 * 5514 * +-----------------------+ 5515 * | ARC | 5516 * +-----------------------+ 5517 * | ^ ^ 5518 * | | | 5519 * l2arc_feed_thread() arc_read() 5520 * | | | 5521 * | l2arc read | 5522 * V | | 5523 * +---------------+ | 5524 * | L2ARC | | 5525 * +---------------+ | 5526 * | ^ | 5527 * l2arc_write() | | 5528 * | | | 5529 * V | | 5530 * +-------+ +-------+ 5531 * | vdev | | vdev | 5532 * | cache | | cache | 5533 * +-------+ +-------+ 5534 * +=========+ .-----. 5535 * : L2ARC : |-_____-| 5536 * : devices : | Disks | 5537 * +=========+ `-_____-' 5538 * 5539 * Read requests are satisfied from the following sources, in order: 5540 * 5541 * 1) ARC 5542 * 2) vdev cache of L2ARC devices 5543 * 3) L2ARC devices 5544 * 4) vdev cache of disks 5545 * 5) disks 5546 * 5547 * Some L2ARC device types exhibit extremely slow write performance. 5548 * To accommodate for this there are some significant differences between 5549 * the L2ARC and traditional cache design: 5550 * 5551 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 5552 * the ARC behave as usual, freeing buffers and placing headers on ghost 5553 * lists. The ARC does not send buffers to the L2ARC during eviction as 5554 * this would add inflated write latencies for all ARC memory pressure. 5555 * 5556 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 5557 * It does this by periodically scanning buffers from the eviction-end of 5558 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 5559 * not already there. It scans until a headroom of buffers is satisfied, 5560 * which itself is a buffer for ARC eviction. If a compressible buffer is 5561 * found during scanning and selected for writing to an L2ARC device, we 5562 * temporarily boost scanning headroom during the next scan cycle to make 5563 * sure we adapt to compression effects (which might significantly reduce 5564 * the data volume we write to L2ARC). The thread that does this is 5565 * l2arc_feed_thread(), illustrated below; example sizes are included to 5566 * provide a better sense of ratio than this diagram: 5567 * 5568 * head --> tail 5569 * +---------------------+----------+ 5570 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 5571 * +---------------------+----------+ | o L2ARC eligible 5572 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 5573 * +---------------------+----------+ | 5574 * 15.9 Gbytes ^ 32 Mbytes | 5575 * headroom | 5576 * l2arc_feed_thread() 5577 * | 5578 * l2arc write hand <--[oooo]--' 5579 * | 8 Mbyte 5580 * | write max 5581 * V 5582 * +==============================+ 5583 * L2ARC dev |####|#|###|###| |####| ... | 5584 * +==============================+ 5585 * 32 Gbytes 5586 * 5587 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 5588 * evicted, then the L2ARC has cached a buffer much sooner than it probably 5589 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 5590 * safe to say that this is an uncommon case, since buffers at the end of 5591 * the ARC lists have moved there due to inactivity. 5592 * 5593 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 5594 * then the L2ARC simply misses copying some buffers. This serves as a 5595 * pressure valve to prevent heavy read workloads from both stalling the ARC 5596 * with waits and clogging the L2ARC with writes. This also helps prevent 5597 * the potential for the L2ARC to churn if it attempts to cache content too 5598 * quickly, such as during backups of the entire pool. 5599 * 5600 * 5. After system boot and before the ARC has filled main memory, there are 5601 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 5602 * lists can remain mostly static. Instead of searching from tail of these 5603 * lists as pictured, the l2arc_feed_thread() will search from the list heads 5604 * for eligible buffers, greatly increasing its chance of finding them. 5605 * 5606 * The L2ARC device write speed is also boosted during this time so that 5607 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 5608 * there are no L2ARC reads, and no fear of degrading read performance 5609 * through increased writes. 5610 * 5611 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 5612 * the vdev queue can aggregate them into larger and fewer writes. Each 5613 * device is written to in a rotor fashion, sweeping writes through 5614 * available space then repeating. 5615 * 5616 * 7. The L2ARC does not store dirty content. It never needs to flush 5617 * write buffers back to disk based storage. 5618 * 5619 * 8. If an ARC buffer is written (and dirtied) which also exists in the 5620 * L2ARC, the now stale L2ARC buffer is immediately dropped. 5621 * 5622 * The performance of the L2ARC can be tweaked by a number of tunables, which 5623 * may be necessary for different workloads: 5624 * 5625 * l2arc_write_max max write bytes per interval 5626 * l2arc_write_boost extra write bytes during device warmup 5627 * l2arc_noprefetch skip caching prefetched buffers 5628 * l2arc_headroom number of max device writes to precache 5629 * l2arc_headroom_boost when we find compressed buffers during ARC 5630 * scanning, we multiply headroom by this 5631 * percentage factor for the next scan cycle, 5632 * since more compressed buffers are likely to 5633 * be present 5634 * l2arc_feed_secs seconds between L2ARC writing 5635 * 5636 * Tunables may be removed or added as future performance improvements are 5637 * integrated, and also may become zpool properties. 5638 * 5639 * There are three key functions that control how the L2ARC warms up: 5640 * 5641 * l2arc_write_eligible() check if a buffer is eligible to cache 5642 * l2arc_write_size() calculate how much to write 5643 * l2arc_write_interval() calculate sleep delay between writes 5644 * 5645 * These three functions determine what to write, how much, and how quickly 5646 * to send writes. 5647 */ 5648 5649static boolean_t 5650l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 5651{ 5652 /* 5653 * A buffer is *not* eligible for the L2ARC if it: 5654 * 1. belongs to a different spa. 5655 * 2. is already cached on the L2ARC. 5656 * 3. has an I/O in progress (it may be an incomplete read). 5657 * 4. is flagged not eligible (zfs property). 5658 */ 5659 if (hdr->b_spa != spa_guid) { 5660 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 5661 return (B_FALSE); 5662 } 5663 if (HDR_HAS_L2HDR(hdr)) { 5664 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 5665 return (B_FALSE); 5666 } 5667 if (HDR_IO_IN_PROGRESS(hdr)) { 5668 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 5669 return (B_FALSE); 5670 } 5671 if (!HDR_L2CACHE(hdr)) { 5672 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 5673 return (B_FALSE); 5674 } 5675 5676 return (B_TRUE); 5677} 5678 5679static uint64_t 5680l2arc_write_size(void) 5681{ 5682 uint64_t size; 5683 5684 /* 5685 * Make sure our globals have meaningful values in case the user 5686 * altered them. 5687 */ 5688 size = l2arc_write_max; 5689 if (size == 0) { 5690 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 5691 "be greater than zero, resetting it to the default (%d)", 5692 L2ARC_WRITE_SIZE); 5693 size = l2arc_write_max = L2ARC_WRITE_SIZE; 5694 } 5695 5696 if (arc_warm == B_FALSE) 5697 size += l2arc_write_boost; 5698 5699 return (size); 5700 5701} 5702 5703static clock_t 5704l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 5705{ 5706 clock_t interval, next, now; 5707 5708 /* 5709 * If the ARC lists are busy, increase our write rate; if the 5710 * lists are stale, idle back. This is achieved by checking 5711 * how much we previously wrote - if it was more than half of 5712 * what we wanted, schedule the next write much sooner. 5713 */ 5714 if (l2arc_feed_again && wrote > (wanted / 2)) 5715 interval = (hz * l2arc_feed_min_ms) / 1000; 5716 else 5717 interval = hz * l2arc_feed_secs; 5718 5719 now = ddi_get_lbolt(); 5720 next = MAX(now, MIN(now + interval, began + interval)); 5721 5722 return (next); 5723} 5724 5725/* 5726 * Cycle through L2ARC devices. This is how L2ARC load balances. 5727 * If a device is returned, this also returns holding the spa config lock. 5728 */ 5729static l2arc_dev_t * 5730l2arc_dev_get_next(void) 5731{ 5732 l2arc_dev_t *first, *next = NULL; 5733 5734 /* 5735 * Lock out the removal of spas (spa_namespace_lock), then removal 5736 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 5737 * both locks will be dropped and a spa config lock held instead. 5738 */ 5739 mutex_enter(&spa_namespace_lock); 5740 mutex_enter(&l2arc_dev_mtx); 5741 5742 /* if there are no vdevs, there is nothing to do */ 5743 if (l2arc_ndev == 0) 5744 goto out; 5745 5746 first = NULL; 5747 next = l2arc_dev_last; 5748 do { 5749 /* loop around the list looking for a non-faulted vdev */ 5750 if (next == NULL) { 5751 next = list_head(l2arc_dev_list); 5752 } else { 5753 next = list_next(l2arc_dev_list, next); 5754 if (next == NULL) 5755 next = list_head(l2arc_dev_list); 5756 } 5757 5758 /* if we have come back to the start, bail out */ 5759 if (first == NULL) 5760 first = next; 5761 else if (next == first) 5762 break; 5763 5764 } while (vdev_is_dead(next->l2ad_vdev)); 5765 5766 /* if we were unable to find any usable vdevs, return NULL */ 5767 if (vdev_is_dead(next->l2ad_vdev)) 5768 next = NULL; 5769 5770 l2arc_dev_last = next; 5771 5772out: 5773 mutex_exit(&l2arc_dev_mtx); 5774 5775 /* 5776 * Grab the config lock to prevent the 'next' device from being 5777 * removed while we are writing to it. 5778 */ 5779 if (next != NULL) 5780 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 5781 mutex_exit(&spa_namespace_lock); 5782 5783 return (next); 5784} 5785 5786/* 5787 * Free buffers that were tagged for destruction. 5788 */ 5789static void 5790l2arc_do_free_on_write() 5791{ 5792 list_t *buflist; 5793 l2arc_data_free_t *df, *df_prev; 5794 5795 mutex_enter(&l2arc_free_on_write_mtx); 5796 buflist = l2arc_free_on_write; 5797 5798 for (df = list_tail(buflist); df; df = df_prev) { 5799 df_prev = list_prev(buflist, df); 5800 ASSERT(df->l2df_data != NULL); 5801 ASSERT(df->l2df_func != NULL); 5802 df->l2df_func(df->l2df_data, df->l2df_size); 5803 list_remove(buflist, df); 5804 kmem_free(df, sizeof (l2arc_data_free_t)); 5805 } 5806 5807 mutex_exit(&l2arc_free_on_write_mtx); 5808} 5809 5810/* 5811 * A write to a cache device has completed. Update all headers to allow 5812 * reads from these buffers to begin. 5813 */ 5814static void 5815l2arc_write_done(zio_t *zio) 5816{ 5817 l2arc_write_callback_t *cb; 5818 l2arc_dev_t *dev; 5819 list_t *buflist; 5820 arc_buf_hdr_t *head, *hdr, *hdr_prev; 5821 kmutex_t *hash_lock; 5822 int64_t bytes_dropped = 0; 5823 5824 cb = zio->io_private; 5825 ASSERT(cb != NULL); 5826 dev = cb->l2wcb_dev; 5827 ASSERT(dev != NULL); 5828 head = cb->l2wcb_head; 5829 ASSERT(head != NULL); 5830 buflist = &dev->l2ad_buflist; 5831 ASSERT(buflist != NULL); 5832 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 5833 l2arc_write_callback_t *, cb); 5834 5835 if (zio->io_error != 0) 5836 ARCSTAT_BUMP(arcstat_l2_writes_error); 5837 5838 /* 5839 * All writes completed, or an error was hit. 5840 */ 5841top: 5842 mutex_enter(&dev->l2ad_mtx); 5843 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 5844 hdr_prev = list_prev(buflist, hdr); 5845 5846 hash_lock = HDR_LOCK(hdr); 5847 5848 /* 5849 * We cannot use mutex_enter or else we can deadlock 5850 * with l2arc_write_buffers (due to swapping the order 5851 * the hash lock and l2ad_mtx are taken). 5852 */ 5853 if (!mutex_tryenter(hash_lock)) { 5854 /* 5855 * Missed the hash lock. We must retry so we 5856 * don't leave the ARC_FLAG_L2_WRITING bit set. 5857 */ 5858 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 5859 5860 /* 5861 * We don't want to rescan the headers we've 5862 * already marked as having been written out, so 5863 * we reinsert the head node so we can pick up 5864 * where we left off. 5865 */ 5866 list_remove(buflist, head); 5867 list_insert_after(buflist, hdr, head); 5868 5869 mutex_exit(&dev->l2ad_mtx); 5870 5871 /* 5872 * We wait for the hash lock to become available 5873 * to try and prevent busy waiting, and increase 5874 * the chance we'll be able to acquire the lock 5875 * the next time around. 5876 */ 5877 mutex_enter(hash_lock); 5878 mutex_exit(hash_lock); 5879 goto top; 5880 } 5881 5882 /* 5883 * We could not have been moved into the arc_l2c_only 5884 * state while in-flight due to our ARC_FLAG_L2_WRITING 5885 * bit being set. Let's just ensure that's being enforced. 5886 */ 5887 ASSERT(HDR_HAS_L1HDR(hdr)); 5888 5889 /* 5890 * We may have allocated a buffer for L2ARC compression, 5891 * we must release it to avoid leaking this data. 5892 */ 5893 l2arc_release_cdata_buf(hdr); 5894 5895 if (zio->io_error != 0) { 5896 /* 5897 * Error - drop L2ARC entry. 5898 */ 5899 list_remove(buflist, hdr); 5900 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 5901 hdr->b_l2hdr.b_daddr, hdr->b_l2hdr.b_asize, 0); 5902 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 5903 5904 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 5905 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5906 5907 bytes_dropped += hdr->b_l2hdr.b_asize; 5908 (void) refcount_remove_many(&dev->l2ad_alloc, 5909 hdr->b_l2hdr.b_asize, hdr); 5910 } 5911 5912 /* 5913 * Allow ARC to begin reads and ghost list evictions to 5914 * this L2ARC entry. 5915 */ 5916 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5917 5918 mutex_exit(hash_lock); 5919 } 5920 5921 atomic_inc_64(&l2arc_writes_done); 5922 list_remove(buflist, head); 5923 ASSERT(!HDR_HAS_L1HDR(head)); 5924 kmem_cache_free(hdr_l2only_cache, head); 5925 mutex_exit(&dev->l2ad_mtx); 5926 5927 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 5928 5929 l2arc_do_free_on_write(); 5930 5931 kmem_free(cb, sizeof (l2arc_write_callback_t)); 5932} 5933 5934/* 5935 * A read to a cache device completed. Validate buffer contents before 5936 * handing over to the regular ARC routines. 5937 */ 5938static void 5939l2arc_read_done(zio_t *zio) 5940{ 5941 l2arc_read_callback_t *cb; 5942 arc_buf_hdr_t *hdr; 5943 arc_buf_t *buf; 5944 kmutex_t *hash_lock; 5945 int equal; 5946 5947 ASSERT(zio->io_vd != NULL); 5948 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 5949 5950 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 5951 5952 cb = zio->io_private; 5953 ASSERT(cb != NULL); 5954 buf = cb->l2rcb_buf; 5955 ASSERT(buf != NULL); 5956 5957 hash_lock = HDR_LOCK(buf->b_hdr); 5958 mutex_enter(hash_lock); 5959 hdr = buf->b_hdr; 5960 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5961 5962 /* 5963 * If the buffer was compressed, decompress it first. 5964 */ 5965 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 5966 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 5967 ASSERT(zio->io_data != NULL); 5968 5969 /* 5970 * Check this survived the L2ARC journey. 5971 */ 5972 equal = arc_cksum_equal(buf); 5973 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 5974 mutex_exit(hash_lock); 5975 zio->io_private = buf; 5976 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 5977 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 5978 arc_read_done(zio); 5979 } else { 5980 mutex_exit(hash_lock); 5981 /* 5982 * Buffer didn't survive caching. Increment stats and 5983 * reissue to the original storage device. 5984 */ 5985 if (zio->io_error != 0) { 5986 ARCSTAT_BUMP(arcstat_l2_io_error); 5987 } else { 5988 zio->io_error = SET_ERROR(EIO); 5989 } 5990 if (!equal) 5991 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 5992 5993 /* 5994 * If there's no waiter, issue an async i/o to the primary 5995 * storage now. If there *is* a waiter, the caller must 5996 * issue the i/o in a context where it's OK to block. 5997 */ 5998 if (zio->io_waiter == NULL) { 5999 zio_t *pio = zio_unique_parent(zio); 6000 6001 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6002 6003 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 6004 buf->b_data, zio->io_size, arc_read_done, buf, 6005 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 6006 } 6007 } 6008 6009 kmem_free(cb, sizeof (l2arc_read_callback_t)); 6010} 6011 6012/* 6013 * This is the list priority from which the L2ARC will search for pages to 6014 * cache. This is used within loops (0..3) to cycle through lists in the 6015 * desired order. This order can have a significant effect on cache 6016 * performance. 6017 * 6018 * Currently the metadata lists are hit first, MFU then MRU, followed by 6019 * the data lists. This function returns a locked list, and also returns 6020 * the lock pointer. 6021 */ 6022static multilist_sublist_t * 6023l2arc_sublist_lock(int list_num) 6024{ 6025 multilist_t *ml = NULL; 6026 unsigned int idx; 6027 6028 ASSERT(list_num >= 0 && list_num <= 3); 6029 6030 switch (list_num) { 6031 case 0: 6032 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6033 break; 6034 case 1: 6035 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 6036 break; 6037 case 2: 6038 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 6039 break; 6040 case 3: 6041 ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; 6042 break; 6043 } 6044 6045 /* 6046 * Return a randomly-selected sublist. This is acceptable 6047 * because the caller feeds only a little bit of data for each 6048 * call (8MB). Subsequent calls will result in different 6049 * sublists being selected. 6050 */ 6051 idx = multilist_get_random_index(ml); 6052 return (multilist_sublist_lock(ml, idx)); 6053} 6054 6055/* 6056 * Evict buffers from the device write hand to the distance specified in 6057 * bytes. This distance may span populated buffers, it may span nothing. 6058 * This is clearing a region on the L2ARC device ready for writing. 6059 * If the 'all' boolean is set, every buffer is evicted. 6060 */ 6061static void 6062l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6063{ 6064 list_t *buflist; 6065 arc_buf_hdr_t *hdr, *hdr_prev; 6066 kmutex_t *hash_lock; 6067 uint64_t taddr; 6068 6069 buflist = &dev->l2ad_buflist; 6070 6071 if (!all && dev->l2ad_first) { 6072 /* 6073 * This is the first sweep through the device. There is 6074 * nothing to evict. 6075 */ 6076 return; 6077 } 6078 6079 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6080 /* 6081 * When nearing the end of the device, evict to the end 6082 * before the device write hand jumps to the start. 6083 */ 6084 taddr = dev->l2ad_end; 6085 } else { 6086 taddr = dev->l2ad_hand + distance; 6087 } 6088 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6089 uint64_t, taddr, boolean_t, all); 6090 6091top: 6092 mutex_enter(&dev->l2ad_mtx); 6093 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 6094 hdr_prev = list_prev(buflist, hdr); 6095 6096 hash_lock = HDR_LOCK(hdr); 6097 6098 /* 6099 * We cannot use mutex_enter or else we can deadlock 6100 * with l2arc_write_buffers (due to swapping the order 6101 * the hash lock and l2ad_mtx are taken). 6102 */ 6103 if (!mutex_tryenter(hash_lock)) { 6104 /* 6105 * Missed the hash lock. Retry. 6106 */ 6107 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 6108 mutex_exit(&dev->l2ad_mtx); 6109 mutex_enter(hash_lock); 6110 mutex_exit(hash_lock); 6111 goto top; 6112 } 6113 6114 if (HDR_L2_WRITE_HEAD(hdr)) { 6115 /* 6116 * We hit a write head node. Leave it for 6117 * l2arc_write_done(). 6118 */ 6119 list_remove(buflist, hdr); 6120 mutex_exit(hash_lock); 6121 continue; 6122 } 6123 6124 if (!all && HDR_HAS_L2HDR(hdr) && 6125 (hdr->b_l2hdr.b_daddr > taddr || 6126 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6127 /* 6128 * We've evicted to the target address, 6129 * or the end of the device. 6130 */ 6131 mutex_exit(hash_lock); 6132 break; 6133 } 6134 6135 ASSERT(HDR_HAS_L2HDR(hdr)); 6136 if (!HDR_HAS_L1HDR(hdr)) { 6137 ASSERT(!HDR_L2_READING(hdr)); 6138 /* 6139 * This doesn't exist in the ARC. Destroy. 6140 * arc_hdr_destroy() will call list_remove() 6141 * and decrement arcstat_l2_size. 6142 */ 6143 arc_change_state(arc_anon, hdr, hash_lock); 6144 arc_hdr_destroy(hdr); 6145 } else { 6146 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 6147 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 6148 /* 6149 * Invalidate issued or about to be issued 6150 * reads, since we may be about to write 6151 * over this location. 6152 */ 6153 if (HDR_L2_READING(hdr)) { 6154 ARCSTAT_BUMP(arcstat_l2_evict_reading); 6155 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 6156 } 6157 6158 /* Ensure this header has finished being written */ 6159 ASSERT(!HDR_L2_WRITING(hdr)); 6160 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 6161 6162 arc_hdr_l2hdr_destroy(hdr); 6163 } 6164 mutex_exit(hash_lock); 6165 } 6166 mutex_exit(&dev->l2ad_mtx); 6167} 6168 6169/* 6170 * Find and write ARC buffers to the L2ARC device. 6171 * 6172 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6173 * for reading until they have completed writing. 6174 * The headroom_boost is an in-out parameter used to maintain headroom boost 6175 * state between calls to this function. 6176 * 6177 * Returns the number of bytes actually written (which may be smaller than 6178 * the delta by which the device hand has changed due to alignment). 6179 */ 6180static uint64_t 6181l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 6182 boolean_t *headroom_boost) 6183{ 6184 arc_buf_hdr_t *hdr, *hdr_prev, *head; 6185 uint64_t write_asize, write_psize, write_sz, headroom, 6186 buf_compress_minsz; 6187 void *buf_data; 6188 boolean_t full; 6189 l2arc_write_callback_t *cb; 6190 zio_t *pio, *wzio; 6191 uint64_t guid = spa_load_guid(spa); 6192 const boolean_t do_headroom_boost = *headroom_boost; 6193 int try; 6194 6195 ASSERT(dev->l2ad_vdev != NULL); 6196 6197 /* Lower the flag now, we might want to raise it again later. */ 6198 *headroom_boost = B_FALSE; 6199 6200 pio = NULL; 6201 write_sz = write_asize = write_psize = 0; 6202 full = B_FALSE; 6203 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 6204 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 6205 head->b_flags |= ARC_FLAG_HAS_L2HDR; 6206 6207 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 6208 /* 6209 * We will want to try to compress buffers that are at least 2x the 6210 * device sector size. 6211 */ 6212 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 6213 6214 /* 6215 * Copy buffers for L2ARC writing. 6216 */ 6217 for (try = 0; try <= 3; try++) { 6218 multilist_sublist_t *mls = l2arc_sublist_lock(try); 6219 uint64_t passed_sz = 0; 6220 6221 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 6222 6223 /* 6224 * L2ARC fast warmup. 6225 * 6226 * Until the ARC is warm and starts to evict, read from the 6227 * head of the ARC lists rather than the tail. 6228 */ 6229 if (arc_warm == B_FALSE) 6230 hdr = multilist_sublist_head(mls); 6231 else 6232 hdr = multilist_sublist_tail(mls); 6233 if (hdr == NULL) 6234 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 6235 6236 headroom = target_sz * l2arc_headroom; 6237 if (do_headroom_boost) 6238 headroom = (headroom * l2arc_headroom_boost) / 100; 6239 6240 for (; hdr; hdr = hdr_prev) { 6241 kmutex_t *hash_lock; 6242 uint64_t buf_sz; 6243 6244 if (arc_warm == B_FALSE) 6245 hdr_prev = multilist_sublist_next(mls, hdr); 6246 else 6247 hdr_prev = multilist_sublist_prev(mls, hdr); 6248 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 6249 6250 hash_lock = HDR_LOCK(hdr); 6251 if (!mutex_tryenter(hash_lock)) { 6252 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 6253 /* 6254 * Skip this buffer rather than waiting. 6255 */ 6256 continue; 6257 } 6258 6259 passed_sz += hdr->b_size; 6260 if (passed_sz > headroom) { 6261 /* 6262 * Searched too far. 6263 */ 6264 mutex_exit(hash_lock); 6265 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 6266 break; 6267 } 6268 6269 if (!l2arc_write_eligible(guid, hdr)) { 6270 mutex_exit(hash_lock); 6271 continue; 6272 } 6273 6274 if ((write_sz + hdr->b_size) > target_sz) { 6275 full = B_TRUE; 6276 mutex_exit(hash_lock); 6277 ARCSTAT_BUMP(arcstat_l2_write_full); 6278 break; 6279 } 6280 6281 if (pio == NULL) { 6282 /* 6283 * Insert a dummy header on the buflist so 6284 * l2arc_write_done() can find where the 6285 * write buffers begin without searching. 6286 */ 6287 mutex_enter(&dev->l2ad_mtx); 6288 list_insert_head(&dev->l2ad_buflist, head); 6289 mutex_exit(&dev->l2ad_mtx); 6290 6291 cb = kmem_alloc( 6292 sizeof (l2arc_write_callback_t), KM_SLEEP); 6293 cb->l2wcb_dev = dev; 6294 cb->l2wcb_head = head; 6295 pio = zio_root(spa, l2arc_write_done, cb, 6296 ZIO_FLAG_CANFAIL); 6297 ARCSTAT_BUMP(arcstat_l2_write_pios); 6298 } 6299 6300 /* 6301 * Create and add a new L2ARC header. 6302 */ 6303 hdr->b_l2hdr.b_dev = dev; 6304 hdr->b_flags |= ARC_FLAG_L2_WRITING; 6305 /* 6306 * Temporarily stash the data buffer in b_tmp_cdata. 6307 * The subsequent write step will pick it up from 6308 * there. This is because can't access b_l1hdr.b_buf 6309 * without holding the hash_lock, which we in turn 6310 * can't access without holding the ARC list locks 6311 * (which we want to avoid during compression/writing). 6312 */ 6313 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 6314 hdr->b_l2hdr.b_asize = hdr->b_size; 6315 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data; 6316 6317 /* 6318 * Explicitly set the b_daddr field to a known 6319 * value which means "invalid address". This 6320 * enables us to differentiate which stage of 6321 * l2arc_write_buffers() the particular header 6322 * is in (e.g. this loop, or the one below). 6323 * ARC_FLAG_L2_WRITING is not enough to make 6324 * this distinction, and we need to know in 6325 * order to do proper l2arc vdev accounting in 6326 * arc_release() and arc_hdr_destroy(). 6327 * 6328 * Note, we can't use a new flag to distinguish 6329 * the two stages because we don't hold the 6330 * header's hash_lock below, in the second stage 6331 * of this function. Thus, we can't simply 6332 * change the b_flags field to denote that the 6333 * IO has been sent. We can change the b_daddr 6334 * field of the L2 portion, though, since we'll 6335 * be holding the l2ad_mtx; which is why we're 6336 * using it to denote the header's state change. 6337 */ 6338 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET; 6339 6340 buf_sz = hdr->b_size; 6341 hdr->b_flags |= ARC_FLAG_HAS_L2HDR; 6342 6343 mutex_enter(&dev->l2ad_mtx); 6344 list_insert_head(&dev->l2ad_buflist, hdr); 6345 mutex_exit(&dev->l2ad_mtx); 6346 6347 /* 6348 * Compute and store the buffer cksum before 6349 * writing. On debug the cksum is verified first. 6350 */ 6351 arc_cksum_verify(hdr->b_l1hdr.b_buf); 6352 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE); 6353 6354 mutex_exit(hash_lock); 6355 6356 write_sz += buf_sz; 6357 } 6358 6359 multilist_sublist_unlock(mls); 6360 6361 if (full == B_TRUE) 6362 break; 6363 } 6364 6365 /* No buffers selected for writing? */ 6366 if (pio == NULL) { 6367 ASSERT0(write_sz); 6368 ASSERT(!HDR_HAS_L1HDR(head)); 6369 kmem_cache_free(hdr_l2only_cache, head); 6370 return (0); 6371 } 6372 6373 mutex_enter(&dev->l2ad_mtx); 6374 6375 /* 6376 * Now start writing the buffers. We're starting at the write head 6377 * and work backwards, retracing the course of the buffer selector 6378 * loop above. 6379 */ 6380 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr; 6381 hdr = list_prev(&dev->l2ad_buflist, hdr)) { 6382 uint64_t buf_sz; 6383 6384 /* 6385 * We rely on the L1 portion of the header below, so 6386 * it's invalid for this header to have been evicted out 6387 * of the ghost cache, prior to being written out. The 6388 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 6389 */ 6390 ASSERT(HDR_HAS_L1HDR(hdr)); 6391 6392 /* 6393 * We shouldn't need to lock the buffer here, since we flagged 6394 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 6395 * take care to only access its L2 cache parameters. In 6396 * particular, hdr->l1hdr.b_buf may be invalid by now due to 6397 * ARC eviction. 6398 */ 6399 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 6400 6401 if ((HDR_L2COMPRESS(hdr)) && 6402 hdr->b_l2hdr.b_asize >= buf_compress_minsz) { 6403 if (l2arc_compress_buf(hdr)) { 6404 /* 6405 * If compression succeeded, enable headroom 6406 * boost on the next scan cycle. 6407 */ 6408 *headroom_boost = B_TRUE; 6409 } 6410 } 6411 6412 /* 6413 * Pick up the buffer data we had previously stashed away 6414 * (and now potentially also compressed). 6415 */ 6416 buf_data = hdr->b_l1hdr.b_tmp_cdata; 6417 buf_sz = hdr->b_l2hdr.b_asize; 6418 6419 /* 6420 * We need to do this regardless if buf_sz is zero or 6421 * not, otherwise, when this l2hdr is evicted we'll 6422 * remove a reference that was never added. 6423 */ 6424 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr); 6425 6426 /* Compression may have squashed the buffer to zero length. */ 6427 if (buf_sz != 0) { 6428 uint64_t buf_p_sz; 6429 6430 wzio = zio_write_phys(pio, dev->l2ad_vdev, 6431 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 6432 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 6433 ZIO_FLAG_CANFAIL, B_FALSE); 6434 6435 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 6436 zio_t *, wzio); 6437 (void) zio_nowait(wzio); 6438 6439 write_asize += buf_sz; 6440 6441 /* 6442 * Keep the clock hand suitably device-aligned. 6443 */ 6444 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 6445 write_psize += buf_p_sz; 6446 dev->l2ad_hand += buf_p_sz; 6447 } 6448 } 6449 6450 mutex_exit(&dev->l2ad_mtx); 6451 6452 ASSERT3U(write_asize, <=, target_sz); 6453 ARCSTAT_BUMP(arcstat_l2_writes_sent); 6454 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 6455 ARCSTAT_INCR(arcstat_l2_size, write_sz); 6456 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 6457 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 6458 6459 /* 6460 * Bump device hand to the device start if it is approaching the end. 6461 * l2arc_evict() will already have evicted ahead for this case. 6462 */ 6463 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 6464 dev->l2ad_hand = dev->l2ad_start; 6465 dev->l2ad_first = B_FALSE; 6466 } 6467 6468 dev->l2ad_writing = B_TRUE; 6469 (void) zio_wait(pio); 6470 dev->l2ad_writing = B_FALSE; 6471 6472 return (write_asize); 6473} 6474 6475/* 6476 * Compresses an L2ARC buffer. 6477 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its 6478 * size in l2hdr->b_asize. This routine tries to compress the data and 6479 * depending on the compression result there are three possible outcomes: 6480 * *) The buffer was incompressible. The original l2hdr contents were left 6481 * untouched and are ready for writing to an L2 device. 6482 * *) The buffer was all-zeros, so there is no need to write it to an L2 6483 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 6484 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 6485 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 6486 * data buffer which holds the compressed data to be written, and b_asize 6487 * tells us how much data there is. b_compress is set to the appropriate 6488 * compression algorithm. Once writing is done, invoke 6489 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 6490 * 6491 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 6492 * buffer was incompressible). 6493 */ 6494static boolean_t 6495l2arc_compress_buf(arc_buf_hdr_t *hdr) 6496{ 6497 void *cdata; 6498 size_t csize, len, rounded; 6499 ASSERT(HDR_HAS_L2HDR(hdr)); 6500 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 6501 6502 ASSERT(HDR_HAS_L1HDR(hdr)); 6503 ASSERT(HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF); 6504 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6505 6506 len = l2hdr->b_asize; 6507 cdata = zio_data_buf_alloc(len); 6508 ASSERT3P(cdata, !=, NULL); 6509 csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata, 6510 cdata, l2hdr->b_asize); 6511 6512 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); 6513 if (rounded > csize) { 6514 bzero((char *)cdata + csize, rounded - csize); 6515 csize = rounded; 6516 } 6517 6518 if (csize == 0) { 6519 /* zero block, indicate that there's nothing to write */ 6520 zio_data_buf_free(cdata, len); 6521 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_EMPTY); 6522 l2hdr->b_asize = 0; 6523 hdr->b_l1hdr.b_tmp_cdata = NULL; 6524 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 6525 return (B_TRUE); 6526 } else if (csize > 0 && csize < len) { 6527 /* 6528 * Compression succeeded, we'll keep the cdata around for 6529 * writing and release it afterwards. 6530 */ 6531 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_LZ4); 6532 l2hdr->b_asize = csize; 6533 hdr->b_l1hdr.b_tmp_cdata = cdata; 6534 ARCSTAT_BUMP(arcstat_l2_compress_successes); 6535 return (B_TRUE); 6536 } else { 6537 /* 6538 * Compression failed, release the compressed buffer. 6539 * l2hdr will be left unmodified. 6540 */ 6541 zio_data_buf_free(cdata, len); 6542 ARCSTAT_BUMP(arcstat_l2_compress_failures); 6543 return (B_FALSE); 6544 } 6545} 6546 6547/* 6548 * Decompresses a zio read back from an l2arc device. On success, the 6549 * underlying zio's io_data buffer is overwritten by the uncompressed 6550 * version. On decompression error (corrupt compressed stream), the 6551 * zio->io_error value is set to signal an I/O error. 6552 * 6553 * Please note that the compressed data stream is not checksummed, so 6554 * if the underlying device is experiencing data corruption, we may feed 6555 * corrupt data to the decompressor, so the decompressor needs to be 6556 * able to handle this situation (LZ4 does). 6557 */ 6558static void 6559l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 6560{ 6561 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 6562 6563 if (zio->io_error != 0) { 6564 /* 6565 * An io error has occured, just restore the original io 6566 * size in preparation for a main pool read. 6567 */ 6568 zio->io_orig_size = zio->io_size = hdr->b_size; 6569 return; 6570 } 6571 6572 if (c == ZIO_COMPRESS_EMPTY) { 6573 /* 6574 * An empty buffer results in a null zio, which means we 6575 * need to fill its io_data after we're done restoring the 6576 * buffer's contents. 6577 */ 6578 ASSERT(hdr->b_l1hdr.b_buf != NULL); 6579 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size); 6580 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data; 6581 } else { 6582 ASSERT(zio->io_data != NULL); 6583 /* 6584 * We copy the compressed data from the start of the arc buffer 6585 * (the zio_read will have pulled in only what we need, the 6586 * rest is garbage which we will overwrite at decompression) 6587 * and then decompress back to the ARC data buffer. This way we 6588 * can minimize copying by simply decompressing back over the 6589 * original compressed data (rather than decompressing to an 6590 * aux buffer and then copying back the uncompressed buffer, 6591 * which is likely to be much larger). 6592 */ 6593 uint64_t csize; 6594 void *cdata; 6595 6596 csize = zio->io_size; 6597 cdata = zio_data_buf_alloc(csize); 6598 bcopy(zio->io_data, cdata, csize); 6599 if (zio_decompress_data(c, cdata, zio->io_data, csize, 6600 hdr->b_size) != 0) 6601 zio->io_error = EIO; 6602 zio_data_buf_free(cdata, csize); 6603 } 6604 6605 /* Restore the expected uncompressed IO size. */ 6606 zio->io_orig_size = zio->io_size = hdr->b_size; 6607} 6608 6609/* 6610 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 6611 * This buffer serves as a temporary holder of compressed data while 6612 * the buffer entry is being written to an l2arc device. Once that is 6613 * done, we can dispose of it. 6614 */ 6615static void 6616l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 6617{ 6618 enum zio_compress comp = HDR_GET_COMPRESS(hdr); 6619 6620 ASSERT(HDR_HAS_L1HDR(hdr)); 6621 ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp)); 6622 6623 if (comp == ZIO_COMPRESS_OFF) { 6624 /* 6625 * In this case, b_tmp_cdata points to the same buffer 6626 * as the arc_buf_t's b_data field. We don't want to 6627 * free it, since the arc_buf_t will handle that. 6628 */ 6629 hdr->b_l1hdr.b_tmp_cdata = NULL; 6630 } else if (comp == ZIO_COMPRESS_EMPTY) { 6631 /* 6632 * In this case, b_tmp_cdata was compressed to an empty 6633 * buffer, thus there's nothing to free and b_tmp_cdata 6634 * should have been set to NULL in l2arc_write_buffers(). 6635 */ 6636 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 6637 } else { 6638 /* 6639 * If the data was compressed, then we've allocated a 6640 * temporary buffer for it, so now we need to release it. 6641 */ 6642 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6643 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, 6644 hdr->b_size); 6645 hdr->b_l1hdr.b_tmp_cdata = NULL; 6646 } 6647 6648} 6649 6650/* 6651 * This thread feeds the L2ARC at regular intervals. This is the beating 6652 * heart of the L2ARC. 6653 */ 6654static void 6655l2arc_feed_thread(void *dummy __unused) 6656{ 6657 callb_cpr_t cpr; 6658 l2arc_dev_t *dev; 6659 spa_t *spa; 6660 uint64_t size, wrote; 6661 clock_t begin, next = ddi_get_lbolt(); 6662 boolean_t headroom_boost = B_FALSE; 6663 6664 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 6665 6666 mutex_enter(&l2arc_feed_thr_lock); 6667 6668 while (l2arc_thread_exit == 0) { 6669 CALLB_CPR_SAFE_BEGIN(&cpr); 6670 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 6671 next - ddi_get_lbolt()); 6672 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 6673 next = ddi_get_lbolt() + hz; 6674 6675 /* 6676 * Quick check for L2ARC devices. 6677 */ 6678 mutex_enter(&l2arc_dev_mtx); 6679 if (l2arc_ndev == 0) { 6680 mutex_exit(&l2arc_dev_mtx); 6681 continue; 6682 } 6683 mutex_exit(&l2arc_dev_mtx); 6684 begin = ddi_get_lbolt(); 6685 6686 /* 6687 * This selects the next l2arc device to write to, and in 6688 * doing so the next spa to feed from: dev->l2ad_spa. This 6689 * will return NULL if there are now no l2arc devices or if 6690 * they are all faulted. 6691 * 6692 * If a device is returned, its spa's config lock is also 6693 * held to prevent device removal. l2arc_dev_get_next() 6694 * will grab and release l2arc_dev_mtx. 6695 */ 6696 if ((dev = l2arc_dev_get_next()) == NULL) 6697 continue; 6698 6699 spa = dev->l2ad_spa; 6700 ASSERT(spa != NULL); 6701 6702 /* 6703 * If the pool is read-only then force the feed thread to 6704 * sleep a little longer. 6705 */ 6706 if (!spa_writeable(spa)) { 6707 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 6708 spa_config_exit(spa, SCL_L2ARC, dev); 6709 continue; 6710 } 6711 6712 /* 6713 * Avoid contributing to memory pressure. 6714 */ 6715 if (arc_reclaim_needed()) { 6716 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 6717 spa_config_exit(spa, SCL_L2ARC, dev); 6718 continue; 6719 } 6720 6721 ARCSTAT_BUMP(arcstat_l2_feeds); 6722 6723 size = l2arc_write_size(); 6724 6725 /* 6726 * Evict L2ARC buffers that will be overwritten. 6727 */ 6728 l2arc_evict(dev, size, B_FALSE); 6729 6730 /* 6731 * Write ARC buffers. 6732 */ 6733 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 6734 6735 /* 6736 * Calculate interval between writes. 6737 */ 6738 next = l2arc_write_interval(begin, size, wrote); 6739 spa_config_exit(spa, SCL_L2ARC, dev); 6740 } 6741 6742 l2arc_thread_exit = 0; 6743 cv_broadcast(&l2arc_feed_thr_cv); 6744 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 6745 thread_exit(); 6746} 6747 6748boolean_t 6749l2arc_vdev_present(vdev_t *vd) 6750{ 6751 l2arc_dev_t *dev; 6752 6753 mutex_enter(&l2arc_dev_mtx); 6754 for (dev = list_head(l2arc_dev_list); dev != NULL; 6755 dev = list_next(l2arc_dev_list, dev)) { 6756 if (dev->l2ad_vdev == vd) 6757 break; 6758 } 6759 mutex_exit(&l2arc_dev_mtx); 6760 6761 return (dev != NULL); 6762} 6763 6764/* 6765 * Add a vdev for use by the L2ARC. By this point the spa has already 6766 * validated the vdev and opened it. 6767 */ 6768void 6769l2arc_add_vdev(spa_t *spa, vdev_t *vd) 6770{ 6771 l2arc_dev_t *adddev; 6772 6773 ASSERT(!l2arc_vdev_present(vd)); 6774 6775 vdev_ashift_optimize(vd); 6776 6777 /* 6778 * Create a new l2arc device entry. 6779 */ 6780 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 6781 adddev->l2ad_spa = spa; 6782 adddev->l2ad_vdev = vd; 6783 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 6784 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 6785 adddev->l2ad_hand = adddev->l2ad_start; 6786 adddev->l2ad_first = B_TRUE; 6787 adddev->l2ad_writing = B_FALSE; 6788 6789 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 6790 /* 6791 * This is a list of all ARC buffers that are still valid on the 6792 * device. 6793 */ 6794 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 6795 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 6796 6797 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 6798 refcount_create(&adddev->l2ad_alloc); 6799 6800 /* 6801 * Add device to global list 6802 */ 6803 mutex_enter(&l2arc_dev_mtx); 6804 list_insert_head(l2arc_dev_list, adddev); 6805 atomic_inc_64(&l2arc_ndev); 6806 mutex_exit(&l2arc_dev_mtx); 6807} 6808 6809/* 6810 * Remove a vdev from the L2ARC. 6811 */ 6812void 6813l2arc_remove_vdev(vdev_t *vd) 6814{ 6815 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 6816 6817 /* 6818 * Find the device by vdev 6819 */ 6820 mutex_enter(&l2arc_dev_mtx); 6821 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 6822 nextdev = list_next(l2arc_dev_list, dev); 6823 if (vd == dev->l2ad_vdev) { 6824 remdev = dev; 6825 break; 6826 } 6827 } 6828 ASSERT(remdev != NULL); 6829 6830 /* 6831 * Remove device from global list 6832 */ 6833 list_remove(l2arc_dev_list, remdev); 6834 l2arc_dev_last = NULL; /* may have been invalidated */ 6835 atomic_dec_64(&l2arc_ndev); 6836 mutex_exit(&l2arc_dev_mtx); 6837 6838 /* 6839 * Clear all buflists and ARC references. L2ARC device flush. 6840 */ 6841 l2arc_evict(remdev, 0, B_TRUE); 6842 list_destroy(&remdev->l2ad_buflist); 6843 mutex_destroy(&remdev->l2ad_mtx); 6844 refcount_destroy(&remdev->l2ad_alloc); 6845 kmem_free(remdev, sizeof (l2arc_dev_t)); 6846} 6847 6848void 6849l2arc_init(void) 6850{ 6851 l2arc_thread_exit = 0; 6852 l2arc_ndev = 0; 6853 l2arc_writes_sent = 0; 6854 l2arc_writes_done = 0; 6855 6856 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 6857 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 6858 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 6859 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 6860 6861 l2arc_dev_list = &L2ARC_dev_list; 6862 l2arc_free_on_write = &L2ARC_free_on_write; 6863 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 6864 offsetof(l2arc_dev_t, l2ad_node)); 6865 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 6866 offsetof(l2arc_data_free_t, l2df_list_node)); 6867} 6868 6869void 6870l2arc_fini(void) 6871{ 6872 /* 6873 * This is called from dmu_fini(), which is called from spa_fini(); 6874 * Because of this, we can assume that all l2arc devices have 6875 * already been removed when the pools themselves were removed. 6876 */ 6877 6878 l2arc_do_free_on_write(); 6879 6880 mutex_destroy(&l2arc_feed_thr_lock); 6881 cv_destroy(&l2arc_feed_thr_cv); 6882 mutex_destroy(&l2arc_dev_mtx); 6883 mutex_destroy(&l2arc_free_on_write_mtx); 6884 6885 list_destroy(l2arc_dev_list); 6886 list_destroy(l2arc_free_on_write); 6887} 6888 6889void 6890l2arc_start(void) 6891{ 6892 if (!(spa_mode_global & FWRITE)) 6893 return; 6894 6895 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 6896 TS_RUN, minclsyspri); 6897} 6898 6899void 6900l2arc_stop(void) 6901{ 6902 if (!(spa_mode_global & FWRITE)) 6903 return; 6904 6905 mutex_enter(&l2arc_feed_thr_lock); 6906 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 6907 l2arc_thread_exit = 1; 6908 while (l2arc_thread_exit != 0) 6909 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 6910 mutex_exit(&l2arc_feed_thr_lock); 6911} 6912