arc.c revision 332525
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2012, Joyent, Inc. All rights reserved. 24 * Copyright (c) 2011, 2017 by Delphix. All rights reserved. 25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 26 * Copyright 2017 Nexenta Systems, Inc. All rights reserved. 27 */ 28 29/* 30 * DVA-based Adjustable Replacement Cache 31 * 32 * While much of the theory of operation used here is 33 * based on the self-tuning, low overhead replacement cache 34 * presented by Megiddo and Modha at FAST 2003, there are some 35 * significant differences: 36 * 37 * 1. The Megiddo and Modha model assumes any page is evictable. 38 * Pages in its cache cannot be "locked" into memory. This makes 39 * the eviction algorithm simple: evict the last page in the list. 40 * This also make the performance characteristics easy to reason 41 * about. Our cache is not so simple. At any given moment, some 42 * subset of the blocks in the cache are un-evictable because we 43 * have handed out a reference to them. Blocks are only evictable 44 * when there are no external references active. This makes 45 * eviction far more problematic: we choose to evict the evictable 46 * blocks that are the "lowest" in the list. 47 * 48 * There are times when it is not possible to evict the requested 49 * space. In these circumstances we are unable to adjust the cache 50 * size. To prevent the cache growing unbounded at these times we 51 * implement a "cache throttle" that slows the flow of new data 52 * into the cache until we can make space available. 53 * 54 * 2. The Megiddo and Modha model assumes a fixed cache size. 55 * Pages are evicted when the cache is full and there is a cache 56 * miss. Our model has a variable sized cache. It grows with 57 * high use, but also tries to react to memory pressure from the 58 * operating system: decreasing its size when system memory is 59 * tight. 60 * 61 * 3. The Megiddo and Modha model assumes a fixed page size. All 62 * elements of the cache are therefore exactly the same size. So 63 * when adjusting the cache size following a cache miss, its simply 64 * a matter of choosing a single page to evict. In our model, we 65 * have variable sized cache blocks (rangeing from 512 bytes to 66 * 128K bytes). We therefore choose a set of blocks to evict to make 67 * space for a cache miss that approximates as closely as possible 68 * the space used by the new block. 69 * 70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 71 * by N. Megiddo & D. Modha, FAST 2003 72 */ 73 74/* 75 * The locking model: 76 * 77 * A new reference to a cache buffer can be obtained in two 78 * ways: 1) via a hash table lookup using the DVA as a key, 79 * or 2) via one of the ARC lists. The arc_read() interface 80 * uses method 1, while the internal ARC algorithms for 81 * adjusting the cache use method 2. We therefore provide two 82 * types of locks: 1) the hash table lock array, and 2) the 83 * ARC list locks. 84 * 85 * Buffers do not have their own mutexes, rather they rely on the 86 * hash table mutexes for the bulk of their protection (i.e. most 87 * fields in the arc_buf_hdr_t are protected by these mutexes). 88 * 89 * buf_hash_find() returns the appropriate mutex (held) when it 90 * locates the requested buffer in the hash table. It returns 91 * NULL for the mutex if the buffer was not in the table. 92 * 93 * buf_hash_remove() expects the appropriate hash mutex to be 94 * already held before it is invoked. 95 * 96 * Each ARC state also has a mutex which is used to protect the 97 * buffer list associated with the state. When attempting to 98 * obtain a hash table lock while holding an ARC list lock you 99 * must use: mutex_tryenter() to avoid deadlock. Also note that 100 * the active state mutex must be held before the ghost state mutex. 101 * 102 * Note that the majority of the performance stats are manipulated 103 * with atomic operations. 104 * 105 * The L2ARC uses the l2ad_mtx on each vdev for the following: 106 * 107 * - L2ARC buflist creation 108 * - L2ARC buflist eviction 109 * - L2ARC write completion, which walks L2ARC buflists 110 * - ARC header destruction, as it removes from L2ARC buflists 111 * - ARC header release, as it removes from L2ARC buflists 112 */ 113 114/* 115 * ARC operation: 116 * 117 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure. 118 * This structure can point either to a block that is still in the cache or to 119 * one that is only accessible in an L2 ARC device, or it can provide 120 * information about a block that was recently evicted. If a block is 121 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough 122 * information to retrieve it from the L2ARC device. This information is 123 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block 124 * that is in this state cannot access the data directly. 125 * 126 * Blocks that are actively being referenced or have not been evicted 127 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within 128 * the arc_buf_hdr_t that will point to the data block in memory. A block can 129 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC 130 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and 131 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd). 132 * 133 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the 134 * ability to store the physical data (b_pabd) associated with the DVA of the 135 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block, 136 * it will match its on-disk compression characteristics. This behavior can be 137 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the 138 * compressed ARC functionality is disabled, the b_pabd will point to an 139 * uncompressed version of the on-disk data. 140 * 141 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each 142 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it. 143 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC 144 * consumer. The ARC will provide references to this data and will keep it 145 * cached until it is no longer in use. The ARC caches only the L1ARC's physical 146 * data block and will evict any arc_buf_t that is no longer referenced. The 147 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the 148 * "overhead_size" kstat. 149 * 150 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or 151 * compressed form. The typical case is that consumers will want uncompressed 152 * data, and when that happens a new data buffer is allocated where the data is 153 * decompressed for them to use. Currently the only consumer who wants 154 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it 155 * exists on disk. When this happens, the arc_buf_t's data buffer is shared 156 * with the arc_buf_hdr_t. 157 * 158 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The 159 * first one is owned by a compressed send consumer (and therefore references 160 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be 161 * used by any other consumer (and has its own uncompressed copy of the data 162 * buffer). 163 * 164 * arc_buf_hdr_t 165 * +-----------+ 166 * | fields | 167 * | common to | 168 * | L1- and | 169 * | L2ARC | 170 * +-----------+ 171 * | l2arc_buf_hdr_t 172 * | | 173 * +-----------+ 174 * | l1arc_buf_hdr_t 175 * | | arc_buf_t 176 * | b_buf +------------>+-----------+ arc_buf_t 177 * | b_pabd +-+ |b_next +---->+-----------+ 178 * +-----------+ | |-----------| |b_next +-->NULL 179 * | |b_comp = T | +-----------+ 180 * | |b_data +-+ |b_comp = F | 181 * | +-----------+ | |b_data +-+ 182 * +->+------+ | +-----------+ | 183 * compressed | | | | 184 * data | |<--------------+ | uncompressed 185 * +------+ compressed, | data 186 * shared +-->+------+ 187 * data | | 188 * | | 189 * +------+ 190 * 191 * When a consumer reads a block, the ARC must first look to see if the 192 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new 193 * arc_buf_t and either copies uncompressed data into a new data buffer from an 194 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a 195 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the 196 * hdr is compressed and the desired compression characteristics of the 197 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the 198 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be 199 * the last buffer in the hdr's b_buf list, however a shared compressed buf can 200 * be anywhere in the hdr's list. 201 * 202 * The diagram below shows an example of an uncompressed ARC hdr that is 203 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is 204 * the last element in the buf list): 205 * 206 * arc_buf_hdr_t 207 * +-----------+ 208 * | | 209 * | | 210 * | | 211 * +-----------+ 212 * l2arc_buf_hdr_t| | 213 * | | 214 * +-----------+ 215 * l1arc_buf_hdr_t| | 216 * | | arc_buf_t (shared) 217 * | b_buf +------------>+---------+ arc_buf_t 218 * | | |b_next +---->+---------+ 219 * | b_pabd +-+ |---------| |b_next +-->NULL 220 * +-----------+ | | | +---------+ 221 * | |b_data +-+ | | 222 * | +---------+ | |b_data +-+ 223 * +->+------+ | +---------+ | 224 * | | | | 225 * uncompressed | | | | 226 * data +------+ | | 227 * ^ +->+------+ | 228 * | uncompressed | | | 229 * | data | | | 230 * | +------+ | 231 * +---------------------------------+ 232 * 233 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd 234 * since the physical block is about to be rewritten. The new data contents 235 * will be contained in the arc_buf_t. As the I/O pipeline performs the write, 236 * it may compress the data before writing it to disk. The ARC will be called 237 * with the transformed data and will bcopy the transformed on-disk block into 238 * a newly allocated b_pabd. Writes are always done into buffers which have 239 * either been loaned (and hence are new and don't have other readers) or 240 * buffers which have been released (and hence have their own hdr, if there 241 * were originally other readers of the buf's original hdr). This ensures that 242 * the ARC only needs to update a single buf and its hdr after a write occurs. 243 * 244 * When the L2ARC is in use, it will also take advantage of the b_pabd. The 245 * L2ARC will always write the contents of b_pabd to the L2ARC. This means 246 * that when compressed ARC is enabled that the L2ARC blocks are identical 247 * to the on-disk block in the main data pool. This provides a significant 248 * advantage since the ARC can leverage the bp's checksum when reading from the 249 * L2ARC to determine if the contents are valid. However, if the compressed 250 * ARC is disabled, then the L2ARC's block must be transformed to look 251 * like the physical block in the main data pool before comparing the 252 * checksum and determining its validity. 253 */ 254 255#include <sys/spa.h> 256#include <sys/zio.h> 257#include <sys/spa_impl.h> 258#include <sys/zio_compress.h> 259#include <sys/zio_checksum.h> 260#include <sys/zfs_context.h> 261#include <sys/arc.h> 262#include <sys/refcount.h> 263#include <sys/vdev.h> 264#include <sys/vdev_impl.h> 265#include <sys/dsl_pool.h> 266#include <sys/zio_checksum.h> 267#include <sys/multilist.h> 268#include <sys/abd.h> 269#ifdef _KERNEL 270#include <sys/dnlc.h> 271#include <sys/racct.h> 272#endif 273#include <sys/callb.h> 274#include <sys/kstat.h> 275#include <sys/trim_map.h> 276#include <zfs_fletcher.h> 277#include <sys/sdt.h> 278 279#include <machine/vmparam.h> 280 281#ifdef illumos 282#ifndef _KERNEL 283/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 284boolean_t arc_watch = B_FALSE; 285int arc_procfd; 286#endif 287#endif /* illumos */ 288 289static kmutex_t arc_reclaim_lock; 290static kcondvar_t arc_reclaim_thread_cv; 291static boolean_t arc_reclaim_thread_exit; 292static kcondvar_t arc_reclaim_waiters_cv; 293 294static kmutex_t arc_dnlc_evicts_lock; 295static kcondvar_t arc_dnlc_evicts_cv; 296static boolean_t arc_dnlc_evicts_thread_exit; 297 298uint_t arc_reduce_dnlc_percent = 3; 299 300/* 301 * The number of headers to evict in arc_evict_state_impl() before 302 * dropping the sublist lock and evicting from another sublist. A lower 303 * value means we're more likely to evict the "correct" header (i.e. the 304 * oldest header in the arc state), but comes with higher overhead 305 * (i.e. more invocations of arc_evict_state_impl()). 306 */ 307int zfs_arc_evict_batch_limit = 10; 308 309/* number of seconds before growing cache again */ 310static int arc_grow_retry = 60; 311 312/* shift of arc_c for calculating overflow limit in arc_get_data_impl */ 313int zfs_arc_overflow_shift = 8; 314 315/* shift of arc_c for calculating both min and max arc_p */ 316static int arc_p_min_shift = 4; 317 318/* log2(fraction of arc to reclaim) */ 319static int arc_shrink_shift = 7; 320 321/* 322 * log2(fraction of ARC which must be free to allow growing). 323 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 324 * when reading a new block into the ARC, we will evict an equal-sized block 325 * from the ARC. 326 * 327 * This must be less than arc_shrink_shift, so that when we shrink the ARC, 328 * we will still not allow it to grow. 329 */ 330int arc_no_grow_shift = 5; 331 332 333/* 334 * minimum lifespan of a prefetch block in clock ticks 335 * (initialized in arc_init()) 336 */ 337static int arc_min_prefetch_lifespan; 338 339/* 340 * If this percent of memory is free, don't throttle. 341 */ 342int arc_lotsfree_percent = 10; 343 344static int arc_dead; 345extern boolean_t zfs_prefetch_disable; 346 347/* 348 * The arc has filled available memory and has now warmed up. 349 */ 350static boolean_t arc_warm; 351 352/* 353 * log2 fraction of the zio arena to keep free. 354 */ 355int arc_zio_arena_free_shift = 2; 356 357/* 358 * These tunables are for performance analysis. 359 */ 360uint64_t zfs_arc_max; 361uint64_t zfs_arc_min; 362uint64_t zfs_arc_meta_limit = 0; 363uint64_t zfs_arc_meta_min = 0; 364int zfs_arc_grow_retry = 0; 365int zfs_arc_shrink_shift = 0; 366int zfs_arc_no_grow_shift = 0; 367int zfs_arc_p_min_shift = 0; 368uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 369u_int zfs_arc_free_target = 0; 370 371/* Absolute min for arc min / max is 16MB. */ 372static uint64_t arc_abs_min = 16 << 20; 373 374boolean_t zfs_compressed_arc_enabled = B_TRUE; 375 376static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 377static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 378static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS); 379static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS); 380static int sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS); 381 382#if defined(__FreeBSD__) && defined(_KERNEL) 383static void 384arc_free_target_init(void *unused __unused) 385{ 386 387 zfs_arc_free_target = vm_pageout_wakeup_thresh; 388} 389SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 390 arc_free_target_init, NULL); 391 392TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 393TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 394TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 395TUNABLE_INT("vfs.zfs.arc_grow_retry", &zfs_arc_grow_retry); 396TUNABLE_INT("vfs.zfs.arc_no_grow_shift", &zfs_arc_no_grow_shift); 397SYSCTL_DECL(_vfs_zfs); 398SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN, 399 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size"); 400SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN, 401 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size"); 402SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_no_grow_shift, CTLTYPE_U32 | CTLFLAG_RWTUN, 403 0, sizeof(uint32_t), sysctl_vfs_zfs_arc_no_grow_shift, "U", 404 "log2(fraction of ARC which must be free to allow growing)"); 405SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 406 &zfs_arc_average_blocksize, 0, 407 "ARC average blocksize"); 408SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 409 &arc_shrink_shift, 0, 410 "log2(fraction of arc to reclaim)"); 411SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_grow_retry, CTLFLAG_RW, 412 &arc_grow_retry, 0, 413 "Wait in seconds before considering growing ARC"); 414SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN, 415 &zfs_compressed_arc_enabled, 0, "Enable compressed ARC"); 416 417/* 418 * We don't have a tunable for arc_free_target due to the dependency on 419 * pagedaemon initialisation. 420 */ 421SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 422 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 423 sysctl_vfs_zfs_arc_free_target, "IU", 424 "Desired number of free pages below which ARC triggers reclaim"); 425 426static int 427sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 428{ 429 u_int val; 430 int err; 431 432 val = zfs_arc_free_target; 433 err = sysctl_handle_int(oidp, &val, 0, req); 434 if (err != 0 || req->newptr == NULL) 435 return (err); 436 437 if (val < minfree) 438 return (EINVAL); 439 if (val > vm_cnt.v_page_count) 440 return (EINVAL); 441 442 zfs_arc_free_target = val; 443 444 return (0); 445} 446 447/* 448 * Must be declared here, before the definition of corresponding kstat 449 * macro which uses the same names will confuse the compiler. 450 */ 451SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 452 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 453 sysctl_vfs_zfs_arc_meta_limit, "QU", 454 "ARC metadata limit"); 455#endif 456 457/* 458 * Note that buffers can be in one of 6 states: 459 * ARC_anon - anonymous (discussed below) 460 * ARC_mru - recently used, currently cached 461 * ARC_mru_ghost - recentely used, no longer in cache 462 * ARC_mfu - frequently used, currently cached 463 * ARC_mfu_ghost - frequently used, no longer in cache 464 * ARC_l2c_only - exists in L2ARC but not other states 465 * When there are no active references to the buffer, they are 466 * are linked onto a list in one of these arc states. These are 467 * the only buffers that can be evicted or deleted. Within each 468 * state there are multiple lists, one for meta-data and one for 469 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 470 * etc.) is tracked separately so that it can be managed more 471 * explicitly: favored over data, limited explicitly. 472 * 473 * Anonymous buffers are buffers that are not associated with 474 * a DVA. These are buffers that hold dirty block copies 475 * before they are written to stable storage. By definition, 476 * they are "ref'd" and are considered part of arc_mru 477 * that cannot be freed. Generally, they will aquire a DVA 478 * as they are written and migrate onto the arc_mru list. 479 * 480 * The ARC_l2c_only state is for buffers that are in the second 481 * level ARC but no longer in any of the ARC_m* lists. The second 482 * level ARC itself may also contain buffers that are in any of 483 * the ARC_m* states - meaning that a buffer can exist in two 484 * places. The reason for the ARC_l2c_only state is to keep the 485 * buffer header in the hash table, so that reads that hit the 486 * second level ARC benefit from these fast lookups. 487 */ 488 489typedef struct arc_state { 490 /* 491 * list of evictable buffers 492 */ 493 multilist_t *arcs_list[ARC_BUFC_NUMTYPES]; 494 /* 495 * total amount of evictable data in this state 496 */ 497 refcount_t arcs_esize[ARC_BUFC_NUMTYPES]; 498 /* 499 * total amount of data in this state; this includes: evictable, 500 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 501 */ 502 refcount_t arcs_size; 503} arc_state_t; 504 505/* The 6 states: */ 506static arc_state_t ARC_anon; 507static arc_state_t ARC_mru; 508static arc_state_t ARC_mru_ghost; 509static arc_state_t ARC_mfu; 510static arc_state_t ARC_mfu_ghost; 511static arc_state_t ARC_l2c_only; 512 513typedef struct arc_stats { 514 kstat_named_t arcstat_hits; 515 kstat_named_t arcstat_misses; 516 kstat_named_t arcstat_demand_data_hits; 517 kstat_named_t arcstat_demand_data_misses; 518 kstat_named_t arcstat_demand_metadata_hits; 519 kstat_named_t arcstat_demand_metadata_misses; 520 kstat_named_t arcstat_prefetch_data_hits; 521 kstat_named_t arcstat_prefetch_data_misses; 522 kstat_named_t arcstat_prefetch_metadata_hits; 523 kstat_named_t arcstat_prefetch_metadata_misses; 524 kstat_named_t arcstat_mru_hits; 525 kstat_named_t arcstat_mru_ghost_hits; 526 kstat_named_t arcstat_mfu_hits; 527 kstat_named_t arcstat_mfu_ghost_hits; 528 kstat_named_t arcstat_allocated; 529 kstat_named_t arcstat_deleted; 530 /* 531 * Number of buffers that could not be evicted because the hash lock 532 * was held by another thread. The lock may not necessarily be held 533 * by something using the same buffer, since hash locks are shared 534 * by multiple buffers. 535 */ 536 kstat_named_t arcstat_mutex_miss; 537 /* 538 * Number of buffers skipped because they have I/O in progress, are 539 * indrect prefetch buffers that have not lived long enough, or are 540 * not from the spa we're trying to evict from. 541 */ 542 kstat_named_t arcstat_evict_skip; 543 /* 544 * Number of times arc_evict_state() was unable to evict enough 545 * buffers to reach it's target amount. 546 */ 547 kstat_named_t arcstat_evict_not_enough; 548 kstat_named_t arcstat_evict_l2_cached; 549 kstat_named_t arcstat_evict_l2_eligible; 550 kstat_named_t arcstat_evict_l2_ineligible; 551 kstat_named_t arcstat_evict_l2_skip; 552 kstat_named_t arcstat_hash_elements; 553 kstat_named_t arcstat_hash_elements_max; 554 kstat_named_t arcstat_hash_collisions; 555 kstat_named_t arcstat_hash_chains; 556 kstat_named_t arcstat_hash_chain_max; 557 kstat_named_t arcstat_p; 558 kstat_named_t arcstat_c; 559 kstat_named_t arcstat_c_min; 560 kstat_named_t arcstat_c_max; 561 kstat_named_t arcstat_size; 562 /* 563 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd. 564 * Note that the compressed bytes may match the uncompressed bytes 565 * if the block is either not compressed or compressed arc is disabled. 566 */ 567 kstat_named_t arcstat_compressed_size; 568 /* 569 * Uncompressed size of the data stored in b_pabd. If compressed 570 * arc is disabled then this value will be identical to the stat 571 * above. 572 */ 573 kstat_named_t arcstat_uncompressed_size; 574 /* 575 * Number of bytes stored in all the arc_buf_t's. This is classified 576 * as "overhead" since this data is typically short-lived and will 577 * be evicted from the arc when it becomes unreferenced unless the 578 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level 579 * values have been set (see comment in dbuf.c for more information). 580 */ 581 kstat_named_t arcstat_overhead_size; 582 /* 583 * Number of bytes consumed by internal ARC structures necessary 584 * for tracking purposes; these structures are not actually 585 * backed by ARC buffers. This includes arc_buf_hdr_t structures 586 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 587 * caches), and arc_buf_t structures (allocated via arc_buf_t 588 * cache). 589 */ 590 kstat_named_t arcstat_hdr_size; 591 /* 592 * Number of bytes consumed by ARC buffers of type equal to 593 * ARC_BUFC_DATA. This is generally consumed by buffers backing 594 * on disk user data (e.g. plain file contents). 595 */ 596 kstat_named_t arcstat_data_size; 597 /* 598 * Number of bytes consumed by ARC buffers of type equal to 599 * ARC_BUFC_METADATA. This is generally consumed by buffers 600 * backing on disk data that is used for internal ZFS 601 * structures (e.g. ZAP, dnode, indirect blocks, etc). 602 */ 603 kstat_named_t arcstat_metadata_size; 604 /* 605 * Number of bytes consumed by various buffers and structures 606 * not actually backed with ARC buffers. This includes bonus 607 * buffers (allocated directly via zio_buf_* functions), 608 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 609 * cache), and dnode_t structures (allocated via dnode_t cache). 610 */ 611 kstat_named_t arcstat_other_size; 612 /* 613 * Total number of bytes consumed by ARC buffers residing in the 614 * arc_anon state. This includes *all* buffers in the arc_anon 615 * state; e.g. data, metadata, evictable, and unevictable buffers 616 * are all included in this value. 617 */ 618 kstat_named_t arcstat_anon_size; 619 /* 620 * Number of bytes consumed by ARC buffers that meet the 621 * following criteria: backing buffers of type ARC_BUFC_DATA, 622 * residing in the arc_anon state, and are eligible for eviction 623 * (e.g. have no outstanding holds on the buffer). 624 */ 625 kstat_named_t arcstat_anon_evictable_data; 626 /* 627 * Number of bytes consumed by ARC buffers that meet the 628 * following criteria: backing buffers of type ARC_BUFC_METADATA, 629 * residing in the arc_anon state, and are eligible for eviction 630 * (e.g. have no outstanding holds on the buffer). 631 */ 632 kstat_named_t arcstat_anon_evictable_metadata; 633 /* 634 * Total number of bytes consumed by ARC buffers residing in the 635 * arc_mru state. This includes *all* buffers in the arc_mru 636 * state; e.g. data, metadata, evictable, and unevictable buffers 637 * are all included in this value. 638 */ 639 kstat_named_t arcstat_mru_size; 640 /* 641 * Number of bytes consumed by ARC buffers that meet the 642 * following criteria: backing buffers of type ARC_BUFC_DATA, 643 * residing in the arc_mru state, and are eligible for eviction 644 * (e.g. have no outstanding holds on the buffer). 645 */ 646 kstat_named_t arcstat_mru_evictable_data; 647 /* 648 * Number of bytes consumed by ARC buffers that meet the 649 * following criteria: backing buffers of type ARC_BUFC_METADATA, 650 * residing in the arc_mru state, and are eligible for eviction 651 * (e.g. have no outstanding holds on the buffer). 652 */ 653 kstat_named_t arcstat_mru_evictable_metadata; 654 /* 655 * Total number of bytes that *would have been* consumed by ARC 656 * buffers in the arc_mru_ghost state. The key thing to note 657 * here, is the fact that this size doesn't actually indicate 658 * RAM consumption. The ghost lists only consist of headers and 659 * don't actually have ARC buffers linked off of these headers. 660 * Thus, *if* the headers had associated ARC buffers, these 661 * buffers *would have* consumed this number of bytes. 662 */ 663 kstat_named_t arcstat_mru_ghost_size; 664 /* 665 * Number of bytes that *would have been* consumed by ARC 666 * buffers that are eligible for eviction, of type 667 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 668 */ 669 kstat_named_t arcstat_mru_ghost_evictable_data; 670 /* 671 * Number of bytes that *would have been* consumed by ARC 672 * buffers that are eligible for eviction, of type 673 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 674 */ 675 kstat_named_t arcstat_mru_ghost_evictable_metadata; 676 /* 677 * Total number of bytes consumed by ARC buffers residing in the 678 * arc_mfu state. This includes *all* buffers in the arc_mfu 679 * state; e.g. data, metadata, evictable, and unevictable buffers 680 * are all included in this value. 681 */ 682 kstat_named_t arcstat_mfu_size; 683 /* 684 * Number of bytes consumed by ARC buffers that are eligible for 685 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 686 * state. 687 */ 688 kstat_named_t arcstat_mfu_evictable_data; 689 /* 690 * Number of bytes consumed by ARC buffers that are eligible for 691 * eviction, of type ARC_BUFC_METADATA, and reside in the 692 * arc_mfu state. 693 */ 694 kstat_named_t arcstat_mfu_evictable_metadata; 695 /* 696 * Total number of bytes that *would have been* consumed by ARC 697 * buffers in the arc_mfu_ghost state. See the comment above 698 * arcstat_mru_ghost_size for more details. 699 */ 700 kstat_named_t arcstat_mfu_ghost_size; 701 /* 702 * Number of bytes that *would have been* consumed by ARC 703 * buffers that are eligible for eviction, of type 704 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 705 */ 706 kstat_named_t arcstat_mfu_ghost_evictable_data; 707 /* 708 * Number of bytes that *would have been* consumed by ARC 709 * buffers that are eligible for eviction, of type 710 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 711 */ 712 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 713 kstat_named_t arcstat_l2_hits; 714 kstat_named_t arcstat_l2_misses; 715 kstat_named_t arcstat_l2_feeds; 716 kstat_named_t arcstat_l2_rw_clash; 717 kstat_named_t arcstat_l2_read_bytes; 718 kstat_named_t arcstat_l2_write_bytes; 719 kstat_named_t arcstat_l2_writes_sent; 720 kstat_named_t arcstat_l2_writes_done; 721 kstat_named_t arcstat_l2_writes_error; 722 kstat_named_t arcstat_l2_writes_lock_retry; 723 kstat_named_t arcstat_l2_evict_lock_retry; 724 kstat_named_t arcstat_l2_evict_reading; 725 kstat_named_t arcstat_l2_evict_l1cached; 726 kstat_named_t arcstat_l2_free_on_write; 727 kstat_named_t arcstat_l2_abort_lowmem; 728 kstat_named_t arcstat_l2_cksum_bad; 729 kstat_named_t arcstat_l2_io_error; 730 kstat_named_t arcstat_l2_lsize; 731 kstat_named_t arcstat_l2_psize; 732 kstat_named_t arcstat_l2_hdr_size; 733 kstat_named_t arcstat_l2_write_trylock_fail; 734 kstat_named_t arcstat_l2_write_passed_headroom; 735 kstat_named_t arcstat_l2_write_spa_mismatch; 736 kstat_named_t arcstat_l2_write_in_l2; 737 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 738 kstat_named_t arcstat_l2_write_not_cacheable; 739 kstat_named_t arcstat_l2_write_full; 740 kstat_named_t arcstat_l2_write_buffer_iter; 741 kstat_named_t arcstat_l2_write_pios; 742 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 743 kstat_named_t arcstat_l2_write_buffer_list_iter; 744 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 745 kstat_named_t arcstat_memory_throttle_count; 746 kstat_named_t arcstat_meta_used; 747 kstat_named_t arcstat_meta_limit; 748 kstat_named_t arcstat_meta_max; 749 kstat_named_t arcstat_meta_min; 750 kstat_named_t arcstat_sync_wait_for_async; 751 kstat_named_t arcstat_demand_hit_predictive_prefetch; 752} arc_stats_t; 753 754static arc_stats_t arc_stats = { 755 { "hits", KSTAT_DATA_UINT64 }, 756 { "misses", KSTAT_DATA_UINT64 }, 757 { "demand_data_hits", KSTAT_DATA_UINT64 }, 758 { "demand_data_misses", KSTAT_DATA_UINT64 }, 759 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 760 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 761 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 762 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 763 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 764 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 765 { "mru_hits", KSTAT_DATA_UINT64 }, 766 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 767 { "mfu_hits", KSTAT_DATA_UINT64 }, 768 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 769 { "allocated", KSTAT_DATA_UINT64 }, 770 { "deleted", KSTAT_DATA_UINT64 }, 771 { "mutex_miss", KSTAT_DATA_UINT64 }, 772 { "evict_skip", KSTAT_DATA_UINT64 }, 773 { "evict_not_enough", KSTAT_DATA_UINT64 }, 774 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 775 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 776 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 777 { "evict_l2_skip", KSTAT_DATA_UINT64 }, 778 { "hash_elements", KSTAT_DATA_UINT64 }, 779 { "hash_elements_max", KSTAT_DATA_UINT64 }, 780 { "hash_collisions", KSTAT_DATA_UINT64 }, 781 { "hash_chains", KSTAT_DATA_UINT64 }, 782 { "hash_chain_max", KSTAT_DATA_UINT64 }, 783 { "p", KSTAT_DATA_UINT64 }, 784 { "c", KSTAT_DATA_UINT64 }, 785 { "c_min", KSTAT_DATA_UINT64 }, 786 { "c_max", KSTAT_DATA_UINT64 }, 787 { "size", KSTAT_DATA_UINT64 }, 788 { "compressed_size", KSTAT_DATA_UINT64 }, 789 { "uncompressed_size", KSTAT_DATA_UINT64 }, 790 { "overhead_size", KSTAT_DATA_UINT64 }, 791 { "hdr_size", KSTAT_DATA_UINT64 }, 792 { "data_size", KSTAT_DATA_UINT64 }, 793 { "metadata_size", KSTAT_DATA_UINT64 }, 794 { "other_size", KSTAT_DATA_UINT64 }, 795 { "anon_size", KSTAT_DATA_UINT64 }, 796 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 797 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 798 { "mru_size", KSTAT_DATA_UINT64 }, 799 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 800 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 801 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 802 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 803 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 804 { "mfu_size", KSTAT_DATA_UINT64 }, 805 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 806 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 807 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 808 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 809 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 810 { "l2_hits", KSTAT_DATA_UINT64 }, 811 { "l2_misses", KSTAT_DATA_UINT64 }, 812 { "l2_feeds", KSTAT_DATA_UINT64 }, 813 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 814 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 815 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 816 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 817 { "l2_writes_done", KSTAT_DATA_UINT64 }, 818 { "l2_writes_error", KSTAT_DATA_UINT64 }, 819 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 820 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 821 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 822 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 823 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 824 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 825 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 826 { "l2_io_error", KSTAT_DATA_UINT64 }, 827 { "l2_size", KSTAT_DATA_UINT64 }, 828 { "l2_asize", KSTAT_DATA_UINT64 }, 829 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 830 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 831 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 832 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 833 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 834 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 835 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 836 { "l2_write_full", KSTAT_DATA_UINT64 }, 837 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 838 { "l2_write_pios", KSTAT_DATA_UINT64 }, 839 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 840 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 841 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 842 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 843 { "arc_meta_used", KSTAT_DATA_UINT64 }, 844 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 845 { "arc_meta_max", KSTAT_DATA_UINT64 }, 846 { "arc_meta_min", KSTAT_DATA_UINT64 }, 847 { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 848 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 849}; 850 851#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 852 853#define ARCSTAT_INCR(stat, val) \ 854 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 855 856#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 857#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 858 859#define ARCSTAT_MAX(stat, val) { \ 860 uint64_t m; \ 861 while ((val) > (m = arc_stats.stat.value.ui64) && \ 862 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 863 continue; \ 864} 865 866#define ARCSTAT_MAXSTAT(stat) \ 867 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 868 869/* 870 * We define a macro to allow ARC hits/misses to be easily broken down by 871 * two separate conditions, giving a total of four different subtypes for 872 * each of hits and misses (so eight statistics total). 873 */ 874#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 875 if (cond1) { \ 876 if (cond2) { \ 877 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 878 } else { \ 879 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 880 } \ 881 } else { \ 882 if (cond2) { \ 883 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 884 } else { \ 885 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 886 } \ 887 } 888 889kstat_t *arc_ksp; 890static arc_state_t *arc_anon; 891static arc_state_t *arc_mru; 892static arc_state_t *arc_mru_ghost; 893static arc_state_t *arc_mfu; 894static arc_state_t *arc_mfu_ghost; 895static arc_state_t *arc_l2c_only; 896 897/* 898 * There are several ARC variables that are critical to export as kstats -- 899 * but we don't want to have to grovel around in the kstat whenever we wish to 900 * manipulate them. For these variables, we therefore define them to be in 901 * terms of the statistic variable. This assures that we are not introducing 902 * the possibility of inconsistency by having shadow copies of the variables, 903 * while still allowing the code to be readable. 904 */ 905#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 906#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 907#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 908#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 909#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 910#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 911#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 912#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 913#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 914 915/* compressed size of entire arc */ 916#define arc_compressed_size ARCSTAT(arcstat_compressed_size) 917/* uncompressed size of entire arc */ 918#define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size) 919/* number of bytes in the arc from arc_buf_t's */ 920#define arc_overhead_size ARCSTAT(arcstat_overhead_size) 921 922static int arc_no_grow; /* Don't try to grow cache size */ 923static uint64_t arc_tempreserve; 924static uint64_t arc_loaned_bytes; 925 926typedef struct arc_callback arc_callback_t; 927 928struct arc_callback { 929 void *acb_private; 930 arc_done_func_t *acb_done; 931 arc_buf_t *acb_buf; 932 boolean_t acb_compressed; 933 zio_t *acb_zio_dummy; 934 arc_callback_t *acb_next; 935}; 936 937typedef struct arc_write_callback arc_write_callback_t; 938 939struct arc_write_callback { 940 void *awcb_private; 941 arc_done_func_t *awcb_ready; 942 arc_done_func_t *awcb_children_ready; 943 arc_done_func_t *awcb_physdone; 944 arc_done_func_t *awcb_done; 945 arc_buf_t *awcb_buf; 946}; 947 948/* 949 * ARC buffers are separated into multiple structs as a memory saving measure: 950 * - Common fields struct, always defined, and embedded within it: 951 * - L2-only fields, always allocated but undefined when not in L2ARC 952 * - L1-only fields, only allocated when in L1ARC 953 * 954 * Buffer in L1 Buffer only in L2 955 * +------------------------+ +------------------------+ 956 * | arc_buf_hdr_t | | arc_buf_hdr_t | 957 * | | | | 958 * | | | | 959 * | | | | 960 * +------------------------+ +------------------------+ 961 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 962 * | (undefined if L1-only) | | | 963 * +------------------------+ +------------------------+ 964 * | l1arc_buf_hdr_t | 965 * | | 966 * | | 967 * | | 968 * | | 969 * +------------------------+ 970 * 971 * Because it's possible for the L2ARC to become extremely large, we can wind 972 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 973 * is minimized by only allocating the fields necessary for an L1-cached buffer 974 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 975 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 976 * words in pointers. arc_hdr_realloc() is used to switch a header between 977 * these two allocation states. 978 */ 979typedef struct l1arc_buf_hdr { 980 kmutex_t b_freeze_lock; 981 zio_cksum_t *b_freeze_cksum; 982#ifdef ZFS_DEBUG 983 /* 984 * Used for debugging with kmem_flags - by allocating and freeing 985 * b_thawed when the buffer is thawed, we get a record of the stack 986 * trace that thawed it. 987 */ 988 void *b_thawed; 989#endif 990 991 arc_buf_t *b_buf; 992 uint32_t b_bufcnt; 993 /* for waiting on writes to complete */ 994 kcondvar_t b_cv; 995 uint8_t b_byteswap; 996 997 /* protected by arc state mutex */ 998 arc_state_t *b_state; 999 multilist_node_t b_arc_node; 1000 1001 /* updated atomically */ 1002 clock_t b_arc_access; 1003 1004 /* self protecting */ 1005 refcount_t b_refcnt; 1006 1007 arc_callback_t *b_acb; 1008 abd_t *b_pabd; 1009} l1arc_buf_hdr_t; 1010 1011typedef struct l2arc_dev l2arc_dev_t; 1012 1013typedef struct l2arc_buf_hdr { 1014 /* protected by arc_buf_hdr mutex */ 1015 l2arc_dev_t *b_dev; /* L2ARC device */ 1016 uint64_t b_daddr; /* disk address, offset byte */ 1017 1018 list_node_t b_l2node; 1019} l2arc_buf_hdr_t; 1020 1021struct arc_buf_hdr { 1022 /* protected by hash lock */ 1023 dva_t b_dva; 1024 uint64_t b_birth; 1025 1026 arc_buf_contents_t b_type; 1027 arc_buf_hdr_t *b_hash_next; 1028 arc_flags_t b_flags; 1029 1030 /* 1031 * This field stores the size of the data buffer after 1032 * compression, and is set in the arc's zio completion handlers. 1033 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes). 1034 * 1035 * While the block pointers can store up to 32MB in their psize 1036 * field, we can only store up to 32MB minus 512B. This is due 1037 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e. 1038 * a field of zeros represents 512B in the bp). We can't use a 1039 * bias of 1 since we need to reserve a psize of zero, here, to 1040 * represent holes and embedded blocks. 1041 * 1042 * This isn't a problem in practice, since the maximum size of a 1043 * buffer is limited to 16MB, so we never need to store 32MB in 1044 * this field. Even in the upstream illumos code base, the 1045 * maximum size of a buffer is limited to 16MB. 1046 */ 1047 uint16_t b_psize; 1048 1049 /* 1050 * This field stores the size of the data buffer before 1051 * compression, and cannot change once set. It is in units 1052 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes) 1053 */ 1054 uint16_t b_lsize; /* immutable */ 1055 uint64_t b_spa; /* immutable */ 1056 1057 /* L2ARC fields. Undefined when not in L2ARC. */ 1058 l2arc_buf_hdr_t b_l2hdr; 1059 /* L1ARC fields. Undefined when in l2arc_only state */ 1060 l1arc_buf_hdr_t b_l1hdr; 1061}; 1062 1063#if defined(__FreeBSD__) && defined(_KERNEL) 1064static int 1065sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 1066{ 1067 uint64_t val; 1068 int err; 1069 1070 val = arc_meta_limit; 1071 err = sysctl_handle_64(oidp, &val, 0, req); 1072 if (err != 0 || req->newptr == NULL) 1073 return (err); 1074 1075 if (val <= 0 || val > arc_c_max) 1076 return (EINVAL); 1077 1078 arc_meta_limit = val; 1079 return (0); 1080} 1081 1082static int 1083sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS) 1084{ 1085 uint32_t val; 1086 int err; 1087 1088 val = arc_no_grow_shift; 1089 err = sysctl_handle_32(oidp, &val, 0, req); 1090 if (err != 0 || req->newptr == NULL) 1091 return (err); 1092 1093 if (val >= arc_shrink_shift) 1094 return (EINVAL); 1095 1096 arc_no_grow_shift = val; 1097 return (0); 1098} 1099 1100static int 1101sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS) 1102{ 1103 uint64_t val; 1104 int err; 1105 1106 val = zfs_arc_max; 1107 err = sysctl_handle_64(oidp, &val, 0, req); 1108 if (err != 0 || req->newptr == NULL) 1109 return (err); 1110 1111 if (zfs_arc_max == 0) { 1112 /* Loader tunable so blindly set */ 1113 zfs_arc_max = val; 1114 return (0); 1115 } 1116 1117 if (val < arc_abs_min || val > kmem_size()) 1118 return (EINVAL); 1119 if (val < arc_c_min) 1120 return (EINVAL); 1121 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit) 1122 return (EINVAL); 1123 1124 arc_c_max = val; 1125 1126 arc_c = arc_c_max; 1127 arc_p = (arc_c >> 1); 1128 1129 if (zfs_arc_meta_limit == 0) { 1130 /* limit meta-data to 1/4 of the arc capacity */ 1131 arc_meta_limit = arc_c_max / 4; 1132 } 1133 1134 /* if kmem_flags are set, lets try to use less memory */ 1135 if (kmem_debugging()) 1136 arc_c = arc_c / 2; 1137 1138 zfs_arc_max = arc_c; 1139 1140 return (0); 1141} 1142 1143static int 1144sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS) 1145{ 1146 uint64_t val; 1147 int err; 1148 1149 val = zfs_arc_min; 1150 err = sysctl_handle_64(oidp, &val, 0, req); 1151 if (err != 0 || req->newptr == NULL) 1152 return (err); 1153 1154 if (zfs_arc_min == 0) { 1155 /* Loader tunable so blindly set */ 1156 zfs_arc_min = val; 1157 return (0); 1158 } 1159 1160 if (val < arc_abs_min || val > arc_c_max) 1161 return (EINVAL); 1162 1163 arc_c_min = val; 1164 1165 if (zfs_arc_meta_min == 0) 1166 arc_meta_min = arc_c_min / 2; 1167 1168 if (arc_c < arc_c_min) 1169 arc_c = arc_c_min; 1170 1171 zfs_arc_min = arc_c_min; 1172 1173 return (0); 1174} 1175#endif 1176 1177#define GHOST_STATE(state) \ 1178 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 1179 (state) == arc_l2c_only) 1180 1181#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 1182#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 1183#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 1184#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 1185#define HDR_COMPRESSION_ENABLED(hdr) \ 1186 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC) 1187 1188#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 1189#define HDR_L2_READING(hdr) \ 1190 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 1191 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 1192#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 1193#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 1194#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 1195#define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA) 1196 1197#define HDR_ISTYPE_METADATA(hdr) \ 1198 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 1199#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 1200 1201#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 1202#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 1203 1204/* For storing compression mode in b_flags */ 1205#define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1) 1206 1207#define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \ 1208 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS)) 1209#define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \ 1210 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp)); 1211 1212#define ARC_BUF_LAST(buf) ((buf)->b_next == NULL) 1213#define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED) 1214#define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED) 1215 1216/* 1217 * Other sizes 1218 */ 1219 1220#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 1221#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 1222 1223/* 1224 * Hash table routines 1225 */ 1226 1227#define HT_LOCK_PAD CACHE_LINE_SIZE 1228 1229struct ht_lock { 1230 kmutex_t ht_lock; 1231#ifdef _KERNEL 1232 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 1233#endif 1234}; 1235 1236#define BUF_LOCKS 256 1237typedef struct buf_hash_table { 1238 uint64_t ht_mask; 1239 arc_buf_hdr_t **ht_table; 1240 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 1241} buf_hash_table_t; 1242 1243static buf_hash_table_t buf_hash_table; 1244 1245#define BUF_HASH_INDEX(spa, dva, birth) \ 1246 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 1247#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 1248#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 1249#define HDR_LOCK(hdr) \ 1250 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 1251 1252uint64_t zfs_crc64_table[256]; 1253 1254/* 1255 * Level 2 ARC 1256 */ 1257 1258#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 1259#define L2ARC_HEADROOM 2 /* num of writes */ 1260/* 1261 * If we discover during ARC scan any buffers to be compressed, we boost 1262 * our headroom for the next scanning cycle by this percentage multiple. 1263 */ 1264#define L2ARC_HEADROOM_BOOST 200 1265#define L2ARC_FEED_SECS 1 /* caching interval secs */ 1266#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 1267 1268#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 1269#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 1270 1271/* L2ARC Performance Tunables */ 1272uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1273uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1274uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1275uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1276uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1277uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1278boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1279boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1280boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1281 1282SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 1283 &l2arc_write_max, 0, "max write size"); 1284SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 1285 &l2arc_write_boost, 0, "extra write during warmup"); 1286SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 1287 &l2arc_headroom, 0, "number of dev writes"); 1288SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 1289 &l2arc_feed_secs, 0, "interval seconds"); 1290SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 1291 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 1292 1293SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 1294 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 1295SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 1296 &l2arc_feed_again, 0, "turbo warmup"); 1297SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 1298 &l2arc_norw, 0, "no reads during writes"); 1299 1300SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 1301 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state"); 1302SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD, 1303 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1304 "size of anonymous state"); 1305SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD, 1306 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1307 "size of anonymous state"); 1308 1309SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 1310 &ARC_mru.arcs_size.rc_count, 0, "size of mru state"); 1311SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD, 1312 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1313 "size of metadata in mru state"); 1314SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD, 1315 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1316 "size of data in mru state"); 1317 1318SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 1319 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state"); 1320SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD, 1321 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1322 "size of metadata in mru ghost state"); 1323SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD, 1324 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1325 "size of data in mru ghost state"); 1326 1327SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 1328 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state"); 1329SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD, 1330 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1331 "size of metadata in mfu state"); 1332SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD, 1333 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1334 "size of data in mfu state"); 1335 1336SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 1337 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state"); 1338SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD, 1339 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1340 "size of metadata in mfu ghost state"); 1341SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD, 1342 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1343 "size of data in mfu ghost state"); 1344 1345SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 1346 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state"); 1347 1348/* 1349 * L2ARC Internals 1350 */ 1351struct l2arc_dev { 1352 vdev_t *l2ad_vdev; /* vdev */ 1353 spa_t *l2ad_spa; /* spa */ 1354 uint64_t l2ad_hand; /* next write location */ 1355 uint64_t l2ad_start; /* first addr on device */ 1356 uint64_t l2ad_end; /* last addr on device */ 1357 boolean_t l2ad_first; /* first sweep through */ 1358 boolean_t l2ad_writing; /* currently writing */ 1359 kmutex_t l2ad_mtx; /* lock for buffer list */ 1360 list_t l2ad_buflist; /* buffer list */ 1361 list_node_t l2ad_node; /* device list node */ 1362 refcount_t l2ad_alloc; /* allocated bytes */ 1363}; 1364 1365static list_t L2ARC_dev_list; /* device list */ 1366static list_t *l2arc_dev_list; /* device list pointer */ 1367static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1368static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1369static list_t L2ARC_free_on_write; /* free after write buf list */ 1370static list_t *l2arc_free_on_write; /* free after write list ptr */ 1371static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1372static uint64_t l2arc_ndev; /* number of devices */ 1373 1374typedef struct l2arc_read_callback { 1375 arc_buf_hdr_t *l2rcb_hdr; /* read header */ 1376 blkptr_t l2rcb_bp; /* original blkptr */ 1377 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1378 int l2rcb_flags; /* original flags */ 1379 abd_t *l2rcb_abd; /* temporary buffer */ 1380} l2arc_read_callback_t; 1381 1382typedef struct l2arc_write_callback { 1383 l2arc_dev_t *l2wcb_dev; /* device info */ 1384 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1385} l2arc_write_callback_t; 1386 1387typedef struct l2arc_data_free { 1388 /* protected by l2arc_free_on_write_mtx */ 1389 abd_t *l2df_abd; 1390 size_t l2df_size; 1391 arc_buf_contents_t l2df_type; 1392 list_node_t l2df_list_node; 1393} l2arc_data_free_t; 1394 1395static kmutex_t l2arc_feed_thr_lock; 1396static kcondvar_t l2arc_feed_thr_cv; 1397static uint8_t l2arc_thread_exit; 1398 1399static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *); 1400static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); 1401static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *); 1402static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *); 1403static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); 1404static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag); 1405static void arc_hdr_free_pabd(arc_buf_hdr_t *); 1406static void arc_hdr_alloc_pabd(arc_buf_hdr_t *); 1407static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1408static boolean_t arc_is_overflowing(); 1409static void arc_buf_watch(arc_buf_t *); 1410 1411static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1412static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1413static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1414static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1415 1416static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1417static void l2arc_read_done(zio_t *); 1418 1419static void 1420l2arc_trim(const arc_buf_hdr_t *hdr) 1421{ 1422 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1423 1424 ASSERT(HDR_HAS_L2HDR(hdr)); 1425 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 1426 1427 if (HDR_GET_PSIZE(hdr) != 0) { 1428 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr, 1429 HDR_GET_PSIZE(hdr), 0); 1430 } 1431} 1432 1433static uint64_t 1434buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1435{ 1436 uint8_t *vdva = (uint8_t *)dva; 1437 uint64_t crc = -1ULL; 1438 int i; 1439 1440 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1441 1442 for (i = 0; i < sizeof (dva_t); i++) 1443 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1444 1445 crc ^= (spa>>8) ^ birth; 1446 1447 return (crc); 1448} 1449 1450#define HDR_EMPTY(hdr) \ 1451 ((hdr)->b_dva.dva_word[0] == 0 && \ 1452 (hdr)->b_dva.dva_word[1] == 0) 1453 1454#define HDR_EQUAL(spa, dva, birth, hdr) \ 1455 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1456 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1457 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) 1458 1459static void 1460buf_discard_identity(arc_buf_hdr_t *hdr) 1461{ 1462 hdr->b_dva.dva_word[0] = 0; 1463 hdr->b_dva.dva_word[1] = 0; 1464 hdr->b_birth = 0; 1465} 1466 1467static arc_buf_hdr_t * 1468buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1469{ 1470 const dva_t *dva = BP_IDENTITY(bp); 1471 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1472 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1473 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1474 arc_buf_hdr_t *hdr; 1475 1476 mutex_enter(hash_lock); 1477 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1478 hdr = hdr->b_hash_next) { 1479 if (HDR_EQUAL(spa, dva, birth, hdr)) { 1480 *lockp = hash_lock; 1481 return (hdr); 1482 } 1483 } 1484 mutex_exit(hash_lock); 1485 *lockp = NULL; 1486 return (NULL); 1487} 1488 1489/* 1490 * Insert an entry into the hash table. If there is already an element 1491 * equal to elem in the hash table, then the already existing element 1492 * will be returned and the new element will not be inserted. 1493 * Otherwise returns NULL. 1494 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1495 */ 1496static arc_buf_hdr_t * 1497buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1498{ 1499 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1500 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1501 arc_buf_hdr_t *fhdr; 1502 uint32_t i; 1503 1504 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1505 ASSERT(hdr->b_birth != 0); 1506 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1507 1508 if (lockp != NULL) { 1509 *lockp = hash_lock; 1510 mutex_enter(hash_lock); 1511 } else { 1512 ASSERT(MUTEX_HELD(hash_lock)); 1513 } 1514 1515 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1516 fhdr = fhdr->b_hash_next, i++) { 1517 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1518 return (fhdr); 1519 } 1520 1521 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1522 buf_hash_table.ht_table[idx] = hdr; 1523 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1524 1525 /* collect some hash table performance data */ 1526 if (i > 0) { 1527 ARCSTAT_BUMP(arcstat_hash_collisions); 1528 if (i == 1) 1529 ARCSTAT_BUMP(arcstat_hash_chains); 1530 1531 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1532 } 1533 1534 ARCSTAT_BUMP(arcstat_hash_elements); 1535 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1536 1537 return (NULL); 1538} 1539 1540static void 1541buf_hash_remove(arc_buf_hdr_t *hdr) 1542{ 1543 arc_buf_hdr_t *fhdr, **hdrp; 1544 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1545 1546 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1547 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1548 1549 hdrp = &buf_hash_table.ht_table[idx]; 1550 while ((fhdr = *hdrp) != hdr) { 1551 ASSERT3P(fhdr, !=, NULL); 1552 hdrp = &fhdr->b_hash_next; 1553 } 1554 *hdrp = hdr->b_hash_next; 1555 hdr->b_hash_next = NULL; 1556 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1557 1558 /* collect some hash table performance data */ 1559 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1560 1561 if (buf_hash_table.ht_table[idx] && 1562 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1563 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1564} 1565 1566/* 1567 * Global data structures and functions for the buf kmem cache. 1568 */ 1569static kmem_cache_t *hdr_full_cache; 1570static kmem_cache_t *hdr_l2only_cache; 1571static kmem_cache_t *buf_cache; 1572 1573static void 1574buf_fini(void) 1575{ 1576 int i; 1577 1578 kmem_free(buf_hash_table.ht_table, 1579 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1580 for (i = 0; i < BUF_LOCKS; i++) 1581 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1582 kmem_cache_destroy(hdr_full_cache); 1583 kmem_cache_destroy(hdr_l2only_cache); 1584 kmem_cache_destroy(buf_cache); 1585} 1586 1587/* 1588 * Constructor callback - called when the cache is empty 1589 * and a new buf is requested. 1590 */ 1591/* ARGSUSED */ 1592static int 1593hdr_full_cons(void *vbuf, void *unused, int kmflag) 1594{ 1595 arc_buf_hdr_t *hdr = vbuf; 1596 1597 bzero(hdr, HDR_FULL_SIZE); 1598 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1599 refcount_create(&hdr->b_l1hdr.b_refcnt); 1600 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1601 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1602 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1603 1604 return (0); 1605} 1606 1607/* ARGSUSED */ 1608static int 1609hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1610{ 1611 arc_buf_hdr_t *hdr = vbuf; 1612 1613 bzero(hdr, HDR_L2ONLY_SIZE); 1614 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1615 1616 return (0); 1617} 1618 1619/* ARGSUSED */ 1620static int 1621buf_cons(void *vbuf, void *unused, int kmflag) 1622{ 1623 arc_buf_t *buf = vbuf; 1624 1625 bzero(buf, sizeof (arc_buf_t)); 1626 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1627 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1628 1629 return (0); 1630} 1631 1632/* 1633 * Destructor callback - called when a cached buf is 1634 * no longer required. 1635 */ 1636/* ARGSUSED */ 1637static void 1638hdr_full_dest(void *vbuf, void *unused) 1639{ 1640 arc_buf_hdr_t *hdr = vbuf; 1641 1642 ASSERT(HDR_EMPTY(hdr)); 1643 cv_destroy(&hdr->b_l1hdr.b_cv); 1644 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1645 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1646 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1647 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1648} 1649 1650/* ARGSUSED */ 1651static void 1652hdr_l2only_dest(void *vbuf, void *unused) 1653{ 1654 arc_buf_hdr_t *hdr = vbuf; 1655 1656 ASSERT(HDR_EMPTY(hdr)); 1657 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1658} 1659 1660/* ARGSUSED */ 1661static void 1662buf_dest(void *vbuf, void *unused) 1663{ 1664 arc_buf_t *buf = vbuf; 1665 1666 mutex_destroy(&buf->b_evict_lock); 1667 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1668} 1669 1670/* 1671 * Reclaim callback -- invoked when memory is low. 1672 */ 1673/* ARGSUSED */ 1674static void 1675hdr_recl(void *unused) 1676{ 1677 dprintf("hdr_recl called\n"); 1678 /* 1679 * umem calls the reclaim func when we destroy the buf cache, 1680 * which is after we do arc_fini(). 1681 */ 1682 if (!arc_dead) 1683 cv_signal(&arc_reclaim_thread_cv); 1684} 1685 1686static void 1687buf_init(void) 1688{ 1689 uint64_t *ct; 1690 uint64_t hsize = 1ULL << 12; 1691 int i, j; 1692 1693 /* 1694 * The hash table is big enough to fill all of physical memory 1695 * with an average block size of zfs_arc_average_blocksize (default 8K). 1696 * By default, the table will take up 1697 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1698 */ 1699 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1700 hsize <<= 1; 1701retry: 1702 buf_hash_table.ht_mask = hsize - 1; 1703 buf_hash_table.ht_table = 1704 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1705 if (buf_hash_table.ht_table == NULL) { 1706 ASSERT(hsize > (1ULL << 8)); 1707 hsize >>= 1; 1708 goto retry; 1709 } 1710 1711 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1712 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1713 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1714 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1715 NULL, NULL, 0); 1716 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1717 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1718 1719 for (i = 0; i < 256; i++) 1720 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1721 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1722 1723 for (i = 0; i < BUF_LOCKS; i++) { 1724 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1725 NULL, MUTEX_DEFAULT, NULL); 1726 } 1727} 1728 1729/* 1730 * This is the size that the buf occupies in memory. If the buf is compressed, 1731 * it will correspond to the compressed size. You should use this method of 1732 * getting the buf size unless you explicitly need the logical size. 1733 */ 1734int32_t 1735arc_buf_size(arc_buf_t *buf) 1736{ 1737 return (ARC_BUF_COMPRESSED(buf) ? 1738 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr)); 1739} 1740 1741int32_t 1742arc_buf_lsize(arc_buf_t *buf) 1743{ 1744 return (HDR_GET_LSIZE(buf->b_hdr)); 1745} 1746 1747enum zio_compress 1748arc_get_compression(arc_buf_t *buf) 1749{ 1750 return (ARC_BUF_COMPRESSED(buf) ? 1751 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF); 1752} 1753 1754#define ARC_MINTIME (hz>>4) /* 62 ms */ 1755 1756static inline boolean_t 1757arc_buf_is_shared(arc_buf_t *buf) 1758{ 1759 boolean_t shared = (buf->b_data != NULL && 1760 buf->b_hdr->b_l1hdr.b_pabd != NULL && 1761 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) && 1762 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd)); 1763 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); 1764 IMPLY(shared, ARC_BUF_SHARED(buf)); 1765 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf)); 1766 1767 /* 1768 * It would be nice to assert arc_can_share() too, but the "hdr isn't 1769 * already being shared" requirement prevents us from doing that. 1770 */ 1771 1772 return (shared); 1773} 1774 1775/* 1776 * Free the checksum associated with this header. If there is no checksum, this 1777 * is a no-op. 1778 */ 1779static inline void 1780arc_cksum_free(arc_buf_hdr_t *hdr) 1781{ 1782 ASSERT(HDR_HAS_L1HDR(hdr)); 1783 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1784 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1785 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); 1786 hdr->b_l1hdr.b_freeze_cksum = NULL; 1787 } 1788 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1789} 1790 1791/* 1792 * Return true iff at least one of the bufs on hdr is not compressed. 1793 */ 1794static boolean_t 1795arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr) 1796{ 1797 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) { 1798 if (!ARC_BUF_COMPRESSED(b)) { 1799 return (B_TRUE); 1800 } 1801 } 1802 return (B_FALSE); 1803} 1804 1805/* 1806 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data 1807 * matches the checksum that is stored in the hdr. If there is no checksum, 1808 * or if the buf is compressed, this is a no-op. 1809 */ 1810static void 1811arc_cksum_verify(arc_buf_t *buf) 1812{ 1813 arc_buf_hdr_t *hdr = buf->b_hdr; 1814 zio_cksum_t zc; 1815 1816 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1817 return; 1818 1819 if (ARC_BUF_COMPRESSED(buf)) { 1820 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1821 arc_hdr_has_uncompressed_buf(hdr)); 1822 return; 1823 } 1824 1825 ASSERT(HDR_HAS_L1HDR(hdr)); 1826 1827 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1828 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { 1829 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1830 return; 1831 } 1832 1833 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc); 1834 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) 1835 panic("buffer modified while frozen!"); 1836 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1837} 1838 1839static boolean_t 1840arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) 1841{ 1842 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp); 1843 boolean_t valid_cksum; 1844 1845 ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); 1846 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); 1847 1848 /* 1849 * We rely on the blkptr's checksum to determine if the block 1850 * is valid or not. When compressed arc is enabled, the l2arc 1851 * writes the block to the l2arc just as it appears in the pool. 1852 * This allows us to use the blkptr's checksum to validate the 1853 * data that we just read off of the l2arc without having to store 1854 * a separate checksum in the arc_buf_hdr_t. However, if compressed 1855 * arc is disabled, then the data written to the l2arc is always 1856 * uncompressed and won't match the block as it exists in the main 1857 * pool. When this is the case, we must first compress it if it is 1858 * compressed on the main pool before we can validate the checksum. 1859 */ 1860 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) { 1861 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1862 uint64_t lsize = HDR_GET_LSIZE(hdr); 1863 uint64_t csize; 1864 1865 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE); 1866 csize = zio_compress_data(compress, zio->io_abd, 1867 abd_to_buf(cdata), lsize); 1868 1869 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr)); 1870 if (csize < HDR_GET_PSIZE(hdr)) { 1871 /* 1872 * Compressed blocks are always a multiple of the 1873 * smallest ashift in the pool. Ideally, we would 1874 * like to round up the csize to the next 1875 * spa_min_ashift but that value may have changed 1876 * since the block was last written. Instead, 1877 * we rely on the fact that the hdr's psize 1878 * was set to the psize of the block when it was 1879 * last written. We set the csize to that value 1880 * and zero out any part that should not contain 1881 * data. 1882 */ 1883 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize); 1884 csize = HDR_GET_PSIZE(hdr); 1885 } 1886 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL); 1887 } 1888 1889 /* 1890 * Block pointers always store the checksum for the logical data. 1891 * If the block pointer has the gang bit set, then the checksum 1892 * it represents is for the reconstituted data and not for an 1893 * individual gang member. The zio pipeline, however, must be able to 1894 * determine the checksum of each of the gang constituents so it 1895 * treats the checksum comparison differently than what we need 1896 * for l2arc blocks. This prevents us from using the 1897 * zio_checksum_error() interface directly. Instead we must call the 1898 * zio_checksum_error_impl() so that we can ensure the checksum is 1899 * generated using the correct checksum algorithm and accounts for the 1900 * logical I/O size and not just a gang fragment. 1901 */ 1902 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp, 1903 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size, 1904 zio->io_offset, NULL) == 0); 1905 zio_pop_transforms(zio); 1906 return (valid_cksum); 1907} 1908 1909/* 1910 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a 1911 * checksum and attaches it to the buf's hdr so that we can ensure that the buf 1912 * isn't modified later on. If buf is compressed or there is already a checksum 1913 * on the hdr, this is a no-op (we only checksum uncompressed bufs). 1914 */ 1915static void 1916arc_cksum_compute(arc_buf_t *buf) 1917{ 1918 arc_buf_hdr_t *hdr = buf->b_hdr; 1919 1920 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1921 return; 1922 1923 ASSERT(HDR_HAS_L1HDR(hdr)); 1924 1925 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1926 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1927 ASSERT(arc_hdr_has_uncompressed_buf(hdr)); 1928 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1929 return; 1930 } else if (ARC_BUF_COMPRESSED(buf)) { 1931 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1932 return; 1933 } 1934 1935 ASSERT(!ARC_BUF_COMPRESSED(buf)); 1936 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), 1937 KM_SLEEP); 1938 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, 1939 hdr->b_l1hdr.b_freeze_cksum); 1940 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1941#ifdef illumos 1942 arc_buf_watch(buf); 1943#endif 1944} 1945 1946#ifdef illumos 1947#ifndef _KERNEL 1948typedef struct procctl { 1949 long cmd; 1950 prwatch_t prwatch; 1951} procctl_t; 1952#endif 1953 1954/* ARGSUSED */ 1955static void 1956arc_buf_unwatch(arc_buf_t *buf) 1957{ 1958#ifndef _KERNEL 1959 if (arc_watch) { 1960 int result; 1961 procctl_t ctl; 1962 ctl.cmd = PCWATCH; 1963 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1964 ctl.prwatch.pr_size = 0; 1965 ctl.prwatch.pr_wflags = 0; 1966 result = write(arc_procfd, &ctl, sizeof (ctl)); 1967 ASSERT3U(result, ==, sizeof (ctl)); 1968 } 1969#endif 1970} 1971 1972/* ARGSUSED */ 1973static void 1974arc_buf_watch(arc_buf_t *buf) 1975{ 1976#ifndef _KERNEL 1977 if (arc_watch) { 1978 int result; 1979 procctl_t ctl; 1980 ctl.cmd = PCWATCH; 1981 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1982 ctl.prwatch.pr_size = arc_buf_size(buf); 1983 ctl.prwatch.pr_wflags = WA_WRITE; 1984 result = write(arc_procfd, &ctl, sizeof (ctl)); 1985 ASSERT3U(result, ==, sizeof (ctl)); 1986 } 1987#endif 1988} 1989#endif /* illumos */ 1990 1991static arc_buf_contents_t 1992arc_buf_type(arc_buf_hdr_t *hdr) 1993{ 1994 arc_buf_contents_t type; 1995 if (HDR_ISTYPE_METADATA(hdr)) { 1996 type = ARC_BUFC_METADATA; 1997 } else { 1998 type = ARC_BUFC_DATA; 1999 } 2000 VERIFY3U(hdr->b_type, ==, type); 2001 return (type); 2002} 2003 2004boolean_t 2005arc_is_metadata(arc_buf_t *buf) 2006{ 2007 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0); 2008} 2009 2010static uint32_t 2011arc_bufc_to_flags(arc_buf_contents_t type) 2012{ 2013 switch (type) { 2014 case ARC_BUFC_DATA: 2015 /* metadata field is 0 if buffer contains normal data */ 2016 return (0); 2017 case ARC_BUFC_METADATA: 2018 return (ARC_FLAG_BUFC_METADATA); 2019 default: 2020 break; 2021 } 2022 panic("undefined ARC buffer type!"); 2023 return ((uint32_t)-1); 2024} 2025 2026void 2027arc_buf_thaw(arc_buf_t *buf) 2028{ 2029 arc_buf_hdr_t *hdr = buf->b_hdr; 2030 2031 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2032 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2033 2034 arc_cksum_verify(buf); 2035 2036 /* 2037 * Compressed buffers do not manipulate the b_freeze_cksum or 2038 * allocate b_thawed. 2039 */ 2040 if (ARC_BUF_COMPRESSED(buf)) { 2041 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 2042 arc_hdr_has_uncompressed_buf(hdr)); 2043 return; 2044 } 2045 2046 ASSERT(HDR_HAS_L1HDR(hdr)); 2047 arc_cksum_free(hdr); 2048 2049 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 2050#ifdef ZFS_DEBUG 2051 if (zfs_flags & ZFS_DEBUG_MODIFY) { 2052 if (hdr->b_l1hdr.b_thawed != NULL) 2053 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2054 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 2055 } 2056#endif 2057 2058 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 2059 2060#ifdef illumos 2061 arc_buf_unwatch(buf); 2062#endif 2063} 2064 2065void 2066arc_buf_freeze(arc_buf_t *buf) 2067{ 2068 arc_buf_hdr_t *hdr = buf->b_hdr; 2069 kmutex_t *hash_lock; 2070 2071 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 2072 return; 2073 2074 if (ARC_BUF_COMPRESSED(buf)) { 2075 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 2076 arc_hdr_has_uncompressed_buf(hdr)); 2077 return; 2078 } 2079 2080 hash_lock = HDR_LOCK(hdr); 2081 mutex_enter(hash_lock); 2082 2083 ASSERT(HDR_HAS_L1HDR(hdr)); 2084 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL || 2085 hdr->b_l1hdr.b_state == arc_anon); 2086 arc_cksum_compute(buf); 2087 mutex_exit(hash_lock); 2088} 2089 2090/* 2091 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, 2092 * the following functions should be used to ensure that the flags are 2093 * updated in a thread-safe way. When manipulating the flags either 2094 * the hash_lock must be held or the hdr must be undiscoverable. This 2095 * ensures that we're not racing with any other threads when updating 2096 * the flags. 2097 */ 2098static inline void 2099arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 2100{ 2101 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2102 hdr->b_flags |= flags; 2103} 2104 2105static inline void 2106arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 2107{ 2108 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2109 hdr->b_flags &= ~flags; 2110} 2111 2112/* 2113 * Setting the compression bits in the arc_buf_hdr_t's b_flags is 2114 * done in a special way since we have to clear and set bits 2115 * at the same time. Consumers that wish to set the compression bits 2116 * must use this function to ensure that the flags are updated in 2117 * thread-safe manner. 2118 */ 2119static void 2120arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) 2121{ 2122 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2123 2124 /* 2125 * Holes and embedded blocks will always have a psize = 0 so 2126 * we ignore the compression of the blkptr and set the 2127 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF. 2128 * Holes and embedded blocks remain anonymous so we don't 2129 * want to uncompress them. Mark them as uncompressed. 2130 */ 2131 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { 2132 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 2133 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 2134 ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); 2135 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 2136 } else { 2137 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 2138 HDR_SET_COMPRESS(hdr, cmp); 2139 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); 2140 ASSERT(HDR_COMPRESSION_ENABLED(hdr)); 2141 } 2142} 2143 2144/* 2145 * Looks for another buf on the same hdr which has the data decompressed, copies 2146 * from it, and returns true. If no such buf exists, returns false. 2147 */ 2148static boolean_t 2149arc_buf_try_copy_decompressed_data(arc_buf_t *buf) 2150{ 2151 arc_buf_hdr_t *hdr = buf->b_hdr; 2152 boolean_t copied = B_FALSE; 2153 2154 ASSERT(HDR_HAS_L1HDR(hdr)); 2155 ASSERT3P(buf->b_data, !=, NULL); 2156 ASSERT(!ARC_BUF_COMPRESSED(buf)); 2157 2158 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL; 2159 from = from->b_next) { 2160 /* can't use our own data buffer */ 2161 if (from == buf) { 2162 continue; 2163 } 2164 2165 if (!ARC_BUF_COMPRESSED(from)) { 2166 bcopy(from->b_data, buf->b_data, arc_buf_size(buf)); 2167 copied = B_TRUE; 2168 break; 2169 } 2170 } 2171 2172 /* 2173 * There were no decompressed bufs, so there should not be a 2174 * checksum on the hdr either. 2175 */ 2176 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL); 2177 2178 return (copied); 2179} 2180 2181/* 2182 * Given a buf that has a data buffer attached to it, this function will 2183 * efficiently fill the buf with data of the specified compression setting from 2184 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr 2185 * are already sharing a data buf, no copy is performed. 2186 * 2187 * If the buf is marked as compressed but uncompressed data was requested, this 2188 * will allocate a new data buffer for the buf, remove that flag, and fill the 2189 * buf with uncompressed data. You can't request a compressed buf on a hdr with 2190 * uncompressed data, and (since we haven't added support for it yet) if you 2191 * want compressed data your buf must already be marked as compressed and have 2192 * the correct-sized data buffer. 2193 */ 2194static int 2195arc_buf_fill(arc_buf_t *buf, boolean_t compressed) 2196{ 2197 arc_buf_hdr_t *hdr = buf->b_hdr; 2198 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 2199 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; 2200 2201 ASSERT3P(buf->b_data, !=, NULL); 2202 IMPLY(compressed, hdr_compressed); 2203 IMPLY(compressed, ARC_BUF_COMPRESSED(buf)); 2204 2205 if (hdr_compressed == compressed) { 2206 if (!arc_buf_is_shared(buf)) { 2207 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd, 2208 arc_buf_size(buf)); 2209 } 2210 } else { 2211 ASSERT(hdr_compressed); 2212 ASSERT(!compressed); 2213 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); 2214 2215 /* 2216 * If the buf is sharing its data with the hdr, unlink it and 2217 * allocate a new data buffer for the buf. 2218 */ 2219 if (arc_buf_is_shared(buf)) { 2220 ASSERT(ARC_BUF_COMPRESSED(buf)); 2221 2222 /* We need to give the buf it's own b_data */ 2223 buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 2224 buf->b_data = 2225 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 2226 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2227 2228 /* Previously overhead was 0; just add new overhead */ 2229 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 2230 } else if (ARC_BUF_COMPRESSED(buf)) { 2231 /* We need to reallocate the buf's b_data */ 2232 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr), 2233 buf); 2234 buf->b_data = 2235 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 2236 2237 /* We increased the size of b_data; update overhead */ 2238 ARCSTAT_INCR(arcstat_overhead_size, 2239 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr)); 2240 } 2241 2242 /* 2243 * Regardless of the buf's previous compression settings, it 2244 * should not be compressed at the end of this function. 2245 */ 2246 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; 2247 2248 /* 2249 * Try copying the data from another buf which already has a 2250 * decompressed version. If that's not possible, it's time to 2251 * bite the bullet and decompress the data from the hdr. 2252 */ 2253 if (arc_buf_try_copy_decompressed_data(buf)) { 2254 /* Skip byteswapping and checksumming (already done) */ 2255 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL); 2256 return (0); 2257 } else { 2258 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr), 2259 hdr->b_l1hdr.b_pabd, buf->b_data, 2260 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 2261 2262 /* 2263 * Absent hardware errors or software bugs, this should 2264 * be impossible, but log it anyway so we can debug it. 2265 */ 2266 if (error != 0) { 2267 zfs_dbgmsg( 2268 "hdr %p, compress %d, psize %d, lsize %d", 2269 hdr, HDR_GET_COMPRESS(hdr), 2270 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 2271 return (SET_ERROR(EIO)); 2272 } 2273 } 2274 } 2275 2276 /* Byteswap the buf's data if necessary */ 2277 if (bswap != DMU_BSWAP_NUMFUNCS) { 2278 ASSERT(!HDR_SHARED_DATA(hdr)); 2279 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); 2280 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); 2281 } 2282 2283 /* Compute the hdr's checksum if necessary */ 2284 arc_cksum_compute(buf); 2285 2286 return (0); 2287} 2288 2289int 2290arc_decompress(arc_buf_t *buf) 2291{ 2292 return (arc_buf_fill(buf, B_FALSE)); 2293} 2294 2295/* 2296 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t. 2297 */ 2298static uint64_t 2299arc_hdr_size(arc_buf_hdr_t *hdr) 2300{ 2301 uint64_t size; 2302 2303 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 2304 HDR_GET_PSIZE(hdr) > 0) { 2305 size = HDR_GET_PSIZE(hdr); 2306 } else { 2307 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); 2308 size = HDR_GET_LSIZE(hdr); 2309 } 2310 return (size); 2311} 2312 2313/* 2314 * Increment the amount of evictable space in the arc_state_t's refcount. 2315 * We account for the space used by the hdr and the arc buf individually 2316 * so that we can add and remove them from the refcount individually. 2317 */ 2318static void 2319arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) 2320{ 2321 arc_buf_contents_t type = arc_buf_type(hdr); 2322 2323 ASSERT(HDR_HAS_L1HDR(hdr)); 2324 2325 if (GHOST_STATE(state)) { 2326 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2327 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2328 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2329 (void) refcount_add_many(&state->arcs_esize[type], 2330 HDR_GET_LSIZE(hdr), hdr); 2331 return; 2332 } 2333 2334 ASSERT(!GHOST_STATE(state)); 2335 if (hdr->b_l1hdr.b_pabd != NULL) { 2336 (void) refcount_add_many(&state->arcs_esize[type], 2337 arc_hdr_size(hdr), hdr); 2338 } 2339 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2340 buf = buf->b_next) { 2341 if (arc_buf_is_shared(buf)) 2342 continue; 2343 (void) refcount_add_many(&state->arcs_esize[type], 2344 arc_buf_size(buf), buf); 2345 } 2346} 2347 2348/* 2349 * Decrement the amount of evictable space in the arc_state_t's refcount. 2350 * We account for the space used by the hdr and the arc buf individually 2351 * so that we can add and remove them from the refcount individually. 2352 */ 2353static void 2354arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) 2355{ 2356 arc_buf_contents_t type = arc_buf_type(hdr); 2357 2358 ASSERT(HDR_HAS_L1HDR(hdr)); 2359 2360 if (GHOST_STATE(state)) { 2361 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2362 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2363 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2364 (void) refcount_remove_many(&state->arcs_esize[type], 2365 HDR_GET_LSIZE(hdr), hdr); 2366 return; 2367 } 2368 2369 ASSERT(!GHOST_STATE(state)); 2370 if (hdr->b_l1hdr.b_pabd != NULL) { 2371 (void) refcount_remove_many(&state->arcs_esize[type], 2372 arc_hdr_size(hdr), hdr); 2373 } 2374 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2375 buf = buf->b_next) { 2376 if (arc_buf_is_shared(buf)) 2377 continue; 2378 (void) refcount_remove_many(&state->arcs_esize[type], 2379 arc_buf_size(buf), buf); 2380 } 2381} 2382 2383/* 2384 * Add a reference to this hdr indicating that someone is actively 2385 * referencing that memory. When the refcount transitions from 0 to 1, 2386 * we remove it from the respective arc_state_t list to indicate that 2387 * it is not evictable. 2388 */ 2389static void 2390add_reference(arc_buf_hdr_t *hdr, void *tag) 2391{ 2392 ASSERT(HDR_HAS_L1HDR(hdr)); 2393 if (!MUTEX_HELD(HDR_LOCK(hdr))) { 2394 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 2395 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2396 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2397 } 2398 2399 arc_state_t *state = hdr->b_l1hdr.b_state; 2400 2401 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 2402 (state != arc_anon)) { 2403 /* We don't use the L2-only state list. */ 2404 if (state != arc_l2c_only) { 2405 multilist_remove(state->arcs_list[arc_buf_type(hdr)], 2406 hdr); 2407 arc_evictable_space_decrement(hdr, state); 2408 } 2409 /* remove the prefetch flag if we get a reference */ 2410 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 2411 } 2412} 2413 2414/* 2415 * Remove a reference from this hdr. When the reference transitions from 2416 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's 2417 * list making it eligible for eviction. 2418 */ 2419static int 2420remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 2421{ 2422 int cnt; 2423 arc_state_t *state = hdr->b_l1hdr.b_state; 2424 2425 ASSERT(HDR_HAS_L1HDR(hdr)); 2426 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 2427 ASSERT(!GHOST_STATE(state)); 2428 2429 /* 2430 * arc_l2c_only counts as a ghost state so we don't need to explicitly 2431 * check to prevent usage of the arc_l2c_only list. 2432 */ 2433 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 2434 (state != arc_anon)) { 2435 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr); 2436 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 2437 arc_evictable_space_increment(hdr, state); 2438 } 2439 return (cnt); 2440} 2441 2442/* 2443 * Move the supplied buffer to the indicated state. The hash lock 2444 * for the buffer must be held by the caller. 2445 */ 2446static void 2447arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 2448 kmutex_t *hash_lock) 2449{ 2450 arc_state_t *old_state; 2451 int64_t refcnt; 2452 uint32_t bufcnt; 2453 boolean_t update_old, update_new; 2454 arc_buf_contents_t buftype = arc_buf_type(hdr); 2455 2456 /* 2457 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 2458 * in arc_read() when bringing a buffer out of the L2ARC. However, the 2459 * L1 hdr doesn't always exist when we change state to arc_anon before 2460 * destroying a header, in which case reallocating to add the L1 hdr is 2461 * pointless. 2462 */ 2463 if (HDR_HAS_L1HDR(hdr)) { 2464 old_state = hdr->b_l1hdr.b_state; 2465 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 2466 bufcnt = hdr->b_l1hdr.b_bufcnt; 2467 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL); 2468 } else { 2469 old_state = arc_l2c_only; 2470 refcnt = 0; 2471 bufcnt = 0; 2472 update_old = B_FALSE; 2473 } 2474 update_new = update_old; 2475 2476 ASSERT(MUTEX_HELD(hash_lock)); 2477 ASSERT3P(new_state, !=, old_state); 2478 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); 2479 ASSERT(old_state != arc_anon || bufcnt <= 1); 2480 2481 /* 2482 * If this buffer is evictable, transfer it from the 2483 * old state list to the new state list. 2484 */ 2485 if (refcnt == 0) { 2486 if (old_state != arc_anon && old_state != arc_l2c_only) { 2487 ASSERT(HDR_HAS_L1HDR(hdr)); 2488 multilist_remove(old_state->arcs_list[buftype], hdr); 2489 2490 if (GHOST_STATE(old_state)) { 2491 ASSERT0(bufcnt); 2492 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2493 update_old = B_TRUE; 2494 } 2495 arc_evictable_space_decrement(hdr, old_state); 2496 } 2497 if (new_state != arc_anon && new_state != arc_l2c_only) { 2498 2499 /* 2500 * An L1 header always exists here, since if we're 2501 * moving to some L1-cached state (i.e. not l2c_only or 2502 * anonymous), we realloc the header to add an L1hdr 2503 * beforehand. 2504 */ 2505 ASSERT(HDR_HAS_L1HDR(hdr)); 2506 multilist_insert(new_state->arcs_list[buftype], hdr); 2507 2508 if (GHOST_STATE(new_state)) { 2509 ASSERT0(bufcnt); 2510 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2511 update_new = B_TRUE; 2512 } 2513 arc_evictable_space_increment(hdr, new_state); 2514 } 2515 } 2516 2517 ASSERT(!HDR_EMPTY(hdr)); 2518 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 2519 buf_hash_remove(hdr); 2520 2521 /* adjust state sizes (ignore arc_l2c_only) */ 2522 2523 if (update_new && new_state != arc_l2c_only) { 2524 ASSERT(HDR_HAS_L1HDR(hdr)); 2525 if (GHOST_STATE(new_state)) { 2526 ASSERT0(bufcnt); 2527 2528 /* 2529 * When moving a header to a ghost state, we first 2530 * remove all arc buffers. Thus, we'll have a 2531 * bufcnt of zero, and no arc buffer to use for 2532 * the reference. As a result, we use the arc 2533 * header pointer for the reference. 2534 */ 2535 (void) refcount_add_many(&new_state->arcs_size, 2536 HDR_GET_LSIZE(hdr), hdr); 2537 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2538 } else { 2539 uint32_t buffers = 0; 2540 2541 /* 2542 * Each individual buffer holds a unique reference, 2543 * thus we must remove each of these references one 2544 * at a time. 2545 */ 2546 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2547 buf = buf->b_next) { 2548 ASSERT3U(bufcnt, !=, 0); 2549 buffers++; 2550 2551 /* 2552 * When the arc_buf_t is sharing the data 2553 * block with the hdr, the owner of the 2554 * reference belongs to the hdr. Only 2555 * add to the refcount if the arc_buf_t is 2556 * not shared. 2557 */ 2558 if (arc_buf_is_shared(buf)) 2559 continue; 2560 2561 (void) refcount_add_many(&new_state->arcs_size, 2562 arc_buf_size(buf), buf); 2563 } 2564 ASSERT3U(bufcnt, ==, buffers); 2565 2566 if (hdr->b_l1hdr.b_pabd != NULL) { 2567 (void) refcount_add_many(&new_state->arcs_size, 2568 arc_hdr_size(hdr), hdr); 2569 } else { 2570 ASSERT(GHOST_STATE(old_state)); 2571 } 2572 } 2573 } 2574 2575 if (update_old && old_state != arc_l2c_only) { 2576 ASSERT(HDR_HAS_L1HDR(hdr)); 2577 if (GHOST_STATE(old_state)) { 2578 ASSERT0(bufcnt); 2579 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2580 2581 /* 2582 * When moving a header off of a ghost state, 2583 * the header will not contain any arc buffers. 2584 * We use the arc header pointer for the reference 2585 * which is exactly what we did when we put the 2586 * header on the ghost state. 2587 */ 2588 2589 (void) refcount_remove_many(&old_state->arcs_size, 2590 HDR_GET_LSIZE(hdr), hdr); 2591 } else { 2592 uint32_t buffers = 0; 2593 2594 /* 2595 * Each individual buffer holds a unique reference, 2596 * thus we must remove each of these references one 2597 * at a time. 2598 */ 2599 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2600 buf = buf->b_next) { 2601 ASSERT3U(bufcnt, !=, 0); 2602 buffers++; 2603 2604 /* 2605 * When the arc_buf_t is sharing the data 2606 * block with the hdr, the owner of the 2607 * reference belongs to the hdr. Only 2608 * add to the refcount if the arc_buf_t is 2609 * not shared. 2610 */ 2611 if (arc_buf_is_shared(buf)) 2612 continue; 2613 2614 (void) refcount_remove_many( 2615 &old_state->arcs_size, arc_buf_size(buf), 2616 buf); 2617 } 2618 ASSERT3U(bufcnt, ==, buffers); 2619 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2620 (void) refcount_remove_many( 2621 &old_state->arcs_size, arc_hdr_size(hdr), hdr); 2622 } 2623 } 2624 2625 if (HDR_HAS_L1HDR(hdr)) 2626 hdr->b_l1hdr.b_state = new_state; 2627 2628 /* 2629 * L2 headers should never be on the L2 state list since they don't 2630 * have L1 headers allocated. 2631 */ 2632 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 2633 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 2634} 2635 2636void 2637arc_space_consume(uint64_t space, arc_space_type_t type) 2638{ 2639 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2640 2641 switch (type) { 2642 case ARC_SPACE_DATA: 2643 ARCSTAT_INCR(arcstat_data_size, space); 2644 break; 2645 case ARC_SPACE_META: 2646 ARCSTAT_INCR(arcstat_metadata_size, space); 2647 break; 2648 case ARC_SPACE_OTHER: 2649 ARCSTAT_INCR(arcstat_other_size, space); 2650 break; 2651 case ARC_SPACE_HDRS: 2652 ARCSTAT_INCR(arcstat_hdr_size, space); 2653 break; 2654 case ARC_SPACE_L2HDRS: 2655 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 2656 break; 2657 } 2658 2659 if (type != ARC_SPACE_DATA) 2660 ARCSTAT_INCR(arcstat_meta_used, space); 2661 2662 atomic_add_64(&arc_size, space); 2663} 2664 2665void 2666arc_space_return(uint64_t space, arc_space_type_t type) 2667{ 2668 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2669 2670 switch (type) { 2671 case ARC_SPACE_DATA: 2672 ARCSTAT_INCR(arcstat_data_size, -space); 2673 break; 2674 case ARC_SPACE_META: 2675 ARCSTAT_INCR(arcstat_metadata_size, -space); 2676 break; 2677 case ARC_SPACE_OTHER: 2678 ARCSTAT_INCR(arcstat_other_size, -space); 2679 break; 2680 case ARC_SPACE_HDRS: 2681 ARCSTAT_INCR(arcstat_hdr_size, -space); 2682 break; 2683 case ARC_SPACE_L2HDRS: 2684 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 2685 break; 2686 } 2687 2688 if (type != ARC_SPACE_DATA) { 2689 ASSERT(arc_meta_used >= space); 2690 if (arc_meta_max < arc_meta_used) 2691 arc_meta_max = arc_meta_used; 2692 ARCSTAT_INCR(arcstat_meta_used, -space); 2693 } 2694 2695 ASSERT(arc_size >= space); 2696 atomic_add_64(&arc_size, -space); 2697} 2698 2699/* 2700 * Given a hdr and a buf, returns whether that buf can share its b_data buffer 2701 * with the hdr's b_pabd. 2702 */ 2703static boolean_t 2704arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2705{ 2706 /* 2707 * The criteria for sharing a hdr's data are: 2708 * 1. the hdr's compression matches the buf's compression 2709 * 2. the hdr doesn't need to be byteswapped 2710 * 3. the hdr isn't already being shared 2711 * 4. the buf is either compressed or it is the last buf in the hdr list 2712 * 2713 * Criterion #4 maintains the invariant that shared uncompressed 2714 * bufs must be the final buf in the hdr's b_buf list. Reading this, you 2715 * might ask, "if a compressed buf is allocated first, won't that be the 2716 * last thing in the list?", but in that case it's impossible to create 2717 * a shared uncompressed buf anyway (because the hdr must be compressed 2718 * to have the compressed buf). You might also think that #3 is 2719 * sufficient to make this guarantee, however it's possible 2720 * (specifically in the rare L2ARC write race mentioned in 2721 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that 2722 * is sharable, but wasn't at the time of its allocation. Rather than 2723 * allow a new shared uncompressed buf to be created and then shuffle 2724 * the list around to make it the last element, this simply disallows 2725 * sharing if the new buf isn't the first to be added. 2726 */ 2727 ASSERT3P(buf->b_hdr, ==, hdr); 2728 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF; 2729 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0; 2730 return (buf_compressed == hdr_compressed && 2731 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && 2732 !HDR_SHARED_DATA(hdr) && 2733 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf))); 2734} 2735 2736/* 2737 * Allocate a buf for this hdr. If you care about the data that's in the hdr, 2738 * or if you want a compressed buffer, pass those flags in. Returns 0 if the 2739 * copy was made successfully, or an error code otherwise. 2740 */ 2741static int 2742arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed, 2743 boolean_t fill, arc_buf_t **ret) 2744{ 2745 arc_buf_t *buf; 2746 2747 ASSERT(HDR_HAS_L1HDR(hdr)); 2748 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2749 VERIFY(hdr->b_type == ARC_BUFC_DATA || 2750 hdr->b_type == ARC_BUFC_METADATA); 2751 ASSERT3P(ret, !=, NULL); 2752 ASSERT3P(*ret, ==, NULL); 2753 2754 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2755 buf->b_hdr = hdr; 2756 buf->b_data = NULL; 2757 buf->b_next = hdr->b_l1hdr.b_buf; 2758 buf->b_flags = 0; 2759 2760 add_reference(hdr, tag); 2761 2762 /* 2763 * We're about to change the hdr's b_flags. We must either 2764 * hold the hash_lock or be undiscoverable. 2765 */ 2766 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2767 2768 /* 2769 * Only honor requests for compressed bufs if the hdr is actually 2770 * compressed. 2771 */ 2772 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 2773 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; 2774 2775 /* 2776 * If the hdr's data can be shared then we share the data buffer and 2777 * set the appropriate bit in the hdr's b_flags to indicate the hdr is 2778 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new 2779 * buffer to store the buf's data. 2780 * 2781 * There are two additional restrictions here because we're sharing 2782 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be 2783 * actively involved in an L2ARC write, because if this buf is used by 2784 * an arc_write() then the hdr's data buffer will be released when the 2785 * write completes, even though the L2ARC write might still be using it. 2786 * Second, the hdr's ABD must be linear so that the buf's user doesn't 2787 * need to be ABD-aware. 2788 */ 2789 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) && 2790 abd_is_linear(hdr->b_l1hdr.b_pabd); 2791 2792 /* Set up b_data and sharing */ 2793 if (can_share) { 2794 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd); 2795 buf->b_flags |= ARC_BUF_FLAG_SHARED; 2796 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2797 } else { 2798 buf->b_data = 2799 arc_get_data_buf(hdr, arc_buf_size(buf), buf); 2800 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2801 } 2802 VERIFY3P(buf->b_data, !=, NULL); 2803 2804 hdr->b_l1hdr.b_buf = buf; 2805 hdr->b_l1hdr.b_bufcnt += 1; 2806 2807 /* 2808 * If the user wants the data from the hdr, we need to either copy or 2809 * decompress the data. 2810 */ 2811 if (fill) { 2812 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0)); 2813 } 2814 2815 return (0); 2816} 2817 2818static char *arc_onloan_tag = "onloan"; 2819 2820static inline void 2821arc_loaned_bytes_update(int64_t delta) 2822{ 2823 atomic_add_64(&arc_loaned_bytes, delta); 2824 2825 /* assert that it did not wrap around */ 2826 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 2827} 2828 2829/* 2830 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2831 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2832 * buffers must be returned to the arc before they can be used by the DMU or 2833 * freed. 2834 */ 2835arc_buf_t * 2836arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size) 2837{ 2838 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag, 2839 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size); 2840 2841 arc_loaned_bytes_update(size); 2842 2843 return (buf); 2844} 2845 2846arc_buf_t * 2847arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize, 2848 enum zio_compress compression_type) 2849{ 2850 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag, 2851 psize, lsize, compression_type); 2852 2853 arc_loaned_bytes_update(psize); 2854 2855 return (buf); 2856} 2857 2858 2859/* 2860 * Return a loaned arc buffer to the arc. 2861 */ 2862void 2863arc_return_buf(arc_buf_t *buf, void *tag) 2864{ 2865 arc_buf_hdr_t *hdr = buf->b_hdr; 2866 2867 ASSERT3P(buf->b_data, !=, NULL); 2868 ASSERT(HDR_HAS_L1HDR(hdr)); 2869 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2870 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2871 2872 arc_loaned_bytes_update(-arc_buf_size(buf)); 2873} 2874 2875/* Detach an arc_buf from a dbuf (tag) */ 2876void 2877arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2878{ 2879 arc_buf_hdr_t *hdr = buf->b_hdr; 2880 2881 ASSERT3P(buf->b_data, !=, NULL); 2882 ASSERT(HDR_HAS_L1HDR(hdr)); 2883 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2884 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2885 2886 arc_loaned_bytes_update(arc_buf_size(buf)); 2887} 2888 2889static void 2890l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type) 2891{ 2892 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); 2893 2894 df->l2df_abd = abd; 2895 df->l2df_size = size; 2896 df->l2df_type = type; 2897 mutex_enter(&l2arc_free_on_write_mtx); 2898 list_insert_head(l2arc_free_on_write, df); 2899 mutex_exit(&l2arc_free_on_write_mtx); 2900} 2901 2902static void 2903arc_hdr_free_on_write(arc_buf_hdr_t *hdr) 2904{ 2905 arc_state_t *state = hdr->b_l1hdr.b_state; 2906 arc_buf_contents_t type = arc_buf_type(hdr); 2907 uint64_t size = arc_hdr_size(hdr); 2908 2909 /* protected by hash lock, if in the hash table */ 2910 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 2911 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2912 ASSERT(state != arc_anon && state != arc_l2c_only); 2913 2914 (void) refcount_remove_many(&state->arcs_esize[type], 2915 size, hdr); 2916 } 2917 (void) refcount_remove_many(&state->arcs_size, size, hdr); 2918 if (type == ARC_BUFC_METADATA) { 2919 arc_space_return(size, ARC_SPACE_META); 2920 } else { 2921 ASSERT(type == ARC_BUFC_DATA); 2922 arc_space_return(size, ARC_SPACE_DATA); 2923 } 2924 2925 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type); 2926} 2927 2928/* 2929 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the 2930 * data buffer, we transfer the refcount ownership to the hdr and update 2931 * the appropriate kstats. 2932 */ 2933static void 2934arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2935{ 2936 arc_state_t *state = hdr->b_l1hdr.b_state; 2937 2938 ASSERT(arc_can_share(hdr, buf)); 2939 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2940 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2941 2942 /* 2943 * Start sharing the data buffer. We transfer the 2944 * refcount ownership to the hdr since it always owns 2945 * the refcount whenever an arc_buf_t is shared. 2946 */ 2947 refcount_transfer_ownership(&state->arcs_size, buf, hdr); 2948 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf)); 2949 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd, 2950 HDR_ISTYPE_METADATA(hdr)); 2951 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2952 buf->b_flags |= ARC_BUF_FLAG_SHARED; 2953 2954 /* 2955 * Since we've transferred ownership to the hdr we need 2956 * to increment its compressed and uncompressed kstats and 2957 * decrement the overhead size. 2958 */ 2959 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2960 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2961 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf)); 2962} 2963 2964static void 2965arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2966{ 2967 arc_state_t *state = hdr->b_l1hdr.b_state; 2968 2969 ASSERT(arc_buf_is_shared(buf)); 2970 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2971 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2972 2973 /* 2974 * We are no longer sharing this buffer so we need 2975 * to transfer its ownership to the rightful owner. 2976 */ 2977 refcount_transfer_ownership(&state->arcs_size, hdr, buf); 2978 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2979 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd); 2980 abd_put(hdr->b_l1hdr.b_pabd); 2981 hdr->b_l1hdr.b_pabd = NULL; 2982 buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 2983 2984 /* 2985 * Since the buffer is no longer shared between 2986 * the arc buf and the hdr, count it as overhead. 2987 */ 2988 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2989 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2990 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2991} 2992 2993/* 2994 * Remove an arc_buf_t from the hdr's buf list and return the last 2995 * arc_buf_t on the list. If no buffers remain on the list then return 2996 * NULL. 2997 */ 2998static arc_buf_t * 2999arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf) 3000{ 3001 ASSERT(HDR_HAS_L1HDR(hdr)); 3002 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 3003 3004 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf; 3005 arc_buf_t *lastbuf = NULL; 3006 3007 /* 3008 * Remove the buf from the hdr list and locate the last 3009 * remaining buffer on the list. 3010 */ 3011 while (*bufp != NULL) { 3012 if (*bufp == buf) 3013 *bufp = buf->b_next; 3014 3015 /* 3016 * If we've removed a buffer in the middle of 3017 * the list then update the lastbuf and update 3018 * bufp. 3019 */ 3020 if (*bufp != NULL) { 3021 lastbuf = *bufp; 3022 bufp = &(*bufp)->b_next; 3023 } 3024 } 3025 buf->b_next = NULL; 3026 ASSERT3P(lastbuf, !=, buf); 3027 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL); 3028 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL); 3029 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf)); 3030 3031 return (lastbuf); 3032} 3033 3034/* 3035 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's 3036 * list and free it. 3037 */ 3038static void 3039arc_buf_destroy_impl(arc_buf_t *buf) 3040{ 3041 arc_buf_hdr_t *hdr = buf->b_hdr; 3042 3043 /* 3044 * Free up the data associated with the buf but only if we're not 3045 * sharing this with the hdr. If we are sharing it with the hdr, the 3046 * hdr is responsible for doing the free. 3047 */ 3048 if (buf->b_data != NULL) { 3049 /* 3050 * We're about to change the hdr's b_flags. We must either 3051 * hold the hash_lock or be undiscoverable. 3052 */ 3053 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 3054 3055 arc_cksum_verify(buf); 3056#ifdef illumos 3057 arc_buf_unwatch(buf); 3058#endif 3059 3060 if (arc_buf_is_shared(buf)) { 3061 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 3062 } else { 3063 uint64_t size = arc_buf_size(buf); 3064 arc_free_data_buf(hdr, buf->b_data, size, buf); 3065 ARCSTAT_INCR(arcstat_overhead_size, -size); 3066 } 3067 buf->b_data = NULL; 3068 3069 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 3070 hdr->b_l1hdr.b_bufcnt -= 1; 3071 } 3072 3073 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 3074 3075 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) { 3076 /* 3077 * If the current arc_buf_t is sharing its data buffer with the 3078 * hdr, then reassign the hdr's b_pabd to share it with the new 3079 * buffer at the end of the list. The shared buffer is always 3080 * the last one on the hdr's buffer list. 3081 * 3082 * There is an equivalent case for compressed bufs, but since 3083 * they aren't guaranteed to be the last buf in the list and 3084 * that is an exceedingly rare case, we just allow that space be 3085 * wasted temporarily. 3086 */ 3087 if (lastbuf != NULL) { 3088 /* Only one buf can be shared at once */ 3089 VERIFY(!arc_buf_is_shared(lastbuf)); 3090 /* hdr is uncompressed so can't have compressed buf */ 3091 VERIFY(!ARC_BUF_COMPRESSED(lastbuf)); 3092 3093 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 3094 arc_hdr_free_pabd(hdr); 3095 3096 /* 3097 * We must setup a new shared block between the 3098 * last buffer and the hdr. The data would have 3099 * been allocated by the arc buf so we need to transfer 3100 * ownership to the hdr since it's now being shared. 3101 */ 3102 arc_share_buf(hdr, lastbuf); 3103 } 3104 } else if (HDR_SHARED_DATA(hdr)) { 3105 /* 3106 * Uncompressed shared buffers are always at the end 3107 * of the list. Compressed buffers don't have the 3108 * same requirements. This makes it hard to 3109 * simply assert that the lastbuf is shared so 3110 * we rely on the hdr's compression flags to determine 3111 * if we have a compressed, shared buffer. 3112 */ 3113 ASSERT3P(lastbuf, !=, NULL); 3114 ASSERT(arc_buf_is_shared(lastbuf) || 3115 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 3116 } 3117 3118 /* 3119 * Free the checksum if we're removing the last uncompressed buf from 3120 * this hdr. 3121 */ 3122 if (!arc_hdr_has_uncompressed_buf(hdr)) { 3123 arc_cksum_free(hdr); 3124 } 3125 3126 /* clean up the buf */ 3127 buf->b_hdr = NULL; 3128 kmem_cache_free(buf_cache, buf); 3129} 3130 3131static void 3132arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr) 3133{ 3134 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 3135 ASSERT(HDR_HAS_L1HDR(hdr)); 3136 ASSERT(!HDR_SHARED_DATA(hdr)); 3137 3138 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3139 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr); 3140 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 3141 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 3142 3143 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 3144 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 3145} 3146 3147static void 3148arc_hdr_free_pabd(arc_buf_hdr_t *hdr) 3149{ 3150 ASSERT(HDR_HAS_L1HDR(hdr)); 3151 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 3152 3153 /* 3154 * If the hdr is currently being written to the l2arc then 3155 * we defer freeing the data by adding it to the l2arc_free_on_write 3156 * list. The l2arc will free the data once it's finished 3157 * writing it to the l2arc device. 3158 */ 3159 if (HDR_L2_WRITING(hdr)) { 3160 arc_hdr_free_on_write(hdr); 3161 ARCSTAT_BUMP(arcstat_l2_free_on_write); 3162 } else { 3163 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, 3164 arc_hdr_size(hdr), hdr); 3165 } 3166 hdr->b_l1hdr.b_pabd = NULL; 3167 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 3168 3169 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 3170 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 3171} 3172 3173static arc_buf_hdr_t * 3174arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, 3175 enum zio_compress compression_type, arc_buf_contents_t type) 3176{ 3177 arc_buf_hdr_t *hdr; 3178 3179 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); 3180 3181 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 3182 ASSERT(HDR_EMPTY(hdr)); 3183 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3184 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL); 3185 HDR_SET_PSIZE(hdr, psize); 3186 HDR_SET_LSIZE(hdr, lsize); 3187 hdr->b_spa = spa; 3188 hdr->b_type = type; 3189 hdr->b_flags = 0; 3190 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); 3191 arc_hdr_set_compress(hdr, compression_type); 3192 3193 hdr->b_l1hdr.b_state = arc_anon; 3194 hdr->b_l1hdr.b_arc_access = 0; 3195 hdr->b_l1hdr.b_bufcnt = 0; 3196 hdr->b_l1hdr.b_buf = NULL; 3197 3198 /* 3199 * Allocate the hdr's buffer. This will contain either 3200 * the compressed or uncompressed data depending on the block 3201 * it references and compressed arc enablement. 3202 */ 3203 arc_hdr_alloc_pabd(hdr); 3204 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3205 3206 return (hdr); 3207} 3208 3209/* 3210 * Transition between the two allocation states for the arc_buf_hdr struct. 3211 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 3212 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 3213 * version is used when a cache buffer is only in the L2ARC in order to reduce 3214 * memory usage. 3215 */ 3216static arc_buf_hdr_t * 3217arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 3218{ 3219 ASSERT(HDR_HAS_L2HDR(hdr)); 3220 3221 arc_buf_hdr_t *nhdr; 3222 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3223 3224 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 3225 (old == hdr_l2only_cache && new == hdr_full_cache)); 3226 3227 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 3228 3229 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 3230 buf_hash_remove(hdr); 3231 3232 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 3233 3234 if (new == hdr_full_cache) { 3235 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); 3236 /* 3237 * arc_access and arc_change_state need to be aware that a 3238 * header has just come out of L2ARC, so we set its state to 3239 * l2c_only even though it's about to change. 3240 */ 3241 nhdr->b_l1hdr.b_state = arc_l2c_only; 3242 3243 /* Verify previous threads set to NULL before freeing */ 3244 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL); 3245 } else { 3246 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3247 ASSERT0(hdr->b_l1hdr.b_bufcnt); 3248 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3249 3250 /* 3251 * If we've reached here, We must have been called from 3252 * arc_evict_hdr(), as such we should have already been 3253 * removed from any ghost list we were previously on 3254 * (which protects us from racing with arc_evict_state), 3255 * thus no locking is needed during this check. 3256 */ 3257 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3258 3259 /* 3260 * A buffer must not be moved into the arc_l2c_only 3261 * state if it's not finished being written out to the 3262 * l2arc device. Otherwise, the b_l1hdr.b_pabd field 3263 * might try to be accessed, even though it was removed. 3264 */ 3265 VERIFY(!HDR_L2_WRITING(hdr)); 3266 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3267 3268#ifdef ZFS_DEBUG 3269 if (hdr->b_l1hdr.b_thawed != NULL) { 3270 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3271 hdr->b_l1hdr.b_thawed = NULL; 3272 } 3273#endif 3274 3275 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); 3276 } 3277 /* 3278 * The header has been reallocated so we need to re-insert it into any 3279 * lists it was on. 3280 */ 3281 (void) buf_hash_insert(nhdr, NULL); 3282 3283 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 3284 3285 mutex_enter(&dev->l2ad_mtx); 3286 3287 /* 3288 * We must place the realloc'ed header back into the list at 3289 * the same spot. Otherwise, if it's placed earlier in the list, 3290 * l2arc_write_buffers() could find it during the function's 3291 * write phase, and try to write it out to the l2arc. 3292 */ 3293 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 3294 list_remove(&dev->l2ad_buflist, hdr); 3295 3296 mutex_exit(&dev->l2ad_mtx); 3297 3298 /* 3299 * Since we're using the pointer address as the tag when 3300 * incrementing and decrementing the l2ad_alloc refcount, we 3301 * must remove the old pointer (that we're about to destroy) and 3302 * add the new pointer to the refcount. Otherwise we'd remove 3303 * the wrong pointer address when calling arc_hdr_destroy() later. 3304 */ 3305 3306 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); 3307 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr); 3308 3309 buf_discard_identity(hdr); 3310 kmem_cache_free(old, hdr); 3311 3312 return (nhdr); 3313} 3314 3315/* 3316 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. 3317 * The buf is returned thawed since we expect the consumer to modify it. 3318 */ 3319arc_buf_t * 3320arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size) 3321{ 3322 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, 3323 ZIO_COMPRESS_OFF, type); 3324 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 3325 3326 arc_buf_t *buf = NULL; 3327 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf)); 3328 arc_buf_thaw(buf); 3329 3330 return (buf); 3331} 3332 3333/* 3334 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this 3335 * for bufs containing metadata. 3336 */ 3337arc_buf_t * 3338arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize, 3339 enum zio_compress compression_type) 3340{ 3341 ASSERT3U(lsize, >, 0); 3342 ASSERT3U(lsize, >=, psize); 3343 ASSERT(compression_type > ZIO_COMPRESS_OFF); 3344 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS); 3345 3346 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 3347 compression_type, ARC_BUFC_DATA); 3348 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 3349 3350 arc_buf_t *buf = NULL; 3351 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf)); 3352 arc_buf_thaw(buf); 3353 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3354 3355 if (!arc_buf_is_shared(buf)) { 3356 /* 3357 * To ensure that the hdr has the correct data in it if we call 3358 * arc_decompress() on this buf before it's been written to 3359 * disk, it's easiest if we just set up sharing between the 3360 * buf and the hdr. 3361 */ 3362 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd)); 3363 arc_hdr_free_pabd(hdr); 3364 arc_share_buf(hdr, buf); 3365 } 3366 3367 return (buf); 3368} 3369 3370static void 3371arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 3372{ 3373 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 3374 l2arc_dev_t *dev = l2hdr->b_dev; 3375 uint64_t psize = arc_hdr_size(hdr); 3376 3377 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 3378 ASSERT(HDR_HAS_L2HDR(hdr)); 3379 3380 list_remove(&dev->l2ad_buflist, hdr); 3381 3382 ARCSTAT_INCR(arcstat_l2_psize, -psize); 3383 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); 3384 3385 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0); 3386 3387 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr); 3388 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 3389} 3390 3391static void 3392arc_hdr_destroy(arc_buf_hdr_t *hdr) 3393{ 3394 if (HDR_HAS_L1HDR(hdr)) { 3395 ASSERT(hdr->b_l1hdr.b_buf == NULL || 3396 hdr->b_l1hdr.b_bufcnt > 0); 3397 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3398 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 3399 } 3400 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3401 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 3402 3403 if (!HDR_EMPTY(hdr)) 3404 buf_discard_identity(hdr); 3405 3406 if (HDR_HAS_L2HDR(hdr)) { 3407 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3408 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 3409 3410 if (!buflist_held) 3411 mutex_enter(&dev->l2ad_mtx); 3412 3413 /* 3414 * Even though we checked this conditional above, we 3415 * need to check this again now that we have the 3416 * l2ad_mtx. This is because we could be racing with 3417 * another thread calling l2arc_evict() which might have 3418 * destroyed this header's L2 portion as we were waiting 3419 * to acquire the l2ad_mtx. If that happens, we don't 3420 * want to re-destroy the header's L2 portion. 3421 */ 3422 if (HDR_HAS_L2HDR(hdr)) { 3423 l2arc_trim(hdr); 3424 arc_hdr_l2hdr_destroy(hdr); 3425 } 3426 3427 if (!buflist_held) 3428 mutex_exit(&dev->l2ad_mtx); 3429 } 3430 3431 if (HDR_HAS_L1HDR(hdr)) { 3432 arc_cksum_free(hdr); 3433 3434 while (hdr->b_l1hdr.b_buf != NULL) 3435 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf); 3436 3437#ifdef ZFS_DEBUG 3438 if (hdr->b_l1hdr.b_thawed != NULL) { 3439 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3440 hdr->b_l1hdr.b_thawed = NULL; 3441 } 3442#endif 3443 3444 if (hdr->b_l1hdr.b_pabd != NULL) { 3445 arc_hdr_free_pabd(hdr); 3446 } 3447 } 3448 3449 ASSERT3P(hdr->b_hash_next, ==, NULL); 3450 if (HDR_HAS_L1HDR(hdr)) { 3451 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3452 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 3453 kmem_cache_free(hdr_full_cache, hdr); 3454 } else { 3455 kmem_cache_free(hdr_l2only_cache, hdr); 3456 } 3457} 3458 3459void 3460arc_buf_destroy(arc_buf_t *buf, void* tag) 3461{ 3462 arc_buf_hdr_t *hdr = buf->b_hdr; 3463 kmutex_t *hash_lock = HDR_LOCK(hdr); 3464 3465 if (hdr->b_l1hdr.b_state == arc_anon) { 3466 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 3467 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3468 VERIFY0(remove_reference(hdr, NULL, tag)); 3469 arc_hdr_destroy(hdr); 3470 return; 3471 } 3472 3473 mutex_enter(hash_lock); 3474 ASSERT3P(hdr, ==, buf->b_hdr); 3475 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 3476 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3477 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); 3478 ASSERT3P(buf->b_data, !=, NULL); 3479 3480 (void) remove_reference(hdr, hash_lock, tag); 3481 arc_buf_destroy_impl(buf); 3482 mutex_exit(hash_lock); 3483} 3484 3485/* 3486 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 3487 * state of the header is dependent on it's state prior to entering this 3488 * function. The following transitions are possible: 3489 * 3490 * - arc_mru -> arc_mru_ghost 3491 * - arc_mfu -> arc_mfu_ghost 3492 * - arc_mru_ghost -> arc_l2c_only 3493 * - arc_mru_ghost -> deleted 3494 * - arc_mfu_ghost -> arc_l2c_only 3495 * - arc_mfu_ghost -> deleted 3496 */ 3497static int64_t 3498arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3499{ 3500 arc_state_t *evicted_state, *state; 3501 int64_t bytes_evicted = 0; 3502 3503 ASSERT(MUTEX_HELD(hash_lock)); 3504 ASSERT(HDR_HAS_L1HDR(hdr)); 3505 3506 state = hdr->b_l1hdr.b_state; 3507 if (GHOST_STATE(state)) { 3508 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3509 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3510 3511 /* 3512 * l2arc_write_buffers() relies on a header's L1 portion 3513 * (i.e. its b_pabd field) during it's write phase. 3514 * Thus, we cannot push a header onto the arc_l2c_only 3515 * state (removing it's L1 piece) until the header is 3516 * done being written to the l2arc. 3517 */ 3518 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 3519 ARCSTAT_BUMP(arcstat_evict_l2_skip); 3520 return (bytes_evicted); 3521 } 3522 3523 ARCSTAT_BUMP(arcstat_deleted); 3524 bytes_evicted += HDR_GET_LSIZE(hdr); 3525 3526 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 3527 3528 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3529 if (HDR_HAS_L2HDR(hdr)) { 3530 /* 3531 * This buffer is cached on the 2nd Level ARC; 3532 * don't destroy the header. 3533 */ 3534 arc_change_state(arc_l2c_only, hdr, hash_lock); 3535 /* 3536 * dropping from L1+L2 cached to L2-only, 3537 * realloc to remove the L1 header. 3538 */ 3539 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 3540 hdr_l2only_cache); 3541 } else { 3542 arc_change_state(arc_anon, hdr, hash_lock); 3543 arc_hdr_destroy(hdr); 3544 } 3545 return (bytes_evicted); 3546 } 3547 3548 ASSERT(state == arc_mru || state == arc_mfu); 3549 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3550 3551 /* prefetch buffers have a minimum lifespan */ 3552 if (HDR_IO_IN_PROGRESS(hdr) || 3553 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 3554 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 3555 arc_min_prefetch_lifespan)) { 3556 ARCSTAT_BUMP(arcstat_evict_skip); 3557 return (bytes_evicted); 3558 } 3559 3560 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3561 while (hdr->b_l1hdr.b_buf) { 3562 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3563 if (!mutex_tryenter(&buf->b_evict_lock)) { 3564 ARCSTAT_BUMP(arcstat_mutex_miss); 3565 break; 3566 } 3567 if (buf->b_data != NULL) 3568 bytes_evicted += HDR_GET_LSIZE(hdr); 3569 mutex_exit(&buf->b_evict_lock); 3570 arc_buf_destroy_impl(buf); 3571 } 3572 3573 if (HDR_HAS_L2HDR(hdr)) { 3574 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); 3575 } else { 3576 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 3577 ARCSTAT_INCR(arcstat_evict_l2_eligible, 3578 HDR_GET_LSIZE(hdr)); 3579 } else { 3580 ARCSTAT_INCR(arcstat_evict_l2_ineligible, 3581 HDR_GET_LSIZE(hdr)); 3582 } 3583 } 3584 3585 if (hdr->b_l1hdr.b_bufcnt == 0) { 3586 arc_cksum_free(hdr); 3587 3588 bytes_evicted += arc_hdr_size(hdr); 3589 3590 /* 3591 * If this hdr is being evicted and has a compressed 3592 * buffer then we discard it here before we change states. 3593 * This ensures that the accounting is updated correctly 3594 * in arc_free_data_impl(). 3595 */ 3596 arc_hdr_free_pabd(hdr); 3597 3598 arc_change_state(evicted_state, hdr, hash_lock); 3599 ASSERT(HDR_IN_HASH_TABLE(hdr)); 3600 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 3601 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 3602 } 3603 3604 return (bytes_evicted); 3605} 3606 3607static uint64_t 3608arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 3609 uint64_t spa, int64_t bytes) 3610{ 3611 multilist_sublist_t *mls; 3612 uint64_t bytes_evicted = 0; 3613 arc_buf_hdr_t *hdr; 3614 kmutex_t *hash_lock; 3615 int evict_count = 0; 3616 3617 ASSERT3P(marker, !=, NULL); 3618 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3619 3620 mls = multilist_sublist_lock(ml, idx); 3621 3622 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 3623 hdr = multilist_sublist_prev(mls, marker)) { 3624 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 3625 (evict_count >= zfs_arc_evict_batch_limit)) 3626 break; 3627 3628 /* 3629 * To keep our iteration location, move the marker 3630 * forward. Since we're not holding hdr's hash lock, we 3631 * must be very careful and not remove 'hdr' from the 3632 * sublist. Otherwise, other consumers might mistake the 3633 * 'hdr' as not being on a sublist when they call the 3634 * multilist_link_active() function (they all rely on 3635 * the hash lock protecting concurrent insertions and 3636 * removals). multilist_sublist_move_forward() was 3637 * specifically implemented to ensure this is the case 3638 * (only 'marker' will be removed and re-inserted). 3639 */ 3640 multilist_sublist_move_forward(mls, marker); 3641 3642 /* 3643 * The only case where the b_spa field should ever be 3644 * zero, is the marker headers inserted by 3645 * arc_evict_state(). It's possible for multiple threads 3646 * to be calling arc_evict_state() concurrently (e.g. 3647 * dsl_pool_close() and zio_inject_fault()), so we must 3648 * skip any markers we see from these other threads. 3649 */ 3650 if (hdr->b_spa == 0) 3651 continue; 3652 3653 /* we're only interested in evicting buffers of a certain spa */ 3654 if (spa != 0 && hdr->b_spa != spa) { 3655 ARCSTAT_BUMP(arcstat_evict_skip); 3656 continue; 3657 } 3658 3659 hash_lock = HDR_LOCK(hdr); 3660 3661 /* 3662 * We aren't calling this function from any code path 3663 * that would already be holding a hash lock, so we're 3664 * asserting on this assumption to be defensive in case 3665 * this ever changes. Without this check, it would be 3666 * possible to incorrectly increment arcstat_mutex_miss 3667 * below (e.g. if the code changed such that we called 3668 * this function with a hash lock held). 3669 */ 3670 ASSERT(!MUTEX_HELD(hash_lock)); 3671 3672 if (mutex_tryenter(hash_lock)) { 3673 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 3674 mutex_exit(hash_lock); 3675 3676 bytes_evicted += evicted; 3677 3678 /* 3679 * If evicted is zero, arc_evict_hdr() must have 3680 * decided to skip this header, don't increment 3681 * evict_count in this case. 3682 */ 3683 if (evicted != 0) 3684 evict_count++; 3685 3686 /* 3687 * If arc_size isn't overflowing, signal any 3688 * threads that might happen to be waiting. 3689 * 3690 * For each header evicted, we wake up a single 3691 * thread. If we used cv_broadcast, we could 3692 * wake up "too many" threads causing arc_size 3693 * to significantly overflow arc_c; since 3694 * arc_get_data_impl() doesn't check for overflow 3695 * when it's woken up (it doesn't because it's 3696 * possible for the ARC to be overflowing while 3697 * full of un-evictable buffers, and the 3698 * function should proceed in this case). 3699 * 3700 * If threads are left sleeping, due to not 3701 * using cv_broadcast, they will be woken up 3702 * just before arc_reclaim_thread() sleeps. 3703 */ 3704 mutex_enter(&arc_reclaim_lock); 3705 if (!arc_is_overflowing()) 3706 cv_signal(&arc_reclaim_waiters_cv); 3707 mutex_exit(&arc_reclaim_lock); 3708 } else { 3709 ARCSTAT_BUMP(arcstat_mutex_miss); 3710 } 3711 } 3712 3713 multilist_sublist_unlock(mls); 3714 3715 return (bytes_evicted); 3716} 3717 3718/* 3719 * Evict buffers from the given arc state, until we've removed the 3720 * specified number of bytes. Move the removed buffers to the 3721 * appropriate evict state. 3722 * 3723 * This function makes a "best effort". It skips over any buffers 3724 * it can't get a hash_lock on, and so, may not catch all candidates. 3725 * It may also return without evicting as much space as requested. 3726 * 3727 * If bytes is specified using the special value ARC_EVICT_ALL, this 3728 * will evict all available (i.e. unlocked and evictable) buffers from 3729 * the given arc state; which is used by arc_flush(). 3730 */ 3731static uint64_t 3732arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 3733 arc_buf_contents_t type) 3734{ 3735 uint64_t total_evicted = 0; 3736 multilist_t *ml = state->arcs_list[type]; 3737 int num_sublists; 3738 arc_buf_hdr_t **markers; 3739 3740 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3741 3742 num_sublists = multilist_get_num_sublists(ml); 3743 3744 /* 3745 * If we've tried to evict from each sublist, made some 3746 * progress, but still have not hit the target number of bytes 3747 * to evict, we want to keep trying. The markers allow us to 3748 * pick up where we left off for each individual sublist, rather 3749 * than starting from the tail each time. 3750 */ 3751 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 3752 for (int i = 0; i < num_sublists; i++) { 3753 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 3754 3755 /* 3756 * A b_spa of 0 is used to indicate that this header is 3757 * a marker. This fact is used in arc_adjust_type() and 3758 * arc_evict_state_impl(). 3759 */ 3760 markers[i]->b_spa = 0; 3761 3762 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3763 multilist_sublist_insert_tail(mls, markers[i]); 3764 multilist_sublist_unlock(mls); 3765 } 3766 3767 /* 3768 * While we haven't hit our target number of bytes to evict, or 3769 * we're evicting all available buffers. 3770 */ 3771 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 3772 /* 3773 * Start eviction using a randomly selected sublist, 3774 * this is to try and evenly balance eviction across all 3775 * sublists. Always starting at the same sublist 3776 * (e.g. index 0) would cause evictions to favor certain 3777 * sublists over others. 3778 */ 3779 int sublist_idx = multilist_get_random_index(ml); 3780 uint64_t scan_evicted = 0; 3781 3782 for (int i = 0; i < num_sublists; i++) { 3783 uint64_t bytes_remaining; 3784 uint64_t bytes_evicted; 3785 3786 if (bytes == ARC_EVICT_ALL) 3787 bytes_remaining = ARC_EVICT_ALL; 3788 else if (total_evicted < bytes) 3789 bytes_remaining = bytes - total_evicted; 3790 else 3791 break; 3792 3793 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 3794 markers[sublist_idx], spa, bytes_remaining); 3795 3796 scan_evicted += bytes_evicted; 3797 total_evicted += bytes_evicted; 3798 3799 /* we've reached the end, wrap to the beginning */ 3800 if (++sublist_idx >= num_sublists) 3801 sublist_idx = 0; 3802 } 3803 3804 /* 3805 * If we didn't evict anything during this scan, we have 3806 * no reason to believe we'll evict more during another 3807 * scan, so break the loop. 3808 */ 3809 if (scan_evicted == 0) { 3810 /* This isn't possible, let's make that obvious */ 3811 ASSERT3S(bytes, !=, 0); 3812 3813 /* 3814 * When bytes is ARC_EVICT_ALL, the only way to 3815 * break the loop is when scan_evicted is zero. 3816 * In that case, we actually have evicted enough, 3817 * so we don't want to increment the kstat. 3818 */ 3819 if (bytes != ARC_EVICT_ALL) { 3820 ASSERT3S(total_evicted, <, bytes); 3821 ARCSTAT_BUMP(arcstat_evict_not_enough); 3822 } 3823 3824 break; 3825 } 3826 } 3827 3828 for (int i = 0; i < num_sublists; i++) { 3829 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3830 multilist_sublist_remove(mls, markers[i]); 3831 multilist_sublist_unlock(mls); 3832 3833 kmem_cache_free(hdr_full_cache, markers[i]); 3834 } 3835 kmem_free(markers, sizeof (*markers) * num_sublists); 3836 3837 return (total_evicted); 3838} 3839 3840/* 3841 * Flush all "evictable" data of the given type from the arc state 3842 * specified. This will not evict any "active" buffers (i.e. referenced). 3843 * 3844 * When 'retry' is set to B_FALSE, the function will make a single pass 3845 * over the state and evict any buffers that it can. Since it doesn't 3846 * continually retry the eviction, it might end up leaving some buffers 3847 * in the ARC due to lock misses. 3848 * 3849 * When 'retry' is set to B_TRUE, the function will continually retry the 3850 * eviction until *all* evictable buffers have been removed from the 3851 * state. As a result, if concurrent insertions into the state are 3852 * allowed (e.g. if the ARC isn't shutting down), this function might 3853 * wind up in an infinite loop, continually trying to evict buffers. 3854 */ 3855static uint64_t 3856arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 3857 boolean_t retry) 3858{ 3859 uint64_t evicted = 0; 3860 3861 while (refcount_count(&state->arcs_esize[type]) != 0) { 3862 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 3863 3864 if (!retry) 3865 break; 3866 } 3867 3868 return (evicted); 3869} 3870 3871/* 3872 * Evict the specified number of bytes from the state specified, 3873 * restricting eviction to the spa and type given. This function 3874 * prevents us from trying to evict more from a state's list than 3875 * is "evictable", and to skip evicting altogether when passed a 3876 * negative value for "bytes". In contrast, arc_evict_state() will 3877 * evict everything it can, when passed a negative value for "bytes". 3878 */ 3879static uint64_t 3880arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3881 arc_buf_contents_t type) 3882{ 3883 int64_t delta; 3884 3885 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) { 3886 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes); 3887 return (arc_evict_state(state, spa, delta, type)); 3888 } 3889 3890 return (0); 3891} 3892 3893/* 3894 * Evict metadata buffers from the cache, such that arc_meta_used is 3895 * capped by the arc_meta_limit tunable. 3896 */ 3897static uint64_t 3898arc_adjust_meta(void) 3899{ 3900 uint64_t total_evicted = 0; 3901 int64_t target; 3902 3903 /* 3904 * If we're over the meta limit, we want to evict enough 3905 * metadata to get back under the meta limit. We don't want to 3906 * evict so much that we drop the MRU below arc_p, though. If 3907 * we're over the meta limit more than we're over arc_p, we 3908 * evict some from the MRU here, and some from the MFU below. 3909 */ 3910 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3911 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3912 refcount_count(&arc_mru->arcs_size) - arc_p)); 3913 3914 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3915 3916 /* 3917 * Similar to the above, we want to evict enough bytes to get us 3918 * below the meta limit, but not so much as to drop us below the 3919 * space allotted to the MFU (which is defined as arc_c - arc_p). 3920 */ 3921 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3922 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3923 3924 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3925 3926 return (total_evicted); 3927} 3928 3929/* 3930 * Return the type of the oldest buffer in the given arc state 3931 * 3932 * This function will select a random sublist of type ARC_BUFC_DATA and 3933 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3934 * is compared, and the type which contains the "older" buffer will be 3935 * returned. 3936 */ 3937static arc_buf_contents_t 3938arc_adjust_type(arc_state_t *state) 3939{ 3940 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA]; 3941 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA]; 3942 int data_idx = multilist_get_random_index(data_ml); 3943 int meta_idx = multilist_get_random_index(meta_ml); 3944 multilist_sublist_t *data_mls; 3945 multilist_sublist_t *meta_mls; 3946 arc_buf_contents_t type; 3947 arc_buf_hdr_t *data_hdr; 3948 arc_buf_hdr_t *meta_hdr; 3949 3950 /* 3951 * We keep the sublist lock until we're finished, to prevent 3952 * the headers from being destroyed via arc_evict_state(). 3953 */ 3954 data_mls = multilist_sublist_lock(data_ml, data_idx); 3955 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3956 3957 /* 3958 * These two loops are to ensure we skip any markers that 3959 * might be at the tail of the lists due to arc_evict_state(). 3960 */ 3961 3962 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3963 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3964 if (data_hdr->b_spa != 0) 3965 break; 3966 } 3967 3968 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3969 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3970 if (meta_hdr->b_spa != 0) 3971 break; 3972 } 3973 3974 if (data_hdr == NULL && meta_hdr == NULL) { 3975 type = ARC_BUFC_DATA; 3976 } else if (data_hdr == NULL) { 3977 ASSERT3P(meta_hdr, !=, NULL); 3978 type = ARC_BUFC_METADATA; 3979 } else if (meta_hdr == NULL) { 3980 ASSERT3P(data_hdr, !=, NULL); 3981 type = ARC_BUFC_DATA; 3982 } else { 3983 ASSERT3P(data_hdr, !=, NULL); 3984 ASSERT3P(meta_hdr, !=, NULL); 3985 3986 /* The headers can't be on the sublist without an L1 header */ 3987 ASSERT(HDR_HAS_L1HDR(data_hdr)); 3988 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3989 3990 if (data_hdr->b_l1hdr.b_arc_access < 3991 meta_hdr->b_l1hdr.b_arc_access) { 3992 type = ARC_BUFC_DATA; 3993 } else { 3994 type = ARC_BUFC_METADATA; 3995 } 3996 } 3997 3998 multilist_sublist_unlock(meta_mls); 3999 multilist_sublist_unlock(data_mls); 4000 4001 return (type); 4002} 4003 4004/* 4005 * Evict buffers from the cache, such that arc_size is capped by arc_c. 4006 */ 4007static uint64_t 4008arc_adjust(void) 4009{ 4010 uint64_t total_evicted = 0; 4011 uint64_t bytes; 4012 int64_t target; 4013 4014 /* 4015 * If we're over arc_meta_limit, we want to correct that before 4016 * potentially evicting data buffers below. 4017 */ 4018 total_evicted += arc_adjust_meta(); 4019 4020 /* 4021 * Adjust MRU size 4022 * 4023 * If we're over the target cache size, we want to evict enough 4024 * from the list to get back to our target size. We don't want 4025 * to evict too much from the MRU, such that it drops below 4026 * arc_p. So, if we're over our target cache size more than 4027 * the MRU is over arc_p, we'll evict enough to get back to 4028 * arc_p here, and then evict more from the MFU below. 4029 */ 4030 target = MIN((int64_t)(arc_size - arc_c), 4031 (int64_t)(refcount_count(&arc_anon->arcs_size) + 4032 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 4033 4034 /* 4035 * If we're below arc_meta_min, always prefer to evict data. 4036 * Otherwise, try to satisfy the requested number of bytes to 4037 * evict from the type which contains older buffers; in an 4038 * effort to keep newer buffers in the cache regardless of their 4039 * type. If we cannot satisfy the number of bytes from this 4040 * type, spill over into the next type. 4041 */ 4042 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 4043 arc_meta_used > arc_meta_min) { 4044 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 4045 total_evicted += bytes; 4046 4047 /* 4048 * If we couldn't evict our target number of bytes from 4049 * metadata, we try to get the rest from data. 4050 */ 4051 target -= bytes; 4052 4053 total_evicted += 4054 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 4055 } else { 4056 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 4057 total_evicted += bytes; 4058 4059 /* 4060 * If we couldn't evict our target number of bytes from 4061 * data, we try to get the rest from metadata. 4062 */ 4063 target -= bytes; 4064 4065 total_evicted += 4066 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 4067 } 4068 4069 /* 4070 * Adjust MFU size 4071 * 4072 * Now that we've tried to evict enough from the MRU to get its 4073 * size back to arc_p, if we're still above the target cache 4074 * size, we evict the rest from the MFU. 4075 */ 4076 target = arc_size - arc_c; 4077 4078 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 4079 arc_meta_used > arc_meta_min) { 4080 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 4081 total_evicted += bytes; 4082 4083 /* 4084 * If we couldn't evict our target number of bytes from 4085 * metadata, we try to get the rest from data. 4086 */ 4087 target -= bytes; 4088 4089 total_evicted += 4090 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 4091 } else { 4092 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 4093 total_evicted += bytes; 4094 4095 /* 4096 * If we couldn't evict our target number of bytes from 4097 * data, we try to get the rest from data. 4098 */ 4099 target -= bytes; 4100 4101 total_evicted += 4102 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 4103 } 4104 4105 /* 4106 * Adjust ghost lists 4107 * 4108 * In addition to the above, the ARC also defines target values 4109 * for the ghost lists. The sum of the mru list and mru ghost 4110 * list should never exceed the target size of the cache, and 4111 * the sum of the mru list, mfu list, mru ghost list, and mfu 4112 * ghost list should never exceed twice the target size of the 4113 * cache. The following logic enforces these limits on the ghost 4114 * caches, and evicts from them as needed. 4115 */ 4116 target = refcount_count(&arc_mru->arcs_size) + 4117 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 4118 4119 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 4120 total_evicted += bytes; 4121 4122 target -= bytes; 4123 4124 total_evicted += 4125 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 4126 4127 /* 4128 * We assume the sum of the mru list and mfu list is less than 4129 * or equal to arc_c (we enforced this above), which means we 4130 * can use the simpler of the two equations below: 4131 * 4132 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 4133 * mru ghost + mfu ghost <= arc_c 4134 */ 4135 target = refcount_count(&arc_mru_ghost->arcs_size) + 4136 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 4137 4138 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 4139 total_evicted += bytes; 4140 4141 target -= bytes; 4142 4143 total_evicted += 4144 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 4145 4146 return (total_evicted); 4147} 4148 4149void 4150arc_flush(spa_t *spa, boolean_t retry) 4151{ 4152 uint64_t guid = 0; 4153 4154 /* 4155 * If retry is B_TRUE, a spa must not be specified since we have 4156 * no good way to determine if all of a spa's buffers have been 4157 * evicted from an arc state. 4158 */ 4159 ASSERT(!retry || spa == 0); 4160 4161 if (spa != NULL) 4162 guid = spa_load_guid(spa); 4163 4164 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 4165 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 4166 4167 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 4168 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 4169 4170 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 4171 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 4172 4173 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 4174 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 4175} 4176 4177void 4178arc_shrink(int64_t to_free) 4179{ 4180 if (arc_c > arc_c_min) { 4181 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 4182 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 4183 if (arc_c > arc_c_min + to_free) 4184 atomic_add_64(&arc_c, -to_free); 4185 else 4186 arc_c = arc_c_min; 4187 4188 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 4189 if (arc_c > arc_size) 4190 arc_c = MAX(arc_size, arc_c_min); 4191 if (arc_p > arc_c) 4192 arc_p = (arc_c >> 1); 4193 4194 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 4195 arc_p); 4196 4197 ASSERT(arc_c >= arc_c_min); 4198 ASSERT((int64_t)arc_p >= 0); 4199 } 4200 4201 if (arc_size > arc_c) { 4202 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 4203 uint64_t, arc_c); 4204 (void) arc_adjust(); 4205 } 4206} 4207 4208typedef enum free_memory_reason_t { 4209 FMR_UNKNOWN, 4210 FMR_NEEDFREE, 4211 FMR_LOTSFREE, 4212 FMR_SWAPFS_MINFREE, 4213 FMR_PAGES_PP_MAXIMUM, 4214 FMR_HEAP_ARENA, 4215 FMR_ZIO_ARENA, 4216 FMR_ZIO_FRAG, 4217} free_memory_reason_t; 4218 4219int64_t last_free_memory; 4220free_memory_reason_t last_free_reason; 4221 4222/* 4223 * Additional reserve of pages for pp_reserve. 4224 */ 4225int64_t arc_pages_pp_reserve = 64; 4226 4227/* 4228 * Additional reserve of pages for swapfs. 4229 */ 4230int64_t arc_swapfs_reserve = 64; 4231 4232/* 4233 * Return the amount of memory that can be consumed before reclaim will be 4234 * needed. Positive if there is sufficient free memory, negative indicates 4235 * the amount of memory that needs to be freed up. 4236 */ 4237static int64_t 4238arc_available_memory(void) 4239{ 4240 int64_t lowest = INT64_MAX; 4241 int64_t n; 4242 free_memory_reason_t r = FMR_UNKNOWN; 4243 4244#ifdef _KERNEL 4245#ifdef __FreeBSD__ 4246 /* 4247 * Cooperate with pagedaemon when it's time for it to scan 4248 * and reclaim some pages. 4249 */ 4250 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 4251 if (n < lowest) { 4252 lowest = n; 4253 r = FMR_LOTSFREE; 4254 } 4255 4256#else 4257 if (needfree > 0) { 4258 n = PAGESIZE * (-needfree); 4259 if (n < lowest) { 4260 lowest = n; 4261 r = FMR_NEEDFREE; 4262 } 4263 } 4264 4265 /* 4266 * check that we're out of range of the pageout scanner. It starts to 4267 * schedule paging if freemem is less than lotsfree and needfree. 4268 * lotsfree is the high-water mark for pageout, and needfree is the 4269 * number of needed free pages. We add extra pages here to make sure 4270 * the scanner doesn't start up while we're freeing memory. 4271 */ 4272 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 4273 if (n < lowest) { 4274 lowest = n; 4275 r = FMR_LOTSFREE; 4276 } 4277 4278 /* 4279 * check to make sure that swapfs has enough space so that anon 4280 * reservations can still succeed. anon_resvmem() checks that the 4281 * availrmem is greater than swapfs_minfree, and the number of reserved 4282 * swap pages. We also add a bit of extra here just to prevent 4283 * circumstances from getting really dire. 4284 */ 4285 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 4286 desfree - arc_swapfs_reserve); 4287 if (n < lowest) { 4288 lowest = n; 4289 r = FMR_SWAPFS_MINFREE; 4290 } 4291 4292 4293 /* 4294 * Check that we have enough availrmem that memory locking (e.g., via 4295 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 4296 * stores the number of pages that cannot be locked; when availrmem 4297 * drops below pages_pp_maximum, page locking mechanisms such as 4298 * page_pp_lock() will fail.) 4299 */ 4300 n = PAGESIZE * (availrmem - pages_pp_maximum - 4301 arc_pages_pp_reserve); 4302 if (n < lowest) { 4303 lowest = n; 4304 r = FMR_PAGES_PP_MAXIMUM; 4305 } 4306 4307#endif /* __FreeBSD__ */ 4308#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4309 /* 4310 * If we're on an i386 platform, it's possible that we'll exhaust the 4311 * kernel heap space before we ever run out of available physical 4312 * memory. Most checks of the size of the heap_area compare against 4313 * tune.t_minarmem, which is the minimum available real memory that we 4314 * can have in the system. However, this is generally fixed at 25 pages 4315 * which is so low that it's useless. In this comparison, we seek to 4316 * calculate the total heap-size, and reclaim if more than 3/4ths of the 4317 * heap is allocated. (Or, in the calculation, if less than 1/4th is 4318 * free) 4319 */ 4320 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 4321 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 4322 if (n < lowest) { 4323 lowest = n; 4324 r = FMR_HEAP_ARENA; 4325 } 4326#define zio_arena NULL 4327#else 4328#define zio_arena heap_arena 4329#endif 4330 4331 /* 4332 * If zio data pages are being allocated out of a separate heap segment, 4333 * then enforce that the size of available vmem for this arena remains 4334 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free. 4335 * 4336 * Note that reducing the arc_zio_arena_free_shift keeps more virtual 4337 * memory (in the zio_arena) free, which can avoid memory 4338 * fragmentation issues. 4339 */ 4340 if (zio_arena != NULL) { 4341 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 4342 (vmem_size(zio_arena, VMEM_ALLOC) >> 4343 arc_zio_arena_free_shift); 4344 if (n < lowest) { 4345 lowest = n; 4346 r = FMR_ZIO_ARENA; 4347 } 4348 } 4349 4350 /* 4351 * Above limits know nothing about real level of KVA fragmentation. 4352 * Start aggressive reclamation if too little sequential KVA left. 4353 */ 4354 if (lowest > 0) { 4355 n = (vmem_size(heap_arena, VMEM_MAXFREE) < SPA_MAXBLOCKSIZE) ? 4356 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 4357 INT64_MAX; 4358 if (n < lowest) { 4359 lowest = n; 4360 r = FMR_ZIO_FRAG; 4361 } 4362 } 4363 4364#else /* _KERNEL */ 4365 /* Every 100 calls, free a small amount */ 4366 if (spa_get_random(100) == 0) 4367 lowest = -1024; 4368#endif /* _KERNEL */ 4369 4370 last_free_memory = lowest; 4371 last_free_reason = r; 4372 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 4373 return (lowest); 4374} 4375 4376 4377/* 4378 * Determine if the system is under memory pressure and is asking 4379 * to reclaim memory. A return value of B_TRUE indicates that the system 4380 * is under memory pressure and that the arc should adjust accordingly. 4381 */ 4382static boolean_t 4383arc_reclaim_needed(void) 4384{ 4385 return (arc_available_memory() < 0); 4386} 4387 4388extern kmem_cache_t *zio_buf_cache[]; 4389extern kmem_cache_t *zio_data_buf_cache[]; 4390extern kmem_cache_t *range_seg_cache; 4391extern kmem_cache_t *abd_chunk_cache; 4392 4393static __noinline void 4394arc_kmem_reap_now(void) 4395{ 4396 size_t i; 4397 kmem_cache_t *prev_cache = NULL; 4398 kmem_cache_t *prev_data_cache = NULL; 4399 4400 DTRACE_PROBE(arc__kmem_reap_start); 4401#ifdef _KERNEL 4402 if (arc_meta_used >= arc_meta_limit) { 4403 /* 4404 * We are exceeding our meta-data cache limit. 4405 * Purge some DNLC entries to release holds on meta-data. 4406 */ 4407 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 4408 } 4409#if defined(__i386) 4410 /* 4411 * Reclaim unused memory from all kmem caches. 4412 */ 4413 kmem_reap(); 4414#endif 4415#endif 4416 4417 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 4418 if (zio_buf_cache[i] != prev_cache) { 4419 prev_cache = zio_buf_cache[i]; 4420 kmem_cache_reap_now(zio_buf_cache[i]); 4421 } 4422 if (zio_data_buf_cache[i] != prev_data_cache) { 4423 prev_data_cache = zio_data_buf_cache[i]; 4424 kmem_cache_reap_now(zio_data_buf_cache[i]); 4425 } 4426 } 4427 kmem_cache_reap_now(abd_chunk_cache); 4428 kmem_cache_reap_now(buf_cache); 4429 kmem_cache_reap_now(hdr_full_cache); 4430 kmem_cache_reap_now(hdr_l2only_cache); 4431 kmem_cache_reap_now(range_seg_cache); 4432 4433#ifdef illumos 4434 if (zio_arena != NULL) { 4435 /* 4436 * Ask the vmem arena to reclaim unused memory from its 4437 * quantum caches. 4438 */ 4439 vmem_qcache_reap(zio_arena); 4440 } 4441#endif 4442 DTRACE_PROBE(arc__kmem_reap_end); 4443} 4444 4445/* 4446 * Threads can block in arc_get_data_impl() waiting for this thread to evict 4447 * enough data and signal them to proceed. When this happens, the threads in 4448 * arc_get_data_impl() are sleeping while holding the hash lock for their 4449 * particular arc header. Thus, we must be careful to never sleep on a 4450 * hash lock in this thread. This is to prevent the following deadlock: 4451 * 4452 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L", 4453 * waiting for the reclaim thread to signal it. 4454 * 4455 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 4456 * fails, and goes to sleep forever. 4457 * 4458 * This possible deadlock is avoided by always acquiring a hash lock 4459 * using mutex_tryenter() from arc_reclaim_thread(). 4460 */ 4461/* ARGSUSED */ 4462static void 4463arc_reclaim_thread(void *unused __unused) 4464{ 4465 hrtime_t growtime = 0; 4466 callb_cpr_t cpr; 4467 4468 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 4469 4470 mutex_enter(&arc_reclaim_lock); 4471 while (!arc_reclaim_thread_exit) { 4472 uint64_t evicted = 0; 4473 4474 /* 4475 * This is necessary in order for the mdb ::arc dcmd to 4476 * show up to date information. Since the ::arc command 4477 * does not call the kstat's update function, without 4478 * this call, the command may show stale stats for the 4479 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 4480 * with this change, the data might be up to 1 second 4481 * out of date; but that should suffice. The arc_state_t 4482 * structures can be queried directly if more accurate 4483 * information is needed. 4484 */ 4485 if (arc_ksp != NULL) 4486 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 4487 4488 mutex_exit(&arc_reclaim_lock); 4489 4490 /* 4491 * We call arc_adjust() before (possibly) calling 4492 * arc_kmem_reap_now(), so that we can wake up 4493 * arc_get_data_impl() sooner. 4494 */ 4495 evicted = arc_adjust(); 4496 4497 int64_t free_memory = arc_available_memory(); 4498 if (free_memory < 0) { 4499 4500 arc_no_grow = B_TRUE; 4501 arc_warm = B_TRUE; 4502 4503 /* 4504 * Wait at least zfs_grow_retry (default 60) seconds 4505 * before considering growing. 4506 */ 4507 growtime = gethrtime() + SEC2NSEC(arc_grow_retry); 4508 4509 arc_kmem_reap_now(); 4510 4511 /* 4512 * If we are still low on memory, shrink the ARC 4513 * so that we have arc_shrink_min free space. 4514 */ 4515 free_memory = arc_available_memory(); 4516 4517 int64_t to_free = 4518 (arc_c >> arc_shrink_shift) - free_memory; 4519 if (to_free > 0) { 4520#ifdef _KERNEL 4521#ifdef illumos 4522 to_free = MAX(to_free, ptob(needfree)); 4523#endif 4524#endif 4525 arc_shrink(to_free); 4526 } 4527 } else if (free_memory < arc_c >> arc_no_grow_shift) { 4528 arc_no_grow = B_TRUE; 4529 } else if (gethrtime() >= growtime) { 4530 arc_no_grow = B_FALSE; 4531 } 4532 4533 mutex_enter(&arc_reclaim_lock); 4534 4535 /* 4536 * If evicted is zero, we couldn't evict anything via 4537 * arc_adjust(). This could be due to hash lock 4538 * collisions, but more likely due to the majority of 4539 * arc buffers being unevictable. Therefore, even if 4540 * arc_size is above arc_c, another pass is unlikely to 4541 * be helpful and could potentially cause us to enter an 4542 * infinite loop. 4543 */ 4544 if (arc_size <= arc_c || evicted == 0) { 4545 /* 4546 * We're either no longer overflowing, or we 4547 * can't evict anything more, so we should wake 4548 * up any threads before we go to sleep. 4549 */ 4550 cv_broadcast(&arc_reclaim_waiters_cv); 4551 4552 /* 4553 * Block until signaled, or after one second (we 4554 * might need to perform arc_kmem_reap_now() 4555 * even if we aren't being signalled) 4556 */ 4557 CALLB_CPR_SAFE_BEGIN(&cpr); 4558 (void) cv_timedwait_hires(&arc_reclaim_thread_cv, 4559 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 4560 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 4561 } 4562 } 4563 4564 arc_reclaim_thread_exit = B_FALSE; 4565 cv_broadcast(&arc_reclaim_thread_cv); 4566 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 4567 thread_exit(); 4568} 4569 4570static u_int arc_dnlc_evicts_arg; 4571extern struct vfsops zfs_vfsops; 4572 4573static void 4574arc_dnlc_evicts_thread(void *dummy __unused) 4575{ 4576 callb_cpr_t cpr; 4577 u_int percent; 4578 4579 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG); 4580 4581 mutex_enter(&arc_dnlc_evicts_lock); 4582 while (!arc_dnlc_evicts_thread_exit) { 4583 CALLB_CPR_SAFE_BEGIN(&cpr); 4584 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock); 4585 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock); 4586 if (arc_dnlc_evicts_arg != 0) { 4587 percent = arc_dnlc_evicts_arg; 4588 mutex_exit(&arc_dnlc_evicts_lock); 4589#ifdef _KERNEL 4590 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops); 4591#endif 4592 mutex_enter(&arc_dnlc_evicts_lock); 4593 /* 4594 * Clear our token only after vnlru_free() 4595 * pass is done, to avoid false queueing of 4596 * the requests. 4597 */ 4598 arc_dnlc_evicts_arg = 0; 4599 } 4600 } 4601 arc_dnlc_evicts_thread_exit = FALSE; 4602 cv_broadcast(&arc_dnlc_evicts_cv); 4603 CALLB_CPR_EXIT(&cpr); 4604 thread_exit(); 4605} 4606 4607void 4608dnlc_reduce_cache(void *arg) 4609{ 4610 u_int percent; 4611 4612 percent = (u_int)(uintptr_t)arg; 4613 mutex_enter(&arc_dnlc_evicts_lock); 4614 if (arc_dnlc_evicts_arg == 0) { 4615 arc_dnlc_evicts_arg = percent; 4616 cv_broadcast(&arc_dnlc_evicts_cv); 4617 } 4618 mutex_exit(&arc_dnlc_evicts_lock); 4619} 4620 4621/* 4622 * Adapt arc info given the number of bytes we are trying to add and 4623 * the state that we are comming from. This function is only called 4624 * when we are adding new content to the cache. 4625 */ 4626static void 4627arc_adapt(int bytes, arc_state_t *state) 4628{ 4629 int mult; 4630 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 4631 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 4632 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 4633 4634 if (state == arc_l2c_only) 4635 return; 4636 4637 ASSERT(bytes > 0); 4638 /* 4639 * Adapt the target size of the MRU list: 4640 * - if we just hit in the MRU ghost list, then increase 4641 * the target size of the MRU list. 4642 * - if we just hit in the MFU ghost list, then increase 4643 * the target size of the MFU list by decreasing the 4644 * target size of the MRU list. 4645 */ 4646 if (state == arc_mru_ghost) { 4647 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 4648 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 4649 4650 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 4651 } else if (state == arc_mfu_ghost) { 4652 uint64_t delta; 4653 4654 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 4655 mult = MIN(mult, 10); 4656 4657 delta = MIN(bytes * mult, arc_p); 4658 arc_p = MAX(arc_p_min, arc_p - delta); 4659 } 4660 ASSERT((int64_t)arc_p >= 0); 4661 4662 if (arc_reclaim_needed()) { 4663 cv_signal(&arc_reclaim_thread_cv); 4664 return; 4665 } 4666 4667 if (arc_no_grow) 4668 return; 4669 4670 if (arc_c >= arc_c_max) 4671 return; 4672 4673 /* 4674 * If we're within (2 * maxblocksize) bytes of the target 4675 * cache size, increment the target cache size 4676 */ 4677 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 4678 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 4679 atomic_add_64(&arc_c, (int64_t)bytes); 4680 if (arc_c > arc_c_max) 4681 arc_c = arc_c_max; 4682 else if (state == arc_anon) 4683 atomic_add_64(&arc_p, (int64_t)bytes); 4684 if (arc_p > arc_c) 4685 arc_p = arc_c; 4686 } 4687 ASSERT((int64_t)arc_p >= 0); 4688} 4689 4690/* 4691 * Check if arc_size has grown past our upper threshold, determined by 4692 * zfs_arc_overflow_shift. 4693 */ 4694static boolean_t 4695arc_is_overflowing(void) 4696{ 4697 /* Always allow at least one block of overflow */ 4698 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 4699 arc_c >> zfs_arc_overflow_shift); 4700 4701 return (arc_size >= arc_c + overflow); 4702} 4703 4704static abd_t * 4705arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4706{ 4707 arc_buf_contents_t type = arc_buf_type(hdr); 4708 4709 arc_get_data_impl(hdr, size, tag); 4710 if (type == ARC_BUFC_METADATA) { 4711 return (abd_alloc(size, B_TRUE)); 4712 } else { 4713 ASSERT(type == ARC_BUFC_DATA); 4714 return (abd_alloc(size, B_FALSE)); 4715 } 4716} 4717 4718static void * 4719arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4720{ 4721 arc_buf_contents_t type = arc_buf_type(hdr); 4722 4723 arc_get_data_impl(hdr, size, tag); 4724 if (type == ARC_BUFC_METADATA) { 4725 return (zio_buf_alloc(size)); 4726 } else { 4727 ASSERT(type == ARC_BUFC_DATA); 4728 return (zio_data_buf_alloc(size)); 4729 } 4730} 4731 4732/* 4733 * Allocate a block and return it to the caller. If we are hitting the 4734 * hard limit for the cache size, we must sleep, waiting for the eviction 4735 * thread to catch up. If we're past the target size but below the hard 4736 * limit, we'll only signal the reclaim thread and continue on. 4737 */ 4738static void 4739arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4740{ 4741 arc_state_t *state = hdr->b_l1hdr.b_state; 4742 arc_buf_contents_t type = arc_buf_type(hdr); 4743 4744 arc_adapt(size, state); 4745 4746 /* 4747 * If arc_size is currently overflowing, and has grown past our 4748 * upper limit, we must be adding data faster than the evict 4749 * thread can evict. Thus, to ensure we don't compound the 4750 * problem by adding more data and forcing arc_size to grow even 4751 * further past it's target size, we halt and wait for the 4752 * eviction thread to catch up. 4753 * 4754 * It's also possible that the reclaim thread is unable to evict 4755 * enough buffers to get arc_size below the overflow limit (e.g. 4756 * due to buffers being un-evictable, or hash lock collisions). 4757 * In this case, we want to proceed regardless if we're 4758 * overflowing; thus we don't use a while loop here. 4759 */ 4760 if (arc_is_overflowing()) { 4761 mutex_enter(&arc_reclaim_lock); 4762 4763 /* 4764 * Now that we've acquired the lock, we may no longer be 4765 * over the overflow limit, lets check. 4766 * 4767 * We're ignoring the case of spurious wake ups. If that 4768 * were to happen, it'd let this thread consume an ARC 4769 * buffer before it should have (i.e. before we're under 4770 * the overflow limit and were signalled by the reclaim 4771 * thread). As long as that is a rare occurrence, it 4772 * shouldn't cause any harm. 4773 */ 4774 if (arc_is_overflowing()) { 4775 cv_signal(&arc_reclaim_thread_cv); 4776 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 4777 } 4778 4779 mutex_exit(&arc_reclaim_lock); 4780 } 4781 4782 VERIFY3U(hdr->b_type, ==, type); 4783 if (type == ARC_BUFC_METADATA) { 4784 arc_space_consume(size, ARC_SPACE_META); 4785 } else { 4786 arc_space_consume(size, ARC_SPACE_DATA); 4787 } 4788 4789 /* 4790 * Update the state size. Note that ghost states have a 4791 * "ghost size" and so don't need to be updated. 4792 */ 4793 if (!GHOST_STATE(state)) { 4794 4795 (void) refcount_add_many(&state->arcs_size, size, tag); 4796 4797 /* 4798 * If this is reached via arc_read, the link is 4799 * protected by the hash lock. If reached via 4800 * arc_buf_alloc, the header should not be accessed by 4801 * any other thread. And, if reached via arc_read_done, 4802 * the hash lock will protect it if it's found in the 4803 * hash table; otherwise no other thread should be 4804 * trying to [add|remove]_reference it. 4805 */ 4806 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4807 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4808 (void) refcount_add_many(&state->arcs_esize[type], 4809 size, tag); 4810 } 4811 4812 /* 4813 * If we are growing the cache, and we are adding anonymous 4814 * data, and we have outgrown arc_p, update arc_p 4815 */ 4816 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 4817 (refcount_count(&arc_anon->arcs_size) + 4818 refcount_count(&arc_mru->arcs_size) > arc_p)) 4819 arc_p = MIN(arc_c, arc_p + size); 4820 } 4821 ARCSTAT_BUMP(arcstat_allocated); 4822} 4823 4824static void 4825arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag) 4826{ 4827 arc_free_data_impl(hdr, size, tag); 4828 abd_free(abd); 4829} 4830 4831static void 4832arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag) 4833{ 4834 arc_buf_contents_t type = arc_buf_type(hdr); 4835 4836 arc_free_data_impl(hdr, size, tag); 4837 if (type == ARC_BUFC_METADATA) { 4838 zio_buf_free(buf, size); 4839 } else { 4840 ASSERT(type == ARC_BUFC_DATA); 4841 zio_data_buf_free(buf, size); 4842 } 4843} 4844 4845/* 4846 * Free the arc data buffer. 4847 */ 4848static void 4849arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4850{ 4851 arc_state_t *state = hdr->b_l1hdr.b_state; 4852 arc_buf_contents_t type = arc_buf_type(hdr); 4853 4854 /* protected by hash lock, if in the hash table */ 4855 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4856 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4857 ASSERT(state != arc_anon && state != arc_l2c_only); 4858 4859 (void) refcount_remove_many(&state->arcs_esize[type], 4860 size, tag); 4861 } 4862 (void) refcount_remove_many(&state->arcs_size, size, tag); 4863 4864 VERIFY3U(hdr->b_type, ==, type); 4865 if (type == ARC_BUFC_METADATA) { 4866 arc_space_return(size, ARC_SPACE_META); 4867 } else { 4868 ASSERT(type == ARC_BUFC_DATA); 4869 arc_space_return(size, ARC_SPACE_DATA); 4870 } 4871} 4872 4873/* 4874 * This routine is called whenever a buffer is accessed. 4875 * NOTE: the hash lock is dropped in this function. 4876 */ 4877static void 4878arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 4879{ 4880 clock_t now; 4881 4882 ASSERT(MUTEX_HELD(hash_lock)); 4883 ASSERT(HDR_HAS_L1HDR(hdr)); 4884 4885 if (hdr->b_l1hdr.b_state == arc_anon) { 4886 /* 4887 * This buffer is not in the cache, and does not 4888 * appear in our "ghost" list. Add the new buffer 4889 * to the MRU state. 4890 */ 4891 4892 ASSERT0(hdr->b_l1hdr.b_arc_access); 4893 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4894 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4895 arc_change_state(arc_mru, hdr, hash_lock); 4896 4897 } else if (hdr->b_l1hdr.b_state == arc_mru) { 4898 now = ddi_get_lbolt(); 4899 4900 /* 4901 * If this buffer is here because of a prefetch, then either: 4902 * - clear the flag if this is a "referencing" read 4903 * (any subsequent access will bump this into the MFU state). 4904 * or 4905 * - move the buffer to the head of the list if this is 4906 * another prefetch (to make it less likely to be evicted). 4907 */ 4908 if (HDR_PREFETCH(hdr)) { 4909 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4910 /* link protected by hash lock */ 4911 ASSERT(multilist_link_active( 4912 &hdr->b_l1hdr.b_arc_node)); 4913 } else { 4914 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 4915 ARCSTAT_BUMP(arcstat_mru_hits); 4916 } 4917 hdr->b_l1hdr.b_arc_access = now; 4918 return; 4919 } 4920 4921 /* 4922 * This buffer has been "accessed" only once so far, 4923 * but it is still in the cache. Move it to the MFU 4924 * state. 4925 */ 4926 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 4927 /* 4928 * More than 125ms have passed since we 4929 * instantiated this buffer. Move it to the 4930 * most frequently used state. 4931 */ 4932 hdr->b_l1hdr.b_arc_access = now; 4933 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4934 arc_change_state(arc_mfu, hdr, hash_lock); 4935 } 4936 ARCSTAT_BUMP(arcstat_mru_hits); 4937 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 4938 arc_state_t *new_state; 4939 /* 4940 * This buffer has been "accessed" recently, but 4941 * was evicted from the cache. Move it to the 4942 * MFU state. 4943 */ 4944 4945 if (HDR_PREFETCH(hdr)) { 4946 new_state = arc_mru; 4947 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 4948 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 4949 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4950 } else { 4951 new_state = arc_mfu; 4952 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4953 } 4954 4955 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4956 arc_change_state(new_state, hdr, hash_lock); 4957 4958 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 4959 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 4960 /* 4961 * This buffer has been accessed more than once and is 4962 * still in the cache. Keep it in the MFU state. 4963 * 4964 * NOTE: an add_reference() that occurred when we did 4965 * the arc_read() will have kicked this off the list. 4966 * If it was a prefetch, we will explicitly move it to 4967 * the head of the list now. 4968 */ 4969 if ((HDR_PREFETCH(hdr)) != 0) { 4970 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4971 /* link protected by hash_lock */ 4972 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4973 } 4974 ARCSTAT_BUMP(arcstat_mfu_hits); 4975 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4976 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 4977 arc_state_t *new_state = arc_mfu; 4978 /* 4979 * This buffer has been accessed more than once but has 4980 * been evicted from the cache. Move it back to the 4981 * MFU state. 4982 */ 4983 4984 if (HDR_PREFETCH(hdr)) { 4985 /* 4986 * This is a prefetch access... 4987 * move this block back to the MRU state. 4988 */ 4989 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4990 new_state = arc_mru; 4991 } 4992 4993 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4994 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4995 arc_change_state(new_state, hdr, hash_lock); 4996 4997 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 4998 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4999 /* 5000 * This buffer is on the 2nd Level ARC. 5001 */ 5002 5003 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 5004 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 5005 arc_change_state(arc_mfu, hdr, hash_lock); 5006 } else { 5007 ASSERT(!"invalid arc state"); 5008 } 5009} 5010 5011/* a generic arc_done_func_t which you can use */ 5012/* ARGSUSED */ 5013void 5014arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 5015{ 5016 if (zio == NULL || zio->io_error == 0) 5017 bcopy(buf->b_data, arg, arc_buf_size(buf)); 5018 arc_buf_destroy(buf, arg); 5019} 5020 5021/* a generic arc_done_func_t */ 5022void 5023arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 5024{ 5025 arc_buf_t **bufp = arg; 5026 if (zio && zio->io_error) { 5027 arc_buf_destroy(buf, arg); 5028 *bufp = NULL; 5029 } else { 5030 *bufp = buf; 5031 ASSERT(buf->b_data); 5032 } 5033} 5034 5035static void 5036arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) 5037{ 5038 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { 5039 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); 5040 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 5041 } else { 5042 if (HDR_COMPRESSION_ENABLED(hdr)) { 5043 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, 5044 BP_GET_COMPRESS(bp)); 5045 } 5046 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); 5047 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); 5048 } 5049} 5050 5051static void 5052arc_read_done(zio_t *zio) 5053{ 5054 arc_buf_hdr_t *hdr = zio->io_private; 5055 kmutex_t *hash_lock = NULL; 5056 arc_callback_t *callback_list; 5057 arc_callback_t *acb; 5058 boolean_t freeable = B_FALSE; 5059 boolean_t no_zio_error = (zio->io_error == 0); 5060 5061 /* 5062 * The hdr was inserted into hash-table and removed from lists 5063 * prior to starting I/O. We should find this header, since 5064 * it's in the hash table, and it should be legit since it's 5065 * not possible to evict it during the I/O. The only possible 5066 * reason for it not to be found is if we were freed during the 5067 * read. 5068 */ 5069 if (HDR_IN_HASH_TABLE(hdr)) { 5070 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 5071 ASSERT3U(hdr->b_dva.dva_word[0], ==, 5072 BP_IDENTITY(zio->io_bp)->dva_word[0]); 5073 ASSERT3U(hdr->b_dva.dva_word[1], ==, 5074 BP_IDENTITY(zio->io_bp)->dva_word[1]); 5075 5076 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 5077 &hash_lock); 5078 5079 ASSERT((found == hdr && 5080 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 5081 (found == hdr && HDR_L2_READING(hdr))); 5082 ASSERT3P(hash_lock, !=, NULL); 5083 } 5084 5085 if (no_zio_error) { 5086 /* byteswap if necessary */ 5087 if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 5088 if (BP_GET_LEVEL(zio->io_bp) > 0) { 5089 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; 5090 } else { 5091 hdr->b_l1hdr.b_byteswap = 5092 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 5093 } 5094 } else { 5095 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 5096 } 5097 } 5098 5099 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); 5100 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 5101 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); 5102 5103 callback_list = hdr->b_l1hdr.b_acb; 5104 ASSERT3P(callback_list, !=, NULL); 5105 5106 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) { 5107 /* 5108 * Only call arc_access on anonymous buffers. This is because 5109 * if we've issued an I/O for an evicted buffer, we've already 5110 * called arc_access (to prevent any simultaneous readers from 5111 * getting confused). 5112 */ 5113 arc_access(hdr, hash_lock); 5114 } 5115 5116 /* 5117 * If a read request has a callback (i.e. acb_done is not NULL), then we 5118 * make a buf containing the data according to the parameters which were 5119 * passed in. The implementation of arc_buf_alloc_impl() ensures that we 5120 * aren't needlessly decompressing the data multiple times. 5121 */ 5122 int callback_cnt = 0; 5123 for (acb = callback_list; acb != NULL; acb = acb->acb_next) { 5124 if (!acb->acb_done) 5125 continue; 5126 5127 /* This is a demand read since prefetches don't use callbacks */ 5128 callback_cnt++; 5129 5130 int error = arc_buf_alloc_impl(hdr, acb->acb_private, 5131 acb->acb_compressed, no_zio_error, &acb->acb_buf); 5132 if (no_zio_error) { 5133 zio->io_error = error; 5134 } 5135 } 5136 hdr->b_l1hdr.b_acb = NULL; 5137 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5138 if (callback_cnt == 0) { 5139 ASSERT(HDR_PREFETCH(hdr)); 5140 ASSERT0(hdr->b_l1hdr.b_bufcnt); 5141 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5142 } 5143 5144 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 5145 callback_list != NULL); 5146 5147 if (no_zio_error) { 5148 arc_hdr_verify(hdr, zio->io_bp); 5149 } else { 5150 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); 5151 if (hdr->b_l1hdr.b_state != arc_anon) 5152 arc_change_state(arc_anon, hdr, hash_lock); 5153 if (HDR_IN_HASH_TABLE(hdr)) 5154 buf_hash_remove(hdr); 5155 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 5156 } 5157 5158 /* 5159 * Broadcast before we drop the hash_lock to avoid the possibility 5160 * that the hdr (and hence the cv) might be freed before we get to 5161 * the cv_broadcast(). 5162 */ 5163 cv_broadcast(&hdr->b_l1hdr.b_cv); 5164 5165 if (hash_lock != NULL) { 5166 mutex_exit(hash_lock); 5167 } else { 5168 /* 5169 * This block was freed while we waited for the read to 5170 * complete. It has been removed from the hash table and 5171 * moved to the anonymous state (so that it won't show up 5172 * in the cache). 5173 */ 5174 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 5175 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 5176 } 5177 5178 /* execute each callback and free its structure */ 5179 while ((acb = callback_list) != NULL) { 5180 if (acb->acb_done) 5181 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 5182 5183 if (acb->acb_zio_dummy != NULL) { 5184 acb->acb_zio_dummy->io_error = zio->io_error; 5185 zio_nowait(acb->acb_zio_dummy); 5186 } 5187 5188 callback_list = acb->acb_next; 5189 kmem_free(acb, sizeof (arc_callback_t)); 5190 } 5191 5192 if (freeable) 5193 arc_hdr_destroy(hdr); 5194} 5195 5196/* 5197 * "Read" the block at the specified DVA (in bp) via the 5198 * cache. If the block is found in the cache, invoke the provided 5199 * callback immediately and return. Note that the `zio' parameter 5200 * in the callback will be NULL in this case, since no IO was 5201 * required. If the block is not in the cache pass the read request 5202 * on to the spa with a substitute callback function, so that the 5203 * requested block will be added to the cache. 5204 * 5205 * If a read request arrives for a block that has a read in-progress, 5206 * either wait for the in-progress read to complete (and return the 5207 * results); or, if this is a read with a "done" func, add a record 5208 * to the read to invoke the "done" func when the read completes, 5209 * and return; or just return. 5210 * 5211 * arc_read_done() will invoke all the requested "done" functions 5212 * for readers of this block. 5213 */ 5214int 5215arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 5216 void *private, zio_priority_t priority, int zio_flags, 5217 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 5218{ 5219 arc_buf_hdr_t *hdr = NULL; 5220 kmutex_t *hash_lock = NULL; 5221 zio_t *rzio; 5222 uint64_t guid = spa_load_guid(spa); 5223 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0; 5224 5225 ASSERT(!BP_IS_EMBEDDED(bp) || 5226 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 5227 5228top: 5229 if (!BP_IS_EMBEDDED(bp)) { 5230 /* 5231 * Embedded BP's have no DVA and require no I/O to "read". 5232 * Create an anonymous arc buf to back it. 5233 */ 5234 hdr = buf_hash_find(guid, bp, &hash_lock); 5235 } 5236 5237 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) { 5238 arc_buf_t *buf = NULL; 5239 *arc_flags |= ARC_FLAG_CACHED; 5240 5241 if (HDR_IO_IN_PROGRESS(hdr)) { 5242 5243 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 5244 priority == ZIO_PRIORITY_SYNC_READ) { 5245 /* 5246 * This sync read must wait for an 5247 * in-progress async read (e.g. a predictive 5248 * prefetch). Async reads are queued 5249 * separately at the vdev_queue layer, so 5250 * this is a form of priority inversion. 5251 * Ideally, we would "inherit" the demand 5252 * i/o's priority by moving the i/o from 5253 * the async queue to the synchronous queue, 5254 * but there is currently no mechanism to do 5255 * so. Track this so that we can evaluate 5256 * the magnitude of this potential performance 5257 * problem. 5258 * 5259 * Note that if the prefetch i/o is already 5260 * active (has been issued to the device), 5261 * the prefetch improved performance, because 5262 * we issued it sooner than we would have 5263 * without the prefetch. 5264 */ 5265 DTRACE_PROBE1(arc__sync__wait__for__async, 5266 arc_buf_hdr_t *, hdr); 5267 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 5268 } 5269 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5270 arc_hdr_clear_flags(hdr, 5271 ARC_FLAG_PREDICTIVE_PREFETCH); 5272 } 5273 5274 if (*arc_flags & ARC_FLAG_WAIT) { 5275 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 5276 mutex_exit(hash_lock); 5277 goto top; 5278 } 5279 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5280 5281 if (done) { 5282 arc_callback_t *acb = NULL; 5283 5284 acb = kmem_zalloc(sizeof (arc_callback_t), 5285 KM_SLEEP); 5286 acb->acb_done = done; 5287 acb->acb_private = private; 5288 acb->acb_compressed = compressed_read; 5289 if (pio != NULL) 5290 acb->acb_zio_dummy = zio_null(pio, 5291 spa, NULL, NULL, NULL, zio_flags); 5292 5293 ASSERT3P(acb->acb_done, !=, NULL); 5294 acb->acb_next = hdr->b_l1hdr.b_acb; 5295 hdr->b_l1hdr.b_acb = acb; 5296 mutex_exit(hash_lock); 5297 return (0); 5298 } 5299 mutex_exit(hash_lock); 5300 return (0); 5301 } 5302 5303 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 5304 hdr->b_l1hdr.b_state == arc_mfu); 5305 5306 if (done) { 5307 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5308 /* 5309 * This is a demand read which does not have to 5310 * wait for i/o because we did a predictive 5311 * prefetch i/o for it, which has completed. 5312 */ 5313 DTRACE_PROBE1( 5314 arc__demand__hit__predictive__prefetch, 5315 arc_buf_hdr_t *, hdr); 5316 ARCSTAT_BUMP( 5317 arcstat_demand_hit_predictive_prefetch); 5318 arc_hdr_clear_flags(hdr, 5319 ARC_FLAG_PREDICTIVE_PREFETCH); 5320 } 5321 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); 5322 5323 /* Get a buf with the desired data in it. */ 5324 VERIFY0(arc_buf_alloc_impl(hdr, private, 5325 compressed_read, B_TRUE, &buf)); 5326 } else if (*arc_flags & ARC_FLAG_PREFETCH && 5327 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 5328 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5329 } 5330 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 5331 arc_access(hdr, hash_lock); 5332 if (*arc_flags & ARC_FLAG_L2CACHE) 5333 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5334 mutex_exit(hash_lock); 5335 ARCSTAT_BUMP(arcstat_hits); 5336 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5337 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5338 data, metadata, hits); 5339 5340 if (done) 5341 done(NULL, buf, private); 5342 } else { 5343 uint64_t lsize = BP_GET_LSIZE(bp); 5344 uint64_t psize = BP_GET_PSIZE(bp); 5345 arc_callback_t *acb; 5346 vdev_t *vd = NULL; 5347 uint64_t addr = 0; 5348 boolean_t devw = B_FALSE; 5349 uint64_t size; 5350 5351 if (hdr == NULL) { 5352 /* this block is not in the cache */ 5353 arc_buf_hdr_t *exists = NULL; 5354 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 5355 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 5356 BP_GET_COMPRESS(bp), type); 5357 5358 if (!BP_IS_EMBEDDED(bp)) { 5359 hdr->b_dva = *BP_IDENTITY(bp); 5360 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 5361 exists = buf_hash_insert(hdr, &hash_lock); 5362 } 5363 if (exists != NULL) { 5364 /* somebody beat us to the hash insert */ 5365 mutex_exit(hash_lock); 5366 buf_discard_identity(hdr); 5367 arc_hdr_destroy(hdr); 5368 goto top; /* restart the IO request */ 5369 } 5370 } else { 5371 /* 5372 * This block is in the ghost cache. If it was L2-only 5373 * (and thus didn't have an L1 hdr), we realloc the 5374 * header to add an L1 hdr. 5375 */ 5376 if (!HDR_HAS_L1HDR(hdr)) { 5377 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 5378 hdr_full_cache); 5379 } 5380 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5381 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 5382 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5383 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5384 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 5385 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 5386 5387 /* 5388 * This is a delicate dance that we play here. 5389 * This hdr is in the ghost list so we access it 5390 * to move it out of the ghost list before we 5391 * initiate the read. If it's a prefetch then 5392 * it won't have a callback so we'll remove the 5393 * reference that arc_buf_alloc_impl() created. We 5394 * do this after we've called arc_access() to 5395 * avoid hitting an assert in remove_reference(). 5396 */ 5397 arc_access(hdr, hash_lock); 5398 arc_hdr_alloc_pabd(hdr); 5399 } 5400 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5401 size = arc_hdr_size(hdr); 5402 5403 /* 5404 * If compression is enabled on the hdr, then will do 5405 * RAW I/O and will store the compressed data in the hdr's 5406 * data block. Otherwise, the hdr's data block will contain 5407 * the uncompressed data. 5408 */ 5409 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5410 zio_flags |= ZIO_FLAG_RAW; 5411 } 5412 5413 if (*arc_flags & ARC_FLAG_PREFETCH) 5414 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5415 if (*arc_flags & ARC_FLAG_L2CACHE) 5416 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5417 if (BP_GET_LEVEL(bp) > 0) 5418 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); 5419 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 5420 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); 5421 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 5422 5423 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 5424 acb->acb_done = done; 5425 acb->acb_private = private; 5426 acb->acb_compressed = compressed_read; 5427 5428 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5429 hdr->b_l1hdr.b_acb = acb; 5430 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5431 5432 if (HDR_HAS_L2HDR(hdr) && 5433 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 5434 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 5435 addr = hdr->b_l2hdr.b_daddr; 5436 /* 5437 * Lock out L2ARC device removal. 5438 */ 5439 if (vdev_is_dead(vd) || 5440 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 5441 vd = NULL; 5442 } 5443 5444 if (priority == ZIO_PRIORITY_ASYNC_READ) 5445 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5446 else 5447 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5448 5449 if (hash_lock != NULL) 5450 mutex_exit(hash_lock); 5451 5452 /* 5453 * At this point, we have a level 1 cache miss. Try again in 5454 * L2ARC if possible. 5455 */ 5456 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); 5457 5458 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 5459 uint64_t, lsize, zbookmark_phys_t *, zb); 5460 ARCSTAT_BUMP(arcstat_misses); 5461 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5462 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5463 data, metadata, misses); 5464#ifdef _KERNEL 5465#ifdef RACCT 5466 if (racct_enable) { 5467 PROC_LOCK(curproc); 5468 racct_add_force(curproc, RACCT_READBPS, size); 5469 racct_add_force(curproc, RACCT_READIOPS, 1); 5470 PROC_UNLOCK(curproc); 5471 } 5472#endif /* RACCT */ 5473 curthread->td_ru.ru_inblock++; 5474#endif 5475 5476 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 5477 /* 5478 * Read from the L2ARC if the following are true: 5479 * 1. The L2ARC vdev was previously cached. 5480 * 2. This buffer still has L2ARC metadata. 5481 * 3. This buffer isn't currently writing to the L2ARC. 5482 * 4. The L2ARC entry wasn't evicted, which may 5483 * also have invalidated the vdev. 5484 * 5. This isn't prefetch and l2arc_noprefetch is set. 5485 */ 5486 if (HDR_HAS_L2HDR(hdr) && 5487 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 5488 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 5489 l2arc_read_callback_t *cb; 5490 abd_t *abd; 5491 uint64_t asize; 5492 5493 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 5494 ARCSTAT_BUMP(arcstat_l2_hits); 5495 5496 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 5497 KM_SLEEP); 5498 cb->l2rcb_hdr = hdr; 5499 cb->l2rcb_bp = *bp; 5500 cb->l2rcb_zb = *zb; 5501 cb->l2rcb_flags = zio_flags; 5502 5503 asize = vdev_psize_to_asize(vd, size); 5504 if (asize != size) { 5505 abd = abd_alloc_for_io(asize, 5506 HDR_ISTYPE_METADATA(hdr)); 5507 cb->l2rcb_abd = abd; 5508 } else { 5509 abd = hdr->b_l1hdr.b_pabd; 5510 } 5511 5512 ASSERT(addr >= VDEV_LABEL_START_SIZE && 5513 addr + asize <= vd->vdev_psize - 5514 VDEV_LABEL_END_SIZE); 5515 5516 /* 5517 * l2arc read. The SCL_L2ARC lock will be 5518 * released by l2arc_read_done(). 5519 * Issue a null zio if the underlying buffer 5520 * was squashed to zero size by compression. 5521 */ 5522 ASSERT3U(HDR_GET_COMPRESS(hdr), !=, 5523 ZIO_COMPRESS_EMPTY); 5524 rzio = zio_read_phys(pio, vd, addr, 5525 asize, abd, 5526 ZIO_CHECKSUM_OFF, 5527 l2arc_read_done, cb, priority, 5528 zio_flags | ZIO_FLAG_DONT_CACHE | 5529 ZIO_FLAG_CANFAIL | 5530 ZIO_FLAG_DONT_PROPAGATE | 5531 ZIO_FLAG_DONT_RETRY, B_FALSE); 5532 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 5533 zio_t *, rzio); 5534 ARCSTAT_INCR(arcstat_l2_read_bytes, size); 5535 5536 if (*arc_flags & ARC_FLAG_NOWAIT) { 5537 zio_nowait(rzio); 5538 return (0); 5539 } 5540 5541 ASSERT(*arc_flags & ARC_FLAG_WAIT); 5542 if (zio_wait(rzio) == 0) 5543 return (0); 5544 5545 /* l2arc read error; goto zio_read() */ 5546 } else { 5547 DTRACE_PROBE1(l2arc__miss, 5548 arc_buf_hdr_t *, hdr); 5549 ARCSTAT_BUMP(arcstat_l2_misses); 5550 if (HDR_L2_WRITING(hdr)) 5551 ARCSTAT_BUMP(arcstat_l2_rw_clash); 5552 spa_config_exit(spa, SCL_L2ARC, vd); 5553 } 5554 } else { 5555 if (vd != NULL) 5556 spa_config_exit(spa, SCL_L2ARC, vd); 5557 if (l2arc_ndev != 0) { 5558 DTRACE_PROBE1(l2arc__miss, 5559 arc_buf_hdr_t *, hdr); 5560 ARCSTAT_BUMP(arcstat_l2_misses); 5561 } 5562 } 5563 5564 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size, 5565 arc_read_done, hdr, priority, zio_flags, zb); 5566 5567 if (*arc_flags & ARC_FLAG_WAIT) 5568 return (zio_wait(rzio)); 5569 5570 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5571 zio_nowait(rzio); 5572 } 5573 return (0); 5574} 5575 5576/* 5577 * Notify the arc that a block was freed, and thus will never be used again. 5578 */ 5579void 5580arc_freed(spa_t *spa, const blkptr_t *bp) 5581{ 5582 arc_buf_hdr_t *hdr; 5583 kmutex_t *hash_lock; 5584 uint64_t guid = spa_load_guid(spa); 5585 5586 ASSERT(!BP_IS_EMBEDDED(bp)); 5587 5588 hdr = buf_hash_find(guid, bp, &hash_lock); 5589 if (hdr == NULL) 5590 return; 5591 5592 /* 5593 * We might be trying to free a block that is still doing I/O 5594 * (i.e. prefetch) or has a reference (i.e. a dedup-ed, 5595 * dmu_sync-ed block). If this block is being prefetched, then it 5596 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr 5597 * until the I/O completes. A block may also have a reference if it is 5598 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would 5599 * have written the new block to its final resting place on disk but 5600 * without the dedup flag set. This would have left the hdr in the MRU 5601 * state and discoverable. When the txg finally syncs it detects that 5602 * the block was overridden in open context and issues an override I/O. 5603 * Since this is a dedup block, the override I/O will determine if the 5604 * block is already in the DDT. If so, then it will replace the io_bp 5605 * with the bp from the DDT and allow the I/O to finish. When the I/O 5606 * reaches the done callback, dbuf_write_override_done, it will 5607 * check to see if the io_bp and io_bp_override are identical. 5608 * If they are not, then it indicates that the bp was replaced with 5609 * the bp in the DDT and the override bp is freed. This allows 5610 * us to arrive here with a reference on a block that is being 5611 * freed. So if we have an I/O in progress, or a reference to 5612 * this hdr, then we don't destroy the hdr. 5613 */ 5614 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && 5615 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { 5616 arc_change_state(arc_anon, hdr, hash_lock); 5617 arc_hdr_destroy(hdr); 5618 mutex_exit(hash_lock); 5619 } else { 5620 mutex_exit(hash_lock); 5621 } 5622 5623} 5624 5625/* 5626 * Release this buffer from the cache, making it an anonymous buffer. This 5627 * must be done after a read and prior to modifying the buffer contents. 5628 * If the buffer has more than one reference, we must make 5629 * a new hdr for the buffer. 5630 */ 5631void 5632arc_release(arc_buf_t *buf, void *tag) 5633{ 5634 arc_buf_hdr_t *hdr = buf->b_hdr; 5635 5636 /* 5637 * It would be nice to assert that if it's DMU metadata (level > 5638 * 0 || it's the dnode file), then it must be syncing context. 5639 * But we don't know that information at this level. 5640 */ 5641 5642 mutex_enter(&buf->b_evict_lock); 5643 5644 ASSERT(HDR_HAS_L1HDR(hdr)); 5645 5646 /* 5647 * We don't grab the hash lock prior to this check, because if 5648 * the buffer's header is in the arc_anon state, it won't be 5649 * linked into the hash table. 5650 */ 5651 if (hdr->b_l1hdr.b_state == arc_anon) { 5652 mutex_exit(&buf->b_evict_lock); 5653 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5654 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 5655 ASSERT(!HDR_HAS_L2HDR(hdr)); 5656 ASSERT(HDR_EMPTY(hdr)); 5657 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5658 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 5659 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 5660 5661 hdr->b_l1hdr.b_arc_access = 0; 5662 5663 /* 5664 * If the buf is being overridden then it may already 5665 * have a hdr that is not empty. 5666 */ 5667 buf_discard_identity(hdr); 5668 arc_buf_thaw(buf); 5669 5670 return; 5671 } 5672 5673 kmutex_t *hash_lock = HDR_LOCK(hdr); 5674 mutex_enter(hash_lock); 5675 5676 /* 5677 * This assignment is only valid as long as the hash_lock is 5678 * held, we must be careful not to reference state or the 5679 * b_state field after dropping the lock. 5680 */ 5681 arc_state_t *state = hdr->b_l1hdr.b_state; 5682 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5683 ASSERT3P(state, !=, arc_anon); 5684 5685 /* this buffer is not on any list */ 5686 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0); 5687 5688 if (HDR_HAS_L2HDR(hdr)) { 5689 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5690 5691 /* 5692 * We have to recheck this conditional again now that 5693 * we're holding the l2ad_mtx to prevent a race with 5694 * another thread which might be concurrently calling 5695 * l2arc_evict(). In that case, l2arc_evict() might have 5696 * destroyed the header's L2 portion as we were waiting 5697 * to acquire the l2ad_mtx. 5698 */ 5699 if (HDR_HAS_L2HDR(hdr)) { 5700 l2arc_trim(hdr); 5701 arc_hdr_l2hdr_destroy(hdr); 5702 } 5703 5704 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5705 } 5706 5707 /* 5708 * Do we have more than one buf? 5709 */ 5710 if (hdr->b_l1hdr.b_bufcnt > 1) { 5711 arc_buf_hdr_t *nhdr; 5712 uint64_t spa = hdr->b_spa; 5713 uint64_t psize = HDR_GET_PSIZE(hdr); 5714 uint64_t lsize = HDR_GET_LSIZE(hdr); 5715 enum zio_compress compress = HDR_GET_COMPRESS(hdr); 5716 arc_buf_contents_t type = arc_buf_type(hdr); 5717 VERIFY3U(hdr->b_type, ==, type); 5718 5719 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 5720 (void) remove_reference(hdr, hash_lock, tag); 5721 5722 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) { 5723 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5724 ASSERT(ARC_BUF_LAST(buf)); 5725 } 5726 5727 /* 5728 * Pull the data off of this hdr and attach it to 5729 * a new anonymous hdr. Also find the last buffer 5730 * in the hdr's buffer list. 5731 */ 5732 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 5733 ASSERT3P(lastbuf, !=, NULL); 5734 5735 /* 5736 * If the current arc_buf_t and the hdr are sharing their data 5737 * buffer, then we must stop sharing that block. 5738 */ 5739 if (arc_buf_is_shared(buf)) { 5740 VERIFY(!arc_buf_is_shared(lastbuf)); 5741 5742 /* 5743 * First, sever the block sharing relationship between 5744 * buf and the arc_buf_hdr_t. 5745 */ 5746 arc_unshare_buf(hdr, buf); 5747 5748 /* 5749 * Now we need to recreate the hdr's b_pabd. Since we 5750 * have lastbuf handy, we try to share with it, but if 5751 * we can't then we allocate a new b_pabd and copy the 5752 * data from buf into it. 5753 */ 5754 if (arc_can_share(hdr, lastbuf)) { 5755 arc_share_buf(hdr, lastbuf); 5756 } else { 5757 arc_hdr_alloc_pabd(hdr); 5758 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, 5759 buf->b_data, psize); 5760 } 5761 VERIFY3P(lastbuf->b_data, !=, NULL); 5762 } else if (HDR_SHARED_DATA(hdr)) { 5763 /* 5764 * Uncompressed shared buffers are always at the end 5765 * of the list. Compressed buffers don't have the 5766 * same requirements. This makes it hard to 5767 * simply assert that the lastbuf is shared so 5768 * we rely on the hdr's compression flags to determine 5769 * if we have a compressed, shared buffer. 5770 */ 5771 ASSERT(arc_buf_is_shared(lastbuf) || 5772 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 5773 ASSERT(!ARC_BUF_SHARED(buf)); 5774 } 5775 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5776 ASSERT3P(state, !=, arc_l2c_only); 5777 5778 (void) refcount_remove_many(&state->arcs_size, 5779 arc_buf_size(buf), buf); 5780 5781 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 5782 ASSERT3P(state, !=, arc_l2c_only); 5783 (void) refcount_remove_many(&state->arcs_esize[type], 5784 arc_buf_size(buf), buf); 5785 } 5786 5787 hdr->b_l1hdr.b_bufcnt -= 1; 5788 arc_cksum_verify(buf); 5789#ifdef illumos 5790 arc_buf_unwatch(buf); 5791#endif 5792 5793 mutex_exit(hash_lock); 5794 5795 /* 5796 * Allocate a new hdr. The new hdr will contain a b_pabd 5797 * buffer which will be freed in arc_write(). 5798 */ 5799 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type); 5800 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); 5801 ASSERT0(nhdr->b_l1hdr.b_bufcnt); 5802 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt)); 5803 VERIFY3U(nhdr->b_type, ==, type); 5804 ASSERT(!HDR_SHARED_DATA(nhdr)); 5805 5806 nhdr->b_l1hdr.b_buf = buf; 5807 nhdr->b_l1hdr.b_bufcnt = 1; 5808 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 5809 buf->b_hdr = nhdr; 5810 5811 mutex_exit(&buf->b_evict_lock); 5812 (void) refcount_add_many(&arc_anon->arcs_size, 5813 arc_buf_size(buf), buf); 5814 } else { 5815 mutex_exit(&buf->b_evict_lock); 5816 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 5817 /* protected by hash lock, or hdr is on arc_anon */ 5818 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5819 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5820 arc_change_state(arc_anon, hdr, hash_lock); 5821 hdr->b_l1hdr.b_arc_access = 0; 5822 mutex_exit(hash_lock); 5823 5824 buf_discard_identity(hdr); 5825 arc_buf_thaw(buf); 5826 } 5827} 5828 5829int 5830arc_released(arc_buf_t *buf) 5831{ 5832 int released; 5833 5834 mutex_enter(&buf->b_evict_lock); 5835 released = (buf->b_data != NULL && 5836 buf->b_hdr->b_l1hdr.b_state == arc_anon); 5837 mutex_exit(&buf->b_evict_lock); 5838 return (released); 5839} 5840 5841#ifdef ZFS_DEBUG 5842int 5843arc_referenced(arc_buf_t *buf) 5844{ 5845 int referenced; 5846 5847 mutex_enter(&buf->b_evict_lock); 5848 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 5849 mutex_exit(&buf->b_evict_lock); 5850 return (referenced); 5851} 5852#endif 5853 5854static void 5855arc_write_ready(zio_t *zio) 5856{ 5857 arc_write_callback_t *callback = zio->io_private; 5858 arc_buf_t *buf = callback->awcb_buf; 5859 arc_buf_hdr_t *hdr = buf->b_hdr; 5860 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp); 5861 5862 ASSERT(HDR_HAS_L1HDR(hdr)); 5863 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 5864 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 5865 5866 /* 5867 * If we're reexecuting this zio because the pool suspended, then 5868 * cleanup any state that was previously set the first time the 5869 * callback was invoked. 5870 */ 5871 if (zio->io_flags & ZIO_FLAG_REEXECUTED) { 5872 arc_cksum_free(hdr); 5873#ifdef illumos 5874 arc_buf_unwatch(buf); 5875#endif 5876 if (hdr->b_l1hdr.b_pabd != NULL) { 5877 if (arc_buf_is_shared(buf)) { 5878 arc_unshare_buf(hdr, buf); 5879 } else { 5880 arc_hdr_free_pabd(hdr); 5881 } 5882 } 5883 } 5884 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5885 ASSERT(!HDR_SHARED_DATA(hdr)); 5886 ASSERT(!arc_buf_is_shared(buf)); 5887 5888 callback->awcb_ready(zio, buf, callback->awcb_private); 5889 5890 if (HDR_IO_IN_PROGRESS(hdr)) 5891 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); 5892 5893 arc_cksum_compute(buf); 5894 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5895 5896 enum zio_compress compress; 5897 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5898 compress = ZIO_COMPRESS_OFF; 5899 } else { 5900 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp)); 5901 compress = BP_GET_COMPRESS(zio->io_bp); 5902 } 5903 HDR_SET_PSIZE(hdr, psize); 5904 arc_hdr_set_compress(hdr, compress); 5905 5906 5907 /* 5908 * Fill the hdr with data. If the hdr is compressed, the data we want 5909 * is available from the zio, otherwise we can take it from the buf. 5910 * 5911 * We might be able to share the buf's data with the hdr here. However, 5912 * doing so would cause the ARC to be full of linear ABDs if we write a 5913 * lot of shareable data. As a compromise, we check whether scattered 5914 * ABDs are allowed, and assume that if they are then the user wants 5915 * the ARC to be primarily filled with them regardless of the data being 5916 * written. Therefore, if they're allowed then we allocate one and copy 5917 * the data into it; otherwise, we share the data directly if we can. 5918 */ 5919 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) { 5920 arc_hdr_alloc_pabd(hdr); 5921 5922 /* 5923 * Ideally, we would always copy the io_abd into b_pabd, but the 5924 * user may have disabled compressed ARC, thus we must check the 5925 * hdr's compression setting rather than the io_bp's. 5926 */ 5927 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5928 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=, 5929 ZIO_COMPRESS_OFF); 5930 ASSERT3U(psize, >, 0); 5931 5932 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize); 5933 } else { 5934 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr)); 5935 5936 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data, 5937 arc_buf_size(buf)); 5938 } 5939 } else { 5940 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd)); 5941 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf)); 5942 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5943 5944 arc_share_buf(hdr, buf); 5945 } 5946 5947 arc_hdr_verify(hdr, zio->io_bp); 5948} 5949 5950static void 5951arc_write_children_ready(zio_t *zio) 5952{ 5953 arc_write_callback_t *callback = zio->io_private; 5954 arc_buf_t *buf = callback->awcb_buf; 5955 5956 callback->awcb_children_ready(zio, buf, callback->awcb_private); 5957} 5958 5959/* 5960 * The SPA calls this callback for each physical write that happens on behalf 5961 * of a logical write. See the comment in dbuf_write_physdone() for details. 5962 */ 5963static void 5964arc_write_physdone(zio_t *zio) 5965{ 5966 arc_write_callback_t *cb = zio->io_private; 5967 if (cb->awcb_physdone != NULL) 5968 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 5969} 5970 5971static void 5972arc_write_done(zio_t *zio) 5973{ 5974 arc_write_callback_t *callback = zio->io_private; 5975 arc_buf_t *buf = callback->awcb_buf; 5976 arc_buf_hdr_t *hdr = buf->b_hdr; 5977 5978 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5979 5980 if (zio->io_error == 0) { 5981 arc_hdr_verify(hdr, zio->io_bp); 5982 5983 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5984 buf_discard_identity(hdr); 5985 } else { 5986 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 5987 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 5988 } 5989 } else { 5990 ASSERT(HDR_EMPTY(hdr)); 5991 } 5992 5993 /* 5994 * If the block to be written was all-zero or compressed enough to be 5995 * embedded in the BP, no write was performed so there will be no 5996 * dva/birth/checksum. The buffer must therefore remain anonymous 5997 * (and uncached). 5998 */ 5999 if (!HDR_EMPTY(hdr)) { 6000 arc_buf_hdr_t *exists; 6001 kmutex_t *hash_lock; 6002 6003 ASSERT3U(zio->io_error, ==, 0); 6004 6005 arc_cksum_verify(buf); 6006 6007 exists = buf_hash_insert(hdr, &hash_lock); 6008 if (exists != NULL) { 6009 /* 6010 * This can only happen if we overwrite for 6011 * sync-to-convergence, because we remove 6012 * buffers from the hash table when we arc_free(). 6013 */ 6014 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 6015 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 6016 panic("bad overwrite, hdr=%p exists=%p", 6017 (void *)hdr, (void *)exists); 6018 ASSERT(refcount_is_zero( 6019 &exists->b_l1hdr.b_refcnt)); 6020 arc_change_state(arc_anon, exists, hash_lock); 6021 mutex_exit(hash_lock); 6022 arc_hdr_destroy(exists); 6023 exists = buf_hash_insert(hdr, &hash_lock); 6024 ASSERT3P(exists, ==, NULL); 6025 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 6026 /* nopwrite */ 6027 ASSERT(zio->io_prop.zp_nopwrite); 6028 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 6029 panic("bad nopwrite, hdr=%p exists=%p", 6030 (void *)hdr, (void *)exists); 6031 } else { 6032 /* Dedup */ 6033 ASSERT(hdr->b_l1hdr.b_bufcnt == 1); 6034 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 6035 ASSERT(BP_GET_DEDUP(zio->io_bp)); 6036 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 6037 } 6038 } 6039 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 6040 /* if it's not anon, we are doing a scrub */ 6041 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 6042 arc_access(hdr, hash_lock); 6043 mutex_exit(hash_lock); 6044 } else { 6045 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 6046 } 6047 6048 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 6049 callback->awcb_done(zio, buf, callback->awcb_private); 6050 6051 abd_put(zio->io_abd); 6052 kmem_free(callback, sizeof (arc_write_callback_t)); 6053} 6054 6055zio_t * 6056arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 6057 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready, 6058 arc_done_func_t *children_ready, arc_done_func_t *physdone, 6059 arc_done_func_t *done, void *private, zio_priority_t priority, 6060 int zio_flags, const zbookmark_phys_t *zb) 6061{ 6062 arc_buf_hdr_t *hdr = buf->b_hdr; 6063 arc_write_callback_t *callback; 6064 zio_t *zio; 6065 zio_prop_t localprop = *zp; 6066 6067 ASSERT3P(ready, !=, NULL); 6068 ASSERT3P(done, !=, NULL); 6069 ASSERT(!HDR_IO_ERROR(hdr)); 6070 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 6071 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 6072 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 6073 if (l2arc) 6074 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 6075 if (ARC_BUF_COMPRESSED(buf)) { 6076 /* 6077 * We're writing a pre-compressed buffer. Make the 6078 * compression algorithm requested by the zio_prop_t match 6079 * the pre-compressed buffer's compression algorithm. 6080 */ 6081 localprop.zp_compress = HDR_GET_COMPRESS(hdr); 6082 6083 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf)); 6084 zio_flags |= ZIO_FLAG_RAW; 6085 } 6086 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 6087 callback->awcb_ready = ready; 6088 callback->awcb_children_ready = children_ready; 6089 callback->awcb_physdone = physdone; 6090 callback->awcb_done = done; 6091 callback->awcb_private = private; 6092 callback->awcb_buf = buf; 6093 6094 /* 6095 * The hdr's b_pabd is now stale, free it now. A new data block 6096 * will be allocated when the zio pipeline calls arc_write_ready(). 6097 */ 6098 if (hdr->b_l1hdr.b_pabd != NULL) { 6099 /* 6100 * If the buf is currently sharing the data block with 6101 * the hdr then we need to break that relationship here. 6102 * The hdr will remain with a NULL data pointer and the 6103 * buf will take sole ownership of the block. 6104 */ 6105 if (arc_buf_is_shared(buf)) { 6106 arc_unshare_buf(hdr, buf); 6107 } else { 6108 arc_hdr_free_pabd(hdr); 6109 } 6110 VERIFY3P(buf->b_data, !=, NULL); 6111 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); 6112 } 6113 ASSERT(!arc_buf_is_shared(buf)); 6114 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 6115 6116 zio = zio_write(pio, spa, txg, bp, 6117 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)), 6118 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready, 6119 (children_ready != NULL) ? arc_write_children_ready : NULL, 6120 arc_write_physdone, arc_write_done, callback, 6121 priority, zio_flags, zb); 6122 6123 return (zio); 6124} 6125 6126static int 6127arc_memory_throttle(uint64_t reserve, uint64_t txg) 6128{ 6129#ifdef _KERNEL 6130 uint64_t available_memory = ptob(freemem); 6131 static uint64_t page_load = 0; 6132 static uint64_t last_txg = 0; 6133 6134#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 6135 available_memory = 6136 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 6137#endif 6138 6139 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 6140 return (0); 6141 6142 if (txg > last_txg) { 6143 last_txg = txg; 6144 page_load = 0; 6145 } 6146 /* 6147 * If we are in pageout, we know that memory is already tight, 6148 * the arc is already going to be evicting, so we just want to 6149 * continue to let page writes occur as quickly as possible. 6150 */ 6151 if (curproc == pageproc) { 6152 if (page_load > MAX(ptob(minfree), available_memory) / 4) 6153 return (SET_ERROR(ERESTART)); 6154 /* Note: reserve is inflated, so we deflate */ 6155 page_load += reserve / 8; 6156 return (0); 6157 } else if (page_load > 0 && arc_reclaim_needed()) { 6158 /* memory is low, delay before restarting */ 6159 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 6160 return (SET_ERROR(EAGAIN)); 6161 } 6162 page_load = 0; 6163#endif 6164 return (0); 6165} 6166 6167void 6168arc_tempreserve_clear(uint64_t reserve) 6169{ 6170 atomic_add_64(&arc_tempreserve, -reserve); 6171 ASSERT((int64_t)arc_tempreserve >= 0); 6172} 6173 6174int 6175arc_tempreserve_space(uint64_t reserve, uint64_t txg) 6176{ 6177 int error; 6178 uint64_t anon_size; 6179 6180 if (reserve > arc_c/4 && !arc_no_grow) { 6181 arc_c = MIN(arc_c_max, reserve * 4); 6182 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 6183 } 6184 if (reserve > arc_c) 6185 return (SET_ERROR(ENOMEM)); 6186 6187 /* 6188 * Don't count loaned bufs as in flight dirty data to prevent long 6189 * network delays from blocking transactions that are ready to be 6190 * assigned to a txg. 6191 */ 6192 6193 /* assert that it has not wrapped around */ 6194 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 6195 6196 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 6197 arc_loaned_bytes), 0); 6198 6199 /* 6200 * Writes will, almost always, require additional memory allocations 6201 * in order to compress/encrypt/etc the data. We therefore need to 6202 * make sure that there is sufficient available memory for this. 6203 */ 6204 error = arc_memory_throttle(reserve, txg); 6205 if (error != 0) 6206 return (error); 6207 6208 /* 6209 * Throttle writes when the amount of dirty data in the cache 6210 * gets too large. We try to keep the cache less than half full 6211 * of dirty blocks so that our sync times don't grow too large. 6212 * Note: if two requests come in concurrently, we might let them 6213 * both succeed, when one of them should fail. Not a huge deal. 6214 */ 6215 6216 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 6217 anon_size > arc_c / 4) { 6218 uint64_t meta_esize = 6219 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6220 uint64_t data_esize = 6221 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6222 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 6223 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 6224 arc_tempreserve >> 10, meta_esize >> 10, 6225 data_esize >> 10, reserve >> 10, arc_c >> 10); 6226 return (SET_ERROR(ERESTART)); 6227 } 6228 atomic_add_64(&arc_tempreserve, reserve); 6229 return (0); 6230} 6231 6232static void 6233arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 6234 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 6235{ 6236 size->value.ui64 = refcount_count(&state->arcs_size); 6237 evict_data->value.ui64 = 6238 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); 6239 evict_metadata->value.ui64 = 6240 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); 6241} 6242 6243static int 6244arc_kstat_update(kstat_t *ksp, int rw) 6245{ 6246 arc_stats_t *as = ksp->ks_data; 6247 6248 if (rw == KSTAT_WRITE) { 6249 return (EACCES); 6250 } else { 6251 arc_kstat_update_state(arc_anon, 6252 &as->arcstat_anon_size, 6253 &as->arcstat_anon_evictable_data, 6254 &as->arcstat_anon_evictable_metadata); 6255 arc_kstat_update_state(arc_mru, 6256 &as->arcstat_mru_size, 6257 &as->arcstat_mru_evictable_data, 6258 &as->arcstat_mru_evictable_metadata); 6259 arc_kstat_update_state(arc_mru_ghost, 6260 &as->arcstat_mru_ghost_size, 6261 &as->arcstat_mru_ghost_evictable_data, 6262 &as->arcstat_mru_ghost_evictable_metadata); 6263 arc_kstat_update_state(arc_mfu, 6264 &as->arcstat_mfu_size, 6265 &as->arcstat_mfu_evictable_data, 6266 &as->arcstat_mfu_evictable_metadata); 6267 arc_kstat_update_state(arc_mfu_ghost, 6268 &as->arcstat_mfu_ghost_size, 6269 &as->arcstat_mfu_ghost_evictable_data, 6270 &as->arcstat_mfu_ghost_evictable_metadata); 6271 } 6272 6273 return (0); 6274} 6275 6276/* 6277 * This function *must* return indices evenly distributed between all 6278 * sublists of the multilist. This is needed due to how the ARC eviction 6279 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 6280 * distributed between all sublists and uses this assumption when 6281 * deciding which sublist to evict from and how much to evict from it. 6282 */ 6283unsigned int 6284arc_state_multilist_index_func(multilist_t *ml, void *obj) 6285{ 6286 arc_buf_hdr_t *hdr = obj; 6287 6288 /* 6289 * We rely on b_dva to generate evenly distributed index 6290 * numbers using buf_hash below. So, as an added precaution, 6291 * let's make sure we never add empty buffers to the arc lists. 6292 */ 6293 ASSERT(!HDR_EMPTY(hdr)); 6294 6295 /* 6296 * The assumption here, is the hash value for a given 6297 * arc_buf_hdr_t will remain constant throughout it's lifetime 6298 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 6299 * Thus, we don't need to store the header's sublist index 6300 * on insertion, as this index can be recalculated on removal. 6301 * 6302 * Also, the low order bits of the hash value are thought to be 6303 * distributed evenly. Otherwise, in the case that the multilist 6304 * has a power of two number of sublists, each sublists' usage 6305 * would not be evenly distributed. 6306 */ 6307 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 6308 multilist_get_num_sublists(ml)); 6309} 6310 6311#ifdef _KERNEL 6312static eventhandler_tag arc_event_lowmem = NULL; 6313 6314static void 6315arc_lowmem(void *arg __unused, int howto __unused) 6316{ 6317 6318 mutex_enter(&arc_reclaim_lock); 6319 DTRACE_PROBE1(arc__needfree, int64_t, ((int64_t)freemem - zfs_arc_free_target) * PAGESIZE); 6320 cv_signal(&arc_reclaim_thread_cv); 6321 6322 /* 6323 * It is unsafe to block here in arbitrary threads, because we can come 6324 * here from ARC itself and may hold ARC locks and thus risk a deadlock 6325 * with ARC reclaim thread. 6326 */ 6327 if (curproc == pageproc) 6328 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 6329 mutex_exit(&arc_reclaim_lock); 6330} 6331#endif 6332 6333static void 6334arc_state_init(void) 6335{ 6336 arc_anon = &ARC_anon; 6337 arc_mru = &ARC_mru; 6338 arc_mru_ghost = &ARC_mru_ghost; 6339 arc_mfu = &ARC_mfu; 6340 arc_mfu_ghost = &ARC_mfu_ghost; 6341 arc_l2c_only = &ARC_l2c_only; 6342 6343 arc_mru->arcs_list[ARC_BUFC_METADATA] = 6344 multilist_create(sizeof (arc_buf_hdr_t), 6345 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6346 arc_state_multilist_index_func); 6347 arc_mru->arcs_list[ARC_BUFC_DATA] = 6348 multilist_create(sizeof (arc_buf_hdr_t), 6349 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6350 arc_state_multilist_index_func); 6351 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] = 6352 multilist_create(sizeof (arc_buf_hdr_t), 6353 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6354 arc_state_multilist_index_func); 6355 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] = 6356 multilist_create(sizeof (arc_buf_hdr_t), 6357 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6358 arc_state_multilist_index_func); 6359 arc_mfu->arcs_list[ARC_BUFC_METADATA] = 6360 multilist_create(sizeof (arc_buf_hdr_t), 6361 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6362 arc_state_multilist_index_func); 6363 arc_mfu->arcs_list[ARC_BUFC_DATA] = 6364 multilist_create(sizeof (arc_buf_hdr_t), 6365 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6366 arc_state_multilist_index_func); 6367 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] = 6368 multilist_create(sizeof (arc_buf_hdr_t), 6369 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6370 arc_state_multilist_index_func); 6371 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] = 6372 multilist_create(sizeof (arc_buf_hdr_t), 6373 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6374 arc_state_multilist_index_func); 6375 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] = 6376 multilist_create(sizeof (arc_buf_hdr_t), 6377 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6378 arc_state_multilist_index_func); 6379 arc_l2c_only->arcs_list[ARC_BUFC_DATA] = 6380 multilist_create(sizeof (arc_buf_hdr_t), 6381 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6382 arc_state_multilist_index_func); 6383 6384 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6385 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6386 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6387 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6388 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6389 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6390 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6391 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6392 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6393 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6394 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6395 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6396 6397 refcount_create(&arc_anon->arcs_size); 6398 refcount_create(&arc_mru->arcs_size); 6399 refcount_create(&arc_mru_ghost->arcs_size); 6400 refcount_create(&arc_mfu->arcs_size); 6401 refcount_create(&arc_mfu_ghost->arcs_size); 6402 refcount_create(&arc_l2c_only->arcs_size); 6403} 6404 6405static void 6406arc_state_fini(void) 6407{ 6408 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6409 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6410 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6411 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6412 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6413 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6414 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6415 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6416 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6417 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6418 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6419 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6420 6421 refcount_destroy(&arc_anon->arcs_size); 6422 refcount_destroy(&arc_mru->arcs_size); 6423 refcount_destroy(&arc_mru_ghost->arcs_size); 6424 refcount_destroy(&arc_mfu->arcs_size); 6425 refcount_destroy(&arc_mfu_ghost->arcs_size); 6426 refcount_destroy(&arc_l2c_only->arcs_size); 6427 6428 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]); 6429 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 6430 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]); 6431 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 6432 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]); 6433 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 6434 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]); 6435 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 6436} 6437 6438uint64_t 6439arc_max_bytes(void) 6440{ 6441 return (arc_c_max); 6442} 6443 6444void 6445arc_init(void) 6446{ 6447 int i, prefetch_tunable_set = 0; 6448 6449 /* 6450 * allmem is "all memory that we could possibly use". 6451 */ 6452#ifdef illumos 6453#ifdef _KERNEL 6454 uint64_t allmem = ptob(physmem - swapfs_minfree); 6455#else 6456 uint64_t allmem = (physmem * PAGESIZE) / 2; 6457#endif 6458#else 6459 uint64_t allmem = kmem_size(); 6460#endif 6461 6462 6463 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 6464 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 6465 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 6466 6467 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 6468 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL); 6469 6470 /* Convert seconds to clock ticks */ 6471 arc_min_prefetch_lifespan = 1 * hz; 6472 6473 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */ 6474 arc_c_min = MAX(allmem / 32, arc_abs_min); 6475 /* set max to 5/8 of all memory, or all but 1GB, whichever is more */ 6476 if (allmem >= 1 << 30) 6477 arc_c_max = allmem - (1 << 30); 6478 else 6479 arc_c_max = arc_c_min; 6480 arc_c_max = MAX(allmem * 5 / 8, arc_c_max); 6481 6482 /* 6483 * In userland, there's only the memory pressure that we artificially 6484 * create (see arc_available_memory()). Don't let arc_c get too 6485 * small, because it can cause transactions to be larger than 6486 * arc_c, causing arc_tempreserve_space() to fail. 6487 */ 6488#ifndef _KERNEL 6489 arc_c_min = arc_c_max / 2; 6490#endif 6491 6492#ifdef _KERNEL 6493 /* 6494 * Allow the tunables to override our calculations if they are 6495 * reasonable. 6496 */ 6497 if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) { 6498 arc_c_max = zfs_arc_max; 6499 arc_c_min = MIN(arc_c_min, arc_c_max); 6500 } 6501 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max) 6502 arc_c_min = zfs_arc_min; 6503#endif 6504 6505 arc_c = arc_c_max; 6506 arc_p = (arc_c >> 1); 6507 arc_size = 0; 6508 6509 /* limit meta-data to 1/4 of the arc capacity */ 6510 arc_meta_limit = arc_c_max / 4; 6511 6512#ifdef _KERNEL 6513 /* 6514 * Metadata is stored in the kernel's heap. Don't let us 6515 * use more than half the heap for the ARC. 6516 */ 6517 arc_meta_limit = MIN(arc_meta_limit, 6518 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2); 6519#endif 6520 6521 /* Allow the tunable to override if it is reasonable */ 6522 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 6523 arc_meta_limit = zfs_arc_meta_limit; 6524 6525 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 6526 arc_c_min = arc_meta_limit / 2; 6527 6528 if (zfs_arc_meta_min > 0) { 6529 arc_meta_min = zfs_arc_meta_min; 6530 } else { 6531 arc_meta_min = arc_c_min / 2; 6532 } 6533 6534 if (zfs_arc_grow_retry > 0) 6535 arc_grow_retry = zfs_arc_grow_retry; 6536 6537 if (zfs_arc_shrink_shift > 0) 6538 arc_shrink_shift = zfs_arc_shrink_shift; 6539 6540 if (zfs_arc_no_grow_shift > 0) 6541 arc_no_grow_shift = zfs_arc_no_grow_shift; 6542 /* 6543 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 6544 */ 6545 if (arc_no_grow_shift >= arc_shrink_shift) 6546 arc_no_grow_shift = arc_shrink_shift - 1; 6547 6548 if (zfs_arc_p_min_shift > 0) 6549 arc_p_min_shift = zfs_arc_p_min_shift; 6550 6551 /* if kmem_flags are set, lets try to use less memory */ 6552 if (kmem_debugging()) 6553 arc_c = arc_c / 2; 6554 if (arc_c < arc_c_min) 6555 arc_c = arc_c_min; 6556 6557 zfs_arc_min = arc_c_min; 6558 zfs_arc_max = arc_c_max; 6559 6560 arc_state_init(); 6561 buf_init(); 6562 6563 arc_reclaim_thread_exit = B_FALSE; 6564 arc_dnlc_evicts_thread_exit = FALSE; 6565 6566 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 6567 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 6568 6569 if (arc_ksp != NULL) { 6570 arc_ksp->ks_data = &arc_stats; 6571 arc_ksp->ks_update = arc_kstat_update; 6572 kstat_install(arc_ksp); 6573 } 6574 6575 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 6576 TS_RUN, minclsyspri); 6577 6578#ifdef _KERNEL 6579 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 6580 EVENTHANDLER_PRI_FIRST); 6581#endif 6582 6583 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0, 6584 TS_RUN, minclsyspri); 6585 6586 arc_dead = B_FALSE; 6587 arc_warm = B_FALSE; 6588 6589 /* 6590 * Calculate maximum amount of dirty data per pool. 6591 * 6592 * If it has been set by /etc/system, take that. 6593 * Otherwise, use a percentage of physical memory defined by 6594 * zfs_dirty_data_max_percent (default 10%) with a cap at 6595 * zfs_dirty_data_max_max (default 4GB). 6596 */ 6597 if (zfs_dirty_data_max == 0) { 6598 zfs_dirty_data_max = ptob(physmem) * 6599 zfs_dirty_data_max_percent / 100; 6600 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 6601 zfs_dirty_data_max_max); 6602 } 6603 6604#ifdef _KERNEL 6605 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 6606 prefetch_tunable_set = 1; 6607 6608#ifdef __i386__ 6609 if (prefetch_tunable_set == 0) { 6610 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 6611 "-- to enable,\n"); 6612 printf(" add \"vfs.zfs.prefetch_disable=0\" " 6613 "to /boot/loader.conf.\n"); 6614 zfs_prefetch_disable = 1; 6615 } 6616#else 6617 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 6618 prefetch_tunable_set == 0) { 6619 printf("ZFS NOTICE: Prefetch is disabled by default if less " 6620 "than 4GB of RAM is present;\n" 6621 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 6622 "to /boot/loader.conf.\n"); 6623 zfs_prefetch_disable = 1; 6624 } 6625#endif 6626 /* Warn about ZFS memory and address space requirements. */ 6627 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 6628 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 6629 "expect unstable behavior.\n"); 6630 } 6631 if (allmem < 512 * (1 << 20)) { 6632 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 6633 "expect unstable behavior.\n"); 6634 printf(" Consider tuning vm.kmem_size and " 6635 "vm.kmem_size_max\n"); 6636 printf(" in /boot/loader.conf.\n"); 6637 } 6638#endif 6639} 6640 6641void 6642arc_fini(void) 6643{ 6644#ifdef _KERNEL 6645 if (arc_event_lowmem != NULL) 6646 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 6647#endif 6648 6649 mutex_enter(&arc_reclaim_lock); 6650 arc_reclaim_thread_exit = B_TRUE; 6651 /* 6652 * The reclaim thread will set arc_reclaim_thread_exit back to 6653 * B_FALSE when it is finished exiting; we're waiting for that. 6654 */ 6655 while (arc_reclaim_thread_exit) { 6656 cv_signal(&arc_reclaim_thread_cv); 6657 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 6658 } 6659 mutex_exit(&arc_reclaim_lock); 6660 6661 /* Use B_TRUE to ensure *all* buffers are evicted */ 6662 arc_flush(NULL, B_TRUE); 6663 6664 mutex_enter(&arc_dnlc_evicts_lock); 6665 arc_dnlc_evicts_thread_exit = TRUE; 6666 /* 6667 * The user evicts thread will set arc_user_evicts_thread_exit 6668 * to FALSE when it is finished exiting; we're waiting for that. 6669 */ 6670 while (arc_dnlc_evicts_thread_exit) { 6671 cv_signal(&arc_dnlc_evicts_cv); 6672 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock); 6673 } 6674 mutex_exit(&arc_dnlc_evicts_lock); 6675 6676 arc_dead = B_TRUE; 6677 6678 if (arc_ksp != NULL) { 6679 kstat_delete(arc_ksp); 6680 arc_ksp = NULL; 6681 } 6682 6683 mutex_destroy(&arc_reclaim_lock); 6684 cv_destroy(&arc_reclaim_thread_cv); 6685 cv_destroy(&arc_reclaim_waiters_cv); 6686 6687 mutex_destroy(&arc_dnlc_evicts_lock); 6688 cv_destroy(&arc_dnlc_evicts_cv); 6689 6690 arc_state_fini(); 6691 buf_fini(); 6692 6693 ASSERT0(arc_loaned_bytes); 6694} 6695 6696/* 6697 * Level 2 ARC 6698 * 6699 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 6700 * It uses dedicated storage devices to hold cached data, which are populated 6701 * using large infrequent writes. The main role of this cache is to boost 6702 * the performance of random read workloads. The intended L2ARC devices 6703 * include short-stroked disks, solid state disks, and other media with 6704 * substantially faster read latency than disk. 6705 * 6706 * +-----------------------+ 6707 * | ARC | 6708 * +-----------------------+ 6709 * | ^ ^ 6710 * | | | 6711 * l2arc_feed_thread() arc_read() 6712 * | | | 6713 * | l2arc read | 6714 * V | | 6715 * +---------------+ | 6716 * | L2ARC | | 6717 * +---------------+ | 6718 * | ^ | 6719 * l2arc_write() | | 6720 * | | | 6721 * V | | 6722 * +-------+ +-------+ 6723 * | vdev | | vdev | 6724 * | cache | | cache | 6725 * +-------+ +-------+ 6726 * +=========+ .-----. 6727 * : L2ARC : |-_____-| 6728 * : devices : | Disks | 6729 * +=========+ `-_____-' 6730 * 6731 * Read requests are satisfied from the following sources, in order: 6732 * 6733 * 1) ARC 6734 * 2) vdev cache of L2ARC devices 6735 * 3) L2ARC devices 6736 * 4) vdev cache of disks 6737 * 5) disks 6738 * 6739 * Some L2ARC device types exhibit extremely slow write performance. 6740 * To accommodate for this there are some significant differences between 6741 * the L2ARC and traditional cache design: 6742 * 6743 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 6744 * the ARC behave as usual, freeing buffers and placing headers on ghost 6745 * lists. The ARC does not send buffers to the L2ARC during eviction as 6746 * this would add inflated write latencies for all ARC memory pressure. 6747 * 6748 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 6749 * It does this by periodically scanning buffers from the eviction-end of 6750 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 6751 * not already there. It scans until a headroom of buffers is satisfied, 6752 * which itself is a buffer for ARC eviction. If a compressible buffer is 6753 * found during scanning and selected for writing to an L2ARC device, we 6754 * temporarily boost scanning headroom during the next scan cycle to make 6755 * sure we adapt to compression effects (which might significantly reduce 6756 * the data volume we write to L2ARC). The thread that does this is 6757 * l2arc_feed_thread(), illustrated below; example sizes are included to 6758 * provide a better sense of ratio than this diagram: 6759 * 6760 * head --> tail 6761 * +---------------------+----------+ 6762 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 6763 * +---------------------+----------+ | o L2ARC eligible 6764 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 6765 * +---------------------+----------+ | 6766 * 15.9 Gbytes ^ 32 Mbytes | 6767 * headroom | 6768 * l2arc_feed_thread() 6769 * | 6770 * l2arc write hand <--[oooo]--' 6771 * | 8 Mbyte 6772 * | write max 6773 * V 6774 * +==============================+ 6775 * L2ARC dev |####|#|###|###| |####| ... | 6776 * +==============================+ 6777 * 32 Gbytes 6778 * 6779 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 6780 * evicted, then the L2ARC has cached a buffer much sooner than it probably 6781 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 6782 * safe to say that this is an uncommon case, since buffers at the end of 6783 * the ARC lists have moved there due to inactivity. 6784 * 6785 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 6786 * then the L2ARC simply misses copying some buffers. This serves as a 6787 * pressure valve to prevent heavy read workloads from both stalling the ARC 6788 * with waits and clogging the L2ARC with writes. This also helps prevent 6789 * the potential for the L2ARC to churn if it attempts to cache content too 6790 * quickly, such as during backups of the entire pool. 6791 * 6792 * 5. After system boot and before the ARC has filled main memory, there are 6793 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 6794 * lists can remain mostly static. Instead of searching from tail of these 6795 * lists as pictured, the l2arc_feed_thread() will search from the list heads 6796 * for eligible buffers, greatly increasing its chance of finding them. 6797 * 6798 * The L2ARC device write speed is also boosted during this time so that 6799 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 6800 * there are no L2ARC reads, and no fear of degrading read performance 6801 * through increased writes. 6802 * 6803 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 6804 * the vdev queue can aggregate them into larger and fewer writes. Each 6805 * device is written to in a rotor fashion, sweeping writes through 6806 * available space then repeating. 6807 * 6808 * 7. The L2ARC does not store dirty content. It never needs to flush 6809 * write buffers back to disk based storage. 6810 * 6811 * 8. If an ARC buffer is written (and dirtied) which also exists in the 6812 * L2ARC, the now stale L2ARC buffer is immediately dropped. 6813 * 6814 * The performance of the L2ARC can be tweaked by a number of tunables, which 6815 * may be necessary for different workloads: 6816 * 6817 * l2arc_write_max max write bytes per interval 6818 * l2arc_write_boost extra write bytes during device warmup 6819 * l2arc_noprefetch skip caching prefetched buffers 6820 * l2arc_headroom number of max device writes to precache 6821 * l2arc_headroom_boost when we find compressed buffers during ARC 6822 * scanning, we multiply headroom by this 6823 * percentage factor for the next scan cycle, 6824 * since more compressed buffers are likely to 6825 * be present 6826 * l2arc_feed_secs seconds between L2ARC writing 6827 * 6828 * Tunables may be removed or added as future performance improvements are 6829 * integrated, and also may become zpool properties. 6830 * 6831 * There are three key functions that control how the L2ARC warms up: 6832 * 6833 * l2arc_write_eligible() check if a buffer is eligible to cache 6834 * l2arc_write_size() calculate how much to write 6835 * l2arc_write_interval() calculate sleep delay between writes 6836 * 6837 * These three functions determine what to write, how much, and how quickly 6838 * to send writes. 6839 */ 6840 6841static boolean_t 6842l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 6843{ 6844 /* 6845 * A buffer is *not* eligible for the L2ARC if it: 6846 * 1. belongs to a different spa. 6847 * 2. is already cached on the L2ARC. 6848 * 3. has an I/O in progress (it may be an incomplete read). 6849 * 4. is flagged not eligible (zfs property). 6850 */ 6851 if (hdr->b_spa != spa_guid) { 6852 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 6853 return (B_FALSE); 6854 } 6855 if (HDR_HAS_L2HDR(hdr)) { 6856 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 6857 return (B_FALSE); 6858 } 6859 if (HDR_IO_IN_PROGRESS(hdr)) { 6860 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 6861 return (B_FALSE); 6862 } 6863 if (!HDR_L2CACHE(hdr)) { 6864 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 6865 return (B_FALSE); 6866 } 6867 6868 return (B_TRUE); 6869} 6870 6871static uint64_t 6872l2arc_write_size(void) 6873{ 6874 uint64_t size; 6875 6876 /* 6877 * Make sure our globals have meaningful values in case the user 6878 * altered them. 6879 */ 6880 size = l2arc_write_max; 6881 if (size == 0) { 6882 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 6883 "be greater than zero, resetting it to the default (%d)", 6884 L2ARC_WRITE_SIZE); 6885 size = l2arc_write_max = L2ARC_WRITE_SIZE; 6886 } 6887 6888 if (arc_warm == B_FALSE) 6889 size += l2arc_write_boost; 6890 6891 return (size); 6892 6893} 6894 6895static clock_t 6896l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 6897{ 6898 clock_t interval, next, now; 6899 6900 /* 6901 * If the ARC lists are busy, increase our write rate; if the 6902 * lists are stale, idle back. This is achieved by checking 6903 * how much we previously wrote - if it was more than half of 6904 * what we wanted, schedule the next write much sooner. 6905 */ 6906 if (l2arc_feed_again && wrote > (wanted / 2)) 6907 interval = (hz * l2arc_feed_min_ms) / 1000; 6908 else 6909 interval = hz * l2arc_feed_secs; 6910 6911 now = ddi_get_lbolt(); 6912 next = MAX(now, MIN(now + interval, began + interval)); 6913 6914 return (next); 6915} 6916 6917/* 6918 * Cycle through L2ARC devices. This is how L2ARC load balances. 6919 * If a device is returned, this also returns holding the spa config lock. 6920 */ 6921static l2arc_dev_t * 6922l2arc_dev_get_next(void) 6923{ 6924 l2arc_dev_t *first, *next = NULL; 6925 6926 /* 6927 * Lock out the removal of spas (spa_namespace_lock), then removal 6928 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 6929 * both locks will be dropped and a spa config lock held instead. 6930 */ 6931 mutex_enter(&spa_namespace_lock); 6932 mutex_enter(&l2arc_dev_mtx); 6933 6934 /* if there are no vdevs, there is nothing to do */ 6935 if (l2arc_ndev == 0) 6936 goto out; 6937 6938 first = NULL; 6939 next = l2arc_dev_last; 6940 do { 6941 /* loop around the list looking for a non-faulted vdev */ 6942 if (next == NULL) { 6943 next = list_head(l2arc_dev_list); 6944 } else { 6945 next = list_next(l2arc_dev_list, next); 6946 if (next == NULL) 6947 next = list_head(l2arc_dev_list); 6948 } 6949 6950 /* if we have come back to the start, bail out */ 6951 if (first == NULL) 6952 first = next; 6953 else if (next == first) 6954 break; 6955 6956 } while (vdev_is_dead(next->l2ad_vdev)); 6957 6958 /* if we were unable to find any usable vdevs, return NULL */ 6959 if (vdev_is_dead(next->l2ad_vdev)) 6960 next = NULL; 6961 6962 l2arc_dev_last = next; 6963 6964out: 6965 mutex_exit(&l2arc_dev_mtx); 6966 6967 /* 6968 * Grab the config lock to prevent the 'next' device from being 6969 * removed while we are writing to it. 6970 */ 6971 if (next != NULL) 6972 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 6973 mutex_exit(&spa_namespace_lock); 6974 6975 return (next); 6976} 6977 6978/* 6979 * Free buffers that were tagged for destruction. 6980 */ 6981static void 6982l2arc_do_free_on_write() 6983{ 6984 list_t *buflist; 6985 l2arc_data_free_t *df, *df_prev; 6986 6987 mutex_enter(&l2arc_free_on_write_mtx); 6988 buflist = l2arc_free_on_write; 6989 6990 for (df = list_tail(buflist); df; df = df_prev) { 6991 df_prev = list_prev(buflist, df); 6992 ASSERT3P(df->l2df_abd, !=, NULL); 6993 abd_free(df->l2df_abd); 6994 list_remove(buflist, df); 6995 kmem_free(df, sizeof (l2arc_data_free_t)); 6996 } 6997 6998 mutex_exit(&l2arc_free_on_write_mtx); 6999} 7000 7001/* 7002 * A write to a cache device has completed. Update all headers to allow 7003 * reads from these buffers to begin. 7004 */ 7005static void 7006l2arc_write_done(zio_t *zio) 7007{ 7008 l2arc_write_callback_t *cb; 7009 l2arc_dev_t *dev; 7010 list_t *buflist; 7011 arc_buf_hdr_t *head, *hdr, *hdr_prev; 7012 kmutex_t *hash_lock; 7013 int64_t bytes_dropped = 0; 7014 7015 cb = zio->io_private; 7016 ASSERT3P(cb, !=, NULL); 7017 dev = cb->l2wcb_dev; 7018 ASSERT3P(dev, !=, NULL); 7019 head = cb->l2wcb_head; 7020 ASSERT3P(head, !=, NULL); 7021 buflist = &dev->l2ad_buflist; 7022 ASSERT3P(buflist, !=, NULL); 7023 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 7024 l2arc_write_callback_t *, cb); 7025 7026 if (zio->io_error != 0) 7027 ARCSTAT_BUMP(arcstat_l2_writes_error); 7028 7029 /* 7030 * All writes completed, or an error was hit. 7031 */ 7032top: 7033 mutex_enter(&dev->l2ad_mtx); 7034 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 7035 hdr_prev = list_prev(buflist, hdr); 7036 7037 hash_lock = HDR_LOCK(hdr); 7038 7039 /* 7040 * We cannot use mutex_enter or else we can deadlock 7041 * with l2arc_write_buffers (due to swapping the order 7042 * the hash lock and l2ad_mtx are taken). 7043 */ 7044 if (!mutex_tryenter(hash_lock)) { 7045 /* 7046 * Missed the hash lock. We must retry so we 7047 * don't leave the ARC_FLAG_L2_WRITING bit set. 7048 */ 7049 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 7050 7051 /* 7052 * We don't want to rescan the headers we've 7053 * already marked as having been written out, so 7054 * we reinsert the head node so we can pick up 7055 * where we left off. 7056 */ 7057 list_remove(buflist, head); 7058 list_insert_after(buflist, hdr, head); 7059 7060 mutex_exit(&dev->l2ad_mtx); 7061 7062 /* 7063 * We wait for the hash lock to become available 7064 * to try and prevent busy waiting, and increase 7065 * the chance we'll be able to acquire the lock 7066 * the next time around. 7067 */ 7068 mutex_enter(hash_lock); 7069 mutex_exit(hash_lock); 7070 goto top; 7071 } 7072 7073 /* 7074 * We could not have been moved into the arc_l2c_only 7075 * state while in-flight due to our ARC_FLAG_L2_WRITING 7076 * bit being set. Let's just ensure that's being enforced. 7077 */ 7078 ASSERT(HDR_HAS_L1HDR(hdr)); 7079 7080 if (zio->io_error != 0) { 7081 /* 7082 * Error - drop L2ARC entry. 7083 */ 7084 list_remove(buflist, hdr); 7085 l2arc_trim(hdr); 7086 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 7087 7088 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr)); 7089 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); 7090 7091 bytes_dropped += arc_hdr_size(hdr); 7092 (void) refcount_remove_many(&dev->l2ad_alloc, 7093 arc_hdr_size(hdr), hdr); 7094 } 7095 7096 /* 7097 * Allow ARC to begin reads and ghost list evictions to 7098 * this L2ARC entry. 7099 */ 7100 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); 7101 7102 mutex_exit(hash_lock); 7103 } 7104 7105 atomic_inc_64(&l2arc_writes_done); 7106 list_remove(buflist, head); 7107 ASSERT(!HDR_HAS_L1HDR(head)); 7108 kmem_cache_free(hdr_l2only_cache, head); 7109 mutex_exit(&dev->l2ad_mtx); 7110 7111 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 7112 7113 l2arc_do_free_on_write(); 7114 7115 kmem_free(cb, sizeof (l2arc_write_callback_t)); 7116} 7117 7118/* 7119 * A read to a cache device completed. Validate buffer contents before 7120 * handing over to the regular ARC routines. 7121 */ 7122static void 7123l2arc_read_done(zio_t *zio) 7124{ 7125 l2arc_read_callback_t *cb; 7126 arc_buf_hdr_t *hdr; 7127 kmutex_t *hash_lock; 7128 boolean_t valid_cksum; 7129 7130 ASSERT3P(zio->io_vd, !=, NULL); 7131 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 7132 7133 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 7134 7135 cb = zio->io_private; 7136 ASSERT3P(cb, !=, NULL); 7137 hdr = cb->l2rcb_hdr; 7138 ASSERT3P(hdr, !=, NULL); 7139 7140 hash_lock = HDR_LOCK(hdr); 7141 mutex_enter(hash_lock); 7142 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 7143 7144 /* 7145 * If the data was read into a temporary buffer, 7146 * move it and free the buffer. 7147 */ 7148 if (cb->l2rcb_abd != NULL) { 7149 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size); 7150 if (zio->io_error == 0) { 7151 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd, 7152 arc_hdr_size(hdr)); 7153 } 7154 7155 /* 7156 * The following must be done regardless of whether 7157 * there was an error: 7158 * - free the temporary buffer 7159 * - point zio to the real ARC buffer 7160 * - set zio size accordingly 7161 * These are required because zio is either re-used for 7162 * an I/O of the block in the case of the error 7163 * or the zio is passed to arc_read_done() and it 7164 * needs real data. 7165 */ 7166 abd_free(cb->l2rcb_abd); 7167 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr); 7168 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd; 7169 } 7170 7171 ASSERT3P(zio->io_abd, !=, NULL); 7172 7173 /* 7174 * Check this survived the L2ARC journey. 7175 */ 7176 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd); 7177 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 7178 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 7179 7180 valid_cksum = arc_cksum_is_equal(hdr, zio); 7181 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 7182 mutex_exit(hash_lock); 7183 zio->io_private = hdr; 7184 arc_read_done(zio); 7185 } else { 7186 mutex_exit(hash_lock); 7187 /* 7188 * Buffer didn't survive caching. Increment stats and 7189 * reissue to the original storage device. 7190 */ 7191 if (zio->io_error != 0) { 7192 ARCSTAT_BUMP(arcstat_l2_io_error); 7193 } else { 7194 zio->io_error = SET_ERROR(EIO); 7195 } 7196 if (!valid_cksum) 7197 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 7198 7199 /* 7200 * If there's no waiter, issue an async i/o to the primary 7201 * storage now. If there *is* a waiter, the caller must 7202 * issue the i/o in a context where it's OK to block. 7203 */ 7204 if (zio->io_waiter == NULL) { 7205 zio_t *pio = zio_unique_parent(zio); 7206 7207 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 7208 7209 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, 7210 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done, 7211 hdr, zio->io_priority, cb->l2rcb_flags, 7212 &cb->l2rcb_zb)); 7213 } 7214 } 7215 7216 kmem_free(cb, sizeof (l2arc_read_callback_t)); 7217} 7218 7219/* 7220 * This is the list priority from which the L2ARC will search for pages to 7221 * cache. This is used within loops (0..3) to cycle through lists in the 7222 * desired order. This order can have a significant effect on cache 7223 * performance. 7224 * 7225 * Currently the metadata lists are hit first, MFU then MRU, followed by 7226 * the data lists. This function returns a locked list, and also returns 7227 * the lock pointer. 7228 */ 7229static multilist_sublist_t * 7230l2arc_sublist_lock(int list_num) 7231{ 7232 multilist_t *ml = NULL; 7233 unsigned int idx; 7234 7235 ASSERT(list_num >= 0 && list_num <= 3); 7236 7237 switch (list_num) { 7238 case 0: 7239 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA]; 7240 break; 7241 case 1: 7242 ml = arc_mru->arcs_list[ARC_BUFC_METADATA]; 7243 break; 7244 case 2: 7245 ml = arc_mfu->arcs_list[ARC_BUFC_DATA]; 7246 break; 7247 case 3: 7248 ml = arc_mru->arcs_list[ARC_BUFC_DATA]; 7249 break; 7250 } 7251 7252 /* 7253 * Return a randomly-selected sublist. This is acceptable 7254 * because the caller feeds only a little bit of data for each 7255 * call (8MB). Subsequent calls will result in different 7256 * sublists being selected. 7257 */ 7258 idx = multilist_get_random_index(ml); 7259 return (multilist_sublist_lock(ml, idx)); 7260} 7261 7262/* 7263 * Evict buffers from the device write hand to the distance specified in 7264 * bytes. This distance may span populated buffers, it may span nothing. 7265 * This is clearing a region on the L2ARC device ready for writing. 7266 * If the 'all' boolean is set, every buffer is evicted. 7267 */ 7268static void 7269l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 7270{ 7271 list_t *buflist; 7272 arc_buf_hdr_t *hdr, *hdr_prev; 7273 kmutex_t *hash_lock; 7274 uint64_t taddr; 7275 7276 buflist = &dev->l2ad_buflist; 7277 7278 if (!all && dev->l2ad_first) { 7279 /* 7280 * This is the first sweep through the device. There is 7281 * nothing to evict. 7282 */ 7283 return; 7284 } 7285 7286 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 7287 /* 7288 * When nearing the end of the device, evict to the end 7289 * before the device write hand jumps to the start. 7290 */ 7291 taddr = dev->l2ad_end; 7292 } else { 7293 taddr = dev->l2ad_hand + distance; 7294 } 7295 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 7296 uint64_t, taddr, boolean_t, all); 7297 7298top: 7299 mutex_enter(&dev->l2ad_mtx); 7300 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 7301 hdr_prev = list_prev(buflist, hdr); 7302 7303 hash_lock = HDR_LOCK(hdr); 7304 7305 /* 7306 * We cannot use mutex_enter or else we can deadlock 7307 * with l2arc_write_buffers (due to swapping the order 7308 * the hash lock and l2ad_mtx are taken). 7309 */ 7310 if (!mutex_tryenter(hash_lock)) { 7311 /* 7312 * Missed the hash lock. Retry. 7313 */ 7314 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 7315 mutex_exit(&dev->l2ad_mtx); 7316 mutex_enter(hash_lock); 7317 mutex_exit(hash_lock); 7318 goto top; 7319 } 7320 7321 /* 7322 * A header can't be on this list if it doesn't have L2 header. 7323 */ 7324 ASSERT(HDR_HAS_L2HDR(hdr)); 7325 7326 /* Ensure this header has finished being written. */ 7327 ASSERT(!HDR_L2_WRITING(hdr)); 7328 ASSERT(!HDR_L2_WRITE_HEAD(hdr)); 7329 7330 if (!all && (hdr->b_l2hdr.b_daddr >= taddr || 7331 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 7332 /* 7333 * We've evicted to the target address, 7334 * or the end of the device. 7335 */ 7336 mutex_exit(hash_lock); 7337 break; 7338 } 7339 7340 if (!HDR_HAS_L1HDR(hdr)) { 7341 ASSERT(!HDR_L2_READING(hdr)); 7342 /* 7343 * This doesn't exist in the ARC. Destroy. 7344 * arc_hdr_destroy() will call list_remove() 7345 * and decrement arcstat_l2_lsize. 7346 */ 7347 arc_change_state(arc_anon, hdr, hash_lock); 7348 arc_hdr_destroy(hdr); 7349 } else { 7350 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 7351 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 7352 /* 7353 * Invalidate issued or about to be issued 7354 * reads, since we may be about to write 7355 * over this location. 7356 */ 7357 if (HDR_L2_READING(hdr)) { 7358 ARCSTAT_BUMP(arcstat_l2_evict_reading); 7359 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); 7360 } 7361 7362 arc_hdr_l2hdr_destroy(hdr); 7363 } 7364 mutex_exit(hash_lock); 7365 } 7366 mutex_exit(&dev->l2ad_mtx); 7367} 7368 7369/* 7370 * Find and write ARC buffers to the L2ARC device. 7371 * 7372 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 7373 * for reading until they have completed writing. 7374 * The headroom_boost is an in-out parameter used to maintain headroom boost 7375 * state between calls to this function. 7376 * 7377 * Returns the number of bytes actually written (which may be smaller than 7378 * the delta by which the device hand has changed due to alignment). 7379 */ 7380static uint64_t 7381l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 7382{ 7383 arc_buf_hdr_t *hdr, *hdr_prev, *head; 7384 uint64_t write_asize, write_psize, write_lsize, headroom; 7385 boolean_t full; 7386 l2arc_write_callback_t *cb; 7387 zio_t *pio, *wzio; 7388 uint64_t guid = spa_load_guid(spa); 7389 int try; 7390 7391 ASSERT3P(dev->l2ad_vdev, !=, NULL); 7392 7393 pio = NULL; 7394 write_lsize = write_asize = write_psize = 0; 7395 full = B_FALSE; 7396 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 7397 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); 7398 7399 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 7400 /* 7401 * Copy buffers for L2ARC writing. 7402 */ 7403 for (try = 0; try <= 3; try++) { 7404 multilist_sublist_t *mls = l2arc_sublist_lock(try); 7405 uint64_t passed_sz = 0; 7406 7407 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 7408 7409 /* 7410 * L2ARC fast warmup. 7411 * 7412 * Until the ARC is warm and starts to evict, read from the 7413 * head of the ARC lists rather than the tail. 7414 */ 7415 if (arc_warm == B_FALSE) 7416 hdr = multilist_sublist_head(mls); 7417 else 7418 hdr = multilist_sublist_tail(mls); 7419 if (hdr == NULL) 7420 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 7421 7422 headroom = target_sz * l2arc_headroom; 7423 if (zfs_compressed_arc_enabled) 7424 headroom = (headroom * l2arc_headroom_boost) / 100; 7425 7426 for (; hdr; hdr = hdr_prev) { 7427 kmutex_t *hash_lock; 7428 7429 if (arc_warm == B_FALSE) 7430 hdr_prev = multilist_sublist_next(mls, hdr); 7431 else 7432 hdr_prev = multilist_sublist_prev(mls, hdr); 7433 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, 7434 HDR_GET_LSIZE(hdr)); 7435 7436 hash_lock = HDR_LOCK(hdr); 7437 if (!mutex_tryenter(hash_lock)) { 7438 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 7439 /* 7440 * Skip this buffer rather than waiting. 7441 */ 7442 continue; 7443 } 7444 7445 passed_sz += HDR_GET_LSIZE(hdr); 7446 if (passed_sz > headroom) { 7447 /* 7448 * Searched too far. 7449 */ 7450 mutex_exit(hash_lock); 7451 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 7452 break; 7453 } 7454 7455 if (!l2arc_write_eligible(guid, hdr)) { 7456 mutex_exit(hash_lock); 7457 continue; 7458 } 7459 7460 /* 7461 * We rely on the L1 portion of the header below, so 7462 * it's invalid for this header to have been evicted out 7463 * of the ghost cache, prior to being written out. The 7464 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 7465 */ 7466 ASSERT(HDR_HAS_L1HDR(hdr)); 7467 7468 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); 7469 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 7470 ASSERT3U(arc_hdr_size(hdr), >, 0); 7471 uint64_t psize = arc_hdr_size(hdr); 7472 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, 7473 psize); 7474 7475 if ((write_asize + asize) > target_sz) { 7476 full = B_TRUE; 7477 mutex_exit(hash_lock); 7478 ARCSTAT_BUMP(arcstat_l2_write_full); 7479 break; 7480 } 7481 7482 if (pio == NULL) { 7483 /* 7484 * Insert a dummy header on the buflist so 7485 * l2arc_write_done() can find where the 7486 * write buffers begin without searching. 7487 */ 7488 mutex_enter(&dev->l2ad_mtx); 7489 list_insert_head(&dev->l2ad_buflist, head); 7490 mutex_exit(&dev->l2ad_mtx); 7491 7492 cb = kmem_alloc( 7493 sizeof (l2arc_write_callback_t), KM_SLEEP); 7494 cb->l2wcb_dev = dev; 7495 cb->l2wcb_head = head; 7496 pio = zio_root(spa, l2arc_write_done, cb, 7497 ZIO_FLAG_CANFAIL); 7498 ARCSTAT_BUMP(arcstat_l2_write_pios); 7499 } 7500 7501 hdr->b_l2hdr.b_dev = dev; 7502 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 7503 arc_hdr_set_flags(hdr, 7504 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR); 7505 7506 mutex_enter(&dev->l2ad_mtx); 7507 list_insert_head(&dev->l2ad_buflist, hdr); 7508 mutex_exit(&dev->l2ad_mtx); 7509 7510 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr); 7511 7512 /* 7513 * Normally the L2ARC can use the hdr's data, but if 7514 * we're sharing data between the hdr and one of its 7515 * bufs, L2ARC needs its own copy of the data so that 7516 * the ZIO below can't race with the buf consumer. 7517 * Another case where we need to create a copy of the 7518 * data is when the buffer size is not device-aligned 7519 * and we need to pad the block to make it such. 7520 * That also keeps the clock hand suitably aligned. 7521 * 7522 * To ensure that the copy will be available for the 7523 * lifetime of the ZIO and be cleaned up afterwards, we 7524 * add it to the l2arc_free_on_write queue. 7525 */ 7526 abd_t *to_write; 7527 if (!HDR_SHARED_DATA(hdr) && psize == asize) { 7528 to_write = hdr->b_l1hdr.b_pabd; 7529 } else { 7530 to_write = abd_alloc_for_io(asize, 7531 HDR_ISTYPE_METADATA(hdr)); 7532 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize); 7533 if (asize != psize) { 7534 abd_zero_off(to_write, psize, 7535 asize - psize); 7536 } 7537 l2arc_free_abd_on_write(to_write, asize, 7538 arc_buf_type(hdr)); 7539 } 7540 wzio = zio_write_phys(pio, dev->l2ad_vdev, 7541 hdr->b_l2hdr.b_daddr, asize, to_write, 7542 ZIO_CHECKSUM_OFF, NULL, hdr, 7543 ZIO_PRIORITY_ASYNC_WRITE, 7544 ZIO_FLAG_CANFAIL, B_FALSE); 7545 7546 write_lsize += HDR_GET_LSIZE(hdr); 7547 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 7548 zio_t *, wzio); 7549 7550 write_psize += psize; 7551 write_asize += asize; 7552 dev->l2ad_hand += asize; 7553 7554 mutex_exit(hash_lock); 7555 7556 (void) zio_nowait(wzio); 7557 } 7558 7559 multilist_sublist_unlock(mls); 7560 7561 if (full == B_TRUE) 7562 break; 7563 } 7564 7565 /* No buffers selected for writing? */ 7566 if (pio == NULL) { 7567 ASSERT0(write_lsize); 7568 ASSERT(!HDR_HAS_L1HDR(head)); 7569 kmem_cache_free(hdr_l2only_cache, head); 7570 return (0); 7571 } 7572 7573 ASSERT3U(write_psize, <=, target_sz); 7574 ARCSTAT_BUMP(arcstat_l2_writes_sent); 7575 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize); 7576 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize); 7577 ARCSTAT_INCR(arcstat_l2_psize, write_psize); 7578 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 7579 7580 /* 7581 * Bump device hand to the device start if it is approaching the end. 7582 * l2arc_evict() will already have evicted ahead for this case. 7583 */ 7584 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 7585 dev->l2ad_hand = dev->l2ad_start; 7586 dev->l2ad_first = B_FALSE; 7587 } 7588 7589 dev->l2ad_writing = B_TRUE; 7590 (void) zio_wait(pio); 7591 dev->l2ad_writing = B_FALSE; 7592 7593 return (write_asize); 7594} 7595 7596/* 7597 * This thread feeds the L2ARC at regular intervals. This is the beating 7598 * heart of the L2ARC. 7599 */ 7600/* ARGSUSED */ 7601static void 7602l2arc_feed_thread(void *unused __unused) 7603{ 7604 callb_cpr_t cpr; 7605 l2arc_dev_t *dev; 7606 spa_t *spa; 7607 uint64_t size, wrote; 7608 clock_t begin, next = ddi_get_lbolt(); 7609 7610 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 7611 7612 mutex_enter(&l2arc_feed_thr_lock); 7613 7614 while (l2arc_thread_exit == 0) { 7615 CALLB_CPR_SAFE_BEGIN(&cpr); 7616 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 7617 next - ddi_get_lbolt()); 7618 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 7619 next = ddi_get_lbolt() + hz; 7620 7621 /* 7622 * Quick check for L2ARC devices. 7623 */ 7624 mutex_enter(&l2arc_dev_mtx); 7625 if (l2arc_ndev == 0) { 7626 mutex_exit(&l2arc_dev_mtx); 7627 continue; 7628 } 7629 mutex_exit(&l2arc_dev_mtx); 7630 begin = ddi_get_lbolt(); 7631 7632 /* 7633 * This selects the next l2arc device to write to, and in 7634 * doing so the next spa to feed from: dev->l2ad_spa. This 7635 * will return NULL if there are now no l2arc devices or if 7636 * they are all faulted. 7637 * 7638 * If a device is returned, its spa's config lock is also 7639 * held to prevent device removal. l2arc_dev_get_next() 7640 * will grab and release l2arc_dev_mtx. 7641 */ 7642 if ((dev = l2arc_dev_get_next()) == NULL) 7643 continue; 7644 7645 spa = dev->l2ad_spa; 7646 ASSERT3P(spa, !=, NULL); 7647 7648 /* 7649 * If the pool is read-only then force the feed thread to 7650 * sleep a little longer. 7651 */ 7652 if (!spa_writeable(spa)) { 7653 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 7654 spa_config_exit(spa, SCL_L2ARC, dev); 7655 continue; 7656 } 7657 7658 /* 7659 * Avoid contributing to memory pressure. 7660 */ 7661 if (arc_reclaim_needed()) { 7662 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 7663 spa_config_exit(spa, SCL_L2ARC, dev); 7664 continue; 7665 } 7666 7667 ARCSTAT_BUMP(arcstat_l2_feeds); 7668 7669 size = l2arc_write_size(); 7670 7671 /* 7672 * Evict L2ARC buffers that will be overwritten. 7673 */ 7674 l2arc_evict(dev, size, B_FALSE); 7675 7676 /* 7677 * Write ARC buffers. 7678 */ 7679 wrote = l2arc_write_buffers(spa, dev, size); 7680 7681 /* 7682 * Calculate interval between writes. 7683 */ 7684 next = l2arc_write_interval(begin, size, wrote); 7685 spa_config_exit(spa, SCL_L2ARC, dev); 7686 } 7687 7688 l2arc_thread_exit = 0; 7689 cv_broadcast(&l2arc_feed_thr_cv); 7690 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 7691 thread_exit(); 7692} 7693 7694boolean_t 7695l2arc_vdev_present(vdev_t *vd) 7696{ 7697 l2arc_dev_t *dev; 7698 7699 mutex_enter(&l2arc_dev_mtx); 7700 for (dev = list_head(l2arc_dev_list); dev != NULL; 7701 dev = list_next(l2arc_dev_list, dev)) { 7702 if (dev->l2ad_vdev == vd) 7703 break; 7704 } 7705 mutex_exit(&l2arc_dev_mtx); 7706 7707 return (dev != NULL); 7708} 7709 7710/* 7711 * Add a vdev for use by the L2ARC. By this point the spa has already 7712 * validated the vdev and opened it. 7713 */ 7714void 7715l2arc_add_vdev(spa_t *spa, vdev_t *vd) 7716{ 7717 l2arc_dev_t *adddev; 7718 7719 ASSERT(!l2arc_vdev_present(vd)); 7720 7721 vdev_ashift_optimize(vd); 7722 7723 /* 7724 * Create a new l2arc device entry. 7725 */ 7726 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 7727 adddev->l2ad_spa = spa; 7728 adddev->l2ad_vdev = vd; 7729 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 7730 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 7731 adddev->l2ad_hand = adddev->l2ad_start; 7732 adddev->l2ad_first = B_TRUE; 7733 adddev->l2ad_writing = B_FALSE; 7734 7735 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 7736 /* 7737 * This is a list of all ARC buffers that are still valid on the 7738 * device. 7739 */ 7740 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 7741 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 7742 7743 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 7744 refcount_create(&adddev->l2ad_alloc); 7745 7746 /* 7747 * Add device to global list 7748 */ 7749 mutex_enter(&l2arc_dev_mtx); 7750 list_insert_head(l2arc_dev_list, adddev); 7751 atomic_inc_64(&l2arc_ndev); 7752 mutex_exit(&l2arc_dev_mtx); 7753} 7754 7755/* 7756 * Remove a vdev from the L2ARC. 7757 */ 7758void 7759l2arc_remove_vdev(vdev_t *vd) 7760{ 7761 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 7762 7763 /* 7764 * Find the device by vdev 7765 */ 7766 mutex_enter(&l2arc_dev_mtx); 7767 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 7768 nextdev = list_next(l2arc_dev_list, dev); 7769 if (vd == dev->l2ad_vdev) { 7770 remdev = dev; 7771 break; 7772 } 7773 } 7774 ASSERT3P(remdev, !=, NULL); 7775 7776 /* 7777 * Remove device from global list 7778 */ 7779 list_remove(l2arc_dev_list, remdev); 7780 l2arc_dev_last = NULL; /* may have been invalidated */ 7781 atomic_dec_64(&l2arc_ndev); 7782 mutex_exit(&l2arc_dev_mtx); 7783 7784 /* 7785 * Clear all buflists and ARC references. L2ARC device flush. 7786 */ 7787 l2arc_evict(remdev, 0, B_TRUE); 7788 list_destroy(&remdev->l2ad_buflist); 7789 mutex_destroy(&remdev->l2ad_mtx); 7790 refcount_destroy(&remdev->l2ad_alloc); 7791 kmem_free(remdev, sizeof (l2arc_dev_t)); 7792} 7793 7794void 7795l2arc_init(void) 7796{ 7797 l2arc_thread_exit = 0; 7798 l2arc_ndev = 0; 7799 l2arc_writes_sent = 0; 7800 l2arc_writes_done = 0; 7801 7802 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7803 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7804 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7805 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7806 7807 l2arc_dev_list = &L2ARC_dev_list; 7808 l2arc_free_on_write = &L2ARC_free_on_write; 7809 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7810 offsetof(l2arc_dev_t, l2ad_node)); 7811 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7812 offsetof(l2arc_data_free_t, l2df_list_node)); 7813} 7814 7815void 7816l2arc_fini(void) 7817{ 7818 /* 7819 * This is called from dmu_fini(), which is called from spa_fini(); 7820 * Because of this, we can assume that all l2arc devices have 7821 * already been removed when the pools themselves were removed. 7822 */ 7823 7824 l2arc_do_free_on_write(); 7825 7826 mutex_destroy(&l2arc_feed_thr_lock); 7827 cv_destroy(&l2arc_feed_thr_cv); 7828 mutex_destroy(&l2arc_dev_mtx); 7829 mutex_destroy(&l2arc_free_on_write_mtx); 7830 7831 list_destroy(l2arc_dev_list); 7832 list_destroy(l2arc_free_on_write); 7833} 7834 7835void 7836l2arc_start(void) 7837{ 7838 if (!(spa_mode_global & FWRITE)) 7839 return; 7840 7841 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7842 TS_RUN, minclsyspri); 7843} 7844 7845void 7846l2arc_stop(void) 7847{ 7848 if (!(spa_mode_global & FWRITE)) 7849 return; 7850 7851 mutex_enter(&l2arc_feed_thr_lock); 7852 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7853 l2arc_thread_exit = 1; 7854 while (l2arc_thread_exit != 0) 7855 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7856 mutex_exit(&l2arc_feed_thr_lock); 7857} 7858