arc.c revision 332785
1139749Simp/* 2113584Ssimokawa * CDDL HEADER START 3103285Sikob * 4103285Sikob * The contents of this file are subject to the terms of the 5103285Sikob * Common Development and Distribution License (the "License"). 6103285Sikob * You may not use this file except in compliance with the License. 7103285Sikob * 8103285Sikob * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9103285Sikob * or http://www.opensolaris.org/os/licensing. 10103285Sikob * See the License for the specific language governing permissions 11103285Sikob * and limitations under the License. 12103285Sikob * 13103285Sikob * When distributing Covered Code, include this CDDL HEADER in each 14103285Sikob * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15103285Sikob * If applicable, add the following below this CDDL HEADER, with the 16103285Sikob * fields enclosed by brackets "[]" replaced with your own identifying 17103285Sikob * information: Portions Copyright [yyyy] [name of copyright owner] 18103285Sikob * 19103285Sikob * CDDL HEADER END 20103285Sikob */ 21103285Sikob/* 22103285Sikob * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23103285Sikob * Copyright (c) 2018, Joyent, Inc. 24103285Sikob * Copyright (c) 2011, 2017 by Delphix. All rights reserved. 25103285Sikob * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 26103285Sikob * Copyright 2017 Nexenta Systems, Inc. All rights reserved. 27103285Sikob */ 28103285Sikob 29103285Sikob/* 30103285Sikob * DVA-based Adjustable Replacement Cache 31103285Sikob * 32103285Sikob * While much of the theory of operation used here is 33103285Sikob * based on the self-tuning, low overhead replacement cache 34103285Sikob * presented by Megiddo and Modha at FAST 2003, there are some 35103285Sikob * significant differences: 36103285Sikob * 37103285Sikob * 1. The Megiddo and Modha model assumes any page is evictable. 38127468Ssimokawa * Pages in its cache cannot be "locked" into memory. This makes 39127468Ssimokawa * the eviction algorithm simple: evict the last page in the list. 40127468Ssimokawa * This also make the performance characteristics easy to reason 41127468Ssimokawa * about. Our cache is not so simple. At any given moment, some 42103285Sikob * subset of the blocks in the cache are un-evictable because we 43103285Sikob * have handed out a reference to them. Blocks are only evictable 44103285Sikob * when there are no external references active. This makes 45103285Sikob * eviction far more problematic: we choose to evict the evictable 46103285Sikob * blocks that are the "lowest" in the list. 47103285Sikob * 48103285Sikob * There are times when it is not possible to evict the requested 49113584Ssimokawa * space. In these circumstances we are unable to adjust the cache 50113584Ssimokawa * size. To prevent the cache growing unbounded at these times we 51103285Sikob * implement a "cache throttle" that slows the flow of new data 52103285Sikob * into the cache until we can make space available. 53169130Ssimokawa * 54169130Ssimokawa * 2. The Megiddo and Modha model assumes a fixed cache size. 55103285Sikob * Pages are evicted when the cache is full and there is a cache 56129585Sdfr * miss. Our model has a variable sized cache. It grows with 57103285Sikob * high use, but also tries to react to memory pressure from the 58129585Sdfr * operating system: decreasing its size when system memory is 59129585Sdfr * tight. 60129585Sdfr * 61129585Sdfr * 3. The Megiddo and Modha model assumes a fixed page size. All 62103285Sikob * elements of the cache are therefore exactly the same size. So 63103285Sikob * when adjusting the cache size following a cache miss, its simply 64103285Sikob * a matter of choosing a single page to evict. In our model, we 65129585Sdfr * have variable sized cache blocks (rangeing from 512 bytes to 66103285Sikob * 128K bytes). We therefore choose a set of blocks to evict to make 67106810Ssimokawa * space for a cache miss that approximates as closely as possible 68129585Sdfr * the space used by the new block. 69103285Sikob * 70103285Sikob * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 71103285Sikob * by N. Megiddo & D. Modha, FAST 2003 72110193Ssimokawa */ 73103285Sikob 74109645Ssimokawa/* 75103285Sikob * The locking model: 76127468Ssimokawa * 77130585Sphk * A new reference to a cache buffer can be obtained in two 78103285Sikob * ways: 1) via a hash table lookup using the DVA as a key, 79103285Sikob * or 2) via one of the ARC lists. The arc_read() interface 80103285Sikob * uses method 1, while the internal ARC algorithms for 81109645Ssimokawa * adjusting the cache use method 2. We therefore provide two 82103285Sikob * types of locks: 1) the hash table lock array, and 2) the 83103285Sikob * ARC list locks. 84103285Sikob * 85109645Ssimokawa * Buffers do not have their own mutexes, rather they rely on the 86103285Sikob * hash table mutexes for the bulk of their protection (i.e. most 87103285Sikob * fields in the arc_buf_hdr_t are protected by these mutexes). 88103285Sikob * 89124169Ssimokawa * buf_hash_find() returns the appropriate mutex (held) when it 90124169Ssimokawa * locates the requested buffer in the hash table. It returns 91103285Sikob * NULL for the mutex if the buffer was not in the table. 92103285Sikob * 93103285Sikob * buf_hash_remove() expects the appropriate hash mutex to be 94103285Sikob * already held before it is invoked. 95103285Sikob * 96103285Sikob * Each ARC state also has a mutex which is used to protect the 97103285Sikob * buffer list associated with the state. When attempting to 98103285Sikob * obtain a hash table lock while holding an ARC list lock you 99103285Sikob * must use: mutex_tryenter() to avoid deadlock. Also note that 100103285Sikob * the active state mutex must be held before the ghost state mutex. 101103285Sikob * 102103285Sikob * Note that the majority of the performance stats are manipulated 103103285Sikob * with atomic operations. 104103285Sikob * 105103285Sikob * The L2ARC uses the l2ad_mtx on each vdev for the following: 106129585Sdfr * 107103285Sikob * - L2ARC buflist creation 108103285Sikob * - L2ARC buflist eviction 109103285Sikob * - L2ARC write completion, which walks L2ARC buflists 110103285Sikob * - ARC header destruction, as it removes from L2ARC buflists 111103285Sikob * - ARC header release, as it removes from L2ARC buflists 112103285Sikob */ 113103285Sikob 114103285Sikob/* 115103285Sikob * ARC operation: 116129585Sdfr * 117169806Ssimokawa * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure. 118116978Ssimokawa * This structure can point either to a block that is still in the cache or to 119103285Sikob * one that is only accessible in an L2 ARC device, or it can provide 120103285Sikob * information about a block that was recently evicted. If a block is 121103285Sikob * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough 122103285Sikob * information to retrieve it from the L2ARC device. This information is 123103285Sikob * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block 124103285Sikob * that is in this state cannot access the data directly. 125103285Sikob * 126103285Sikob * Blocks that are actively being referenced or have not been evicted 127103285Sikob * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within 128103285Sikob * the arc_buf_hdr_t that will point to the data block in memory. A block can 129109814Ssimokawa * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC 130103285Sikob * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and 131103285Sikob * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd). 132169130Ssimokawa * 133103285Sikob * The L1ARC's data pointer may or may not be uncompressed. The ARC has the 134110193Ssimokawa * ability to store the physical data (b_pabd) associated with the DVA of the 135103285Sikob * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block, 136103285Sikob * it will match its on-disk compression characteristics. This behavior can be 137129585Sdfr * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the 138103285Sikob * compressed ARC functionality is disabled, the b_pabd will point to an 139129585Sdfr * uncompressed version of the on-disk data. 140116376Ssimokawa * 141116376Ssimokawa * Data in the L1ARC is not accessed by consumers of the ARC directly. Each 142116376Ssimokawa * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it. 143103285Sikob * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC 144103285Sikob * consumer. The ARC will provide references to this data and will keep it 145108853Ssimokawa * cached until it is no longer in use. The ARC caches only the L1ARC's physical 146110193Ssimokawa * data block and will evict any arc_buf_t that is no longer referenced. The 147110193Ssimokawa * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the 148129585Sdfr * "overhead_size" kstat. 149124169Ssimokawa * 150129585Sdfr * Depending on the consumer, an arc_buf_t can be requested in uncompressed or 151130585Sphk * compressed form. The typical case is that consumers will want uncompressed 152124169Ssimokawa * data, and when that happens a new data buffer is allocated where the data is 153124169Ssimokawa * decompressed for them to use. Currently the only consumer who wants 154124169Ssimokawa * compressed arc_buf_t's is "zfs send", when it streams data exactly as it 155124169Ssimokawa * exists on disk. When this happens, the arc_buf_t's data buffer is shared 156124169Ssimokawa * with the arc_buf_hdr_t. 157124169Ssimokawa * 158124169Ssimokawa * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The 159129585Sdfr * first one is owned by a compressed send consumer (and therefore references 160129585Sdfr * the same compressed data buffer as the arc_buf_hdr_t) and the second could be 161103285Sikob * used by any other consumer (and has its own uncompressed copy of the data 162113584Ssimokawa * buffer). 163103285Sikob * 164103285Sikob * arc_buf_hdr_t 165103285Sikob * +-----------+ 166109645Ssimokawa * | fields | 167109645Ssimokawa * | common to | 168109645Ssimokawa * | L1- and | 169109645Ssimokawa * | L2ARC | 170109645Ssimokawa * +-----------+ 171109645Ssimokawa * | l2arc_buf_hdr_t 172109645Ssimokawa * | | 173109645Ssimokawa * +-----------+ 174109645Ssimokawa * | l1arc_buf_hdr_t 175109645Ssimokawa * | | arc_buf_t 176109645Ssimokawa * | b_buf +------------>+-----------+ arc_buf_t 177109645Ssimokawa * | b_pabd +-+ |b_next +---->+-----------+ 178109645Ssimokawa * +-----------+ | |-----------| |b_next +-->NULL 179109645Ssimokawa * | |b_comp = T | +-----------+ 180124169Ssimokawa * | |b_data +-+ |b_comp = F | 181118293Ssimokawa * | +-----------+ | |b_data +-+ 182169130Ssimokawa * +->+------+ | +-----------+ | 183109645Ssimokawa * compressed | | | | 184109645Ssimokawa * data | |<--------------+ | uncompressed 185109645Ssimokawa * +------+ compressed, | data 186113584Ssimokawa * shared +-->+------+ 187109645Ssimokawa * data | | 188109645Ssimokawa * | | 189109645Ssimokawa * +------+ 190109645Ssimokawa * 191109645Ssimokawa * When a consumer reads a block, the ARC must first look to see if the 192109890Ssimokawa * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new 193109645Ssimokawa * arc_buf_t and either copies uncompressed data into a new data buffer from an 194109645Ssimokawa * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a 195109645Ssimokawa * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the 196124169Ssimokawa * hdr is compressed and the desired compression characteristics of the 197109645Ssimokawa * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the 198109645Ssimokawa * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be 199109645Ssimokawa * the last buffer in the hdr's b_buf list, however a shared compressed buf can 200113584Ssimokawa * be anywhere in the hdr's list. 201111942Ssimokawa * 202109645Ssimokawa * The diagram below shows an example of an uncompressed ARC hdr that is 203109645Ssimokawa * sharing its data with an arc_buf_t (note that the shared uncompressed buf is 204109645Ssimokawa * the last element in the buf list): 205111942Ssimokawa * 206109645Ssimokawa * arc_buf_hdr_t 207109645Ssimokawa * +-----------+ 208109645Ssimokawa * | | 209120660Ssimokawa * | | 210120660Ssimokawa * | | 211169130Ssimokawa * +-----------+ 212109645Ssimokawa * l2arc_buf_hdr_t| | 213109645Ssimokawa * | | 214169130Ssimokawa * +-----------+ 215109645Ssimokawa * l1arc_buf_hdr_t| | 216109645Ssimokawa * | | arc_buf_t (shared) 217103285Sikob * | b_buf +------------>+---------+ arc_buf_t 218103285Sikob * | | |b_next +---->+---------+ 219103285Sikob * | b_pabd +-+ |---------| |b_next +-->NULL 220103285Sikob * +-----------+ | | | +---------+ 221110577Ssimokawa * | |b_data +-+ | | 222113584Ssimokawa * | +---------+ | |b_data +-+ 223103285Sikob * +->+------+ | +---------+ | 224103285Sikob * | | | | 225103285Sikob * uncompressed | | | | 226103285Sikob * data +------+ | | 227103285Sikob * ^ +->+------+ | 228103285Sikob * | uncompressed | | | 229103285Sikob * | data | | | 230129585Sdfr * | +------+ | 231169119Ssimokawa * +---------------------------------+ 232167632Ssimokawa * 233103285Sikob * Writing to the ARC requires that the ARC first discard the hdr's b_pabd 234120660Ssimokawa * since the physical block is about to be rewritten. The new data contents 235129585Sdfr * will be contained in the arc_buf_t. As the I/O pipeline performs the write, 236129585Sdfr * it may compress the data before writing it to disk. The ARC will be called 237129585Sdfr * with the transformed data and will bcopy the transformed on-disk block into 238103285Sikob * a newly allocated b_pabd. Writes are always done into buffers which have 239103285Sikob * either been loaned (and hence are new and don't have other readers) or 240103285Sikob * buffers which have been released (and hence have their own hdr, if there 241169119Ssimokawa * were originally other readers of the buf's original hdr). This ensures that 242110269Ssimokawa * the ARC only needs to update a single buf and its hdr after a write occurs. 243103285Sikob * 244120660Ssimokawa * When the L2ARC is in use, it will also take advantage of the b_pabd. The 245120660Ssimokawa * L2ARC will always write the contents of b_pabd to the L2ARC. This means 246120660Ssimokawa * that when compressed ARC is enabled that the L2ARC blocks are identical 247120660Ssimokawa * to the on-disk block in the main data pool. This provides a significant 248120660Ssimokawa * advantage since the ARC can leverage the bp's checksum when reading from the 249120660Ssimokawa * L2ARC to determine if the contents are valid. However, if the compressed 250129585Sdfr * ARC is disabled, then the L2ARC's block must be transformed to look 251120660Ssimokawa * like the physical block in the main data pool before comparing the 252120660Ssimokawa * checksum and determining its validity. 253129585Sdfr */ 254124169Ssimokawa 255124169Ssimokawa#include <sys/spa.h> 256124169Ssimokawa#include <sys/zio.h> 257124169Ssimokawa#include <sys/spa_impl.h> 258124169Ssimokawa#include <sys/zio_compress.h> 259124169Ssimokawa#include <sys/zio_checksum.h> 260124169Ssimokawa#include <sys/zfs_context.h> 261124169Ssimokawa#include <sys/arc.h> 262124169Ssimokawa#include <sys/refcount.h> 263124169Ssimokawa#include <sys/vdev.h> 264124169Ssimokawa#include <sys/vdev_impl.h> 265169130Ssimokawa#include <sys/dsl_pool.h> 266169130Ssimokawa#include <sys/zio_checksum.h> 267169130Ssimokawa#include <sys/multilist.h> 268124169Ssimokawa#include <sys/abd.h> 269169117Ssimokawa#ifdef _KERNEL 270129585Sdfr#include <sys/dnlc.h> 271124169Ssimokawa#include <sys/racct.h> 272124169Ssimokawa#endif 273124169Ssimokawa#include <sys/callb.h> 274124169Ssimokawa#include <sys/kstat.h> 275124169Ssimokawa#include <sys/trim_map.h> 276124169Ssimokawa#include <zfs_fletcher.h> 277129585Sdfr#include <sys/sdt.h> 278124169Ssimokawa#include <sys/aggsum.h> 279124169Ssimokawa#include <sys/cityhash.h> 280124169Ssimokawa 281148868Srwatson#include <machine/vmparam.h> 282103285Sikob 283103285Sikob#ifdef illumos 284103285Sikob#ifndef _KERNEL 285103285Sikob/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 286127468Ssimokawaboolean_t arc_watch = B_FALSE; 287127468Ssimokawaint arc_procfd; 288127468Ssimokawa#endif 289103285Sikob#endif /* illumos */ 290127468Ssimokawa 291103285Sikobstatic kmutex_t arc_reclaim_lock; 292127468Ssimokawastatic kcondvar_t arc_reclaim_thread_cv; 293127468Ssimokawastatic boolean_t arc_reclaim_thread_exit; 294127468Ssimokawastatic kcondvar_t arc_reclaim_waiters_cv; 295111615Ssimokawa 296111615Ssimokawastatic kmutex_t arc_dnlc_evicts_lock; 297111615Ssimokawastatic kcondvar_t arc_dnlc_evicts_cv; 298127468Ssimokawastatic boolean_t arc_dnlc_evicts_thread_exit; 299120660Ssimokawa 300120660Ssimokawauint_t arc_reduce_dnlc_percent = 3; 301120660Ssimokawa 302120660Ssimokawa/* 303120660Ssimokawa * The number of headers to evict in arc_evict_state_impl() before 304120660Ssimokawa * dropping the sublist lock and evicting from another sublist. A lower 305120660Ssimokawa * value means we're more likely to evict the "correct" header (i.e. the 306120660Ssimokawa * oldest header in the arc state), but comes with higher overhead 307120660Ssimokawa * (i.e. more invocations of arc_evict_state_impl()). 308120660Ssimokawa */ 309120660Ssimokawaint zfs_arc_evict_batch_limit = 10; 310120660Ssimokawa 311120660Ssimokawa/* number of seconds before growing cache again */ 312120660Ssimokawastatic int arc_grow_retry = 60; 313120660Ssimokawa 314121780Ssimokawa/* number of milliseconds before attempting a kmem-cache-reap */ 315120660Ssimokawastatic int arc_kmem_cache_reap_retry_ms = 1000; 316120660Ssimokawa 317110195Ssimokawa/* shift of arc_c for calculating overflow limit in arc_get_data_impl */ 318110269Ssimokawaint zfs_arc_overflow_shift = 8; 319 320/* shift of arc_c for calculating both min and max arc_p */ 321static int arc_p_min_shift = 4; 322 323/* log2(fraction of arc to reclaim) */ 324static int arc_shrink_shift = 7; 325 326/* 327 * log2(fraction of ARC which must be free to allow growing). 328 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 329 * when reading a new block into the ARC, we will evict an equal-sized block 330 * from the ARC. 331 * 332 * This must be less than arc_shrink_shift, so that when we shrink the ARC, 333 * we will still not allow it to grow. 334 */ 335int arc_no_grow_shift = 5; 336 337 338/* 339 * minimum lifespan of a prefetch block in clock ticks 340 * (initialized in arc_init()) 341 */ 342static int arc_min_prefetch_lifespan; 343 344/* 345 * If this percent of memory is free, don't throttle. 346 */ 347int arc_lotsfree_percent = 10; 348 349static int arc_dead; 350extern boolean_t zfs_prefetch_disable; 351 352/* 353 * The arc has filled available memory and has now warmed up. 354 */ 355static boolean_t arc_warm; 356 357/* 358 * log2 fraction of the zio arena to keep free. 359 */ 360int arc_zio_arena_free_shift = 2; 361 362/* 363 * These tunables are for performance analysis. 364 */ 365uint64_t zfs_arc_max; 366uint64_t zfs_arc_min; 367uint64_t zfs_arc_meta_limit = 0; 368uint64_t zfs_arc_meta_min = 0; 369int zfs_arc_grow_retry = 0; 370int zfs_arc_shrink_shift = 0; 371int zfs_arc_no_grow_shift = 0; 372int zfs_arc_p_min_shift = 0; 373uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 374u_int zfs_arc_free_target = 0; 375 376/* Absolute min for arc min / max is 16MB. */ 377static uint64_t arc_abs_min = 16 << 20; 378 379boolean_t zfs_compressed_arc_enabled = B_TRUE; 380 381static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 382static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 383static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS); 384static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS); 385static int sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS); 386 387#if defined(__FreeBSD__) && defined(_KERNEL) 388static void 389arc_free_target_init(void *unused __unused) 390{ 391 392 zfs_arc_free_target = vm_pageout_wakeup_thresh; 393} 394SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 395 arc_free_target_init, NULL); 396 397TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 398TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 399TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 400TUNABLE_INT("vfs.zfs.arc_grow_retry", &zfs_arc_grow_retry); 401TUNABLE_INT("vfs.zfs.arc_no_grow_shift", &zfs_arc_no_grow_shift); 402SYSCTL_DECL(_vfs_zfs); 403SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN, 404 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size"); 405SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN, 406 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size"); 407SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_no_grow_shift, CTLTYPE_U32 | CTLFLAG_RWTUN, 408 0, sizeof(uint32_t), sysctl_vfs_zfs_arc_no_grow_shift, "U", 409 "log2(fraction of ARC which must be free to allow growing)"); 410SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 411 &zfs_arc_average_blocksize, 0, 412 "ARC average blocksize"); 413SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 414 &arc_shrink_shift, 0, 415 "log2(fraction of arc to reclaim)"); 416SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_grow_retry, CTLFLAG_RW, 417 &arc_grow_retry, 0, 418 "Wait in seconds before considering growing ARC"); 419SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN, 420 &zfs_compressed_arc_enabled, 0, "Enable compressed ARC"); 421 422/* 423 * We don't have a tunable for arc_free_target due to the dependency on 424 * pagedaemon initialisation. 425 */ 426SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 427 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 428 sysctl_vfs_zfs_arc_free_target, "IU", 429 "Desired number of free pages below which ARC triggers reclaim"); 430 431static int 432sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 433{ 434 u_int val; 435 int err; 436 437 val = zfs_arc_free_target; 438 err = sysctl_handle_int(oidp, &val, 0, req); 439 if (err != 0 || req->newptr == NULL) 440 return (err); 441 442 if (val < minfree) 443 return (EINVAL); 444 if (val > vm_cnt.v_page_count) 445 return (EINVAL); 446 447 zfs_arc_free_target = val; 448 449 return (0); 450} 451 452/* 453 * Must be declared here, before the definition of corresponding kstat 454 * macro which uses the same names will confuse the compiler. 455 */ 456SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 457 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 458 sysctl_vfs_zfs_arc_meta_limit, "QU", 459 "ARC metadata limit"); 460#endif 461 462/* 463 * Note that buffers can be in one of 6 states: 464 * ARC_anon - anonymous (discussed below) 465 * ARC_mru - recently used, currently cached 466 * ARC_mru_ghost - recentely used, no longer in cache 467 * ARC_mfu - frequently used, currently cached 468 * ARC_mfu_ghost - frequently used, no longer in cache 469 * ARC_l2c_only - exists in L2ARC but not other states 470 * When there are no active references to the buffer, they are 471 * are linked onto a list in one of these arc states. These are 472 * the only buffers that can be evicted or deleted. Within each 473 * state there are multiple lists, one for meta-data and one for 474 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 475 * etc.) is tracked separately so that it can be managed more 476 * explicitly: favored over data, limited explicitly. 477 * 478 * Anonymous buffers are buffers that are not associated with 479 * a DVA. These are buffers that hold dirty block copies 480 * before they are written to stable storage. By definition, 481 * they are "ref'd" and are considered part of arc_mru 482 * that cannot be freed. Generally, they will aquire a DVA 483 * as they are written and migrate onto the arc_mru list. 484 * 485 * The ARC_l2c_only state is for buffers that are in the second 486 * level ARC but no longer in any of the ARC_m* lists. The second 487 * level ARC itself may also contain buffers that are in any of 488 * the ARC_m* states - meaning that a buffer can exist in two 489 * places. The reason for the ARC_l2c_only state is to keep the 490 * buffer header in the hash table, so that reads that hit the 491 * second level ARC benefit from these fast lookups. 492 */ 493 494typedef struct arc_state { 495 /* 496 * list of evictable buffers 497 */ 498 multilist_t *arcs_list[ARC_BUFC_NUMTYPES]; 499 /* 500 * total amount of evictable data in this state 501 */ 502 refcount_t arcs_esize[ARC_BUFC_NUMTYPES]; 503 /* 504 * total amount of data in this state; this includes: evictable, 505 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 506 */ 507 refcount_t arcs_size; 508} arc_state_t; 509 510/* The 6 states: */ 511static arc_state_t ARC_anon; 512static arc_state_t ARC_mru; 513static arc_state_t ARC_mru_ghost; 514static arc_state_t ARC_mfu; 515static arc_state_t ARC_mfu_ghost; 516static arc_state_t ARC_l2c_only; 517 518typedef struct arc_stats { 519 kstat_named_t arcstat_hits; 520 kstat_named_t arcstat_misses; 521 kstat_named_t arcstat_demand_data_hits; 522 kstat_named_t arcstat_demand_data_misses; 523 kstat_named_t arcstat_demand_metadata_hits; 524 kstat_named_t arcstat_demand_metadata_misses; 525 kstat_named_t arcstat_prefetch_data_hits; 526 kstat_named_t arcstat_prefetch_data_misses; 527 kstat_named_t arcstat_prefetch_metadata_hits; 528 kstat_named_t arcstat_prefetch_metadata_misses; 529 kstat_named_t arcstat_mru_hits; 530 kstat_named_t arcstat_mru_ghost_hits; 531 kstat_named_t arcstat_mfu_hits; 532 kstat_named_t arcstat_mfu_ghost_hits; 533 kstat_named_t arcstat_allocated; 534 kstat_named_t arcstat_deleted; 535 /* 536 * Number of buffers that could not be evicted because the hash lock 537 * was held by another thread. The lock may not necessarily be held 538 * by something using the same buffer, since hash locks are shared 539 * by multiple buffers. 540 */ 541 kstat_named_t arcstat_mutex_miss; 542 /* 543 * Number of buffers skipped when updating the access state due to the 544 * header having already been released after acquiring the hash lock. 545 */ 546 kstat_named_t arcstat_access_skip; 547 /* 548 * Number of buffers skipped because they have I/O in progress, are 549 * indirect prefetch buffers that have not lived long enough, or are 550 * not from the spa we're trying to evict from. 551 */ 552 kstat_named_t arcstat_evict_skip; 553 /* 554 * Number of times arc_evict_state() was unable to evict enough 555 * buffers to reach it's target amount. 556 */ 557 kstat_named_t arcstat_evict_not_enough; 558 kstat_named_t arcstat_evict_l2_cached; 559 kstat_named_t arcstat_evict_l2_eligible; 560 kstat_named_t arcstat_evict_l2_ineligible; 561 kstat_named_t arcstat_evict_l2_skip; 562 kstat_named_t arcstat_hash_elements; 563 kstat_named_t arcstat_hash_elements_max; 564 kstat_named_t arcstat_hash_collisions; 565 kstat_named_t arcstat_hash_chains; 566 kstat_named_t arcstat_hash_chain_max; 567 kstat_named_t arcstat_p; 568 kstat_named_t arcstat_c; 569 kstat_named_t arcstat_c_min; 570 kstat_named_t arcstat_c_max; 571 /* Not updated directly; only synced in arc_kstat_update. */ 572 kstat_named_t arcstat_size; 573 /* 574 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd. 575 * Note that the compressed bytes may match the uncompressed bytes 576 * if the block is either not compressed or compressed arc is disabled. 577 */ 578 kstat_named_t arcstat_compressed_size; 579 /* 580 * Uncompressed size of the data stored in b_pabd. If compressed 581 * arc is disabled then this value will be identical to the stat 582 * above. 583 */ 584 kstat_named_t arcstat_uncompressed_size; 585 /* 586 * Number of bytes stored in all the arc_buf_t's. This is classified 587 * as "overhead" since this data is typically short-lived and will 588 * be evicted from the arc when it becomes unreferenced unless the 589 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level 590 * values have been set (see comment in dbuf.c for more information). 591 */ 592 kstat_named_t arcstat_overhead_size; 593 /* 594 * Number of bytes consumed by internal ARC structures necessary 595 * for tracking purposes; these structures are not actually 596 * backed by ARC buffers. This includes arc_buf_hdr_t structures 597 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 598 * caches), and arc_buf_t structures (allocated via arc_buf_t 599 * cache). 600 * Not updated directly; only synced in arc_kstat_update. 601 */ 602 kstat_named_t arcstat_hdr_size; 603 /* 604 * Number of bytes consumed by ARC buffers of type equal to 605 * ARC_BUFC_DATA. This is generally consumed by buffers backing 606 * on disk user data (e.g. plain file contents). 607 * Not updated directly; only synced in arc_kstat_update. 608 */ 609 kstat_named_t arcstat_data_size; 610 /* 611 * Number of bytes consumed by ARC buffers of type equal to 612 * ARC_BUFC_METADATA. This is generally consumed by buffers 613 * backing on disk data that is used for internal ZFS 614 * structures (e.g. ZAP, dnode, indirect blocks, etc). 615 * Not updated directly; only synced in arc_kstat_update. 616 */ 617 kstat_named_t arcstat_metadata_size; 618 /* 619 * Number of bytes consumed by various buffers and structures 620 * not actually backed with ARC buffers. This includes bonus 621 * buffers (allocated directly via zio_buf_* functions), 622 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 623 * cache), and dnode_t structures (allocated via dnode_t cache). 624 * Not updated directly; only synced in arc_kstat_update. 625 */ 626 kstat_named_t arcstat_other_size; 627 /* 628 * Total number of bytes consumed by ARC buffers residing in the 629 * arc_anon state. This includes *all* buffers in the arc_anon 630 * state; e.g. data, metadata, evictable, and unevictable buffers 631 * are all included in this value. 632 * Not updated directly; only synced in arc_kstat_update. 633 */ 634 kstat_named_t arcstat_anon_size; 635 /* 636 * Number of bytes consumed by ARC buffers that meet the 637 * following criteria: backing buffers of type ARC_BUFC_DATA, 638 * residing in the arc_anon state, and are eligible for eviction 639 * (e.g. have no outstanding holds on the buffer). 640 * Not updated directly; only synced in arc_kstat_update. 641 */ 642 kstat_named_t arcstat_anon_evictable_data; 643 /* 644 * Number of bytes consumed by ARC buffers that meet the 645 * following criteria: backing buffers of type ARC_BUFC_METADATA, 646 * residing in the arc_anon state, and are eligible for eviction 647 * (e.g. have no outstanding holds on the buffer). 648 * Not updated directly; only synced in arc_kstat_update. 649 */ 650 kstat_named_t arcstat_anon_evictable_metadata; 651 /* 652 * Total number of bytes consumed by ARC buffers residing in the 653 * arc_mru state. This includes *all* buffers in the arc_mru 654 * state; e.g. data, metadata, evictable, and unevictable buffers 655 * are all included in this value. 656 * Not updated directly; only synced in arc_kstat_update. 657 */ 658 kstat_named_t arcstat_mru_size; 659 /* 660 * Number of bytes consumed by ARC buffers that meet the 661 * following criteria: backing buffers of type ARC_BUFC_DATA, 662 * residing in the arc_mru state, and are eligible for eviction 663 * (e.g. have no outstanding holds on the buffer). 664 * Not updated directly; only synced in arc_kstat_update. 665 */ 666 kstat_named_t arcstat_mru_evictable_data; 667 /* 668 * Number of bytes consumed by ARC buffers that meet the 669 * following criteria: backing buffers of type ARC_BUFC_METADATA, 670 * residing in the arc_mru state, and are eligible for eviction 671 * (e.g. have no outstanding holds on the buffer). 672 * Not updated directly; only synced in arc_kstat_update. 673 */ 674 kstat_named_t arcstat_mru_evictable_metadata; 675 /* 676 * Total number of bytes that *would have been* consumed by ARC 677 * buffers in the arc_mru_ghost state. The key thing to note 678 * here, is the fact that this size doesn't actually indicate 679 * RAM consumption. The ghost lists only consist of headers and 680 * don't actually have ARC buffers linked off of these headers. 681 * Thus, *if* the headers had associated ARC buffers, these 682 * buffers *would have* consumed this number of bytes. 683 * Not updated directly; only synced in arc_kstat_update. 684 */ 685 kstat_named_t arcstat_mru_ghost_size; 686 /* 687 * Number of bytes that *would have been* consumed by ARC 688 * buffers that are eligible for eviction, of type 689 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 690 * Not updated directly; only synced in arc_kstat_update. 691 */ 692 kstat_named_t arcstat_mru_ghost_evictable_data; 693 /* 694 * Number of bytes that *would have been* consumed by ARC 695 * buffers that are eligible for eviction, of type 696 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 697 * Not updated directly; only synced in arc_kstat_update. 698 */ 699 kstat_named_t arcstat_mru_ghost_evictable_metadata; 700 /* 701 * Total number of bytes consumed by ARC buffers residing in the 702 * arc_mfu state. This includes *all* buffers in the arc_mfu 703 * state; e.g. data, metadata, evictable, and unevictable buffers 704 * are all included in this value. 705 * Not updated directly; only synced in arc_kstat_update. 706 */ 707 kstat_named_t arcstat_mfu_size; 708 /* 709 * Number of bytes consumed by ARC buffers that are eligible for 710 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 711 * state. 712 * Not updated directly; only synced in arc_kstat_update. 713 */ 714 kstat_named_t arcstat_mfu_evictable_data; 715 /* 716 * Number of bytes consumed by ARC buffers that are eligible for 717 * eviction, of type ARC_BUFC_METADATA, and reside in the 718 * arc_mfu state. 719 * Not updated directly; only synced in arc_kstat_update. 720 */ 721 kstat_named_t arcstat_mfu_evictable_metadata; 722 /* 723 * Total number of bytes that *would have been* consumed by ARC 724 * buffers in the arc_mfu_ghost state. See the comment above 725 * arcstat_mru_ghost_size for more details. 726 * Not updated directly; only synced in arc_kstat_update. 727 */ 728 kstat_named_t arcstat_mfu_ghost_size; 729 /* 730 * Number of bytes that *would have been* consumed by ARC 731 * buffers that are eligible for eviction, of type 732 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 733 * Not updated directly; only synced in arc_kstat_update. 734 */ 735 kstat_named_t arcstat_mfu_ghost_evictable_data; 736 /* 737 * Number of bytes that *would have been* consumed by ARC 738 * buffers that are eligible for eviction, of type 739 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 740 * Not updated directly; only synced in arc_kstat_update. 741 */ 742 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 743 kstat_named_t arcstat_l2_hits; 744 kstat_named_t arcstat_l2_misses; 745 kstat_named_t arcstat_l2_feeds; 746 kstat_named_t arcstat_l2_rw_clash; 747 kstat_named_t arcstat_l2_read_bytes; 748 kstat_named_t arcstat_l2_write_bytes; 749 kstat_named_t arcstat_l2_writes_sent; 750 kstat_named_t arcstat_l2_writes_done; 751 kstat_named_t arcstat_l2_writes_error; 752 kstat_named_t arcstat_l2_writes_lock_retry; 753 kstat_named_t arcstat_l2_evict_lock_retry; 754 kstat_named_t arcstat_l2_evict_reading; 755 kstat_named_t arcstat_l2_evict_l1cached; 756 kstat_named_t arcstat_l2_free_on_write; 757 kstat_named_t arcstat_l2_abort_lowmem; 758 kstat_named_t arcstat_l2_cksum_bad; 759 kstat_named_t arcstat_l2_io_error; 760 kstat_named_t arcstat_l2_lsize; 761 kstat_named_t arcstat_l2_psize; 762 /* Not updated directly; only synced in arc_kstat_update. */ 763 kstat_named_t arcstat_l2_hdr_size; 764 kstat_named_t arcstat_l2_write_trylock_fail; 765 kstat_named_t arcstat_l2_write_passed_headroom; 766 kstat_named_t arcstat_l2_write_spa_mismatch; 767 kstat_named_t arcstat_l2_write_in_l2; 768 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 769 kstat_named_t arcstat_l2_write_not_cacheable; 770 kstat_named_t arcstat_l2_write_full; 771 kstat_named_t arcstat_l2_write_buffer_iter; 772 kstat_named_t arcstat_l2_write_pios; 773 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 774 kstat_named_t arcstat_l2_write_buffer_list_iter; 775 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 776 kstat_named_t arcstat_memory_throttle_count; 777 /* Not updated directly; only synced in arc_kstat_update. */ 778 kstat_named_t arcstat_meta_used; 779 kstat_named_t arcstat_meta_limit; 780 kstat_named_t arcstat_meta_max; 781 kstat_named_t arcstat_meta_min; 782 kstat_named_t arcstat_sync_wait_for_async; 783 kstat_named_t arcstat_demand_hit_predictive_prefetch; 784} arc_stats_t; 785 786static arc_stats_t arc_stats = { 787 { "hits", KSTAT_DATA_UINT64 }, 788 { "misses", KSTAT_DATA_UINT64 }, 789 { "demand_data_hits", KSTAT_DATA_UINT64 }, 790 { "demand_data_misses", KSTAT_DATA_UINT64 }, 791 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 792 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 793 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 794 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 795 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 796 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 797 { "mru_hits", KSTAT_DATA_UINT64 }, 798 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 799 { "mfu_hits", KSTAT_DATA_UINT64 }, 800 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 801 { "allocated", KSTAT_DATA_UINT64 }, 802 { "deleted", KSTAT_DATA_UINT64 }, 803 { "mutex_miss", KSTAT_DATA_UINT64 }, 804 { "access_skip", KSTAT_DATA_UINT64 }, 805 { "evict_skip", KSTAT_DATA_UINT64 }, 806 { "evict_not_enough", KSTAT_DATA_UINT64 }, 807 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 808 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 809 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 810 { "evict_l2_skip", KSTAT_DATA_UINT64 }, 811 { "hash_elements", KSTAT_DATA_UINT64 }, 812 { "hash_elements_max", KSTAT_DATA_UINT64 }, 813 { "hash_collisions", KSTAT_DATA_UINT64 }, 814 { "hash_chains", KSTAT_DATA_UINT64 }, 815 { "hash_chain_max", KSTAT_DATA_UINT64 }, 816 { "p", KSTAT_DATA_UINT64 }, 817 { "c", KSTAT_DATA_UINT64 }, 818 { "c_min", KSTAT_DATA_UINT64 }, 819 { "c_max", KSTAT_DATA_UINT64 }, 820 { "size", KSTAT_DATA_UINT64 }, 821 { "compressed_size", KSTAT_DATA_UINT64 }, 822 { "uncompressed_size", KSTAT_DATA_UINT64 }, 823 { "overhead_size", KSTAT_DATA_UINT64 }, 824 { "hdr_size", KSTAT_DATA_UINT64 }, 825 { "data_size", KSTAT_DATA_UINT64 }, 826 { "metadata_size", KSTAT_DATA_UINT64 }, 827 { "other_size", KSTAT_DATA_UINT64 }, 828 { "anon_size", KSTAT_DATA_UINT64 }, 829 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 830 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 831 { "mru_size", KSTAT_DATA_UINT64 }, 832 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 833 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 834 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 835 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 836 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 837 { "mfu_size", KSTAT_DATA_UINT64 }, 838 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 839 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 840 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 841 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 842 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 843 { "l2_hits", KSTAT_DATA_UINT64 }, 844 { "l2_misses", KSTAT_DATA_UINT64 }, 845 { "l2_feeds", KSTAT_DATA_UINT64 }, 846 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 847 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 848 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 849 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 850 { "l2_writes_done", KSTAT_DATA_UINT64 }, 851 { "l2_writes_error", KSTAT_DATA_UINT64 }, 852 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 853 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 854 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 855 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 856 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 857 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 858 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 859 { "l2_io_error", KSTAT_DATA_UINT64 }, 860 { "l2_size", KSTAT_DATA_UINT64 }, 861 { "l2_asize", KSTAT_DATA_UINT64 }, 862 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 863 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 864 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 865 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 866 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 867 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 868 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 869 { "l2_write_full", KSTAT_DATA_UINT64 }, 870 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 871 { "l2_write_pios", KSTAT_DATA_UINT64 }, 872 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 873 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 874 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 875 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 876 { "arc_meta_used", KSTAT_DATA_UINT64 }, 877 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 878 { "arc_meta_max", KSTAT_DATA_UINT64 }, 879 { "arc_meta_min", KSTAT_DATA_UINT64 }, 880 { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 881 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 882}; 883 884#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 885 886#define ARCSTAT_INCR(stat, val) \ 887 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 888 889#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 890#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 891 892#define ARCSTAT_MAX(stat, val) { \ 893 uint64_t m; \ 894 while ((val) > (m = arc_stats.stat.value.ui64) && \ 895 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 896 continue; \ 897} 898 899#define ARCSTAT_MAXSTAT(stat) \ 900 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 901 902/* 903 * We define a macro to allow ARC hits/misses to be easily broken down by 904 * two separate conditions, giving a total of four different subtypes for 905 * each of hits and misses (so eight statistics total). 906 */ 907#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 908 if (cond1) { \ 909 if (cond2) { \ 910 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 911 } else { \ 912 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 913 } \ 914 } else { \ 915 if (cond2) { \ 916 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 917 } else { \ 918 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 919 } \ 920 } 921 922kstat_t *arc_ksp; 923static arc_state_t *arc_anon; 924static arc_state_t *arc_mru; 925static arc_state_t *arc_mru_ghost; 926static arc_state_t *arc_mfu; 927static arc_state_t *arc_mfu_ghost; 928static arc_state_t *arc_l2c_only; 929 930/* 931 * There are several ARC variables that are critical to export as kstats -- 932 * but we don't want to have to grovel around in the kstat whenever we wish to 933 * manipulate them. For these variables, we therefore define them to be in 934 * terms of the statistic variable. This assures that we are not introducing 935 * the possibility of inconsistency by having shadow copies of the variables, 936 * while still allowing the code to be readable. 937 */ 938#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 939#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 940#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 941#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 942#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 943#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 944#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 945 946/* compressed size of entire arc */ 947#define arc_compressed_size ARCSTAT(arcstat_compressed_size) 948/* uncompressed size of entire arc */ 949#define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size) 950/* number of bytes in the arc from arc_buf_t's */ 951#define arc_overhead_size ARCSTAT(arcstat_overhead_size) 952 953/* 954 * There are also some ARC variables that we want to export, but that are 955 * updated so often that having the canonical representation be the statistic 956 * variable causes a performance bottleneck. We want to use aggsum_t's for these 957 * instead, but still be able to export the kstat in the same way as before. 958 * The solution is to always use the aggsum version, except in the kstat update 959 * callback. 960 */ 961aggsum_t arc_size; 962aggsum_t arc_meta_used; 963aggsum_t astat_data_size; 964aggsum_t astat_metadata_size; 965aggsum_t astat_hdr_size; 966aggsum_t astat_other_size; 967aggsum_t astat_l2_hdr_size; 968 969static int arc_no_grow; /* Don't try to grow cache size */ 970static uint64_t arc_tempreserve; 971static uint64_t arc_loaned_bytes; 972 973typedef struct arc_callback arc_callback_t; 974 975struct arc_callback { 976 void *acb_private; 977 arc_done_func_t *acb_done; 978 arc_buf_t *acb_buf; 979 boolean_t acb_compressed; 980 zio_t *acb_zio_dummy; 981 arc_callback_t *acb_next; 982}; 983 984typedef struct arc_write_callback arc_write_callback_t; 985 986struct arc_write_callback { 987 void *awcb_private; 988 arc_done_func_t *awcb_ready; 989 arc_done_func_t *awcb_children_ready; 990 arc_done_func_t *awcb_physdone; 991 arc_done_func_t *awcb_done; 992 arc_buf_t *awcb_buf; 993}; 994 995/* 996 * ARC buffers are separated into multiple structs as a memory saving measure: 997 * - Common fields struct, always defined, and embedded within it: 998 * - L2-only fields, always allocated but undefined when not in L2ARC 999 * - L1-only fields, only allocated when in L1ARC 1000 * 1001 * Buffer in L1 Buffer only in L2 1002 * +------------------------+ +------------------------+ 1003 * | arc_buf_hdr_t | | arc_buf_hdr_t | 1004 * | | | | 1005 * | | | | 1006 * | | | | 1007 * +------------------------+ +------------------------+ 1008 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 1009 * | (undefined if L1-only) | | | 1010 * +------------------------+ +------------------------+ 1011 * | l1arc_buf_hdr_t | 1012 * | | 1013 * | | 1014 * | | 1015 * | | 1016 * +------------------------+ 1017 * 1018 * Because it's possible for the L2ARC to become extremely large, we can wind 1019 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 1020 * is minimized by only allocating the fields necessary for an L1-cached buffer 1021 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 1022 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 1023 * words in pointers. arc_hdr_realloc() is used to switch a header between 1024 * these two allocation states. 1025 */ 1026typedef struct l1arc_buf_hdr { 1027 kmutex_t b_freeze_lock; 1028 zio_cksum_t *b_freeze_cksum; 1029#ifdef ZFS_DEBUG 1030 /* 1031 * Used for debugging with kmem_flags - by allocating and freeing 1032 * b_thawed when the buffer is thawed, we get a record of the stack 1033 * trace that thawed it. 1034 */ 1035 void *b_thawed; 1036#endif 1037 1038 arc_buf_t *b_buf; 1039 uint32_t b_bufcnt; 1040 /* for waiting on writes to complete */ 1041 kcondvar_t b_cv; 1042 uint8_t b_byteswap; 1043 1044 /* protected by arc state mutex */ 1045 arc_state_t *b_state; 1046 multilist_node_t b_arc_node; 1047 1048 /* updated atomically */ 1049 clock_t b_arc_access; 1050 1051 /* self protecting */ 1052 refcount_t b_refcnt; 1053 1054 arc_callback_t *b_acb; 1055 abd_t *b_pabd; 1056} l1arc_buf_hdr_t; 1057 1058typedef struct l2arc_dev l2arc_dev_t; 1059 1060typedef struct l2arc_buf_hdr { 1061 /* protected by arc_buf_hdr mutex */ 1062 l2arc_dev_t *b_dev; /* L2ARC device */ 1063 uint64_t b_daddr; /* disk address, offset byte */ 1064 1065 list_node_t b_l2node; 1066} l2arc_buf_hdr_t; 1067 1068struct arc_buf_hdr { 1069 /* protected by hash lock */ 1070 dva_t b_dva; 1071 uint64_t b_birth; 1072 1073 arc_buf_contents_t b_type; 1074 arc_buf_hdr_t *b_hash_next; 1075 arc_flags_t b_flags; 1076 1077 /* 1078 * This field stores the size of the data buffer after 1079 * compression, and is set in the arc's zio completion handlers. 1080 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes). 1081 * 1082 * While the block pointers can store up to 32MB in their psize 1083 * field, we can only store up to 32MB minus 512B. This is due 1084 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e. 1085 * a field of zeros represents 512B in the bp). We can't use a 1086 * bias of 1 since we need to reserve a psize of zero, here, to 1087 * represent holes and embedded blocks. 1088 * 1089 * This isn't a problem in practice, since the maximum size of a 1090 * buffer is limited to 16MB, so we never need to store 32MB in 1091 * this field. Even in the upstream illumos code base, the 1092 * maximum size of a buffer is limited to 16MB. 1093 */ 1094 uint16_t b_psize; 1095 1096 /* 1097 * This field stores the size of the data buffer before 1098 * compression, and cannot change once set. It is in units 1099 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes) 1100 */ 1101 uint16_t b_lsize; /* immutable */ 1102 uint64_t b_spa; /* immutable */ 1103 1104 /* L2ARC fields. Undefined when not in L2ARC. */ 1105 l2arc_buf_hdr_t b_l2hdr; 1106 /* L1ARC fields. Undefined when in l2arc_only state */ 1107 l1arc_buf_hdr_t b_l1hdr; 1108}; 1109 1110#if defined(__FreeBSD__) && defined(_KERNEL) 1111static int 1112sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 1113{ 1114 uint64_t val; 1115 int err; 1116 1117 val = arc_meta_limit; 1118 err = sysctl_handle_64(oidp, &val, 0, req); 1119 if (err != 0 || req->newptr == NULL) 1120 return (err); 1121 1122 if (val <= 0 || val > arc_c_max) 1123 return (EINVAL); 1124 1125 arc_meta_limit = val; 1126 return (0); 1127} 1128 1129static int 1130sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS) 1131{ 1132 uint32_t val; 1133 int err; 1134 1135 val = arc_no_grow_shift; 1136 err = sysctl_handle_32(oidp, &val, 0, req); 1137 if (err != 0 || req->newptr == NULL) 1138 return (err); 1139 1140 if (val >= arc_shrink_shift) 1141 return (EINVAL); 1142 1143 arc_no_grow_shift = val; 1144 return (0); 1145} 1146 1147static int 1148sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS) 1149{ 1150 uint64_t val; 1151 int err; 1152 1153 val = zfs_arc_max; 1154 err = sysctl_handle_64(oidp, &val, 0, req); 1155 if (err != 0 || req->newptr == NULL) 1156 return (err); 1157 1158 if (zfs_arc_max == 0) { 1159 /* Loader tunable so blindly set */ 1160 zfs_arc_max = val; 1161 return (0); 1162 } 1163 1164 if (val < arc_abs_min || val > kmem_size()) 1165 return (EINVAL); 1166 if (val < arc_c_min) 1167 return (EINVAL); 1168 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit) 1169 return (EINVAL); 1170 1171 arc_c_max = val; 1172 1173 arc_c = arc_c_max; 1174 arc_p = (arc_c >> 1); 1175 1176 if (zfs_arc_meta_limit == 0) { 1177 /* limit meta-data to 1/4 of the arc capacity */ 1178 arc_meta_limit = arc_c_max / 4; 1179 } 1180 1181 /* if kmem_flags are set, lets try to use less memory */ 1182 if (kmem_debugging()) 1183 arc_c = arc_c / 2; 1184 1185 zfs_arc_max = arc_c; 1186 1187 return (0); 1188} 1189 1190static int 1191sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS) 1192{ 1193 uint64_t val; 1194 int err; 1195 1196 val = zfs_arc_min; 1197 err = sysctl_handle_64(oidp, &val, 0, req); 1198 if (err != 0 || req->newptr == NULL) 1199 return (err); 1200 1201 if (zfs_arc_min == 0) { 1202 /* Loader tunable so blindly set */ 1203 zfs_arc_min = val; 1204 return (0); 1205 } 1206 1207 if (val < arc_abs_min || val > arc_c_max) 1208 return (EINVAL); 1209 1210 arc_c_min = val; 1211 1212 if (zfs_arc_meta_min == 0) 1213 arc_meta_min = arc_c_min / 2; 1214 1215 if (arc_c < arc_c_min) 1216 arc_c = arc_c_min; 1217 1218 zfs_arc_min = arc_c_min; 1219 1220 return (0); 1221} 1222#endif 1223 1224#define GHOST_STATE(state) \ 1225 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 1226 (state) == arc_l2c_only) 1227 1228#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 1229#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 1230#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 1231#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 1232#define HDR_COMPRESSION_ENABLED(hdr) \ 1233 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC) 1234 1235#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 1236#define HDR_L2_READING(hdr) \ 1237 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 1238 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 1239#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 1240#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 1241#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 1242#define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA) 1243 1244#define HDR_ISTYPE_METADATA(hdr) \ 1245 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 1246#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 1247 1248#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 1249#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 1250 1251/* For storing compression mode in b_flags */ 1252#define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1) 1253 1254#define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \ 1255 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS)) 1256#define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \ 1257 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp)); 1258 1259#define ARC_BUF_LAST(buf) ((buf)->b_next == NULL) 1260#define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED) 1261#define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED) 1262 1263/* 1264 * Other sizes 1265 */ 1266 1267#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 1268#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 1269 1270/* 1271 * Hash table routines 1272 */ 1273 1274#define HT_LOCK_PAD CACHE_LINE_SIZE 1275 1276struct ht_lock { 1277 kmutex_t ht_lock; 1278#ifdef _KERNEL 1279 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 1280#endif 1281}; 1282 1283#define BUF_LOCKS 256 1284typedef struct buf_hash_table { 1285 uint64_t ht_mask; 1286 arc_buf_hdr_t **ht_table; 1287 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 1288} buf_hash_table_t; 1289 1290static buf_hash_table_t buf_hash_table; 1291 1292#define BUF_HASH_INDEX(spa, dva, birth) \ 1293 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 1294#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 1295#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 1296#define HDR_LOCK(hdr) \ 1297 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 1298 1299uint64_t zfs_crc64_table[256]; 1300 1301/* 1302 * Level 2 ARC 1303 */ 1304 1305#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 1306#define L2ARC_HEADROOM 2 /* num of writes */ 1307/* 1308 * If we discover during ARC scan any buffers to be compressed, we boost 1309 * our headroom for the next scanning cycle by this percentage multiple. 1310 */ 1311#define L2ARC_HEADROOM_BOOST 200 1312#define L2ARC_FEED_SECS 1 /* caching interval secs */ 1313#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 1314 1315#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 1316#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 1317 1318/* L2ARC Performance Tunables */ 1319uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1320uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1321uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1322uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1323uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1324uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1325boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1326boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1327boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1328 1329SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 1330 &l2arc_write_max, 0, "max write size"); 1331SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 1332 &l2arc_write_boost, 0, "extra write during warmup"); 1333SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 1334 &l2arc_headroom, 0, "number of dev writes"); 1335SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 1336 &l2arc_feed_secs, 0, "interval seconds"); 1337SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 1338 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 1339 1340SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 1341 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 1342SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 1343 &l2arc_feed_again, 0, "turbo warmup"); 1344SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 1345 &l2arc_norw, 0, "no reads during writes"); 1346 1347SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 1348 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state"); 1349SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD, 1350 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1351 "size of anonymous state"); 1352SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD, 1353 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1354 "size of anonymous state"); 1355 1356SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 1357 &ARC_mru.arcs_size.rc_count, 0, "size of mru state"); 1358SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD, 1359 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1360 "size of metadata in mru state"); 1361SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD, 1362 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1363 "size of data in mru state"); 1364 1365SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 1366 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state"); 1367SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD, 1368 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1369 "size of metadata in mru ghost state"); 1370SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD, 1371 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1372 "size of data in mru ghost state"); 1373 1374SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 1375 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state"); 1376SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD, 1377 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1378 "size of metadata in mfu state"); 1379SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD, 1380 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1381 "size of data in mfu state"); 1382 1383SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 1384 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state"); 1385SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD, 1386 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1387 "size of metadata in mfu ghost state"); 1388SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD, 1389 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1390 "size of data in mfu ghost state"); 1391 1392SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 1393 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state"); 1394 1395/* 1396 * L2ARC Internals 1397 */ 1398struct l2arc_dev { 1399 vdev_t *l2ad_vdev; /* vdev */ 1400 spa_t *l2ad_spa; /* spa */ 1401 uint64_t l2ad_hand; /* next write location */ 1402 uint64_t l2ad_start; /* first addr on device */ 1403 uint64_t l2ad_end; /* last addr on device */ 1404 boolean_t l2ad_first; /* first sweep through */ 1405 boolean_t l2ad_writing; /* currently writing */ 1406 kmutex_t l2ad_mtx; /* lock for buffer list */ 1407 list_t l2ad_buflist; /* buffer list */ 1408 list_node_t l2ad_node; /* device list node */ 1409 refcount_t l2ad_alloc; /* allocated bytes */ 1410}; 1411 1412static list_t L2ARC_dev_list; /* device list */ 1413static list_t *l2arc_dev_list; /* device list pointer */ 1414static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1415static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1416static list_t L2ARC_free_on_write; /* free after write buf list */ 1417static list_t *l2arc_free_on_write; /* free after write list ptr */ 1418static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1419static uint64_t l2arc_ndev; /* number of devices */ 1420 1421typedef struct l2arc_read_callback { 1422 arc_buf_hdr_t *l2rcb_hdr; /* read header */ 1423 blkptr_t l2rcb_bp; /* original blkptr */ 1424 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1425 int l2rcb_flags; /* original flags */ 1426 abd_t *l2rcb_abd; /* temporary buffer */ 1427} l2arc_read_callback_t; 1428 1429typedef struct l2arc_write_callback { 1430 l2arc_dev_t *l2wcb_dev; /* device info */ 1431 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1432} l2arc_write_callback_t; 1433 1434typedef struct l2arc_data_free { 1435 /* protected by l2arc_free_on_write_mtx */ 1436 abd_t *l2df_abd; 1437 size_t l2df_size; 1438 arc_buf_contents_t l2df_type; 1439 list_node_t l2df_list_node; 1440} l2arc_data_free_t; 1441 1442static kmutex_t l2arc_feed_thr_lock; 1443static kcondvar_t l2arc_feed_thr_cv; 1444static uint8_t l2arc_thread_exit; 1445 1446static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *); 1447static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); 1448static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *); 1449static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *); 1450static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); 1451static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag); 1452static void arc_hdr_free_pabd(arc_buf_hdr_t *); 1453static void arc_hdr_alloc_pabd(arc_buf_hdr_t *); 1454static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1455static boolean_t arc_is_overflowing(); 1456static void arc_buf_watch(arc_buf_t *); 1457 1458static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1459static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1460static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1461static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1462 1463static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1464static void l2arc_read_done(zio_t *); 1465 1466static void 1467l2arc_trim(const arc_buf_hdr_t *hdr) 1468{ 1469 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1470 1471 ASSERT(HDR_HAS_L2HDR(hdr)); 1472 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 1473 1474 if (HDR_GET_PSIZE(hdr) != 0) { 1475 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr, 1476 HDR_GET_PSIZE(hdr), 0); 1477 } 1478} 1479 1480/* 1481 * We use Cityhash for this. It's fast, and has good hash properties without 1482 * requiring any large static buffers. 1483 */ 1484static uint64_t 1485buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1486{ 1487 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth)); 1488} 1489 1490#define HDR_EMPTY(hdr) \ 1491 ((hdr)->b_dva.dva_word[0] == 0 && \ 1492 (hdr)->b_dva.dva_word[1] == 0) 1493 1494#define HDR_EQUAL(spa, dva, birth, hdr) \ 1495 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1496 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1497 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) 1498 1499static void 1500buf_discard_identity(arc_buf_hdr_t *hdr) 1501{ 1502 hdr->b_dva.dva_word[0] = 0; 1503 hdr->b_dva.dva_word[1] = 0; 1504 hdr->b_birth = 0; 1505} 1506 1507static arc_buf_hdr_t * 1508buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1509{ 1510 const dva_t *dva = BP_IDENTITY(bp); 1511 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1512 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1513 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1514 arc_buf_hdr_t *hdr; 1515 1516 mutex_enter(hash_lock); 1517 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1518 hdr = hdr->b_hash_next) { 1519 if (HDR_EQUAL(spa, dva, birth, hdr)) { 1520 *lockp = hash_lock; 1521 return (hdr); 1522 } 1523 } 1524 mutex_exit(hash_lock); 1525 *lockp = NULL; 1526 return (NULL); 1527} 1528 1529/* 1530 * Insert an entry into the hash table. If there is already an element 1531 * equal to elem in the hash table, then the already existing element 1532 * will be returned and the new element will not be inserted. 1533 * Otherwise returns NULL. 1534 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1535 */ 1536static arc_buf_hdr_t * 1537buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1538{ 1539 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1540 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1541 arc_buf_hdr_t *fhdr; 1542 uint32_t i; 1543 1544 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1545 ASSERT(hdr->b_birth != 0); 1546 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1547 1548 if (lockp != NULL) { 1549 *lockp = hash_lock; 1550 mutex_enter(hash_lock); 1551 } else { 1552 ASSERT(MUTEX_HELD(hash_lock)); 1553 } 1554 1555 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1556 fhdr = fhdr->b_hash_next, i++) { 1557 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1558 return (fhdr); 1559 } 1560 1561 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1562 buf_hash_table.ht_table[idx] = hdr; 1563 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1564 1565 /* collect some hash table performance data */ 1566 if (i > 0) { 1567 ARCSTAT_BUMP(arcstat_hash_collisions); 1568 if (i == 1) 1569 ARCSTAT_BUMP(arcstat_hash_chains); 1570 1571 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1572 } 1573 1574 ARCSTAT_BUMP(arcstat_hash_elements); 1575 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1576 1577 return (NULL); 1578} 1579 1580static void 1581buf_hash_remove(arc_buf_hdr_t *hdr) 1582{ 1583 arc_buf_hdr_t *fhdr, **hdrp; 1584 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1585 1586 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1587 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1588 1589 hdrp = &buf_hash_table.ht_table[idx]; 1590 while ((fhdr = *hdrp) != hdr) { 1591 ASSERT3P(fhdr, !=, NULL); 1592 hdrp = &fhdr->b_hash_next; 1593 } 1594 *hdrp = hdr->b_hash_next; 1595 hdr->b_hash_next = NULL; 1596 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1597 1598 /* collect some hash table performance data */ 1599 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1600 1601 if (buf_hash_table.ht_table[idx] && 1602 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1603 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1604} 1605 1606/* 1607 * Global data structures and functions for the buf kmem cache. 1608 */ 1609static kmem_cache_t *hdr_full_cache; 1610static kmem_cache_t *hdr_l2only_cache; 1611static kmem_cache_t *buf_cache; 1612 1613static void 1614buf_fini(void) 1615{ 1616 int i; 1617 1618 kmem_free(buf_hash_table.ht_table, 1619 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1620 for (i = 0; i < BUF_LOCKS; i++) 1621 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1622 kmem_cache_destroy(hdr_full_cache); 1623 kmem_cache_destroy(hdr_l2only_cache); 1624 kmem_cache_destroy(buf_cache); 1625} 1626 1627/* 1628 * Constructor callback - called when the cache is empty 1629 * and a new buf is requested. 1630 */ 1631/* ARGSUSED */ 1632static int 1633hdr_full_cons(void *vbuf, void *unused, int kmflag) 1634{ 1635 arc_buf_hdr_t *hdr = vbuf; 1636 1637 bzero(hdr, HDR_FULL_SIZE); 1638 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1639 refcount_create(&hdr->b_l1hdr.b_refcnt); 1640 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1641 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1642 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1643 1644 return (0); 1645} 1646 1647/* ARGSUSED */ 1648static int 1649hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1650{ 1651 arc_buf_hdr_t *hdr = vbuf; 1652 1653 bzero(hdr, HDR_L2ONLY_SIZE); 1654 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1655 1656 return (0); 1657} 1658 1659/* ARGSUSED */ 1660static int 1661buf_cons(void *vbuf, void *unused, int kmflag) 1662{ 1663 arc_buf_t *buf = vbuf; 1664 1665 bzero(buf, sizeof (arc_buf_t)); 1666 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1667 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1668 1669 return (0); 1670} 1671 1672/* 1673 * Destructor callback - called when a cached buf is 1674 * no longer required. 1675 */ 1676/* ARGSUSED */ 1677static void 1678hdr_full_dest(void *vbuf, void *unused) 1679{ 1680 arc_buf_hdr_t *hdr = vbuf; 1681 1682 ASSERT(HDR_EMPTY(hdr)); 1683 cv_destroy(&hdr->b_l1hdr.b_cv); 1684 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1685 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1686 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1687 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1688} 1689 1690/* ARGSUSED */ 1691static void 1692hdr_l2only_dest(void *vbuf, void *unused) 1693{ 1694 arc_buf_hdr_t *hdr = vbuf; 1695 1696 ASSERT(HDR_EMPTY(hdr)); 1697 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1698} 1699 1700/* ARGSUSED */ 1701static void 1702buf_dest(void *vbuf, void *unused) 1703{ 1704 arc_buf_t *buf = vbuf; 1705 1706 mutex_destroy(&buf->b_evict_lock); 1707 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1708} 1709 1710/* 1711 * Reclaim callback -- invoked when memory is low. 1712 */ 1713/* ARGSUSED */ 1714static void 1715hdr_recl(void *unused) 1716{ 1717 dprintf("hdr_recl called\n"); 1718 /* 1719 * umem calls the reclaim func when we destroy the buf cache, 1720 * which is after we do arc_fini(). 1721 */ 1722 if (!arc_dead) 1723 cv_signal(&arc_reclaim_thread_cv); 1724} 1725 1726static void 1727buf_init(void) 1728{ 1729 uint64_t *ct; 1730 uint64_t hsize = 1ULL << 12; 1731 int i, j; 1732 1733 /* 1734 * The hash table is big enough to fill all of physical memory 1735 * with an average block size of zfs_arc_average_blocksize (default 8K). 1736 * By default, the table will take up 1737 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1738 */ 1739 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1740 hsize <<= 1; 1741retry: 1742 buf_hash_table.ht_mask = hsize - 1; 1743 buf_hash_table.ht_table = 1744 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1745 if (buf_hash_table.ht_table == NULL) { 1746 ASSERT(hsize > (1ULL << 8)); 1747 hsize >>= 1; 1748 goto retry; 1749 } 1750 1751 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1752 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1753 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1754 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1755 NULL, NULL, 0); 1756 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1757 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1758 1759 for (i = 0; i < 256; i++) 1760 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1761 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1762 1763 for (i = 0; i < BUF_LOCKS; i++) { 1764 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1765 NULL, MUTEX_DEFAULT, NULL); 1766 } 1767} 1768 1769/* 1770 * This is the size that the buf occupies in memory. If the buf is compressed, 1771 * it will correspond to the compressed size. You should use this method of 1772 * getting the buf size unless you explicitly need the logical size. 1773 */ 1774int32_t 1775arc_buf_size(arc_buf_t *buf) 1776{ 1777 return (ARC_BUF_COMPRESSED(buf) ? 1778 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr)); 1779} 1780 1781int32_t 1782arc_buf_lsize(arc_buf_t *buf) 1783{ 1784 return (HDR_GET_LSIZE(buf->b_hdr)); 1785} 1786 1787enum zio_compress 1788arc_get_compression(arc_buf_t *buf) 1789{ 1790 return (ARC_BUF_COMPRESSED(buf) ? 1791 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF); 1792} 1793 1794#define ARC_MINTIME (hz>>4) /* 62 ms */ 1795 1796static inline boolean_t 1797arc_buf_is_shared(arc_buf_t *buf) 1798{ 1799 boolean_t shared = (buf->b_data != NULL && 1800 buf->b_hdr->b_l1hdr.b_pabd != NULL && 1801 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) && 1802 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd)); 1803 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); 1804 IMPLY(shared, ARC_BUF_SHARED(buf)); 1805 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf)); 1806 1807 /* 1808 * It would be nice to assert arc_can_share() too, but the "hdr isn't 1809 * already being shared" requirement prevents us from doing that. 1810 */ 1811 1812 return (shared); 1813} 1814 1815/* 1816 * Free the checksum associated with this header. If there is no checksum, this 1817 * is a no-op. 1818 */ 1819static inline void 1820arc_cksum_free(arc_buf_hdr_t *hdr) 1821{ 1822 ASSERT(HDR_HAS_L1HDR(hdr)); 1823 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1824 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1825 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); 1826 hdr->b_l1hdr.b_freeze_cksum = NULL; 1827 } 1828 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1829} 1830 1831/* 1832 * Return true iff at least one of the bufs on hdr is not compressed. 1833 */ 1834static boolean_t 1835arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr) 1836{ 1837 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) { 1838 if (!ARC_BUF_COMPRESSED(b)) { 1839 return (B_TRUE); 1840 } 1841 } 1842 return (B_FALSE); 1843} 1844 1845/* 1846 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data 1847 * matches the checksum that is stored in the hdr. If there is no checksum, 1848 * or if the buf is compressed, this is a no-op. 1849 */ 1850static void 1851arc_cksum_verify(arc_buf_t *buf) 1852{ 1853 arc_buf_hdr_t *hdr = buf->b_hdr; 1854 zio_cksum_t zc; 1855 1856 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1857 return; 1858 1859 if (ARC_BUF_COMPRESSED(buf)) { 1860 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1861 arc_hdr_has_uncompressed_buf(hdr)); 1862 return; 1863 } 1864 1865 ASSERT(HDR_HAS_L1HDR(hdr)); 1866 1867 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1868 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { 1869 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1870 return; 1871 } 1872 1873 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc); 1874 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) 1875 panic("buffer modified while frozen!"); 1876 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1877} 1878 1879static boolean_t 1880arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) 1881{ 1882 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp); 1883 boolean_t valid_cksum; 1884 1885 ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); 1886 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); 1887 1888 /* 1889 * We rely on the blkptr's checksum to determine if the block 1890 * is valid or not. When compressed arc is enabled, the l2arc 1891 * writes the block to the l2arc just as it appears in the pool. 1892 * This allows us to use the blkptr's checksum to validate the 1893 * data that we just read off of the l2arc without having to store 1894 * a separate checksum in the arc_buf_hdr_t. However, if compressed 1895 * arc is disabled, then the data written to the l2arc is always 1896 * uncompressed and won't match the block as it exists in the main 1897 * pool. When this is the case, we must first compress it if it is 1898 * compressed on the main pool before we can validate the checksum. 1899 */ 1900 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) { 1901 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1902 uint64_t lsize = HDR_GET_LSIZE(hdr); 1903 uint64_t csize; 1904 1905 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE); 1906 csize = zio_compress_data(compress, zio->io_abd, 1907 abd_to_buf(cdata), lsize); 1908 1909 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr)); 1910 if (csize < HDR_GET_PSIZE(hdr)) { 1911 /* 1912 * Compressed blocks are always a multiple of the 1913 * smallest ashift in the pool. Ideally, we would 1914 * like to round up the csize to the next 1915 * spa_min_ashift but that value may have changed 1916 * since the block was last written. Instead, 1917 * we rely on the fact that the hdr's psize 1918 * was set to the psize of the block when it was 1919 * last written. We set the csize to that value 1920 * and zero out any part that should not contain 1921 * data. 1922 */ 1923 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize); 1924 csize = HDR_GET_PSIZE(hdr); 1925 } 1926 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL); 1927 } 1928 1929 /* 1930 * Block pointers always store the checksum for the logical data. 1931 * If the block pointer has the gang bit set, then the checksum 1932 * it represents is for the reconstituted data and not for an 1933 * individual gang member. The zio pipeline, however, must be able to 1934 * determine the checksum of each of the gang constituents so it 1935 * treats the checksum comparison differently than what we need 1936 * for l2arc blocks. This prevents us from using the 1937 * zio_checksum_error() interface directly. Instead we must call the 1938 * zio_checksum_error_impl() so that we can ensure the checksum is 1939 * generated using the correct checksum algorithm and accounts for the 1940 * logical I/O size and not just a gang fragment. 1941 */ 1942 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp, 1943 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size, 1944 zio->io_offset, NULL) == 0); 1945 zio_pop_transforms(zio); 1946 return (valid_cksum); 1947} 1948 1949/* 1950 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a 1951 * checksum and attaches it to the buf's hdr so that we can ensure that the buf 1952 * isn't modified later on. If buf is compressed or there is already a checksum 1953 * on the hdr, this is a no-op (we only checksum uncompressed bufs). 1954 */ 1955static void 1956arc_cksum_compute(arc_buf_t *buf) 1957{ 1958 arc_buf_hdr_t *hdr = buf->b_hdr; 1959 1960 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1961 return; 1962 1963 ASSERT(HDR_HAS_L1HDR(hdr)); 1964 1965 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1966 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1967 ASSERT(arc_hdr_has_uncompressed_buf(hdr)); 1968 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1969 return; 1970 } else if (ARC_BUF_COMPRESSED(buf)) { 1971 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1972 return; 1973 } 1974 1975 ASSERT(!ARC_BUF_COMPRESSED(buf)); 1976 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), 1977 KM_SLEEP); 1978 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, 1979 hdr->b_l1hdr.b_freeze_cksum); 1980 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1981#ifdef illumos 1982 arc_buf_watch(buf); 1983#endif 1984} 1985 1986#ifdef illumos 1987#ifndef _KERNEL 1988typedef struct procctl { 1989 long cmd; 1990 prwatch_t prwatch; 1991} procctl_t; 1992#endif 1993 1994/* ARGSUSED */ 1995static void 1996arc_buf_unwatch(arc_buf_t *buf) 1997{ 1998#ifndef _KERNEL 1999 if (arc_watch) { 2000 int result; 2001 procctl_t ctl; 2002 ctl.cmd = PCWATCH; 2003 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 2004 ctl.prwatch.pr_size = 0; 2005 ctl.prwatch.pr_wflags = 0; 2006 result = write(arc_procfd, &ctl, sizeof (ctl)); 2007 ASSERT3U(result, ==, sizeof (ctl)); 2008 } 2009#endif 2010} 2011 2012/* ARGSUSED */ 2013static void 2014arc_buf_watch(arc_buf_t *buf) 2015{ 2016#ifndef _KERNEL 2017 if (arc_watch) { 2018 int result; 2019 procctl_t ctl; 2020 ctl.cmd = PCWATCH; 2021 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 2022 ctl.prwatch.pr_size = arc_buf_size(buf); 2023 ctl.prwatch.pr_wflags = WA_WRITE; 2024 result = write(arc_procfd, &ctl, sizeof (ctl)); 2025 ASSERT3U(result, ==, sizeof (ctl)); 2026 } 2027#endif 2028} 2029#endif /* illumos */ 2030 2031static arc_buf_contents_t 2032arc_buf_type(arc_buf_hdr_t *hdr) 2033{ 2034 arc_buf_contents_t type; 2035 if (HDR_ISTYPE_METADATA(hdr)) { 2036 type = ARC_BUFC_METADATA; 2037 } else { 2038 type = ARC_BUFC_DATA; 2039 } 2040 VERIFY3U(hdr->b_type, ==, type); 2041 return (type); 2042} 2043 2044boolean_t 2045arc_is_metadata(arc_buf_t *buf) 2046{ 2047 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0); 2048} 2049 2050static uint32_t 2051arc_bufc_to_flags(arc_buf_contents_t type) 2052{ 2053 switch (type) { 2054 case ARC_BUFC_DATA: 2055 /* metadata field is 0 if buffer contains normal data */ 2056 return (0); 2057 case ARC_BUFC_METADATA: 2058 return (ARC_FLAG_BUFC_METADATA); 2059 default: 2060 break; 2061 } 2062 panic("undefined ARC buffer type!"); 2063 return ((uint32_t)-1); 2064} 2065 2066void 2067arc_buf_thaw(arc_buf_t *buf) 2068{ 2069 arc_buf_hdr_t *hdr = buf->b_hdr; 2070 2071 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2072 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2073 2074 arc_cksum_verify(buf); 2075 2076 /* 2077 * Compressed buffers do not manipulate the b_freeze_cksum or 2078 * allocate b_thawed. 2079 */ 2080 if (ARC_BUF_COMPRESSED(buf)) { 2081 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 2082 arc_hdr_has_uncompressed_buf(hdr)); 2083 return; 2084 } 2085 2086 ASSERT(HDR_HAS_L1HDR(hdr)); 2087 arc_cksum_free(hdr); 2088 2089 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 2090#ifdef ZFS_DEBUG 2091 if (zfs_flags & ZFS_DEBUG_MODIFY) { 2092 if (hdr->b_l1hdr.b_thawed != NULL) 2093 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2094 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 2095 } 2096#endif 2097 2098 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 2099 2100#ifdef illumos 2101 arc_buf_unwatch(buf); 2102#endif 2103} 2104 2105void 2106arc_buf_freeze(arc_buf_t *buf) 2107{ 2108 arc_buf_hdr_t *hdr = buf->b_hdr; 2109 kmutex_t *hash_lock; 2110 2111 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 2112 return; 2113 2114 if (ARC_BUF_COMPRESSED(buf)) { 2115 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 2116 arc_hdr_has_uncompressed_buf(hdr)); 2117 return; 2118 } 2119 2120 hash_lock = HDR_LOCK(hdr); 2121 mutex_enter(hash_lock); 2122 2123 ASSERT(HDR_HAS_L1HDR(hdr)); 2124 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL || 2125 hdr->b_l1hdr.b_state == arc_anon); 2126 arc_cksum_compute(buf); 2127 mutex_exit(hash_lock); 2128} 2129 2130/* 2131 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, 2132 * the following functions should be used to ensure that the flags are 2133 * updated in a thread-safe way. When manipulating the flags either 2134 * the hash_lock must be held or the hdr must be undiscoverable. This 2135 * ensures that we're not racing with any other threads when updating 2136 * the flags. 2137 */ 2138static inline void 2139arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 2140{ 2141 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2142 hdr->b_flags |= flags; 2143} 2144 2145static inline void 2146arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 2147{ 2148 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2149 hdr->b_flags &= ~flags; 2150} 2151 2152/* 2153 * Setting the compression bits in the arc_buf_hdr_t's b_flags is 2154 * done in a special way since we have to clear and set bits 2155 * at the same time. Consumers that wish to set the compression bits 2156 * must use this function to ensure that the flags are updated in 2157 * thread-safe manner. 2158 */ 2159static void 2160arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) 2161{ 2162 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2163 2164 /* 2165 * Holes and embedded blocks will always have a psize = 0 so 2166 * we ignore the compression of the blkptr and set the 2167 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF. 2168 * Holes and embedded blocks remain anonymous so we don't 2169 * want to uncompress them. Mark them as uncompressed. 2170 */ 2171 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { 2172 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 2173 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 2174 ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); 2175 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 2176 } else { 2177 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 2178 HDR_SET_COMPRESS(hdr, cmp); 2179 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); 2180 ASSERT(HDR_COMPRESSION_ENABLED(hdr)); 2181 } 2182} 2183 2184/* 2185 * Looks for another buf on the same hdr which has the data decompressed, copies 2186 * from it, and returns true. If no such buf exists, returns false. 2187 */ 2188static boolean_t 2189arc_buf_try_copy_decompressed_data(arc_buf_t *buf) 2190{ 2191 arc_buf_hdr_t *hdr = buf->b_hdr; 2192 boolean_t copied = B_FALSE; 2193 2194 ASSERT(HDR_HAS_L1HDR(hdr)); 2195 ASSERT3P(buf->b_data, !=, NULL); 2196 ASSERT(!ARC_BUF_COMPRESSED(buf)); 2197 2198 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL; 2199 from = from->b_next) { 2200 /* can't use our own data buffer */ 2201 if (from == buf) { 2202 continue; 2203 } 2204 2205 if (!ARC_BUF_COMPRESSED(from)) { 2206 bcopy(from->b_data, buf->b_data, arc_buf_size(buf)); 2207 copied = B_TRUE; 2208 break; 2209 } 2210 } 2211 2212 /* 2213 * There were no decompressed bufs, so there should not be a 2214 * checksum on the hdr either. 2215 */ 2216 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL); 2217 2218 return (copied); 2219} 2220 2221/* 2222 * Given a buf that has a data buffer attached to it, this function will 2223 * efficiently fill the buf with data of the specified compression setting from 2224 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr 2225 * are already sharing a data buf, no copy is performed. 2226 * 2227 * If the buf is marked as compressed but uncompressed data was requested, this 2228 * will allocate a new data buffer for the buf, remove that flag, and fill the 2229 * buf with uncompressed data. You can't request a compressed buf on a hdr with 2230 * uncompressed data, and (since we haven't added support for it yet) if you 2231 * want compressed data your buf must already be marked as compressed and have 2232 * the correct-sized data buffer. 2233 */ 2234static int 2235arc_buf_fill(arc_buf_t *buf, boolean_t compressed) 2236{ 2237 arc_buf_hdr_t *hdr = buf->b_hdr; 2238 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 2239 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; 2240 2241 ASSERT3P(buf->b_data, !=, NULL); 2242 IMPLY(compressed, hdr_compressed); 2243 IMPLY(compressed, ARC_BUF_COMPRESSED(buf)); 2244 2245 if (hdr_compressed == compressed) { 2246 if (!arc_buf_is_shared(buf)) { 2247 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd, 2248 arc_buf_size(buf)); 2249 } 2250 } else { 2251 ASSERT(hdr_compressed); 2252 ASSERT(!compressed); 2253 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); 2254 2255 /* 2256 * If the buf is sharing its data with the hdr, unlink it and 2257 * allocate a new data buffer for the buf. 2258 */ 2259 if (arc_buf_is_shared(buf)) { 2260 ASSERT(ARC_BUF_COMPRESSED(buf)); 2261 2262 /* We need to give the buf it's own b_data */ 2263 buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 2264 buf->b_data = 2265 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 2266 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2267 2268 /* Previously overhead was 0; just add new overhead */ 2269 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 2270 } else if (ARC_BUF_COMPRESSED(buf)) { 2271 /* We need to reallocate the buf's b_data */ 2272 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr), 2273 buf); 2274 buf->b_data = 2275 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 2276 2277 /* We increased the size of b_data; update overhead */ 2278 ARCSTAT_INCR(arcstat_overhead_size, 2279 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr)); 2280 } 2281 2282 /* 2283 * Regardless of the buf's previous compression settings, it 2284 * should not be compressed at the end of this function. 2285 */ 2286 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; 2287 2288 /* 2289 * Try copying the data from another buf which already has a 2290 * decompressed version. If that's not possible, it's time to 2291 * bite the bullet and decompress the data from the hdr. 2292 */ 2293 if (arc_buf_try_copy_decompressed_data(buf)) { 2294 /* Skip byteswapping and checksumming (already done) */ 2295 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL); 2296 return (0); 2297 } else { 2298 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr), 2299 hdr->b_l1hdr.b_pabd, buf->b_data, 2300 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 2301 2302 /* 2303 * Absent hardware errors or software bugs, this should 2304 * be impossible, but log it anyway so we can debug it. 2305 */ 2306 if (error != 0) { 2307 zfs_dbgmsg( 2308 "hdr %p, compress %d, psize %d, lsize %d", 2309 hdr, HDR_GET_COMPRESS(hdr), 2310 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 2311 return (SET_ERROR(EIO)); 2312 } 2313 } 2314 } 2315 2316 /* Byteswap the buf's data if necessary */ 2317 if (bswap != DMU_BSWAP_NUMFUNCS) { 2318 ASSERT(!HDR_SHARED_DATA(hdr)); 2319 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); 2320 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); 2321 } 2322 2323 /* Compute the hdr's checksum if necessary */ 2324 arc_cksum_compute(buf); 2325 2326 return (0); 2327} 2328 2329int 2330arc_decompress(arc_buf_t *buf) 2331{ 2332 return (arc_buf_fill(buf, B_FALSE)); 2333} 2334 2335/* 2336 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t. 2337 */ 2338static uint64_t 2339arc_hdr_size(arc_buf_hdr_t *hdr) 2340{ 2341 uint64_t size; 2342 2343 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 2344 HDR_GET_PSIZE(hdr) > 0) { 2345 size = HDR_GET_PSIZE(hdr); 2346 } else { 2347 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); 2348 size = HDR_GET_LSIZE(hdr); 2349 } 2350 return (size); 2351} 2352 2353/* 2354 * Increment the amount of evictable space in the arc_state_t's refcount. 2355 * We account for the space used by the hdr and the arc buf individually 2356 * so that we can add and remove them from the refcount individually. 2357 */ 2358static void 2359arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) 2360{ 2361 arc_buf_contents_t type = arc_buf_type(hdr); 2362 2363 ASSERT(HDR_HAS_L1HDR(hdr)); 2364 2365 if (GHOST_STATE(state)) { 2366 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2367 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2368 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2369 (void) refcount_add_many(&state->arcs_esize[type], 2370 HDR_GET_LSIZE(hdr), hdr); 2371 return; 2372 } 2373 2374 ASSERT(!GHOST_STATE(state)); 2375 if (hdr->b_l1hdr.b_pabd != NULL) { 2376 (void) refcount_add_many(&state->arcs_esize[type], 2377 arc_hdr_size(hdr), hdr); 2378 } 2379 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2380 buf = buf->b_next) { 2381 if (arc_buf_is_shared(buf)) 2382 continue; 2383 (void) refcount_add_many(&state->arcs_esize[type], 2384 arc_buf_size(buf), buf); 2385 } 2386} 2387 2388/* 2389 * Decrement the amount of evictable space in the arc_state_t's refcount. 2390 * We account for the space used by the hdr and the arc buf individually 2391 * so that we can add and remove them from the refcount individually. 2392 */ 2393static void 2394arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) 2395{ 2396 arc_buf_contents_t type = arc_buf_type(hdr); 2397 2398 ASSERT(HDR_HAS_L1HDR(hdr)); 2399 2400 if (GHOST_STATE(state)) { 2401 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2402 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2403 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2404 (void) refcount_remove_many(&state->arcs_esize[type], 2405 HDR_GET_LSIZE(hdr), hdr); 2406 return; 2407 } 2408 2409 ASSERT(!GHOST_STATE(state)); 2410 if (hdr->b_l1hdr.b_pabd != NULL) { 2411 (void) refcount_remove_many(&state->arcs_esize[type], 2412 arc_hdr_size(hdr), hdr); 2413 } 2414 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2415 buf = buf->b_next) { 2416 if (arc_buf_is_shared(buf)) 2417 continue; 2418 (void) refcount_remove_many(&state->arcs_esize[type], 2419 arc_buf_size(buf), buf); 2420 } 2421} 2422 2423/* 2424 * Add a reference to this hdr indicating that someone is actively 2425 * referencing that memory. When the refcount transitions from 0 to 1, 2426 * we remove it from the respective arc_state_t list to indicate that 2427 * it is not evictable. 2428 */ 2429static void 2430add_reference(arc_buf_hdr_t *hdr, void *tag) 2431{ 2432 ASSERT(HDR_HAS_L1HDR(hdr)); 2433 if (!MUTEX_HELD(HDR_LOCK(hdr))) { 2434 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 2435 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2436 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2437 } 2438 2439 arc_state_t *state = hdr->b_l1hdr.b_state; 2440 2441 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 2442 (state != arc_anon)) { 2443 /* We don't use the L2-only state list. */ 2444 if (state != arc_l2c_only) { 2445 multilist_remove(state->arcs_list[arc_buf_type(hdr)], 2446 hdr); 2447 arc_evictable_space_decrement(hdr, state); 2448 } 2449 /* remove the prefetch flag if we get a reference */ 2450 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 2451 } 2452} 2453 2454/* 2455 * Remove a reference from this hdr. When the reference transitions from 2456 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's 2457 * list making it eligible for eviction. 2458 */ 2459static int 2460remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 2461{ 2462 int cnt; 2463 arc_state_t *state = hdr->b_l1hdr.b_state; 2464 2465 ASSERT(HDR_HAS_L1HDR(hdr)); 2466 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 2467 ASSERT(!GHOST_STATE(state)); 2468 2469 /* 2470 * arc_l2c_only counts as a ghost state so we don't need to explicitly 2471 * check to prevent usage of the arc_l2c_only list. 2472 */ 2473 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 2474 (state != arc_anon)) { 2475 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr); 2476 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 2477 arc_evictable_space_increment(hdr, state); 2478 } 2479 return (cnt); 2480} 2481 2482/* 2483 * Move the supplied buffer to the indicated state. The hash lock 2484 * for the buffer must be held by the caller. 2485 */ 2486static void 2487arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 2488 kmutex_t *hash_lock) 2489{ 2490 arc_state_t *old_state; 2491 int64_t refcnt; 2492 uint32_t bufcnt; 2493 boolean_t update_old, update_new; 2494 arc_buf_contents_t buftype = arc_buf_type(hdr); 2495 2496 /* 2497 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 2498 * in arc_read() when bringing a buffer out of the L2ARC. However, the 2499 * L1 hdr doesn't always exist when we change state to arc_anon before 2500 * destroying a header, in which case reallocating to add the L1 hdr is 2501 * pointless. 2502 */ 2503 if (HDR_HAS_L1HDR(hdr)) { 2504 old_state = hdr->b_l1hdr.b_state; 2505 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 2506 bufcnt = hdr->b_l1hdr.b_bufcnt; 2507 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL); 2508 } else { 2509 old_state = arc_l2c_only; 2510 refcnt = 0; 2511 bufcnt = 0; 2512 update_old = B_FALSE; 2513 } 2514 update_new = update_old; 2515 2516 ASSERT(MUTEX_HELD(hash_lock)); 2517 ASSERT3P(new_state, !=, old_state); 2518 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); 2519 ASSERT(old_state != arc_anon || bufcnt <= 1); 2520 2521 /* 2522 * If this buffer is evictable, transfer it from the 2523 * old state list to the new state list. 2524 */ 2525 if (refcnt == 0) { 2526 if (old_state != arc_anon && old_state != arc_l2c_only) { 2527 ASSERT(HDR_HAS_L1HDR(hdr)); 2528 multilist_remove(old_state->arcs_list[buftype], hdr); 2529 2530 if (GHOST_STATE(old_state)) { 2531 ASSERT0(bufcnt); 2532 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2533 update_old = B_TRUE; 2534 } 2535 arc_evictable_space_decrement(hdr, old_state); 2536 } 2537 if (new_state != arc_anon && new_state != arc_l2c_only) { 2538 2539 /* 2540 * An L1 header always exists here, since if we're 2541 * moving to some L1-cached state (i.e. not l2c_only or 2542 * anonymous), we realloc the header to add an L1hdr 2543 * beforehand. 2544 */ 2545 ASSERT(HDR_HAS_L1HDR(hdr)); 2546 multilist_insert(new_state->arcs_list[buftype], hdr); 2547 2548 if (GHOST_STATE(new_state)) { 2549 ASSERT0(bufcnt); 2550 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2551 update_new = B_TRUE; 2552 } 2553 arc_evictable_space_increment(hdr, new_state); 2554 } 2555 } 2556 2557 ASSERT(!HDR_EMPTY(hdr)); 2558 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 2559 buf_hash_remove(hdr); 2560 2561 /* adjust state sizes (ignore arc_l2c_only) */ 2562 2563 if (update_new && new_state != arc_l2c_only) { 2564 ASSERT(HDR_HAS_L1HDR(hdr)); 2565 if (GHOST_STATE(new_state)) { 2566 ASSERT0(bufcnt); 2567 2568 /* 2569 * When moving a header to a ghost state, we first 2570 * remove all arc buffers. Thus, we'll have a 2571 * bufcnt of zero, and no arc buffer to use for 2572 * the reference. As a result, we use the arc 2573 * header pointer for the reference. 2574 */ 2575 (void) refcount_add_many(&new_state->arcs_size, 2576 HDR_GET_LSIZE(hdr), hdr); 2577 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2578 } else { 2579 uint32_t buffers = 0; 2580 2581 /* 2582 * Each individual buffer holds a unique reference, 2583 * thus we must remove each of these references one 2584 * at a time. 2585 */ 2586 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2587 buf = buf->b_next) { 2588 ASSERT3U(bufcnt, !=, 0); 2589 buffers++; 2590 2591 /* 2592 * When the arc_buf_t is sharing the data 2593 * block with the hdr, the owner of the 2594 * reference belongs to the hdr. Only 2595 * add to the refcount if the arc_buf_t is 2596 * not shared. 2597 */ 2598 if (arc_buf_is_shared(buf)) 2599 continue; 2600 2601 (void) refcount_add_many(&new_state->arcs_size, 2602 arc_buf_size(buf), buf); 2603 } 2604 ASSERT3U(bufcnt, ==, buffers); 2605 2606 if (hdr->b_l1hdr.b_pabd != NULL) { 2607 (void) refcount_add_many(&new_state->arcs_size, 2608 arc_hdr_size(hdr), hdr); 2609 } else { 2610 ASSERT(GHOST_STATE(old_state)); 2611 } 2612 } 2613 } 2614 2615 if (update_old && old_state != arc_l2c_only) { 2616 ASSERT(HDR_HAS_L1HDR(hdr)); 2617 if (GHOST_STATE(old_state)) { 2618 ASSERT0(bufcnt); 2619 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2620 2621 /* 2622 * When moving a header off of a ghost state, 2623 * the header will not contain any arc buffers. 2624 * We use the arc header pointer for the reference 2625 * which is exactly what we did when we put the 2626 * header on the ghost state. 2627 */ 2628 2629 (void) refcount_remove_many(&old_state->arcs_size, 2630 HDR_GET_LSIZE(hdr), hdr); 2631 } else { 2632 uint32_t buffers = 0; 2633 2634 /* 2635 * Each individual buffer holds a unique reference, 2636 * thus we must remove each of these references one 2637 * at a time. 2638 */ 2639 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2640 buf = buf->b_next) { 2641 ASSERT3U(bufcnt, !=, 0); 2642 buffers++; 2643 2644 /* 2645 * When the arc_buf_t is sharing the data 2646 * block with the hdr, the owner of the 2647 * reference belongs to the hdr. Only 2648 * add to the refcount if the arc_buf_t is 2649 * not shared. 2650 */ 2651 if (arc_buf_is_shared(buf)) 2652 continue; 2653 2654 (void) refcount_remove_many( 2655 &old_state->arcs_size, arc_buf_size(buf), 2656 buf); 2657 } 2658 ASSERT3U(bufcnt, ==, buffers); 2659 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2660 (void) refcount_remove_many( 2661 &old_state->arcs_size, arc_hdr_size(hdr), hdr); 2662 } 2663 } 2664 2665 if (HDR_HAS_L1HDR(hdr)) 2666 hdr->b_l1hdr.b_state = new_state; 2667 2668 /* 2669 * L2 headers should never be on the L2 state list since they don't 2670 * have L1 headers allocated. 2671 */ 2672 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 2673 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 2674} 2675 2676void 2677arc_space_consume(uint64_t space, arc_space_type_t type) 2678{ 2679 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2680 2681 switch (type) { 2682 case ARC_SPACE_DATA: 2683 aggsum_add(&astat_data_size, space); 2684 break; 2685 case ARC_SPACE_META: 2686 aggsum_add(&astat_metadata_size, space); 2687 break; 2688 case ARC_SPACE_OTHER: 2689 aggsum_add(&astat_other_size, space); 2690 break; 2691 case ARC_SPACE_HDRS: 2692 aggsum_add(&astat_hdr_size, space); 2693 break; 2694 case ARC_SPACE_L2HDRS: 2695 aggsum_add(&astat_l2_hdr_size, space); 2696 break; 2697 } 2698 2699 if (type != ARC_SPACE_DATA) 2700 aggsum_add(&arc_meta_used, space); 2701 2702 aggsum_add(&arc_size, space); 2703} 2704 2705void 2706arc_space_return(uint64_t space, arc_space_type_t type) 2707{ 2708 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2709 2710 switch (type) { 2711 case ARC_SPACE_DATA: 2712 aggsum_add(&astat_data_size, -space); 2713 break; 2714 case ARC_SPACE_META: 2715 aggsum_add(&astat_metadata_size, -space); 2716 break; 2717 case ARC_SPACE_OTHER: 2718 aggsum_add(&astat_other_size, -space); 2719 break; 2720 case ARC_SPACE_HDRS: 2721 aggsum_add(&astat_hdr_size, -space); 2722 break; 2723 case ARC_SPACE_L2HDRS: 2724 aggsum_add(&astat_l2_hdr_size, -space); 2725 break; 2726 } 2727 2728 if (type != ARC_SPACE_DATA) { 2729 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0); 2730 /* 2731 * We use the upper bound here rather than the precise value 2732 * because the arc_meta_max value doesn't need to be 2733 * precise. It's only consumed by humans via arcstats. 2734 */ 2735 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used)) 2736 arc_meta_max = aggsum_upper_bound(&arc_meta_used); 2737 aggsum_add(&arc_meta_used, -space); 2738 } 2739 2740 ASSERT(aggsum_compare(&arc_size, space) >= 0); 2741 aggsum_add(&arc_size, -space); 2742} 2743 2744/* 2745 * Given a hdr and a buf, returns whether that buf can share its b_data buffer 2746 * with the hdr's b_pabd. 2747 */ 2748static boolean_t 2749arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2750{ 2751 /* 2752 * The criteria for sharing a hdr's data are: 2753 * 1. the hdr's compression matches the buf's compression 2754 * 2. the hdr doesn't need to be byteswapped 2755 * 3. the hdr isn't already being shared 2756 * 4. the buf is either compressed or it is the last buf in the hdr list 2757 * 2758 * Criterion #4 maintains the invariant that shared uncompressed 2759 * bufs must be the final buf in the hdr's b_buf list. Reading this, you 2760 * might ask, "if a compressed buf is allocated first, won't that be the 2761 * last thing in the list?", but in that case it's impossible to create 2762 * a shared uncompressed buf anyway (because the hdr must be compressed 2763 * to have the compressed buf). You might also think that #3 is 2764 * sufficient to make this guarantee, however it's possible 2765 * (specifically in the rare L2ARC write race mentioned in 2766 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that 2767 * is sharable, but wasn't at the time of its allocation. Rather than 2768 * allow a new shared uncompressed buf to be created and then shuffle 2769 * the list around to make it the last element, this simply disallows 2770 * sharing if the new buf isn't the first to be added. 2771 */ 2772 ASSERT3P(buf->b_hdr, ==, hdr); 2773 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF; 2774 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0; 2775 return (buf_compressed == hdr_compressed && 2776 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && 2777 !HDR_SHARED_DATA(hdr) && 2778 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf))); 2779} 2780 2781/* 2782 * Allocate a buf for this hdr. If you care about the data that's in the hdr, 2783 * or if you want a compressed buffer, pass those flags in. Returns 0 if the 2784 * copy was made successfully, or an error code otherwise. 2785 */ 2786static int 2787arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed, 2788 boolean_t fill, arc_buf_t **ret) 2789{ 2790 arc_buf_t *buf; 2791 2792 ASSERT(HDR_HAS_L1HDR(hdr)); 2793 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2794 VERIFY(hdr->b_type == ARC_BUFC_DATA || 2795 hdr->b_type == ARC_BUFC_METADATA); 2796 ASSERT3P(ret, !=, NULL); 2797 ASSERT3P(*ret, ==, NULL); 2798 2799 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2800 buf->b_hdr = hdr; 2801 buf->b_data = NULL; 2802 buf->b_next = hdr->b_l1hdr.b_buf; 2803 buf->b_flags = 0; 2804 2805 add_reference(hdr, tag); 2806 2807 /* 2808 * We're about to change the hdr's b_flags. We must either 2809 * hold the hash_lock or be undiscoverable. 2810 */ 2811 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2812 2813 /* 2814 * Only honor requests for compressed bufs if the hdr is actually 2815 * compressed. 2816 */ 2817 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 2818 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; 2819 2820 /* 2821 * If the hdr's data can be shared then we share the data buffer and 2822 * set the appropriate bit in the hdr's b_flags to indicate the hdr is 2823 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new 2824 * buffer to store the buf's data. 2825 * 2826 * There are two additional restrictions here because we're sharing 2827 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be 2828 * actively involved in an L2ARC write, because if this buf is used by 2829 * an arc_write() then the hdr's data buffer will be released when the 2830 * write completes, even though the L2ARC write might still be using it. 2831 * Second, the hdr's ABD must be linear so that the buf's user doesn't 2832 * need to be ABD-aware. 2833 */ 2834 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) && 2835 abd_is_linear(hdr->b_l1hdr.b_pabd); 2836 2837 /* Set up b_data and sharing */ 2838 if (can_share) { 2839 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd); 2840 buf->b_flags |= ARC_BUF_FLAG_SHARED; 2841 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2842 } else { 2843 buf->b_data = 2844 arc_get_data_buf(hdr, arc_buf_size(buf), buf); 2845 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2846 } 2847 VERIFY3P(buf->b_data, !=, NULL); 2848 2849 hdr->b_l1hdr.b_buf = buf; 2850 hdr->b_l1hdr.b_bufcnt += 1; 2851 2852 /* 2853 * If the user wants the data from the hdr, we need to either copy or 2854 * decompress the data. 2855 */ 2856 if (fill) { 2857 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0)); 2858 } 2859 2860 return (0); 2861} 2862 2863static char *arc_onloan_tag = "onloan"; 2864 2865static inline void 2866arc_loaned_bytes_update(int64_t delta) 2867{ 2868 atomic_add_64(&arc_loaned_bytes, delta); 2869 2870 /* assert that it did not wrap around */ 2871 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 2872} 2873 2874/* 2875 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2876 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2877 * buffers must be returned to the arc before they can be used by the DMU or 2878 * freed. 2879 */ 2880arc_buf_t * 2881arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size) 2882{ 2883 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag, 2884 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size); 2885 2886 arc_loaned_bytes_update(arc_buf_size(buf)); 2887 2888 return (buf); 2889} 2890 2891arc_buf_t * 2892arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize, 2893 enum zio_compress compression_type) 2894{ 2895 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag, 2896 psize, lsize, compression_type); 2897 2898 arc_loaned_bytes_update(arc_buf_size(buf)); 2899 2900 return (buf); 2901} 2902 2903 2904/* 2905 * Return a loaned arc buffer to the arc. 2906 */ 2907void 2908arc_return_buf(arc_buf_t *buf, void *tag) 2909{ 2910 arc_buf_hdr_t *hdr = buf->b_hdr; 2911 2912 ASSERT3P(buf->b_data, !=, NULL); 2913 ASSERT(HDR_HAS_L1HDR(hdr)); 2914 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2915 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2916 2917 arc_loaned_bytes_update(-arc_buf_size(buf)); 2918} 2919 2920/* Detach an arc_buf from a dbuf (tag) */ 2921void 2922arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2923{ 2924 arc_buf_hdr_t *hdr = buf->b_hdr; 2925 2926 ASSERT3P(buf->b_data, !=, NULL); 2927 ASSERT(HDR_HAS_L1HDR(hdr)); 2928 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2929 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2930 2931 arc_loaned_bytes_update(arc_buf_size(buf)); 2932} 2933 2934static void 2935l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type) 2936{ 2937 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); 2938 2939 df->l2df_abd = abd; 2940 df->l2df_size = size; 2941 df->l2df_type = type; 2942 mutex_enter(&l2arc_free_on_write_mtx); 2943 list_insert_head(l2arc_free_on_write, df); 2944 mutex_exit(&l2arc_free_on_write_mtx); 2945} 2946 2947static void 2948arc_hdr_free_on_write(arc_buf_hdr_t *hdr) 2949{ 2950 arc_state_t *state = hdr->b_l1hdr.b_state; 2951 arc_buf_contents_t type = arc_buf_type(hdr); 2952 uint64_t size = arc_hdr_size(hdr); 2953 2954 /* protected by hash lock, if in the hash table */ 2955 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 2956 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2957 ASSERT(state != arc_anon && state != arc_l2c_only); 2958 2959 (void) refcount_remove_many(&state->arcs_esize[type], 2960 size, hdr); 2961 } 2962 (void) refcount_remove_many(&state->arcs_size, size, hdr); 2963 if (type == ARC_BUFC_METADATA) { 2964 arc_space_return(size, ARC_SPACE_META); 2965 } else { 2966 ASSERT(type == ARC_BUFC_DATA); 2967 arc_space_return(size, ARC_SPACE_DATA); 2968 } 2969 2970 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type); 2971} 2972 2973/* 2974 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the 2975 * data buffer, we transfer the refcount ownership to the hdr and update 2976 * the appropriate kstats. 2977 */ 2978static void 2979arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2980{ 2981 arc_state_t *state = hdr->b_l1hdr.b_state; 2982 2983 ASSERT(arc_can_share(hdr, buf)); 2984 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2985 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2986 2987 /* 2988 * Start sharing the data buffer. We transfer the 2989 * refcount ownership to the hdr since it always owns 2990 * the refcount whenever an arc_buf_t is shared. 2991 */ 2992 refcount_transfer_ownership(&state->arcs_size, buf, hdr); 2993 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf)); 2994 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd, 2995 HDR_ISTYPE_METADATA(hdr)); 2996 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2997 buf->b_flags |= ARC_BUF_FLAG_SHARED; 2998 2999 /* 3000 * Since we've transferred ownership to the hdr we need 3001 * to increment its compressed and uncompressed kstats and 3002 * decrement the overhead size. 3003 */ 3004 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 3005 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 3006 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf)); 3007} 3008 3009static void 3010arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 3011{ 3012 arc_state_t *state = hdr->b_l1hdr.b_state; 3013 3014 ASSERT(arc_buf_is_shared(buf)); 3015 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 3016 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 3017 3018 /* 3019 * We are no longer sharing this buffer so we need 3020 * to transfer its ownership to the rightful owner. 3021 */ 3022 refcount_transfer_ownership(&state->arcs_size, hdr, buf); 3023 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 3024 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd); 3025 abd_put(hdr->b_l1hdr.b_pabd); 3026 hdr->b_l1hdr.b_pabd = NULL; 3027 buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 3028 3029 /* 3030 * Since the buffer is no longer shared between 3031 * the arc buf and the hdr, count it as overhead. 3032 */ 3033 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 3034 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 3035 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 3036} 3037 3038/* 3039 * Remove an arc_buf_t from the hdr's buf list and return the last 3040 * arc_buf_t on the list. If no buffers remain on the list then return 3041 * NULL. 3042 */ 3043static arc_buf_t * 3044arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf) 3045{ 3046 ASSERT(HDR_HAS_L1HDR(hdr)); 3047 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 3048 3049 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf; 3050 arc_buf_t *lastbuf = NULL; 3051 3052 /* 3053 * Remove the buf from the hdr list and locate the last 3054 * remaining buffer on the list. 3055 */ 3056 while (*bufp != NULL) { 3057 if (*bufp == buf) 3058 *bufp = buf->b_next; 3059 3060 /* 3061 * If we've removed a buffer in the middle of 3062 * the list then update the lastbuf and update 3063 * bufp. 3064 */ 3065 if (*bufp != NULL) { 3066 lastbuf = *bufp; 3067 bufp = &(*bufp)->b_next; 3068 } 3069 } 3070 buf->b_next = NULL; 3071 ASSERT3P(lastbuf, !=, buf); 3072 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL); 3073 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL); 3074 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf)); 3075 3076 return (lastbuf); 3077} 3078 3079/* 3080 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's 3081 * list and free it. 3082 */ 3083static void 3084arc_buf_destroy_impl(arc_buf_t *buf) 3085{ 3086 arc_buf_hdr_t *hdr = buf->b_hdr; 3087 3088 /* 3089 * Free up the data associated with the buf but only if we're not 3090 * sharing this with the hdr. If we are sharing it with the hdr, the 3091 * hdr is responsible for doing the free. 3092 */ 3093 if (buf->b_data != NULL) { 3094 /* 3095 * We're about to change the hdr's b_flags. We must either 3096 * hold the hash_lock or be undiscoverable. 3097 */ 3098 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 3099 3100 arc_cksum_verify(buf); 3101#ifdef illumos 3102 arc_buf_unwatch(buf); 3103#endif 3104 3105 if (arc_buf_is_shared(buf)) { 3106 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 3107 } else { 3108 uint64_t size = arc_buf_size(buf); 3109 arc_free_data_buf(hdr, buf->b_data, size, buf); 3110 ARCSTAT_INCR(arcstat_overhead_size, -size); 3111 } 3112 buf->b_data = NULL; 3113 3114 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 3115 hdr->b_l1hdr.b_bufcnt -= 1; 3116 } 3117 3118 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 3119 3120 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) { 3121 /* 3122 * If the current arc_buf_t is sharing its data buffer with the 3123 * hdr, then reassign the hdr's b_pabd to share it with the new 3124 * buffer at the end of the list. The shared buffer is always 3125 * the last one on the hdr's buffer list. 3126 * 3127 * There is an equivalent case for compressed bufs, but since 3128 * they aren't guaranteed to be the last buf in the list and 3129 * that is an exceedingly rare case, we just allow that space be 3130 * wasted temporarily. 3131 */ 3132 if (lastbuf != NULL) { 3133 /* Only one buf can be shared at once */ 3134 VERIFY(!arc_buf_is_shared(lastbuf)); 3135 /* hdr is uncompressed so can't have compressed buf */ 3136 VERIFY(!ARC_BUF_COMPRESSED(lastbuf)); 3137 3138 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 3139 arc_hdr_free_pabd(hdr); 3140 3141 /* 3142 * We must setup a new shared block between the 3143 * last buffer and the hdr. The data would have 3144 * been allocated by the arc buf so we need to transfer 3145 * ownership to the hdr since it's now being shared. 3146 */ 3147 arc_share_buf(hdr, lastbuf); 3148 } 3149 } else if (HDR_SHARED_DATA(hdr)) { 3150 /* 3151 * Uncompressed shared buffers are always at the end 3152 * of the list. Compressed buffers don't have the 3153 * same requirements. This makes it hard to 3154 * simply assert that the lastbuf is shared so 3155 * we rely on the hdr's compression flags to determine 3156 * if we have a compressed, shared buffer. 3157 */ 3158 ASSERT3P(lastbuf, !=, NULL); 3159 ASSERT(arc_buf_is_shared(lastbuf) || 3160 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 3161 } 3162 3163 /* 3164 * Free the checksum if we're removing the last uncompressed buf from 3165 * this hdr. 3166 */ 3167 if (!arc_hdr_has_uncompressed_buf(hdr)) { 3168 arc_cksum_free(hdr); 3169 } 3170 3171 /* clean up the buf */ 3172 buf->b_hdr = NULL; 3173 kmem_cache_free(buf_cache, buf); 3174} 3175 3176static void 3177arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr) 3178{ 3179 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 3180 ASSERT(HDR_HAS_L1HDR(hdr)); 3181 ASSERT(!HDR_SHARED_DATA(hdr)); 3182 3183 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3184 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr); 3185 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 3186 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 3187 3188 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 3189 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 3190} 3191 3192static void 3193arc_hdr_free_pabd(arc_buf_hdr_t *hdr) 3194{ 3195 ASSERT(HDR_HAS_L1HDR(hdr)); 3196 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 3197 3198 /* 3199 * If the hdr is currently being written to the l2arc then 3200 * we defer freeing the data by adding it to the l2arc_free_on_write 3201 * list. The l2arc will free the data once it's finished 3202 * writing it to the l2arc device. 3203 */ 3204 if (HDR_L2_WRITING(hdr)) { 3205 arc_hdr_free_on_write(hdr); 3206 ARCSTAT_BUMP(arcstat_l2_free_on_write); 3207 } else { 3208 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, 3209 arc_hdr_size(hdr), hdr); 3210 } 3211 hdr->b_l1hdr.b_pabd = NULL; 3212 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 3213 3214 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 3215 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 3216} 3217 3218static arc_buf_hdr_t * 3219arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, 3220 enum zio_compress compression_type, arc_buf_contents_t type) 3221{ 3222 arc_buf_hdr_t *hdr; 3223 3224 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); 3225 3226 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 3227 ASSERT(HDR_EMPTY(hdr)); 3228 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3229 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL); 3230 HDR_SET_PSIZE(hdr, psize); 3231 HDR_SET_LSIZE(hdr, lsize); 3232 hdr->b_spa = spa; 3233 hdr->b_type = type; 3234 hdr->b_flags = 0; 3235 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); 3236 arc_hdr_set_compress(hdr, compression_type); 3237 3238 hdr->b_l1hdr.b_state = arc_anon; 3239 hdr->b_l1hdr.b_arc_access = 0; 3240 hdr->b_l1hdr.b_bufcnt = 0; 3241 hdr->b_l1hdr.b_buf = NULL; 3242 3243 /* 3244 * Allocate the hdr's buffer. This will contain either 3245 * the compressed or uncompressed data depending on the block 3246 * it references and compressed arc enablement. 3247 */ 3248 arc_hdr_alloc_pabd(hdr); 3249 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3250 3251 return (hdr); 3252} 3253 3254/* 3255 * Transition between the two allocation states for the arc_buf_hdr struct. 3256 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 3257 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 3258 * version is used when a cache buffer is only in the L2ARC in order to reduce 3259 * memory usage. 3260 */ 3261static arc_buf_hdr_t * 3262arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 3263{ 3264 ASSERT(HDR_HAS_L2HDR(hdr)); 3265 3266 arc_buf_hdr_t *nhdr; 3267 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3268 3269 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 3270 (old == hdr_l2only_cache && new == hdr_full_cache)); 3271 3272 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 3273 3274 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 3275 buf_hash_remove(hdr); 3276 3277 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 3278 3279 if (new == hdr_full_cache) { 3280 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); 3281 /* 3282 * arc_access and arc_change_state need to be aware that a 3283 * header has just come out of L2ARC, so we set its state to 3284 * l2c_only even though it's about to change. 3285 */ 3286 nhdr->b_l1hdr.b_state = arc_l2c_only; 3287 3288 /* Verify previous threads set to NULL before freeing */ 3289 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL); 3290 } else { 3291 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3292 ASSERT0(hdr->b_l1hdr.b_bufcnt); 3293 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3294 3295 /* 3296 * If we've reached here, We must have been called from 3297 * arc_evict_hdr(), as such we should have already been 3298 * removed from any ghost list we were previously on 3299 * (which protects us from racing with arc_evict_state), 3300 * thus no locking is needed during this check. 3301 */ 3302 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3303 3304 /* 3305 * A buffer must not be moved into the arc_l2c_only 3306 * state if it's not finished being written out to the 3307 * l2arc device. Otherwise, the b_l1hdr.b_pabd field 3308 * might try to be accessed, even though it was removed. 3309 */ 3310 VERIFY(!HDR_L2_WRITING(hdr)); 3311 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3312 3313#ifdef ZFS_DEBUG 3314 if (hdr->b_l1hdr.b_thawed != NULL) { 3315 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3316 hdr->b_l1hdr.b_thawed = NULL; 3317 } 3318#endif 3319 3320 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); 3321 } 3322 /* 3323 * The header has been reallocated so we need to re-insert it into any 3324 * lists it was on. 3325 */ 3326 (void) buf_hash_insert(nhdr, NULL); 3327 3328 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 3329 3330 mutex_enter(&dev->l2ad_mtx); 3331 3332 /* 3333 * We must place the realloc'ed header back into the list at 3334 * the same spot. Otherwise, if it's placed earlier in the list, 3335 * l2arc_write_buffers() could find it during the function's 3336 * write phase, and try to write it out to the l2arc. 3337 */ 3338 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 3339 list_remove(&dev->l2ad_buflist, hdr); 3340 3341 mutex_exit(&dev->l2ad_mtx); 3342 3343 /* 3344 * Since we're using the pointer address as the tag when 3345 * incrementing and decrementing the l2ad_alloc refcount, we 3346 * must remove the old pointer (that we're about to destroy) and 3347 * add the new pointer to the refcount. Otherwise we'd remove 3348 * the wrong pointer address when calling arc_hdr_destroy() later. 3349 */ 3350 3351 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); 3352 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr); 3353 3354 buf_discard_identity(hdr); 3355 kmem_cache_free(old, hdr); 3356 3357 return (nhdr); 3358} 3359 3360/* 3361 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. 3362 * The buf is returned thawed since we expect the consumer to modify it. 3363 */ 3364arc_buf_t * 3365arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size) 3366{ 3367 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, 3368 ZIO_COMPRESS_OFF, type); 3369 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 3370 3371 arc_buf_t *buf = NULL; 3372 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf)); 3373 arc_buf_thaw(buf); 3374 3375 return (buf); 3376} 3377 3378/* 3379 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this 3380 * for bufs containing metadata. 3381 */ 3382arc_buf_t * 3383arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize, 3384 enum zio_compress compression_type) 3385{ 3386 ASSERT3U(lsize, >, 0); 3387 ASSERT3U(lsize, >=, psize); 3388 ASSERT(compression_type > ZIO_COMPRESS_OFF); 3389 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS); 3390 3391 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 3392 compression_type, ARC_BUFC_DATA); 3393 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 3394 3395 arc_buf_t *buf = NULL; 3396 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf)); 3397 arc_buf_thaw(buf); 3398 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3399 3400 if (!arc_buf_is_shared(buf)) { 3401 /* 3402 * To ensure that the hdr has the correct data in it if we call 3403 * arc_decompress() on this buf before it's been written to 3404 * disk, it's easiest if we just set up sharing between the 3405 * buf and the hdr. 3406 */ 3407 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd)); 3408 arc_hdr_free_pabd(hdr); 3409 arc_share_buf(hdr, buf); 3410 } 3411 3412 return (buf); 3413} 3414 3415static void 3416arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 3417{ 3418 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 3419 l2arc_dev_t *dev = l2hdr->b_dev; 3420 uint64_t psize = arc_hdr_size(hdr); 3421 3422 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 3423 ASSERT(HDR_HAS_L2HDR(hdr)); 3424 3425 list_remove(&dev->l2ad_buflist, hdr); 3426 3427 ARCSTAT_INCR(arcstat_l2_psize, -psize); 3428 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); 3429 3430 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0); 3431 3432 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr); 3433 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 3434} 3435 3436static void 3437arc_hdr_destroy(arc_buf_hdr_t *hdr) 3438{ 3439 if (HDR_HAS_L1HDR(hdr)) { 3440 ASSERT(hdr->b_l1hdr.b_buf == NULL || 3441 hdr->b_l1hdr.b_bufcnt > 0); 3442 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3443 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 3444 } 3445 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3446 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 3447 3448 if (!HDR_EMPTY(hdr)) 3449 buf_discard_identity(hdr); 3450 3451 if (HDR_HAS_L2HDR(hdr)) { 3452 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3453 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 3454 3455 if (!buflist_held) 3456 mutex_enter(&dev->l2ad_mtx); 3457 3458 /* 3459 * Even though we checked this conditional above, we 3460 * need to check this again now that we have the 3461 * l2ad_mtx. This is because we could be racing with 3462 * another thread calling l2arc_evict() which might have 3463 * destroyed this header's L2 portion as we were waiting 3464 * to acquire the l2ad_mtx. If that happens, we don't 3465 * want to re-destroy the header's L2 portion. 3466 */ 3467 if (HDR_HAS_L2HDR(hdr)) { 3468 l2arc_trim(hdr); 3469 arc_hdr_l2hdr_destroy(hdr); 3470 } 3471 3472 if (!buflist_held) 3473 mutex_exit(&dev->l2ad_mtx); 3474 } 3475 3476 if (HDR_HAS_L1HDR(hdr)) { 3477 arc_cksum_free(hdr); 3478 3479 while (hdr->b_l1hdr.b_buf != NULL) 3480 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf); 3481 3482#ifdef ZFS_DEBUG 3483 if (hdr->b_l1hdr.b_thawed != NULL) { 3484 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3485 hdr->b_l1hdr.b_thawed = NULL; 3486 } 3487#endif 3488 3489 if (hdr->b_l1hdr.b_pabd != NULL) { 3490 arc_hdr_free_pabd(hdr); 3491 } 3492 } 3493 3494 ASSERT3P(hdr->b_hash_next, ==, NULL); 3495 if (HDR_HAS_L1HDR(hdr)) { 3496 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3497 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 3498 kmem_cache_free(hdr_full_cache, hdr); 3499 } else { 3500 kmem_cache_free(hdr_l2only_cache, hdr); 3501 } 3502} 3503 3504void 3505arc_buf_destroy(arc_buf_t *buf, void* tag) 3506{ 3507 arc_buf_hdr_t *hdr = buf->b_hdr; 3508 kmutex_t *hash_lock = HDR_LOCK(hdr); 3509 3510 if (hdr->b_l1hdr.b_state == arc_anon) { 3511 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 3512 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3513 VERIFY0(remove_reference(hdr, NULL, tag)); 3514 arc_hdr_destroy(hdr); 3515 return; 3516 } 3517 3518 mutex_enter(hash_lock); 3519 ASSERT3P(hdr, ==, buf->b_hdr); 3520 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 3521 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3522 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); 3523 ASSERT3P(buf->b_data, !=, NULL); 3524 3525 (void) remove_reference(hdr, hash_lock, tag); 3526 arc_buf_destroy_impl(buf); 3527 mutex_exit(hash_lock); 3528} 3529 3530/* 3531 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 3532 * state of the header is dependent on it's state prior to entering this 3533 * function. The following transitions are possible: 3534 * 3535 * - arc_mru -> arc_mru_ghost 3536 * - arc_mfu -> arc_mfu_ghost 3537 * - arc_mru_ghost -> arc_l2c_only 3538 * - arc_mru_ghost -> deleted 3539 * - arc_mfu_ghost -> arc_l2c_only 3540 * - arc_mfu_ghost -> deleted 3541 */ 3542static int64_t 3543arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3544{ 3545 arc_state_t *evicted_state, *state; 3546 int64_t bytes_evicted = 0; 3547 3548 ASSERT(MUTEX_HELD(hash_lock)); 3549 ASSERT(HDR_HAS_L1HDR(hdr)); 3550 3551 state = hdr->b_l1hdr.b_state; 3552 if (GHOST_STATE(state)) { 3553 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3554 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3555 3556 /* 3557 * l2arc_write_buffers() relies on a header's L1 portion 3558 * (i.e. its b_pabd field) during it's write phase. 3559 * Thus, we cannot push a header onto the arc_l2c_only 3560 * state (removing it's L1 piece) until the header is 3561 * done being written to the l2arc. 3562 */ 3563 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 3564 ARCSTAT_BUMP(arcstat_evict_l2_skip); 3565 return (bytes_evicted); 3566 } 3567 3568 ARCSTAT_BUMP(arcstat_deleted); 3569 bytes_evicted += HDR_GET_LSIZE(hdr); 3570 3571 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 3572 3573 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3574 if (HDR_HAS_L2HDR(hdr)) { 3575 /* 3576 * This buffer is cached on the 2nd Level ARC; 3577 * don't destroy the header. 3578 */ 3579 arc_change_state(arc_l2c_only, hdr, hash_lock); 3580 /* 3581 * dropping from L1+L2 cached to L2-only, 3582 * realloc to remove the L1 header. 3583 */ 3584 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 3585 hdr_l2only_cache); 3586 } else { 3587 arc_change_state(arc_anon, hdr, hash_lock); 3588 arc_hdr_destroy(hdr); 3589 } 3590 return (bytes_evicted); 3591 } 3592 3593 ASSERT(state == arc_mru || state == arc_mfu); 3594 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3595 3596 /* prefetch buffers have a minimum lifespan */ 3597 if (HDR_IO_IN_PROGRESS(hdr) || 3598 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 3599 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 3600 arc_min_prefetch_lifespan)) { 3601 ARCSTAT_BUMP(arcstat_evict_skip); 3602 return (bytes_evicted); 3603 } 3604 3605 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3606 while (hdr->b_l1hdr.b_buf) { 3607 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3608 if (!mutex_tryenter(&buf->b_evict_lock)) { 3609 ARCSTAT_BUMP(arcstat_mutex_miss); 3610 break; 3611 } 3612 if (buf->b_data != NULL) 3613 bytes_evicted += HDR_GET_LSIZE(hdr); 3614 mutex_exit(&buf->b_evict_lock); 3615 arc_buf_destroy_impl(buf); 3616 } 3617 3618 if (HDR_HAS_L2HDR(hdr)) { 3619 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); 3620 } else { 3621 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 3622 ARCSTAT_INCR(arcstat_evict_l2_eligible, 3623 HDR_GET_LSIZE(hdr)); 3624 } else { 3625 ARCSTAT_INCR(arcstat_evict_l2_ineligible, 3626 HDR_GET_LSIZE(hdr)); 3627 } 3628 } 3629 3630 if (hdr->b_l1hdr.b_bufcnt == 0) { 3631 arc_cksum_free(hdr); 3632 3633 bytes_evicted += arc_hdr_size(hdr); 3634 3635 /* 3636 * If this hdr is being evicted and has a compressed 3637 * buffer then we discard it here before we change states. 3638 * This ensures that the accounting is updated correctly 3639 * in arc_free_data_impl(). 3640 */ 3641 arc_hdr_free_pabd(hdr); 3642 3643 arc_change_state(evicted_state, hdr, hash_lock); 3644 ASSERT(HDR_IN_HASH_TABLE(hdr)); 3645 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 3646 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 3647 } 3648 3649 return (bytes_evicted); 3650} 3651 3652static uint64_t 3653arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 3654 uint64_t spa, int64_t bytes) 3655{ 3656 multilist_sublist_t *mls; 3657 uint64_t bytes_evicted = 0; 3658 arc_buf_hdr_t *hdr; 3659 kmutex_t *hash_lock; 3660 int evict_count = 0; 3661 3662 ASSERT3P(marker, !=, NULL); 3663 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3664 3665 mls = multilist_sublist_lock(ml, idx); 3666 3667 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 3668 hdr = multilist_sublist_prev(mls, marker)) { 3669 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 3670 (evict_count >= zfs_arc_evict_batch_limit)) 3671 break; 3672 3673 /* 3674 * To keep our iteration location, move the marker 3675 * forward. Since we're not holding hdr's hash lock, we 3676 * must be very careful and not remove 'hdr' from the 3677 * sublist. Otherwise, other consumers might mistake the 3678 * 'hdr' as not being on a sublist when they call the 3679 * multilist_link_active() function (they all rely on 3680 * the hash lock protecting concurrent insertions and 3681 * removals). multilist_sublist_move_forward() was 3682 * specifically implemented to ensure this is the case 3683 * (only 'marker' will be removed and re-inserted). 3684 */ 3685 multilist_sublist_move_forward(mls, marker); 3686 3687 /* 3688 * The only case where the b_spa field should ever be 3689 * zero, is the marker headers inserted by 3690 * arc_evict_state(). It's possible for multiple threads 3691 * to be calling arc_evict_state() concurrently (e.g. 3692 * dsl_pool_close() and zio_inject_fault()), so we must 3693 * skip any markers we see from these other threads. 3694 */ 3695 if (hdr->b_spa == 0) 3696 continue; 3697 3698 /* we're only interested in evicting buffers of a certain spa */ 3699 if (spa != 0 && hdr->b_spa != spa) { 3700 ARCSTAT_BUMP(arcstat_evict_skip); 3701 continue; 3702 } 3703 3704 hash_lock = HDR_LOCK(hdr); 3705 3706 /* 3707 * We aren't calling this function from any code path 3708 * that would already be holding a hash lock, so we're 3709 * asserting on this assumption to be defensive in case 3710 * this ever changes. Without this check, it would be 3711 * possible to incorrectly increment arcstat_mutex_miss 3712 * below (e.g. if the code changed such that we called 3713 * this function with a hash lock held). 3714 */ 3715 ASSERT(!MUTEX_HELD(hash_lock)); 3716 3717 if (mutex_tryenter(hash_lock)) { 3718 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 3719 mutex_exit(hash_lock); 3720 3721 bytes_evicted += evicted; 3722 3723 /* 3724 * If evicted is zero, arc_evict_hdr() must have 3725 * decided to skip this header, don't increment 3726 * evict_count in this case. 3727 */ 3728 if (evicted != 0) 3729 evict_count++; 3730 3731 /* 3732 * If arc_size isn't overflowing, signal any 3733 * threads that might happen to be waiting. 3734 * 3735 * For each header evicted, we wake up a single 3736 * thread. If we used cv_broadcast, we could 3737 * wake up "too many" threads causing arc_size 3738 * to significantly overflow arc_c; since 3739 * arc_get_data_impl() doesn't check for overflow 3740 * when it's woken up (it doesn't because it's 3741 * possible for the ARC to be overflowing while 3742 * full of un-evictable buffers, and the 3743 * function should proceed in this case). 3744 * 3745 * If threads are left sleeping, due to not 3746 * using cv_broadcast, they will be woken up 3747 * just before arc_reclaim_thread() sleeps. 3748 */ 3749 mutex_enter(&arc_reclaim_lock); 3750 if (!arc_is_overflowing()) 3751 cv_signal(&arc_reclaim_waiters_cv); 3752 mutex_exit(&arc_reclaim_lock); 3753 } else { 3754 ARCSTAT_BUMP(arcstat_mutex_miss); 3755 } 3756 } 3757 3758 multilist_sublist_unlock(mls); 3759 3760 return (bytes_evicted); 3761} 3762 3763/* 3764 * Evict buffers from the given arc state, until we've removed the 3765 * specified number of bytes. Move the removed buffers to the 3766 * appropriate evict state. 3767 * 3768 * This function makes a "best effort". It skips over any buffers 3769 * it can't get a hash_lock on, and so, may not catch all candidates. 3770 * It may also return without evicting as much space as requested. 3771 * 3772 * If bytes is specified using the special value ARC_EVICT_ALL, this 3773 * will evict all available (i.e. unlocked and evictable) buffers from 3774 * the given arc state; which is used by arc_flush(). 3775 */ 3776static uint64_t 3777arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 3778 arc_buf_contents_t type) 3779{ 3780 uint64_t total_evicted = 0; 3781 multilist_t *ml = state->arcs_list[type]; 3782 int num_sublists; 3783 arc_buf_hdr_t **markers; 3784 3785 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3786 3787 num_sublists = multilist_get_num_sublists(ml); 3788 3789 /* 3790 * If we've tried to evict from each sublist, made some 3791 * progress, but still have not hit the target number of bytes 3792 * to evict, we want to keep trying. The markers allow us to 3793 * pick up where we left off for each individual sublist, rather 3794 * than starting from the tail each time. 3795 */ 3796 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 3797 for (int i = 0; i < num_sublists; i++) { 3798 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 3799 3800 /* 3801 * A b_spa of 0 is used to indicate that this header is 3802 * a marker. This fact is used in arc_adjust_type() and 3803 * arc_evict_state_impl(). 3804 */ 3805 markers[i]->b_spa = 0; 3806 3807 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3808 multilist_sublist_insert_tail(mls, markers[i]); 3809 multilist_sublist_unlock(mls); 3810 } 3811 3812 /* 3813 * While we haven't hit our target number of bytes to evict, or 3814 * we're evicting all available buffers. 3815 */ 3816 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 3817 /* 3818 * Start eviction using a randomly selected sublist, 3819 * this is to try and evenly balance eviction across all 3820 * sublists. Always starting at the same sublist 3821 * (e.g. index 0) would cause evictions to favor certain 3822 * sublists over others. 3823 */ 3824 int sublist_idx = multilist_get_random_index(ml); 3825 uint64_t scan_evicted = 0; 3826 3827 for (int i = 0; i < num_sublists; i++) { 3828 uint64_t bytes_remaining; 3829 uint64_t bytes_evicted; 3830 3831 if (bytes == ARC_EVICT_ALL) 3832 bytes_remaining = ARC_EVICT_ALL; 3833 else if (total_evicted < bytes) 3834 bytes_remaining = bytes - total_evicted; 3835 else 3836 break; 3837 3838 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 3839 markers[sublist_idx], spa, bytes_remaining); 3840 3841 scan_evicted += bytes_evicted; 3842 total_evicted += bytes_evicted; 3843 3844 /* we've reached the end, wrap to the beginning */ 3845 if (++sublist_idx >= num_sublists) 3846 sublist_idx = 0; 3847 } 3848 3849 /* 3850 * If we didn't evict anything during this scan, we have 3851 * no reason to believe we'll evict more during another 3852 * scan, so break the loop. 3853 */ 3854 if (scan_evicted == 0) { 3855 /* This isn't possible, let's make that obvious */ 3856 ASSERT3S(bytes, !=, 0); 3857 3858 /* 3859 * When bytes is ARC_EVICT_ALL, the only way to 3860 * break the loop is when scan_evicted is zero. 3861 * In that case, we actually have evicted enough, 3862 * so we don't want to increment the kstat. 3863 */ 3864 if (bytes != ARC_EVICT_ALL) { 3865 ASSERT3S(total_evicted, <, bytes); 3866 ARCSTAT_BUMP(arcstat_evict_not_enough); 3867 } 3868 3869 break; 3870 } 3871 } 3872 3873 for (int i = 0; i < num_sublists; i++) { 3874 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3875 multilist_sublist_remove(mls, markers[i]); 3876 multilist_sublist_unlock(mls); 3877 3878 kmem_cache_free(hdr_full_cache, markers[i]); 3879 } 3880 kmem_free(markers, sizeof (*markers) * num_sublists); 3881 3882 return (total_evicted); 3883} 3884 3885/* 3886 * Flush all "evictable" data of the given type from the arc state 3887 * specified. This will not evict any "active" buffers (i.e. referenced). 3888 * 3889 * When 'retry' is set to B_FALSE, the function will make a single pass 3890 * over the state and evict any buffers that it can. Since it doesn't 3891 * continually retry the eviction, it might end up leaving some buffers 3892 * in the ARC due to lock misses. 3893 * 3894 * When 'retry' is set to B_TRUE, the function will continually retry the 3895 * eviction until *all* evictable buffers have been removed from the 3896 * state. As a result, if concurrent insertions into the state are 3897 * allowed (e.g. if the ARC isn't shutting down), this function might 3898 * wind up in an infinite loop, continually trying to evict buffers. 3899 */ 3900static uint64_t 3901arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 3902 boolean_t retry) 3903{ 3904 uint64_t evicted = 0; 3905 3906 while (refcount_count(&state->arcs_esize[type]) != 0) { 3907 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 3908 3909 if (!retry) 3910 break; 3911 } 3912 3913 return (evicted); 3914} 3915 3916/* 3917 * Evict the specified number of bytes from the state specified, 3918 * restricting eviction to the spa and type given. This function 3919 * prevents us from trying to evict more from a state's list than 3920 * is "evictable", and to skip evicting altogether when passed a 3921 * negative value for "bytes". In contrast, arc_evict_state() will 3922 * evict everything it can, when passed a negative value for "bytes". 3923 */ 3924static uint64_t 3925arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3926 arc_buf_contents_t type) 3927{ 3928 int64_t delta; 3929 3930 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) { 3931 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes); 3932 return (arc_evict_state(state, spa, delta, type)); 3933 } 3934 3935 return (0); 3936} 3937 3938/* 3939 * Evict metadata buffers from the cache, such that arc_meta_used is 3940 * capped by the arc_meta_limit tunable. 3941 */ 3942static uint64_t 3943arc_adjust_meta(uint64_t meta_used) 3944{ 3945 uint64_t total_evicted = 0; 3946 int64_t target; 3947 3948 /* 3949 * If we're over the meta limit, we want to evict enough 3950 * metadata to get back under the meta limit. We don't want to 3951 * evict so much that we drop the MRU below arc_p, though. If 3952 * we're over the meta limit more than we're over arc_p, we 3953 * evict some from the MRU here, and some from the MFU below. 3954 */ 3955 target = MIN((int64_t)(meta_used - arc_meta_limit), 3956 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3957 refcount_count(&arc_mru->arcs_size) - arc_p)); 3958 3959 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3960 3961 /* 3962 * Similar to the above, we want to evict enough bytes to get us 3963 * below the meta limit, but not so much as to drop us below the 3964 * space allotted to the MFU (which is defined as arc_c - arc_p). 3965 */ 3966 target = MIN((int64_t)(meta_used - arc_meta_limit), 3967 (int64_t)(refcount_count(&arc_mfu->arcs_size) - 3968 (arc_c - arc_p))); 3969 3970 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3971 3972 return (total_evicted); 3973} 3974 3975/* 3976 * Return the type of the oldest buffer in the given arc state 3977 * 3978 * This function will select a random sublist of type ARC_BUFC_DATA and 3979 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3980 * is compared, and the type which contains the "older" buffer will be 3981 * returned. 3982 */ 3983static arc_buf_contents_t 3984arc_adjust_type(arc_state_t *state) 3985{ 3986 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA]; 3987 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA]; 3988 int data_idx = multilist_get_random_index(data_ml); 3989 int meta_idx = multilist_get_random_index(meta_ml); 3990 multilist_sublist_t *data_mls; 3991 multilist_sublist_t *meta_mls; 3992 arc_buf_contents_t type; 3993 arc_buf_hdr_t *data_hdr; 3994 arc_buf_hdr_t *meta_hdr; 3995 3996 /* 3997 * We keep the sublist lock until we're finished, to prevent 3998 * the headers from being destroyed via arc_evict_state(). 3999 */ 4000 data_mls = multilist_sublist_lock(data_ml, data_idx); 4001 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 4002 4003 /* 4004 * These two loops are to ensure we skip any markers that 4005 * might be at the tail of the lists due to arc_evict_state(). 4006 */ 4007 4008 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 4009 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 4010 if (data_hdr->b_spa != 0) 4011 break; 4012 } 4013 4014 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 4015 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 4016 if (meta_hdr->b_spa != 0) 4017 break; 4018 } 4019 4020 if (data_hdr == NULL && meta_hdr == NULL) { 4021 type = ARC_BUFC_DATA; 4022 } else if (data_hdr == NULL) { 4023 ASSERT3P(meta_hdr, !=, NULL); 4024 type = ARC_BUFC_METADATA; 4025 } else if (meta_hdr == NULL) { 4026 ASSERT3P(data_hdr, !=, NULL); 4027 type = ARC_BUFC_DATA; 4028 } else { 4029 ASSERT3P(data_hdr, !=, NULL); 4030 ASSERT3P(meta_hdr, !=, NULL); 4031 4032 /* The headers can't be on the sublist without an L1 header */ 4033 ASSERT(HDR_HAS_L1HDR(data_hdr)); 4034 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 4035 4036 if (data_hdr->b_l1hdr.b_arc_access < 4037 meta_hdr->b_l1hdr.b_arc_access) { 4038 type = ARC_BUFC_DATA; 4039 } else { 4040 type = ARC_BUFC_METADATA; 4041 } 4042 } 4043 4044 multilist_sublist_unlock(meta_mls); 4045 multilist_sublist_unlock(data_mls); 4046 4047 return (type); 4048} 4049 4050/* 4051 * Evict buffers from the cache, such that arc_size is capped by arc_c. 4052 */ 4053static uint64_t 4054arc_adjust(void) 4055{ 4056 uint64_t total_evicted = 0; 4057 uint64_t bytes; 4058 int64_t target; 4059 uint64_t asize = aggsum_value(&arc_size); 4060 uint64_t ameta = aggsum_value(&arc_meta_used); 4061 4062 /* 4063 * If we're over arc_meta_limit, we want to correct that before 4064 * potentially evicting data buffers below. 4065 */ 4066 total_evicted += arc_adjust_meta(ameta); 4067 4068 /* 4069 * Adjust MRU size 4070 * 4071 * If we're over the target cache size, we want to evict enough 4072 * from the list to get back to our target size. We don't want 4073 * to evict too much from the MRU, such that it drops below 4074 * arc_p. So, if we're over our target cache size more than 4075 * the MRU is over arc_p, we'll evict enough to get back to 4076 * arc_p here, and then evict more from the MFU below. 4077 */ 4078 target = MIN((int64_t)(asize - arc_c), 4079 (int64_t)(refcount_count(&arc_anon->arcs_size) + 4080 refcount_count(&arc_mru->arcs_size) + ameta - arc_p)); 4081 4082 /* 4083 * If we're below arc_meta_min, always prefer to evict data. 4084 * Otherwise, try to satisfy the requested number of bytes to 4085 * evict from the type which contains older buffers; in an 4086 * effort to keep newer buffers in the cache regardless of their 4087 * type. If we cannot satisfy the number of bytes from this 4088 * type, spill over into the next type. 4089 */ 4090 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 4091 ameta > arc_meta_min) { 4092 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 4093 total_evicted += bytes; 4094 4095 /* 4096 * If we couldn't evict our target number of bytes from 4097 * metadata, we try to get the rest from data. 4098 */ 4099 target -= bytes; 4100 4101 total_evicted += 4102 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 4103 } else { 4104 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 4105 total_evicted += bytes; 4106 4107 /* 4108 * If we couldn't evict our target number of bytes from 4109 * data, we try to get the rest from metadata. 4110 */ 4111 target -= bytes; 4112 4113 total_evicted += 4114 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 4115 } 4116 4117 /* 4118 * Adjust MFU size 4119 * 4120 * Now that we've tried to evict enough from the MRU to get its 4121 * size back to arc_p, if we're still above the target cache 4122 * size, we evict the rest from the MFU. 4123 */ 4124 target = asize - arc_c; 4125 4126 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 4127 ameta > arc_meta_min) { 4128 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 4129 total_evicted += bytes; 4130 4131 /* 4132 * If we couldn't evict our target number of bytes from 4133 * metadata, we try to get the rest from data. 4134 */ 4135 target -= bytes; 4136 4137 total_evicted += 4138 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 4139 } else { 4140 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 4141 total_evicted += bytes; 4142 4143 /* 4144 * If we couldn't evict our target number of bytes from 4145 * data, we try to get the rest from data. 4146 */ 4147 target -= bytes; 4148 4149 total_evicted += 4150 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 4151 } 4152 4153 /* 4154 * Adjust ghost lists 4155 * 4156 * In addition to the above, the ARC also defines target values 4157 * for the ghost lists. The sum of the mru list and mru ghost 4158 * list should never exceed the target size of the cache, and 4159 * the sum of the mru list, mfu list, mru ghost list, and mfu 4160 * ghost list should never exceed twice the target size of the 4161 * cache. The following logic enforces these limits on the ghost 4162 * caches, and evicts from them as needed. 4163 */ 4164 target = refcount_count(&arc_mru->arcs_size) + 4165 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 4166 4167 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 4168 total_evicted += bytes; 4169 4170 target -= bytes; 4171 4172 total_evicted += 4173 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 4174 4175 /* 4176 * We assume the sum of the mru list and mfu list is less than 4177 * or equal to arc_c (we enforced this above), which means we 4178 * can use the simpler of the two equations below: 4179 * 4180 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 4181 * mru ghost + mfu ghost <= arc_c 4182 */ 4183 target = refcount_count(&arc_mru_ghost->arcs_size) + 4184 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 4185 4186 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 4187 total_evicted += bytes; 4188 4189 target -= bytes; 4190 4191 total_evicted += 4192 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 4193 4194 return (total_evicted); 4195} 4196 4197void 4198arc_flush(spa_t *spa, boolean_t retry) 4199{ 4200 uint64_t guid = 0; 4201 4202 /* 4203 * If retry is B_TRUE, a spa must not be specified since we have 4204 * no good way to determine if all of a spa's buffers have been 4205 * evicted from an arc state. 4206 */ 4207 ASSERT(!retry || spa == 0); 4208 4209 if (spa != NULL) 4210 guid = spa_load_guid(spa); 4211 4212 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 4213 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 4214 4215 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 4216 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 4217 4218 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 4219 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 4220 4221 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 4222 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 4223} 4224 4225void 4226arc_shrink(int64_t to_free) 4227{ 4228 uint64_t asize = aggsum_value(&arc_size); 4229 if (arc_c > arc_c_min) { 4230 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 4231 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 4232 if (arc_c > arc_c_min + to_free) 4233 atomic_add_64(&arc_c, -to_free); 4234 else 4235 arc_c = arc_c_min; 4236 4237 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 4238 if (asize < arc_c) 4239 arc_c = MAX(asize, arc_c_min); 4240 if (arc_p > arc_c) 4241 arc_p = (arc_c >> 1); 4242 4243 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 4244 arc_p); 4245 4246 ASSERT(arc_c >= arc_c_min); 4247 ASSERT((int64_t)arc_p >= 0); 4248 } 4249 4250 if (asize > arc_c) { 4251 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, asize, 4252 uint64_t, arc_c); 4253 (void) arc_adjust(); 4254 } 4255} 4256 4257typedef enum free_memory_reason_t { 4258 FMR_UNKNOWN, 4259 FMR_NEEDFREE, 4260 FMR_LOTSFREE, 4261 FMR_SWAPFS_MINFREE, 4262 FMR_PAGES_PP_MAXIMUM, 4263 FMR_HEAP_ARENA, 4264 FMR_ZIO_ARENA, 4265 FMR_ZIO_FRAG, 4266} free_memory_reason_t; 4267 4268int64_t last_free_memory; 4269free_memory_reason_t last_free_reason; 4270 4271/* 4272 * Additional reserve of pages for pp_reserve. 4273 */ 4274int64_t arc_pages_pp_reserve = 64; 4275 4276/* 4277 * Additional reserve of pages for swapfs. 4278 */ 4279int64_t arc_swapfs_reserve = 64; 4280 4281/* 4282 * Return the amount of memory that can be consumed before reclaim will be 4283 * needed. Positive if there is sufficient free memory, negative indicates 4284 * the amount of memory that needs to be freed up. 4285 */ 4286static int64_t 4287arc_available_memory(void) 4288{ 4289 int64_t lowest = INT64_MAX; 4290 int64_t n; 4291 free_memory_reason_t r = FMR_UNKNOWN; 4292 4293#ifdef _KERNEL 4294#ifdef __FreeBSD__ 4295 /* 4296 * Cooperate with pagedaemon when it's time for it to scan 4297 * and reclaim some pages. 4298 */ 4299 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 4300 if (n < lowest) { 4301 lowest = n; 4302 r = FMR_LOTSFREE; 4303 } 4304 4305#else 4306 if (needfree > 0) { 4307 n = PAGESIZE * (-needfree); 4308 if (n < lowest) { 4309 lowest = n; 4310 r = FMR_NEEDFREE; 4311 } 4312 } 4313 4314 /* 4315 * check that we're out of range of the pageout scanner. It starts to 4316 * schedule paging if freemem is less than lotsfree and needfree. 4317 * lotsfree is the high-water mark for pageout, and needfree is the 4318 * number of needed free pages. We add extra pages here to make sure 4319 * the scanner doesn't start up while we're freeing memory. 4320 */ 4321 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 4322 if (n < lowest) { 4323 lowest = n; 4324 r = FMR_LOTSFREE; 4325 } 4326 4327 /* 4328 * check to make sure that swapfs has enough space so that anon 4329 * reservations can still succeed. anon_resvmem() checks that the 4330 * availrmem is greater than swapfs_minfree, and the number of reserved 4331 * swap pages. We also add a bit of extra here just to prevent 4332 * circumstances from getting really dire. 4333 */ 4334 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 4335 desfree - arc_swapfs_reserve); 4336 if (n < lowest) { 4337 lowest = n; 4338 r = FMR_SWAPFS_MINFREE; 4339 } 4340 4341 4342 /* 4343 * Check that we have enough availrmem that memory locking (e.g., via 4344 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 4345 * stores the number of pages that cannot be locked; when availrmem 4346 * drops below pages_pp_maximum, page locking mechanisms such as 4347 * page_pp_lock() will fail.) 4348 */ 4349 n = PAGESIZE * (availrmem - pages_pp_maximum - 4350 arc_pages_pp_reserve); 4351 if (n < lowest) { 4352 lowest = n; 4353 r = FMR_PAGES_PP_MAXIMUM; 4354 } 4355 4356#endif /* __FreeBSD__ */ 4357#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4358 /* 4359 * If we're on an i386 platform, it's possible that we'll exhaust the 4360 * kernel heap space before we ever run out of available physical 4361 * memory. Most checks of the size of the heap_area compare against 4362 * tune.t_minarmem, which is the minimum available real memory that we 4363 * can have in the system. However, this is generally fixed at 25 pages 4364 * which is so low that it's useless. In this comparison, we seek to 4365 * calculate the total heap-size, and reclaim if more than 3/4ths of the 4366 * heap is allocated. (Or, in the calculation, if less than 1/4th is 4367 * free) 4368 */ 4369 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 4370 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 4371 if (n < lowest) { 4372 lowest = n; 4373 r = FMR_HEAP_ARENA; 4374 } 4375#define zio_arena NULL 4376#else 4377#define zio_arena heap_arena 4378#endif 4379 4380 /* 4381 * If zio data pages are being allocated out of a separate heap segment, 4382 * then enforce that the size of available vmem for this arena remains 4383 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free. 4384 * 4385 * Note that reducing the arc_zio_arena_free_shift keeps more virtual 4386 * memory (in the zio_arena) free, which can avoid memory 4387 * fragmentation issues. 4388 */ 4389 if (zio_arena != NULL) { 4390 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 4391 (vmem_size(zio_arena, VMEM_ALLOC) >> 4392 arc_zio_arena_free_shift); 4393 if (n < lowest) { 4394 lowest = n; 4395 r = FMR_ZIO_ARENA; 4396 } 4397 } 4398 4399 /* 4400 * Above limits know nothing about real level of KVA fragmentation. 4401 * Start aggressive reclamation if too little sequential KVA left. 4402 */ 4403 if (lowest > 0) { 4404 n = (vmem_size(heap_arena, VMEM_MAXFREE) < SPA_MAXBLOCKSIZE) ? 4405 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 4406 INT64_MAX; 4407 if (n < lowest) { 4408 lowest = n; 4409 r = FMR_ZIO_FRAG; 4410 } 4411 } 4412 4413#else /* _KERNEL */ 4414 /* Every 100 calls, free a small amount */ 4415 if (spa_get_random(100) == 0) 4416 lowest = -1024; 4417#endif /* _KERNEL */ 4418 4419 last_free_memory = lowest; 4420 last_free_reason = r; 4421 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 4422 return (lowest); 4423} 4424 4425 4426/* 4427 * Determine if the system is under memory pressure and is asking 4428 * to reclaim memory. A return value of B_TRUE indicates that the system 4429 * is under memory pressure and that the arc should adjust accordingly. 4430 */ 4431static boolean_t 4432arc_reclaim_needed(void) 4433{ 4434 return (arc_available_memory() < 0); 4435} 4436 4437extern kmem_cache_t *zio_buf_cache[]; 4438extern kmem_cache_t *zio_data_buf_cache[]; 4439extern kmem_cache_t *range_seg_cache; 4440extern kmem_cache_t *abd_chunk_cache; 4441 4442static __noinline void 4443arc_kmem_reap_now(void) 4444{ 4445 size_t i; 4446 kmem_cache_t *prev_cache = NULL; 4447 kmem_cache_t *prev_data_cache = NULL; 4448 4449 DTRACE_PROBE(arc__kmem_reap_start); 4450#ifdef _KERNEL 4451 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) { 4452 /* 4453 * We are exceeding our meta-data cache limit. 4454 * Purge some DNLC entries to release holds on meta-data. 4455 */ 4456 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 4457 } 4458#if defined(__i386) 4459 /* 4460 * Reclaim unused memory from all kmem caches. 4461 */ 4462 kmem_reap(); 4463#endif 4464#endif 4465 4466 /* 4467 * If a kmem reap is already active, don't schedule more. We must 4468 * check for this because kmem_cache_reap_soon() won't actually 4469 * block on the cache being reaped (this is to prevent callers from 4470 * becoming implicitly blocked by a system-wide kmem reap -- which, 4471 * on a system with many, many full magazines, can take minutes). 4472 */ 4473 if (kmem_cache_reap_active()) 4474 return; 4475 4476 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 4477 if (zio_buf_cache[i] != prev_cache) { 4478 prev_cache = zio_buf_cache[i]; 4479 kmem_cache_reap_soon(zio_buf_cache[i]); 4480 } 4481 if (zio_data_buf_cache[i] != prev_data_cache) { 4482 prev_data_cache = zio_data_buf_cache[i]; 4483 kmem_cache_reap_soon(zio_data_buf_cache[i]); 4484 } 4485 } 4486 kmem_cache_reap_soon(abd_chunk_cache); 4487 kmem_cache_reap_soon(buf_cache); 4488 kmem_cache_reap_soon(hdr_full_cache); 4489 kmem_cache_reap_soon(hdr_l2only_cache); 4490 kmem_cache_reap_soon(range_seg_cache); 4491 4492#ifdef illumos 4493 if (zio_arena != NULL) { 4494 /* 4495 * Ask the vmem arena to reclaim unused memory from its 4496 * quantum caches. 4497 */ 4498 vmem_qcache_reap(zio_arena); 4499 } 4500#endif 4501 DTRACE_PROBE(arc__kmem_reap_end); 4502} 4503 4504/* 4505 * Threads can block in arc_get_data_impl() waiting for this thread to evict 4506 * enough data and signal them to proceed. When this happens, the threads in 4507 * arc_get_data_impl() are sleeping while holding the hash lock for their 4508 * particular arc header. Thus, we must be careful to never sleep on a 4509 * hash lock in this thread. This is to prevent the following deadlock: 4510 * 4511 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L", 4512 * waiting for the reclaim thread to signal it. 4513 * 4514 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 4515 * fails, and goes to sleep forever. 4516 * 4517 * This possible deadlock is avoided by always acquiring a hash lock 4518 * using mutex_tryenter() from arc_reclaim_thread(). 4519 */ 4520/* ARGSUSED */ 4521static void 4522arc_reclaim_thread(void *unused __unused) 4523{ 4524 hrtime_t growtime = 0; 4525 hrtime_t kmem_reap_time = 0; 4526 callb_cpr_t cpr; 4527 4528 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 4529 4530 mutex_enter(&arc_reclaim_lock); 4531 while (!arc_reclaim_thread_exit) { 4532 uint64_t evicted = 0; 4533 4534 /* 4535 * This is necessary in order for the mdb ::arc dcmd to 4536 * show up to date information. Since the ::arc command 4537 * does not call the kstat's update function, without 4538 * this call, the command may show stale stats for the 4539 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 4540 * with this change, the data might be up to 1 second 4541 * out of date; but that should suffice. The arc_state_t 4542 * structures can be queried directly if more accurate 4543 * information is needed. 4544 */ 4545 if (arc_ksp != NULL) 4546 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 4547 4548 mutex_exit(&arc_reclaim_lock); 4549 4550 /* 4551 * We call arc_adjust() before (possibly) calling 4552 * arc_kmem_reap_now(), so that we can wake up 4553 * arc_get_data_impl() sooner. 4554 */ 4555 evicted = arc_adjust(); 4556 4557 int64_t free_memory = arc_available_memory(); 4558 if (free_memory < 0) { 4559 hrtime_t curtime = gethrtime(); 4560 arc_no_grow = B_TRUE; 4561 arc_warm = B_TRUE; 4562 4563 /* 4564 * Wait at least zfs_grow_retry (default 60) seconds 4565 * before considering growing. 4566 */ 4567 growtime = curtime + SEC2NSEC(arc_grow_retry); 4568 4569 /* 4570 * Wait at least arc_kmem_cache_reap_retry_ms 4571 * between arc_kmem_reap_now() calls. Without 4572 * this check it is possible to end up in a 4573 * situation where we spend lots of time 4574 * reaping caches, while we're near arc_c_min. 4575 */ 4576 if (curtime >= kmem_reap_time) { 4577 arc_kmem_reap_now(); 4578 kmem_reap_time = gethrtime() + 4579 MSEC2NSEC(arc_kmem_cache_reap_retry_ms); 4580 } 4581 4582 /* 4583 * If we are still low on memory, shrink the ARC 4584 * so that we have arc_shrink_min free space. 4585 */ 4586 free_memory = arc_available_memory(); 4587 4588 int64_t to_free = 4589 (arc_c >> arc_shrink_shift) - free_memory; 4590 if (to_free > 0) { 4591#ifdef _KERNEL 4592#ifdef illumos 4593 to_free = MAX(to_free, ptob(needfree)); 4594#endif 4595#endif 4596 arc_shrink(to_free); 4597 } 4598 } else if (free_memory < arc_c >> arc_no_grow_shift) { 4599 arc_no_grow = B_TRUE; 4600 } else if (gethrtime() >= growtime) { 4601 arc_no_grow = B_FALSE; 4602 } 4603 4604 mutex_enter(&arc_reclaim_lock); 4605 4606 /* 4607 * If evicted is zero, we couldn't evict anything via 4608 * arc_adjust(). This could be due to hash lock 4609 * collisions, but more likely due to the majority of 4610 * arc buffers being unevictable. Therefore, even if 4611 * arc_size is above arc_c, another pass is unlikely to 4612 * be helpful and could potentially cause us to enter an 4613 * infinite loop. 4614 */ 4615 if (aggsum_compare(&arc_size, arc_c) <= 0|| evicted == 0) { 4616 /* 4617 * We're either no longer overflowing, or we 4618 * can't evict anything more, so we should wake 4619 * up any threads before we go to sleep. 4620 */ 4621 cv_broadcast(&arc_reclaim_waiters_cv); 4622 4623 /* 4624 * Block until signaled, or after one second (we 4625 * might need to perform arc_kmem_reap_now() 4626 * even if we aren't being signalled) 4627 */ 4628 CALLB_CPR_SAFE_BEGIN(&cpr); 4629 (void) cv_timedwait_hires(&arc_reclaim_thread_cv, 4630 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 4631 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 4632 } 4633 } 4634 4635 arc_reclaim_thread_exit = B_FALSE; 4636 cv_broadcast(&arc_reclaim_thread_cv); 4637 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 4638 thread_exit(); 4639} 4640 4641static u_int arc_dnlc_evicts_arg; 4642extern struct vfsops zfs_vfsops; 4643 4644static void 4645arc_dnlc_evicts_thread(void *dummy __unused) 4646{ 4647 callb_cpr_t cpr; 4648 u_int percent; 4649 4650 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG); 4651 4652 mutex_enter(&arc_dnlc_evicts_lock); 4653 while (!arc_dnlc_evicts_thread_exit) { 4654 CALLB_CPR_SAFE_BEGIN(&cpr); 4655 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock); 4656 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock); 4657 if (arc_dnlc_evicts_arg != 0) { 4658 percent = arc_dnlc_evicts_arg; 4659 mutex_exit(&arc_dnlc_evicts_lock); 4660#ifdef _KERNEL 4661 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops); 4662#endif 4663 mutex_enter(&arc_dnlc_evicts_lock); 4664 /* 4665 * Clear our token only after vnlru_free() 4666 * pass is done, to avoid false queueing of 4667 * the requests. 4668 */ 4669 arc_dnlc_evicts_arg = 0; 4670 } 4671 } 4672 arc_dnlc_evicts_thread_exit = FALSE; 4673 cv_broadcast(&arc_dnlc_evicts_cv); 4674 CALLB_CPR_EXIT(&cpr); 4675 thread_exit(); 4676} 4677 4678void 4679dnlc_reduce_cache(void *arg) 4680{ 4681 u_int percent; 4682 4683 percent = (u_int)(uintptr_t)arg; 4684 mutex_enter(&arc_dnlc_evicts_lock); 4685 if (arc_dnlc_evicts_arg == 0) { 4686 arc_dnlc_evicts_arg = percent; 4687 cv_broadcast(&arc_dnlc_evicts_cv); 4688 } 4689 mutex_exit(&arc_dnlc_evicts_lock); 4690} 4691 4692/* 4693 * Adapt arc info given the number of bytes we are trying to add and 4694 * the state that we are comming from. This function is only called 4695 * when we are adding new content to the cache. 4696 */ 4697static void 4698arc_adapt(int bytes, arc_state_t *state) 4699{ 4700 int mult; 4701 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 4702 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 4703 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 4704 4705 if (state == arc_l2c_only) 4706 return; 4707 4708 ASSERT(bytes > 0); 4709 /* 4710 * Adapt the target size of the MRU list: 4711 * - if we just hit in the MRU ghost list, then increase 4712 * the target size of the MRU list. 4713 * - if we just hit in the MFU ghost list, then increase 4714 * the target size of the MFU list by decreasing the 4715 * target size of the MRU list. 4716 */ 4717 if (state == arc_mru_ghost) { 4718 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 4719 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 4720 4721 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 4722 } else if (state == arc_mfu_ghost) { 4723 uint64_t delta; 4724 4725 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 4726 mult = MIN(mult, 10); 4727 4728 delta = MIN(bytes * mult, arc_p); 4729 arc_p = MAX(arc_p_min, arc_p - delta); 4730 } 4731 ASSERT((int64_t)arc_p >= 0); 4732 4733 if (arc_reclaim_needed()) { 4734 cv_signal(&arc_reclaim_thread_cv); 4735 return; 4736 } 4737 4738 if (arc_no_grow) 4739 return; 4740 4741 if (arc_c >= arc_c_max) 4742 return; 4743 4744 /* 4745 * If we're within (2 * maxblocksize) bytes of the target 4746 * cache size, increment the target cache size 4747 */ 4748 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) > 4749 0) { 4750 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 4751 atomic_add_64(&arc_c, (int64_t)bytes); 4752 if (arc_c > arc_c_max) 4753 arc_c = arc_c_max; 4754 else if (state == arc_anon) 4755 atomic_add_64(&arc_p, (int64_t)bytes); 4756 if (arc_p > arc_c) 4757 arc_p = arc_c; 4758 } 4759 ASSERT((int64_t)arc_p >= 0); 4760} 4761 4762/* 4763 * Check if arc_size has grown past our upper threshold, determined by 4764 * zfs_arc_overflow_shift. 4765 */ 4766static boolean_t 4767arc_is_overflowing(void) 4768{ 4769 /* Always allow at least one block of overflow */ 4770 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 4771 arc_c >> zfs_arc_overflow_shift); 4772 4773 /* 4774 * We just compare the lower bound here for performance reasons. Our 4775 * primary goals are to make sure that the arc never grows without 4776 * bound, and that it can reach its maximum size. This check 4777 * accomplishes both goals. The maximum amount we could run over by is 4778 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block 4779 * in the ARC. In practice, that's in the tens of MB, which is low 4780 * enough to be safe. 4781 */ 4782 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow); 4783} 4784 4785static abd_t * 4786arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4787{ 4788 arc_buf_contents_t type = arc_buf_type(hdr); 4789 4790 arc_get_data_impl(hdr, size, tag); 4791 if (type == ARC_BUFC_METADATA) { 4792 return (abd_alloc(size, B_TRUE)); 4793 } else { 4794 ASSERT(type == ARC_BUFC_DATA); 4795 return (abd_alloc(size, B_FALSE)); 4796 } 4797} 4798 4799static void * 4800arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4801{ 4802 arc_buf_contents_t type = arc_buf_type(hdr); 4803 4804 arc_get_data_impl(hdr, size, tag); 4805 if (type == ARC_BUFC_METADATA) { 4806 return (zio_buf_alloc(size)); 4807 } else { 4808 ASSERT(type == ARC_BUFC_DATA); 4809 return (zio_data_buf_alloc(size)); 4810 } 4811} 4812 4813/* 4814 * Allocate a block and return it to the caller. If we are hitting the 4815 * hard limit for the cache size, we must sleep, waiting for the eviction 4816 * thread to catch up. If we're past the target size but below the hard 4817 * limit, we'll only signal the reclaim thread and continue on. 4818 */ 4819static void 4820arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4821{ 4822 arc_state_t *state = hdr->b_l1hdr.b_state; 4823 arc_buf_contents_t type = arc_buf_type(hdr); 4824 4825 arc_adapt(size, state); 4826 4827 /* 4828 * If arc_size is currently overflowing, and has grown past our 4829 * upper limit, we must be adding data faster than the evict 4830 * thread can evict. Thus, to ensure we don't compound the 4831 * problem by adding more data and forcing arc_size to grow even 4832 * further past it's target size, we halt and wait for the 4833 * eviction thread to catch up. 4834 * 4835 * It's also possible that the reclaim thread is unable to evict 4836 * enough buffers to get arc_size below the overflow limit (e.g. 4837 * due to buffers being un-evictable, or hash lock collisions). 4838 * In this case, we want to proceed regardless if we're 4839 * overflowing; thus we don't use a while loop here. 4840 */ 4841 if (arc_is_overflowing()) { 4842 mutex_enter(&arc_reclaim_lock); 4843 4844 /* 4845 * Now that we've acquired the lock, we may no longer be 4846 * over the overflow limit, lets check. 4847 * 4848 * We're ignoring the case of spurious wake ups. If that 4849 * were to happen, it'd let this thread consume an ARC 4850 * buffer before it should have (i.e. before we're under 4851 * the overflow limit and were signalled by the reclaim 4852 * thread). As long as that is a rare occurrence, it 4853 * shouldn't cause any harm. 4854 */ 4855 if (arc_is_overflowing()) { 4856 cv_signal(&arc_reclaim_thread_cv); 4857 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 4858 } 4859 4860 mutex_exit(&arc_reclaim_lock); 4861 } 4862 4863 VERIFY3U(hdr->b_type, ==, type); 4864 if (type == ARC_BUFC_METADATA) { 4865 arc_space_consume(size, ARC_SPACE_META); 4866 } else { 4867 arc_space_consume(size, ARC_SPACE_DATA); 4868 } 4869 4870 /* 4871 * Update the state size. Note that ghost states have a 4872 * "ghost size" and so don't need to be updated. 4873 */ 4874 if (!GHOST_STATE(state)) { 4875 4876 (void) refcount_add_many(&state->arcs_size, size, tag); 4877 4878 /* 4879 * If this is reached via arc_read, the link is 4880 * protected by the hash lock. If reached via 4881 * arc_buf_alloc, the header should not be accessed by 4882 * any other thread. And, if reached via arc_read_done, 4883 * the hash lock will protect it if it's found in the 4884 * hash table; otherwise no other thread should be 4885 * trying to [add|remove]_reference it. 4886 */ 4887 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4888 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4889 (void) refcount_add_many(&state->arcs_esize[type], 4890 size, tag); 4891 } 4892 4893 /* 4894 * If we are growing the cache, and we are adding anonymous 4895 * data, and we have outgrown arc_p, update arc_p 4896 */ 4897 if (aggsum_compare(&arc_size, arc_c) < 0 && 4898 hdr->b_l1hdr.b_state == arc_anon && 4899 (refcount_count(&arc_anon->arcs_size) + 4900 refcount_count(&arc_mru->arcs_size) > arc_p)) 4901 arc_p = MIN(arc_c, arc_p + size); 4902 } 4903 ARCSTAT_BUMP(arcstat_allocated); 4904} 4905 4906static void 4907arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag) 4908{ 4909 arc_free_data_impl(hdr, size, tag); 4910 abd_free(abd); 4911} 4912 4913static void 4914arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag) 4915{ 4916 arc_buf_contents_t type = arc_buf_type(hdr); 4917 4918 arc_free_data_impl(hdr, size, tag); 4919 if (type == ARC_BUFC_METADATA) { 4920 zio_buf_free(buf, size); 4921 } else { 4922 ASSERT(type == ARC_BUFC_DATA); 4923 zio_data_buf_free(buf, size); 4924 } 4925} 4926 4927/* 4928 * Free the arc data buffer. 4929 */ 4930static void 4931arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4932{ 4933 arc_state_t *state = hdr->b_l1hdr.b_state; 4934 arc_buf_contents_t type = arc_buf_type(hdr); 4935 4936 /* protected by hash lock, if in the hash table */ 4937 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4938 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4939 ASSERT(state != arc_anon && state != arc_l2c_only); 4940 4941 (void) refcount_remove_many(&state->arcs_esize[type], 4942 size, tag); 4943 } 4944 (void) refcount_remove_many(&state->arcs_size, size, tag); 4945 4946 VERIFY3U(hdr->b_type, ==, type); 4947 if (type == ARC_BUFC_METADATA) { 4948 arc_space_return(size, ARC_SPACE_META); 4949 } else { 4950 ASSERT(type == ARC_BUFC_DATA); 4951 arc_space_return(size, ARC_SPACE_DATA); 4952 } 4953} 4954 4955/* 4956 * This routine is called whenever a buffer is accessed. 4957 * NOTE: the hash lock is dropped in this function. 4958 */ 4959static void 4960arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 4961{ 4962 clock_t now; 4963 4964 ASSERT(MUTEX_HELD(hash_lock)); 4965 ASSERT(HDR_HAS_L1HDR(hdr)); 4966 4967 if (hdr->b_l1hdr.b_state == arc_anon) { 4968 /* 4969 * This buffer is not in the cache, and does not 4970 * appear in our "ghost" list. Add the new buffer 4971 * to the MRU state. 4972 */ 4973 4974 ASSERT0(hdr->b_l1hdr.b_arc_access); 4975 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4976 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4977 arc_change_state(arc_mru, hdr, hash_lock); 4978 4979 } else if (hdr->b_l1hdr.b_state == arc_mru) { 4980 now = ddi_get_lbolt(); 4981 4982 /* 4983 * If this buffer is here because of a prefetch, then either: 4984 * - clear the flag if this is a "referencing" read 4985 * (any subsequent access will bump this into the MFU state). 4986 * or 4987 * - move the buffer to the head of the list if this is 4988 * another prefetch (to make it less likely to be evicted). 4989 */ 4990 if (HDR_PREFETCH(hdr)) { 4991 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4992 /* link protected by hash lock */ 4993 ASSERT(multilist_link_active( 4994 &hdr->b_l1hdr.b_arc_node)); 4995 } else { 4996 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 4997 ARCSTAT_BUMP(arcstat_mru_hits); 4998 } 4999 hdr->b_l1hdr.b_arc_access = now; 5000 return; 5001 } 5002 5003 /* 5004 * This buffer has been "accessed" only once so far, 5005 * but it is still in the cache. Move it to the MFU 5006 * state. 5007 */ 5008 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 5009 /* 5010 * More than 125ms have passed since we 5011 * instantiated this buffer. Move it to the 5012 * most frequently used state. 5013 */ 5014 hdr->b_l1hdr.b_arc_access = now; 5015 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 5016 arc_change_state(arc_mfu, hdr, hash_lock); 5017 } 5018 ARCSTAT_BUMP(arcstat_mru_hits); 5019 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 5020 arc_state_t *new_state; 5021 /* 5022 * This buffer has been "accessed" recently, but 5023 * was evicted from the cache. Move it to the 5024 * MFU state. 5025 */ 5026 5027 if (HDR_PREFETCH(hdr)) { 5028 new_state = arc_mru; 5029 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 5030 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 5031 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 5032 } else { 5033 new_state = arc_mfu; 5034 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 5035 } 5036 5037 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 5038 arc_change_state(new_state, hdr, hash_lock); 5039 5040 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 5041 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 5042 /* 5043 * This buffer has been accessed more than once and is 5044 * still in the cache. Keep it in the MFU state. 5045 * 5046 * NOTE: an add_reference() that occurred when we did 5047 * the arc_read() will have kicked this off the list. 5048 * If it was a prefetch, we will explicitly move it to 5049 * the head of the list now. 5050 */ 5051 if ((HDR_PREFETCH(hdr)) != 0) { 5052 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5053 /* link protected by hash_lock */ 5054 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5055 } 5056 ARCSTAT_BUMP(arcstat_mfu_hits); 5057 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 5058 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 5059 arc_state_t *new_state = arc_mfu; 5060 /* 5061 * This buffer has been accessed more than once but has 5062 * been evicted from the cache. Move it back to the 5063 * MFU state. 5064 */ 5065 5066 if (HDR_PREFETCH(hdr)) { 5067 /* 5068 * This is a prefetch access... 5069 * move this block back to the MRU state. 5070 */ 5071 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 5072 new_state = arc_mru; 5073 } 5074 5075 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 5076 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 5077 arc_change_state(new_state, hdr, hash_lock); 5078 5079 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 5080 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 5081 /* 5082 * This buffer is on the 2nd Level ARC. 5083 */ 5084 5085 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 5086 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 5087 arc_change_state(arc_mfu, hdr, hash_lock); 5088 } else { 5089 ASSERT(!"invalid arc state"); 5090 } 5091} 5092 5093/* 5094 * This routine is called by dbuf_hold() to update the arc_access() state 5095 * which otherwise would be skipped for entries in the dbuf cache. 5096 */ 5097void 5098arc_buf_access(arc_buf_t *buf) 5099{ 5100 mutex_enter(&buf->b_evict_lock); 5101 arc_buf_hdr_t *hdr = buf->b_hdr; 5102 5103 /* 5104 * Avoid taking the hash_lock when possible as an optimization. 5105 * The header must be checked again under the hash_lock in order 5106 * to handle the case where it is concurrently being released. 5107 */ 5108 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) { 5109 mutex_exit(&buf->b_evict_lock); 5110 ARCSTAT_BUMP(arcstat_access_skip); 5111 return; 5112 } 5113 5114 kmutex_t *hash_lock = HDR_LOCK(hdr); 5115 mutex_enter(hash_lock); 5116 5117 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) { 5118 mutex_exit(hash_lock); 5119 mutex_exit(&buf->b_evict_lock); 5120 ARCSTAT_BUMP(arcstat_access_skip); 5121 return; 5122 } 5123 5124 mutex_exit(&buf->b_evict_lock); 5125 5126 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 5127 hdr->b_l1hdr.b_state == arc_mfu); 5128 5129 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 5130 arc_access(hdr, hash_lock); 5131 mutex_exit(hash_lock); 5132 5133 ARCSTAT_BUMP(arcstat_hits); 5134 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5135 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits); 5136} 5137 5138/* a generic arc_done_func_t which you can use */ 5139/* ARGSUSED */ 5140void 5141arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 5142{ 5143 if (zio == NULL || zio->io_error == 0) 5144 bcopy(buf->b_data, arg, arc_buf_size(buf)); 5145 arc_buf_destroy(buf, arg); 5146} 5147 5148/* a generic arc_done_func_t */ 5149void 5150arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 5151{ 5152 arc_buf_t **bufp = arg; 5153 if (zio && zio->io_error) { 5154 arc_buf_destroy(buf, arg); 5155 *bufp = NULL; 5156 } else { 5157 *bufp = buf; 5158 ASSERT(buf->b_data); 5159 } 5160} 5161 5162static void 5163arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) 5164{ 5165 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { 5166 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); 5167 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 5168 } else { 5169 if (HDR_COMPRESSION_ENABLED(hdr)) { 5170 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, 5171 BP_GET_COMPRESS(bp)); 5172 } 5173 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); 5174 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); 5175 } 5176} 5177 5178static void 5179arc_read_done(zio_t *zio) 5180{ 5181 arc_buf_hdr_t *hdr = zio->io_private; 5182 kmutex_t *hash_lock = NULL; 5183 arc_callback_t *callback_list; 5184 arc_callback_t *acb; 5185 boolean_t freeable = B_FALSE; 5186 boolean_t no_zio_error = (zio->io_error == 0); 5187 5188 /* 5189 * The hdr was inserted into hash-table and removed from lists 5190 * prior to starting I/O. We should find this header, since 5191 * it's in the hash table, and it should be legit since it's 5192 * not possible to evict it during the I/O. The only possible 5193 * reason for it not to be found is if we were freed during the 5194 * read. 5195 */ 5196 if (HDR_IN_HASH_TABLE(hdr)) { 5197 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 5198 ASSERT3U(hdr->b_dva.dva_word[0], ==, 5199 BP_IDENTITY(zio->io_bp)->dva_word[0]); 5200 ASSERT3U(hdr->b_dva.dva_word[1], ==, 5201 BP_IDENTITY(zio->io_bp)->dva_word[1]); 5202 5203 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 5204 &hash_lock); 5205 5206 ASSERT((found == hdr && 5207 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 5208 (found == hdr && HDR_L2_READING(hdr))); 5209 ASSERT3P(hash_lock, !=, NULL); 5210 } 5211 5212 if (no_zio_error) { 5213 /* byteswap if necessary */ 5214 if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 5215 if (BP_GET_LEVEL(zio->io_bp) > 0) { 5216 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; 5217 } else { 5218 hdr->b_l1hdr.b_byteswap = 5219 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 5220 } 5221 } else { 5222 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 5223 } 5224 } 5225 5226 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); 5227 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 5228 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); 5229 5230 callback_list = hdr->b_l1hdr.b_acb; 5231 ASSERT3P(callback_list, !=, NULL); 5232 5233 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) { 5234 /* 5235 * Only call arc_access on anonymous buffers. This is because 5236 * if we've issued an I/O for an evicted buffer, we've already 5237 * called arc_access (to prevent any simultaneous readers from 5238 * getting confused). 5239 */ 5240 arc_access(hdr, hash_lock); 5241 } 5242 5243 /* 5244 * If a read request has a callback (i.e. acb_done is not NULL), then we 5245 * make a buf containing the data according to the parameters which were 5246 * passed in. The implementation of arc_buf_alloc_impl() ensures that we 5247 * aren't needlessly decompressing the data multiple times. 5248 */ 5249 int callback_cnt = 0; 5250 for (acb = callback_list; acb != NULL; acb = acb->acb_next) { 5251 if (!acb->acb_done) 5252 continue; 5253 5254 /* This is a demand read since prefetches don't use callbacks */ 5255 callback_cnt++; 5256 5257 int error = arc_buf_alloc_impl(hdr, acb->acb_private, 5258 acb->acb_compressed, no_zio_error, &acb->acb_buf); 5259 if (no_zio_error) { 5260 zio->io_error = error; 5261 } 5262 } 5263 hdr->b_l1hdr.b_acb = NULL; 5264 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5265 if (callback_cnt == 0) { 5266 ASSERT(HDR_PREFETCH(hdr)); 5267 ASSERT0(hdr->b_l1hdr.b_bufcnt); 5268 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5269 } 5270 5271 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 5272 callback_list != NULL); 5273 5274 if (no_zio_error) { 5275 arc_hdr_verify(hdr, zio->io_bp); 5276 } else { 5277 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); 5278 if (hdr->b_l1hdr.b_state != arc_anon) 5279 arc_change_state(arc_anon, hdr, hash_lock); 5280 if (HDR_IN_HASH_TABLE(hdr)) 5281 buf_hash_remove(hdr); 5282 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 5283 } 5284 5285 /* 5286 * Broadcast before we drop the hash_lock to avoid the possibility 5287 * that the hdr (and hence the cv) might be freed before we get to 5288 * the cv_broadcast(). 5289 */ 5290 cv_broadcast(&hdr->b_l1hdr.b_cv); 5291 5292 if (hash_lock != NULL) { 5293 mutex_exit(hash_lock); 5294 } else { 5295 /* 5296 * This block was freed while we waited for the read to 5297 * complete. It has been removed from the hash table and 5298 * moved to the anonymous state (so that it won't show up 5299 * in the cache). 5300 */ 5301 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 5302 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 5303 } 5304 5305 /* execute each callback and free its structure */ 5306 while ((acb = callback_list) != NULL) { 5307 if (acb->acb_done) 5308 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 5309 5310 if (acb->acb_zio_dummy != NULL) { 5311 acb->acb_zio_dummy->io_error = zio->io_error; 5312 zio_nowait(acb->acb_zio_dummy); 5313 } 5314 5315 callback_list = acb->acb_next; 5316 kmem_free(acb, sizeof (arc_callback_t)); 5317 } 5318 5319 if (freeable) 5320 arc_hdr_destroy(hdr); 5321} 5322 5323/* 5324 * "Read" the block at the specified DVA (in bp) via the 5325 * cache. If the block is found in the cache, invoke the provided 5326 * callback immediately and return. Note that the `zio' parameter 5327 * in the callback will be NULL in this case, since no IO was 5328 * required. If the block is not in the cache pass the read request 5329 * on to the spa with a substitute callback function, so that the 5330 * requested block will be added to the cache. 5331 * 5332 * If a read request arrives for a block that has a read in-progress, 5333 * either wait for the in-progress read to complete (and return the 5334 * results); or, if this is a read with a "done" func, add a record 5335 * to the read to invoke the "done" func when the read completes, 5336 * and return; or just return. 5337 * 5338 * arc_read_done() will invoke all the requested "done" functions 5339 * for readers of this block. 5340 */ 5341int 5342arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 5343 void *private, zio_priority_t priority, int zio_flags, 5344 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 5345{ 5346 arc_buf_hdr_t *hdr = NULL; 5347 kmutex_t *hash_lock = NULL; 5348 zio_t *rzio; 5349 uint64_t guid = spa_load_guid(spa); 5350 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0; 5351 5352 ASSERT(!BP_IS_EMBEDDED(bp) || 5353 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 5354 5355top: 5356 if (!BP_IS_EMBEDDED(bp)) { 5357 /* 5358 * Embedded BP's have no DVA and require no I/O to "read". 5359 * Create an anonymous arc buf to back it. 5360 */ 5361 hdr = buf_hash_find(guid, bp, &hash_lock); 5362 } 5363 5364 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) { 5365 arc_buf_t *buf = NULL; 5366 *arc_flags |= ARC_FLAG_CACHED; 5367 5368 if (HDR_IO_IN_PROGRESS(hdr)) { 5369 5370 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 5371 priority == ZIO_PRIORITY_SYNC_READ) { 5372 /* 5373 * This sync read must wait for an 5374 * in-progress async read (e.g. a predictive 5375 * prefetch). Async reads are queued 5376 * separately at the vdev_queue layer, so 5377 * this is a form of priority inversion. 5378 * Ideally, we would "inherit" the demand 5379 * i/o's priority by moving the i/o from 5380 * the async queue to the synchronous queue, 5381 * but there is currently no mechanism to do 5382 * so. Track this so that we can evaluate 5383 * the magnitude of this potential performance 5384 * problem. 5385 * 5386 * Note that if the prefetch i/o is already 5387 * active (has been issued to the device), 5388 * the prefetch improved performance, because 5389 * we issued it sooner than we would have 5390 * without the prefetch. 5391 */ 5392 DTRACE_PROBE1(arc__sync__wait__for__async, 5393 arc_buf_hdr_t *, hdr); 5394 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 5395 } 5396 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5397 arc_hdr_clear_flags(hdr, 5398 ARC_FLAG_PREDICTIVE_PREFETCH); 5399 } 5400 5401 if (*arc_flags & ARC_FLAG_WAIT) { 5402 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 5403 mutex_exit(hash_lock); 5404 goto top; 5405 } 5406 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5407 5408 if (done) { 5409 arc_callback_t *acb = NULL; 5410 5411 acb = kmem_zalloc(sizeof (arc_callback_t), 5412 KM_SLEEP); 5413 acb->acb_done = done; 5414 acb->acb_private = private; 5415 acb->acb_compressed = compressed_read; 5416 if (pio != NULL) 5417 acb->acb_zio_dummy = zio_null(pio, 5418 spa, NULL, NULL, NULL, zio_flags); 5419 5420 ASSERT3P(acb->acb_done, !=, NULL); 5421 acb->acb_next = hdr->b_l1hdr.b_acb; 5422 hdr->b_l1hdr.b_acb = acb; 5423 mutex_exit(hash_lock); 5424 return (0); 5425 } 5426 mutex_exit(hash_lock); 5427 return (0); 5428 } 5429 5430 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 5431 hdr->b_l1hdr.b_state == arc_mfu); 5432 5433 if (done) { 5434 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5435 /* 5436 * This is a demand read which does not have to 5437 * wait for i/o because we did a predictive 5438 * prefetch i/o for it, which has completed. 5439 */ 5440 DTRACE_PROBE1( 5441 arc__demand__hit__predictive__prefetch, 5442 arc_buf_hdr_t *, hdr); 5443 ARCSTAT_BUMP( 5444 arcstat_demand_hit_predictive_prefetch); 5445 arc_hdr_clear_flags(hdr, 5446 ARC_FLAG_PREDICTIVE_PREFETCH); 5447 } 5448 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); 5449 5450 /* Get a buf with the desired data in it. */ 5451 VERIFY0(arc_buf_alloc_impl(hdr, private, 5452 compressed_read, B_TRUE, &buf)); 5453 } else if (*arc_flags & ARC_FLAG_PREFETCH && 5454 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 5455 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5456 } 5457 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 5458 arc_access(hdr, hash_lock); 5459 if (*arc_flags & ARC_FLAG_L2CACHE) 5460 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5461 mutex_exit(hash_lock); 5462 ARCSTAT_BUMP(arcstat_hits); 5463 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5464 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5465 data, metadata, hits); 5466 5467 if (done) 5468 done(NULL, buf, private); 5469 } else { 5470 uint64_t lsize = BP_GET_LSIZE(bp); 5471 uint64_t psize = BP_GET_PSIZE(bp); 5472 arc_callback_t *acb; 5473 vdev_t *vd = NULL; 5474 uint64_t addr = 0; 5475 boolean_t devw = B_FALSE; 5476 uint64_t size; 5477 5478 if (hdr == NULL) { 5479 /* this block is not in the cache */ 5480 arc_buf_hdr_t *exists = NULL; 5481 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 5482 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 5483 BP_GET_COMPRESS(bp), type); 5484 5485 if (!BP_IS_EMBEDDED(bp)) { 5486 hdr->b_dva = *BP_IDENTITY(bp); 5487 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 5488 exists = buf_hash_insert(hdr, &hash_lock); 5489 } 5490 if (exists != NULL) { 5491 /* somebody beat us to the hash insert */ 5492 mutex_exit(hash_lock); 5493 buf_discard_identity(hdr); 5494 arc_hdr_destroy(hdr); 5495 goto top; /* restart the IO request */ 5496 } 5497 } else { 5498 /* 5499 * This block is in the ghost cache. If it was L2-only 5500 * (and thus didn't have an L1 hdr), we realloc the 5501 * header to add an L1 hdr. 5502 */ 5503 if (!HDR_HAS_L1HDR(hdr)) { 5504 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 5505 hdr_full_cache); 5506 } 5507 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5508 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 5509 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5510 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5511 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 5512 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 5513 5514 /* 5515 * This is a delicate dance that we play here. 5516 * This hdr is in the ghost list so we access it 5517 * to move it out of the ghost list before we 5518 * initiate the read. If it's a prefetch then 5519 * it won't have a callback so we'll remove the 5520 * reference that arc_buf_alloc_impl() created. We 5521 * do this after we've called arc_access() to 5522 * avoid hitting an assert in remove_reference(). 5523 */ 5524 arc_access(hdr, hash_lock); 5525 arc_hdr_alloc_pabd(hdr); 5526 } 5527 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5528 size = arc_hdr_size(hdr); 5529 5530 /* 5531 * If compression is enabled on the hdr, then will do 5532 * RAW I/O and will store the compressed data in the hdr's 5533 * data block. Otherwise, the hdr's data block will contain 5534 * the uncompressed data. 5535 */ 5536 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5537 zio_flags |= ZIO_FLAG_RAW; 5538 } 5539 5540 if (*arc_flags & ARC_FLAG_PREFETCH) 5541 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5542 if (*arc_flags & ARC_FLAG_L2CACHE) 5543 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5544 if (BP_GET_LEVEL(bp) > 0) 5545 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); 5546 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 5547 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); 5548 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 5549 5550 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 5551 acb->acb_done = done; 5552 acb->acb_private = private; 5553 acb->acb_compressed = compressed_read; 5554 5555 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5556 hdr->b_l1hdr.b_acb = acb; 5557 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5558 5559 if (HDR_HAS_L2HDR(hdr) && 5560 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 5561 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 5562 addr = hdr->b_l2hdr.b_daddr; 5563 /* 5564 * Lock out L2ARC device removal. 5565 */ 5566 if (vdev_is_dead(vd) || 5567 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 5568 vd = NULL; 5569 } 5570 5571 if (priority == ZIO_PRIORITY_ASYNC_READ) 5572 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5573 else 5574 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5575 5576 if (hash_lock != NULL) 5577 mutex_exit(hash_lock); 5578 5579 /* 5580 * At this point, we have a level 1 cache miss. Try again in 5581 * L2ARC if possible. 5582 */ 5583 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); 5584 5585 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 5586 uint64_t, lsize, zbookmark_phys_t *, zb); 5587 ARCSTAT_BUMP(arcstat_misses); 5588 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5589 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5590 data, metadata, misses); 5591#ifdef _KERNEL 5592#ifdef RACCT 5593 if (racct_enable) { 5594 PROC_LOCK(curproc); 5595 racct_add_force(curproc, RACCT_READBPS, size); 5596 racct_add_force(curproc, RACCT_READIOPS, 1); 5597 PROC_UNLOCK(curproc); 5598 } 5599#endif /* RACCT */ 5600 curthread->td_ru.ru_inblock++; 5601#endif 5602 5603 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 5604 /* 5605 * Read from the L2ARC if the following are true: 5606 * 1. The L2ARC vdev was previously cached. 5607 * 2. This buffer still has L2ARC metadata. 5608 * 3. This buffer isn't currently writing to the L2ARC. 5609 * 4. The L2ARC entry wasn't evicted, which may 5610 * also have invalidated the vdev. 5611 * 5. This isn't prefetch and l2arc_noprefetch is set. 5612 */ 5613 if (HDR_HAS_L2HDR(hdr) && 5614 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 5615 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 5616 l2arc_read_callback_t *cb; 5617 abd_t *abd; 5618 uint64_t asize; 5619 5620 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 5621 ARCSTAT_BUMP(arcstat_l2_hits); 5622 5623 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 5624 KM_SLEEP); 5625 cb->l2rcb_hdr = hdr; 5626 cb->l2rcb_bp = *bp; 5627 cb->l2rcb_zb = *zb; 5628 cb->l2rcb_flags = zio_flags; 5629 5630 asize = vdev_psize_to_asize(vd, size); 5631 if (asize != size) { 5632 abd = abd_alloc_for_io(asize, 5633 HDR_ISTYPE_METADATA(hdr)); 5634 cb->l2rcb_abd = abd; 5635 } else { 5636 abd = hdr->b_l1hdr.b_pabd; 5637 } 5638 5639 ASSERT(addr >= VDEV_LABEL_START_SIZE && 5640 addr + asize <= vd->vdev_psize - 5641 VDEV_LABEL_END_SIZE); 5642 5643 /* 5644 * l2arc read. The SCL_L2ARC lock will be 5645 * released by l2arc_read_done(). 5646 * Issue a null zio if the underlying buffer 5647 * was squashed to zero size by compression. 5648 */ 5649 ASSERT3U(HDR_GET_COMPRESS(hdr), !=, 5650 ZIO_COMPRESS_EMPTY); 5651 rzio = zio_read_phys(pio, vd, addr, 5652 asize, abd, 5653 ZIO_CHECKSUM_OFF, 5654 l2arc_read_done, cb, priority, 5655 zio_flags | ZIO_FLAG_DONT_CACHE | 5656 ZIO_FLAG_CANFAIL | 5657 ZIO_FLAG_DONT_PROPAGATE | 5658 ZIO_FLAG_DONT_RETRY, B_FALSE); 5659 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 5660 zio_t *, rzio); 5661 ARCSTAT_INCR(arcstat_l2_read_bytes, size); 5662 5663 if (*arc_flags & ARC_FLAG_NOWAIT) { 5664 zio_nowait(rzio); 5665 return (0); 5666 } 5667 5668 ASSERT(*arc_flags & ARC_FLAG_WAIT); 5669 if (zio_wait(rzio) == 0) 5670 return (0); 5671 5672 /* l2arc read error; goto zio_read() */ 5673 } else { 5674 DTRACE_PROBE1(l2arc__miss, 5675 arc_buf_hdr_t *, hdr); 5676 ARCSTAT_BUMP(arcstat_l2_misses); 5677 if (HDR_L2_WRITING(hdr)) 5678 ARCSTAT_BUMP(arcstat_l2_rw_clash); 5679 spa_config_exit(spa, SCL_L2ARC, vd); 5680 } 5681 } else { 5682 if (vd != NULL) 5683 spa_config_exit(spa, SCL_L2ARC, vd); 5684 if (l2arc_ndev != 0) { 5685 DTRACE_PROBE1(l2arc__miss, 5686 arc_buf_hdr_t *, hdr); 5687 ARCSTAT_BUMP(arcstat_l2_misses); 5688 } 5689 } 5690 5691 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size, 5692 arc_read_done, hdr, priority, zio_flags, zb); 5693 5694 if (*arc_flags & ARC_FLAG_WAIT) 5695 return (zio_wait(rzio)); 5696 5697 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5698 zio_nowait(rzio); 5699 } 5700 return (0); 5701} 5702 5703/* 5704 * Notify the arc that a block was freed, and thus will never be used again. 5705 */ 5706void 5707arc_freed(spa_t *spa, const blkptr_t *bp) 5708{ 5709 arc_buf_hdr_t *hdr; 5710 kmutex_t *hash_lock; 5711 uint64_t guid = spa_load_guid(spa); 5712 5713 ASSERT(!BP_IS_EMBEDDED(bp)); 5714 5715 hdr = buf_hash_find(guid, bp, &hash_lock); 5716 if (hdr == NULL) 5717 return; 5718 5719 /* 5720 * We might be trying to free a block that is still doing I/O 5721 * (i.e. prefetch) or has a reference (i.e. a dedup-ed, 5722 * dmu_sync-ed block). If this block is being prefetched, then it 5723 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr 5724 * until the I/O completes. A block may also have a reference if it is 5725 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would 5726 * have written the new block to its final resting place on disk but 5727 * without the dedup flag set. This would have left the hdr in the MRU 5728 * state and discoverable. When the txg finally syncs it detects that 5729 * the block was overridden in open context and issues an override I/O. 5730 * Since this is a dedup block, the override I/O will determine if the 5731 * block is already in the DDT. If so, then it will replace the io_bp 5732 * with the bp from the DDT and allow the I/O to finish. When the I/O 5733 * reaches the done callback, dbuf_write_override_done, it will 5734 * check to see if the io_bp and io_bp_override are identical. 5735 * If they are not, then it indicates that the bp was replaced with 5736 * the bp in the DDT and the override bp is freed. This allows 5737 * us to arrive here with a reference on a block that is being 5738 * freed. So if we have an I/O in progress, or a reference to 5739 * this hdr, then we don't destroy the hdr. 5740 */ 5741 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && 5742 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { 5743 arc_change_state(arc_anon, hdr, hash_lock); 5744 arc_hdr_destroy(hdr); 5745 mutex_exit(hash_lock); 5746 } else { 5747 mutex_exit(hash_lock); 5748 } 5749 5750} 5751 5752/* 5753 * Release this buffer from the cache, making it an anonymous buffer. This 5754 * must be done after a read and prior to modifying the buffer contents. 5755 * If the buffer has more than one reference, we must make 5756 * a new hdr for the buffer. 5757 */ 5758void 5759arc_release(arc_buf_t *buf, void *tag) 5760{ 5761 arc_buf_hdr_t *hdr = buf->b_hdr; 5762 5763 /* 5764 * It would be nice to assert that if it's DMU metadata (level > 5765 * 0 || it's the dnode file), then it must be syncing context. 5766 * But we don't know that information at this level. 5767 */ 5768 5769 mutex_enter(&buf->b_evict_lock); 5770 5771 ASSERT(HDR_HAS_L1HDR(hdr)); 5772 5773 /* 5774 * We don't grab the hash lock prior to this check, because if 5775 * the buffer's header is in the arc_anon state, it won't be 5776 * linked into the hash table. 5777 */ 5778 if (hdr->b_l1hdr.b_state == arc_anon) { 5779 mutex_exit(&buf->b_evict_lock); 5780 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5781 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 5782 ASSERT(!HDR_HAS_L2HDR(hdr)); 5783 ASSERT(HDR_EMPTY(hdr)); 5784 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5785 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 5786 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 5787 5788 hdr->b_l1hdr.b_arc_access = 0; 5789 5790 /* 5791 * If the buf is being overridden then it may already 5792 * have a hdr that is not empty. 5793 */ 5794 buf_discard_identity(hdr); 5795 arc_buf_thaw(buf); 5796 5797 return; 5798 } 5799 5800 kmutex_t *hash_lock = HDR_LOCK(hdr); 5801 mutex_enter(hash_lock); 5802 5803 /* 5804 * This assignment is only valid as long as the hash_lock is 5805 * held, we must be careful not to reference state or the 5806 * b_state field after dropping the lock. 5807 */ 5808 arc_state_t *state = hdr->b_l1hdr.b_state; 5809 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5810 ASSERT3P(state, !=, arc_anon); 5811 5812 /* this buffer is not on any list */ 5813 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0); 5814 5815 if (HDR_HAS_L2HDR(hdr)) { 5816 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5817 5818 /* 5819 * We have to recheck this conditional again now that 5820 * we're holding the l2ad_mtx to prevent a race with 5821 * another thread which might be concurrently calling 5822 * l2arc_evict(). In that case, l2arc_evict() might have 5823 * destroyed the header's L2 portion as we were waiting 5824 * to acquire the l2ad_mtx. 5825 */ 5826 if (HDR_HAS_L2HDR(hdr)) { 5827 l2arc_trim(hdr); 5828 arc_hdr_l2hdr_destroy(hdr); 5829 } 5830 5831 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5832 } 5833 5834 /* 5835 * Do we have more than one buf? 5836 */ 5837 if (hdr->b_l1hdr.b_bufcnt > 1) { 5838 arc_buf_hdr_t *nhdr; 5839 uint64_t spa = hdr->b_spa; 5840 uint64_t psize = HDR_GET_PSIZE(hdr); 5841 uint64_t lsize = HDR_GET_LSIZE(hdr); 5842 enum zio_compress compress = HDR_GET_COMPRESS(hdr); 5843 arc_buf_contents_t type = arc_buf_type(hdr); 5844 VERIFY3U(hdr->b_type, ==, type); 5845 5846 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 5847 (void) remove_reference(hdr, hash_lock, tag); 5848 5849 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) { 5850 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5851 ASSERT(ARC_BUF_LAST(buf)); 5852 } 5853 5854 /* 5855 * Pull the data off of this hdr and attach it to 5856 * a new anonymous hdr. Also find the last buffer 5857 * in the hdr's buffer list. 5858 */ 5859 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 5860 ASSERT3P(lastbuf, !=, NULL); 5861 5862 /* 5863 * If the current arc_buf_t and the hdr are sharing their data 5864 * buffer, then we must stop sharing that block. 5865 */ 5866 if (arc_buf_is_shared(buf)) { 5867 VERIFY(!arc_buf_is_shared(lastbuf)); 5868 5869 /* 5870 * First, sever the block sharing relationship between 5871 * buf and the arc_buf_hdr_t. 5872 */ 5873 arc_unshare_buf(hdr, buf); 5874 5875 /* 5876 * Now we need to recreate the hdr's b_pabd. Since we 5877 * have lastbuf handy, we try to share with it, but if 5878 * we can't then we allocate a new b_pabd and copy the 5879 * data from buf into it. 5880 */ 5881 if (arc_can_share(hdr, lastbuf)) { 5882 arc_share_buf(hdr, lastbuf); 5883 } else { 5884 arc_hdr_alloc_pabd(hdr); 5885 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, 5886 buf->b_data, psize); 5887 } 5888 VERIFY3P(lastbuf->b_data, !=, NULL); 5889 } else if (HDR_SHARED_DATA(hdr)) { 5890 /* 5891 * Uncompressed shared buffers are always at the end 5892 * of the list. Compressed buffers don't have the 5893 * same requirements. This makes it hard to 5894 * simply assert that the lastbuf is shared so 5895 * we rely on the hdr's compression flags to determine 5896 * if we have a compressed, shared buffer. 5897 */ 5898 ASSERT(arc_buf_is_shared(lastbuf) || 5899 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 5900 ASSERT(!ARC_BUF_SHARED(buf)); 5901 } 5902 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5903 ASSERT3P(state, !=, arc_l2c_only); 5904 5905 (void) refcount_remove_many(&state->arcs_size, 5906 arc_buf_size(buf), buf); 5907 5908 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 5909 ASSERT3P(state, !=, arc_l2c_only); 5910 (void) refcount_remove_many(&state->arcs_esize[type], 5911 arc_buf_size(buf), buf); 5912 } 5913 5914 hdr->b_l1hdr.b_bufcnt -= 1; 5915 arc_cksum_verify(buf); 5916#ifdef illumos 5917 arc_buf_unwatch(buf); 5918#endif 5919 5920 mutex_exit(hash_lock); 5921 5922 /* 5923 * Allocate a new hdr. The new hdr will contain a b_pabd 5924 * buffer which will be freed in arc_write(). 5925 */ 5926 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type); 5927 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); 5928 ASSERT0(nhdr->b_l1hdr.b_bufcnt); 5929 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt)); 5930 VERIFY3U(nhdr->b_type, ==, type); 5931 ASSERT(!HDR_SHARED_DATA(nhdr)); 5932 5933 nhdr->b_l1hdr.b_buf = buf; 5934 nhdr->b_l1hdr.b_bufcnt = 1; 5935 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 5936 buf->b_hdr = nhdr; 5937 5938 mutex_exit(&buf->b_evict_lock); 5939 (void) refcount_add_many(&arc_anon->arcs_size, 5940 arc_buf_size(buf), buf); 5941 } else { 5942 mutex_exit(&buf->b_evict_lock); 5943 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 5944 /* protected by hash lock, or hdr is on arc_anon */ 5945 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5946 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5947 arc_change_state(arc_anon, hdr, hash_lock); 5948 hdr->b_l1hdr.b_arc_access = 0; 5949 mutex_exit(hash_lock); 5950 5951 buf_discard_identity(hdr); 5952 arc_buf_thaw(buf); 5953 } 5954} 5955 5956int 5957arc_released(arc_buf_t *buf) 5958{ 5959 int released; 5960 5961 mutex_enter(&buf->b_evict_lock); 5962 released = (buf->b_data != NULL && 5963 buf->b_hdr->b_l1hdr.b_state == arc_anon); 5964 mutex_exit(&buf->b_evict_lock); 5965 return (released); 5966} 5967 5968#ifdef ZFS_DEBUG 5969int 5970arc_referenced(arc_buf_t *buf) 5971{ 5972 int referenced; 5973 5974 mutex_enter(&buf->b_evict_lock); 5975 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 5976 mutex_exit(&buf->b_evict_lock); 5977 return (referenced); 5978} 5979#endif 5980 5981static void 5982arc_write_ready(zio_t *zio) 5983{ 5984 arc_write_callback_t *callback = zio->io_private; 5985 arc_buf_t *buf = callback->awcb_buf; 5986 arc_buf_hdr_t *hdr = buf->b_hdr; 5987 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp); 5988 5989 ASSERT(HDR_HAS_L1HDR(hdr)); 5990 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 5991 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 5992 5993 /* 5994 * If we're reexecuting this zio because the pool suspended, then 5995 * cleanup any state that was previously set the first time the 5996 * callback was invoked. 5997 */ 5998 if (zio->io_flags & ZIO_FLAG_REEXECUTED) { 5999 arc_cksum_free(hdr); 6000#ifdef illumos 6001 arc_buf_unwatch(buf); 6002#endif 6003 if (hdr->b_l1hdr.b_pabd != NULL) { 6004 if (arc_buf_is_shared(buf)) { 6005 arc_unshare_buf(hdr, buf); 6006 } else { 6007 arc_hdr_free_pabd(hdr); 6008 } 6009 } 6010 } 6011 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 6012 ASSERT(!HDR_SHARED_DATA(hdr)); 6013 ASSERT(!arc_buf_is_shared(buf)); 6014 6015 callback->awcb_ready(zio, buf, callback->awcb_private); 6016 6017 if (HDR_IO_IN_PROGRESS(hdr)) 6018 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); 6019 6020 arc_cksum_compute(buf); 6021 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 6022 6023 enum zio_compress compress; 6024 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 6025 compress = ZIO_COMPRESS_OFF; 6026 } else { 6027 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp)); 6028 compress = BP_GET_COMPRESS(zio->io_bp); 6029 } 6030 HDR_SET_PSIZE(hdr, psize); 6031 arc_hdr_set_compress(hdr, compress); 6032 6033 6034 /* 6035 * Fill the hdr with data. If the hdr is compressed, the data we want 6036 * is available from the zio, otherwise we can take it from the buf. 6037 * 6038 * We might be able to share the buf's data with the hdr here. However, 6039 * doing so would cause the ARC to be full of linear ABDs if we write a 6040 * lot of shareable data. As a compromise, we check whether scattered 6041 * ABDs are allowed, and assume that if they are then the user wants 6042 * the ARC to be primarily filled with them regardless of the data being 6043 * written. Therefore, if they're allowed then we allocate one and copy 6044 * the data into it; otherwise, we share the data directly if we can. 6045 */ 6046 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) { 6047 arc_hdr_alloc_pabd(hdr); 6048 6049 /* 6050 * Ideally, we would always copy the io_abd into b_pabd, but the 6051 * user may have disabled compressed ARC, thus we must check the 6052 * hdr's compression setting rather than the io_bp's. 6053 */ 6054 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 6055 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=, 6056 ZIO_COMPRESS_OFF); 6057 ASSERT3U(psize, >, 0); 6058 6059 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize); 6060 } else { 6061 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr)); 6062 6063 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data, 6064 arc_buf_size(buf)); 6065 } 6066 } else { 6067 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd)); 6068 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf)); 6069 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 6070 6071 arc_share_buf(hdr, buf); 6072 } 6073 6074 arc_hdr_verify(hdr, zio->io_bp); 6075} 6076 6077static void 6078arc_write_children_ready(zio_t *zio) 6079{ 6080 arc_write_callback_t *callback = zio->io_private; 6081 arc_buf_t *buf = callback->awcb_buf; 6082 6083 callback->awcb_children_ready(zio, buf, callback->awcb_private); 6084} 6085 6086/* 6087 * The SPA calls this callback for each physical write that happens on behalf 6088 * of a logical write. See the comment in dbuf_write_physdone() for details. 6089 */ 6090static void 6091arc_write_physdone(zio_t *zio) 6092{ 6093 arc_write_callback_t *cb = zio->io_private; 6094 if (cb->awcb_physdone != NULL) 6095 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 6096} 6097 6098static void 6099arc_write_done(zio_t *zio) 6100{ 6101 arc_write_callback_t *callback = zio->io_private; 6102 arc_buf_t *buf = callback->awcb_buf; 6103 arc_buf_hdr_t *hdr = buf->b_hdr; 6104 6105 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 6106 6107 if (zio->io_error == 0) { 6108 arc_hdr_verify(hdr, zio->io_bp); 6109 6110 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 6111 buf_discard_identity(hdr); 6112 } else { 6113 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 6114 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 6115 } 6116 } else { 6117 ASSERT(HDR_EMPTY(hdr)); 6118 } 6119 6120 /* 6121 * If the block to be written was all-zero or compressed enough to be 6122 * embedded in the BP, no write was performed so there will be no 6123 * dva/birth/checksum. The buffer must therefore remain anonymous 6124 * (and uncached). 6125 */ 6126 if (!HDR_EMPTY(hdr)) { 6127 arc_buf_hdr_t *exists; 6128 kmutex_t *hash_lock; 6129 6130 ASSERT3U(zio->io_error, ==, 0); 6131 6132 arc_cksum_verify(buf); 6133 6134 exists = buf_hash_insert(hdr, &hash_lock); 6135 if (exists != NULL) { 6136 /* 6137 * This can only happen if we overwrite for 6138 * sync-to-convergence, because we remove 6139 * buffers from the hash table when we arc_free(). 6140 */ 6141 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 6142 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 6143 panic("bad overwrite, hdr=%p exists=%p", 6144 (void *)hdr, (void *)exists); 6145 ASSERT(refcount_is_zero( 6146 &exists->b_l1hdr.b_refcnt)); 6147 arc_change_state(arc_anon, exists, hash_lock); 6148 mutex_exit(hash_lock); 6149 arc_hdr_destroy(exists); 6150 exists = buf_hash_insert(hdr, &hash_lock); 6151 ASSERT3P(exists, ==, NULL); 6152 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 6153 /* nopwrite */ 6154 ASSERT(zio->io_prop.zp_nopwrite); 6155 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 6156 panic("bad nopwrite, hdr=%p exists=%p", 6157 (void *)hdr, (void *)exists); 6158 } else { 6159 /* Dedup */ 6160 ASSERT(hdr->b_l1hdr.b_bufcnt == 1); 6161 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 6162 ASSERT(BP_GET_DEDUP(zio->io_bp)); 6163 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 6164 } 6165 } 6166 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 6167 /* if it's not anon, we are doing a scrub */ 6168 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 6169 arc_access(hdr, hash_lock); 6170 mutex_exit(hash_lock); 6171 } else { 6172 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 6173 } 6174 6175 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 6176 callback->awcb_done(zio, buf, callback->awcb_private); 6177 6178 abd_put(zio->io_abd); 6179 kmem_free(callback, sizeof (arc_write_callback_t)); 6180} 6181 6182zio_t * 6183arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 6184 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready, 6185 arc_done_func_t *children_ready, arc_done_func_t *physdone, 6186 arc_done_func_t *done, void *private, zio_priority_t priority, 6187 int zio_flags, const zbookmark_phys_t *zb) 6188{ 6189 arc_buf_hdr_t *hdr = buf->b_hdr; 6190 arc_write_callback_t *callback; 6191 zio_t *zio; 6192 zio_prop_t localprop = *zp; 6193 6194 ASSERT3P(ready, !=, NULL); 6195 ASSERT3P(done, !=, NULL); 6196 ASSERT(!HDR_IO_ERROR(hdr)); 6197 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 6198 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 6199 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 6200 if (l2arc) 6201 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 6202 if (ARC_BUF_COMPRESSED(buf)) { 6203 /* 6204 * We're writing a pre-compressed buffer. Make the 6205 * compression algorithm requested by the zio_prop_t match 6206 * the pre-compressed buffer's compression algorithm. 6207 */ 6208 localprop.zp_compress = HDR_GET_COMPRESS(hdr); 6209 6210 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf)); 6211 zio_flags |= ZIO_FLAG_RAW; 6212 } 6213 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 6214 callback->awcb_ready = ready; 6215 callback->awcb_children_ready = children_ready; 6216 callback->awcb_physdone = physdone; 6217 callback->awcb_done = done; 6218 callback->awcb_private = private; 6219 callback->awcb_buf = buf; 6220 6221 /* 6222 * The hdr's b_pabd is now stale, free it now. A new data block 6223 * will be allocated when the zio pipeline calls arc_write_ready(). 6224 */ 6225 if (hdr->b_l1hdr.b_pabd != NULL) { 6226 /* 6227 * If the buf is currently sharing the data block with 6228 * the hdr then we need to break that relationship here. 6229 * The hdr will remain with a NULL data pointer and the 6230 * buf will take sole ownership of the block. 6231 */ 6232 if (arc_buf_is_shared(buf)) { 6233 arc_unshare_buf(hdr, buf); 6234 } else { 6235 arc_hdr_free_pabd(hdr); 6236 } 6237 VERIFY3P(buf->b_data, !=, NULL); 6238 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); 6239 } 6240 ASSERT(!arc_buf_is_shared(buf)); 6241 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 6242 6243 zio = zio_write(pio, spa, txg, bp, 6244 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)), 6245 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready, 6246 (children_ready != NULL) ? arc_write_children_ready : NULL, 6247 arc_write_physdone, arc_write_done, callback, 6248 priority, zio_flags, zb); 6249 6250 return (zio); 6251} 6252 6253static int 6254arc_memory_throttle(uint64_t reserve, uint64_t txg) 6255{ 6256#ifdef _KERNEL 6257 uint64_t available_memory = ptob(freemem); 6258 static uint64_t page_load = 0; 6259 static uint64_t last_txg = 0; 6260 6261#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 6262 available_memory = 6263 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 6264#endif 6265 6266 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 6267 return (0); 6268 6269 if (txg > last_txg) { 6270 last_txg = txg; 6271 page_load = 0; 6272 } 6273 /* 6274 * If we are in pageout, we know that memory is already tight, 6275 * the arc is already going to be evicting, so we just want to 6276 * continue to let page writes occur as quickly as possible. 6277 */ 6278 if (curproc == pageproc) { 6279 if (page_load > MAX(ptob(minfree), available_memory) / 4) 6280 return (SET_ERROR(ERESTART)); 6281 /* Note: reserve is inflated, so we deflate */ 6282 page_load += reserve / 8; 6283 return (0); 6284 } else if (page_load > 0 && arc_reclaim_needed()) { 6285 /* memory is low, delay before restarting */ 6286 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 6287 return (SET_ERROR(EAGAIN)); 6288 } 6289 page_load = 0; 6290#endif 6291 return (0); 6292} 6293 6294void 6295arc_tempreserve_clear(uint64_t reserve) 6296{ 6297 atomic_add_64(&arc_tempreserve, -reserve); 6298 ASSERT((int64_t)arc_tempreserve >= 0); 6299} 6300 6301int 6302arc_tempreserve_space(uint64_t reserve, uint64_t txg) 6303{ 6304 int error; 6305 uint64_t anon_size; 6306 6307 if (reserve > arc_c/4 && !arc_no_grow) { 6308 arc_c = MIN(arc_c_max, reserve * 4); 6309 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 6310 } 6311 if (reserve > arc_c) 6312 return (SET_ERROR(ENOMEM)); 6313 6314 /* 6315 * Don't count loaned bufs as in flight dirty data to prevent long 6316 * network delays from blocking transactions that are ready to be 6317 * assigned to a txg. 6318 */ 6319 6320 /* assert that it has not wrapped around */ 6321 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 6322 6323 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 6324 arc_loaned_bytes), 0); 6325 6326 /* 6327 * Writes will, almost always, require additional memory allocations 6328 * in order to compress/encrypt/etc the data. We therefore need to 6329 * make sure that there is sufficient available memory for this. 6330 */ 6331 error = arc_memory_throttle(reserve, txg); 6332 if (error != 0) 6333 return (error); 6334 6335 /* 6336 * Throttle writes when the amount of dirty data in the cache 6337 * gets too large. We try to keep the cache less than half full 6338 * of dirty blocks so that our sync times don't grow too large. 6339 * Note: if two requests come in concurrently, we might let them 6340 * both succeed, when one of them should fail. Not a huge deal. 6341 */ 6342 6343 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 6344 anon_size > arc_c / 4) { 6345 uint64_t meta_esize = 6346 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6347 uint64_t data_esize = 6348 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6349 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 6350 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 6351 arc_tempreserve >> 10, meta_esize >> 10, 6352 data_esize >> 10, reserve >> 10, arc_c >> 10); 6353 return (SET_ERROR(ERESTART)); 6354 } 6355 atomic_add_64(&arc_tempreserve, reserve); 6356 return (0); 6357} 6358 6359static void 6360arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 6361 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 6362{ 6363 size->value.ui64 = refcount_count(&state->arcs_size); 6364 evict_data->value.ui64 = 6365 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); 6366 evict_metadata->value.ui64 = 6367 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); 6368} 6369 6370static int 6371arc_kstat_update(kstat_t *ksp, int rw) 6372{ 6373 arc_stats_t *as = ksp->ks_data; 6374 6375 if (rw == KSTAT_WRITE) { 6376 return (EACCES); 6377 } else { 6378 arc_kstat_update_state(arc_anon, 6379 &as->arcstat_anon_size, 6380 &as->arcstat_anon_evictable_data, 6381 &as->arcstat_anon_evictable_metadata); 6382 arc_kstat_update_state(arc_mru, 6383 &as->arcstat_mru_size, 6384 &as->arcstat_mru_evictable_data, 6385 &as->arcstat_mru_evictable_metadata); 6386 arc_kstat_update_state(arc_mru_ghost, 6387 &as->arcstat_mru_ghost_size, 6388 &as->arcstat_mru_ghost_evictable_data, 6389 &as->arcstat_mru_ghost_evictable_metadata); 6390 arc_kstat_update_state(arc_mfu, 6391 &as->arcstat_mfu_size, 6392 &as->arcstat_mfu_evictable_data, 6393 &as->arcstat_mfu_evictable_metadata); 6394 arc_kstat_update_state(arc_mfu_ghost, 6395 &as->arcstat_mfu_ghost_size, 6396 &as->arcstat_mfu_ghost_evictable_data, 6397 &as->arcstat_mfu_ghost_evictable_metadata); 6398 6399 ARCSTAT(arcstat_size) = aggsum_value(&arc_size); 6400 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used); 6401 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size); 6402 ARCSTAT(arcstat_metadata_size) = 6403 aggsum_value(&astat_metadata_size); 6404 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size); 6405 ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size); 6406 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size); 6407 } 6408 6409 return (0); 6410} 6411 6412/* 6413 * This function *must* return indices evenly distributed between all 6414 * sublists of the multilist. This is needed due to how the ARC eviction 6415 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 6416 * distributed between all sublists and uses this assumption when 6417 * deciding which sublist to evict from and how much to evict from it. 6418 */ 6419unsigned int 6420arc_state_multilist_index_func(multilist_t *ml, void *obj) 6421{ 6422 arc_buf_hdr_t *hdr = obj; 6423 6424 /* 6425 * We rely on b_dva to generate evenly distributed index 6426 * numbers using buf_hash below. So, as an added precaution, 6427 * let's make sure we never add empty buffers to the arc lists. 6428 */ 6429 ASSERT(!HDR_EMPTY(hdr)); 6430 6431 /* 6432 * The assumption here, is the hash value for a given 6433 * arc_buf_hdr_t will remain constant throughout it's lifetime 6434 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 6435 * Thus, we don't need to store the header's sublist index 6436 * on insertion, as this index can be recalculated on removal. 6437 * 6438 * Also, the low order bits of the hash value are thought to be 6439 * distributed evenly. Otherwise, in the case that the multilist 6440 * has a power of two number of sublists, each sublists' usage 6441 * would not be evenly distributed. 6442 */ 6443 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 6444 multilist_get_num_sublists(ml)); 6445} 6446 6447#ifdef _KERNEL 6448static eventhandler_tag arc_event_lowmem = NULL; 6449 6450static void 6451arc_lowmem(void *arg __unused, int howto __unused) 6452{ 6453 6454 mutex_enter(&arc_reclaim_lock); 6455 DTRACE_PROBE1(arc__needfree, int64_t, ((int64_t)freemem - zfs_arc_free_target) * PAGESIZE); 6456 cv_signal(&arc_reclaim_thread_cv); 6457 6458 /* 6459 * It is unsafe to block here in arbitrary threads, because we can come 6460 * here from ARC itself and may hold ARC locks and thus risk a deadlock 6461 * with ARC reclaim thread. 6462 */ 6463 if (curproc == pageproc) 6464 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 6465 mutex_exit(&arc_reclaim_lock); 6466} 6467#endif 6468 6469static void 6470arc_state_init(void) 6471{ 6472 arc_anon = &ARC_anon; 6473 arc_mru = &ARC_mru; 6474 arc_mru_ghost = &ARC_mru_ghost; 6475 arc_mfu = &ARC_mfu; 6476 arc_mfu_ghost = &ARC_mfu_ghost; 6477 arc_l2c_only = &ARC_l2c_only; 6478 6479 arc_mru->arcs_list[ARC_BUFC_METADATA] = 6480 multilist_create(sizeof (arc_buf_hdr_t), 6481 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6482 arc_state_multilist_index_func); 6483 arc_mru->arcs_list[ARC_BUFC_DATA] = 6484 multilist_create(sizeof (arc_buf_hdr_t), 6485 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6486 arc_state_multilist_index_func); 6487 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] = 6488 multilist_create(sizeof (arc_buf_hdr_t), 6489 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6490 arc_state_multilist_index_func); 6491 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] = 6492 multilist_create(sizeof (arc_buf_hdr_t), 6493 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6494 arc_state_multilist_index_func); 6495 arc_mfu->arcs_list[ARC_BUFC_METADATA] = 6496 multilist_create(sizeof (arc_buf_hdr_t), 6497 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6498 arc_state_multilist_index_func); 6499 arc_mfu->arcs_list[ARC_BUFC_DATA] = 6500 multilist_create(sizeof (arc_buf_hdr_t), 6501 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6502 arc_state_multilist_index_func); 6503 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] = 6504 multilist_create(sizeof (arc_buf_hdr_t), 6505 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6506 arc_state_multilist_index_func); 6507 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] = 6508 multilist_create(sizeof (arc_buf_hdr_t), 6509 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6510 arc_state_multilist_index_func); 6511 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] = 6512 multilist_create(sizeof (arc_buf_hdr_t), 6513 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6514 arc_state_multilist_index_func); 6515 arc_l2c_only->arcs_list[ARC_BUFC_DATA] = 6516 multilist_create(sizeof (arc_buf_hdr_t), 6517 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6518 arc_state_multilist_index_func); 6519 6520 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6521 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6522 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6523 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6524 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6525 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6526 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6527 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6528 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6529 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6530 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6531 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6532 6533 refcount_create(&arc_anon->arcs_size); 6534 refcount_create(&arc_mru->arcs_size); 6535 refcount_create(&arc_mru_ghost->arcs_size); 6536 refcount_create(&arc_mfu->arcs_size); 6537 refcount_create(&arc_mfu_ghost->arcs_size); 6538 refcount_create(&arc_l2c_only->arcs_size); 6539 6540 aggsum_init(&arc_meta_used, 0); 6541 aggsum_init(&arc_size, 0); 6542 aggsum_init(&astat_data_size, 0); 6543 aggsum_init(&astat_metadata_size, 0); 6544 aggsum_init(&astat_hdr_size, 0); 6545 aggsum_init(&astat_other_size, 0); 6546 aggsum_init(&astat_l2_hdr_size, 0); 6547} 6548 6549static void 6550arc_state_fini(void) 6551{ 6552 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6553 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6554 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6555 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6556 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6557 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6558 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6559 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6560 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6561 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6562 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6563 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6564 6565 refcount_destroy(&arc_anon->arcs_size); 6566 refcount_destroy(&arc_mru->arcs_size); 6567 refcount_destroy(&arc_mru_ghost->arcs_size); 6568 refcount_destroy(&arc_mfu->arcs_size); 6569 refcount_destroy(&arc_mfu_ghost->arcs_size); 6570 refcount_destroy(&arc_l2c_only->arcs_size); 6571 6572 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]); 6573 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 6574 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]); 6575 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 6576 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]); 6577 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 6578 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]); 6579 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 6580} 6581 6582uint64_t 6583arc_max_bytes(void) 6584{ 6585 return (arc_c_max); 6586} 6587 6588void 6589arc_init(void) 6590{ 6591 int i, prefetch_tunable_set = 0; 6592 6593 /* 6594 * allmem is "all memory that we could possibly use". 6595 */ 6596#ifdef illumos 6597#ifdef _KERNEL 6598 uint64_t allmem = ptob(physmem - swapfs_minfree); 6599#else 6600 uint64_t allmem = (physmem * PAGESIZE) / 2; 6601#endif 6602#else 6603 uint64_t allmem = kmem_size(); 6604#endif 6605 6606 6607 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 6608 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 6609 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 6610 6611 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 6612 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL); 6613 6614 /* Convert seconds to clock ticks */ 6615 arc_min_prefetch_lifespan = 1 * hz; 6616 6617 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */ 6618 arc_c_min = MAX(allmem / 32, arc_abs_min); 6619 /* set max to 5/8 of all memory, or all but 1GB, whichever is more */ 6620 if (allmem >= 1 << 30) 6621 arc_c_max = allmem - (1 << 30); 6622 else 6623 arc_c_max = arc_c_min; 6624 arc_c_max = MAX(allmem * 5 / 8, arc_c_max); 6625 6626 /* 6627 * In userland, there's only the memory pressure that we artificially 6628 * create (see arc_available_memory()). Don't let arc_c get too 6629 * small, because it can cause transactions to be larger than 6630 * arc_c, causing arc_tempreserve_space() to fail. 6631 */ 6632#ifndef _KERNEL 6633 arc_c_min = arc_c_max / 2; 6634#endif 6635 6636#ifdef _KERNEL 6637 /* 6638 * Allow the tunables to override our calculations if they are 6639 * reasonable. 6640 */ 6641 if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) { 6642 arc_c_max = zfs_arc_max; 6643 arc_c_min = MIN(arc_c_min, arc_c_max); 6644 } 6645 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max) 6646 arc_c_min = zfs_arc_min; 6647#endif 6648 6649 arc_c = arc_c_max; 6650 arc_p = (arc_c >> 1); 6651 6652 /* limit meta-data to 1/4 of the arc capacity */ 6653 arc_meta_limit = arc_c_max / 4; 6654 6655#ifdef _KERNEL 6656 /* 6657 * Metadata is stored in the kernel's heap. Don't let us 6658 * use more than half the heap for the ARC. 6659 */ 6660 arc_meta_limit = MIN(arc_meta_limit, 6661 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2); 6662#endif 6663 6664 /* Allow the tunable to override if it is reasonable */ 6665 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 6666 arc_meta_limit = zfs_arc_meta_limit; 6667 6668 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 6669 arc_c_min = arc_meta_limit / 2; 6670 6671 if (zfs_arc_meta_min > 0) { 6672 arc_meta_min = zfs_arc_meta_min; 6673 } else { 6674 arc_meta_min = arc_c_min / 2; 6675 } 6676 6677 if (zfs_arc_grow_retry > 0) 6678 arc_grow_retry = zfs_arc_grow_retry; 6679 6680 if (zfs_arc_shrink_shift > 0) 6681 arc_shrink_shift = zfs_arc_shrink_shift; 6682 6683 if (zfs_arc_no_grow_shift > 0) 6684 arc_no_grow_shift = zfs_arc_no_grow_shift; 6685 /* 6686 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 6687 */ 6688 if (arc_no_grow_shift >= arc_shrink_shift) 6689 arc_no_grow_shift = arc_shrink_shift - 1; 6690 6691 if (zfs_arc_p_min_shift > 0) 6692 arc_p_min_shift = zfs_arc_p_min_shift; 6693 6694 /* if kmem_flags are set, lets try to use less memory */ 6695 if (kmem_debugging()) 6696 arc_c = arc_c / 2; 6697 if (arc_c < arc_c_min) 6698 arc_c = arc_c_min; 6699 6700 zfs_arc_min = arc_c_min; 6701 zfs_arc_max = arc_c_max; 6702 6703 arc_state_init(); 6704 buf_init(); 6705 6706 arc_reclaim_thread_exit = B_FALSE; 6707 arc_dnlc_evicts_thread_exit = FALSE; 6708 6709 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 6710 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 6711 6712 if (arc_ksp != NULL) { 6713 arc_ksp->ks_data = &arc_stats; 6714 arc_ksp->ks_update = arc_kstat_update; 6715 kstat_install(arc_ksp); 6716 } 6717 6718 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 6719 TS_RUN, minclsyspri); 6720 6721#ifdef _KERNEL 6722 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 6723 EVENTHANDLER_PRI_FIRST); 6724#endif 6725 6726 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0, 6727 TS_RUN, minclsyspri); 6728 6729 arc_dead = B_FALSE; 6730 arc_warm = B_FALSE; 6731 6732 /* 6733 * Calculate maximum amount of dirty data per pool. 6734 * 6735 * If it has been set by /etc/system, take that. 6736 * Otherwise, use a percentage of physical memory defined by 6737 * zfs_dirty_data_max_percent (default 10%) with a cap at 6738 * zfs_dirty_data_max_max (default 4GB). 6739 */ 6740 if (zfs_dirty_data_max == 0) { 6741 zfs_dirty_data_max = ptob(physmem) * 6742 zfs_dirty_data_max_percent / 100; 6743 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 6744 zfs_dirty_data_max_max); 6745 } 6746 6747#ifdef _KERNEL 6748 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 6749 prefetch_tunable_set = 1; 6750 6751#ifdef __i386__ 6752 if (prefetch_tunable_set == 0) { 6753 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 6754 "-- to enable,\n"); 6755 printf(" add \"vfs.zfs.prefetch_disable=0\" " 6756 "to /boot/loader.conf.\n"); 6757 zfs_prefetch_disable = 1; 6758 } 6759#else 6760 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 6761 prefetch_tunable_set == 0) { 6762 printf("ZFS NOTICE: Prefetch is disabled by default if less " 6763 "than 4GB of RAM is present;\n" 6764 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 6765 "to /boot/loader.conf.\n"); 6766 zfs_prefetch_disable = 1; 6767 } 6768#endif 6769 /* Warn about ZFS memory and address space requirements. */ 6770 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 6771 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 6772 "expect unstable behavior.\n"); 6773 } 6774 if (allmem < 512 * (1 << 20)) { 6775 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 6776 "expect unstable behavior.\n"); 6777 printf(" Consider tuning vm.kmem_size and " 6778 "vm.kmem_size_max\n"); 6779 printf(" in /boot/loader.conf.\n"); 6780 } 6781#endif 6782} 6783 6784void 6785arc_fini(void) 6786{ 6787#ifdef _KERNEL 6788 if (arc_event_lowmem != NULL) 6789 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 6790#endif 6791 6792 mutex_enter(&arc_reclaim_lock); 6793 arc_reclaim_thread_exit = B_TRUE; 6794 /* 6795 * The reclaim thread will set arc_reclaim_thread_exit back to 6796 * B_FALSE when it is finished exiting; we're waiting for that. 6797 */ 6798 while (arc_reclaim_thread_exit) { 6799 cv_signal(&arc_reclaim_thread_cv); 6800 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 6801 } 6802 mutex_exit(&arc_reclaim_lock); 6803 6804 /* Use B_TRUE to ensure *all* buffers are evicted */ 6805 arc_flush(NULL, B_TRUE); 6806 6807 mutex_enter(&arc_dnlc_evicts_lock); 6808 arc_dnlc_evicts_thread_exit = TRUE; 6809 /* 6810 * The user evicts thread will set arc_user_evicts_thread_exit 6811 * to FALSE when it is finished exiting; we're waiting for that. 6812 */ 6813 while (arc_dnlc_evicts_thread_exit) { 6814 cv_signal(&arc_dnlc_evicts_cv); 6815 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock); 6816 } 6817 mutex_exit(&arc_dnlc_evicts_lock); 6818 6819 arc_dead = B_TRUE; 6820 6821 if (arc_ksp != NULL) { 6822 kstat_delete(arc_ksp); 6823 arc_ksp = NULL; 6824 } 6825 6826 mutex_destroy(&arc_reclaim_lock); 6827 cv_destroy(&arc_reclaim_thread_cv); 6828 cv_destroy(&arc_reclaim_waiters_cv); 6829 6830 mutex_destroy(&arc_dnlc_evicts_lock); 6831 cv_destroy(&arc_dnlc_evicts_cv); 6832 6833 arc_state_fini(); 6834 buf_fini(); 6835 6836 ASSERT0(arc_loaned_bytes); 6837} 6838 6839/* 6840 * Level 2 ARC 6841 * 6842 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 6843 * It uses dedicated storage devices to hold cached data, which are populated 6844 * using large infrequent writes. The main role of this cache is to boost 6845 * the performance of random read workloads. The intended L2ARC devices 6846 * include short-stroked disks, solid state disks, and other media with 6847 * substantially faster read latency than disk. 6848 * 6849 * +-----------------------+ 6850 * | ARC | 6851 * +-----------------------+ 6852 * | ^ ^ 6853 * | | | 6854 * l2arc_feed_thread() arc_read() 6855 * | | | 6856 * | l2arc read | 6857 * V | | 6858 * +---------------+ | 6859 * | L2ARC | | 6860 * +---------------+ | 6861 * | ^ | 6862 * l2arc_write() | | 6863 * | | | 6864 * V | | 6865 * +-------+ +-------+ 6866 * | vdev | | vdev | 6867 * | cache | | cache | 6868 * +-------+ +-------+ 6869 * +=========+ .-----. 6870 * : L2ARC : |-_____-| 6871 * : devices : | Disks | 6872 * +=========+ `-_____-' 6873 * 6874 * Read requests are satisfied from the following sources, in order: 6875 * 6876 * 1) ARC 6877 * 2) vdev cache of L2ARC devices 6878 * 3) L2ARC devices 6879 * 4) vdev cache of disks 6880 * 5) disks 6881 * 6882 * Some L2ARC device types exhibit extremely slow write performance. 6883 * To accommodate for this there are some significant differences between 6884 * the L2ARC and traditional cache design: 6885 * 6886 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 6887 * the ARC behave as usual, freeing buffers and placing headers on ghost 6888 * lists. The ARC does not send buffers to the L2ARC during eviction as 6889 * this would add inflated write latencies for all ARC memory pressure. 6890 * 6891 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 6892 * It does this by periodically scanning buffers from the eviction-end of 6893 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 6894 * not already there. It scans until a headroom of buffers is satisfied, 6895 * which itself is a buffer for ARC eviction. If a compressible buffer is 6896 * found during scanning and selected for writing to an L2ARC device, we 6897 * temporarily boost scanning headroom during the next scan cycle to make 6898 * sure we adapt to compression effects (which might significantly reduce 6899 * the data volume we write to L2ARC). The thread that does this is 6900 * l2arc_feed_thread(), illustrated below; example sizes are included to 6901 * provide a better sense of ratio than this diagram: 6902 * 6903 * head --> tail 6904 * +---------------------+----------+ 6905 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 6906 * +---------------------+----------+ | o L2ARC eligible 6907 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 6908 * +---------------------+----------+ | 6909 * 15.9 Gbytes ^ 32 Mbytes | 6910 * headroom | 6911 * l2arc_feed_thread() 6912 * | 6913 * l2arc write hand <--[oooo]--' 6914 * | 8 Mbyte 6915 * | write max 6916 * V 6917 * +==============================+ 6918 * L2ARC dev |####|#|###|###| |####| ... | 6919 * +==============================+ 6920 * 32 Gbytes 6921 * 6922 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 6923 * evicted, then the L2ARC has cached a buffer much sooner than it probably 6924 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 6925 * safe to say that this is an uncommon case, since buffers at the end of 6926 * the ARC lists have moved there due to inactivity. 6927 * 6928 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 6929 * then the L2ARC simply misses copying some buffers. This serves as a 6930 * pressure valve to prevent heavy read workloads from both stalling the ARC 6931 * with waits and clogging the L2ARC with writes. This also helps prevent 6932 * the potential for the L2ARC to churn if it attempts to cache content too 6933 * quickly, such as during backups of the entire pool. 6934 * 6935 * 5. After system boot and before the ARC has filled main memory, there are 6936 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 6937 * lists can remain mostly static. Instead of searching from tail of these 6938 * lists as pictured, the l2arc_feed_thread() will search from the list heads 6939 * for eligible buffers, greatly increasing its chance of finding them. 6940 * 6941 * The L2ARC device write speed is also boosted during this time so that 6942 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 6943 * there are no L2ARC reads, and no fear of degrading read performance 6944 * through increased writes. 6945 * 6946 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 6947 * the vdev queue can aggregate them into larger and fewer writes. Each 6948 * device is written to in a rotor fashion, sweeping writes through 6949 * available space then repeating. 6950 * 6951 * 7. The L2ARC does not store dirty content. It never needs to flush 6952 * write buffers back to disk based storage. 6953 * 6954 * 8. If an ARC buffer is written (and dirtied) which also exists in the 6955 * L2ARC, the now stale L2ARC buffer is immediately dropped. 6956 * 6957 * The performance of the L2ARC can be tweaked by a number of tunables, which 6958 * may be necessary for different workloads: 6959 * 6960 * l2arc_write_max max write bytes per interval 6961 * l2arc_write_boost extra write bytes during device warmup 6962 * l2arc_noprefetch skip caching prefetched buffers 6963 * l2arc_headroom number of max device writes to precache 6964 * l2arc_headroom_boost when we find compressed buffers during ARC 6965 * scanning, we multiply headroom by this 6966 * percentage factor for the next scan cycle, 6967 * since more compressed buffers are likely to 6968 * be present 6969 * l2arc_feed_secs seconds between L2ARC writing 6970 * 6971 * Tunables may be removed or added as future performance improvements are 6972 * integrated, and also may become zpool properties. 6973 * 6974 * There are three key functions that control how the L2ARC warms up: 6975 * 6976 * l2arc_write_eligible() check if a buffer is eligible to cache 6977 * l2arc_write_size() calculate how much to write 6978 * l2arc_write_interval() calculate sleep delay between writes 6979 * 6980 * These three functions determine what to write, how much, and how quickly 6981 * to send writes. 6982 */ 6983 6984static boolean_t 6985l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 6986{ 6987 /* 6988 * A buffer is *not* eligible for the L2ARC if it: 6989 * 1. belongs to a different spa. 6990 * 2. is already cached on the L2ARC. 6991 * 3. has an I/O in progress (it may be an incomplete read). 6992 * 4. is flagged not eligible (zfs property). 6993 */ 6994 if (hdr->b_spa != spa_guid) { 6995 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 6996 return (B_FALSE); 6997 } 6998 if (HDR_HAS_L2HDR(hdr)) { 6999 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 7000 return (B_FALSE); 7001 } 7002 if (HDR_IO_IN_PROGRESS(hdr)) { 7003 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 7004 return (B_FALSE); 7005 } 7006 if (!HDR_L2CACHE(hdr)) { 7007 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 7008 return (B_FALSE); 7009 } 7010 7011 return (B_TRUE); 7012} 7013 7014static uint64_t 7015l2arc_write_size(void) 7016{ 7017 uint64_t size; 7018 7019 /* 7020 * Make sure our globals have meaningful values in case the user 7021 * altered them. 7022 */ 7023 size = l2arc_write_max; 7024 if (size == 0) { 7025 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 7026 "be greater than zero, resetting it to the default (%d)", 7027 L2ARC_WRITE_SIZE); 7028 size = l2arc_write_max = L2ARC_WRITE_SIZE; 7029 } 7030 7031 if (arc_warm == B_FALSE) 7032 size += l2arc_write_boost; 7033 7034 return (size); 7035 7036} 7037 7038static clock_t 7039l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 7040{ 7041 clock_t interval, next, now; 7042 7043 /* 7044 * If the ARC lists are busy, increase our write rate; if the 7045 * lists are stale, idle back. This is achieved by checking 7046 * how much we previously wrote - if it was more than half of 7047 * what we wanted, schedule the next write much sooner. 7048 */ 7049 if (l2arc_feed_again && wrote > (wanted / 2)) 7050 interval = (hz * l2arc_feed_min_ms) / 1000; 7051 else 7052 interval = hz * l2arc_feed_secs; 7053 7054 now = ddi_get_lbolt(); 7055 next = MAX(now, MIN(now + interval, began + interval)); 7056 7057 return (next); 7058} 7059 7060/* 7061 * Cycle through L2ARC devices. This is how L2ARC load balances. 7062 * If a device is returned, this also returns holding the spa config lock. 7063 */ 7064static l2arc_dev_t * 7065l2arc_dev_get_next(void) 7066{ 7067 l2arc_dev_t *first, *next = NULL; 7068 7069 /* 7070 * Lock out the removal of spas (spa_namespace_lock), then removal 7071 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 7072 * both locks will be dropped and a spa config lock held instead. 7073 */ 7074 mutex_enter(&spa_namespace_lock); 7075 mutex_enter(&l2arc_dev_mtx); 7076 7077 /* if there are no vdevs, there is nothing to do */ 7078 if (l2arc_ndev == 0) 7079 goto out; 7080 7081 first = NULL; 7082 next = l2arc_dev_last; 7083 do { 7084 /* loop around the list looking for a non-faulted vdev */ 7085 if (next == NULL) { 7086 next = list_head(l2arc_dev_list); 7087 } else { 7088 next = list_next(l2arc_dev_list, next); 7089 if (next == NULL) 7090 next = list_head(l2arc_dev_list); 7091 } 7092 7093 /* if we have come back to the start, bail out */ 7094 if (first == NULL) 7095 first = next; 7096 else if (next == first) 7097 break; 7098 7099 } while (vdev_is_dead(next->l2ad_vdev)); 7100 7101 /* if we were unable to find any usable vdevs, return NULL */ 7102 if (vdev_is_dead(next->l2ad_vdev)) 7103 next = NULL; 7104 7105 l2arc_dev_last = next; 7106 7107out: 7108 mutex_exit(&l2arc_dev_mtx); 7109 7110 /* 7111 * Grab the config lock to prevent the 'next' device from being 7112 * removed while we are writing to it. 7113 */ 7114 if (next != NULL) 7115 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 7116 mutex_exit(&spa_namespace_lock); 7117 7118 return (next); 7119} 7120 7121/* 7122 * Free buffers that were tagged for destruction. 7123 */ 7124static void 7125l2arc_do_free_on_write() 7126{ 7127 list_t *buflist; 7128 l2arc_data_free_t *df, *df_prev; 7129 7130 mutex_enter(&l2arc_free_on_write_mtx); 7131 buflist = l2arc_free_on_write; 7132 7133 for (df = list_tail(buflist); df; df = df_prev) { 7134 df_prev = list_prev(buflist, df); 7135 ASSERT3P(df->l2df_abd, !=, NULL); 7136 abd_free(df->l2df_abd); 7137 list_remove(buflist, df); 7138 kmem_free(df, sizeof (l2arc_data_free_t)); 7139 } 7140 7141 mutex_exit(&l2arc_free_on_write_mtx); 7142} 7143 7144/* 7145 * A write to a cache device has completed. Update all headers to allow 7146 * reads from these buffers to begin. 7147 */ 7148static void 7149l2arc_write_done(zio_t *zio) 7150{ 7151 l2arc_write_callback_t *cb; 7152 l2arc_dev_t *dev; 7153 list_t *buflist; 7154 arc_buf_hdr_t *head, *hdr, *hdr_prev; 7155 kmutex_t *hash_lock; 7156 int64_t bytes_dropped = 0; 7157 7158 cb = zio->io_private; 7159 ASSERT3P(cb, !=, NULL); 7160 dev = cb->l2wcb_dev; 7161 ASSERT3P(dev, !=, NULL); 7162 head = cb->l2wcb_head; 7163 ASSERT3P(head, !=, NULL); 7164 buflist = &dev->l2ad_buflist; 7165 ASSERT3P(buflist, !=, NULL); 7166 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 7167 l2arc_write_callback_t *, cb); 7168 7169 if (zio->io_error != 0) 7170 ARCSTAT_BUMP(arcstat_l2_writes_error); 7171 7172 /* 7173 * All writes completed, or an error was hit. 7174 */ 7175top: 7176 mutex_enter(&dev->l2ad_mtx); 7177 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 7178 hdr_prev = list_prev(buflist, hdr); 7179 7180 hash_lock = HDR_LOCK(hdr); 7181 7182 /* 7183 * We cannot use mutex_enter or else we can deadlock 7184 * with l2arc_write_buffers (due to swapping the order 7185 * the hash lock and l2ad_mtx are taken). 7186 */ 7187 if (!mutex_tryenter(hash_lock)) { 7188 /* 7189 * Missed the hash lock. We must retry so we 7190 * don't leave the ARC_FLAG_L2_WRITING bit set. 7191 */ 7192 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 7193 7194 /* 7195 * We don't want to rescan the headers we've 7196 * already marked as having been written out, so 7197 * we reinsert the head node so we can pick up 7198 * where we left off. 7199 */ 7200 list_remove(buflist, head); 7201 list_insert_after(buflist, hdr, head); 7202 7203 mutex_exit(&dev->l2ad_mtx); 7204 7205 /* 7206 * We wait for the hash lock to become available 7207 * to try and prevent busy waiting, and increase 7208 * the chance we'll be able to acquire the lock 7209 * the next time around. 7210 */ 7211 mutex_enter(hash_lock); 7212 mutex_exit(hash_lock); 7213 goto top; 7214 } 7215 7216 /* 7217 * We could not have been moved into the arc_l2c_only 7218 * state while in-flight due to our ARC_FLAG_L2_WRITING 7219 * bit being set. Let's just ensure that's being enforced. 7220 */ 7221 ASSERT(HDR_HAS_L1HDR(hdr)); 7222 7223 if (zio->io_error != 0) { 7224 /* 7225 * Error - drop L2ARC entry. 7226 */ 7227 list_remove(buflist, hdr); 7228 l2arc_trim(hdr); 7229 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 7230 7231 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr)); 7232 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); 7233 7234 bytes_dropped += arc_hdr_size(hdr); 7235 (void) refcount_remove_many(&dev->l2ad_alloc, 7236 arc_hdr_size(hdr), hdr); 7237 } 7238 7239 /* 7240 * Allow ARC to begin reads and ghost list evictions to 7241 * this L2ARC entry. 7242 */ 7243 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); 7244 7245 mutex_exit(hash_lock); 7246 } 7247 7248 atomic_inc_64(&l2arc_writes_done); 7249 list_remove(buflist, head); 7250 ASSERT(!HDR_HAS_L1HDR(head)); 7251 kmem_cache_free(hdr_l2only_cache, head); 7252 mutex_exit(&dev->l2ad_mtx); 7253 7254 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 7255 7256 l2arc_do_free_on_write(); 7257 7258 kmem_free(cb, sizeof (l2arc_write_callback_t)); 7259} 7260 7261/* 7262 * A read to a cache device completed. Validate buffer contents before 7263 * handing over to the regular ARC routines. 7264 */ 7265static void 7266l2arc_read_done(zio_t *zio) 7267{ 7268 l2arc_read_callback_t *cb; 7269 arc_buf_hdr_t *hdr; 7270 kmutex_t *hash_lock; 7271 boolean_t valid_cksum; 7272 7273 ASSERT3P(zio->io_vd, !=, NULL); 7274 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 7275 7276 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 7277 7278 cb = zio->io_private; 7279 ASSERT3P(cb, !=, NULL); 7280 hdr = cb->l2rcb_hdr; 7281 ASSERT3P(hdr, !=, NULL); 7282 7283 hash_lock = HDR_LOCK(hdr); 7284 mutex_enter(hash_lock); 7285 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 7286 7287 /* 7288 * If the data was read into a temporary buffer, 7289 * move it and free the buffer. 7290 */ 7291 if (cb->l2rcb_abd != NULL) { 7292 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size); 7293 if (zio->io_error == 0) { 7294 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd, 7295 arc_hdr_size(hdr)); 7296 } 7297 7298 /* 7299 * The following must be done regardless of whether 7300 * there was an error: 7301 * - free the temporary buffer 7302 * - point zio to the real ARC buffer 7303 * - set zio size accordingly 7304 * These are required because zio is either re-used for 7305 * an I/O of the block in the case of the error 7306 * or the zio is passed to arc_read_done() and it 7307 * needs real data. 7308 */ 7309 abd_free(cb->l2rcb_abd); 7310 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr); 7311 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd; 7312 } 7313 7314 ASSERT3P(zio->io_abd, !=, NULL); 7315 7316 /* 7317 * Check this survived the L2ARC journey. 7318 */ 7319 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd); 7320 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 7321 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 7322 7323 valid_cksum = arc_cksum_is_equal(hdr, zio); 7324 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 7325 mutex_exit(hash_lock); 7326 zio->io_private = hdr; 7327 arc_read_done(zio); 7328 } else { 7329 mutex_exit(hash_lock); 7330 /* 7331 * Buffer didn't survive caching. Increment stats and 7332 * reissue to the original storage device. 7333 */ 7334 if (zio->io_error != 0) { 7335 ARCSTAT_BUMP(arcstat_l2_io_error); 7336 } else { 7337 zio->io_error = SET_ERROR(EIO); 7338 } 7339 if (!valid_cksum) 7340 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 7341 7342 /* 7343 * If there's no waiter, issue an async i/o to the primary 7344 * storage now. If there *is* a waiter, the caller must 7345 * issue the i/o in a context where it's OK to block. 7346 */ 7347 if (zio->io_waiter == NULL) { 7348 zio_t *pio = zio_unique_parent(zio); 7349 7350 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 7351 7352 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, 7353 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done, 7354 hdr, zio->io_priority, cb->l2rcb_flags, 7355 &cb->l2rcb_zb)); 7356 } 7357 } 7358 7359 kmem_free(cb, sizeof (l2arc_read_callback_t)); 7360} 7361 7362/* 7363 * This is the list priority from which the L2ARC will search for pages to 7364 * cache. This is used within loops (0..3) to cycle through lists in the 7365 * desired order. This order can have a significant effect on cache 7366 * performance. 7367 * 7368 * Currently the metadata lists are hit first, MFU then MRU, followed by 7369 * the data lists. This function returns a locked list, and also returns 7370 * the lock pointer. 7371 */ 7372static multilist_sublist_t * 7373l2arc_sublist_lock(int list_num) 7374{ 7375 multilist_t *ml = NULL; 7376 unsigned int idx; 7377 7378 ASSERT(list_num >= 0 && list_num <= 3); 7379 7380 switch (list_num) { 7381 case 0: 7382 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA]; 7383 break; 7384 case 1: 7385 ml = arc_mru->arcs_list[ARC_BUFC_METADATA]; 7386 break; 7387 case 2: 7388 ml = arc_mfu->arcs_list[ARC_BUFC_DATA]; 7389 break; 7390 case 3: 7391 ml = arc_mru->arcs_list[ARC_BUFC_DATA]; 7392 break; 7393 } 7394 7395 /* 7396 * Return a randomly-selected sublist. This is acceptable 7397 * because the caller feeds only a little bit of data for each 7398 * call (8MB). Subsequent calls will result in different 7399 * sublists being selected. 7400 */ 7401 idx = multilist_get_random_index(ml); 7402 return (multilist_sublist_lock(ml, idx)); 7403} 7404 7405/* 7406 * Evict buffers from the device write hand to the distance specified in 7407 * bytes. This distance may span populated buffers, it may span nothing. 7408 * This is clearing a region on the L2ARC device ready for writing. 7409 * If the 'all' boolean is set, every buffer is evicted. 7410 */ 7411static void 7412l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 7413{ 7414 list_t *buflist; 7415 arc_buf_hdr_t *hdr, *hdr_prev; 7416 kmutex_t *hash_lock; 7417 uint64_t taddr; 7418 7419 buflist = &dev->l2ad_buflist; 7420 7421 if (!all && dev->l2ad_first) { 7422 /* 7423 * This is the first sweep through the device. There is 7424 * nothing to evict. 7425 */ 7426 return; 7427 } 7428 7429 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 7430 /* 7431 * When nearing the end of the device, evict to the end 7432 * before the device write hand jumps to the start. 7433 */ 7434 taddr = dev->l2ad_end; 7435 } else { 7436 taddr = dev->l2ad_hand + distance; 7437 } 7438 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 7439 uint64_t, taddr, boolean_t, all); 7440 7441top: 7442 mutex_enter(&dev->l2ad_mtx); 7443 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 7444 hdr_prev = list_prev(buflist, hdr); 7445 7446 hash_lock = HDR_LOCK(hdr); 7447 7448 /* 7449 * We cannot use mutex_enter or else we can deadlock 7450 * with l2arc_write_buffers (due to swapping the order 7451 * the hash lock and l2ad_mtx are taken). 7452 */ 7453 if (!mutex_tryenter(hash_lock)) { 7454 /* 7455 * Missed the hash lock. Retry. 7456 */ 7457 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 7458 mutex_exit(&dev->l2ad_mtx); 7459 mutex_enter(hash_lock); 7460 mutex_exit(hash_lock); 7461 goto top; 7462 } 7463 7464 /* 7465 * A header can't be on this list if it doesn't have L2 header. 7466 */ 7467 ASSERT(HDR_HAS_L2HDR(hdr)); 7468 7469 /* Ensure this header has finished being written. */ 7470 ASSERT(!HDR_L2_WRITING(hdr)); 7471 ASSERT(!HDR_L2_WRITE_HEAD(hdr)); 7472 7473 if (!all && (hdr->b_l2hdr.b_daddr >= taddr || 7474 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 7475 /* 7476 * We've evicted to the target address, 7477 * or the end of the device. 7478 */ 7479 mutex_exit(hash_lock); 7480 break; 7481 } 7482 7483 if (!HDR_HAS_L1HDR(hdr)) { 7484 ASSERT(!HDR_L2_READING(hdr)); 7485 /* 7486 * This doesn't exist in the ARC. Destroy. 7487 * arc_hdr_destroy() will call list_remove() 7488 * and decrement arcstat_l2_lsize. 7489 */ 7490 arc_change_state(arc_anon, hdr, hash_lock); 7491 arc_hdr_destroy(hdr); 7492 } else { 7493 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 7494 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 7495 /* 7496 * Invalidate issued or about to be issued 7497 * reads, since we may be about to write 7498 * over this location. 7499 */ 7500 if (HDR_L2_READING(hdr)) { 7501 ARCSTAT_BUMP(arcstat_l2_evict_reading); 7502 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); 7503 } 7504 7505 arc_hdr_l2hdr_destroy(hdr); 7506 } 7507 mutex_exit(hash_lock); 7508 } 7509 mutex_exit(&dev->l2ad_mtx); 7510} 7511 7512/* 7513 * Find and write ARC buffers to the L2ARC device. 7514 * 7515 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 7516 * for reading until they have completed writing. 7517 * The headroom_boost is an in-out parameter used to maintain headroom boost 7518 * state between calls to this function. 7519 * 7520 * Returns the number of bytes actually written (which may be smaller than 7521 * the delta by which the device hand has changed due to alignment). 7522 */ 7523static uint64_t 7524l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 7525{ 7526 arc_buf_hdr_t *hdr, *hdr_prev, *head; 7527 uint64_t write_asize, write_psize, write_lsize, headroom; 7528 boolean_t full; 7529 l2arc_write_callback_t *cb; 7530 zio_t *pio, *wzio; 7531 uint64_t guid = spa_load_guid(spa); 7532 int try; 7533 7534 ASSERT3P(dev->l2ad_vdev, !=, NULL); 7535 7536 pio = NULL; 7537 write_lsize = write_asize = write_psize = 0; 7538 full = B_FALSE; 7539 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 7540 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); 7541 7542 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 7543 /* 7544 * Copy buffers for L2ARC writing. 7545 */ 7546 for (try = 0; try <= 3; try++) { 7547 multilist_sublist_t *mls = l2arc_sublist_lock(try); 7548 uint64_t passed_sz = 0; 7549 7550 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 7551 7552 /* 7553 * L2ARC fast warmup. 7554 * 7555 * Until the ARC is warm and starts to evict, read from the 7556 * head of the ARC lists rather than the tail. 7557 */ 7558 if (arc_warm == B_FALSE) 7559 hdr = multilist_sublist_head(mls); 7560 else 7561 hdr = multilist_sublist_tail(mls); 7562 if (hdr == NULL) 7563 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 7564 7565 headroom = target_sz * l2arc_headroom; 7566 if (zfs_compressed_arc_enabled) 7567 headroom = (headroom * l2arc_headroom_boost) / 100; 7568 7569 for (; hdr; hdr = hdr_prev) { 7570 kmutex_t *hash_lock; 7571 7572 if (arc_warm == B_FALSE) 7573 hdr_prev = multilist_sublist_next(mls, hdr); 7574 else 7575 hdr_prev = multilist_sublist_prev(mls, hdr); 7576 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, 7577 HDR_GET_LSIZE(hdr)); 7578 7579 hash_lock = HDR_LOCK(hdr); 7580 if (!mutex_tryenter(hash_lock)) { 7581 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 7582 /* 7583 * Skip this buffer rather than waiting. 7584 */ 7585 continue; 7586 } 7587 7588 passed_sz += HDR_GET_LSIZE(hdr); 7589 if (passed_sz > headroom) { 7590 /* 7591 * Searched too far. 7592 */ 7593 mutex_exit(hash_lock); 7594 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 7595 break; 7596 } 7597 7598 if (!l2arc_write_eligible(guid, hdr)) { 7599 mutex_exit(hash_lock); 7600 continue; 7601 } 7602 7603 /* 7604 * We rely on the L1 portion of the header below, so 7605 * it's invalid for this header to have been evicted out 7606 * of the ghost cache, prior to being written out. The 7607 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 7608 */ 7609 ASSERT(HDR_HAS_L1HDR(hdr)); 7610 7611 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); 7612 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 7613 ASSERT3U(arc_hdr_size(hdr), >, 0); 7614 uint64_t psize = arc_hdr_size(hdr); 7615 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, 7616 psize); 7617 7618 if ((write_asize + asize) > target_sz) { 7619 full = B_TRUE; 7620 mutex_exit(hash_lock); 7621 ARCSTAT_BUMP(arcstat_l2_write_full); 7622 break; 7623 } 7624 7625 if (pio == NULL) { 7626 /* 7627 * Insert a dummy header on the buflist so 7628 * l2arc_write_done() can find where the 7629 * write buffers begin without searching. 7630 */ 7631 mutex_enter(&dev->l2ad_mtx); 7632 list_insert_head(&dev->l2ad_buflist, head); 7633 mutex_exit(&dev->l2ad_mtx); 7634 7635 cb = kmem_alloc( 7636 sizeof (l2arc_write_callback_t), KM_SLEEP); 7637 cb->l2wcb_dev = dev; 7638 cb->l2wcb_head = head; 7639 pio = zio_root(spa, l2arc_write_done, cb, 7640 ZIO_FLAG_CANFAIL); 7641 ARCSTAT_BUMP(arcstat_l2_write_pios); 7642 } 7643 7644 hdr->b_l2hdr.b_dev = dev; 7645 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 7646 arc_hdr_set_flags(hdr, 7647 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR); 7648 7649 mutex_enter(&dev->l2ad_mtx); 7650 list_insert_head(&dev->l2ad_buflist, hdr); 7651 mutex_exit(&dev->l2ad_mtx); 7652 7653 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr); 7654 7655 /* 7656 * Normally the L2ARC can use the hdr's data, but if 7657 * we're sharing data between the hdr and one of its 7658 * bufs, L2ARC needs its own copy of the data so that 7659 * the ZIO below can't race with the buf consumer. 7660 * Another case where we need to create a copy of the 7661 * data is when the buffer size is not device-aligned 7662 * and we need to pad the block to make it such. 7663 * That also keeps the clock hand suitably aligned. 7664 * 7665 * To ensure that the copy will be available for the 7666 * lifetime of the ZIO and be cleaned up afterwards, we 7667 * add it to the l2arc_free_on_write queue. 7668 */ 7669 abd_t *to_write; 7670 if (!HDR_SHARED_DATA(hdr) && psize == asize) { 7671 to_write = hdr->b_l1hdr.b_pabd; 7672 } else { 7673 to_write = abd_alloc_for_io(asize, 7674 HDR_ISTYPE_METADATA(hdr)); 7675 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize); 7676 if (asize != psize) { 7677 abd_zero_off(to_write, psize, 7678 asize - psize); 7679 } 7680 l2arc_free_abd_on_write(to_write, asize, 7681 arc_buf_type(hdr)); 7682 } 7683 wzio = zio_write_phys(pio, dev->l2ad_vdev, 7684 hdr->b_l2hdr.b_daddr, asize, to_write, 7685 ZIO_CHECKSUM_OFF, NULL, hdr, 7686 ZIO_PRIORITY_ASYNC_WRITE, 7687 ZIO_FLAG_CANFAIL, B_FALSE); 7688 7689 write_lsize += HDR_GET_LSIZE(hdr); 7690 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 7691 zio_t *, wzio); 7692 7693 write_psize += psize; 7694 write_asize += asize; 7695 dev->l2ad_hand += asize; 7696 7697 mutex_exit(hash_lock); 7698 7699 (void) zio_nowait(wzio); 7700 } 7701 7702 multilist_sublist_unlock(mls); 7703 7704 if (full == B_TRUE) 7705 break; 7706 } 7707 7708 /* No buffers selected for writing? */ 7709 if (pio == NULL) { 7710 ASSERT0(write_lsize); 7711 ASSERT(!HDR_HAS_L1HDR(head)); 7712 kmem_cache_free(hdr_l2only_cache, head); 7713 return (0); 7714 } 7715 7716 ASSERT3U(write_psize, <=, target_sz); 7717 ARCSTAT_BUMP(arcstat_l2_writes_sent); 7718 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize); 7719 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize); 7720 ARCSTAT_INCR(arcstat_l2_psize, write_psize); 7721 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 7722 7723 /* 7724 * Bump device hand to the device start if it is approaching the end. 7725 * l2arc_evict() will already have evicted ahead for this case. 7726 */ 7727 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 7728 dev->l2ad_hand = dev->l2ad_start; 7729 dev->l2ad_first = B_FALSE; 7730 } 7731 7732 dev->l2ad_writing = B_TRUE; 7733 (void) zio_wait(pio); 7734 dev->l2ad_writing = B_FALSE; 7735 7736 return (write_asize); 7737} 7738 7739/* 7740 * This thread feeds the L2ARC at regular intervals. This is the beating 7741 * heart of the L2ARC. 7742 */ 7743/* ARGSUSED */ 7744static void 7745l2arc_feed_thread(void *unused __unused) 7746{ 7747 callb_cpr_t cpr; 7748 l2arc_dev_t *dev; 7749 spa_t *spa; 7750 uint64_t size, wrote; 7751 clock_t begin, next = ddi_get_lbolt(); 7752 7753 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 7754 7755 mutex_enter(&l2arc_feed_thr_lock); 7756 7757 while (l2arc_thread_exit == 0) { 7758 CALLB_CPR_SAFE_BEGIN(&cpr); 7759 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 7760 next - ddi_get_lbolt()); 7761 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 7762 next = ddi_get_lbolt() + hz; 7763 7764 /* 7765 * Quick check for L2ARC devices. 7766 */ 7767 mutex_enter(&l2arc_dev_mtx); 7768 if (l2arc_ndev == 0) { 7769 mutex_exit(&l2arc_dev_mtx); 7770 continue; 7771 } 7772 mutex_exit(&l2arc_dev_mtx); 7773 begin = ddi_get_lbolt(); 7774 7775 /* 7776 * This selects the next l2arc device to write to, and in 7777 * doing so the next spa to feed from: dev->l2ad_spa. This 7778 * will return NULL if there are now no l2arc devices or if 7779 * they are all faulted. 7780 * 7781 * If a device is returned, its spa's config lock is also 7782 * held to prevent device removal. l2arc_dev_get_next() 7783 * will grab and release l2arc_dev_mtx. 7784 */ 7785 if ((dev = l2arc_dev_get_next()) == NULL) 7786 continue; 7787 7788 spa = dev->l2ad_spa; 7789 ASSERT3P(spa, !=, NULL); 7790 7791 /* 7792 * If the pool is read-only then force the feed thread to 7793 * sleep a little longer. 7794 */ 7795 if (!spa_writeable(spa)) { 7796 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 7797 spa_config_exit(spa, SCL_L2ARC, dev); 7798 continue; 7799 } 7800 7801 /* 7802 * Avoid contributing to memory pressure. 7803 */ 7804 if (arc_reclaim_needed()) { 7805 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 7806 spa_config_exit(spa, SCL_L2ARC, dev); 7807 continue; 7808 } 7809 7810 ARCSTAT_BUMP(arcstat_l2_feeds); 7811 7812 size = l2arc_write_size(); 7813 7814 /* 7815 * Evict L2ARC buffers that will be overwritten. 7816 */ 7817 l2arc_evict(dev, size, B_FALSE); 7818 7819 /* 7820 * Write ARC buffers. 7821 */ 7822 wrote = l2arc_write_buffers(spa, dev, size); 7823 7824 /* 7825 * Calculate interval between writes. 7826 */ 7827 next = l2arc_write_interval(begin, size, wrote); 7828 spa_config_exit(spa, SCL_L2ARC, dev); 7829 } 7830 7831 l2arc_thread_exit = 0; 7832 cv_broadcast(&l2arc_feed_thr_cv); 7833 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 7834 thread_exit(); 7835} 7836 7837boolean_t 7838l2arc_vdev_present(vdev_t *vd) 7839{ 7840 l2arc_dev_t *dev; 7841 7842 mutex_enter(&l2arc_dev_mtx); 7843 for (dev = list_head(l2arc_dev_list); dev != NULL; 7844 dev = list_next(l2arc_dev_list, dev)) { 7845 if (dev->l2ad_vdev == vd) 7846 break; 7847 } 7848 mutex_exit(&l2arc_dev_mtx); 7849 7850 return (dev != NULL); 7851} 7852 7853/* 7854 * Add a vdev for use by the L2ARC. By this point the spa has already 7855 * validated the vdev and opened it. 7856 */ 7857void 7858l2arc_add_vdev(spa_t *spa, vdev_t *vd) 7859{ 7860 l2arc_dev_t *adddev; 7861 7862 ASSERT(!l2arc_vdev_present(vd)); 7863 7864 vdev_ashift_optimize(vd); 7865 7866 /* 7867 * Create a new l2arc device entry. 7868 */ 7869 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 7870 adddev->l2ad_spa = spa; 7871 adddev->l2ad_vdev = vd; 7872 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 7873 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 7874 adddev->l2ad_hand = adddev->l2ad_start; 7875 adddev->l2ad_first = B_TRUE; 7876 adddev->l2ad_writing = B_FALSE; 7877 7878 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 7879 /* 7880 * This is a list of all ARC buffers that are still valid on the 7881 * device. 7882 */ 7883 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 7884 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 7885 7886 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 7887 refcount_create(&adddev->l2ad_alloc); 7888 7889 /* 7890 * Add device to global list 7891 */ 7892 mutex_enter(&l2arc_dev_mtx); 7893 list_insert_head(l2arc_dev_list, adddev); 7894 atomic_inc_64(&l2arc_ndev); 7895 mutex_exit(&l2arc_dev_mtx); 7896} 7897 7898/* 7899 * Remove a vdev from the L2ARC. 7900 */ 7901void 7902l2arc_remove_vdev(vdev_t *vd) 7903{ 7904 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 7905 7906 /* 7907 * Find the device by vdev 7908 */ 7909 mutex_enter(&l2arc_dev_mtx); 7910 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 7911 nextdev = list_next(l2arc_dev_list, dev); 7912 if (vd == dev->l2ad_vdev) { 7913 remdev = dev; 7914 break; 7915 } 7916 } 7917 ASSERT3P(remdev, !=, NULL); 7918 7919 /* 7920 * Remove device from global list 7921 */ 7922 list_remove(l2arc_dev_list, remdev); 7923 l2arc_dev_last = NULL; /* may have been invalidated */ 7924 atomic_dec_64(&l2arc_ndev); 7925 mutex_exit(&l2arc_dev_mtx); 7926 7927 /* 7928 * Clear all buflists and ARC references. L2ARC device flush. 7929 */ 7930 l2arc_evict(remdev, 0, B_TRUE); 7931 list_destroy(&remdev->l2ad_buflist); 7932 mutex_destroy(&remdev->l2ad_mtx); 7933 refcount_destroy(&remdev->l2ad_alloc); 7934 kmem_free(remdev, sizeof (l2arc_dev_t)); 7935} 7936 7937void 7938l2arc_init(void) 7939{ 7940 l2arc_thread_exit = 0; 7941 l2arc_ndev = 0; 7942 l2arc_writes_sent = 0; 7943 l2arc_writes_done = 0; 7944 7945 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7946 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7947 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7948 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7949 7950 l2arc_dev_list = &L2ARC_dev_list; 7951 l2arc_free_on_write = &L2ARC_free_on_write; 7952 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7953 offsetof(l2arc_dev_t, l2ad_node)); 7954 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7955 offsetof(l2arc_data_free_t, l2df_list_node)); 7956} 7957 7958void 7959l2arc_fini(void) 7960{ 7961 /* 7962 * This is called from dmu_fini(), which is called from spa_fini(); 7963 * Because of this, we can assume that all l2arc devices have 7964 * already been removed when the pools themselves were removed. 7965 */ 7966 7967 l2arc_do_free_on_write(); 7968 7969 mutex_destroy(&l2arc_feed_thr_lock); 7970 cv_destroy(&l2arc_feed_thr_cv); 7971 mutex_destroy(&l2arc_dev_mtx); 7972 mutex_destroy(&l2arc_free_on_write_mtx); 7973 7974 list_destroy(l2arc_dev_list); 7975 list_destroy(l2arc_free_on_write); 7976} 7977 7978void 7979l2arc_start(void) 7980{ 7981 if (!(spa_mode_global & FWRITE)) 7982 return; 7983 7984 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7985 TS_RUN, minclsyspri); 7986} 7987 7988void 7989l2arc_stop(void) 7990{ 7991 if (!(spa_mode_global & FWRITE)) 7992 return; 7993 7994 mutex_enter(&l2arc_feed_thr_lock); 7995 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7996 l2arc_thread_exit = 1; 7997 while (l2arc_thread_exit != 0) 7998 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7999 mutex_exit(&l2arc_feed_thr_lock); 8000} 8001