1fa9e4066Sahrens /* 2fa9e4066Sahrens * CDDL HEADER START 3fa9e4066Sahrens * 4fa9e4066Sahrens * The contents of this file are subject to the terms of the 5033f9833Sek * Common Development and Distribution License (the "License"). 6033f9833Sek * You may not use this file except in compliance with the License. 7fa9e4066Sahrens * 8fa9e4066Sahrens * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9fa9e4066Sahrens * or http://www.opensolaris.org/os/licensing. 10fa9e4066Sahrens * See the License for the specific language governing permissions 11fa9e4066Sahrens * and limitations under the License. 12fa9e4066Sahrens * 13fa9e4066Sahrens * When distributing Covered Code, include this CDDL HEADER in each 14fa9e4066Sahrens * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15fa9e4066Sahrens * If applicable, add the following below this CDDL HEADER, with the 16fa9e4066Sahrens * fields enclosed by brackets "[]" replaced with your own identifying 17fa9e4066Sahrens * information: Portions Copyright [yyyy] [name of copyright owner] 18fa9e4066Sahrens * 19fa9e4066Sahrens * CDDL HEADER END 20fa9e4066Sahrens */ 21fa9e4066Sahrens /* 223f9d6ad7SLin Ling * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 2336a64e62STim Kordas * Copyright (c) 2018, Joyent, Inc. 24*fa98e487SMatthew Ahrens * Copyright (c) 2011, 2018 by Delphix. All rights reserved. 2571cb1b74SSaso Kiselkov * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 2601a059eeSRoman Strashkin * Copyright 2017 Nexenta Systems, Inc. All rights reserved. 27fa9e4066Sahrens */ 28fa9e4066Sahrens 29fa9e4066Sahrens /* 3044cb6abcSbmc * DVA-based Adjustable Replacement Cache 31fa9e4066Sahrens * 32ea8dc4b6Seschrock * While much of the theory of operation used here is 33ea8dc4b6Seschrock * based on the self-tuning, low overhead replacement cache 34fa9e4066Sahrens * presented by Megiddo and Modha at FAST 2003, there are some 35fa9e4066Sahrens * significant differences: 36fa9e4066Sahrens * 37fa9e4066Sahrens * 1. The Megiddo and Modha model assumes any page is evictable. 38fa9e4066Sahrens * Pages in its cache cannot be "locked" into memory. This makes 39fa9e4066Sahrens * the eviction algorithm simple: evict the last page in the list. 40fa9e4066Sahrens * This also make the performance characteristics easy to reason 41fa9e4066Sahrens * about. Our cache is not so simple. At any given moment, some 42fa9e4066Sahrens * subset of the blocks in the cache are un-evictable because we 43fa9e4066Sahrens * have handed out a reference to them. Blocks are only evictable 44fa9e4066Sahrens * when there are no external references active. This makes 45fa9e4066Sahrens * eviction far more problematic: we choose to evict the evictable 46fa9e4066Sahrens * blocks that are the "lowest" in the list. 47fa9e4066Sahrens * 48fa9e4066Sahrens * There are times when it is not possible to evict the requested 49fa9e4066Sahrens * space. In these circumstances we are unable to adjust the cache 50fa9e4066Sahrens * size. To prevent the cache growing unbounded at these times we 51fa94a07fSbrendan * implement a "cache throttle" that slows the flow of new data 52fa94a07fSbrendan * into the cache until we can make space available. 53fa9e4066Sahrens * 54fa9e4066Sahrens * 2. The Megiddo and Modha model assumes a fixed cache size. 55fa9e4066Sahrens * Pages are evicted when the cache is full and there is a cache 56fa9e4066Sahrens * miss. Our model has a variable sized cache. It grows with 57fa94a07fSbrendan * high use, but also tries to react to memory pressure from the 58fa9e4066Sahrens * operating system: decreasing its size when system memory is 59fa9e4066Sahrens * tight. 60fa9e4066Sahrens * 61fa9e4066Sahrens * 3. The Megiddo and Modha model assumes a fixed page size. All 62f7170741SWill Andrews * elements of the cache are therefore exactly the same size. So 63fa9e4066Sahrens * when adjusting the cache size following a cache miss, its simply 64fa9e4066Sahrens * a matter of choosing a single page to evict. In our model, we 65fa9e4066Sahrens * have variable sized cache blocks (rangeing from 512 bytes to 66f7170741SWill Andrews * 128K bytes). We therefore choose a set of blocks to evict to make 67fa9e4066Sahrens * space for a cache miss that approximates as closely as possible 68fa9e4066Sahrens * the space used by the new block. 69fa9e4066Sahrens * 70fa9e4066Sahrens * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 71fa9e4066Sahrens * by N. Megiddo & D. Modha, FAST 2003 72fa9e4066Sahrens */ 73fa9e4066Sahrens 74fa9e4066Sahrens /* 75fa9e4066Sahrens * The locking model: 76fa9e4066Sahrens * 77fa9e4066Sahrens * A new reference to a cache buffer can be obtained in two 78fa9e4066Sahrens * ways: 1) via a hash table lookup using the DVA as a key, 79fa94a07fSbrendan * or 2) via one of the ARC lists. The arc_read() interface 805602294fSDan Kimmel * uses method 1, while the internal ARC algorithms for 81f7170741SWill Andrews * adjusting the cache use method 2. We therefore provide two 82fa9e4066Sahrens * types of locks: 1) the hash table lock array, and 2) the 835602294fSDan Kimmel * ARC list locks. 84fa9e4066Sahrens * 85fc98fea5SBart Coddens * Buffers do not have their own mutexes, rather they rely on the 86fc98fea5SBart Coddens * hash table mutexes for the bulk of their protection (i.e. most 87fc98fea5SBart Coddens * fields in the arc_buf_hdr_t are protected by these mutexes). 88fa9e4066Sahrens * 89fa9e4066Sahrens * buf_hash_find() returns the appropriate mutex (held) when it 90fa9e4066Sahrens * locates the requested buffer in the hash table. It returns 91fa9e4066Sahrens * NULL for the mutex if the buffer was not in the table. 92fa9e4066Sahrens * 93fa9e4066Sahrens * buf_hash_remove() expects the appropriate hash mutex to be 94fa9e4066Sahrens * already held before it is invoked. 95fa9e4066Sahrens * 965602294fSDan Kimmel * Each ARC state also has a mutex which is used to protect the 97fa9e4066Sahrens * buffer list associated with the state. When attempting to 985602294fSDan Kimmel * obtain a hash table lock while holding an ARC list lock you 99fa9e4066Sahrens * must use: mutex_tryenter() to avoid deadlock. Also note that 10044eda4d7Smaybee * the active state mutex must be held before the ghost state mutex. 101fa9e4066Sahrens * 102fa9e4066Sahrens * Note that the majority of the performance stats are manipulated 103fa9e4066Sahrens * with atomic operations. 104fa94a07fSbrendan * 10589c86e32SChris Williamson * The L2ARC uses the l2ad_mtx on each vdev for the following: 106fa94a07fSbrendan * 107fa94a07fSbrendan * - L2ARC buflist creation 108fa94a07fSbrendan * - L2ARC buflist eviction 109fa94a07fSbrendan * - L2ARC write completion, which walks L2ARC buflists 110fa94a07fSbrendan * - ARC header destruction, as it removes from L2ARC buflists 111fa94a07fSbrendan * - ARC header release, as it removes from L2ARC buflists 112fa9e4066Sahrens */ 113fa9e4066Sahrens 114dcbf3bd6SGeorge Wilson /* 115dcbf3bd6SGeorge Wilson * ARC operation: 116dcbf3bd6SGeorge Wilson * 117dcbf3bd6SGeorge Wilson * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure. 118dcbf3bd6SGeorge Wilson * This structure can point either to a block that is still in the cache or to 119dcbf3bd6SGeorge Wilson * one that is only accessible in an L2 ARC device, or it can provide 120dcbf3bd6SGeorge Wilson * information about a block that was recently evicted. If a block is 121dcbf3bd6SGeorge Wilson * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough 122dcbf3bd6SGeorge Wilson * information to retrieve it from the L2ARC device. This information is 123dcbf3bd6SGeorge Wilson * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block 124dcbf3bd6SGeorge Wilson * that is in this state cannot access the data directly. 125dcbf3bd6SGeorge Wilson * 126dcbf3bd6SGeorge Wilson * Blocks that are actively being referenced or have not been evicted 127dcbf3bd6SGeorge Wilson * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within 128dcbf3bd6SGeorge Wilson * the arc_buf_hdr_t that will point to the data block in memory. A block can 129dcbf3bd6SGeorge Wilson * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC 1305602294fSDan Kimmel * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and 131770499e1SDan Kimmel * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd). 1325602294fSDan Kimmel * 1335602294fSDan Kimmel * The L1ARC's data pointer may or may not be uncompressed. The ARC has the 134770499e1SDan Kimmel * ability to store the physical data (b_pabd) associated with the DVA of the 135770499e1SDan Kimmel * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block, 1365602294fSDan Kimmel * it will match its on-disk compression characteristics. This behavior can be 1375602294fSDan Kimmel * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the 138770499e1SDan Kimmel * compressed ARC functionality is disabled, the b_pabd will point to an 1395602294fSDan Kimmel * uncompressed version of the on-disk data. 1405602294fSDan Kimmel * 1415602294fSDan Kimmel * Data in the L1ARC is not accessed by consumers of the ARC directly. Each 1425602294fSDan Kimmel * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it. 1435602294fSDan Kimmel * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC 1445602294fSDan Kimmel * consumer. The ARC will provide references to this data and will keep it 1455602294fSDan Kimmel * cached until it is no longer in use. The ARC caches only the L1ARC's physical 1465602294fSDan Kimmel * data block and will evict any arc_buf_t that is no longer referenced. The 1475602294fSDan Kimmel * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the 148dcbf3bd6SGeorge Wilson * "overhead_size" kstat. 149dcbf3bd6SGeorge Wilson * 1505602294fSDan Kimmel * Depending on the consumer, an arc_buf_t can be requested in uncompressed or 1515602294fSDan Kimmel * compressed form. The typical case is that consumers will want uncompressed 1525602294fSDan Kimmel * data, and when that happens a new data buffer is allocated where the data is 1535602294fSDan Kimmel * decompressed for them to use. Currently the only consumer who wants 1545602294fSDan Kimmel * compressed arc_buf_t's is "zfs send", when it streams data exactly as it 1555602294fSDan Kimmel * exists on disk. When this happens, the arc_buf_t's data buffer is shared 1565602294fSDan Kimmel * with the arc_buf_hdr_t. 157dcbf3bd6SGeorge Wilson * 1585602294fSDan Kimmel * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The 1595602294fSDan Kimmel * first one is owned by a compressed send consumer (and therefore references 1605602294fSDan Kimmel * the same compressed data buffer as the arc_buf_hdr_t) and the second could be 1615602294fSDan Kimmel * used by any other consumer (and has its own uncompressed copy of the data 1625602294fSDan Kimmel * buffer). 163dcbf3bd6SGeorge Wilson * 1645602294fSDan Kimmel * arc_buf_hdr_t 1655602294fSDan Kimmel * +-----------+ 1665602294fSDan Kimmel * | fields | 1675602294fSDan Kimmel * | common to | 1685602294fSDan Kimmel * | L1- and | 1695602294fSDan Kimmel * | L2ARC | 1705602294fSDan Kimmel * +-----------+ 1715602294fSDan Kimmel * | l2arc_buf_hdr_t 1725602294fSDan Kimmel * | | 1735602294fSDan Kimmel * +-----------+ 1745602294fSDan Kimmel * | l1arc_buf_hdr_t 1755602294fSDan Kimmel * | | arc_buf_t 1765602294fSDan Kimmel * | b_buf +------------>+-----------+ arc_buf_t 177770499e1SDan Kimmel * | b_pabd +-+ |b_next +---->+-----------+ 1785602294fSDan Kimmel * +-----------+ | |-----------| |b_next +-->NULL 1795602294fSDan Kimmel * | |b_comp = T | +-----------+ 1805602294fSDan Kimmel * | |b_data +-+ |b_comp = F | 1815602294fSDan Kimmel * | +-----------+ | |b_data +-+ 1825602294fSDan Kimmel * +->+------+ | +-----------+ | 1835602294fSDan Kimmel * compressed | | | | 1845602294fSDan Kimmel * data | |<--------------+ | uncompressed 1855602294fSDan Kimmel * +------+ compressed, | data 1865602294fSDan Kimmel * shared +-->+------+ 1875602294fSDan Kimmel * data | | 1885602294fSDan Kimmel * | | 1895602294fSDan Kimmel * +------+ 190dcbf3bd6SGeorge Wilson * 191dcbf3bd6SGeorge Wilson * When a consumer reads a block, the ARC must first look to see if the 1925602294fSDan Kimmel * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new 1935602294fSDan Kimmel * arc_buf_t and either copies uncompressed data into a new data buffer from an 194770499e1SDan Kimmel * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a 195770499e1SDan Kimmel * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the 1965602294fSDan Kimmel * hdr is compressed and the desired compression characteristics of the 1975602294fSDan Kimmel * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the 1985602294fSDan Kimmel * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be 1995602294fSDan Kimmel * the last buffer in the hdr's b_buf list, however a shared compressed buf can 2005602294fSDan Kimmel * be anywhere in the hdr's list. 201dcbf3bd6SGeorge Wilson * 202dcbf3bd6SGeorge Wilson * The diagram below shows an example of an uncompressed ARC hdr that is 2035602294fSDan Kimmel * sharing its data with an arc_buf_t (note that the shared uncompressed buf is 2045602294fSDan Kimmel * the last element in the buf list): 205dcbf3bd6SGeorge Wilson * 206dcbf3bd6SGeorge Wilson * arc_buf_hdr_t 207dcbf3bd6SGeorge Wilson * +-----------+ 208dcbf3bd6SGeorge Wilson * | | 209dcbf3bd6SGeorge Wilson * | | 210dcbf3bd6SGeorge Wilson * | | 211dcbf3bd6SGeorge Wilson * +-----------+ 212dcbf3bd6SGeorge Wilson * l2arc_buf_hdr_t| | 213dcbf3bd6SGeorge Wilson * | | 214dcbf3bd6SGeorge Wilson * +-----------+ 215dcbf3bd6SGeorge Wilson * l1arc_buf_hdr_t| | 216dcbf3bd6SGeorge Wilson * | | arc_buf_t (shared) 217dcbf3bd6SGeorge Wilson * | b_buf +------------>+---------+ arc_buf_t 218dcbf3bd6SGeorge Wilson * | | |b_next +---->+---------+ 219770499e1SDan Kimmel * | b_pabd +-+ |---------| |b_next +-->NULL 220dcbf3bd6SGeorge Wilson * +-----------+ | | | +---------+ 221dcbf3bd6SGeorge Wilson * | |b_data +-+ | | 222dcbf3bd6SGeorge Wilson * | +---------+ | |b_data +-+ 223dcbf3bd6SGeorge Wilson * +->+------+ | +---------+ | 224dcbf3bd6SGeorge Wilson * | | | | 225dcbf3bd6SGeorge Wilson * uncompressed | | | | 226dcbf3bd6SGeorge Wilson * data +------+ | | 227dcbf3bd6SGeorge Wilson * ^ +->+------+ | 228dcbf3bd6SGeorge Wilson * | uncompressed | | | 229dcbf3bd6SGeorge Wilson * | data | | | 230dcbf3bd6SGeorge Wilson * | +------+ | 231dcbf3bd6SGeorge Wilson * +---------------------------------+ 232dcbf3bd6SGeorge Wilson * 233770499e1SDan Kimmel * Writing to the ARC requires that the ARC first discard the hdr's b_pabd 234dcbf3bd6SGeorge Wilson * since the physical block is about to be rewritten. The new data contents 2355602294fSDan Kimmel * will be contained in the arc_buf_t. As the I/O pipeline performs the write, 2365602294fSDan Kimmel * it may compress the data before writing it to disk. The ARC will be called 2375602294fSDan Kimmel * with the transformed data and will bcopy the transformed on-disk block into 238770499e1SDan Kimmel * a newly allocated b_pabd. Writes are always done into buffers which have 2395602294fSDan Kimmel * either been loaned (and hence are new and don't have other readers) or 2405602294fSDan Kimmel * buffers which have been released (and hence have their own hdr, if there 2415602294fSDan Kimmel * were originally other readers of the buf's original hdr). This ensures that 2425602294fSDan Kimmel * the ARC only needs to update a single buf and its hdr after a write occurs. 243dcbf3bd6SGeorge Wilson * 244770499e1SDan Kimmel * When the L2ARC is in use, it will also take advantage of the b_pabd. The 245770499e1SDan Kimmel * L2ARC will always write the contents of b_pabd to the L2ARC. This means 2465602294fSDan Kimmel * that when compressed ARC is enabled that the L2ARC blocks are identical 247dcbf3bd6SGeorge Wilson * to the on-disk block in the main data pool. This provides a significant 248dcbf3bd6SGeorge Wilson * advantage since the ARC can leverage the bp's checksum when reading from the 249dcbf3bd6SGeorge Wilson * L2ARC to determine if the contents are valid. However, if the compressed 2505602294fSDan Kimmel * ARC is disabled, then the L2ARC's block must be transformed to look 251dcbf3bd6SGeorge Wilson * like the physical block in the main data pool before comparing the 252dcbf3bd6SGeorge Wilson * checksum and determining its validity. 253dcbf3bd6SGeorge Wilson */ 254dcbf3bd6SGeorge Wilson 255fa9e4066Sahrens #include <sys/spa.h> 256fa9e4066Sahrens #include <sys/zio.h> 257dcbf3bd6SGeorge Wilson #include <sys/spa_impl.h> 258aad02571SSaso Kiselkov #include <sys/zio_compress.h> 259dcbf3bd6SGeorge Wilson #include <sys/zio_checksum.h> 260fa9e4066Sahrens #include <sys/zfs_context.h> 261fa9e4066Sahrens #include <sys/arc.h> 262fa9e4066Sahrens #include <sys/refcount.h> 263c5904d13Seschrock #include <sys/vdev.h> 264573ca77eSGeorge Wilson #include <sys/vdev_impl.h> 26569962b56SMatthew Ahrens #include <sys/dsl_pool.h> 266770499e1SDan Kimmel #include <sys/zio_checksum.h> 267244781f1SPrakash Surya #include <sys/multilist.h> 268770499e1SDan Kimmel #include <sys/abd.h> 269fa9e4066Sahrens #ifdef _KERNEL 270fa9e4066Sahrens #include <sys/vmsystm.h> 271fa9e4066Sahrens #include <vm/anon.h> 272fa9e4066Sahrens #include <sys/fs/swapnode.h> 273033f9833Sek #include <sys/dnlc.h> 274fa9e4066Sahrens #endif 275fa9e4066Sahrens #include <sys/callb.h> 27644cb6abcSbmc #include <sys/kstat.h> 277de753e34SBrad Lewis #include <sys/zthr.h> 278b24ab676SJeff Bonwick #include <zfs_fletcher.h> 2793a2d8a1bSPaul Dagnelie #include <sys/aggsum.h> 2803a2d8a1bSPaul Dagnelie #include <sys/cityhash.h> 281fa9e4066Sahrens 282cd1c8b85SMatthew Ahrens #ifndef _KERNEL 283cd1c8b85SMatthew Ahrens /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 284cd1c8b85SMatthew Ahrens boolean_t arc_watch = B_FALSE; 285cd1c8b85SMatthew Ahrens int arc_procfd; 286cd1c8b85SMatthew Ahrens #endif 287cd1c8b85SMatthew Ahrens 288de753e34SBrad Lewis /* 289de753e34SBrad Lewis * This thread's job is to keep enough free memory in the system, by 290de753e34SBrad Lewis * calling arc_kmem_reap_now() plus arc_shrink(), which improves 291de753e34SBrad Lewis * arc_available_memory(). 292de753e34SBrad Lewis */ 293de753e34SBrad Lewis static zthr_t *arc_reap_zthr; 294de753e34SBrad Lewis 295de753e34SBrad Lewis /* 296de753e34SBrad Lewis * This thread's job is to keep arc_size under arc_c, by calling 297de753e34SBrad Lewis * arc_adjust(), which improves arc_is_overflowing(). 298de753e34SBrad Lewis */ 299de753e34SBrad Lewis static zthr_t *arc_adjust_zthr; 300de753e34SBrad Lewis 301de753e34SBrad Lewis static kmutex_t arc_adjust_lock; 302de753e34SBrad Lewis static kcondvar_t arc_adjust_waiters_cv; 303de753e34SBrad Lewis static boolean_t arc_adjust_needed = B_FALSE; 304244781f1SPrakash Surya 3052ec99e3eSMatthew Ahrens uint_t arc_reduce_dnlc_percent = 3; 306fa9e4066Sahrens 30769962b56SMatthew Ahrens /* 308244781f1SPrakash Surya * The number of headers to evict in arc_evict_state_impl() before 309244781f1SPrakash Surya * dropping the sublist lock and evicting from another sublist. A lower 310244781f1SPrakash Surya * value means we're more likely to evict the "correct" header (i.e. the 311244781f1SPrakash Surya * oldest header in the arc state), but comes with higher overhead 312244781f1SPrakash Surya * (i.e. more invocations of arc_evict_state_impl()). 313244781f1SPrakash Surya */ 314244781f1SPrakash Surya int zfs_arc_evict_batch_limit = 10; 315244781f1SPrakash Surya 316fa9e4066Sahrens /* number of seconds before growing cache again */ 317de753e34SBrad Lewis int arc_grow_retry = 60; 318fa9e4066Sahrens 319de753e34SBrad Lewis /* 320de753e34SBrad Lewis * Minimum time between calls to arc_kmem_reap_soon(). Note that this will 321de753e34SBrad Lewis * be converted to ticks, so with the default hz=100, a setting of 15 ms 322de753e34SBrad Lewis * will actually wait 2 ticks, or 20ms. 323de753e34SBrad Lewis */ 324de753e34SBrad Lewis int arc_kmem_cache_reap_retry_ms = 1000; 32536a64e62STim Kordas 326770499e1SDan Kimmel /* shift of arc_c for calculating overflow limit in arc_get_data_impl */ 327de753e34SBrad Lewis int zfs_arc_overflow_shift = 8; 328244781f1SPrakash Surya 3295a98e54bSBrendan Gregg - Sun Microsystems /* shift of arc_c for calculating both min and max arc_p */ 330de753e34SBrad Lewis int arc_p_min_shift = 4; 3315a98e54bSBrendan Gregg - Sun Microsystems 3325a98e54bSBrendan Gregg - Sun Microsystems /* log2(fraction of arc to reclaim) */ 333de753e34SBrad Lewis int arc_shrink_shift = 7; 3342ec99e3eSMatthew Ahrens 3352ec99e3eSMatthew Ahrens /* 3362ec99e3eSMatthew Ahrens * log2(fraction of ARC which must be free to allow growing). 3372ec99e3eSMatthew Ahrens * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 3382ec99e3eSMatthew Ahrens * when reading a new block into the ARC, we will evict an equal-sized block 3392ec99e3eSMatthew Ahrens * from the ARC. 3402ec99e3eSMatthew Ahrens * 3412ec99e3eSMatthew Ahrens * This must be less than arc_shrink_shift, so that when we shrink the ARC, 3422ec99e3eSMatthew Ahrens * we will still not allow it to grow. 3432ec99e3eSMatthew Ahrens */ 3442ec99e3eSMatthew Ahrens int arc_no_grow_shift = 5; 3452ec99e3eSMatthew Ahrens 3465a98e54bSBrendan Gregg - Sun Microsystems 34713506d1eSmaybee /* 348b19a79ecSperrin * minimum lifespan of a prefetch block in clock ticks 349b19a79ecSperrin * (initialized in arc_init()) 35013506d1eSmaybee */ 351b19a79ecSperrin static int arc_min_prefetch_lifespan; 35213506d1eSmaybee 35369962b56SMatthew Ahrens /* 35469962b56SMatthew Ahrens * If this percent of memory is free, don't throttle. 35569962b56SMatthew Ahrens */ 35669962b56SMatthew Ahrens int arc_lotsfree_percent = 10; 35769962b56SMatthew Ahrens 358de753e34SBrad Lewis static boolean_t arc_initialized; 359fa9e4066Sahrens 3603a737e0dSbrendan /* 3613a737e0dSbrendan * The arc has filled available memory and has now warmed up. 3623a737e0dSbrendan */ 3633a737e0dSbrendan static boolean_t arc_warm; 3643a737e0dSbrendan 3650dd053d7SPrakash Surya /* 3660dd053d7SPrakash Surya * log2 fraction of the zio arena to keep free. 3670dd053d7SPrakash Surya */ 3680dd053d7SPrakash Surya int arc_zio_arena_free_shift = 2; 3690dd053d7SPrakash Surya 370a2eea2e1Sahrens /* 371a2eea2e1Sahrens * These tunables are for performance analysis. 372a2eea2e1Sahrens */ 373a2eea2e1Sahrens uint64_t zfs_arc_max; 374a2eea2e1Sahrens uint64_t zfs_arc_min; 3751116048bSek uint64_t zfs_arc_meta_limit = 0; 3763a5286a1SMatthew Ahrens uint64_t zfs_arc_meta_min = 0; 3775a98e54bSBrendan Gregg - Sun Microsystems int zfs_arc_grow_retry = 0; 3785a98e54bSBrendan Gregg - Sun Microsystems int zfs_arc_shrink_shift = 0; 3795a98e54bSBrendan Gregg - Sun Microsystems int zfs_arc_p_min_shift = 0; 38063e911b6SMatthew Ahrens int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 381a2eea2e1Sahrens 382dcbf3bd6SGeorge Wilson boolean_t zfs_compressed_arc_enabled = B_TRUE; 383dcbf3bd6SGeorge Wilson 384fa9e4066Sahrens /* 385fa94a07fSbrendan * Note that buffers can be in one of 6 states: 386fa9e4066Sahrens * ARC_anon - anonymous (discussed below) 387ea8dc4b6Seschrock * ARC_mru - recently used, currently cached 388ea8dc4b6Seschrock * ARC_mru_ghost - recentely used, no longer in cache 389ea8dc4b6Seschrock * ARC_mfu - frequently used, currently cached 390ea8dc4b6Seschrock * ARC_mfu_ghost - frequently used, no longer in cache 391fa94a07fSbrendan * ARC_l2c_only - exists in L2ARC but not other states 3920e8c6158Smaybee * When there are no active references to the buffer, they are 3930e8c6158Smaybee * are linked onto a list in one of these arc states. These are 3940e8c6158Smaybee * the only buffers that can be evicted or deleted. Within each 3950e8c6158Smaybee * state there are multiple lists, one for meta-data and one for 3960e8c6158Smaybee * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 3970e8c6158Smaybee * etc.) is tracked separately so that it can be managed more 398fa94a07fSbrendan * explicitly: favored over data, limited explicitly. 399fa9e4066Sahrens * 400fa9e4066Sahrens * Anonymous buffers are buffers that are not associated with 401fa9e4066Sahrens * a DVA. These are buffers that hold dirty block copies 402fa9e4066Sahrens * before they are written to stable storage. By definition, 403ea8dc4b6Seschrock * they are "ref'd" and are considered part of arc_mru 404fa9e4066Sahrens * that cannot be freed. Generally, they will aquire a DVA 405ea8dc4b6Seschrock * as they are written and migrate onto the arc_mru list. 406fa94a07fSbrendan * 407fa94a07fSbrendan * The ARC_l2c_only state is for buffers that are in the second 408fa94a07fSbrendan * level ARC but no longer in any of the ARC_m* lists. The second 409fa94a07fSbrendan * level ARC itself may also contain buffers that are in any of 410fa94a07fSbrendan * the ARC_m* states - meaning that a buffer can exist in two 411fa94a07fSbrendan * places. The reason for the ARC_l2c_only state is to keep the 412fa94a07fSbrendan * buffer header in the hash table, so that reads that hit the 413fa94a07fSbrendan * second level ARC benefit from these fast lookups. 414fa9e4066Sahrens */ 415fa9e4066Sahrens 416fa9e4066Sahrens typedef struct arc_state { 417244781f1SPrakash Surya /* 418244781f1SPrakash Surya * list of evictable buffers 419244781f1SPrakash Surya */ 42094c2d0ebSMatthew Ahrens multilist_t *arcs_list[ARC_BUFC_NUMTYPES]; 421244781f1SPrakash Surya /* 422244781f1SPrakash Surya * total amount of evictable data in this state 423244781f1SPrakash Surya */ 424dcbf3bd6SGeorge Wilson refcount_t arcs_esize[ARC_BUFC_NUMTYPES]; 425244781f1SPrakash Surya /* 426244781f1SPrakash Surya * total amount of data in this state; this includes: evictable, 427244781f1SPrakash Surya * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 428244781f1SPrakash Surya */ 4292fd872a7SPrakash Surya refcount_t arcs_size; 430fa9e4066Sahrens } arc_state_t; 431fa9e4066Sahrens 432fa94a07fSbrendan /* The 6 states: */ 433fa9e4066Sahrens static arc_state_t ARC_anon; 434ea8dc4b6Seschrock static arc_state_t ARC_mru; 435ea8dc4b6Seschrock static arc_state_t ARC_mru_ghost; 436ea8dc4b6Seschrock static arc_state_t ARC_mfu; 437ea8dc4b6Seschrock static arc_state_t ARC_mfu_ghost; 438fa94a07fSbrendan static arc_state_t ARC_l2c_only; 439fa9e4066Sahrens 44044cb6abcSbmc typedef struct arc_stats { 44144cb6abcSbmc kstat_named_t arcstat_hits; 44244cb6abcSbmc kstat_named_t arcstat_misses; 44344cb6abcSbmc kstat_named_t arcstat_demand_data_hits; 44444cb6abcSbmc kstat_named_t arcstat_demand_data_misses; 44544cb6abcSbmc kstat_named_t arcstat_demand_metadata_hits; 44644cb6abcSbmc kstat_named_t arcstat_demand_metadata_misses; 44744cb6abcSbmc kstat_named_t arcstat_prefetch_data_hits; 44844cb6abcSbmc kstat_named_t arcstat_prefetch_data_misses; 44944cb6abcSbmc kstat_named_t arcstat_prefetch_metadata_hits; 45044cb6abcSbmc kstat_named_t arcstat_prefetch_metadata_misses; 45144cb6abcSbmc kstat_named_t arcstat_mru_hits; 45244cb6abcSbmc kstat_named_t arcstat_mru_ghost_hits; 45344cb6abcSbmc kstat_named_t arcstat_mfu_hits; 45444cb6abcSbmc kstat_named_t arcstat_mfu_ghost_hits; 45544cb6abcSbmc kstat_named_t arcstat_deleted; 4563e30c24aSWill Andrews /* 4573e30c24aSWill Andrews * Number of buffers that could not be evicted because the hash lock 4583e30c24aSWill Andrews * was held by another thread. The lock may not necessarily be held 4593e30c24aSWill Andrews * by something using the same buffer, since hash locks are shared 4603e30c24aSWill Andrews * by multiple buffers. 4613e30c24aSWill Andrews */ 46244cb6abcSbmc kstat_named_t arcstat_mutex_miss; 4633e30c24aSWill Andrews /* 4643e30c24aSWill Andrews * Number of buffers skipped because they have I/O in progress, are 4653e30c24aSWill Andrews * indrect prefetch buffers that have not lived long enough, or are 4663e30c24aSWill Andrews * not from the spa we're trying to evict from. 4673e30c24aSWill Andrews */ 46844cb6abcSbmc kstat_named_t arcstat_evict_skip; 469244781f1SPrakash Surya /* 470244781f1SPrakash Surya * Number of times arc_evict_state() was unable to evict enough 471244781f1SPrakash Surya * buffers to reach it's target amount. 472244781f1SPrakash Surya */ 473244781f1SPrakash Surya kstat_named_t arcstat_evict_not_enough; 4745ea40c06SBrendan Gregg - Sun Microsystems kstat_named_t arcstat_evict_l2_cached; 4755ea40c06SBrendan Gregg - Sun Microsystems kstat_named_t arcstat_evict_l2_eligible; 4765ea40c06SBrendan Gregg - Sun Microsystems kstat_named_t arcstat_evict_l2_ineligible; 477244781f1SPrakash Surya kstat_named_t arcstat_evict_l2_skip; 47844cb6abcSbmc kstat_named_t arcstat_hash_elements; 47944cb6abcSbmc kstat_named_t arcstat_hash_elements_max; 48044cb6abcSbmc kstat_named_t arcstat_hash_collisions; 48144cb6abcSbmc kstat_named_t arcstat_hash_chains; 48244cb6abcSbmc kstat_named_t arcstat_hash_chain_max; 48344cb6abcSbmc kstat_named_t arcstat_p; 48444cb6abcSbmc kstat_named_t arcstat_c; 48544cb6abcSbmc kstat_named_t arcstat_c_min; 48644cb6abcSbmc kstat_named_t arcstat_c_max; 4873a2d8a1bSPaul Dagnelie /* Not updated directly; only synced in arc_kstat_update. */ 48844cb6abcSbmc kstat_named_t arcstat_size; 489dcbf3bd6SGeorge Wilson /* 490770499e1SDan Kimmel * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd. 491dcbf3bd6SGeorge Wilson * Note that the compressed bytes may match the uncompressed bytes 492dcbf3bd6SGeorge Wilson * if the block is either not compressed or compressed arc is disabled. 493dcbf3bd6SGeorge Wilson */ 494dcbf3bd6SGeorge Wilson kstat_named_t arcstat_compressed_size; 495dcbf3bd6SGeorge Wilson /* 496770499e1SDan Kimmel * Uncompressed size of the data stored in b_pabd. If compressed 497dcbf3bd6SGeorge Wilson * arc is disabled then this value will be identical to the stat 498dcbf3bd6SGeorge Wilson * above. 499dcbf3bd6SGeorge Wilson */ 500dcbf3bd6SGeorge Wilson kstat_named_t arcstat_uncompressed_size; 501dcbf3bd6SGeorge Wilson /* 502dcbf3bd6SGeorge Wilson * Number of bytes stored in all the arc_buf_t's. This is classified 503dcbf3bd6SGeorge Wilson * as "overhead" since this data is typically short-lived and will 504dcbf3bd6SGeorge Wilson * be evicted from the arc when it becomes unreferenced unless the 505dcbf3bd6SGeorge Wilson * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level 506dcbf3bd6SGeorge Wilson * values have been set (see comment in dbuf.c for more information). 507dcbf3bd6SGeorge Wilson */ 508dcbf3bd6SGeorge Wilson kstat_named_t arcstat_overhead_size; 5094076b1bfSPrakash Surya /* 5104076b1bfSPrakash Surya * Number of bytes consumed by internal ARC structures necessary 5114076b1bfSPrakash Surya * for tracking purposes; these structures are not actually 5124076b1bfSPrakash Surya * backed by ARC buffers. This includes arc_buf_hdr_t structures 5134076b1bfSPrakash Surya * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 5144076b1bfSPrakash Surya * caches), and arc_buf_t structures (allocated via arc_buf_t 5154076b1bfSPrakash Surya * cache). 5163a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 5174076b1bfSPrakash Surya */ 518fa94a07fSbrendan kstat_named_t arcstat_hdr_size; 5194076b1bfSPrakash Surya /* 5204076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers of type equal to 5214076b1bfSPrakash Surya * ARC_BUFC_DATA. This is generally consumed by buffers backing 5224076b1bfSPrakash Surya * on disk user data (e.g. plain file contents). 5233a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 5244076b1bfSPrakash Surya */ 5255a98e54bSBrendan Gregg - Sun Microsystems kstat_named_t arcstat_data_size; 5264076b1bfSPrakash Surya /* 5274076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers of type equal to 5284076b1bfSPrakash Surya * ARC_BUFC_METADATA. This is generally consumed by buffers 5294076b1bfSPrakash Surya * backing on disk data that is used for internal ZFS 5304076b1bfSPrakash Surya * structures (e.g. ZAP, dnode, indirect blocks, etc). 5313a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 5324076b1bfSPrakash Surya */ 5334076b1bfSPrakash Surya kstat_named_t arcstat_metadata_size; 5344076b1bfSPrakash Surya /* 5354076b1bfSPrakash Surya * Number of bytes consumed by various buffers and structures 5364076b1bfSPrakash Surya * not actually backed with ARC buffers. This includes bonus 5374076b1bfSPrakash Surya * buffers (allocated directly via zio_buf_* functions), 5384076b1bfSPrakash Surya * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 5394076b1bfSPrakash Surya * cache), and dnode_t structures (allocated via dnode_t cache). 5403a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 5414076b1bfSPrakash Surya */ 5425a98e54bSBrendan Gregg - Sun Microsystems kstat_named_t arcstat_other_size; 5434076b1bfSPrakash Surya /* 5444076b1bfSPrakash Surya * Total number of bytes consumed by ARC buffers residing in the 5454076b1bfSPrakash Surya * arc_anon state. This includes *all* buffers in the arc_anon 5464076b1bfSPrakash Surya * state; e.g. data, metadata, evictable, and unevictable buffers 5474076b1bfSPrakash Surya * are all included in this value. 5483a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 5494076b1bfSPrakash Surya */ 5504076b1bfSPrakash Surya kstat_named_t arcstat_anon_size; 5514076b1bfSPrakash Surya /* 5524076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that meet the 5534076b1bfSPrakash Surya * following criteria: backing buffers of type ARC_BUFC_DATA, 5544076b1bfSPrakash Surya * residing in the arc_anon state, and are eligible for eviction 5554076b1bfSPrakash Surya * (e.g. have no outstanding holds on the buffer). 5563a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 5574076b1bfSPrakash Surya */ 5584076b1bfSPrakash Surya kstat_named_t arcstat_anon_evictable_data; 5594076b1bfSPrakash Surya /* 5604076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that meet the 5614076b1bfSPrakash Surya * following criteria: backing buffers of type ARC_BUFC_METADATA, 5624076b1bfSPrakash Surya * residing in the arc_anon state, and are eligible for eviction 5634076b1bfSPrakash Surya * (e.g. have no outstanding holds on the buffer). 5643a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 5654076b1bfSPrakash Surya */ 5664076b1bfSPrakash Surya kstat_named_t arcstat_anon_evictable_metadata; 5674076b1bfSPrakash Surya /* 5684076b1bfSPrakash Surya * Total number of bytes consumed by ARC buffers residing in the 5694076b1bfSPrakash Surya * arc_mru state. This includes *all* buffers in the arc_mru 5704076b1bfSPrakash Surya * state; e.g. data, metadata, evictable, and unevictable buffers 5714076b1bfSPrakash Surya * are all included in this value. 5723a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 5734076b1bfSPrakash Surya */ 5744076b1bfSPrakash Surya kstat_named_t arcstat_mru_size; 5754076b1bfSPrakash Surya /* 5764076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that meet the 5774076b1bfSPrakash Surya * following criteria: backing buffers of type ARC_BUFC_DATA, 5784076b1bfSPrakash Surya * residing in the arc_mru state, and are eligible for eviction 5794076b1bfSPrakash Surya * (e.g. have no outstanding holds on the buffer). 5803a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 5814076b1bfSPrakash Surya */ 5824076b1bfSPrakash Surya kstat_named_t arcstat_mru_evictable_data; 5834076b1bfSPrakash Surya /* 5844076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that meet the 5854076b1bfSPrakash Surya * following criteria: backing buffers of type ARC_BUFC_METADATA, 5864076b1bfSPrakash Surya * residing in the arc_mru state, and are eligible for eviction 5874076b1bfSPrakash Surya * (e.g. have no outstanding holds on the buffer). 5883a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 5894076b1bfSPrakash Surya */ 5904076b1bfSPrakash Surya kstat_named_t arcstat_mru_evictable_metadata; 5914076b1bfSPrakash Surya /* 5924076b1bfSPrakash Surya * Total number of bytes that *would have been* consumed by ARC 5934076b1bfSPrakash Surya * buffers in the arc_mru_ghost state. The key thing to note 5944076b1bfSPrakash Surya * here, is the fact that this size doesn't actually indicate 5954076b1bfSPrakash Surya * RAM consumption. The ghost lists only consist of headers and 5964076b1bfSPrakash Surya * don't actually have ARC buffers linked off of these headers. 5974076b1bfSPrakash Surya * Thus, *if* the headers had associated ARC buffers, these 5984076b1bfSPrakash Surya * buffers *would have* consumed this number of bytes. 5993a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 6004076b1bfSPrakash Surya */ 6014076b1bfSPrakash Surya kstat_named_t arcstat_mru_ghost_size; 6024076b1bfSPrakash Surya /* 6034076b1bfSPrakash Surya * Number of bytes that *would have been* consumed by ARC 6044076b1bfSPrakash Surya * buffers that are eligible for eviction, of type 6054076b1bfSPrakash Surya * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 6063a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 6074076b1bfSPrakash Surya */ 6084076b1bfSPrakash Surya kstat_named_t arcstat_mru_ghost_evictable_data; 6094076b1bfSPrakash Surya /* 6104076b1bfSPrakash Surya * Number of bytes that *would have been* consumed by ARC 6114076b1bfSPrakash Surya * buffers that are eligible for eviction, of type 6124076b1bfSPrakash Surya * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 6133a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 6144076b1bfSPrakash Surya */ 6154076b1bfSPrakash Surya kstat_named_t arcstat_mru_ghost_evictable_metadata; 6164076b1bfSPrakash Surya /* 6174076b1bfSPrakash Surya * Total number of bytes consumed by ARC buffers residing in the 6184076b1bfSPrakash Surya * arc_mfu state. This includes *all* buffers in the arc_mfu 6194076b1bfSPrakash Surya * state; e.g. data, metadata, evictable, and unevictable buffers 6204076b1bfSPrakash Surya * are all included in this value. 6213a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 6224076b1bfSPrakash Surya */ 6234076b1bfSPrakash Surya kstat_named_t arcstat_mfu_size; 6244076b1bfSPrakash Surya /* 6254076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that are eligible for 6264076b1bfSPrakash Surya * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 6274076b1bfSPrakash Surya * state. 6283a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 6294076b1bfSPrakash Surya */ 6304076b1bfSPrakash Surya kstat_named_t arcstat_mfu_evictable_data; 6314076b1bfSPrakash Surya /* 6324076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that are eligible for 6334076b1bfSPrakash Surya * eviction, of type ARC_BUFC_METADATA, and reside in the 6344076b1bfSPrakash Surya * arc_mfu state. 6353a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 6364076b1bfSPrakash Surya */ 6374076b1bfSPrakash Surya kstat_named_t arcstat_mfu_evictable_metadata; 6384076b1bfSPrakash Surya /* 6394076b1bfSPrakash Surya * Total number of bytes that *would have been* consumed by ARC 6404076b1bfSPrakash Surya * buffers in the arc_mfu_ghost state. See the comment above 6414076b1bfSPrakash Surya * arcstat_mru_ghost_size for more details. 6423a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 6434076b1bfSPrakash Surya */ 6444076b1bfSPrakash Surya kstat_named_t arcstat_mfu_ghost_size; 6454076b1bfSPrakash Surya /* 6464076b1bfSPrakash Surya * Number of bytes that *would have been* consumed by ARC 6474076b1bfSPrakash Surya * buffers that are eligible for eviction, of type 6484076b1bfSPrakash Surya * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 6493a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 6504076b1bfSPrakash Surya */ 6514076b1bfSPrakash Surya kstat_named_t arcstat_mfu_ghost_evictable_data; 6524076b1bfSPrakash Surya /* 6534076b1bfSPrakash Surya * Number of bytes that *would have been* consumed by ARC 6544076b1bfSPrakash Surya * buffers that are eligible for eviction, of type 6554076b1bfSPrakash Surya * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 6563a2d8a1bSPaul Dagnelie * Not updated directly; only synced in arc_kstat_update. 6574076b1bfSPrakash Surya */ 6584076b1bfSPrakash Surya kstat_named_t arcstat_mfu_ghost_evictable_metadata; 659fa94a07fSbrendan kstat_named_t arcstat_l2_hits; 660fa94a07fSbrendan kstat_named_t arcstat_l2_misses; 661fa94a07fSbrendan kstat_named_t arcstat_l2_feeds; 662fa94a07fSbrendan kstat_named_t arcstat_l2_rw_clash; 6635a98e54bSBrendan Gregg - Sun Microsystems kstat_named_t arcstat_l2_read_bytes; 6645a98e54bSBrendan Gregg - Sun Microsystems kstat_named_t arcstat_l2_write_bytes; 665fa94a07fSbrendan kstat_named_t arcstat_l2_writes_sent; 666fa94a07fSbrendan kstat_named_t arcstat_l2_writes_done; 667fa94a07fSbrendan kstat_named_t arcstat_l2_writes_error; 668244781f1SPrakash Surya kstat_named_t arcstat_l2_writes_lock_retry; 669fa94a07fSbrendan kstat_named_t arcstat_l2_evict_lock_retry; 670fa94a07fSbrendan kstat_named_t arcstat_l2_evict_reading; 67189c86e32SChris Williamson kstat_named_t arcstat_l2_evict_l1cached; 672fa94a07fSbrendan kstat_named_t arcstat_l2_free_on_write; 673fa94a07fSbrendan kstat_named_t arcstat_l2_abort_lowmem; 674fa94a07fSbrendan kstat_named_t arcstat_l2_cksum_bad; 675fa94a07fSbrendan kstat_named_t arcstat_l2_io_error; 67616a7e5acSAndriy Gapon kstat_named_t arcstat_l2_lsize; 67716a7e5acSAndriy Gapon kstat_named_t arcstat_l2_psize; 6783a2d8a1bSPaul Dagnelie /* Not updated directly; only synced in arc_kstat_update. */ 679fa94a07fSbrendan kstat_named_t arcstat_l2_hdr_size; 6801ab7f2deSmaybee kstat_named_t arcstat_memory_throttle_count; 6813a2d8a1bSPaul Dagnelie /* Not updated directly; only synced in arc_kstat_update. */ 68220128a08SGeorge Wilson kstat_named_t arcstat_meta_used; 68320128a08SGeorge Wilson kstat_named_t arcstat_meta_limit; 68420128a08SGeorge Wilson kstat_named_t arcstat_meta_max; 6853a5286a1SMatthew Ahrens kstat_named_t arcstat_meta_min; 686cf6106c8SMatthew Ahrens kstat_named_t arcstat_sync_wait_for_async; 687cf6106c8SMatthew Ahrens kstat_named_t arcstat_demand_hit_predictive_prefetch; 68844cb6abcSbmc } arc_stats_t; 68944cb6abcSbmc 69044cb6abcSbmc static arc_stats_t arc_stats = { 69144cb6abcSbmc { "hits", KSTAT_DATA_UINT64 }, 69244cb6abcSbmc { "misses", KSTAT_DATA_UINT64 }, 69344cb6abcSbmc { "demand_data_hits", KSTAT_DATA_UINT64 }, 69444cb6abcSbmc { "demand_data_misses", KSTAT_DATA_UINT64 }, 69544cb6abcSbmc { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 69644cb6abcSbmc { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 69744cb6abcSbmc { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 69844cb6abcSbmc { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 69944cb6abcSbmc { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 70044cb6abcSbmc { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 70144cb6abcSbmc { "mru_hits", KSTAT_DATA_UINT64 }, 70244cb6abcSbmc { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 70344cb6abcSbmc { "mfu_hits", KSTAT_DATA_UINT64 }, 70444cb6abcSbmc { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 70544cb6abcSbmc { "deleted", KSTAT_DATA_UINT64 }, 70644cb6abcSbmc { "mutex_miss", KSTAT_DATA_UINT64 }, 70744cb6abcSbmc { "evict_skip", KSTAT_DATA_UINT64 }, 708244781f1SPrakash Surya { "evict_not_enough", KSTAT_DATA_UINT64 }, 7095ea40c06SBrendan Gregg - Sun Microsystems { "evict_l2_cached", KSTAT_DATA_UINT64 }, 7105ea40c06SBrendan Gregg - Sun Microsystems { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 7115ea40c06SBrendan Gregg - Sun Microsystems { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 712244781f1SPrakash Surya { "evict_l2_skip", KSTAT_DATA_UINT64 }, 71344cb6abcSbmc { "hash_elements", KSTAT_DATA_UINT64 }, 71444cb6abcSbmc { "hash_elements_max", KSTAT_DATA_UINT64 }, 71544cb6abcSbmc { "hash_collisions", KSTAT_DATA_UINT64 }, 71644cb6abcSbmc { "hash_chains", KSTAT_DATA_UINT64 }, 71744cb6abcSbmc { "hash_chain_max", KSTAT_DATA_UINT64 }, 71844cb6abcSbmc { "p", KSTAT_DATA_UINT64 }, 71944cb6abcSbmc { "c", KSTAT_DATA_UINT64 }, 72044cb6abcSbmc { "c_min", KSTAT_DATA_UINT64 }, 72144cb6abcSbmc { "c_max", KSTAT_DATA_UINT64 }, 722fa94a07fSbrendan { "size", KSTAT_DATA_UINT64 }, 723dcbf3bd6SGeorge Wilson { "compressed_size", KSTAT_DATA_UINT64 }, 724dcbf3bd6SGeorge Wilson { "uncompressed_size", KSTAT_DATA_UINT64 }, 725dcbf3bd6SGeorge Wilson { "overhead_size", KSTAT_DATA_UINT64 }, 726fa94a07fSbrendan { "hdr_size", KSTAT_DATA_UINT64 }, 7275a98e54bSBrendan Gregg - Sun Microsystems { "data_size", KSTAT_DATA_UINT64 }, 7284076b1bfSPrakash Surya { "metadata_size", KSTAT_DATA_UINT64 }, 7295a98e54bSBrendan Gregg - Sun Microsystems { "other_size", KSTAT_DATA_UINT64 }, 7304076b1bfSPrakash Surya { "anon_size", KSTAT_DATA_UINT64 }, 7314076b1bfSPrakash Surya { "anon_evictable_data", KSTAT_DATA_UINT64 }, 7324076b1bfSPrakash Surya { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 7334076b1bfSPrakash Surya { "mru_size", KSTAT_DATA_UINT64 }, 7344076b1bfSPrakash Surya { "mru_evictable_data", KSTAT_DATA_UINT64 }, 7354076b1bfSPrakash Surya { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 7364076b1bfSPrakash Surya { "mru_ghost_size", KSTAT_DATA_UINT64 }, 7374076b1bfSPrakash Surya { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 7384076b1bfSPrakash Surya { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 7394076b1bfSPrakash Surya { "mfu_size", KSTAT_DATA_UINT64 }, 7404076b1bfSPrakash Surya { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 7414076b1bfSPrakash Surya { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 7424076b1bfSPrakash Surya { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 7434076b1bfSPrakash Surya { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 7444076b1bfSPrakash Surya { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 745fa94a07fSbrendan { "l2_hits", KSTAT_DATA_UINT64 }, 746fa94a07fSbrendan { "l2_misses", KSTAT_DATA_UINT64 }, 747fa94a07fSbrendan { "l2_feeds", KSTAT_DATA_UINT64 }, 748fa94a07fSbrendan { "l2_rw_clash", KSTAT_DATA_UINT64 }, 7495a98e54bSBrendan Gregg - Sun Microsystems { "l2_read_bytes", KSTAT_DATA_UINT64 }, 7505a98e54bSBrendan Gregg - Sun Microsystems { "l2_write_bytes", KSTAT_DATA_UINT64 }, 751fa94a07fSbrendan { "l2_writes_sent", KSTAT_DATA_UINT64 }, 752fa94a07fSbrendan { "l2_writes_done", KSTAT_DATA_UINT64 }, 753fa94a07fSbrendan { "l2_writes_error", KSTAT_DATA_UINT64 }, 754244781f1SPrakash Surya { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 755fa94a07fSbrendan { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 756fa94a07fSbrendan { "l2_evict_reading", KSTAT_DATA_UINT64 }, 75789c86e32SChris Williamson { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 758fa94a07fSbrendan { "l2_free_on_write", KSTAT_DATA_UINT64 }, 759fa94a07fSbrendan { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 760fa94a07fSbrendan { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 761fa94a07fSbrendan { "l2_io_error", KSTAT_DATA_UINT64 }, 762fa94a07fSbrendan { "l2_size", KSTAT_DATA_UINT64 }, 763aad02571SSaso Kiselkov { "l2_asize", KSTAT_DATA_UINT64 }, 7641ab7f2deSmaybee { "l2_hdr_size", KSTAT_DATA_UINT64 }, 7659253d63dSGeorge Wilson { "memory_throttle_count", KSTAT_DATA_UINT64 }, 76620128a08SGeorge Wilson { "arc_meta_used", KSTAT_DATA_UINT64 }, 76720128a08SGeorge Wilson { "arc_meta_limit", KSTAT_DATA_UINT64 }, 7683a5286a1SMatthew Ahrens { "arc_meta_max", KSTAT_DATA_UINT64 }, 769cf6106c8SMatthew Ahrens { "arc_meta_min", KSTAT_DATA_UINT64 }, 770cf6106c8SMatthew Ahrens { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 771cf6106c8SMatthew Ahrens { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 77244cb6abcSbmc }; 77344cb6abcSbmc 77444cb6abcSbmc #define ARCSTAT(stat) (arc_stats.stat.value.ui64) 77544cb6abcSbmc 77644cb6abcSbmc #define ARCSTAT_INCR(stat, val) \ 777f7170741SWill Andrews atomic_add_64(&arc_stats.stat.value.ui64, (val)) 77844cb6abcSbmc 779b24ab676SJeff Bonwick #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 78044cb6abcSbmc #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 78144cb6abcSbmc 78244cb6abcSbmc #define ARCSTAT_MAX(stat, val) { \ 78344cb6abcSbmc uint64_t m; \ 78444cb6abcSbmc while ((val) > (m = arc_stats.stat.value.ui64) && \ 78544cb6abcSbmc (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 78644cb6abcSbmc continue; \ 78744cb6abcSbmc } 78844cb6abcSbmc 78944cb6abcSbmc #define ARCSTAT_MAXSTAT(stat) \ 79044cb6abcSbmc ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 79144cb6abcSbmc 79244cb6abcSbmc /* 79344cb6abcSbmc * We define a macro to allow ARC hits/misses to be easily broken down by 79444cb6abcSbmc * two separate conditions, giving a total of four different subtypes for 79544cb6abcSbmc * each of hits and misses (so eight statistics total). 79644cb6abcSbmc */ 79744cb6abcSbmc #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 79844cb6abcSbmc if (cond1) { \ 79944cb6abcSbmc if (cond2) { \ 80044cb6abcSbmc ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 80144cb6abcSbmc } else { \ 80244cb6abcSbmc ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 80344cb6abcSbmc } \ 80444cb6abcSbmc } else { \ 80544cb6abcSbmc if (cond2) { \ 80644cb6abcSbmc ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 80744cb6abcSbmc } else { \ 80844cb6abcSbmc ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 80944cb6abcSbmc } \ 81044cb6abcSbmc } 81144cb6abcSbmc 81244cb6abcSbmc kstat_t *arc_ksp; 813b24ab676SJeff Bonwick static arc_state_t *arc_anon; 81444cb6abcSbmc static arc_state_t *arc_mru; 81544cb6abcSbmc static arc_state_t *arc_mru_ghost; 81644cb6abcSbmc static arc_state_t *arc_mfu; 81744cb6abcSbmc static arc_state_t *arc_mfu_ghost; 818fa94a07fSbrendan static arc_state_t *arc_l2c_only; 81944cb6abcSbmc 82044cb6abcSbmc /* 82144cb6abcSbmc * There are several ARC variables that are critical to export as kstats -- 82244cb6abcSbmc * but we don't want to have to grovel around in the kstat whenever we wish to 82344cb6abcSbmc * manipulate them. For these variables, we therefore define them to be in 82444cb6abcSbmc * terms of the statistic variable. This assures that we are not introducing 82544cb6abcSbmc * the possibility of inconsistency by having shadow copies of the variables, 82644cb6abcSbmc * while still allowing the code to be readable. 82744cb6abcSbmc */ 82844cb6abcSbmc #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 82944cb6abcSbmc #define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 83044cb6abcSbmc #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 83144cb6abcSbmc #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 83220128a08SGeorge Wilson #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 8333a5286a1SMatthew Ahrens #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 83420128a08SGeorge Wilson #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 83544cb6abcSbmc 836dcbf3bd6SGeorge Wilson /* compressed size of entire arc */ 837dcbf3bd6SGeorge Wilson #define arc_compressed_size ARCSTAT(arcstat_compressed_size) 838dcbf3bd6SGeorge Wilson /* uncompressed size of entire arc */ 839dcbf3bd6SGeorge Wilson #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size) 840dcbf3bd6SGeorge Wilson /* number of bytes in the arc from arc_buf_t's */ 841dcbf3bd6SGeorge Wilson #define arc_overhead_size ARCSTAT(arcstat_overhead_size) 842aad02571SSaso Kiselkov 8433a2d8a1bSPaul Dagnelie /* 8443a2d8a1bSPaul Dagnelie * There are also some ARC variables that we want to export, but that are 8453a2d8a1bSPaul Dagnelie * updated so often that having the canonical representation be the statistic 8463a2d8a1bSPaul Dagnelie * variable causes a performance bottleneck. We want to use aggsum_t's for these 8473a2d8a1bSPaul Dagnelie * instead, but still be able to export the kstat in the same way as before. 8483a2d8a1bSPaul Dagnelie * The solution is to always use the aggsum version, except in the kstat update 8493a2d8a1bSPaul Dagnelie * callback. 8503a2d8a1bSPaul Dagnelie */ 8513a2d8a1bSPaul Dagnelie aggsum_t arc_size; 8523a2d8a1bSPaul Dagnelie aggsum_t arc_meta_used; 8533a2d8a1bSPaul Dagnelie aggsum_t astat_data_size; 8543a2d8a1bSPaul Dagnelie aggsum_t astat_metadata_size; 8553a2d8a1bSPaul Dagnelie aggsum_t astat_hdr_size; 8563a2d8a1bSPaul Dagnelie aggsum_t astat_other_size; 8573a2d8a1bSPaul Dagnelie aggsum_t astat_l2_hdr_size; 8583a2d8a1bSPaul Dagnelie 85944cb6abcSbmc static int arc_no_grow; /* Don't try to grow cache size */ 860de753e34SBrad Lewis static hrtime_t arc_growtime; 86144cb6abcSbmc static uint64_t arc_tempreserve; 8622fdbea25SAleksandr Guzovskiy static uint64_t arc_loaned_bytes; 863fa9e4066Sahrens 864fa9e4066Sahrens typedef struct arc_callback arc_callback_t; 865fa9e4066Sahrens 866fa9e4066Sahrens struct arc_callback { 867fa9e4066Sahrens void *acb_private; 868c717a561Smaybee arc_done_func_t *acb_done; 869fa9e4066Sahrens arc_buf_t *acb_buf; 8705602294fSDan Kimmel boolean_t acb_compressed; 871fa9e4066Sahrens zio_t *acb_zio_dummy; 872fa9e4066Sahrens arc_callback_t *acb_next; 873fa9e4066Sahrens }; 874fa9e4066Sahrens 875c717a561Smaybee typedef struct arc_write_callback arc_write_callback_t; 876c717a561Smaybee 877c717a561Smaybee struct arc_write_callback { 878c717a561Smaybee void *awcb_private; 879c717a561Smaybee arc_done_func_t *awcb_ready; 8808df0bcf0SPaul Dagnelie arc_done_func_t *awcb_children_ready; 88169962b56SMatthew Ahrens arc_done_func_t *awcb_physdone; 882c717a561Smaybee arc_done_func_t *awcb_done; 883c717a561Smaybee arc_buf_t *awcb_buf; 884c717a561Smaybee }; 885c717a561Smaybee 88689c86e32SChris Williamson /* 88789c86e32SChris Williamson * ARC buffers are separated into multiple structs as a memory saving measure: 88889c86e32SChris Williamson * - Common fields struct, always defined, and embedded within it: 88989c86e32SChris Williamson * - L2-only fields, always allocated but undefined when not in L2ARC 89089c86e32SChris Williamson * - L1-only fields, only allocated when in L1ARC 89189c86e32SChris Williamson * 89289c86e32SChris Williamson * Buffer in L1 Buffer only in L2 89389c86e32SChris Williamson * +------------------------+ +------------------------+ 89489c86e32SChris Williamson * | arc_buf_hdr_t | | arc_buf_hdr_t | 89589c86e32SChris Williamson * | | | | 89689c86e32SChris Williamson * | | | | 89789c86e32SChris Williamson * | | | | 89889c86e32SChris Williamson * +------------------------+ +------------------------+ 89989c86e32SChris Williamson * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 90089c86e32SChris Williamson * | (undefined if L1-only) | | | 90189c86e32SChris Williamson * +------------------------+ +------------------------+ 90289c86e32SChris Williamson * | l1arc_buf_hdr_t | 90389c86e32SChris Williamson * | | 90489c86e32SChris Williamson * | | 90589c86e32SChris Williamson * | | 90689c86e32SChris Williamson * | | 90789c86e32SChris Williamson * +------------------------+ 90889c86e32SChris Williamson * 90989c86e32SChris Williamson * Because it's possible for the L2ARC to become extremely large, we can wind 91089c86e32SChris Williamson * up eating a lot of memory in L2ARC buffer headers, so the size of a header 91189c86e32SChris Williamson * is minimized by only allocating the fields necessary for an L1-cached buffer 91289c86e32SChris Williamson * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 91389c86e32SChris Williamson * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 91489c86e32SChris Williamson * words in pointers. arc_hdr_realloc() is used to switch a header between 91589c86e32SChris Williamson * these two allocation states. 91689c86e32SChris Williamson */ 91789c86e32SChris Williamson typedef struct l1arc_buf_hdr { 9186b4acc8bSahrens kmutex_t b_freeze_lock; 919dcbf3bd6SGeorge Wilson zio_cksum_t *b_freeze_cksum; 92089c86e32SChris Williamson #ifdef ZFS_DEBUG 92189c86e32SChris Williamson /* 9225602294fSDan Kimmel * Used for debugging with kmem_flags - by allocating and freeing 92389c86e32SChris Williamson * b_thawed when the buffer is thawed, we get a record of the stack 92489c86e32SChris Williamson * trace that thawed it. 92589c86e32SChris Williamson */ 9263f9d6ad7SLin Ling void *b_thawed; 92789c86e32SChris Williamson #endif 9286b4acc8bSahrens 929fa9e4066Sahrens arc_buf_t *b_buf; 930dcbf3bd6SGeorge Wilson uint32_t b_bufcnt; 93189c86e32SChris Williamson /* for waiting on writes to complete */ 932ad23a2dbSjohansen kcondvar_t b_cv; 933dcbf3bd6SGeorge Wilson uint8_t b_byteswap; 934ad23a2dbSjohansen 935fa9e4066Sahrens /* protected by arc state mutex */ 936fa9e4066Sahrens arc_state_t *b_state; 937244781f1SPrakash Surya multilist_node_t b_arc_node; 938fa9e4066Sahrens 939fa9e4066Sahrens /* updated atomically */ 940fa9e4066Sahrens clock_t b_arc_access; 941fa9e4066Sahrens 942fa9e4066Sahrens /* self protecting */ 943fa9e4066Sahrens refcount_t b_refcnt; 944fa94a07fSbrendan 94589c86e32SChris Williamson arc_callback_t *b_acb; 946770499e1SDan Kimmel abd_t *b_pabd; 94789c86e32SChris Williamson } l1arc_buf_hdr_t; 94889c86e32SChris Williamson 94989c86e32SChris Williamson typedef struct l2arc_dev l2arc_dev_t; 95089c86e32SChris Williamson 95189c86e32SChris Williamson typedef struct l2arc_buf_hdr { 95289c86e32SChris Williamson /* protected by arc_buf_hdr mutex */ 95389c86e32SChris Williamson l2arc_dev_t *b_dev; /* L2ARC device */ 95489c86e32SChris Williamson uint64_t b_daddr; /* disk address, offset byte */ 95589c86e32SChris Williamson 956fa94a07fSbrendan list_node_t b_l2node; 95789c86e32SChris Williamson } l2arc_buf_hdr_t; 95889c86e32SChris Williamson 95989c86e32SChris Williamson struct arc_buf_hdr { 96089c86e32SChris Williamson /* protected by hash lock */ 96189c86e32SChris Williamson dva_t b_dva; 96289c86e32SChris Williamson uint64_t b_birth; 96389c86e32SChris Williamson 964dcbf3bd6SGeorge Wilson arc_buf_contents_t b_type; 96589c86e32SChris Williamson arc_buf_hdr_t *b_hash_next; 96689c86e32SChris Williamson arc_flags_t b_flags; 96789c86e32SChris Williamson 968dcbf3bd6SGeorge Wilson /* 969dcbf3bd6SGeorge Wilson * This field stores the size of the data buffer after 970dcbf3bd6SGeorge Wilson * compression, and is set in the arc's zio completion handlers. 971dcbf3bd6SGeorge Wilson * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes). 972dcbf3bd6SGeorge Wilson * 973dcbf3bd6SGeorge Wilson * While the block pointers can store up to 32MB in their psize 974dcbf3bd6SGeorge Wilson * field, we can only store up to 32MB minus 512B. This is due 975dcbf3bd6SGeorge Wilson * to the bp using a bias of 1, whereas we use a bias of 0 (i.e. 976dcbf3bd6SGeorge Wilson * a field of zeros represents 512B in the bp). We can't use a 977dcbf3bd6SGeorge Wilson * bias of 1 since we need to reserve a psize of zero, here, to 978dcbf3bd6SGeorge Wilson * represent holes and embedded blocks. 979dcbf3bd6SGeorge Wilson * 980dcbf3bd6SGeorge Wilson * This isn't a problem in practice, since the maximum size of a 981dcbf3bd6SGeorge Wilson * buffer is limited to 16MB, so we never need to store 32MB in 982dcbf3bd6SGeorge Wilson * this field. Even in the upstream illumos code base, the 983dcbf3bd6SGeorge Wilson * maximum size of a buffer is limited to 16MB. 984dcbf3bd6SGeorge Wilson */ 985dcbf3bd6SGeorge Wilson uint16_t b_psize; 986dcbf3bd6SGeorge Wilson 987dcbf3bd6SGeorge Wilson /* 988dcbf3bd6SGeorge Wilson * This field stores the size of the data buffer before 989dcbf3bd6SGeorge Wilson * compression, and cannot change once set. It is in units 990dcbf3bd6SGeorge Wilson * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes) 991dcbf3bd6SGeorge Wilson */ 992dcbf3bd6SGeorge Wilson uint16_t b_lsize; /* immutable */ 993dcbf3bd6SGeorge Wilson uint64_t b_spa; /* immutable */ 99489c86e32SChris Williamson 99589c86e32SChris Williamson /* L2ARC fields. Undefined when not in L2ARC. */ 99689c86e32SChris Williamson l2arc_buf_hdr_t b_l2hdr; 99789c86e32SChris Williamson /* L1ARC fields. Undefined when in l2arc_only state */ 99889c86e32SChris Williamson l1arc_buf_hdr_t b_l1hdr; 999fa9e4066Sahrens }; 1000fa9e4066Sahrens 1001ea8dc4b6Seschrock #define GHOST_STATE(state) \ 1002fa94a07fSbrendan ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 1003fa94a07fSbrendan (state) == arc_l2c_only) 1004ea8dc4b6Seschrock 10057adb730bSGeorge Wilson #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 10067adb730bSGeorge Wilson #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 10077adb730bSGeorge Wilson #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 10087adb730bSGeorge Wilson #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 1009dcbf3bd6SGeorge Wilson #define HDR_COMPRESSION_ENABLED(hdr) \ 1010dcbf3bd6SGeorge Wilson ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC) 101189c86e32SChris Williamson 10127adb730bSGeorge Wilson #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 10137adb730bSGeorge Wilson #define HDR_L2_READING(hdr) \ 1014dcbf3bd6SGeorge Wilson (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 1015dcbf3bd6SGeorge Wilson ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 10167adb730bSGeorge Wilson #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 10177adb730bSGeorge Wilson #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 10187adb730bSGeorge Wilson #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 1019dcbf3bd6SGeorge Wilson #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA) 1020fa9e4066Sahrens 102189c86e32SChris Williamson #define HDR_ISTYPE_METADATA(hdr) \ 1022dcbf3bd6SGeorge Wilson ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 102389c86e32SChris Williamson #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 102489c86e32SChris Williamson 102589c86e32SChris Williamson #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 102689c86e32SChris Williamson #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 102789c86e32SChris Williamson 1028dcbf3bd6SGeorge Wilson /* For storing compression mode in b_flags */ 1029dcbf3bd6SGeorge Wilson #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1) 1030dcbf3bd6SGeorge Wilson 1031dcbf3bd6SGeorge Wilson #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \ 1032dcbf3bd6SGeorge Wilson HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS)) 1033dcbf3bd6SGeorge Wilson #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \ 1034dcbf3bd6SGeorge Wilson HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp)); 1035dcbf3bd6SGeorge Wilson 1036dcbf3bd6SGeorge Wilson #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL) 10375602294fSDan Kimmel #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED) 10385602294fSDan Kimmel #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED) 1039dcbf3bd6SGeorge Wilson 1040e6c728e1Sbrendan /* 1041e6c728e1Sbrendan * Other sizes 1042e6c728e1Sbrendan */ 1043e6c728e1Sbrendan 104489c86e32SChris Williamson #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 104589c86e32SChris Williamson #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 1046e6c728e1Sbrendan 1047fa9e4066Sahrens /* 1048fa9e4066Sahrens * Hash table routines 1049fa9e4066Sahrens */ 1050fa9e4066Sahrens 1051fa9e4066Sahrens #define HT_LOCK_PAD 64 1052fa9e4066Sahrens 1053fa9e4066Sahrens struct ht_lock { 1054fa9e4066Sahrens kmutex_t ht_lock; 1055fa9e4066Sahrens #ifdef _KERNEL 1056fa9e4066Sahrens unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 1057fa9e4066Sahrens #endif 1058fa9e4066Sahrens }; 1059fa9e4066Sahrens 1060fa9e4066Sahrens #define BUF_LOCKS 256 1061fa9e4066Sahrens typedef struct buf_hash_table { 1062fa9e4066Sahrens uint64_t ht_mask; 1063fa9e4066Sahrens arc_buf_hdr_t **ht_table; 1064fa9e4066Sahrens struct ht_lock ht_locks[BUF_LOCKS]; 1065fa9e4066Sahrens } buf_hash_table_t; 1066fa9e4066Sahrens 1067fa9e4066Sahrens static buf_hash_table_t buf_hash_table; 1068fa9e4066Sahrens 1069fa9e4066Sahrens #define BUF_HASH_INDEX(spa, dva, birth) \ 1070fa9e4066Sahrens (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 1071fa9e4066Sahrens #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 1072fa9e4066Sahrens #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 10733f9d6ad7SLin Ling #define HDR_LOCK(hdr) \ 10743f9d6ad7SLin Ling (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 1075fa9e4066Sahrens 1076fa9e4066Sahrens uint64_t zfs_crc64_table[256]; 1077fa9e4066Sahrens 1078fa94a07fSbrendan /* 1079fa94a07fSbrendan * Level 2 ARC 1080fa94a07fSbrendan */ 1081fa94a07fSbrendan 1082fa94a07fSbrendan #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 1083aad02571SSaso Kiselkov #define L2ARC_HEADROOM 2 /* num of writes */ 1084aad02571SSaso Kiselkov /* 1085aad02571SSaso Kiselkov * If we discover during ARC scan any buffers to be compressed, we boost 1086aad02571SSaso Kiselkov * our headroom for the next scanning cycle by this percentage multiple. 1087aad02571SSaso Kiselkov */ 1088aad02571SSaso Kiselkov #define L2ARC_HEADROOM_BOOST 200 10895a98e54bSBrendan Gregg - Sun Microsystems #define L2ARC_FEED_SECS 1 /* caching interval secs */ 10905a98e54bSBrendan Gregg - Sun Microsystems #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 1091fa94a07fSbrendan 1092fa94a07fSbrendan #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 1093fa94a07fSbrendan #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 1094fa94a07fSbrendan 1095f7170741SWill Andrews /* L2ARC Performance Tunables */ 1096fa94a07fSbrendan uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 10973a737e0dSbrendan uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1098fa94a07fSbrendan uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1099aad02571SSaso Kiselkov uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1100fa94a07fSbrendan uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 11015a98e54bSBrendan Gregg - Sun Microsystems uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1102fa94a07fSbrendan boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 11035a98e54bSBrendan Gregg - Sun Microsystems boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 11045a98e54bSBrendan Gregg - Sun Microsystems boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1105fa94a07fSbrendan 1106fa94a07fSbrendan /* 1107fa94a07fSbrendan * L2ARC Internals 1108fa94a07fSbrendan */ 110989c86e32SChris Williamson struct l2arc_dev { 1110fa94a07fSbrendan vdev_t *l2ad_vdev; /* vdev */ 1111fa94a07fSbrendan spa_t *l2ad_spa; /* spa */ 1112fa94a07fSbrendan uint64_t l2ad_hand; /* next write location */ 1113fa94a07fSbrendan uint64_t l2ad_start; /* first addr on device */ 1114fa94a07fSbrendan uint64_t l2ad_end; /* last addr on device */ 1115fa94a07fSbrendan boolean_t l2ad_first; /* first sweep through */ 11165a98e54bSBrendan Gregg - Sun Microsystems boolean_t l2ad_writing; /* currently writing */ 111789c86e32SChris Williamson kmutex_t l2ad_mtx; /* lock for buffer list */ 111889c86e32SChris Williamson list_t l2ad_buflist; /* buffer list */ 1119fa94a07fSbrendan list_node_t l2ad_node; /* device list node */ 1120a52fc310SPrakash Surya refcount_t l2ad_alloc; /* allocated bytes */ 112189c86e32SChris Williamson }; 1122fa94a07fSbrendan 1123fa94a07fSbrendan static list_t L2ARC_dev_list; /* device list */ 1124fa94a07fSbrendan static list_t *l2arc_dev_list; /* device list pointer */ 1125fa94a07fSbrendan static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1126fa94a07fSbrendan static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1127fa94a07fSbrendan static list_t L2ARC_free_on_write; /* free after write buf list */ 1128fa94a07fSbrendan static list_t *l2arc_free_on_write; /* free after write list ptr */ 1129fa94a07fSbrendan static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1130fa94a07fSbrendan static uint64_t l2arc_ndev; /* number of devices */ 1131fa94a07fSbrendan 1132fa94a07fSbrendan typedef struct l2arc_read_callback { 11335602294fSDan Kimmel arc_buf_hdr_t *l2rcb_hdr; /* read header */ 1134aad02571SSaso Kiselkov blkptr_t l2rcb_bp; /* original blkptr */ 11357802d7bfSMatthew Ahrens zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1136aad02571SSaso Kiselkov int l2rcb_flags; /* original flags */ 1137403a8da7SAndriy Gapon abd_t *l2rcb_abd; /* temporary buffer */ 1138fa94a07fSbrendan } l2arc_read_callback_t; 1139fa94a07fSbrendan 1140fa94a07fSbrendan typedef struct l2arc_write_callback { 1141fa94a07fSbrendan l2arc_dev_t *l2wcb_dev; /* device info */ 1142fa94a07fSbrendan arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1143fa94a07fSbrendan } l2arc_write_callback_t; 1144fa94a07fSbrendan 1145fa94a07fSbrendan typedef struct l2arc_data_free { 1146fa94a07fSbrendan /* protected by l2arc_free_on_write_mtx */ 1147770499e1SDan Kimmel abd_t *l2df_abd; 1148fa94a07fSbrendan size_t l2df_size; 1149dcbf3bd6SGeorge Wilson arc_buf_contents_t l2df_type; 1150fa94a07fSbrendan list_node_t l2df_list_node; 1151fa94a07fSbrendan } l2arc_data_free_t; 1152fa94a07fSbrendan 1153fa94a07fSbrendan static kmutex_t l2arc_feed_thr_lock; 1154fa94a07fSbrendan static kcondvar_t l2arc_feed_thr_cv; 1155fa94a07fSbrendan static uint8_t l2arc_thread_exit; 1156fa94a07fSbrendan 1157770499e1SDan Kimmel static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *); 1158dcbf3bd6SGeorge Wilson static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); 1159770499e1SDan Kimmel static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *); 1160770499e1SDan Kimmel static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *); 1161dcbf3bd6SGeorge Wilson static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); 1162770499e1SDan Kimmel static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag); 1163770499e1SDan Kimmel static void arc_hdr_free_pabd(arc_buf_hdr_t *); 1164770499e1SDan Kimmel static void arc_hdr_alloc_pabd(arc_buf_hdr_t *); 11657adb730bSGeorge Wilson static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1166244781f1SPrakash Surya static boolean_t arc_is_overflowing(); 11677adb730bSGeorge Wilson static void arc_buf_watch(arc_buf_t *); 11687adb730bSGeorge Wilson 116989c86e32SChris Williamson static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 117089c86e32SChris Williamson static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1171dcbf3bd6SGeorge Wilson static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1172dcbf3bd6SGeorge Wilson static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 117389c86e32SChris Williamson 11747adb730bSGeorge Wilson static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 11757adb730bSGeorge Wilson static void l2arc_read_done(zio_t *); 1176fa94a07fSbrendan 11773a2d8a1bSPaul Dagnelie 11783a2d8a1bSPaul Dagnelie /* 11793a2d8a1bSPaul Dagnelie * We use Cityhash for this. It's fast, and has good hash properties without 11803a2d8a1bSPaul Dagnelie * requiring any large static buffers. 11813a2d8a1bSPaul Dagnelie */ 1182fa9e4066Sahrens static uint64_t 1183ac05c741SMark Maybee buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1184fa9e4066Sahrens { 11853a2d8a1bSPaul Dagnelie return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth)); 1186fa9e4066Sahrens } 1187fa9e4066Sahrens 1188dcbf3bd6SGeorge Wilson #define HDR_EMPTY(hdr) \ 1189dcbf3bd6SGeorge Wilson ((hdr)->b_dva.dva_word[0] == 0 && \ 1190dcbf3bd6SGeorge Wilson (hdr)->b_dva.dva_word[1] == 0) 1191fa9e4066Sahrens 1192dcbf3bd6SGeorge Wilson #define HDR_EQUAL(spa, dva, birth, hdr) \ 1193dcbf3bd6SGeorge Wilson ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1194dcbf3bd6SGeorge Wilson ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1195dcbf3bd6SGeorge Wilson ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) 1196fa9e4066Sahrens 11973f9d6ad7SLin Ling static void 11983f9d6ad7SLin Ling buf_discard_identity(arc_buf_hdr_t *hdr) 11993f9d6ad7SLin Ling { 12003f9d6ad7SLin Ling hdr->b_dva.dva_word[0] = 0; 12013f9d6ad7SLin Ling hdr->b_dva.dva_word[1] = 0; 12023f9d6ad7SLin Ling hdr->b_birth = 0; 12033f9d6ad7SLin Ling } 12043f9d6ad7SLin Ling 1205fa9e4066Sahrens static arc_buf_hdr_t * 12065d7b4d43SMatthew Ahrens buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1207fa9e4066Sahrens { 12085d7b4d43SMatthew Ahrens const dva_t *dva = BP_IDENTITY(bp); 12095d7b4d43SMatthew Ahrens uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1210fa9e4066Sahrens uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1211fa9e4066Sahrens kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 12127adb730bSGeorge Wilson arc_buf_hdr_t *hdr; 1213fa9e4066Sahrens 1214fa9e4066Sahrens mutex_enter(hash_lock); 12157adb730bSGeorge Wilson for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 12167adb730bSGeorge Wilson hdr = hdr->b_hash_next) { 1217dcbf3bd6SGeorge Wilson if (HDR_EQUAL(spa, dva, birth, hdr)) { 1218fa9e4066Sahrens *lockp = hash_lock; 12197adb730bSGeorge Wilson return (hdr); 1220fa9e4066Sahrens } 1221fa9e4066Sahrens } 1222fa9e4066Sahrens mutex_exit(hash_lock); 1223fa9e4066Sahrens *lockp = NULL; 1224fa9e4066Sahrens return (NULL); 1225fa9e4066Sahrens } 1226fa9e4066Sahrens 1227fa9e4066Sahrens /* 1228fa9e4066Sahrens * Insert an entry into the hash table. If there is already an element 1229fa9e4066Sahrens * equal to elem in the hash table, then the already existing element 1230fa9e4066Sahrens * will be returned and the new element will not be inserted. 1231fa9e4066Sahrens * Otherwise returns NULL. 123289c86e32SChris Williamson * If lockp == NULL, the caller is assumed to already hold the hash lock. 1233fa9e4066Sahrens */ 1234fa9e4066Sahrens static arc_buf_hdr_t * 12357adb730bSGeorge Wilson buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1236fa9e4066Sahrens { 12377adb730bSGeorge Wilson uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1238fa9e4066Sahrens kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 12397adb730bSGeorge Wilson arc_buf_hdr_t *fhdr; 124044cb6abcSbmc uint32_t i; 1241fa9e4066Sahrens 12427adb730bSGeorge Wilson ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 12437adb730bSGeorge Wilson ASSERT(hdr->b_birth != 0); 12447adb730bSGeorge Wilson ASSERT(!HDR_IN_HASH_TABLE(hdr)); 124589c86e32SChris Williamson 124689c86e32SChris Williamson if (lockp != NULL) { 124789c86e32SChris Williamson *lockp = hash_lock; 124889c86e32SChris Williamson mutex_enter(hash_lock); 124989c86e32SChris Williamson } else { 125089c86e32SChris Williamson ASSERT(MUTEX_HELD(hash_lock)); 125189c86e32SChris Williamson } 125289c86e32SChris Williamson 12537adb730bSGeorge Wilson for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 12547adb730bSGeorge Wilson fhdr = fhdr->b_hash_next, i++) { 1255dcbf3bd6SGeorge Wilson if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 12567adb730bSGeorge Wilson return (fhdr); 1257fa9e4066Sahrens } 1258fa9e4066Sahrens 12597adb730bSGeorge Wilson hdr->b_hash_next = buf_hash_table.ht_table[idx]; 12607adb730bSGeorge Wilson buf_hash_table.ht_table[idx] = hdr; 1261dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1262fa9e4066Sahrens 1263fa9e4066Sahrens /* collect some hash table performance data */ 1264fa9e4066Sahrens if (i > 0) { 126544cb6abcSbmc ARCSTAT_BUMP(arcstat_hash_collisions); 1266fa9e4066Sahrens if (i == 1) 126744cb6abcSbmc ARCSTAT_BUMP(arcstat_hash_chains); 126844cb6abcSbmc 126944cb6abcSbmc ARCSTAT_MAX(arcstat_hash_chain_max, i); 1270fa9e4066Sahrens } 127144cb6abcSbmc 127244cb6abcSbmc ARCSTAT_BUMP(arcstat_hash_elements); 127344cb6abcSbmc ARCSTAT_MAXSTAT(arcstat_hash_elements); 1274fa9e4066Sahrens 1275fa9e4066Sahrens return (NULL); 1276fa9e4066Sahrens } 1277fa9e4066Sahrens 1278fa9e4066Sahrens static void 12797adb730bSGeorge Wilson buf_hash_remove(arc_buf_hdr_t *hdr) 1280fa9e4066Sahrens { 12817adb730bSGeorge Wilson arc_buf_hdr_t *fhdr, **hdrp; 12827adb730bSGeorge Wilson uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1283fa9e4066Sahrens 1284fa9e4066Sahrens ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 12857adb730bSGeorge Wilson ASSERT(HDR_IN_HASH_TABLE(hdr)); 1286fa9e4066Sahrens 12877adb730bSGeorge Wilson hdrp = &buf_hash_table.ht_table[idx]; 12887adb730bSGeorge Wilson while ((fhdr = *hdrp) != hdr) { 1289dcbf3bd6SGeorge Wilson ASSERT3P(fhdr, !=, NULL); 12907adb730bSGeorge Wilson hdrp = &fhdr->b_hash_next; 1291fa9e4066Sahrens } 12927adb730bSGeorge Wilson *hdrp = hdr->b_hash_next; 12937adb730bSGeorge Wilson hdr->b_hash_next = NULL; 1294dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1295fa9e4066Sahrens 1296fa9e4066Sahrens /* collect some hash table performance data */ 129744cb6abcSbmc ARCSTAT_BUMPDOWN(arcstat_hash_elements); 129844cb6abcSbmc 1299fa9e4066Sahrens if (buf_hash_table.ht_table[idx] && 1300fa9e4066Sahrens buf_hash_table.ht_table[idx]->b_hash_next == NULL) 130144cb6abcSbmc ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1302fa9e4066Sahrens } 1303fa9e4066Sahrens 1304fa9e4066Sahrens /* 1305fa9e4066Sahrens * Global data structures and functions for the buf kmem cache. 1306fa9e4066Sahrens */ 130789c86e32SChris Williamson static kmem_cache_t *hdr_full_cache; 130889c86e32SChris Williamson static kmem_cache_t *hdr_l2only_cache; 1309fa9e4066Sahrens static kmem_cache_t *buf_cache; 1310fa9e4066Sahrens 1311fa9e4066Sahrens static void 1312fa9e4066Sahrens buf_fini(void) 1313fa9e4066Sahrens { 1314fa9e4066Sahrens int i; 1315fa9e4066Sahrens 1316fa9e4066Sahrens kmem_free(buf_hash_table.ht_table, 1317fa9e4066Sahrens (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1318fa9e4066Sahrens for (i = 0; i < BUF_LOCKS; i++) 1319fa9e4066Sahrens mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 132089c86e32SChris Williamson kmem_cache_destroy(hdr_full_cache); 132189c86e32SChris Williamson kmem_cache_destroy(hdr_l2only_cache); 1322fa9e4066Sahrens kmem_cache_destroy(buf_cache); 1323fa9e4066Sahrens } 1324fa9e4066Sahrens 1325fa9e4066Sahrens /* 1326fa9e4066Sahrens * Constructor callback - called when the cache is empty 1327fa9e4066Sahrens * and a new buf is requested. 1328fa9e4066Sahrens */ 1329fa9e4066Sahrens /* ARGSUSED */ 1330fa9e4066Sahrens static int 133189c86e32SChris Williamson hdr_full_cons(void *vbuf, void *unused, int kmflag) 1332fa9e4066Sahrens { 13337adb730bSGeorge Wilson arc_buf_hdr_t *hdr = vbuf; 1334fa9e4066Sahrens 133589c86e32SChris Williamson bzero(hdr, HDR_FULL_SIZE); 133689c86e32SChris Williamson cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 133789c86e32SChris Williamson refcount_create(&hdr->b_l1hdr.b_refcnt); 133889c86e32SChris Williamson mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1339244781f1SPrakash Surya multilist_link_init(&hdr->b_l1hdr.b_arc_node); 134089c86e32SChris Williamson arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 134189c86e32SChris Williamson 134289c86e32SChris Williamson return (0); 134389c86e32SChris Williamson } 134489c86e32SChris Williamson 134589c86e32SChris Williamson /* ARGSUSED */ 134689c86e32SChris Williamson static int 134789c86e32SChris Williamson hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 134889c86e32SChris Williamson { 134989c86e32SChris Williamson arc_buf_hdr_t *hdr = vbuf; 135089c86e32SChris Williamson 135189c86e32SChris Williamson bzero(hdr, HDR_L2ONLY_SIZE); 135289c86e32SChris Williamson arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1353fa94a07fSbrendan 1354fa9e4066Sahrens return (0); 1355fa9e4066Sahrens } 1356fa9e4066Sahrens 13576f83844dSMark Maybee /* ARGSUSED */ 13586f83844dSMark Maybee static int 13596f83844dSMark Maybee buf_cons(void *vbuf, void *unused, int kmflag) 13606f83844dSMark Maybee { 13616f83844dSMark Maybee arc_buf_t *buf = vbuf; 13626f83844dSMark Maybee 13636f83844dSMark Maybee bzero(buf, sizeof (arc_buf_t)); 13643f9d6ad7SLin Ling mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 13655a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 13665a98e54bSBrendan Gregg - Sun Microsystems 13676f83844dSMark Maybee return (0); 13686f83844dSMark Maybee } 13696f83844dSMark Maybee 1370fa9e4066Sahrens /* 1371fa9e4066Sahrens * Destructor callback - called when a cached buf is 1372fa9e4066Sahrens * no longer required. 1373fa9e4066Sahrens */ 1374fa9e4066Sahrens /* ARGSUSED */ 1375fa9e4066Sahrens static void 137689c86e32SChris Williamson hdr_full_dest(void *vbuf, void *unused) 137789c86e32SChris Williamson { 137889c86e32SChris Williamson arc_buf_hdr_t *hdr = vbuf; 137989c86e32SChris Williamson 1380dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 138189c86e32SChris Williamson cv_destroy(&hdr->b_l1hdr.b_cv); 138289c86e32SChris Williamson refcount_destroy(&hdr->b_l1hdr.b_refcnt); 138389c86e32SChris Williamson mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1384244781f1SPrakash Surya ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 138589c86e32SChris Williamson arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 138689c86e32SChris Williamson } 138789c86e32SChris Williamson 138889c86e32SChris Williamson /* ARGSUSED */ 138989c86e32SChris Williamson static void 139089c86e32SChris Williamson hdr_l2only_dest(void *vbuf, void *unused) 1391fa9e4066Sahrens { 13927adb730bSGeorge Wilson arc_buf_hdr_t *hdr = vbuf; 1393fa9e4066Sahrens 1394dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 139589c86e32SChris Williamson arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1396fa9e4066Sahrens } 1397fa9e4066Sahrens 13986f83844dSMark Maybee /* ARGSUSED */ 13996f83844dSMark Maybee static void 14006f83844dSMark Maybee buf_dest(void *vbuf, void *unused) 14016f83844dSMark Maybee { 14026f83844dSMark Maybee arc_buf_t *buf = vbuf; 14036f83844dSMark Maybee 14043f9d6ad7SLin Ling mutex_destroy(&buf->b_evict_lock); 14055a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 14066f83844dSMark Maybee } 14076f83844dSMark Maybee 1408fa9e4066Sahrens /* 1409fa9e4066Sahrens * Reclaim callback -- invoked when memory is low. 1410fa9e4066Sahrens */ 1411fa9e4066Sahrens /* ARGSUSED */ 1412fa9e4066Sahrens static void 1413fa9e4066Sahrens hdr_recl(void *unused) 1414fa9e4066Sahrens { 1415fa9e4066Sahrens dprintf("hdr_recl called\n"); 141649e3519aSmaybee /* 141749e3519aSmaybee * umem calls the reclaim func when we destroy the buf cache, 141849e3519aSmaybee * which is after we do arc_fini(). 141949e3519aSmaybee */ 1420de753e34SBrad Lewis if (arc_initialized) 1421de753e34SBrad Lewis zthr_wakeup(arc_reap_zthr); 1422fa9e4066Sahrens } 1423fa9e4066Sahrens 1424fa9e4066Sahrens static void 1425fa9e4066Sahrens buf_init(void) 1426fa9e4066Sahrens { 1427fa9e4066Sahrens uint64_t *ct; 1428ea8dc4b6Seschrock uint64_t hsize = 1ULL << 12; 1429fa9e4066Sahrens int i, j; 1430fa9e4066Sahrens 1431fa9e4066Sahrens /* 1432fa9e4066Sahrens * The hash table is big enough to fill all of physical memory 143363e911b6SMatthew Ahrens * with an average block size of zfs_arc_average_blocksize (default 8K). 143463e911b6SMatthew Ahrens * By default, the table will take up 143563e911b6SMatthew Ahrens * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1436fa9e4066Sahrens */ 143763e911b6SMatthew Ahrens while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE) 1438fa9e4066Sahrens hsize <<= 1; 1439ea8dc4b6Seschrock retry: 1440fa9e4066Sahrens buf_hash_table.ht_mask = hsize - 1; 1441ea8dc4b6Seschrock buf_hash_table.ht_table = 1442ea8dc4b6Seschrock kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1443ea8dc4b6Seschrock if (buf_hash_table.ht_table == NULL) { 1444ea8dc4b6Seschrock ASSERT(hsize > (1ULL << 8)); 1445ea8dc4b6Seschrock hsize >>= 1; 1446ea8dc4b6Seschrock goto retry; 1447ea8dc4b6Seschrock } 1448fa9e4066Sahrens 144989c86e32SChris Williamson hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 145089c86e32SChris Williamson 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 145189c86e32SChris Williamson hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 145289c86e32SChris Williamson HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 145389c86e32SChris Williamson NULL, NULL, 0); 1454fa9e4066Sahrens buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 14556f83844dSMark Maybee 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1456fa9e4066Sahrens 1457fa9e4066Sahrens for (i = 0; i < 256; i++) 1458fa9e4066Sahrens for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1459fa9e4066Sahrens *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1460fa9e4066Sahrens 1461fa9e4066Sahrens for (i = 0; i < BUF_LOCKS; i++) { 1462fa9e4066Sahrens mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1463fa9e4066Sahrens NULL, MUTEX_DEFAULT, NULL); 1464fa9e4066Sahrens } 1465fa9e4066Sahrens } 1466fa9e4066Sahrens 14675602294fSDan Kimmel /* 14685602294fSDan Kimmel * This is the size that the buf occupies in memory. If the buf is compressed, 14695602294fSDan Kimmel * it will correspond to the compressed size. You should use this method of 14705602294fSDan Kimmel * getting the buf size unless you explicitly need the logical size. 14715602294fSDan Kimmel */ 14725602294fSDan Kimmel int32_t 14735602294fSDan Kimmel arc_buf_size(arc_buf_t *buf) 14745602294fSDan Kimmel { 14755602294fSDan Kimmel return (ARC_BUF_COMPRESSED(buf) ? 14765602294fSDan Kimmel HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr)); 14775602294fSDan Kimmel } 14785602294fSDan Kimmel 14795602294fSDan Kimmel int32_t 14805602294fSDan Kimmel arc_buf_lsize(arc_buf_t *buf) 14815602294fSDan Kimmel { 14825602294fSDan Kimmel return (HDR_GET_LSIZE(buf->b_hdr)); 14835602294fSDan Kimmel } 14845602294fSDan Kimmel 14855602294fSDan Kimmel enum zio_compress 14865602294fSDan Kimmel arc_get_compression(arc_buf_t *buf) 14875602294fSDan Kimmel { 14885602294fSDan Kimmel return (ARC_BUF_COMPRESSED(buf) ? 14895602294fSDan Kimmel HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF); 14905602294fSDan Kimmel } 14915602294fSDan Kimmel 1492dcbf3bd6SGeorge Wilson #define ARC_MINTIME (hz>>4) /* 62 ms */ 1493244781f1SPrakash Surya 1494dcbf3bd6SGeorge Wilson static inline boolean_t 1495dcbf3bd6SGeorge Wilson arc_buf_is_shared(arc_buf_t *buf) 1496dcbf3bd6SGeorge Wilson { 1497dcbf3bd6SGeorge Wilson boolean_t shared = (buf->b_data != NULL && 1498770499e1SDan Kimmel buf->b_hdr->b_l1hdr.b_pabd != NULL && 1499770499e1SDan Kimmel abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) && 1500770499e1SDan Kimmel buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd)); 1501dcbf3bd6SGeorge Wilson IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); 15025602294fSDan Kimmel IMPLY(shared, ARC_BUF_SHARED(buf)); 15035602294fSDan Kimmel IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf)); 15045602294fSDan Kimmel 15055602294fSDan Kimmel /* 15065602294fSDan Kimmel * It would be nice to assert arc_can_share() too, but the "hdr isn't 15075602294fSDan Kimmel * already being shared" requirement prevents us from doing that. 15085602294fSDan Kimmel */ 15095602294fSDan Kimmel 1510dcbf3bd6SGeorge Wilson return (shared); 1511dcbf3bd6SGeorge Wilson } 1512c546f36aSArne Jansen 15135602294fSDan Kimmel /* 15145602294fSDan Kimmel * Free the checksum associated with this header. If there is no checksum, this 15155602294fSDan Kimmel * is a no-op. 15165602294fSDan Kimmel */ 1517dcbf3bd6SGeorge Wilson static inline void 1518dcbf3bd6SGeorge Wilson arc_cksum_free(arc_buf_hdr_t *hdr) 1519dcbf3bd6SGeorge Wilson { 1520dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1521dcbf3bd6SGeorge Wilson mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1522dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1523dcbf3bd6SGeorge Wilson kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); 1524dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_freeze_cksum = NULL; 152589c86e32SChris Williamson } 1526dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 152789c86e32SChris Williamson } 152889c86e32SChris Williamson 15295602294fSDan Kimmel /* 15305602294fSDan Kimmel * Return true iff at least one of the bufs on hdr is not compressed. 15315602294fSDan Kimmel */ 15325602294fSDan Kimmel static boolean_t 15335602294fSDan Kimmel arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr) 15345602294fSDan Kimmel { 15355602294fSDan Kimmel for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) { 15365602294fSDan Kimmel if (!ARC_BUF_COMPRESSED(b)) { 15375602294fSDan Kimmel return (B_TRUE); 15385602294fSDan Kimmel } 15395602294fSDan Kimmel } 15405602294fSDan Kimmel return (B_FALSE); 15415602294fSDan Kimmel } 15425602294fSDan Kimmel 15435602294fSDan Kimmel /* 15445602294fSDan Kimmel * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data 15455602294fSDan Kimmel * matches the checksum that is stored in the hdr. If there is no checksum, 15465602294fSDan Kimmel * or if the buf is compressed, this is a no-op. 15475602294fSDan Kimmel */ 15486b4acc8bSahrens static void 15496b4acc8bSahrens arc_cksum_verify(arc_buf_t *buf) 15506b4acc8bSahrens { 1551dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 15526b4acc8bSahrens zio_cksum_t zc; 15536b4acc8bSahrens 1554cc60fd72Sahrens if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 15556b4acc8bSahrens return; 15566b4acc8bSahrens 15575602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 15585602294fSDan Kimmel ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 15595602294fSDan Kimmel arc_hdr_has_uncompressed_buf(hdr)); 15605602294fSDan Kimmel return; 15615602294fSDan Kimmel } 15625602294fSDan Kimmel 1563dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1564dcbf3bd6SGeorge Wilson 1565dcbf3bd6SGeorge Wilson mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1566dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { 1567dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 15686b4acc8bSahrens return; 15696b4acc8bSahrens } 15705602294fSDan Kimmel 15715602294fSDan Kimmel fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc); 1572dcbf3bd6SGeorge Wilson if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) 15736b4acc8bSahrens panic("buffer modified while frozen!"); 1574dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 15756b4acc8bSahrens } 15766b4acc8bSahrens 1577dcbf3bd6SGeorge Wilson static boolean_t 1578dcbf3bd6SGeorge Wilson arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) 1579fa94a07fSbrendan { 1580dcbf3bd6SGeorge Wilson enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp); 1581dcbf3bd6SGeorge Wilson boolean_t valid_cksum; 1582fa94a07fSbrendan 1583dcbf3bd6SGeorge Wilson ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); 1584dcbf3bd6SGeorge Wilson VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); 1585dcbf3bd6SGeorge Wilson 1586dcbf3bd6SGeorge Wilson /* 1587dcbf3bd6SGeorge Wilson * We rely on the blkptr's checksum to determine if the block 1588dcbf3bd6SGeorge Wilson * is valid or not. When compressed arc is enabled, the l2arc 1589dcbf3bd6SGeorge Wilson * writes the block to the l2arc just as it appears in the pool. 1590dcbf3bd6SGeorge Wilson * This allows us to use the blkptr's checksum to validate the 1591dcbf3bd6SGeorge Wilson * data that we just read off of the l2arc without having to store 1592dcbf3bd6SGeorge Wilson * a separate checksum in the arc_buf_hdr_t. However, if compressed 1593dcbf3bd6SGeorge Wilson * arc is disabled, then the data written to the l2arc is always 1594dcbf3bd6SGeorge Wilson * uncompressed and won't match the block as it exists in the main 1595dcbf3bd6SGeorge Wilson * pool. When this is the case, we must first compress it if it is 1596dcbf3bd6SGeorge Wilson * compressed on the main pool before we can validate the checksum. 1597dcbf3bd6SGeorge Wilson */ 1598dcbf3bd6SGeorge Wilson if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) { 1599dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1600dcbf3bd6SGeorge Wilson uint64_t lsize = HDR_GET_LSIZE(hdr); 1601dcbf3bd6SGeorge Wilson uint64_t csize; 1602dcbf3bd6SGeorge Wilson 160301a059eeSRoman Strashkin abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE); 160401a059eeSRoman Strashkin csize = zio_compress_data(compress, zio->io_abd, 160501a059eeSRoman Strashkin abd_to_buf(cdata), lsize); 1606770499e1SDan Kimmel 1607dcbf3bd6SGeorge Wilson ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr)); 1608dcbf3bd6SGeorge Wilson if (csize < HDR_GET_PSIZE(hdr)) { 1609dcbf3bd6SGeorge Wilson /* 1610dcbf3bd6SGeorge Wilson * Compressed blocks are always a multiple of the 1611dcbf3bd6SGeorge Wilson * smallest ashift in the pool. Ideally, we would 1612dcbf3bd6SGeorge Wilson * like to round up the csize to the next 1613dcbf3bd6SGeorge Wilson * spa_min_ashift but that value may have changed 1614dcbf3bd6SGeorge Wilson * since the block was last written. Instead, 1615dcbf3bd6SGeorge Wilson * we rely on the fact that the hdr's psize 1616dcbf3bd6SGeorge Wilson * was set to the psize of the block when it was 1617dcbf3bd6SGeorge Wilson * last written. We set the csize to that value 1618dcbf3bd6SGeorge Wilson * and zero out any part that should not contain 1619dcbf3bd6SGeorge Wilson * data. 1620dcbf3bd6SGeorge Wilson */ 162101a059eeSRoman Strashkin abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize); 1622dcbf3bd6SGeorge Wilson csize = HDR_GET_PSIZE(hdr); 1623dcbf3bd6SGeorge Wilson } 162401a059eeSRoman Strashkin zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL); 1625dcbf3bd6SGeorge Wilson } 1626fa94a07fSbrendan 1627dcbf3bd6SGeorge Wilson /* 1628dcbf3bd6SGeorge Wilson * Block pointers always store the checksum for the logical data. 1629dcbf3bd6SGeorge Wilson * If the block pointer has the gang bit set, then the checksum 1630dcbf3bd6SGeorge Wilson * it represents is for the reconstituted data and not for an 1631dcbf3bd6SGeorge Wilson * individual gang member. The zio pipeline, however, must be able to 1632dcbf3bd6SGeorge Wilson * determine the checksum of each of the gang constituents so it 1633dcbf3bd6SGeorge Wilson * treats the checksum comparison differently than what we need 1634dcbf3bd6SGeorge Wilson * for l2arc blocks. This prevents us from using the 1635dcbf3bd6SGeorge Wilson * zio_checksum_error() interface directly. Instead we must call the 1636dcbf3bd6SGeorge Wilson * zio_checksum_error_impl() so that we can ensure the checksum is 1637dcbf3bd6SGeorge Wilson * generated using the correct checksum algorithm and accounts for the 1638dcbf3bd6SGeorge Wilson * logical I/O size and not just a gang fragment. 1639dcbf3bd6SGeorge Wilson */ 1640dcbf3bd6SGeorge Wilson valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp, 1641770499e1SDan Kimmel BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size, 1642dcbf3bd6SGeorge Wilson zio->io_offset, NULL) == 0); 1643dcbf3bd6SGeorge Wilson zio_pop_transforms(zio); 1644dcbf3bd6SGeorge Wilson return (valid_cksum); 1645fa94a07fSbrendan } 1646fa94a07fSbrendan 16475602294fSDan Kimmel /* 16485602294fSDan Kimmel * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a 16495602294fSDan Kimmel * checksum and attaches it to the buf's hdr so that we can ensure that the buf 16505602294fSDan Kimmel * isn't modified later on. If buf is compressed or there is already a checksum 16515602294fSDan Kimmel * on the hdr, this is a no-op (we only checksum uncompressed bufs). 16525602294fSDan Kimmel */ 16536b4acc8bSahrens static void 1654dcbf3bd6SGeorge Wilson arc_cksum_compute(arc_buf_t *buf) 16556b4acc8bSahrens { 1656dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 1657dcbf3bd6SGeorge Wilson 1658dcbf3bd6SGeorge Wilson if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 16596b4acc8bSahrens return; 16606b4acc8bSahrens 1661dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 16625602294fSDan Kimmel 166389c86e32SChris Williamson mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1664dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 16655602294fSDan Kimmel ASSERT(arc_hdr_has_uncompressed_buf(hdr)); 16665602294fSDan Kimmel mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 16675602294fSDan Kimmel return; 16685602294fSDan Kimmel } else if (ARC_BUF_COMPRESSED(buf)) { 1669dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 16706b4acc8bSahrens return; 16716b4acc8bSahrens } 16725602294fSDan Kimmel 16735602294fSDan Kimmel ASSERT(!ARC_BUF_COMPRESSED(buf)); 1674dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), 1675dcbf3bd6SGeorge Wilson KM_SLEEP); 16765602294fSDan Kimmel fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, 1677dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_freeze_cksum); 1678dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1679cd1c8b85SMatthew Ahrens arc_buf_watch(buf); 1680cd1c8b85SMatthew Ahrens } 1681cd1c8b85SMatthew Ahrens 1682cd1c8b85SMatthew Ahrens #ifndef _KERNEL 1683cd1c8b85SMatthew Ahrens typedef struct procctl { 1684cd1c8b85SMatthew Ahrens long cmd; 1685cd1c8b85SMatthew Ahrens prwatch_t prwatch; 1686cd1c8b85SMatthew Ahrens } procctl_t; 1687cd1c8b85SMatthew Ahrens #endif 1688cd1c8b85SMatthew Ahrens 1689cd1c8b85SMatthew Ahrens /* ARGSUSED */ 1690cd1c8b85SMatthew Ahrens static void 1691cd1c8b85SMatthew Ahrens arc_buf_unwatch(arc_buf_t *buf) 1692cd1c8b85SMatthew Ahrens { 1693cd1c8b85SMatthew Ahrens #ifndef _KERNEL 1694cd1c8b85SMatthew Ahrens if (arc_watch) { 1695cd1c8b85SMatthew Ahrens int result; 1696cd1c8b85SMatthew Ahrens procctl_t ctl; 1697cd1c8b85SMatthew Ahrens ctl.cmd = PCWATCH; 1698cd1c8b85SMatthew Ahrens ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1699cd1c8b85SMatthew Ahrens ctl.prwatch.pr_size = 0; 1700cd1c8b85SMatthew Ahrens ctl.prwatch.pr_wflags = 0; 1701cd1c8b85SMatthew Ahrens result = write(arc_procfd, &ctl, sizeof (ctl)); 1702cd1c8b85SMatthew Ahrens ASSERT3U(result, ==, sizeof (ctl)); 1703cd1c8b85SMatthew Ahrens } 1704cd1c8b85SMatthew Ahrens #endif 1705cd1c8b85SMatthew Ahrens } 1706cd1c8b85SMatthew Ahrens 1707cd1c8b85SMatthew Ahrens /* ARGSUSED */ 1708cd1c8b85SMatthew Ahrens static void 1709cd1c8b85SMatthew Ahrens arc_buf_watch(arc_buf_t *buf) 1710cd1c8b85SMatthew Ahrens { 1711cd1c8b85SMatthew Ahrens #ifndef _KERNEL 1712cd1c8b85SMatthew Ahrens if (arc_watch) { 1713cd1c8b85SMatthew Ahrens int result; 1714cd1c8b85SMatthew Ahrens procctl_t ctl; 1715cd1c8b85SMatthew Ahrens ctl.cmd = PCWATCH; 1716cd1c8b85SMatthew Ahrens ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 17175602294fSDan Kimmel ctl.prwatch.pr_size = arc_buf_size(buf); 1718cd1c8b85SMatthew Ahrens ctl.prwatch.pr_wflags = WA_WRITE; 1719cd1c8b85SMatthew Ahrens result = write(arc_procfd, &ctl, sizeof (ctl)); 1720cd1c8b85SMatthew Ahrens ASSERT3U(result, ==, sizeof (ctl)); 1721cd1c8b85SMatthew Ahrens } 1722cd1c8b85SMatthew Ahrens #endif 17236b4acc8bSahrens } 17246b4acc8bSahrens 172589c86e32SChris Williamson static arc_buf_contents_t 172689c86e32SChris Williamson arc_buf_type(arc_buf_hdr_t *hdr) 172789c86e32SChris Williamson { 1728dcbf3bd6SGeorge Wilson arc_buf_contents_t type; 172989c86e32SChris Williamson if (HDR_ISTYPE_METADATA(hdr)) { 1730dcbf3bd6SGeorge Wilson type = ARC_BUFC_METADATA; 173189c86e32SChris Williamson } else { 1732dcbf3bd6SGeorge Wilson type = ARC_BUFC_DATA; 173389c86e32SChris Williamson } 1734dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 1735dcbf3bd6SGeorge Wilson return (type); 173689c86e32SChris Williamson } 173789c86e32SChris Williamson 17385602294fSDan Kimmel boolean_t 17395602294fSDan Kimmel arc_is_metadata(arc_buf_t *buf) 17405602294fSDan Kimmel { 17415602294fSDan Kimmel return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0); 17425602294fSDan Kimmel } 17435602294fSDan Kimmel 174489c86e32SChris Williamson static uint32_t 174589c86e32SChris Williamson arc_bufc_to_flags(arc_buf_contents_t type) 174689c86e32SChris Williamson { 174789c86e32SChris Williamson switch (type) { 174889c86e32SChris Williamson case ARC_BUFC_DATA: 174989c86e32SChris Williamson /* metadata field is 0 if buffer contains normal data */ 175089c86e32SChris Williamson return (0); 175189c86e32SChris Williamson case ARC_BUFC_METADATA: 175289c86e32SChris Williamson return (ARC_FLAG_BUFC_METADATA); 175389c86e32SChris Williamson default: 175489c86e32SChris Williamson break; 175589c86e32SChris Williamson } 175689c86e32SChris Williamson panic("undefined ARC buffer type!"); 175789c86e32SChris Williamson return ((uint32_t)-1); 175889c86e32SChris Williamson } 175989c86e32SChris Williamson 17606b4acc8bSahrens void 17616b4acc8bSahrens arc_buf_thaw(arc_buf_t *buf) 17626b4acc8bSahrens { 1763dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 1764dcbf3bd6SGeorge Wilson 17655602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 17665602294fSDan Kimmel ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 17675602294fSDan Kimmel 17685602294fSDan Kimmel arc_cksum_verify(buf); 17695602294fSDan Kimmel 17705602294fSDan Kimmel /* 17715602294fSDan Kimmel * Compressed buffers do not manipulate the b_freeze_cksum or 17725602294fSDan Kimmel * allocate b_thawed. 17735602294fSDan Kimmel */ 17745602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 17755602294fSDan Kimmel ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 17765602294fSDan Kimmel arc_hdr_has_uncompressed_buf(hdr)); 17775602294fSDan Kimmel return; 1778fa94a07fSbrendan } 17796b4acc8bSahrens 1780dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1781dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 17823f9d6ad7SLin Ling 1783dcbf3bd6SGeorge Wilson mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 178489c86e32SChris Williamson #ifdef ZFS_DEBUG 17853f9d6ad7SLin Ling if (zfs_flags & ZFS_DEBUG_MODIFY) { 1786dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_thawed != NULL) 1787dcbf3bd6SGeorge Wilson kmem_free(hdr->b_l1hdr.b_thawed, 1); 1788dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 17893f9d6ad7SLin Ling } 179089c86e32SChris Williamson #endif 17913f9d6ad7SLin Ling 1792dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1793cd1c8b85SMatthew Ahrens 1794cd1c8b85SMatthew Ahrens arc_buf_unwatch(buf); 17956b4acc8bSahrens } 17966b4acc8bSahrens 17976b4acc8bSahrens void 17986b4acc8bSahrens arc_buf_freeze(arc_buf_t *buf) 17996b4acc8bSahrens { 1800dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 18013f9d6ad7SLin Ling kmutex_t *hash_lock; 18023f9d6ad7SLin Ling 1803cc60fd72Sahrens if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1804cc60fd72Sahrens return; 1805cc60fd72Sahrens 18065602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 18075602294fSDan Kimmel ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 18085602294fSDan Kimmel arc_hdr_has_uncompressed_buf(hdr)); 18095602294fSDan Kimmel return; 18105602294fSDan Kimmel } 18115602294fSDan Kimmel 1812dcbf3bd6SGeorge Wilson hash_lock = HDR_LOCK(hdr); 18133f9d6ad7SLin Ling mutex_enter(hash_lock); 18143f9d6ad7SLin Ling 1815dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1816dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL || 1817dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_state == arc_anon); 1818dcbf3bd6SGeorge Wilson arc_cksum_compute(buf); 18193f9d6ad7SLin Ling mutex_exit(hash_lock); 18206b4acc8bSahrens } 18216b4acc8bSahrens 1822dcbf3bd6SGeorge Wilson /* 1823dcbf3bd6SGeorge Wilson * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, 1824dcbf3bd6SGeorge Wilson * the following functions should be used to ensure that the flags are 1825dcbf3bd6SGeorge Wilson * updated in a thread-safe way. When manipulating the flags either 1826dcbf3bd6SGeorge Wilson * the hash_lock must be held or the hdr must be undiscoverable. This 1827dcbf3bd6SGeorge Wilson * ensures that we're not racing with any other threads when updating 1828dcbf3bd6SGeorge Wilson * the flags. 1829dcbf3bd6SGeorge Wilson */ 1830dcbf3bd6SGeorge Wilson static inline void 1831dcbf3bd6SGeorge Wilson arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1832dcbf3bd6SGeorge Wilson { 1833dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1834dcbf3bd6SGeorge Wilson hdr->b_flags |= flags; 1835dcbf3bd6SGeorge Wilson } 1836dcbf3bd6SGeorge Wilson 1837dcbf3bd6SGeorge Wilson static inline void 1838dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1839dcbf3bd6SGeorge Wilson { 1840dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1841dcbf3bd6SGeorge Wilson hdr->b_flags &= ~flags; 1842dcbf3bd6SGeorge Wilson } 1843dcbf3bd6SGeorge Wilson 1844dcbf3bd6SGeorge Wilson /* 1845dcbf3bd6SGeorge Wilson * Setting the compression bits in the arc_buf_hdr_t's b_flags is 1846dcbf3bd6SGeorge Wilson * done in a special way since we have to clear and set bits 1847dcbf3bd6SGeorge Wilson * at the same time. Consumers that wish to set the compression bits 1848dcbf3bd6SGeorge Wilson * must use this function to ensure that the flags are updated in 1849dcbf3bd6SGeorge Wilson * thread-safe manner. 1850dcbf3bd6SGeorge Wilson */ 1851fa9e4066Sahrens static void 1852dcbf3bd6SGeorge Wilson arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) 1853fa9e4066Sahrens { 1854dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1855fa9e4066Sahrens 1856dcbf3bd6SGeorge Wilson /* 1857dcbf3bd6SGeorge Wilson * Holes and embedded blocks will always have a psize = 0 so 1858dcbf3bd6SGeorge Wilson * we ignore the compression of the blkptr and set the 1859dcbf3bd6SGeorge Wilson * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF. 1860dcbf3bd6SGeorge Wilson * Holes and embedded blocks remain anonymous so we don't 1861dcbf3bd6SGeorge Wilson * want to uncompress them. Mark them as uncompressed. 1862dcbf3bd6SGeorge Wilson */ 1863dcbf3bd6SGeorge Wilson if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { 1864dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1865dcbf3bd6SGeorge Wilson HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 1866dcbf3bd6SGeorge Wilson ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); 1867dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1868dcbf3bd6SGeorge Wilson } else { 1869dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1870dcbf3bd6SGeorge Wilson HDR_SET_COMPRESS(hdr, cmp); 1871dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); 1872dcbf3bd6SGeorge Wilson ASSERT(HDR_COMPRESSION_ENABLED(hdr)); 1873dcbf3bd6SGeorge Wilson } 1874dcbf3bd6SGeorge Wilson } 1875244781f1SPrakash Surya 18765602294fSDan Kimmel /* 18775602294fSDan Kimmel * Looks for another buf on the same hdr which has the data decompressed, copies 18785602294fSDan Kimmel * from it, and returns true. If no such buf exists, returns false. 18795602294fSDan Kimmel */ 18805602294fSDan Kimmel static boolean_t 18815602294fSDan Kimmel arc_buf_try_copy_decompressed_data(arc_buf_t *buf) 18825602294fSDan Kimmel { 18835602294fSDan Kimmel arc_buf_hdr_t *hdr = buf->b_hdr; 18845602294fSDan Kimmel boolean_t copied = B_FALSE; 18855602294fSDan Kimmel 18865602294fSDan Kimmel ASSERT(HDR_HAS_L1HDR(hdr)); 18875602294fSDan Kimmel ASSERT3P(buf->b_data, !=, NULL); 18885602294fSDan Kimmel ASSERT(!ARC_BUF_COMPRESSED(buf)); 18895602294fSDan Kimmel 18905602294fSDan Kimmel for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL; 18915602294fSDan Kimmel from = from->b_next) { 18925602294fSDan Kimmel /* can't use our own data buffer */ 18935602294fSDan Kimmel if (from == buf) { 18945602294fSDan Kimmel continue; 18955602294fSDan Kimmel } 18965602294fSDan Kimmel 18975602294fSDan Kimmel if (!ARC_BUF_COMPRESSED(from)) { 18985602294fSDan Kimmel bcopy(from->b_data, buf->b_data, arc_buf_size(buf)); 18995602294fSDan Kimmel copied = B_TRUE; 19005602294fSDan Kimmel break; 19015602294fSDan Kimmel } 19025602294fSDan Kimmel } 19035602294fSDan Kimmel 19045602294fSDan Kimmel /* 19055602294fSDan Kimmel * There were no decompressed bufs, so there should not be a 19065602294fSDan Kimmel * checksum on the hdr either. 19075602294fSDan Kimmel */ 19085602294fSDan Kimmel EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL); 19095602294fSDan Kimmel 19105602294fSDan Kimmel return (copied); 19115602294fSDan Kimmel } 19125602294fSDan Kimmel 19135602294fSDan Kimmel /* 19145602294fSDan Kimmel * Given a buf that has a data buffer attached to it, this function will 19155602294fSDan Kimmel * efficiently fill the buf with data of the specified compression setting from 19165602294fSDan Kimmel * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr 19175602294fSDan Kimmel * are already sharing a data buf, no copy is performed. 19185602294fSDan Kimmel * 19195602294fSDan Kimmel * If the buf is marked as compressed but uncompressed data was requested, this 19205602294fSDan Kimmel * will allocate a new data buffer for the buf, remove that flag, and fill the 19215602294fSDan Kimmel * buf with uncompressed data. You can't request a compressed buf on a hdr with 19225602294fSDan Kimmel * uncompressed data, and (since we haven't added support for it yet) if you 19235602294fSDan Kimmel * want compressed data your buf must already be marked as compressed and have 19245602294fSDan Kimmel * the correct-sized data buffer. 19255602294fSDan Kimmel */ 1926dcbf3bd6SGeorge Wilson static int 19275602294fSDan Kimmel arc_buf_fill(arc_buf_t *buf, boolean_t compressed) 1928dcbf3bd6SGeorge Wilson { 1929dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 19305602294fSDan Kimmel boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 1931dcbf3bd6SGeorge Wilson dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; 193289c86e32SChris Williamson 19335602294fSDan Kimmel ASSERT3P(buf->b_data, !=, NULL); 19345602294fSDan Kimmel IMPLY(compressed, hdr_compressed); 19355602294fSDan Kimmel IMPLY(compressed, ARC_BUF_COMPRESSED(buf)); 19365602294fSDan Kimmel 19375602294fSDan Kimmel if (hdr_compressed == compressed) { 19385602294fSDan Kimmel if (!arc_buf_is_shared(buf)) { 1939770499e1SDan Kimmel abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd, 19405602294fSDan Kimmel arc_buf_size(buf)); 19415602294fSDan Kimmel } 1942dcbf3bd6SGeorge Wilson } else { 19435602294fSDan Kimmel ASSERT(hdr_compressed); 19445602294fSDan Kimmel ASSERT(!compressed); 1945dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); 19465602294fSDan Kimmel 19475602294fSDan Kimmel /* 19485602294fSDan Kimmel * If the buf is sharing its data with the hdr, unlink it and 19495602294fSDan Kimmel * allocate a new data buffer for the buf. 19505602294fSDan Kimmel */ 19515602294fSDan Kimmel if (arc_buf_is_shared(buf)) { 19525602294fSDan Kimmel ASSERT(ARC_BUF_COMPRESSED(buf)); 19535602294fSDan Kimmel 19545602294fSDan Kimmel /* We need to give the buf it's own b_data */ 19555602294fSDan Kimmel buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 19565602294fSDan Kimmel buf->b_data = 19575602294fSDan Kimmel arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 19585602294fSDan Kimmel arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 19595602294fSDan Kimmel 19605602294fSDan Kimmel /* Previously overhead was 0; just add new overhead */ 19615602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 19625602294fSDan Kimmel } else if (ARC_BUF_COMPRESSED(buf)) { 19635602294fSDan Kimmel /* We need to reallocate the buf's b_data */ 19645602294fSDan Kimmel arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr), 19655602294fSDan Kimmel buf); 19665602294fSDan Kimmel buf->b_data = 19675602294fSDan Kimmel arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 19685602294fSDan Kimmel 19695602294fSDan Kimmel /* We increased the size of b_data; update overhead */ 19705602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, 19715602294fSDan Kimmel HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr)); 19725602294fSDan Kimmel } 19735602294fSDan Kimmel 19745602294fSDan Kimmel /* 19755602294fSDan Kimmel * Regardless of the buf's previous compression settings, it 19765602294fSDan Kimmel * should not be compressed at the end of this function. 19775602294fSDan Kimmel */ 19785602294fSDan Kimmel buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; 19795602294fSDan Kimmel 19805602294fSDan Kimmel /* 19815602294fSDan Kimmel * Try copying the data from another buf which already has a 19825602294fSDan Kimmel * decompressed version. If that's not possible, it's time to 19835602294fSDan Kimmel * bite the bullet and decompress the data from the hdr. 19845602294fSDan Kimmel */ 19855602294fSDan Kimmel if (arc_buf_try_copy_decompressed_data(buf)) { 19865602294fSDan Kimmel /* Skip byteswapping and checksumming (already done) */ 19875602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL); 19885602294fSDan Kimmel return (0); 19895602294fSDan Kimmel } else { 19905602294fSDan Kimmel int error = zio_decompress_data(HDR_GET_COMPRESS(hdr), 1991770499e1SDan Kimmel hdr->b_l1hdr.b_pabd, buf->b_data, 19925602294fSDan Kimmel HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 19935602294fSDan Kimmel 19945602294fSDan Kimmel /* 19955602294fSDan Kimmel * Absent hardware errors or software bugs, this should 19965602294fSDan Kimmel * be impossible, but log it anyway so we can debug it. 19975602294fSDan Kimmel */ 19985602294fSDan Kimmel if (error != 0) { 19995602294fSDan Kimmel zfs_dbgmsg( 20005602294fSDan Kimmel "hdr %p, compress %d, psize %d, lsize %d", 20015602294fSDan Kimmel hdr, HDR_GET_COMPRESS(hdr), 20025602294fSDan Kimmel HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 20035602294fSDan Kimmel return (SET_ERROR(EIO)); 20045602294fSDan Kimmel } 2005ea8dc4b6Seschrock } 2006fa9e4066Sahrens } 20075602294fSDan Kimmel 20085602294fSDan Kimmel /* Byteswap the buf's data if necessary */ 2009dcbf3bd6SGeorge Wilson if (bswap != DMU_BSWAP_NUMFUNCS) { 2010dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(hdr)); 2011dcbf3bd6SGeorge Wilson ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); 2012dcbf3bd6SGeorge Wilson dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); 2013dcbf3bd6SGeorge Wilson } 20145602294fSDan Kimmel 20155602294fSDan Kimmel /* Compute the hdr's checksum if necessary */ 2016dcbf3bd6SGeorge Wilson arc_cksum_compute(buf); 20175602294fSDan Kimmel 2018dcbf3bd6SGeorge Wilson return (0); 2019fa9e4066Sahrens } 2020fa9e4066Sahrens 20215602294fSDan Kimmel int 20225602294fSDan Kimmel arc_decompress(arc_buf_t *buf) 20235602294fSDan Kimmel { 20245602294fSDan Kimmel return (arc_buf_fill(buf, B_FALSE)); 20255602294fSDan Kimmel } 20265602294fSDan Kimmel 2027dcbf3bd6SGeorge Wilson /* 2028770499e1SDan Kimmel * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t. 2029dcbf3bd6SGeorge Wilson */ 2030dcbf3bd6SGeorge Wilson static uint64_t 2031dcbf3bd6SGeorge Wilson arc_hdr_size(arc_buf_hdr_t *hdr) 2032fa9e4066Sahrens { 2033dcbf3bd6SGeorge Wilson uint64_t size; 2034fa9e4066Sahrens 2035dcbf3bd6SGeorge Wilson if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 2036dcbf3bd6SGeorge Wilson HDR_GET_PSIZE(hdr) > 0) { 2037dcbf3bd6SGeorge Wilson size = HDR_GET_PSIZE(hdr); 2038dcbf3bd6SGeorge Wilson } else { 2039dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); 2040dcbf3bd6SGeorge Wilson size = HDR_GET_LSIZE(hdr); 2041dcbf3bd6SGeorge Wilson } 2042dcbf3bd6SGeorge Wilson return (size); 2043dcbf3bd6SGeorge Wilson } 2044fa9e4066Sahrens 2045dcbf3bd6SGeorge Wilson /* 2046dcbf3bd6SGeorge Wilson * Increment the amount of evictable space in the arc_state_t's refcount. 2047dcbf3bd6SGeorge Wilson * We account for the space used by the hdr and the arc buf individually 2048dcbf3bd6SGeorge Wilson * so that we can add and remove them from the refcount individually. 2049dcbf3bd6SGeorge Wilson */ 2050dcbf3bd6SGeorge Wilson static void 2051dcbf3bd6SGeorge Wilson arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) 2052dcbf3bd6SGeorge Wilson { 2053dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 2054244781f1SPrakash Surya 2055dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 20560e8c6158Smaybee 2057dcbf3bd6SGeorge Wilson if (GHOST_STATE(state)) { 2058dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 2059dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2060770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 20615602294fSDan Kimmel (void) refcount_add_many(&state->arcs_esize[type], 20625602294fSDan Kimmel HDR_GET_LSIZE(hdr), hdr); 2063dcbf3bd6SGeorge Wilson return; 2064dcbf3bd6SGeorge Wilson } 2065dcbf3bd6SGeorge Wilson 2066dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(state)); 2067770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 2068dcbf3bd6SGeorge Wilson (void) refcount_add_many(&state->arcs_esize[type], 2069dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2070dcbf3bd6SGeorge Wilson } 2071dcbf3bd6SGeorge Wilson for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2072dcbf3bd6SGeorge Wilson buf = buf->b_next) { 20735602294fSDan Kimmel if (arc_buf_is_shared(buf)) 2074dcbf3bd6SGeorge Wilson continue; 20755602294fSDan Kimmel (void) refcount_add_many(&state->arcs_esize[type], 20765602294fSDan Kimmel arc_buf_size(buf), buf); 2077fa9e4066Sahrens } 2078fa9e4066Sahrens } 2079fa9e4066Sahrens 2080fa9e4066Sahrens /* 2081dcbf3bd6SGeorge Wilson * Decrement the amount of evictable space in the arc_state_t's refcount. 2082dcbf3bd6SGeorge Wilson * We account for the space used by the hdr and the arc buf individually 2083dcbf3bd6SGeorge Wilson * so that we can add and remove them from the refcount individually. 2084fa9e4066Sahrens */ 2085fa9e4066Sahrens static void 20865602294fSDan Kimmel arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) 2087fa9e4066Sahrens { 2088dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 2089dcbf3bd6SGeorge Wilson 2090dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2091dcbf3bd6SGeorge Wilson 2092dcbf3bd6SGeorge Wilson if (GHOST_STATE(state)) { 2093dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 2094dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2095770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2096dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 20975602294fSDan Kimmel HDR_GET_LSIZE(hdr), hdr); 2098dcbf3bd6SGeorge Wilson return; 2099dcbf3bd6SGeorge Wilson } 2100dcbf3bd6SGeorge Wilson 2101dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(state)); 2102770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 2103dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 2104dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2105dcbf3bd6SGeorge Wilson } 2106dcbf3bd6SGeorge Wilson for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2107dcbf3bd6SGeorge Wilson buf = buf->b_next) { 21085602294fSDan Kimmel if (arc_buf_is_shared(buf)) 2109dcbf3bd6SGeorge Wilson continue; 2110dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 21115602294fSDan Kimmel arc_buf_size(buf), buf); 2112dcbf3bd6SGeorge Wilson } 2113dcbf3bd6SGeorge Wilson } 2114dcbf3bd6SGeorge Wilson 2115dcbf3bd6SGeorge Wilson /* 2116dcbf3bd6SGeorge Wilson * Add a reference to this hdr indicating that someone is actively 2117dcbf3bd6SGeorge Wilson * referencing that memory. When the refcount transitions from 0 to 1, 2118dcbf3bd6SGeorge Wilson * we remove it from the respective arc_state_t list to indicate that 2119dcbf3bd6SGeorge Wilson * it is not evictable. 2120dcbf3bd6SGeorge Wilson */ 2121dcbf3bd6SGeorge Wilson static void 2122dcbf3bd6SGeorge Wilson add_reference(arc_buf_hdr_t *hdr, void *tag) 2123dcbf3bd6SGeorge Wilson { 2124dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2125dcbf3bd6SGeorge Wilson if (!MUTEX_HELD(HDR_LOCK(hdr))) { 2126dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_state == arc_anon); 2127dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2128dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2129dcbf3bd6SGeorge Wilson } 2130dcbf3bd6SGeorge Wilson 2131dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2132dcbf3bd6SGeorge Wilson 2133dcbf3bd6SGeorge Wilson if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 2134dcbf3bd6SGeorge Wilson (state != arc_anon)) { 2135dcbf3bd6SGeorge Wilson /* We don't use the L2-only state list. */ 2136dcbf3bd6SGeorge Wilson if (state != arc_l2c_only) { 213794c2d0ebSMatthew Ahrens multilist_remove(state->arcs_list[arc_buf_type(hdr)], 2138dcbf3bd6SGeorge Wilson hdr); 21395602294fSDan Kimmel arc_evictable_space_decrement(hdr, state); 2140dcbf3bd6SGeorge Wilson } 2141dcbf3bd6SGeorge Wilson /* remove the prefetch flag if we get a reference */ 2142dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 2143dcbf3bd6SGeorge Wilson } 2144dcbf3bd6SGeorge Wilson } 2145dcbf3bd6SGeorge Wilson 2146dcbf3bd6SGeorge Wilson /* 2147dcbf3bd6SGeorge Wilson * Remove a reference from this hdr. When the reference transitions from 2148dcbf3bd6SGeorge Wilson * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's 2149dcbf3bd6SGeorge Wilson * list making it eligible for eviction. 2150dcbf3bd6SGeorge Wilson */ 2151dcbf3bd6SGeorge Wilson static int 2152dcbf3bd6SGeorge Wilson remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 2153dcbf3bd6SGeorge Wilson { 2154dcbf3bd6SGeorge Wilson int cnt; 2155dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2156dcbf3bd6SGeorge Wilson 2157dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2158dcbf3bd6SGeorge Wilson ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 2159dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(state)); 2160dcbf3bd6SGeorge Wilson 2161dcbf3bd6SGeorge Wilson /* 2162dcbf3bd6SGeorge Wilson * arc_l2c_only counts as a ghost state so we don't need to explicitly 2163dcbf3bd6SGeorge Wilson * check to prevent usage of the arc_l2c_only list. 2164dcbf3bd6SGeorge Wilson */ 2165dcbf3bd6SGeorge Wilson if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 2166dcbf3bd6SGeorge Wilson (state != arc_anon)) { 216794c2d0ebSMatthew Ahrens multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr); 2168dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 2169dcbf3bd6SGeorge Wilson arc_evictable_space_increment(hdr, state); 2170dcbf3bd6SGeorge Wilson } 2171dcbf3bd6SGeorge Wilson return (cnt); 2172dcbf3bd6SGeorge Wilson } 2173dcbf3bd6SGeorge Wilson 2174dcbf3bd6SGeorge Wilson /* 2175dcbf3bd6SGeorge Wilson * Move the supplied buffer to the indicated state. The hash lock 2176dcbf3bd6SGeorge Wilson * for the buffer must be held by the caller. 2177dcbf3bd6SGeorge Wilson */ 2178dcbf3bd6SGeorge Wilson static void 2179dcbf3bd6SGeorge Wilson arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 2180dcbf3bd6SGeorge Wilson kmutex_t *hash_lock) 2181dcbf3bd6SGeorge Wilson { 2182dcbf3bd6SGeorge Wilson arc_state_t *old_state; 2183dcbf3bd6SGeorge Wilson int64_t refcnt; 2184dcbf3bd6SGeorge Wilson uint32_t bufcnt; 2185dcbf3bd6SGeorge Wilson boolean_t update_old, update_new; 2186dcbf3bd6SGeorge Wilson arc_buf_contents_t buftype = arc_buf_type(hdr); 218789c86e32SChris Williamson 218889c86e32SChris Williamson /* 218989c86e32SChris Williamson * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 219089c86e32SChris Williamson * in arc_read() when bringing a buffer out of the L2ARC. However, the 219189c86e32SChris Williamson * L1 hdr doesn't always exist when we change state to arc_anon before 219289c86e32SChris Williamson * destroying a header, in which case reallocating to add the L1 hdr is 219389c86e32SChris Williamson * pointless. 219489c86e32SChris Williamson */ 219589c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) { 219689c86e32SChris Williamson old_state = hdr->b_l1hdr.b_state; 219789c86e32SChris Williamson refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 2198dcbf3bd6SGeorge Wilson bufcnt = hdr->b_l1hdr.b_bufcnt; 2199770499e1SDan Kimmel update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL); 220089c86e32SChris Williamson } else { 220189c86e32SChris Williamson old_state = arc_l2c_only; 220289c86e32SChris Williamson refcnt = 0; 2203dcbf3bd6SGeorge Wilson bufcnt = 0; 2204dcbf3bd6SGeorge Wilson update_old = B_FALSE; 220589c86e32SChris Williamson } 2206dcbf3bd6SGeorge Wilson update_new = update_old; 2207fa9e4066Sahrens 2208fa9e4066Sahrens ASSERT(MUTEX_HELD(hash_lock)); 220969962b56SMatthew Ahrens ASSERT3P(new_state, !=, old_state); 2210dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); 2211dcbf3bd6SGeorge Wilson ASSERT(old_state != arc_anon || bufcnt <= 1); 2212fa9e4066Sahrens 2213fa9e4066Sahrens /* 2214fa9e4066Sahrens * If this buffer is evictable, transfer it from the 2215fa9e4066Sahrens * old state list to the new state list. 2216fa9e4066Sahrens */ 2217ea8dc4b6Seschrock if (refcnt == 0) { 221889c86e32SChris Williamson if (old_state != arc_anon && old_state != arc_l2c_only) { 221989c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 222094c2d0ebSMatthew Ahrens multilist_remove(old_state->arcs_list[buftype], hdr); 2221ea8dc4b6Seschrock 2222dcbf3bd6SGeorge Wilson if (GHOST_STATE(old_state)) { 2223dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 2224dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2225dcbf3bd6SGeorge Wilson update_old = B_TRUE; 2226ea8dc4b6Seschrock } 22275602294fSDan Kimmel arc_evictable_space_decrement(hdr, old_state); 2228fa9e4066Sahrens } 222989c86e32SChris Williamson if (new_state != arc_anon && new_state != arc_l2c_only) { 2230fa9e4066Sahrens 223189c86e32SChris Williamson /* 223289c86e32SChris Williamson * An L1 header always exists here, since if we're 223389c86e32SChris Williamson * moving to some L1-cached state (i.e. not l2c_only or 223489c86e32SChris Williamson * anonymous), we realloc the header to add an L1hdr 223589c86e32SChris Williamson * beforehand. 223689c86e32SChris Williamson */ 223789c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 223894c2d0ebSMatthew Ahrens multilist_insert(new_state->arcs_list[buftype], hdr); 2239ea8dc4b6Seschrock 2240ea8dc4b6Seschrock if (GHOST_STATE(new_state)) { 2241dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 2242dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2243dcbf3bd6SGeorge Wilson update_new = B_TRUE; 2244ea8dc4b6Seschrock } 2245dcbf3bd6SGeorge Wilson arc_evictable_space_increment(hdr, new_state); 2246fa9e4066Sahrens } 2247fa9e4066Sahrens } 2248fa9e4066Sahrens 2249dcbf3bd6SGeorge Wilson ASSERT(!HDR_EMPTY(hdr)); 22507adb730bSGeorge Wilson if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 22517adb730bSGeorge Wilson buf_hash_remove(hdr); 2252fa9e4066Sahrens 225389c86e32SChris Williamson /* adjust state sizes (ignore arc_l2c_only) */ 22542fd872a7SPrakash Surya 2255dcbf3bd6SGeorge Wilson if (update_new && new_state != arc_l2c_only) { 22562fd872a7SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 22572fd872a7SPrakash Surya if (GHOST_STATE(new_state)) { 2258dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 22592fd872a7SPrakash Surya 22602fd872a7SPrakash Surya /* 2261dcbf3bd6SGeorge Wilson * When moving a header to a ghost state, we first 22622fd872a7SPrakash Surya * remove all arc buffers. Thus, we'll have a 2263dcbf3bd6SGeorge Wilson * bufcnt of zero, and no arc buffer to use for 22642fd872a7SPrakash Surya * the reference. As a result, we use the arc 22652fd872a7SPrakash Surya * header pointer for the reference. 22662fd872a7SPrakash Surya */ 22672fd872a7SPrakash Surya (void) refcount_add_many(&new_state->arcs_size, 2268dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr), hdr); 2269770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 22702fd872a7SPrakash Surya } else { 2271dcbf3bd6SGeorge Wilson uint32_t buffers = 0; 22722fd872a7SPrakash Surya 22732fd872a7SPrakash Surya /* 22742fd872a7SPrakash Surya * Each individual buffer holds a unique reference, 22752fd872a7SPrakash Surya * thus we must remove each of these references one 22762fd872a7SPrakash Surya * at a time. 22772fd872a7SPrakash Surya */ 22782fd872a7SPrakash Surya for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 22792fd872a7SPrakash Surya buf = buf->b_next) { 2280dcbf3bd6SGeorge Wilson ASSERT3U(bufcnt, !=, 0); 2281dcbf3bd6SGeorge Wilson buffers++; 2282dcbf3bd6SGeorge Wilson 2283dcbf3bd6SGeorge Wilson /* 2284dcbf3bd6SGeorge Wilson * When the arc_buf_t is sharing the data 2285dcbf3bd6SGeorge Wilson * block with the hdr, the owner of the 2286dcbf3bd6SGeorge Wilson * reference belongs to the hdr. Only 2287dcbf3bd6SGeorge Wilson * add to the refcount if the arc_buf_t is 2288dcbf3bd6SGeorge Wilson * not shared. 2289dcbf3bd6SGeorge Wilson */ 22905602294fSDan Kimmel if (arc_buf_is_shared(buf)) 2291dcbf3bd6SGeorge Wilson continue; 2292dcbf3bd6SGeorge Wilson 22932fd872a7SPrakash Surya (void) refcount_add_many(&new_state->arcs_size, 22945602294fSDan Kimmel arc_buf_size(buf), buf); 2295dcbf3bd6SGeorge Wilson } 2296dcbf3bd6SGeorge Wilson ASSERT3U(bufcnt, ==, buffers); 2297dcbf3bd6SGeorge Wilson 2298770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 2299dcbf3bd6SGeorge Wilson (void) refcount_add_many(&new_state->arcs_size, 2300dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2301dcbf3bd6SGeorge Wilson } else { 2302dcbf3bd6SGeorge Wilson ASSERT(GHOST_STATE(old_state)); 23032fd872a7SPrakash Surya } 23042fd872a7SPrakash Surya } 23052fd872a7SPrakash Surya } 23062fd872a7SPrakash Surya 2307dcbf3bd6SGeorge Wilson if (update_old && old_state != arc_l2c_only) { 23082fd872a7SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 23092fd872a7SPrakash Surya if (GHOST_STATE(old_state)) { 2310dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 2311770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2312dcbf3bd6SGeorge Wilson 23132fd872a7SPrakash Surya /* 23142fd872a7SPrakash Surya * When moving a header off of a ghost state, 2315dcbf3bd6SGeorge Wilson * the header will not contain any arc buffers. 2316dcbf3bd6SGeorge Wilson * We use the arc header pointer for the reference 2317dcbf3bd6SGeorge Wilson * which is exactly what we did when we put the 2318dcbf3bd6SGeorge Wilson * header on the ghost state. 23192fd872a7SPrakash Surya */ 23202fd872a7SPrakash Surya 23212fd872a7SPrakash Surya (void) refcount_remove_many(&old_state->arcs_size, 2322dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr), hdr); 23232fd872a7SPrakash Surya } else { 2324dcbf3bd6SGeorge Wilson uint32_t buffers = 0; 23252fd872a7SPrakash Surya 23262fd872a7SPrakash Surya /* 23272fd872a7SPrakash Surya * Each individual buffer holds a unique reference, 23282fd872a7SPrakash Surya * thus we must remove each of these references one 23292fd872a7SPrakash Surya * at a time. 23302fd872a7SPrakash Surya */ 23312fd872a7SPrakash Surya for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 23322fd872a7SPrakash Surya buf = buf->b_next) { 23335602294fSDan Kimmel ASSERT3U(bufcnt, !=, 0); 2334dcbf3bd6SGeorge Wilson buffers++; 2335dcbf3bd6SGeorge Wilson 2336dcbf3bd6SGeorge Wilson /* 2337dcbf3bd6SGeorge Wilson * When the arc_buf_t is sharing the data 2338dcbf3bd6SGeorge Wilson * block with the hdr, the owner of the 2339dcbf3bd6SGeorge Wilson * reference belongs to the hdr. Only 2340dcbf3bd6SGeorge Wilson * add to the refcount if the arc_buf_t is 2341dcbf3bd6SGeorge Wilson * not shared. 2342dcbf3bd6SGeorge Wilson */ 23435602294fSDan Kimmel if (arc_buf_is_shared(buf)) 2344dcbf3bd6SGeorge Wilson continue; 2345dcbf3bd6SGeorge Wilson 23462fd872a7SPrakash Surya (void) refcount_remove_many( 23475602294fSDan Kimmel &old_state->arcs_size, arc_buf_size(buf), 2348dcbf3bd6SGeorge Wilson buf); 23492fd872a7SPrakash Surya } 2350dcbf3bd6SGeorge Wilson ASSERT3U(bufcnt, ==, buffers); 2351770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2352dcbf3bd6SGeorge Wilson (void) refcount_remove_many( 2353dcbf3bd6SGeorge Wilson &old_state->arcs_size, arc_hdr_size(hdr), hdr); 23542fd872a7SPrakash Surya } 2355fa9e4066Sahrens } 23562fd872a7SPrakash Surya 235789c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) 235889c86e32SChris Williamson hdr->b_l1hdr.b_state = new_state; 2359fa94a07fSbrendan 236089c86e32SChris Williamson /* 236189c86e32SChris Williamson * L2 headers should never be on the L2 state list since they don't 236289c86e32SChris Williamson * have L1 headers allocated. 236389c86e32SChris Williamson */ 236494c2d0ebSMatthew Ahrens ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 236594c2d0ebSMatthew Ahrens multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 2366fa9e4066Sahrens } 2367fa9e4066Sahrens 23680e8c6158Smaybee void 23695a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(uint64_t space, arc_space_type_t type) 23700e8c6158Smaybee { 23715a98e54bSBrendan Gregg - Sun Microsystems ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 23725a98e54bSBrendan Gregg - Sun Microsystems 23735a98e54bSBrendan Gregg - Sun Microsystems switch (type) { 23745a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_DATA: 23753a2d8a1bSPaul Dagnelie aggsum_add(&astat_data_size, space); 23765a98e54bSBrendan Gregg - Sun Microsystems break; 23774076b1bfSPrakash Surya case ARC_SPACE_META: 23783a2d8a1bSPaul Dagnelie aggsum_add(&astat_metadata_size, space); 23794076b1bfSPrakash Surya break; 23805a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_OTHER: 23813a2d8a1bSPaul Dagnelie aggsum_add(&astat_other_size, space); 23825a98e54bSBrendan Gregg - Sun Microsystems break; 23835a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_HDRS: 23843a2d8a1bSPaul Dagnelie aggsum_add(&astat_hdr_size, space); 23855a98e54bSBrendan Gregg - Sun Microsystems break; 23865a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_L2HDRS: 23873a2d8a1bSPaul Dagnelie aggsum_add(&astat_l2_hdr_size, space); 23885a98e54bSBrendan Gregg - Sun Microsystems break; 23895a98e54bSBrendan Gregg - Sun Microsystems } 23905a98e54bSBrendan Gregg - Sun Microsystems 23914076b1bfSPrakash Surya if (type != ARC_SPACE_DATA) 23923a2d8a1bSPaul Dagnelie aggsum_add(&arc_meta_used, space); 23934076b1bfSPrakash Surya 23943a2d8a1bSPaul Dagnelie aggsum_add(&arc_size, space); 23950e8c6158Smaybee } 23960e8c6158Smaybee 23970e8c6158Smaybee void 23985a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(uint64_t space, arc_space_type_t type) 23990e8c6158Smaybee { 24005a98e54bSBrendan Gregg - Sun Microsystems ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 24015a98e54bSBrendan Gregg - Sun Microsystems 24025a98e54bSBrendan Gregg - Sun Microsystems switch (type) { 24035a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_DATA: 24043a2d8a1bSPaul Dagnelie aggsum_add(&astat_data_size, -space); 24055a98e54bSBrendan Gregg - Sun Microsystems break; 24064076b1bfSPrakash Surya case ARC_SPACE_META: 24073a2d8a1bSPaul Dagnelie aggsum_add(&astat_metadata_size, -space); 24084076b1bfSPrakash Surya break; 24095a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_OTHER: 24103a2d8a1bSPaul Dagnelie aggsum_add(&astat_other_size, -space); 24115a98e54bSBrendan Gregg - Sun Microsystems break; 24125a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_HDRS: 24133a2d8a1bSPaul Dagnelie aggsum_add(&astat_hdr_size, -space); 24145a98e54bSBrendan Gregg - Sun Microsystems break; 24155a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_L2HDRS: 24163a2d8a1bSPaul Dagnelie aggsum_add(&astat_l2_hdr_size, -space); 24175a98e54bSBrendan Gregg - Sun Microsystems break; 24185a98e54bSBrendan Gregg - Sun Microsystems } 24195a98e54bSBrendan Gregg - Sun Microsystems 24204076b1bfSPrakash Surya if (type != ARC_SPACE_DATA) { 24213a2d8a1bSPaul Dagnelie ASSERT(aggsum_compare(&arc_meta_used, space) >= 0); 24223a2d8a1bSPaul Dagnelie /* 24233a2d8a1bSPaul Dagnelie * We use the upper bound here rather than the precise value 24243a2d8a1bSPaul Dagnelie * because the arc_meta_max value doesn't need to be 24253a2d8a1bSPaul Dagnelie * precise. It's only consumed by humans via arcstats. 24263a2d8a1bSPaul Dagnelie */ 24273a2d8a1bSPaul Dagnelie if (arc_meta_max < aggsum_upper_bound(&arc_meta_used)) 24283a2d8a1bSPaul Dagnelie arc_meta_max = aggsum_upper_bound(&arc_meta_used); 24293a2d8a1bSPaul Dagnelie aggsum_add(&arc_meta_used, -space); 24304076b1bfSPrakash Surya } 24314076b1bfSPrakash Surya 24323a2d8a1bSPaul Dagnelie ASSERT(aggsum_compare(&arc_size, space) >= 0); 24333a2d8a1bSPaul Dagnelie aggsum_add(&arc_size, -space); 24340e8c6158Smaybee } 24350e8c6158Smaybee 2436dcbf3bd6SGeorge Wilson /* 24375602294fSDan Kimmel * Given a hdr and a buf, returns whether that buf can share its b_data buffer 2438770499e1SDan Kimmel * with the hdr's b_pabd. 2439dcbf3bd6SGeorge Wilson */ 24405602294fSDan Kimmel static boolean_t 24415602294fSDan Kimmel arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf) 24425602294fSDan Kimmel { 24435602294fSDan Kimmel /* 24445602294fSDan Kimmel * The criteria for sharing a hdr's data are: 24455602294fSDan Kimmel * 1. the hdr's compression matches the buf's compression 24465602294fSDan Kimmel * 2. the hdr doesn't need to be byteswapped 24475602294fSDan Kimmel * 3. the hdr isn't already being shared 24485602294fSDan Kimmel * 4. the buf is either compressed or it is the last buf in the hdr list 24495602294fSDan Kimmel * 24505602294fSDan Kimmel * Criterion #4 maintains the invariant that shared uncompressed 24515602294fSDan Kimmel * bufs must be the final buf in the hdr's b_buf list. Reading this, you 24525602294fSDan Kimmel * might ask, "if a compressed buf is allocated first, won't that be the 24535602294fSDan Kimmel * last thing in the list?", but in that case it's impossible to create 24545602294fSDan Kimmel * a shared uncompressed buf anyway (because the hdr must be compressed 24555602294fSDan Kimmel * to have the compressed buf). You might also think that #3 is 24565602294fSDan Kimmel * sufficient to make this guarantee, however it's possible 24575602294fSDan Kimmel * (specifically in the rare L2ARC write race mentioned in 24585602294fSDan Kimmel * arc_buf_alloc_impl()) there will be an existing uncompressed buf that 24595602294fSDan Kimmel * is sharable, but wasn't at the time of its allocation. Rather than 24605602294fSDan Kimmel * allow a new shared uncompressed buf to be created and then shuffle 24615602294fSDan Kimmel * the list around to make it the last element, this simply disallows 24625602294fSDan Kimmel * sharing if the new buf isn't the first to be added. 24635602294fSDan Kimmel */ 24645602294fSDan Kimmel ASSERT3P(buf->b_hdr, ==, hdr); 24655602294fSDan Kimmel boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF; 24665602294fSDan Kimmel boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0; 24675602294fSDan Kimmel return (buf_compressed == hdr_compressed && 24685602294fSDan Kimmel hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && 24695602294fSDan Kimmel !HDR_SHARED_DATA(hdr) && 24705602294fSDan Kimmel (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf))); 24715602294fSDan Kimmel } 24725602294fSDan Kimmel 24735602294fSDan Kimmel /* 24745602294fSDan Kimmel * Allocate a buf for this hdr. If you care about the data that's in the hdr, 24755602294fSDan Kimmel * or if you want a compressed buffer, pass those flags in. Returns 0 if the 24765602294fSDan Kimmel * copy was made successfully, or an error code otherwise. 24775602294fSDan Kimmel */ 24785602294fSDan Kimmel static int 24795602294fSDan Kimmel arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed, 24805602294fSDan Kimmel boolean_t fill, arc_buf_t **ret) 2481fa9e4066Sahrens { 2482fa9e4066Sahrens arc_buf_t *buf; 2483fa9e4066Sahrens 2484dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2485dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2486dcbf3bd6SGeorge Wilson VERIFY(hdr->b_type == ARC_BUFC_DATA || 2487dcbf3bd6SGeorge Wilson hdr->b_type == ARC_BUFC_METADATA); 24885602294fSDan Kimmel ASSERT3P(ret, !=, NULL); 24895602294fSDan Kimmel ASSERT3P(*ret, ==, NULL); 2490dcbf3bd6SGeorge Wilson 24915602294fSDan Kimmel buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2492fa9e4066Sahrens buf->b_hdr = hdr; 249344eda4d7Smaybee buf->b_data = NULL; 24945602294fSDan Kimmel buf->b_next = hdr->b_l1hdr.b_buf; 24955602294fSDan Kimmel buf->b_flags = 0; 249689c86e32SChris Williamson 2497dcbf3bd6SGeorge Wilson add_reference(hdr, tag); 2498dcbf3bd6SGeorge Wilson 2499dcbf3bd6SGeorge Wilson /* 2500dcbf3bd6SGeorge Wilson * We're about to change the hdr's b_flags. We must either 2501dcbf3bd6SGeorge Wilson * hold the hash_lock or be undiscoverable. 2502dcbf3bd6SGeorge Wilson */ 2503dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2504dcbf3bd6SGeorge Wilson 2505dcbf3bd6SGeorge Wilson /* 25065602294fSDan Kimmel * Only honor requests for compressed bufs if the hdr is actually 25075602294fSDan Kimmel * compressed. 25085602294fSDan Kimmel */ 25095602294fSDan Kimmel if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 25105602294fSDan Kimmel buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; 25115602294fSDan Kimmel 25125602294fSDan Kimmel /* 25135602294fSDan Kimmel * If the hdr's data can be shared then we share the data buffer and 25145602294fSDan Kimmel * set the appropriate bit in the hdr's b_flags to indicate the hdr is 2515770499e1SDan Kimmel * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new 25165602294fSDan Kimmel * buffer to store the buf's data. 25175602294fSDan Kimmel * 2518770499e1SDan Kimmel * There are two additional restrictions here because we're sharing 2519770499e1SDan Kimmel * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be 2520770499e1SDan Kimmel * actively involved in an L2ARC write, because if this buf is used by 2521770499e1SDan Kimmel * an arc_write() then the hdr's data buffer will be released when the 25225602294fSDan Kimmel * write completes, even though the L2ARC write might still be using it. 2523770499e1SDan Kimmel * Second, the hdr's ABD must be linear so that the buf's user doesn't 2524770499e1SDan Kimmel * need to be ABD-aware. 2525dcbf3bd6SGeorge Wilson */ 2526770499e1SDan Kimmel boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) && 2527770499e1SDan Kimmel abd_is_linear(hdr->b_l1hdr.b_pabd); 25285602294fSDan Kimmel 25295602294fSDan Kimmel /* Set up b_data and sharing */ 25305602294fSDan Kimmel if (can_share) { 2531770499e1SDan Kimmel buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd); 25325602294fSDan Kimmel buf->b_flags |= ARC_BUF_FLAG_SHARED; 2533dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2534dcbf3bd6SGeorge Wilson } else { 25355602294fSDan Kimmel buf->b_data = 25365602294fSDan Kimmel arc_get_data_buf(hdr, arc_buf_size(buf), buf); 25375602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2538dcbf3bd6SGeorge Wilson } 2539dcbf3bd6SGeorge Wilson VERIFY3P(buf->b_data, !=, NULL); 254089c86e32SChris Williamson 254189c86e32SChris Williamson hdr->b_l1hdr.b_buf = buf; 2542dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt += 1; 254389c86e32SChris Williamson 25445602294fSDan Kimmel /* 25455602294fSDan Kimmel * If the user wants the data from the hdr, we need to either copy or 25465602294fSDan Kimmel * decompress the data. 25475602294fSDan Kimmel */ 25485602294fSDan Kimmel if (fill) { 25495602294fSDan Kimmel return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0)); 25505602294fSDan Kimmel } 2551dcbf3bd6SGeorge Wilson 25525602294fSDan Kimmel return (0); 25535602294fSDan Kimmel } 2554dcbf3bd6SGeorge Wilson 25555602294fSDan Kimmel static char *arc_onloan_tag = "onloan"; 2556fa9e4066Sahrens 25575602294fSDan Kimmel static inline void 25585602294fSDan Kimmel arc_loaned_bytes_update(int64_t delta) 25595602294fSDan Kimmel { 25605602294fSDan Kimmel atomic_add_64(&arc_loaned_bytes, delta); 2561dcbf3bd6SGeorge Wilson 25625602294fSDan Kimmel /* assert that it did not wrap around */ 25635602294fSDan Kimmel ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 2564fa9e4066Sahrens } 2565fa9e4066Sahrens 25662fdbea25SAleksandr Guzovskiy /* 25672fdbea25SAleksandr Guzovskiy * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 25682fdbea25SAleksandr Guzovskiy * flight data by arc_tempreserve_space() until they are "returned". Loaned 25692fdbea25SAleksandr Guzovskiy * buffers must be returned to the arc before they can be used by the DMU or 25702fdbea25SAleksandr Guzovskiy * freed. 25712fdbea25SAleksandr Guzovskiy */ 25722fdbea25SAleksandr Guzovskiy arc_buf_t * 25735602294fSDan Kimmel arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size) 25742fdbea25SAleksandr Guzovskiy { 25755602294fSDan Kimmel arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag, 25765602294fSDan Kimmel is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size); 25775602294fSDan Kimmel 25789be12bd7SAllan Jude arc_loaned_bytes_update(arc_buf_size(buf)); 25795602294fSDan Kimmel 25805602294fSDan Kimmel return (buf); 25815602294fSDan Kimmel } 25822fdbea25SAleksandr Guzovskiy 25835602294fSDan Kimmel arc_buf_t * 25845602294fSDan Kimmel arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize, 25855602294fSDan Kimmel enum zio_compress compression_type) 25865602294fSDan Kimmel { 25875602294fSDan Kimmel arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag, 25885602294fSDan Kimmel psize, lsize, compression_type); 25895602294fSDan Kimmel 25909be12bd7SAllan Jude arc_loaned_bytes_update(arc_buf_size(buf)); 25912fdbea25SAleksandr Guzovskiy 25922fdbea25SAleksandr Guzovskiy return (buf); 25932fdbea25SAleksandr Guzovskiy } 25942fdbea25SAleksandr Guzovskiy 25955602294fSDan Kimmel 25962fdbea25SAleksandr Guzovskiy /* 25972fdbea25SAleksandr Guzovskiy * Return a loaned arc buffer to the arc. 25982fdbea25SAleksandr Guzovskiy */ 25992fdbea25SAleksandr Guzovskiy void 26002fdbea25SAleksandr Guzovskiy arc_return_buf(arc_buf_t *buf, void *tag) 26012fdbea25SAleksandr Guzovskiy { 26022fdbea25SAleksandr Guzovskiy arc_buf_hdr_t *hdr = buf->b_hdr; 26032fdbea25SAleksandr Guzovskiy 2604dcbf3bd6SGeorge Wilson ASSERT3P(buf->b_data, !=, NULL); 260589c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 260689c86e32SChris Williamson (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 260789c86e32SChris Williamson (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 26082fdbea25SAleksandr Guzovskiy 26095602294fSDan Kimmel arc_loaned_bytes_update(-arc_buf_size(buf)); 26102fdbea25SAleksandr Guzovskiy } 26112fdbea25SAleksandr Guzovskiy 2612c242f9a0Schunli zhang - Sun Microsystems - Irvine United States /* Detach an arc_buf from a dbuf (tag) */ 2613c242f9a0Schunli zhang - Sun Microsystems - Irvine United States void 2614c242f9a0Schunli zhang - Sun Microsystems - Irvine United States arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2615c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 261689c86e32SChris Williamson arc_buf_hdr_t *hdr = buf->b_hdr; 2617c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 2618dcbf3bd6SGeorge Wilson ASSERT3P(buf->b_data, !=, NULL); 261989c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 262089c86e32SChris Williamson (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 262189c86e32SChris Williamson (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2622c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 26235602294fSDan Kimmel arc_loaned_bytes_update(arc_buf_size(buf)); 2624c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 2625c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 2626dcbf3bd6SGeorge Wilson static void 2627770499e1SDan Kimmel l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type) 2628ea8dc4b6Seschrock { 2629dcbf3bd6SGeorge Wilson l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); 2630ea8dc4b6Seschrock 2631770499e1SDan Kimmel df->l2df_abd = abd; 2632dcbf3bd6SGeorge Wilson df->l2df_size = size; 2633dcbf3bd6SGeorge Wilson df->l2df_type = type; 2634dcbf3bd6SGeorge Wilson mutex_enter(&l2arc_free_on_write_mtx); 2635dcbf3bd6SGeorge Wilson list_insert_head(l2arc_free_on_write, df); 2636dcbf3bd6SGeorge Wilson mutex_exit(&l2arc_free_on_write_mtx); 2637dcbf3bd6SGeorge Wilson } 2638b24ab676SJeff Bonwick 2639dcbf3bd6SGeorge Wilson static void 2640dcbf3bd6SGeorge Wilson arc_hdr_free_on_write(arc_buf_hdr_t *hdr) 2641dcbf3bd6SGeorge Wilson { 2642dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2643dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 2644dcbf3bd6SGeorge Wilson uint64_t size = arc_hdr_size(hdr); 2645dcbf3bd6SGeorge Wilson 2646dcbf3bd6SGeorge Wilson /* protected by hash lock, if in the hash table */ 2647dcbf3bd6SGeorge Wilson if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 2648dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2649dcbf3bd6SGeorge Wilson ASSERT(state != arc_anon && state != arc_l2c_only); 2650dcbf3bd6SGeorge Wilson 2651dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 2652dcbf3bd6SGeorge Wilson size, hdr); 2653dcbf3bd6SGeorge Wilson } 2654dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_size, size, hdr); 26556de76ce2SAndriy Gapon if (type == ARC_BUFC_METADATA) { 26566de76ce2SAndriy Gapon arc_space_return(size, ARC_SPACE_META); 26576de76ce2SAndriy Gapon } else { 26586de76ce2SAndriy Gapon ASSERT(type == ARC_BUFC_DATA); 26596de76ce2SAndriy Gapon arc_space_return(size, ARC_SPACE_DATA); 26606de76ce2SAndriy Gapon } 2661dcbf3bd6SGeorge Wilson 2662770499e1SDan Kimmel l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type); 2663dcbf3bd6SGeorge Wilson } 2664dcbf3bd6SGeorge Wilson 2665dcbf3bd6SGeorge Wilson /* 2666dcbf3bd6SGeorge Wilson * Share the arc_buf_t's data with the hdr. Whenever we are sharing the 2667dcbf3bd6SGeorge Wilson * data buffer, we transfer the refcount ownership to the hdr and update 2668dcbf3bd6SGeorge Wilson * the appropriate kstats. 2669dcbf3bd6SGeorge Wilson */ 2670dcbf3bd6SGeorge Wilson static void 2671dcbf3bd6SGeorge Wilson arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2672dcbf3bd6SGeorge Wilson { 2673dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2674dcbf3bd6SGeorge Wilson 26755602294fSDan Kimmel ASSERT(arc_can_share(hdr, buf)); 2676770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2677dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 26789253d63dSGeorge Wilson 26799253d63dSGeorge Wilson /* 2680dcbf3bd6SGeorge Wilson * Start sharing the data buffer. We transfer the 2681dcbf3bd6SGeorge Wilson * refcount ownership to the hdr since it always owns 2682dcbf3bd6SGeorge Wilson * the refcount whenever an arc_buf_t is shared. 26839253d63dSGeorge Wilson */ 2684dcbf3bd6SGeorge Wilson refcount_transfer_ownership(&state->arcs_size, buf, hdr); 2685770499e1SDan Kimmel hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf)); 2686770499e1SDan Kimmel abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd, 2687770499e1SDan Kimmel HDR_ISTYPE_METADATA(hdr)); 2688dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 26895602294fSDan Kimmel buf->b_flags |= ARC_BUF_FLAG_SHARED; 2690dcbf3bd6SGeorge Wilson 2691dcbf3bd6SGeorge Wilson /* 2692dcbf3bd6SGeorge Wilson * Since we've transferred ownership to the hdr we need 2693dcbf3bd6SGeorge Wilson * to increment its compressed and uncompressed kstats and 2694dcbf3bd6SGeorge Wilson * decrement the overhead size. 2695dcbf3bd6SGeorge Wilson */ 2696dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2697dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 26985602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf)); 2699ea8dc4b6Seschrock } 2700ea8dc4b6Seschrock 2701dcbf3bd6SGeorge Wilson static void 2702dcbf3bd6SGeorge Wilson arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2703ea8dc4b6Seschrock { 2704dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2705dcbf3bd6SGeorge Wilson 2706dcbf3bd6SGeorge Wilson ASSERT(arc_buf_is_shared(buf)); 2707770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2708dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2709ea8dc4b6Seschrock 27109b23f181Smaybee /* 2711dcbf3bd6SGeorge Wilson * We are no longer sharing this buffer so we need 2712dcbf3bd6SGeorge Wilson * to transfer its ownership to the rightful owner. 27139b23f181Smaybee */ 2714dcbf3bd6SGeorge Wilson refcount_transfer_ownership(&state->arcs_size, hdr, buf); 2715dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2716770499e1SDan Kimmel abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd); 2717770499e1SDan Kimmel abd_put(hdr->b_l1hdr.b_pabd); 2718770499e1SDan Kimmel hdr->b_l1hdr.b_pabd = NULL; 27195602294fSDan Kimmel buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 2720dcbf3bd6SGeorge Wilson 2721dcbf3bd6SGeorge Wilson /* 2722dcbf3bd6SGeorge Wilson * Since the buffer is no longer shared between 2723dcbf3bd6SGeorge Wilson * the arc buf and the hdr, count it as overhead. 2724dcbf3bd6SGeorge Wilson */ 2725dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2726dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 27275602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2728dcbf3bd6SGeorge Wilson } 2729dcbf3bd6SGeorge Wilson 2730dcbf3bd6SGeorge Wilson /* 27315602294fSDan Kimmel * Remove an arc_buf_t from the hdr's buf list and return the last 27325602294fSDan Kimmel * arc_buf_t on the list. If no buffers remain on the list then return 27335602294fSDan Kimmel * NULL. 27345602294fSDan Kimmel */ 27355602294fSDan Kimmel static arc_buf_t * 27365602294fSDan Kimmel arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf) 27375602294fSDan Kimmel { 27385602294fSDan Kimmel ASSERT(HDR_HAS_L1HDR(hdr)); 27395602294fSDan Kimmel ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 27405602294fSDan Kimmel 27415602294fSDan Kimmel arc_buf_t **bufp = &hdr->b_l1hdr.b_buf; 27425602294fSDan Kimmel arc_buf_t *lastbuf = NULL; 27435602294fSDan Kimmel 27445602294fSDan Kimmel /* 27455602294fSDan Kimmel * Remove the buf from the hdr list and locate the last 27465602294fSDan Kimmel * remaining buffer on the list. 27475602294fSDan Kimmel */ 27485602294fSDan Kimmel while (*bufp != NULL) { 27495602294fSDan Kimmel if (*bufp == buf) 27505602294fSDan Kimmel *bufp = buf->b_next; 27515602294fSDan Kimmel 27525602294fSDan Kimmel /* 27535602294fSDan Kimmel * If we've removed a buffer in the middle of 27545602294fSDan Kimmel * the list then update the lastbuf and update 27555602294fSDan Kimmel * bufp. 27565602294fSDan Kimmel */ 27575602294fSDan Kimmel if (*bufp != NULL) { 27585602294fSDan Kimmel lastbuf = *bufp; 27595602294fSDan Kimmel bufp = &(*bufp)->b_next; 27605602294fSDan Kimmel } 27615602294fSDan Kimmel } 27625602294fSDan Kimmel buf->b_next = NULL; 27635602294fSDan Kimmel ASSERT3P(lastbuf, !=, buf); 27645602294fSDan Kimmel IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL); 27655602294fSDan Kimmel IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL); 27665602294fSDan Kimmel IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf)); 27675602294fSDan Kimmel 27685602294fSDan Kimmel return (lastbuf); 27695602294fSDan Kimmel } 27705602294fSDan Kimmel 27715602294fSDan Kimmel /* 27725602294fSDan Kimmel * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's 27735602294fSDan Kimmel * list and free it. 2774dcbf3bd6SGeorge Wilson */ 2775dcbf3bd6SGeorge Wilson static void 27765602294fSDan Kimmel arc_buf_destroy_impl(arc_buf_t *buf) 2777dcbf3bd6SGeorge Wilson { 2778dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 2779dcbf3bd6SGeorge Wilson 2780dcbf3bd6SGeorge Wilson /* 27815602294fSDan Kimmel * Free up the data associated with the buf but only if we're not 27825602294fSDan Kimmel * sharing this with the hdr. If we are sharing it with the hdr, the 27835602294fSDan Kimmel * hdr is responsible for doing the free. 2784dcbf3bd6SGeorge Wilson */ 2785dcbf3bd6SGeorge Wilson if (buf->b_data != NULL) { 2786dcbf3bd6SGeorge Wilson /* 2787dcbf3bd6SGeorge Wilson * We're about to change the hdr's b_flags. We must either 2788dcbf3bd6SGeorge Wilson * hold the hash_lock or be undiscoverable. 2789dcbf3bd6SGeorge Wilson */ 2790dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2791dcbf3bd6SGeorge Wilson 2792dcbf3bd6SGeorge Wilson arc_cksum_verify(buf); 2793dcbf3bd6SGeorge Wilson arc_buf_unwatch(buf); 2794dcbf3bd6SGeorge Wilson 27955602294fSDan Kimmel if (arc_buf_is_shared(buf)) { 2796dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2797dcbf3bd6SGeorge Wilson } else { 27985602294fSDan Kimmel uint64_t size = arc_buf_size(buf); 2799dcbf3bd6SGeorge Wilson arc_free_data_buf(hdr, buf->b_data, size, buf); 2800dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_overhead_size, -size); 2801dcbf3bd6SGeorge Wilson } 2802dcbf3bd6SGeorge Wilson buf->b_data = NULL; 2803dcbf3bd6SGeorge Wilson 2804dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 2805dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt -= 1; 2806dcbf3bd6SGeorge Wilson } 2807dcbf3bd6SGeorge Wilson 28085602294fSDan Kimmel arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 2809dcbf3bd6SGeorge Wilson 28105602294fSDan Kimmel if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) { 2811dcbf3bd6SGeorge Wilson /* 28125602294fSDan Kimmel * If the current arc_buf_t is sharing its data buffer with the 2813770499e1SDan Kimmel * hdr, then reassign the hdr's b_pabd to share it with the new 28145602294fSDan Kimmel * buffer at the end of the list. The shared buffer is always 28155602294fSDan Kimmel * the last one on the hdr's buffer list. 28165602294fSDan Kimmel * 28175602294fSDan Kimmel * There is an equivalent case for compressed bufs, but since 28185602294fSDan Kimmel * they aren't guaranteed to be the last buf in the list and 28195602294fSDan Kimmel * that is an exceedingly rare case, we just allow that space be 28205602294fSDan Kimmel * wasted temporarily. 2821dcbf3bd6SGeorge Wilson */ 28225602294fSDan Kimmel if (lastbuf != NULL) { 28235602294fSDan Kimmel /* Only one buf can be shared at once */ 28245602294fSDan Kimmel VERIFY(!arc_buf_is_shared(lastbuf)); 28255602294fSDan Kimmel /* hdr is uncompressed so can't have compressed buf */ 28265602294fSDan Kimmel VERIFY(!ARC_BUF_COMPRESSED(lastbuf)); 2827dcbf3bd6SGeorge Wilson 2828770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2829770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 2830dcbf3bd6SGeorge Wilson 28315602294fSDan Kimmel /* 28325602294fSDan Kimmel * We must setup a new shared block between the 28335602294fSDan Kimmel * last buffer and the hdr. The data would have 28345602294fSDan Kimmel * been allocated by the arc buf so we need to transfer 28355602294fSDan Kimmel * ownership to the hdr since it's now being shared. 28365602294fSDan Kimmel */ 28375602294fSDan Kimmel arc_share_buf(hdr, lastbuf); 28385602294fSDan Kimmel } 28395602294fSDan Kimmel } else if (HDR_SHARED_DATA(hdr)) { 2840dcbf3bd6SGeorge Wilson /* 28415602294fSDan Kimmel * Uncompressed shared buffers are always at the end 28425602294fSDan Kimmel * of the list. Compressed buffers don't have the 28435602294fSDan Kimmel * same requirements. This makes it hard to 28445602294fSDan Kimmel * simply assert that the lastbuf is shared so 28455602294fSDan Kimmel * we rely on the hdr's compression flags to determine 28465602294fSDan Kimmel * if we have a compressed, shared buffer. 2847dcbf3bd6SGeorge Wilson */ 28485602294fSDan Kimmel ASSERT3P(lastbuf, !=, NULL); 28495602294fSDan Kimmel ASSERT(arc_buf_is_shared(lastbuf) || 28505602294fSDan Kimmel HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 2851dcbf3bd6SGeorge Wilson } 2852dcbf3bd6SGeorge Wilson 28535602294fSDan Kimmel /* 28545602294fSDan Kimmel * Free the checksum if we're removing the last uncompressed buf from 28555602294fSDan Kimmel * this hdr. 28565602294fSDan Kimmel */ 28575602294fSDan Kimmel if (!arc_hdr_has_uncompressed_buf(hdr)) { 2858dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 28595602294fSDan Kimmel } 2860dcbf3bd6SGeorge Wilson 2861dcbf3bd6SGeorge Wilson /* clean up the buf */ 2862dcbf3bd6SGeorge Wilson buf->b_hdr = NULL; 2863dcbf3bd6SGeorge Wilson kmem_cache_free(buf_cache, buf); 2864dcbf3bd6SGeorge Wilson } 2865dcbf3bd6SGeorge Wilson 2866dcbf3bd6SGeorge Wilson static void 2867770499e1SDan Kimmel arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr) 2868dcbf3bd6SGeorge Wilson { 2869dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 287089c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 2871dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(hdr)); 2872ea8dc4b6Seschrock 2873770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2874770499e1SDan Kimmel hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr); 2875dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2876770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 287789c86e32SChris Williamson 2878dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2879dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2880ea8dc4b6Seschrock } 2881ea8dc4b6Seschrock 2882244781f1SPrakash Surya static void 2883770499e1SDan Kimmel arc_hdr_free_pabd(arc_buf_hdr_t *hdr) 2884244781f1SPrakash Surya { 2885dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2886770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2887244781f1SPrakash Surya 2888dcbf3bd6SGeorge Wilson /* 2889dcbf3bd6SGeorge Wilson * If the hdr is currently being written to the l2arc then 2890dcbf3bd6SGeorge Wilson * we defer freeing the data by adding it to the l2arc_free_on_write 2891dcbf3bd6SGeorge Wilson * list. The l2arc will free the data once it's finished 2892dcbf3bd6SGeorge Wilson * writing it to the l2arc device. 2893dcbf3bd6SGeorge Wilson */ 2894dcbf3bd6SGeorge Wilson if (HDR_L2_WRITING(hdr)) { 2895dcbf3bd6SGeorge Wilson arc_hdr_free_on_write(hdr); 2896dcbf3bd6SGeorge Wilson ARCSTAT_BUMP(arcstat_l2_free_on_write); 2897dcbf3bd6SGeorge Wilson } else { 2898770499e1SDan Kimmel arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, 2899dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2900dcbf3bd6SGeorge Wilson } 2901770499e1SDan Kimmel hdr->b_l1hdr.b_pabd = NULL; 2902dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2903dcbf3bd6SGeorge Wilson 2904dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2905dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2906dcbf3bd6SGeorge Wilson } 2907dcbf3bd6SGeorge Wilson 2908dcbf3bd6SGeorge Wilson static arc_buf_hdr_t * 2909dcbf3bd6SGeorge Wilson arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, 29105602294fSDan Kimmel enum zio_compress compression_type, arc_buf_contents_t type) 2911dcbf3bd6SGeorge Wilson { 2912dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr; 2913dcbf3bd6SGeorge Wilson 2914dcbf3bd6SGeorge Wilson VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); 2915dcbf3bd6SGeorge Wilson 2916dcbf3bd6SGeorge Wilson hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2917dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 2918dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2919dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL); 2920dcbf3bd6SGeorge Wilson HDR_SET_PSIZE(hdr, psize); 2921dcbf3bd6SGeorge Wilson HDR_SET_LSIZE(hdr, lsize); 2922dcbf3bd6SGeorge Wilson hdr->b_spa = spa; 2923dcbf3bd6SGeorge Wilson hdr->b_type = type; 2924dcbf3bd6SGeorge Wilson hdr->b_flags = 0; 2925dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); 29265602294fSDan Kimmel arc_hdr_set_compress(hdr, compression_type); 2927dcbf3bd6SGeorge Wilson 2928dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_state = arc_anon; 2929dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_arc_access = 0; 2930dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt = 0; 2931dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_buf = NULL; 2932dcbf3bd6SGeorge Wilson 2933dcbf3bd6SGeorge Wilson /* 2934dcbf3bd6SGeorge Wilson * Allocate the hdr's buffer. This will contain either 2935dcbf3bd6SGeorge Wilson * the compressed or uncompressed data depending on the block 2936dcbf3bd6SGeorge Wilson * it references and compressed arc enablement. 2937dcbf3bd6SGeorge Wilson */ 2938770499e1SDan Kimmel arc_hdr_alloc_pabd(hdr); 2939dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2940dcbf3bd6SGeorge Wilson 2941dcbf3bd6SGeorge Wilson return (hdr); 2942244781f1SPrakash Surya } 2943244781f1SPrakash Surya 2944fa94a07fSbrendan /* 2945dcbf3bd6SGeorge Wilson * Transition between the two allocation states for the arc_buf_hdr struct. 2946dcbf3bd6SGeorge Wilson * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 2947dcbf3bd6SGeorge Wilson * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 2948dcbf3bd6SGeorge Wilson * version is used when a cache buffer is only in the L2ARC in order to reduce 2949dcbf3bd6SGeorge Wilson * memory usage. 2950fa94a07fSbrendan */ 2951dcbf3bd6SGeorge Wilson static arc_buf_hdr_t * 2952dcbf3bd6SGeorge Wilson arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 2953fa94a07fSbrendan { 2954dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L2HDR(hdr)); 2955dcbf3bd6SGeorge Wilson 2956dcbf3bd6SGeorge Wilson arc_buf_hdr_t *nhdr; 2957dcbf3bd6SGeorge Wilson l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2958dcbf3bd6SGeorge Wilson 2959dcbf3bd6SGeorge Wilson ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 2960dcbf3bd6SGeorge Wilson (old == hdr_l2only_cache && new == hdr_full_cache)); 2961dcbf3bd6SGeorge Wilson 2962dcbf3bd6SGeorge Wilson nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 2963dcbf3bd6SGeorge Wilson 2964dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 2965dcbf3bd6SGeorge Wilson buf_hash_remove(hdr); 2966dcbf3bd6SGeorge Wilson 2967dcbf3bd6SGeorge Wilson bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 2968dcbf3bd6SGeorge Wilson 2969dcbf3bd6SGeorge Wilson if (new == hdr_full_cache) { 2970dcbf3bd6SGeorge Wilson arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2971dcbf3bd6SGeorge Wilson /* 2972dcbf3bd6SGeorge Wilson * arc_access and arc_change_state need to be aware that a 2973dcbf3bd6SGeorge Wilson * header has just come out of L2ARC, so we set its state to 2974dcbf3bd6SGeorge Wilson * l2c_only even though it's about to change. 2975dcbf3bd6SGeorge Wilson */ 2976dcbf3bd6SGeorge Wilson nhdr->b_l1hdr.b_state = arc_l2c_only; 2977dcbf3bd6SGeorge Wilson 2978dcbf3bd6SGeorge Wilson /* Verify previous threads set to NULL before freeing */ 2979770499e1SDan Kimmel ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL); 2980dcbf3bd6SGeorge Wilson } else { 2981dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2982dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 2983dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2984dcbf3bd6SGeorge Wilson 2985dcbf3bd6SGeorge Wilson /* 2986dcbf3bd6SGeorge Wilson * If we've reached here, We must have been called from 2987dcbf3bd6SGeorge Wilson * arc_evict_hdr(), as such we should have already been 2988dcbf3bd6SGeorge Wilson * removed from any ghost list we were previously on 2989dcbf3bd6SGeorge Wilson * (which protects us from racing with arc_evict_state), 2990dcbf3bd6SGeorge Wilson * thus no locking is needed during this check. 2991dcbf3bd6SGeorge Wilson */ 2992dcbf3bd6SGeorge Wilson ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2993cd1c8b85SMatthew Ahrens 2994dcbf3bd6SGeorge Wilson /* 2995dcbf3bd6SGeorge Wilson * A buffer must not be moved into the arc_l2c_only 2996dcbf3bd6SGeorge Wilson * state if it's not finished being written out to the 2997770499e1SDan Kimmel * l2arc device. Otherwise, the b_l1hdr.b_pabd field 2998dcbf3bd6SGeorge Wilson * might try to be accessed, even though it was removed. 2999dcbf3bd6SGeorge Wilson */ 3000dcbf3bd6SGeorge Wilson VERIFY(!HDR_L2_WRITING(hdr)); 3001770499e1SDan Kimmel VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3002fa94a07fSbrendan 3003dcbf3bd6SGeorge Wilson #ifdef ZFS_DEBUG 3004dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_thawed != NULL) { 3005dcbf3bd6SGeorge Wilson kmem_free(hdr->b_l1hdr.b_thawed, 1); 3006dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_thawed = NULL; 3007dcbf3bd6SGeorge Wilson } 3008dcbf3bd6SGeorge Wilson #endif 3009244781f1SPrakash Surya 3010dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); 3011dcbf3bd6SGeorge Wilson } 3012244781f1SPrakash Surya /* 3013dcbf3bd6SGeorge Wilson * The header has been reallocated so we need to re-insert it into any 3014dcbf3bd6SGeorge Wilson * lists it was on. 3015244781f1SPrakash Surya */ 3016dcbf3bd6SGeorge Wilson (void) buf_hash_insert(nhdr, NULL); 3017244781f1SPrakash Surya 3018dcbf3bd6SGeorge Wilson ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 3019dcbf3bd6SGeorge Wilson 3020dcbf3bd6SGeorge Wilson mutex_enter(&dev->l2ad_mtx); 3021244781f1SPrakash Surya 3022244781f1SPrakash Surya /* 3023dcbf3bd6SGeorge Wilson * We must place the realloc'ed header back into the list at 3024dcbf3bd6SGeorge Wilson * the same spot. Otherwise, if it's placed earlier in the list, 3025dcbf3bd6SGeorge Wilson * l2arc_write_buffers() could find it during the function's 3026dcbf3bd6SGeorge Wilson * write phase, and try to write it out to the l2arc. 3027244781f1SPrakash Surya */ 3028dcbf3bd6SGeorge Wilson list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 3029dcbf3bd6SGeorge Wilson list_remove(&dev->l2ad_buflist, hdr); 3030dcbf3bd6SGeorge Wilson 3031dcbf3bd6SGeorge Wilson mutex_exit(&dev->l2ad_mtx); 3032244781f1SPrakash Surya 3033244781f1SPrakash Surya /* 3034dcbf3bd6SGeorge Wilson * Since we're using the pointer address as the tag when 3035dcbf3bd6SGeorge Wilson * incrementing and decrementing the l2ad_alloc refcount, we 3036dcbf3bd6SGeorge Wilson * must remove the old pointer (that we're about to destroy) and 3037dcbf3bd6SGeorge Wilson * add the new pointer to the refcount. Otherwise we'd remove 3038dcbf3bd6SGeorge Wilson * the wrong pointer address when calling arc_hdr_destroy() later. 3039244781f1SPrakash Surya */ 3040244781f1SPrakash Surya 3041dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); 3042dcbf3bd6SGeorge Wilson (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr); 3043244781f1SPrakash Surya 3044dcbf3bd6SGeorge Wilson buf_discard_identity(hdr); 3045dcbf3bd6SGeorge Wilson kmem_cache_free(old, hdr); 3046244781f1SPrakash Surya 3047dcbf3bd6SGeorge Wilson return (nhdr); 3048244781f1SPrakash Surya } 3049244781f1SPrakash Surya 3050bbfa8ea8SMatthew Ahrens /* 3051dcbf3bd6SGeorge Wilson * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. 3052dcbf3bd6SGeorge Wilson * The buf is returned thawed since we expect the consumer to modify it. 3053bbfa8ea8SMatthew Ahrens */ 3054dcbf3bd6SGeorge Wilson arc_buf_t * 30555602294fSDan Kimmel arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size) 3056ea8dc4b6Seschrock { 3057dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, 3058dcbf3bd6SGeorge Wilson ZIO_COMPRESS_OFF, type); 3059dcbf3bd6SGeorge Wilson ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 30605602294fSDan Kimmel 30615602294fSDan Kimmel arc_buf_t *buf = NULL; 30625602294fSDan Kimmel VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf)); 3063dcbf3bd6SGeorge Wilson arc_buf_thaw(buf); 30645602294fSDan Kimmel 30655602294fSDan Kimmel return (buf); 30665602294fSDan Kimmel } 30675602294fSDan Kimmel 30685602294fSDan Kimmel /* 30695602294fSDan Kimmel * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this 30705602294fSDan Kimmel * for bufs containing metadata. 30715602294fSDan Kimmel */ 30725602294fSDan Kimmel arc_buf_t * 30735602294fSDan Kimmel arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize, 30745602294fSDan Kimmel enum zio_compress compression_type) 30755602294fSDan Kimmel { 30765602294fSDan Kimmel ASSERT3U(lsize, >, 0); 30775602294fSDan Kimmel ASSERT3U(lsize, >=, psize); 30785602294fSDan Kimmel ASSERT(compression_type > ZIO_COMPRESS_OFF); 30795602294fSDan Kimmel ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS); 30805602294fSDan Kimmel 30815602294fSDan Kimmel arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 30825602294fSDan Kimmel compression_type, ARC_BUFC_DATA); 30835602294fSDan Kimmel ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 30845602294fSDan Kimmel 30855602294fSDan Kimmel arc_buf_t *buf = NULL; 30865602294fSDan Kimmel VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf)); 30875602294fSDan Kimmel arc_buf_thaw(buf); 30885602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 30895602294fSDan Kimmel 3090770499e1SDan Kimmel if (!arc_buf_is_shared(buf)) { 3091770499e1SDan Kimmel /* 3092770499e1SDan Kimmel * To ensure that the hdr has the correct data in it if we call 3093770499e1SDan Kimmel * arc_decompress() on this buf before it's been written to 3094770499e1SDan Kimmel * disk, it's easiest if we just set up sharing between the 3095770499e1SDan Kimmel * buf and the hdr. 3096770499e1SDan Kimmel */ 3097770499e1SDan Kimmel ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd)); 3098770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 3099770499e1SDan Kimmel arc_share_buf(hdr, buf); 3100770499e1SDan Kimmel } 3101770499e1SDan Kimmel 3102dcbf3bd6SGeorge Wilson return (buf); 3103ea8dc4b6Seschrock } 3104ea8dc4b6Seschrock 3105a52fc310SPrakash Surya static void 3106a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 3107a52fc310SPrakash Surya { 3108a52fc310SPrakash Surya l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 3109a52fc310SPrakash Surya l2arc_dev_t *dev = l2hdr->b_dev; 311016a7e5acSAndriy Gapon uint64_t psize = arc_hdr_size(hdr); 3111a52fc310SPrakash Surya 3112a52fc310SPrakash Surya ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 3113a52fc310SPrakash Surya ASSERT(HDR_HAS_L2HDR(hdr)); 3114a52fc310SPrakash Surya 3115a52fc310SPrakash Surya list_remove(&dev->l2ad_buflist, hdr); 3116a52fc310SPrakash Surya 311716a7e5acSAndriy Gapon ARCSTAT_INCR(arcstat_l2_psize, -psize); 311816a7e5acSAndriy Gapon ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); 3119a52fc310SPrakash Surya 312016a7e5acSAndriy Gapon vdev_space_update(dev->l2ad_vdev, -psize, 0, 0); 3121a52fc310SPrakash Surya 312216a7e5acSAndriy Gapon (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr); 3123dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 3124a52fc310SPrakash Surya } 3125a52fc310SPrakash Surya 3126fa9e4066Sahrens static void 3127ea8dc4b6Seschrock arc_hdr_destroy(arc_buf_hdr_t *hdr) 3128fa9e4066Sahrens { 312989c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) { 313089c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_buf == NULL || 3131dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt > 0); 313289c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 313389c86e32SChris Williamson ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 313489c86e32SChris Williamson } 3135ea8dc4b6Seschrock ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 313689c86e32SChris Williamson ASSERT(!HDR_IN_HASH_TABLE(hdr)); 313789c86e32SChris Williamson 3138dcbf3bd6SGeorge Wilson if (!HDR_EMPTY(hdr)) 3139dcbf3bd6SGeorge Wilson buf_discard_identity(hdr); 3140dcbf3bd6SGeorge Wilson 314189c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr)) { 3142a52fc310SPrakash Surya l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3143a52fc310SPrakash Surya boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 3144b24ab676SJeff Bonwick 3145a52fc310SPrakash Surya if (!buflist_held) 3146a52fc310SPrakash Surya mutex_enter(&dev->l2ad_mtx); 314789c86e32SChris Williamson 3148244781f1SPrakash Surya /* 3149a52fc310SPrakash Surya * Even though we checked this conditional above, we 3150a52fc310SPrakash Surya * need to check this again now that we have the 3151a52fc310SPrakash Surya * l2ad_mtx. This is because we could be racing with 3152a52fc310SPrakash Surya * another thread calling l2arc_evict() which might have 3153a52fc310SPrakash Surya * destroyed this header's L2 portion as we were waiting 3154a52fc310SPrakash Surya * to acquire the l2ad_mtx. If that happens, we don't 3155a52fc310SPrakash Surya * want to re-destroy the header's L2 portion. 3156244781f1SPrakash Surya */ 3157a52fc310SPrakash Surya if (HDR_HAS_L2HDR(hdr)) 3158a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(hdr); 3159b24ab676SJeff Bonwick 3160b24ab676SJeff Bonwick if (!buflist_held) 3161a52fc310SPrakash Surya mutex_exit(&dev->l2ad_mtx); 3162fa94a07fSbrendan } 3163fa94a07fSbrendan 3164dcbf3bd6SGeorge Wilson if (HDR_HAS_L1HDR(hdr)) { 3165dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 316689c86e32SChris Williamson 3167dcbf3bd6SGeorge Wilson while (hdr->b_l1hdr.b_buf != NULL) 31685602294fSDan Kimmel arc_buf_destroy_impl(hdr->b_l1hdr.b_buf); 316989c86e32SChris Williamson 317089c86e32SChris Williamson #ifdef ZFS_DEBUG 317189c86e32SChris Williamson if (hdr->b_l1hdr.b_thawed != NULL) { 317289c86e32SChris Williamson kmem_free(hdr->b_l1hdr.b_thawed, 1); 317389c86e32SChris Williamson hdr->b_l1hdr.b_thawed = NULL; 317489c86e32SChris Williamson } 317589c86e32SChris Williamson #endif 3176dcbf3bd6SGeorge Wilson 3177770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 3178770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 3179dcbf3bd6SGeorge Wilson } 31803f9d6ad7SLin Ling } 3181ea8dc4b6Seschrock 3182fa9e4066Sahrens ASSERT3P(hdr->b_hash_next, ==, NULL); 318389c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) { 3184244781f1SPrakash Surya ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 318589c86e32SChris Williamson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 318689c86e32SChris Williamson kmem_cache_free(hdr_full_cache, hdr); 318789c86e32SChris Williamson } else { 318889c86e32SChris Williamson kmem_cache_free(hdr_l2only_cache, hdr); 318989c86e32SChris Williamson } 3190fa9e4066Sahrens } 3191fa9e4066Sahrens 3192fa9e4066Sahrens void 3193dcbf3bd6SGeorge Wilson arc_buf_destroy(arc_buf_t *buf, void* tag) 3194ea8dc4b6Seschrock { 3195ea8dc4b6Seschrock arc_buf_hdr_t *hdr = buf->b_hdr; 3196ea8dc4b6Seschrock kmutex_t *hash_lock = HDR_LOCK(hdr); 3197fa9e4066Sahrens 319889c86e32SChris Williamson if (hdr->b_l1hdr.b_state == arc_anon) { 3199dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 3200dcbf3bd6SGeorge Wilson ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3201dcbf3bd6SGeorge Wilson VERIFY0(remove_reference(hdr, NULL, tag)); 3202dcbf3bd6SGeorge Wilson arc_hdr_destroy(hdr); 3203dcbf3bd6SGeorge Wilson return; 3204ea8dc4b6Seschrock } 3205ea8dc4b6Seschrock 3206ea8dc4b6Seschrock mutex_enter(hash_lock); 3207dcbf3bd6SGeorge Wilson ASSERT3P(hdr, ==, buf->b_hdr); 3208dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 32093f9d6ad7SLin Ling ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3210dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); 3211dcbf3bd6SGeorge Wilson ASSERT3P(buf->b_data, !=, NULL); 3212ea8dc4b6Seschrock 3213ea8dc4b6Seschrock (void) remove_reference(hdr, hash_lock, tag); 32145602294fSDan Kimmel arc_buf_destroy_impl(buf); 3215ea8dc4b6Seschrock mutex_exit(hash_lock); 3216fa9e4066Sahrens } 3217fa9e4066Sahrens 3218fa9e4066Sahrens /* 3219244781f1SPrakash Surya * Evict the arc_buf_hdr that is provided as a parameter. The resultant 3220244781f1SPrakash Surya * state of the header is dependent on it's state prior to entering this 3221244781f1SPrakash Surya * function. The following transitions are possible: 3222874395d5Smaybee * 3223244781f1SPrakash Surya * - arc_mru -> arc_mru_ghost 3224244781f1SPrakash Surya * - arc_mfu -> arc_mfu_ghost 3225244781f1SPrakash Surya * - arc_mru_ghost -> arc_l2c_only 3226244781f1SPrakash Surya * - arc_mru_ghost -> deleted 3227244781f1SPrakash Surya * - arc_mfu_ghost -> arc_l2c_only 3228244781f1SPrakash Surya * - arc_mfu_ghost -> deleted 3229fa9e4066Sahrens */ 3230244781f1SPrakash Surya static int64_t 3231244781f1SPrakash Surya arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3232fa9e4066Sahrens { 3233244781f1SPrakash Surya arc_state_t *evicted_state, *state; 3234244781f1SPrakash Surya int64_t bytes_evicted = 0; 3235fa9e4066Sahrens 3236244781f1SPrakash Surya ASSERT(MUTEX_HELD(hash_lock)); 3237244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 3238fa9e4066Sahrens 3239244781f1SPrakash Surya state = hdr->b_l1hdr.b_state; 3240244781f1SPrakash Surya if (GHOST_STATE(state)) { 3241244781f1SPrakash Surya ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3242dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3243fa9e4066Sahrens 3244244781f1SPrakash Surya /* 3245244781f1SPrakash Surya * l2arc_write_buffers() relies on a header's L1 portion 3246770499e1SDan Kimmel * (i.e. its b_pabd field) during it's write phase. 3247244781f1SPrakash Surya * Thus, we cannot push a header onto the arc_l2c_only 3248244781f1SPrakash Surya * state (removing it's L1 piece) until the header is 3249244781f1SPrakash Surya * done being written to the l2arc. 3250244781f1SPrakash Surya */ 3251244781f1SPrakash Surya if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 3252244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_l2_skip); 3253244781f1SPrakash Surya return (bytes_evicted); 32543a5286a1SMatthew Ahrens } 3255244781f1SPrakash Surya 3256244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_deleted); 3257dcbf3bd6SGeorge Wilson bytes_evicted += HDR_GET_LSIZE(hdr); 3258244781f1SPrakash Surya 3259244781f1SPrakash Surya DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 3260244781f1SPrakash Surya 3261770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3262244781f1SPrakash Surya if (HDR_HAS_L2HDR(hdr)) { 3263244781f1SPrakash Surya /* 3264244781f1SPrakash Surya * This buffer is cached on the 2nd Level ARC; 3265244781f1SPrakash Surya * don't destroy the header. 3266244781f1SPrakash Surya */ 3267244781f1SPrakash Surya arc_change_state(arc_l2c_only, hdr, hash_lock); 32683a5286a1SMatthew Ahrens /* 3269244781f1SPrakash Surya * dropping from L1+L2 cached to L2-only, 3270244781f1SPrakash Surya * realloc to remove the L1 header. 32713a5286a1SMatthew Ahrens */ 3272244781f1SPrakash Surya hdr = arc_hdr_realloc(hdr, hdr_full_cache, 3273244781f1SPrakash Surya hdr_l2only_cache); 3274244781f1SPrakash Surya } else { 3275244781f1SPrakash Surya arc_change_state(arc_anon, hdr, hash_lock); 3276244781f1SPrakash Surya arc_hdr_destroy(hdr); 32773a5286a1SMatthew Ahrens } 3278244781f1SPrakash Surya return (bytes_evicted); 32793a5286a1SMatthew Ahrens } 32803a5286a1SMatthew Ahrens 3281244781f1SPrakash Surya ASSERT(state == arc_mru || state == arc_mfu); 3282244781f1SPrakash Surya evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3283244781f1SPrakash Surya 3284244781f1SPrakash Surya /* prefetch buffers have a minimum lifespan */ 3285244781f1SPrakash Surya if (HDR_IO_IN_PROGRESS(hdr) || 3286244781f1SPrakash Surya ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 3287244781f1SPrakash Surya ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 3288244781f1SPrakash Surya arc_min_prefetch_lifespan)) { 3289244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_skip); 3290244781f1SPrakash Surya return (bytes_evicted); 3291244781f1SPrakash Surya } 3292244781f1SPrakash Surya 3293244781f1SPrakash Surya ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3294244781f1SPrakash Surya while (hdr->b_l1hdr.b_buf) { 3295244781f1SPrakash Surya arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3296244781f1SPrakash Surya if (!mutex_tryenter(&buf->b_evict_lock)) { 3297244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_mutex_miss); 3298244781f1SPrakash Surya break; 329913506d1eSmaybee } 3300244781f1SPrakash Surya if (buf->b_data != NULL) 3301dcbf3bd6SGeorge Wilson bytes_evicted += HDR_GET_LSIZE(hdr); 3302dcbf3bd6SGeorge Wilson mutex_exit(&buf->b_evict_lock); 33035602294fSDan Kimmel arc_buf_destroy_impl(buf); 3304244781f1SPrakash Surya } 330569962b56SMatthew Ahrens 3306244781f1SPrakash Surya if (HDR_HAS_L2HDR(hdr)) { 3307dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); 3308244781f1SPrakash Surya } else { 3309dcbf3bd6SGeorge Wilson if (l2arc_write_eligible(hdr->b_spa, hdr)) { 3310dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_evict_l2_eligible, 3311dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr)); 3312dcbf3bd6SGeorge Wilson } else { 3313dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_evict_l2_ineligible, 3314dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr)); 3315dcbf3bd6SGeorge Wilson } 3316244781f1SPrakash Surya } 3317244781f1SPrakash Surya 3318dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_bufcnt == 0) { 3319dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 3320dcbf3bd6SGeorge Wilson 3321dcbf3bd6SGeorge Wilson bytes_evicted += arc_hdr_size(hdr); 3322dcbf3bd6SGeorge Wilson 3323dcbf3bd6SGeorge Wilson /* 3324dcbf3bd6SGeorge Wilson * If this hdr is being evicted and has a compressed 3325dcbf3bd6SGeorge Wilson * buffer then we discard it here before we change states. 3326dcbf3bd6SGeorge Wilson * This ensures that the accounting is updated correctly 3327770499e1SDan Kimmel * in arc_free_data_impl(). 3328dcbf3bd6SGeorge Wilson */ 3329770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 3330dcbf3bd6SGeorge Wilson 3331244781f1SPrakash Surya arc_change_state(evicted_state, hdr, hash_lock); 3332244781f1SPrakash Surya ASSERT(HDR_IN_HASH_TABLE(hdr)); 3333dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 3334244781f1SPrakash Surya DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 3335244781f1SPrakash Surya } 3336244781f1SPrakash Surya 3337244781f1SPrakash Surya return (bytes_evicted); 3338244781f1SPrakash Surya } 3339244781f1SPrakash Surya 3340244781f1SPrakash Surya static uint64_t 3341244781f1SPrakash Surya arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 3342244781f1SPrakash Surya uint64_t spa, int64_t bytes) 3343244781f1SPrakash Surya { 3344244781f1SPrakash Surya multilist_sublist_t *mls; 3345244781f1SPrakash Surya uint64_t bytes_evicted = 0; 3346244781f1SPrakash Surya arc_buf_hdr_t *hdr; 3347244781f1SPrakash Surya kmutex_t *hash_lock; 3348244781f1SPrakash Surya int evict_count = 0; 3349244781f1SPrakash Surya 3350244781f1SPrakash Surya ASSERT3P(marker, !=, NULL); 3351244781f1SPrakash Surya IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3352244781f1SPrakash Surya 3353244781f1SPrakash Surya mls = multilist_sublist_lock(ml, idx); 3354244781f1SPrakash Surya 3355244781f1SPrakash Surya for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 3356244781f1SPrakash Surya hdr = multilist_sublist_prev(mls, marker)) { 3357244781f1SPrakash Surya if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 3358244781f1SPrakash Surya (evict_count >= zfs_arc_evict_batch_limit)) 3359244781f1SPrakash Surya break; 336069962b56SMatthew Ahrens 336169962b56SMatthew Ahrens /* 3362244781f1SPrakash Surya * To keep our iteration location, move the marker 3363244781f1SPrakash Surya * forward. Since we're not holding hdr's hash lock, we 3364244781f1SPrakash Surya * must be very careful and not remove 'hdr' from the 3365244781f1SPrakash Surya * sublist. Otherwise, other consumers might mistake the 3366244781f1SPrakash Surya * 'hdr' as not being on a sublist when they call the 3367244781f1SPrakash Surya * multilist_link_active() function (they all rely on 3368244781f1SPrakash Surya * the hash lock protecting concurrent insertions and 3369244781f1SPrakash Surya * removals). multilist_sublist_move_forward() was 3370244781f1SPrakash Surya * specifically implemented to ensure this is the case 3371244781f1SPrakash Surya * (only 'marker' will be removed and re-inserted). 3372244781f1SPrakash Surya */ 3373244781f1SPrakash Surya multilist_sublist_move_forward(mls, marker); 3374244781f1SPrakash Surya 3375244781f1SPrakash Surya /* 3376244781f1SPrakash Surya * The only case where the b_spa field should ever be 3377244781f1SPrakash Surya * zero, is the marker headers inserted by 3378244781f1SPrakash Surya * arc_evict_state(). It's possible for multiple threads 3379244781f1SPrakash Surya * to be calling arc_evict_state() concurrently (e.g. 3380244781f1SPrakash Surya * dsl_pool_close() and zio_inject_fault()), so we must 3381244781f1SPrakash Surya * skip any markers we see from these other threads. 338269962b56SMatthew Ahrens */ 3383244781f1SPrakash Surya if (hdr->b_spa == 0) 3384244781f1SPrakash Surya continue; 3385244781f1SPrakash Surya 3386244781f1SPrakash Surya /* we're only interested in evicting buffers of a certain spa */ 3387244781f1SPrakash Surya if (spa != 0 && hdr->b_spa != spa) { 3388244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_skip); 338969962b56SMatthew Ahrens continue; 339069962b56SMatthew Ahrens } 339169962b56SMatthew Ahrens 33927adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 33935ea40c06SBrendan Gregg - Sun Microsystems 3394244781f1SPrakash Surya /* 3395244781f1SPrakash Surya * We aren't calling this function from any code path 3396244781f1SPrakash Surya * that would already be holding a hash lock, so we're 3397244781f1SPrakash Surya * asserting on this assumption to be defensive in case 3398244781f1SPrakash Surya * this ever changes. Without this check, it would be 3399244781f1SPrakash Surya * possible to incorrectly increment arcstat_mutex_miss 3400244781f1SPrakash Surya * below (e.g. if the code changed such that we called 3401244781f1SPrakash Surya * this function with a hash lock held). 3402244781f1SPrakash Surya */ 3403244781f1SPrakash Surya ASSERT(!MUTEX_HELD(hash_lock)); 340444cb6abcSbmc 3405244781f1SPrakash Surya if (mutex_tryenter(hash_lock)) { 3406244781f1SPrakash Surya uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 3407244781f1SPrakash Surya mutex_exit(hash_lock); 3408fa9e4066Sahrens 3409244781f1SPrakash Surya bytes_evicted += evicted; 3410fa9e4066Sahrens 3411244781f1SPrakash Surya /* 3412244781f1SPrakash Surya * If evicted is zero, arc_evict_hdr() must have 3413244781f1SPrakash Surya * decided to skip this header, don't increment 3414244781f1SPrakash Surya * evict_count in this case. 3415244781f1SPrakash Surya */ 3416244781f1SPrakash Surya if (evicted != 0) 3417244781f1SPrakash Surya evict_count++; 341844cb6abcSbmc 3419244781f1SPrakash Surya /* 3420244781f1SPrakash Surya * If arc_size isn't overflowing, signal any 3421244781f1SPrakash Surya * threads that might happen to be waiting. 3422244781f1SPrakash Surya * 3423244781f1SPrakash Surya * For each header evicted, we wake up a single 3424244781f1SPrakash Surya * thread. If we used cv_broadcast, we could 3425244781f1SPrakash Surya * wake up "too many" threads causing arc_size 3426244781f1SPrakash Surya * to significantly overflow arc_c; since 3427770499e1SDan Kimmel * arc_get_data_impl() doesn't check for overflow 3428244781f1SPrakash Surya * when it's woken up (it doesn't because it's 3429244781f1SPrakash Surya * possible for the ARC to be overflowing while 3430244781f1SPrakash Surya * full of un-evictable buffers, and the 3431244781f1SPrakash Surya * function should proceed in this case). 3432244781f1SPrakash Surya * 3433244781f1SPrakash Surya * If threads are left sleeping, due to not 3434de753e34SBrad Lewis * using cv_broadcast here, they will be woken 3435de753e34SBrad Lewis * up via cv_broadcast in arc_adjust_cb() just 3436de753e34SBrad Lewis * before arc_adjust_zthr sleeps. 3437244781f1SPrakash Surya */ 3438de753e34SBrad Lewis mutex_enter(&arc_adjust_lock); 3439244781f1SPrakash Surya if (!arc_is_overflowing()) 3440de753e34SBrad Lewis cv_signal(&arc_adjust_waiters_cv); 3441de753e34SBrad Lewis mutex_exit(&arc_adjust_lock); 3442244781f1SPrakash Surya } else { 3443244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_mutex_miss); 3444244781f1SPrakash Surya } 3445244781f1SPrakash Surya } 3446f4d2e9e6Smaybee 3447244781f1SPrakash Surya multilist_sublist_unlock(mls); 344844cb6abcSbmc 3449244781f1SPrakash Surya return (bytes_evicted); 3450fa9e4066Sahrens } 3451fa9e4066Sahrens 3452fa9e4066Sahrens /* 3453244781f1SPrakash Surya * Evict buffers from the given arc state, until we've removed the 3454244781f1SPrakash Surya * specified number of bytes. Move the removed buffers to the 3455244781f1SPrakash Surya * appropriate evict state. 3456244781f1SPrakash Surya * 3457244781f1SPrakash Surya * This function makes a "best effort". It skips over any buffers 3458244781f1SPrakash Surya * it can't get a hash_lock on, and so, may not catch all candidates. 3459244781f1SPrakash Surya * It may also return without evicting as much space as requested. 3460244781f1SPrakash Surya * 3461244781f1SPrakash Surya * If bytes is specified using the special value ARC_EVICT_ALL, this 3462244781f1SPrakash Surya * will evict all available (i.e. unlocked and evictable) buffers from 3463244781f1SPrakash Surya * the given arc state; which is used by arc_flush(). 3464fa9e4066Sahrens */ 3465244781f1SPrakash Surya static uint64_t 3466244781f1SPrakash Surya arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 3467244781f1SPrakash Surya arc_buf_contents_t type) 3468fa9e4066Sahrens { 3469244781f1SPrakash Surya uint64_t total_evicted = 0; 347094c2d0ebSMatthew Ahrens multilist_t *ml = state->arcs_list[type]; 3471244781f1SPrakash Surya int num_sublists; 3472244781f1SPrakash Surya arc_buf_hdr_t **markers; 3473fa9e4066Sahrens 3474244781f1SPrakash Surya IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3475b802aa8cSSanjeev Bagewadi 3476244781f1SPrakash Surya num_sublists = multilist_get_num_sublists(ml); 3477b802aa8cSSanjeev Bagewadi 3478244781f1SPrakash Surya /* 3479244781f1SPrakash Surya * If we've tried to evict from each sublist, made some 3480244781f1SPrakash Surya * progress, but still have not hit the target number of bytes 3481244781f1SPrakash Surya * to evict, we want to keep trying. The markers allow us to 3482244781f1SPrakash Surya * pick up where we left off for each individual sublist, rather 3483244781f1SPrakash Surya * than starting from the tail each time. 3484244781f1SPrakash Surya */ 3485244781f1SPrakash Surya markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 3486244781f1SPrakash Surya for (int i = 0; i < num_sublists; i++) { 3487244781f1SPrakash Surya markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 348869962b56SMatthew Ahrens 348969962b56SMatthew Ahrens /* 3490244781f1SPrakash Surya * A b_spa of 0 is used to indicate that this header is 3491244781f1SPrakash Surya * a marker. This fact is used in arc_adjust_type() and 3492244781f1SPrakash Surya * arc_evict_state_impl(). 349369962b56SMatthew Ahrens */ 3494244781f1SPrakash Surya markers[i]->b_spa = 0; 3495fa94a07fSbrendan 3496244781f1SPrakash Surya multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3497244781f1SPrakash Surya multilist_sublist_insert_tail(mls, markers[i]); 3498244781f1SPrakash Surya multilist_sublist_unlock(mls); 3499244781f1SPrakash Surya } 3500fa94a07fSbrendan 3501244781f1SPrakash Surya /* 3502244781f1SPrakash Surya * While we haven't hit our target number of bytes to evict, or 3503244781f1SPrakash Surya * we're evicting all available buffers. 3504244781f1SPrakash Surya */ 3505244781f1SPrakash Surya while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 3506244781f1SPrakash Surya /* 3507244781f1SPrakash Surya * Start eviction using a randomly selected sublist, 3508244781f1SPrakash Surya * this is to try and evenly balance eviction across all 3509244781f1SPrakash Surya * sublists. Always starting at the same sublist 3510244781f1SPrakash Surya * (e.g. index 0) would cause evictions to favor certain 3511244781f1SPrakash Surya * sublists over others. 3512244781f1SPrakash Surya */ 3513244781f1SPrakash Surya int sublist_idx = multilist_get_random_index(ml); 3514244781f1SPrakash Surya uint64_t scan_evicted = 0; 3515244781f1SPrakash Surya 3516244781f1SPrakash Surya for (int i = 0; i < num_sublists; i++) { 3517244781f1SPrakash Surya uint64_t bytes_remaining; 3518244781f1SPrakash Surya uint64_t bytes_evicted; 3519244781f1SPrakash Surya 3520244781f1SPrakash Surya if (bytes == ARC_EVICT_ALL) 3521244781f1SPrakash Surya bytes_remaining = ARC_EVICT_ALL; 3522244781f1SPrakash Surya else if (total_evicted < bytes) 3523244781f1SPrakash Surya bytes_remaining = bytes - total_evicted; 3524244781f1SPrakash Surya else 3525fa9e4066Sahrens break; 3526244781f1SPrakash Surya 3527244781f1SPrakash Surya bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 3528244781f1SPrakash Surya markers[sublist_idx], spa, bytes_remaining); 3529244781f1SPrakash Surya 3530244781f1SPrakash Surya scan_evicted += bytes_evicted; 3531244781f1SPrakash Surya total_evicted += bytes_evicted; 3532244781f1SPrakash Surya 3533244781f1SPrakash Surya /* we've reached the end, wrap to the beginning */ 3534244781f1SPrakash Surya if (++sublist_idx >= num_sublists) 3535244781f1SPrakash Surya sublist_idx = 0; 3536244781f1SPrakash Surya } 3537244781f1SPrakash Surya 3538244781f1SPrakash Surya /* 3539244781f1SPrakash Surya * If we didn't evict anything during this scan, we have 3540244781f1SPrakash Surya * no reason to believe we'll evict more during another 3541244781f1SPrakash Surya * scan, so break the loop. 3542244781f1SPrakash Surya */ 3543244781f1SPrakash Surya if (scan_evicted == 0) { 3544244781f1SPrakash Surya /* This isn't possible, let's make that obvious */ 3545244781f1SPrakash Surya ASSERT3S(bytes, !=, 0); 3546244781f1SPrakash Surya 3547b802aa8cSSanjeev Bagewadi /* 3548244781f1SPrakash Surya * When bytes is ARC_EVICT_ALL, the only way to 3549244781f1SPrakash Surya * break the loop is when scan_evicted is zero. 3550244781f1SPrakash Surya * In that case, we actually have evicted enough, 3551244781f1SPrakash Surya * so we don't want to increment the kstat. 3552b802aa8cSSanjeev Bagewadi */ 3553244781f1SPrakash Surya if (bytes != ARC_EVICT_ALL) { 3554244781f1SPrakash Surya ASSERT3S(total_evicted, <, bytes); 3555244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_not_enough); 3556244781f1SPrakash Surya } 3557244781f1SPrakash Surya 3558244781f1SPrakash Surya break; 355969962b56SMatthew Ahrens } 3560244781f1SPrakash Surya } 3561244781f1SPrakash Surya 3562244781f1SPrakash Surya for (int i = 0; i < num_sublists; i++) { 3563244781f1SPrakash Surya multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3564244781f1SPrakash Surya multilist_sublist_remove(mls, markers[i]); 3565244781f1SPrakash Surya multilist_sublist_unlock(mls); 356669962b56SMatthew Ahrens 3567244781f1SPrakash Surya kmem_cache_free(hdr_full_cache, markers[i]); 3568fa9e4066Sahrens } 3569244781f1SPrakash Surya kmem_free(markers, sizeof (*markers) * num_sublists); 3570fa9e4066Sahrens 3571244781f1SPrakash Surya return (total_evicted); 3572244781f1SPrakash Surya } 3573244781f1SPrakash Surya 3574244781f1SPrakash Surya /* 3575244781f1SPrakash Surya * Flush all "evictable" data of the given type from the arc state 3576244781f1SPrakash Surya * specified. This will not evict any "active" buffers (i.e. referenced). 3577244781f1SPrakash Surya * 3578dcbf3bd6SGeorge Wilson * When 'retry' is set to B_FALSE, the function will make a single pass 3579244781f1SPrakash Surya * over the state and evict any buffers that it can. Since it doesn't 3580244781f1SPrakash Surya * continually retry the eviction, it might end up leaving some buffers 3581244781f1SPrakash Surya * in the ARC due to lock misses. 3582244781f1SPrakash Surya * 3583dcbf3bd6SGeorge Wilson * When 'retry' is set to B_TRUE, the function will continually retry the 3584244781f1SPrakash Surya * eviction until *all* evictable buffers have been removed from the 3585244781f1SPrakash Surya * state. As a result, if concurrent insertions into the state are 3586244781f1SPrakash Surya * allowed (e.g. if the ARC isn't shutting down), this function might 3587244781f1SPrakash Surya * wind up in an infinite loop, continually trying to evict buffers. 3588244781f1SPrakash Surya */ 3589244781f1SPrakash Surya static uint64_t 3590244781f1SPrakash Surya arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 3591244781f1SPrakash Surya boolean_t retry) 3592244781f1SPrakash Surya { 3593244781f1SPrakash Surya uint64_t evicted = 0; 3594244781f1SPrakash Surya 3595dcbf3bd6SGeorge Wilson while (refcount_count(&state->arcs_esize[type]) != 0) { 3596244781f1SPrakash Surya evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 3597244781f1SPrakash Surya 3598244781f1SPrakash Surya if (!retry) 3599244781f1SPrakash Surya break; 36000e8c6158Smaybee } 36010e8c6158Smaybee 3602244781f1SPrakash Surya return (evicted); 3603244781f1SPrakash Surya } 3604244781f1SPrakash Surya 3605244781f1SPrakash Surya /* 3606244781f1SPrakash Surya * Evict the specified number of bytes from the state specified, 3607244781f1SPrakash Surya * restricting eviction to the spa and type given. This function 3608244781f1SPrakash Surya * prevents us from trying to evict more from a state's list than 3609244781f1SPrakash Surya * is "evictable", and to skip evicting altogether when passed a 3610244781f1SPrakash Surya * negative value for "bytes". In contrast, arc_evict_state() will 3611244781f1SPrakash Surya * evict everything it can, when passed a negative value for "bytes". 3612244781f1SPrakash Surya */ 3613244781f1SPrakash Surya static uint64_t 3614244781f1SPrakash Surya arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3615244781f1SPrakash Surya arc_buf_contents_t type) 3616244781f1SPrakash Surya { 3617244781f1SPrakash Surya int64_t delta; 3618244781f1SPrakash Surya 3619dcbf3bd6SGeorge Wilson if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) { 3620dcbf3bd6SGeorge Wilson delta = MIN(refcount_count(&state->arcs_esize[type]), bytes); 3621244781f1SPrakash Surya return (arc_evict_state(state, spa, delta, type)); 3622fa9e4066Sahrens } 3623fa9e4066Sahrens 3624244781f1SPrakash Surya return (0); 3625fa9e4066Sahrens } 3626fa9e4066Sahrens 3627244781f1SPrakash Surya /* 3628244781f1SPrakash Surya * Evict metadata buffers from the cache, such that arc_meta_used is 3629244781f1SPrakash Surya * capped by the arc_meta_limit tunable. 3630244781f1SPrakash Surya */ 3631244781f1SPrakash Surya static uint64_t 36323a2d8a1bSPaul Dagnelie arc_adjust_meta(uint64_t meta_used) 3633244781f1SPrakash Surya { 3634244781f1SPrakash Surya uint64_t total_evicted = 0; 3635244781f1SPrakash Surya int64_t target; 3636244781f1SPrakash Surya 3637244781f1SPrakash Surya /* 3638244781f1SPrakash Surya * If we're over the meta limit, we want to evict enough 3639244781f1SPrakash Surya * metadata to get back under the meta limit. We don't want to 3640244781f1SPrakash Surya * evict so much that we drop the MRU below arc_p, though. If 3641244781f1SPrakash Surya * we're over the meta limit more than we're over arc_p, we 3642244781f1SPrakash Surya * evict some from the MRU here, and some from the MFU below. 3643244781f1SPrakash Surya */ 36443a2d8a1bSPaul Dagnelie target = MIN((int64_t)(meta_used - arc_meta_limit), 36452fd872a7SPrakash Surya (int64_t)(refcount_count(&arc_anon->arcs_size) + 36462fd872a7SPrakash Surya refcount_count(&arc_mru->arcs_size) - arc_p)); 3647244781f1SPrakash Surya 3648244781f1SPrakash Surya total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3649244781f1SPrakash Surya 3650244781f1SPrakash Surya /* 3651244781f1SPrakash Surya * Similar to the above, we want to evict enough bytes to get us 3652244781f1SPrakash Surya * below the meta limit, but not so much as to drop us below the 36535602294fSDan Kimmel * space allotted to the MFU (which is defined as arc_c - arc_p). 3654244781f1SPrakash Surya */ 36553a2d8a1bSPaul Dagnelie target = MIN((int64_t)(meta_used - arc_meta_limit), 36563a2d8a1bSPaul Dagnelie (int64_t)(refcount_count(&arc_mfu->arcs_size) - 36573a2d8a1bSPaul Dagnelie (arc_c - arc_p))); 3658244781f1SPrakash Surya 3659244781f1SPrakash Surya total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3660244781f1SPrakash Surya 3661244781f1SPrakash Surya return (total_evicted); 3662244781f1SPrakash Surya } 3663244781f1SPrakash Surya 3664244781f1SPrakash Surya /* 3665244781f1SPrakash Surya * Return the type of the oldest buffer in the given arc state 3666244781f1SPrakash Surya * 3667244781f1SPrakash Surya * This function will select a random sublist of type ARC_BUFC_DATA and 3668244781f1SPrakash Surya * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3669244781f1SPrakash Surya * is compared, and the type which contains the "older" buffer will be 3670244781f1SPrakash Surya * returned. 3671244781f1SPrakash Surya */ 3672244781f1SPrakash Surya static arc_buf_contents_t 3673244781f1SPrakash Surya arc_adjust_type(arc_state_t *state) 3674244781f1SPrakash Surya { 367594c2d0ebSMatthew Ahrens multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA]; 367694c2d0ebSMatthew Ahrens multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA]; 3677244781f1SPrakash Surya int data_idx = multilist_get_random_index(data_ml); 3678244781f1SPrakash Surya int meta_idx = multilist_get_random_index(meta_ml); 3679244781f1SPrakash Surya multilist_sublist_t *data_mls; 3680244781f1SPrakash Surya multilist_sublist_t *meta_mls; 3681244781f1SPrakash Surya arc_buf_contents_t type; 3682244781f1SPrakash Surya arc_buf_hdr_t *data_hdr; 3683244781f1SPrakash Surya arc_buf_hdr_t *meta_hdr; 3684244781f1SPrakash Surya 3685244781f1SPrakash Surya /* 3686244781f1SPrakash Surya * We keep the sublist lock until we're finished, to prevent 3687244781f1SPrakash Surya * the headers from being destroyed via arc_evict_state(). 3688244781f1SPrakash Surya */ 3689244781f1SPrakash Surya data_mls = multilist_sublist_lock(data_ml, data_idx); 3690244781f1SPrakash Surya meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3691244781f1SPrakash Surya 3692244781f1SPrakash Surya /* 3693244781f1SPrakash Surya * These two loops are to ensure we skip any markers that 3694244781f1SPrakash Surya * might be at the tail of the lists due to arc_evict_state(). 3695244781f1SPrakash Surya */ 3696244781f1SPrakash Surya 3697244781f1SPrakash Surya for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3698244781f1SPrakash Surya data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3699244781f1SPrakash Surya if (data_hdr->b_spa != 0) 3700244781f1SPrakash Surya break; 3701244781f1SPrakash Surya } 3702244781f1SPrakash Surya 3703244781f1SPrakash Surya for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3704244781f1SPrakash Surya meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3705244781f1SPrakash Surya if (meta_hdr->b_spa != 0) 3706244781f1SPrakash Surya break; 3707244781f1SPrakash Surya } 3708244781f1SPrakash Surya 3709244781f1SPrakash Surya if (data_hdr == NULL && meta_hdr == NULL) { 3710244781f1SPrakash Surya type = ARC_BUFC_DATA; 3711244781f1SPrakash Surya } else if (data_hdr == NULL) { 3712244781f1SPrakash Surya ASSERT3P(meta_hdr, !=, NULL); 3713244781f1SPrakash Surya type = ARC_BUFC_METADATA; 3714244781f1SPrakash Surya } else if (meta_hdr == NULL) { 3715244781f1SPrakash Surya ASSERT3P(data_hdr, !=, NULL); 3716244781f1SPrakash Surya type = ARC_BUFC_DATA; 3717244781f1SPrakash Surya } else { 3718244781f1SPrakash Surya ASSERT3P(data_hdr, !=, NULL); 3719244781f1SPrakash Surya ASSERT3P(meta_hdr, !=, NULL); 3720244781f1SPrakash Surya 3721244781f1SPrakash Surya /* The headers can't be on the sublist without an L1 header */ 3722244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(data_hdr)); 3723244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3724244781f1SPrakash Surya 3725244781f1SPrakash Surya if (data_hdr->b_l1hdr.b_arc_access < 3726244781f1SPrakash Surya meta_hdr->b_l1hdr.b_arc_access) { 3727244781f1SPrakash Surya type = ARC_BUFC_DATA; 3728244781f1SPrakash Surya } else { 3729244781f1SPrakash Surya type = ARC_BUFC_METADATA; 3730244781f1SPrakash Surya } 3731244781f1SPrakash Surya } 3732244781f1SPrakash Surya 3733244781f1SPrakash Surya multilist_sublist_unlock(meta_mls); 3734244781f1SPrakash Surya multilist_sublist_unlock(data_mls); 3735244781f1SPrakash Surya 3736244781f1SPrakash Surya return (type); 3737244781f1SPrakash Surya } 3738244781f1SPrakash Surya 3739244781f1SPrakash Surya /* 3740244781f1SPrakash Surya * Evict buffers from the cache, such that arc_size is capped by arc_c. 3741244781f1SPrakash Surya */ 3742244781f1SPrakash Surya static uint64_t 3743fa9e4066Sahrens arc_adjust(void) 3744fa9e4066Sahrens { 3745244781f1SPrakash Surya uint64_t total_evicted = 0; 3746244781f1SPrakash Surya uint64_t bytes; 3747244781f1SPrakash Surya int64_t target; 37483a2d8a1bSPaul Dagnelie uint64_t asize = aggsum_value(&arc_size); 37493a2d8a1bSPaul Dagnelie uint64_t ameta = aggsum_value(&arc_meta_used); 3750fa9e4066Sahrens 37515a98e54bSBrendan Gregg - Sun Microsystems /* 3752244781f1SPrakash Surya * If we're over arc_meta_limit, we want to correct that before 3753244781f1SPrakash Surya * potentially evicting data buffers below. 37545a98e54bSBrendan Gregg - Sun Microsystems */ 37553a2d8a1bSPaul Dagnelie total_evicted += arc_adjust_meta(ameta); 37565a98e54bSBrendan Gregg - Sun Microsystems 3757244781f1SPrakash Surya /* 3758244781f1SPrakash Surya * Adjust MRU size 3759244781f1SPrakash Surya * 3760244781f1SPrakash Surya * If we're over the target cache size, we want to evict enough 3761244781f1SPrakash Surya * from the list to get back to our target size. We don't want 3762244781f1SPrakash Surya * to evict too much from the MRU, such that it drops below 3763244781f1SPrakash Surya * arc_p. So, if we're over our target cache size more than 3764244781f1SPrakash Surya * the MRU is over arc_p, we'll evict enough to get back to 3765244781f1SPrakash Surya * arc_p here, and then evict more from the MFU below. 3766244781f1SPrakash Surya */ 37673a2d8a1bSPaul Dagnelie target = MIN((int64_t)(asize - arc_c), 37682fd872a7SPrakash Surya (int64_t)(refcount_count(&arc_anon->arcs_size) + 37693a2d8a1bSPaul Dagnelie refcount_count(&arc_mru->arcs_size) + ameta - arc_p)); 3770fa9e4066Sahrens 3771244781f1SPrakash Surya /* 3772244781f1SPrakash Surya * If we're below arc_meta_min, always prefer to evict data. 3773244781f1SPrakash Surya * Otherwise, try to satisfy the requested number of bytes to 3774244781f1SPrakash Surya * evict from the type which contains older buffers; in an 3775244781f1SPrakash Surya * effort to keep newer buffers in the cache regardless of their 3776244781f1SPrakash Surya * type. If we cannot satisfy the number of bytes from this 3777244781f1SPrakash Surya * type, spill over into the next type. 3778244781f1SPrakash Surya */ 3779244781f1SPrakash Surya if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 37803a2d8a1bSPaul Dagnelie ameta > arc_meta_min) { 3781244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3782244781f1SPrakash Surya total_evicted += bytes; 37830e8c6158Smaybee 3784244781f1SPrakash Surya /* 3785244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3786244781f1SPrakash Surya * metadata, we try to get the rest from data. 3787244781f1SPrakash Surya */ 3788244781f1SPrakash Surya target -= bytes; 3789244781f1SPrakash Surya 3790244781f1SPrakash Surya total_evicted += 3791244781f1SPrakash Surya arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3792244781f1SPrakash Surya } else { 3793244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3794244781f1SPrakash Surya total_evicted += bytes; 3795244781f1SPrakash Surya 3796244781f1SPrakash Surya /* 3797244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3798244781f1SPrakash Surya * data, we try to get the rest from metadata. 3799244781f1SPrakash Surya */ 3800244781f1SPrakash Surya target -= bytes; 3801244781f1SPrakash Surya 3802244781f1SPrakash Surya total_evicted += 3803244781f1SPrakash Surya arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3804fa9e4066Sahrens } 3805fa9e4066Sahrens 38065a98e54bSBrendan Gregg - Sun Microsystems /* 38075a98e54bSBrendan Gregg - Sun Microsystems * Adjust MFU size 3808244781f1SPrakash Surya * 3809244781f1SPrakash Surya * Now that we've tried to evict enough from the MRU to get its 3810244781f1SPrakash Surya * size back to arc_p, if we're still above the target cache 3811244781f1SPrakash Surya * size, we evict the rest from the MFU. 38125a98e54bSBrendan Gregg - Sun Microsystems */ 38133a2d8a1bSPaul Dagnelie target = asize - arc_c; 3814fa9e4066Sahrens 381531c46cf2SAlek Pinchuk if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 38163a2d8a1bSPaul Dagnelie ameta > arc_meta_min) { 3817244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3818244781f1SPrakash Surya total_evicted += bytes; 38195a98e54bSBrendan Gregg - Sun Microsystems 3820244781f1SPrakash Surya /* 3821244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3822244781f1SPrakash Surya * metadata, we try to get the rest from data. 3823244781f1SPrakash Surya */ 3824244781f1SPrakash Surya target -= bytes; 3825244781f1SPrakash Surya 3826244781f1SPrakash Surya total_evicted += 3827244781f1SPrakash Surya arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3828244781f1SPrakash Surya } else { 3829244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3830244781f1SPrakash Surya total_evicted += bytes; 3831244781f1SPrakash Surya 3832244781f1SPrakash Surya /* 3833244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3834244781f1SPrakash Surya * data, we try to get the rest from data. 3835244781f1SPrakash Surya */ 3836244781f1SPrakash Surya target -= bytes; 3837fa9e4066Sahrens 3838244781f1SPrakash Surya total_evicted += 3839244781f1SPrakash Surya arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 38405a98e54bSBrendan Gregg - Sun Microsystems } 3841fa9e4066Sahrens 38425a98e54bSBrendan Gregg - Sun Microsystems /* 38435a98e54bSBrendan Gregg - Sun Microsystems * Adjust ghost lists 3844244781f1SPrakash Surya * 3845244781f1SPrakash Surya * In addition to the above, the ARC also defines target values 3846244781f1SPrakash Surya * for the ghost lists. The sum of the mru list and mru ghost 3847244781f1SPrakash Surya * list should never exceed the target size of the cache, and 3848244781f1SPrakash Surya * the sum of the mru list, mfu list, mru ghost list, and mfu 3849244781f1SPrakash Surya * ghost list should never exceed twice the target size of the 3850244781f1SPrakash Surya * cache. The following logic enforces these limits on the ghost 3851244781f1SPrakash Surya * caches, and evicts from them as needed. 38525a98e54bSBrendan Gregg - Sun Microsystems */ 38532fd872a7SPrakash Surya target = refcount_count(&arc_mru->arcs_size) + 38542fd872a7SPrakash Surya refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3855fa9e4066Sahrens 3856244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3857244781f1SPrakash Surya total_evicted += bytes; 3858fa9e4066Sahrens 3859244781f1SPrakash Surya target -= bytes; 38600e8c6158Smaybee 3861244781f1SPrakash Surya total_evicted += 3862244781f1SPrakash Surya arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 38635a98e54bSBrendan Gregg - Sun Microsystems 3864244781f1SPrakash Surya /* 3865244781f1SPrakash Surya * We assume the sum of the mru list and mfu list is less than 3866244781f1SPrakash Surya * or equal to arc_c (we enforced this above), which means we 3867244781f1SPrakash Surya * can use the simpler of the two equations below: 3868244781f1SPrakash Surya * 3869244781f1SPrakash Surya * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3870244781f1SPrakash Surya * mru ghost + mfu ghost <= arc_c 3871244781f1SPrakash Surya */ 38722fd872a7SPrakash Surya target = refcount_count(&arc_mru_ghost->arcs_size) + 38732fd872a7SPrakash Surya refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3874244781f1SPrakash Surya 3875244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3876244781f1SPrakash Surya total_evicted += bytes; 3877244781f1SPrakash Surya 3878244781f1SPrakash Surya target -= bytes; 3879244781f1SPrakash Surya 3880244781f1SPrakash Surya total_evicted += 3881244781f1SPrakash Surya arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3882244781f1SPrakash Surya 3883244781f1SPrakash Surya return (total_evicted); 3884fa9e4066Sahrens } 3885fa9e4066Sahrens 3886fa9e4066Sahrens void 3887244781f1SPrakash Surya arc_flush(spa_t *spa, boolean_t retry) 3888fa9e4066Sahrens { 3889ac05c741SMark Maybee uint64_t guid = 0; 3890ac05c741SMark Maybee 3891244781f1SPrakash Surya /* 3892dcbf3bd6SGeorge Wilson * If retry is B_TRUE, a spa must not be specified since we have 3893244781f1SPrakash Surya * no good way to determine if all of a spa's buffers have been 3894244781f1SPrakash Surya * evicted from an arc state. 3895244781f1SPrakash Surya */ 3896244781f1SPrakash Surya ASSERT(!retry || spa == 0); 3897244781f1SPrakash Surya 389889c86e32SChris Williamson if (spa != NULL) 3899e9103aaeSGarrett D'Amore guid = spa_load_guid(spa); 3900ac05c741SMark Maybee 3901244781f1SPrakash Surya (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3902244781f1SPrakash Surya (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3903244781f1SPrakash Surya 3904244781f1SPrakash Surya (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3905244781f1SPrakash Surya (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3906874395d5Smaybee 3907244781f1SPrakash Surya (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3908244781f1SPrakash Surya (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3909244781f1SPrakash Surya 3910244781f1SPrakash Surya (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3911244781f1SPrakash Surya (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3912fa9e4066Sahrens } 3913fa9e4066Sahrens 3914de753e34SBrad Lewis static void 3915de753e34SBrad Lewis arc_reduce_target_size(int64_t to_free) 3916fa9e4066Sahrens { 39173a2d8a1bSPaul Dagnelie uint64_t asize = aggsum_value(&arc_size); 391844cb6abcSbmc if (arc_c > arc_c_min) { 3919fa9e4066Sahrens 392044cb6abcSbmc if (arc_c > arc_c_min + to_free) 392144cb6abcSbmc atomic_add_64(&arc_c, -to_free); 392249e3519aSmaybee else 392344cb6abcSbmc arc_c = arc_c_min; 392444cb6abcSbmc 392544cb6abcSbmc atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 39263a2d8a1bSPaul Dagnelie if (asize < arc_c) 39273a2d8a1bSPaul Dagnelie arc_c = MAX(asize, arc_c_min); 392844cb6abcSbmc if (arc_p > arc_c) 392944cb6abcSbmc arc_p = (arc_c >> 1); 393044cb6abcSbmc ASSERT(arc_c >= arc_c_min); 393144cb6abcSbmc ASSERT((int64_t)arc_p >= 0); 393249e3519aSmaybee } 3933fa9e4066Sahrens 3934de753e34SBrad Lewis if (asize > arc_c) { 3935de753e34SBrad Lewis /* See comment in arc_adjust_cb_check() on why lock+flag */ 3936de753e34SBrad Lewis mutex_enter(&arc_adjust_lock); 3937de753e34SBrad Lewis arc_adjust_needed = B_TRUE; 3938de753e34SBrad Lewis mutex_exit(&arc_adjust_lock); 3939de753e34SBrad Lewis zthr_wakeup(arc_adjust_zthr); 3940de753e34SBrad Lewis } 3941fa9e4066Sahrens } 3942fa9e4066Sahrens 39432ec99e3eSMatthew Ahrens typedef enum free_memory_reason_t { 39442ec99e3eSMatthew Ahrens FMR_UNKNOWN, 39452ec99e3eSMatthew Ahrens FMR_NEEDFREE, 39462ec99e3eSMatthew Ahrens FMR_LOTSFREE, 39472ec99e3eSMatthew Ahrens FMR_SWAPFS_MINFREE, 39482ec99e3eSMatthew Ahrens FMR_PAGES_PP_MAXIMUM, 39492ec99e3eSMatthew Ahrens FMR_HEAP_ARENA, 39502ec99e3eSMatthew Ahrens FMR_ZIO_ARENA, 39512ec99e3eSMatthew Ahrens } free_memory_reason_t; 39522ec99e3eSMatthew Ahrens 39532ec99e3eSMatthew Ahrens int64_t last_free_memory; 39542ec99e3eSMatthew Ahrens free_memory_reason_t last_free_reason; 39552ec99e3eSMatthew Ahrens 395694dd93aeSGeorge Wilson /* 39572ec99e3eSMatthew Ahrens * Additional reserve of pages for pp_reserve. 395894dd93aeSGeorge Wilson */ 39592ec99e3eSMatthew Ahrens int64_t arc_pages_pp_reserve = 64; 3960fa9e4066Sahrens 39612ec99e3eSMatthew Ahrens /* 39622ec99e3eSMatthew Ahrens * Additional reserve of pages for swapfs. 39632ec99e3eSMatthew Ahrens */ 39642ec99e3eSMatthew Ahrens int64_t arc_swapfs_reserve = 64; 39653cff2f43Sstans 39662ec99e3eSMatthew Ahrens /* 39672ec99e3eSMatthew Ahrens * Return the amount of memory that can be consumed before reclaim will be 39682ec99e3eSMatthew Ahrens * needed. Positive if there is sufficient free memory, negative indicates 39692ec99e3eSMatthew Ahrens * the amount of memory that needs to be freed up. 39702ec99e3eSMatthew Ahrens */ 39712ec99e3eSMatthew Ahrens static int64_t 39722ec99e3eSMatthew Ahrens arc_available_memory(void) 39732ec99e3eSMatthew Ahrens { 39742ec99e3eSMatthew Ahrens int64_t lowest = INT64_MAX; 39752ec99e3eSMatthew Ahrens int64_t n; 39762ec99e3eSMatthew Ahrens free_memory_reason_t r = FMR_UNKNOWN; 39773cff2f43Sstans 39782ec99e3eSMatthew Ahrens #ifdef _KERNEL 39792ec99e3eSMatthew Ahrens if (needfree > 0) { 39802ec99e3eSMatthew Ahrens n = PAGESIZE * (-needfree); 39812ec99e3eSMatthew Ahrens if (n < lowest) { 39822ec99e3eSMatthew Ahrens lowest = n; 39832ec99e3eSMatthew Ahrens r = FMR_NEEDFREE; 39842ec99e3eSMatthew Ahrens } 39852ec99e3eSMatthew Ahrens } 3986fa9e4066Sahrens 3987fa9e4066Sahrens /* 3988fa9e4066Sahrens * check that we're out of range of the pageout scanner. It starts to 3989fa9e4066Sahrens * schedule paging if freemem is less than lotsfree and needfree. 3990fa9e4066Sahrens * lotsfree is the high-water mark for pageout, and needfree is the 3991fa9e4066Sahrens * number of needed free pages. We add extra pages here to make sure 3992fa9e4066Sahrens * the scanner doesn't start up while we're freeing memory. 3993fa9e4066Sahrens */ 39942ec99e3eSMatthew Ahrens n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 39952ec99e3eSMatthew Ahrens if (n < lowest) { 39962ec99e3eSMatthew Ahrens lowest = n; 39972ec99e3eSMatthew Ahrens r = FMR_LOTSFREE; 39982ec99e3eSMatthew Ahrens } 3999fa9e4066Sahrens 4000fa9e4066Sahrens /* 4001fa9e4066Sahrens * check to make sure that swapfs has enough space so that anon 4002fa94a07fSbrendan * reservations can still succeed. anon_resvmem() checks that the 4003fa9e4066Sahrens * availrmem is greater than swapfs_minfree, and the number of reserved 4004fa9e4066Sahrens * swap pages. We also add a bit of extra here just to prevent 4005fa9e4066Sahrens * circumstances from getting really dire. 4006fa9e4066Sahrens */ 40072ec99e3eSMatthew Ahrens n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 40082ec99e3eSMatthew Ahrens desfree - arc_swapfs_reserve); 40092ec99e3eSMatthew Ahrens if (n < lowest) { 40102ec99e3eSMatthew Ahrens lowest = n; 40112ec99e3eSMatthew Ahrens r = FMR_SWAPFS_MINFREE; 40122ec99e3eSMatthew Ahrens } 40132ec99e3eSMatthew Ahrens 4014fa9e4066Sahrens 4015cf746768SBryan Cantrill /* 4016cf746768SBryan Cantrill * Check that we have enough availrmem that memory locking (e.g., via 4017cf746768SBryan Cantrill * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 4018cf746768SBryan Cantrill * stores the number of pages that cannot be locked; when availrmem 4019cf746768SBryan Cantrill * drops below pages_pp_maximum, page locking mechanisms such as 4020cf746768SBryan Cantrill * page_pp_lock() will fail.) 4021cf746768SBryan Cantrill */ 40222ec99e3eSMatthew Ahrens n = PAGESIZE * (availrmem - pages_pp_maximum - 40232ec99e3eSMatthew Ahrens arc_pages_pp_reserve); 40242ec99e3eSMatthew Ahrens if (n < lowest) { 40252ec99e3eSMatthew Ahrens lowest = n; 40262ec99e3eSMatthew Ahrens r = FMR_PAGES_PP_MAXIMUM; 40272ec99e3eSMatthew Ahrens } 4028cf746768SBryan Cantrill 40295dc8af33Smaybee #if defined(__i386) 4030fa9e4066Sahrens /* 4031fa9e4066Sahrens * If we're on an i386 platform, it's possible that we'll exhaust the 4032fa9e4066Sahrens * kernel heap space before we ever run out of available physical 4033fa9e4066Sahrens * memory. Most checks of the size of the heap_area compare against 4034fa9e4066Sahrens * tune.t_minarmem, which is the minimum available real memory that we 4035fa9e4066Sahrens * can have in the system. However, this is generally fixed at 25 pages 4036fa9e4066Sahrens * which is so low that it's useless. In this comparison, we seek to 4037fa9e4066Sahrens * calculate the total heap-size, and reclaim if more than 3/4ths of the 4038fa94a07fSbrendan * heap is allocated. (Or, in the calculation, if less than 1/4th is 4039fa9e4066Sahrens * free) 4040fa9e4066Sahrens */ 40410dd053d7SPrakash Surya n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 40422ec99e3eSMatthew Ahrens (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 40432ec99e3eSMatthew Ahrens if (n < lowest) { 40442ec99e3eSMatthew Ahrens lowest = n; 40452ec99e3eSMatthew Ahrens r = FMR_HEAP_ARENA; 40462ec99e3eSMatthew Ahrens } 4047fa9e4066Sahrens #endif 4048fa9e4066Sahrens 404994dd93aeSGeorge Wilson /* 405094dd93aeSGeorge Wilson * If zio data pages are being allocated out of a separate heap segment, 405194dd93aeSGeorge Wilson * then enforce that the size of available vmem for this arena remains 40520dd053d7SPrakash Surya * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free. 405394dd93aeSGeorge Wilson * 40540dd053d7SPrakash Surya * Note that reducing the arc_zio_arena_free_shift keeps more virtual 40550dd053d7SPrakash Surya * memory (in the zio_arena) free, which can avoid memory 40560dd053d7SPrakash Surya * fragmentation issues. 405794dd93aeSGeorge Wilson */ 40582ec99e3eSMatthew Ahrens if (zio_arena != NULL) { 40590dd053d7SPrakash Surya n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 40600dd053d7SPrakash Surya (vmem_size(zio_arena, VMEM_ALLOC) >> 40610dd053d7SPrakash Surya arc_zio_arena_free_shift); 40622ec99e3eSMatthew Ahrens if (n < lowest) { 40632ec99e3eSMatthew Ahrens lowest = n; 40642ec99e3eSMatthew Ahrens r = FMR_ZIO_ARENA; 40652ec99e3eSMatthew Ahrens } 40662ec99e3eSMatthew Ahrens } 4067fa9e4066Sahrens #else 40682ec99e3eSMatthew Ahrens /* Every 100 calls, free a small amount */ 4069fa9e4066Sahrens if (spa_get_random(100) == 0) 40702ec99e3eSMatthew Ahrens lowest = -1024; 4071fa9e4066Sahrens #endif 40722ec99e3eSMatthew Ahrens 40732ec99e3eSMatthew Ahrens last_free_memory = lowest; 40742ec99e3eSMatthew Ahrens last_free_reason = r; 40752ec99e3eSMatthew Ahrens 40762ec99e3eSMatthew Ahrens return (lowest); 40772ec99e3eSMatthew Ahrens } 40782ec99e3eSMatthew Ahrens 40792ec99e3eSMatthew Ahrens 40802ec99e3eSMatthew Ahrens /* 40812ec99e3eSMatthew Ahrens * Determine if the system is under memory pressure and is asking 4082dcbf3bd6SGeorge Wilson * to reclaim memory. A return value of B_TRUE indicates that the system 40832ec99e3eSMatthew Ahrens * is under memory pressure and that the arc should adjust accordingly. 40842ec99e3eSMatthew Ahrens */ 40852ec99e3eSMatthew Ahrens static boolean_t 40862ec99e3eSMatthew Ahrens arc_reclaim_needed(void) 40872ec99e3eSMatthew Ahrens { 40882ec99e3eSMatthew Ahrens return (arc_available_memory() < 0); 4089fa9e4066Sahrens } 4090fa9e4066Sahrens 4091fa9e4066Sahrens static void 4092de753e34SBrad Lewis arc_kmem_reap_soon(void) 4093fa9e4066Sahrens { 4094fa9e4066Sahrens size_t i; 4095fa9e4066Sahrens kmem_cache_t *prev_cache = NULL; 4096ad23a2dbSjohansen kmem_cache_t *prev_data_cache = NULL; 4097fa9e4066Sahrens extern kmem_cache_t *zio_buf_cache[]; 4098ad23a2dbSjohansen extern kmem_cache_t *zio_data_buf_cache[]; 409983803b51SGeorge Wilson extern kmem_cache_t *range_seg_cache; 4100770499e1SDan Kimmel extern kmem_cache_t *abd_chunk_cache; 4101fa9e4066Sahrens 4102033f9833Sek #ifdef _KERNEL 41033a2d8a1bSPaul Dagnelie if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) { 41040e8c6158Smaybee /* 41050e8c6158Smaybee * We are exceeding our meta-data cache limit. 41060e8c6158Smaybee * Purge some DNLC entries to release holds on meta-data. 41070e8c6158Smaybee */ 41080e8c6158Smaybee dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 41090e8c6158Smaybee } 41105dc8af33Smaybee #if defined(__i386) 41115dc8af33Smaybee /* 41125dc8af33Smaybee * Reclaim unused memory from all kmem caches. 41135dc8af33Smaybee */ 41145dc8af33Smaybee kmem_reap(); 41155dc8af33Smaybee #endif 4116033f9833Sek #endif 4117033f9833Sek 4118fa9e4066Sahrens for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 4119fa9e4066Sahrens if (zio_buf_cache[i] != prev_cache) { 4120fa9e4066Sahrens prev_cache = zio_buf_cache[i]; 412136a64e62STim Kordas kmem_cache_reap_soon(zio_buf_cache[i]); 4122fa9e4066Sahrens } 4123ad23a2dbSjohansen if (zio_data_buf_cache[i] != prev_data_cache) { 4124ad23a2dbSjohansen prev_data_cache = zio_data_buf_cache[i]; 412536a64e62STim Kordas kmem_cache_reap_soon(zio_data_buf_cache[i]); 4126ad23a2dbSjohansen } 4127fa9e4066Sahrens } 412836a64e62STim Kordas kmem_cache_reap_soon(abd_chunk_cache); 412936a64e62STim Kordas kmem_cache_reap_soon(buf_cache); 413036a64e62STim Kordas kmem_cache_reap_soon(hdr_full_cache); 413136a64e62STim Kordas kmem_cache_reap_soon(hdr_l2only_cache); 413236a64e62STim Kordas kmem_cache_reap_soon(range_seg_cache); 413394dd93aeSGeorge Wilson 41342ec99e3eSMatthew Ahrens if (zio_arena != NULL) { 41352ec99e3eSMatthew Ahrens /* 41362ec99e3eSMatthew Ahrens * Ask the vmem arena to reclaim unused memory from its 41372ec99e3eSMatthew Ahrens * quantum caches. 41382ec99e3eSMatthew Ahrens */ 413994dd93aeSGeorge Wilson vmem_qcache_reap(zio_arena); 41402ec99e3eSMatthew Ahrens } 4141fa9e4066Sahrens } 4142fa9e4066Sahrens 4143de753e34SBrad Lewis /* ARGSUSED */ 4144de753e34SBrad Lewis static boolean_t 4145de753e34SBrad Lewis arc_adjust_cb_check(void *arg, zthr_t *zthr) 4146de753e34SBrad Lewis { 4147de753e34SBrad Lewis /* 4148de753e34SBrad Lewis * This is necessary in order for the mdb ::arc dcmd to 4149de753e34SBrad Lewis * show up to date information. Since the ::arc command 4150de753e34SBrad Lewis * does not call the kstat's update function, without 4151de753e34SBrad Lewis * this call, the command may show stale stats for the 4152de753e34SBrad Lewis * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 4153de753e34SBrad Lewis * with this change, the data might be up to 1 second 4154de753e34SBrad Lewis * out of date(the arc_adjust_zthr has a maximum sleep 4155de753e34SBrad Lewis * time of 1 second); but that should suffice. The 4156de753e34SBrad Lewis * arc_state_t structures can be queried directly if more 4157de753e34SBrad Lewis * accurate information is needed. 4158de753e34SBrad Lewis */ 4159de753e34SBrad Lewis if (arc_ksp != NULL) 4160de753e34SBrad Lewis arc_ksp->ks_update(arc_ksp, KSTAT_READ); 4161de753e34SBrad Lewis 4162de753e34SBrad Lewis /* 4163de753e34SBrad Lewis * We have to rely on arc_get_data_impl() to tell us when to adjust, 4164de753e34SBrad Lewis * rather than checking if we are overflowing here, so that we are 4165de753e34SBrad Lewis * sure to not leave arc_get_data_impl() waiting on 4166de753e34SBrad Lewis * arc_adjust_waiters_cv. If we have become "not overflowing" since 4167de753e34SBrad Lewis * arc_get_data_impl() checked, we need to wake it up. We could 4168de753e34SBrad Lewis * broadcast the CV here, but arc_get_data_impl() may have not yet 4169de753e34SBrad Lewis * gone to sleep. We would need to use a mutex to ensure that this 4170de753e34SBrad Lewis * function doesn't broadcast until arc_get_data_impl() has gone to 4171de753e34SBrad Lewis * sleep (e.g. the arc_adjust_lock). However, the lock ordering of 4172de753e34SBrad Lewis * such a lock would necessarily be incorrect with respect to the 4173de753e34SBrad Lewis * zthr_lock, which is held before this function is called, and is 4174de753e34SBrad Lewis * held by arc_get_data_impl() when it calls zthr_wakeup(). 4175de753e34SBrad Lewis */ 4176de753e34SBrad Lewis return (arc_adjust_needed); 4177de753e34SBrad Lewis } 4178de753e34SBrad Lewis 4179244781f1SPrakash Surya /* 4180de753e34SBrad Lewis * Keep arc_size under arc_c by running arc_adjust which evicts data 4181de753e34SBrad Lewis * from the ARC. 4182244781f1SPrakash Surya */ 41833f7978d0SAlan Somers /* ARGSUSED */ 4184de753e34SBrad Lewis static int 4185de753e34SBrad Lewis arc_adjust_cb(void *arg, zthr_t *zthr) 4186fa9e4066Sahrens { 4187de753e34SBrad Lewis uint64_t evicted = 0; 4188fa9e4066Sahrens 4189de753e34SBrad Lewis /* Evict from cache */ 4190de753e34SBrad Lewis evicted = arc_adjust(); 4191244781f1SPrakash Surya 4192de753e34SBrad Lewis /* 4193de753e34SBrad Lewis * If evicted is zero, we couldn't evict anything 4194de753e34SBrad Lewis * via arc_adjust(). This could be due to hash lock 4195de753e34SBrad Lewis * collisions, but more likely due to the majority of 4196de753e34SBrad Lewis * arc buffers being unevictable. Therefore, even if 4197de753e34SBrad Lewis * arc_size is above arc_c, another pass is unlikely to 4198de753e34SBrad Lewis * be helpful and could potentially cause us to enter an 4199de753e34SBrad Lewis * infinite loop. Additionally, zthr_iscancelled() is 4200de753e34SBrad Lewis * checked here so that if the arc is shutting down, the 4201de753e34SBrad Lewis * broadcast will wake any remaining arc adjust waiters. 4202de753e34SBrad Lewis */ 4203de753e34SBrad Lewis mutex_enter(&arc_adjust_lock); 4204de753e34SBrad Lewis arc_adjust_needed = !zthr_iscancelled(arc_adjust_zthr) && 4205de753e34SBrad Lewis evicted > 0 && aggsum_compare(&arc_size, arc_c) > 0; 4206de753e34SBrad Lewis if (!arc_adjust_needed) { 4207dcbf3bd6SGeorge Wilson /* 4208de753e34SBrad Lewis * We're either no longer overflowing, or we 4209de753e34SBrad Lewis * can't evict anything more, so we should wake 4210de753e34SBrad Lewis * up any waiters. 4211dcbf3bd6SGeorge Wilson */ 4212de753e34SBrad Lewis cv_broadcast(&arc_adjust_waiters_cv); 4213de753e34SBrad Lewis } 4214de753e34SBrad Lewis mutex_exit(&arc_adjust_lock); 4215dcbf3bd6SGeorge Wilson 4216de753e34SBrad Lewis return (0); 4217de753e34SBrad Lewis } 4218244781f1SPrakash Surya 4219de753e34SBrad Lewis /* ARGSUSED */ 4220de753e34SBrad Lewis static boolean_t 4221de753e34SBrad Lewis arc_reap_cb_check(void *arg, zthr_t *zthr) 4222de753e34SBrad Lewis { 4223de753e34SBrad Lewis int64_t free_memory = arc_available_memory(); 4224de753e34SBrad Lewis 4225de753e34SBrad Lewis /* 4226de753e34SBrad Lewis * If a kmem reap is already active, don't schedule more. We must 4227de753e34SBrad Lewis * check for this because kmem_cache_reap_soon() won't actually 4228de753e34SBrad Lewis * block on the cache being reaped (this is to prevent callers from 4229de753e34SBrad Lewis * becoming implicitly blocked by a system-wide kmem reap -- which, 4230de753e34SBrad Lewis * on a system with many, many full magazines, can take minutes). 4231de753e34SBrad Lewis */ 4232de753e34SBrad Lewis if (!kmem_cache_reap_active() && 4233de753e34SBrad Lewis free_memory < 0) { 4234de753e34SBrad Lewis arc_no_grow = B_TRUE; 4235de753e34SBrad Lewis arc_warm = B_TRUE; 4236405a5a0fSMatthew Ahrens /* 4237de753e34SBrad Lewis * Wait at least zfs_grow_retry (default 60) seconds 4238de753e34SBrad Lewis * before considering growing. 4239405a5a0fSMatthew Ahrens */ 4240de753e34SBrad Lewis arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry); 4241de753e34SBrad Lewis return (B_TRUE); 4242de753e34SBrad Lewis } else if (free_memory < arc_c >> arc_no_grow_shift) { 4243de753e34SBrad Lewis arc_no_grow = B_TRUE; 4244de753e34SBrad Lewis } else if (gethrtime() >= arc_growtime) { 4245de753e34SBrad Lewis arc_no_grow = B_FALSE; 4246de753e34SBrad Lewis } 4247405a5a0fSMatthew Ahrens 4248de753e34SBrad Lewis return (B_FALSE); 4249de753e34SBrad Lewis } 4250fa9e4066Sahrens 4251de753e34SBrad Lewis /* 4252de753e34SBrad Lewis * Keep enough free memory in the system by reaping the ARC's kmem 4253de753e34SBrad Lewis * caches. To cause more slabs to be reapable, we may reduce the 4254de753e34SBrad Lewis * target size of the cache (arc_c), causing the arc_adjust_cb() 4255de753e34SBrad Lewis * to free more buffers. 4256de753e34SBrad Lewis */ 4257de753e34SBrad Lewis /* ARGSUSED */ 4258de753e34SBrad Lewis static int 4259de753e34SBrad Lewis arc_reap_cb(void *arg, zthr_t *zthr) 4260de753e34SBrad Lewis { 4261de753e34SBrad Lewis int64_t free_memory; 4262fa9e4066Sahrens 4263de753e34SBrad Lewis /* 4264de753e34SBrad Lewis * Kick off asynchronous kmem_reap()'s of all our caches. 4265de753e34SBrad Lewis */ 4266de753e34SBrad Lewis arc_kmem_reap_soon(); 4267fa9e4066Sahrens 4268de753e34SBrad Lewis /* 4269de753e34SBrad Lewis * Wait at least arc_kmem_cache_reap_retry_ms between 4270de753e34SBrad Lewis * arc_kmem_reap_soon() calls. Without this check it is possible to 4271de753e34SBrad Lewis * end up in a situation where we spend lots of time reaping 4272de753e34SBrad Lewis * caches, while we're near arc_c_min. Waiting here also gives the 4273de753e34SBrad Lewis * subsequent free memory check a chance of finding that the 4274de753e34SBrad Lewis * asynchronous reap has already freed enough memory, and we don't 4275de753e34SBrad Lewis * need to call arc_reduce_target_size(). 4276de753e34SBrad Lewis */ 4277de753e34SBrad Lewis delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000); 4278de753e34SBrad Lewis 4279de753e34SBrad Lewis /* 4280de753e34SBrad Lewis * Reduce the target size as needed to maintain the amount of free 4281de753e34SBrad Lewis * memory in the system at a fraction of the arc_size (1/128th by 4282de753e34SBrad Lewis * default). If oversubscribed (free_memory < 0) then reduce the 4283de753e34SBrad Lewis * target arc_size by the deficit amount plus the fractional 4284de753e34SBrad Lewis * amount. If free memory is positive but less then the fractional 4285de753e34SBrad Lewis * amount, reduce by what is needed to hit the fractional amount. 4286de753e34SBrad Lewis */ 4287de753e34SBrad Lewis free_memory = arc_available_memory(); 42882ec99e3eSMatthew Ahrens 4289de753e34SBrad Lewis int64_t to_free = 4290de753e34SBrad Lewis (arc_c >> arc_shrink_shift) - free_memory; 4291de753e34SBrad Lewis if (to_free > 0) { 42922ec99e3eSMatthew Ahrens #ifdef _KERNEL 4293de753e34SBrad Lewis to_free = MAX(to_free, ptob(needfree)); 42942ec99e3eSMatthew Ahrens #endif 4295de753e34SBrad Lewis arc_reduce_target_size(to_free); 4296244781f1SPrakash Surya } 4297244781f1SPrakash Surya 4298de753e34SBrad Lewis return (0); 4299244781f1SPrakash Surya } 4300244781f1SPrakash Surya 4301ea8dc4b6Seschrock /* 4302ea8dc4b6Seschrock * Adapt arc info given the number of bytes we are trying to add and 4303ea8dc4b6Seschrock * the state that we are comming from. This function is only called 4304ea8dc4b6Seschrock * when we are adding new content to the cache. 4305ea8dc4b6Seschrock */ 4306fa9e4066Sahrens static void 4307ea8dc4b6Seschrock arc_adapt(int bytes, arc_state_t *state) 4308fa9e4066Sahrens { 4309ea8dc4b6Seschrock int mult; 43105a98e54bSBrendan Gregg - Sun Microsystems uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 43112fd872a7SPrakash Surya int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 43122fd872a7SPrakash Surya int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 4313ea8dc4b6Seschrock 4314fa94a07fSbrendan if (state == arc_l2c_only) 4315fa94a07fSbrendan return; 4316fa94a07fSbrendan 4317ea8dc4b6Seschrock ASSERT(bytes > 0); 4318fa9e4066Sahrens /* 4319ea8dc4b6Seschrock * Adapt the target size of the MRU list: 4320ea8dc4b6Seschrock * - if we just hit in the MRU ghost list, then increase 4321ea8dc4b6Seschrock * the target size of the MRU list. 4322ea8dc4b6Seschrock * - if we just hit in the MFU ghost list, then increase 4323ea8dc4b6Seschrock * the target size of the MFU list by decreasing the 4324ea8dc4b6Seschrock * target size of the MRU list. 4325fa9e4066Sahrens */ 432644cb6abcSbmc if (state == arc_mru_ghost) { 43272fd872a7SPrakash Surya mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 43283e4e8481STom Erickson mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 4329ea8dc4b6Seschrock 43305a98e54bSBrendan Gregg - Sun Microsystems arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 433144cb6abcSbmc } else if (state == arc_mfu_ghost) { 43325a98e54bSBrendan Gregg - Sun Microsystems uint64_t delta; 43335a98e54bSBrendan Gregg - Sun Microsystems 43342fd872a7SPrakash Surya mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 43353e4e8481STom Erickson mult = MIN(mult, 10); 4336ea8dc4b6Seschrock 43375a98e54bSBrendan Gregg - Sun Microsystems delta = MIN(bytes * mult, arc_p); 43385a98e54bSBrendan Gregg - Sun Microsystems arc_p = MAX(arc_p_min, arc_p - delta); 4339ea8dc4b6Seschrock } 434044cb6abcSbmc ASSERT((int64_t)arc_p >= 0); 4341fa9e4066Sahrens 4342de753e34SBrad Lewis /* 4343de753e34SBrad Lewis * Wake reap thread if we do not have any available memory 4344de753e34SBrad Lewis */ 4345fa9e4066Sahrens if (arc_reclaim_needed()) { 4346de753e34SBrad Lewis zthr_wakeup(arc_reap_zthr); 4347fa9e4066Sahrens return; 4348fa9e4066Sahrens } 4349fa9e4066Sahrens 4350de753e34SBrad Lewis 435144cb6abcSbmc if (arc_no_grow) 4352fa9e4066Sahrens return; 4353fa9e4066Sahrens 435444cb6abcSbmc if (arc_c >= arc_c_max) 4355ea8dc4b6Seschrock return; 4356ea8dc4b6Seschrock 4357fa9e4066Sahrens /* 4358ea8dc4b6Seschrock * If we're within (2 * maxblocksize) bytes of the target 4359ea8dc4b6Seschrock * cache size, increment the target cache size 4360fa9e4066Sahrens */ 43613a2d8a1bSPaul Dagnelie if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) > 43623a2d8a1bSPaul Dagnelie 0) { 436344cb6abcSbmc atomic_add_64(&arc_c, (int64_t)bytes); 436444cb6abcSbmc if (arc_c > arc_c_max) 436544cb6abcSbmc arc_c = arc_c_max; 436644cb6abcSbmc else if (state == arc_anon) 436744cb6abcSbmc atomic_add_64(&arc_p, (int64_t)bytes); 436844cb6abcSbmc if (arc_p > arc_c) 436944cb6abcSbmc arc_p = arc_c; 4370fa9e4066Sahrens } 437144cb6abcSbmc ASSERT((int64_t)arc_p >= 0); 4372fa9e4066Sahrens } 4373fa9e4066Sahrens 4374fa9e4066Sahrens /* 4375244781f1SPrakash Surya * Check if arc_size has grown past our upper threshold, determined by 4376244781f1SPrakash Surya * zfs_arc_overflow_shift. 4377fa9e4066Sahrens */ 4378244781f1SPrakash Surya static boolean_t 4379244781f1SPrakash Surya arc_is_overflowing(void) 4380fa9e4066Sahrens { 4381244781f1SPrakash Surya /* Always allow at least one block of overflow */ 4382244781f1SPrakash Surya uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 4383244781f1SPrakash Surya arc_c >> zfs_arc_overflow_shift); 4384fa9e4066Sahrens 43853a2d8a1bSPaul Dagnelie /* 43863a2d8a1bSPaul Dagnelie * We just compare the lower bound here for performance reasons. Our 43873a2d8a1bSPaul Dagnelie * primary goals are to make sure that the arc never grows without 43883a2d8a1bSPaul Dagnelie * bound, and that it can reach its maximum size. This check 43893a2d8a1bSPaul Dagnelie * accomplishes both goals. The maximum amount we could run over by is 43903a2d8a1bSPaul Dagnelie * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block 43913a2d8a1bSPaul Dagnelie * in the ARC. In practice, that's in the tens of MB, which is low 43923a2d8a1bSPaul Dagnelie * enough to be safe. 43933a2d8a1bSPaul Dagnelie */ 43943a2d8a1bSPaul Dagnelie return (aggsum_lower_bound(&arc_size) >= arc_c + overflow); 4395fa9e4066Sahrens } 4396fa9e4066Sahrens 4397770499e1SDan Kimmel static abd_t * 4398770499e1SDan Kimmel arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4399770499e1SDan Kimmel { 4400770499e1SDan Kimmel arc_buf_contents_t type = arc_buf_type(hdr); 4401770499e1SDan Kimmel 4402770499e1SDan Kimmel arc_get_data_impl(hdr, size, tag); 4403770499e1SDan Kimmel if (type == ARC_BUFC_METADATA) { 4404770499e1SDan Kimmel return (abd_alloc(size, B_TRUE)); 4405770499e1SDan Kimmel } else { 4406770499e1SDan Kimmel ASSERT(type == ARC_BUFC_DATA); 4407770499e1SDan Kimmel return (abd_alloc(size, B_FALSE)); 4408770499e1SDan Kimmel } 4409770499e1SDan Kimmel } 4410770499e1SDan Kimmel 4411770499e1SDan Kimmel static void * 4412770499e1SDan Kimmel arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4413770499e1SDan Kimmel { 4414770499e1SDan Kimmel arc_buf_contents_t type = arc_buf_type(hdr); 4415770499e1SDan Kimmel 4416770499e1SDan Kimmel arc_get_data_impl(hdr, size, tag); 4417770499e1SDan Kimmel if (type == ARC_BUFC_METADATA) { 4418770499e1SDan Kimmel return (zio_buf_alloc(size)); 4419770499e1SDan Kimmel } else { 4420770499e1SDan Kimmel ASSERT(type == ARC_BUFC_DATA); 4421770499e1SDan Kimmel return (zio_data_buf_alloc(size)); 4422770499e1SDan Kimmel } 4423770499e1SDan Kimmel } 4424770499e1SDan Kimmel 4425fa9e4066Sahrens /* 4426dcbf3bd6SGeorge Wilson * Allocate a block and return it to the caller. If we are hitting the 4427dcbf3bd6SGeorge Wilson * hard limit for the cache size, we must sleep, waiting for the eviction 4428dcbf3bd6SGeorge Wilson * thread to catch up. If we're past the target size but below the hard 4429dcbf3bd6SGeorge Wilson * limit, we'll only signal the reclaim thread and continue on. 4430fa9e4066Sahrens */ 4431770499e1SDan Kimmel static void 4432770499e1SDan Kimmel arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4433fa9e4066Sahrens { 4434770499e1SDan Kimmel arc_state_t *state = hdr->b_l1hdr.b_state; 4435770499e1SDan Kimmel arc_buf_contents_t type = arc_buf_type(hdr); 4436fa9e4066Sahrens 443744eda4d7Smaybee arc_adapt(size, state); 4438fa9e4066Sahrens 443944eda4d7Smaybee /* 4440244781f1SPrakash Surya * If arc_size is currently overflowing, and has grown past our 4441244781f1SPrakash Surya * upper limit, we must be adding data faster than the evict 4442244781f1SPrakash Surya * thread can evict. Thus, to ensure we don't compound the 4443244781f1SPrakash Surya * problem by adding more data and forcing arc_size to grow even 4444244781f1SPrakash Surya * further past it's target size, we halt and wait for the 4445244781f1SPrakash Surya * eviction thread to catch up. 4446244781f1SPrakash Surya * 4447244781f1SPrakash Surya * It's also possible that the reclaim thread is unable to evict 4448244781f1SPrakash Surya * enough buffers to get arc_size below the overflow limit (e.g. 4449244781f1SPrakash Surya * due to buffers being un-evictable, or hash lock collisions). 4450244781f1SPrakash Surya * In this case, we want to proceed regardless if we're 4451244781f1SPrakash Surya * overflowing; thus we don't use a while loop here. 445244eda4d7Smaybee */ 4453244781f1SPrakash Surya if (arc_is_overflowing()) { 4454de753e34SBrad Lewis mutex_enter(&arc_adjust_lock); 4455244781f1SPrakash Surya 4456244781f1SPrakash Surya /* 4457244781f1SPrakash Surya * Now that we've acquired the lock, we may no longer be 4458244781f1SPrakash Surya * over the overflow limit, lets check. 4459244781f1SPrakash Surya * 4460244781f1SPrakash Surya * We're ignoring the case of spurious wake ups. If that 4461244781f1SPrakash Surya * were to happen, it'd let this thread consume an ARC 4462244781f1SPrakash Surya * buffer before it should have (i.e. before we're under 4463244781f1SPrakash Surya * the overflow limit and were signalled by the reclaim 4464244781f1SPrakash Surya * thread). As long as that is a rare occurrence, it 4465244781f1SPrakash Surya * shouldn't cause any harm. 4466244781f1SPrakash Surya */ 4467244781f1SPrakash Surya if (arc_is_overflowing()) { 4468de753e34SBrad Lewis arc_adjust_needed = B_TRUE; 4469de753e34SBrad Lewis zthr_wakeup(arc_adjust_zthr); 4470de753e34SBrad Lewis (void) cv_wait(&arc_adjust_waiters_cv, 4471de753e34SBrad Lewis &arc_adjust_lock); 4472ad23a2dbSjohansen } 4473de753e34SBrad Lewis mutex_exit(&arc_adjust_lock); 447444eda4d7Smaybee } 447544eda4d7Smaybee 4476dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 4477244781f1SPrakash Surya if (type == ARC_BUFC_METADATA) { 4478244781f1SPrakash Surya arc_space_consume(size, ARC_SPACE_META); 4479fa9e4066Sahrens } else { 4480244781f1SPrakash Surya arc_space_consume(size, ARC_SPACE_DATA); 448144eda4d7Smaybee } 4482244781f1SPrakash Surya 448344eda4d7Smaybee /* 448444eda4d7Smaybee * Update the state size. Note that ghost states have a 448544eda4d7Smaybee * "ghost size" and so don't need to be updated. 448644eda4d7Smaybee */ 4487dcbf3bd6SGeorge Wilson if (!GHOST_STATE(state)) { 448844eda4d7Smaybee 4489dcbf3bd6SGeorge Wilson (void) refcount_add_many(&state->arcs_size, size, tag); 4490244781f1SPrakash Surya 4491244781f1SPrakash Surya /* 4492244781f1SPrakash Surya * If this is reached via arc_read, the link is 4493244781f1SPrakash Surya * protected by the hash lock. If reached via 4494244781f1SPrakash Surya * arc_buf_alloc, the header should not be accessed by 4495244781f1SPrakash Surya * any other thread. And, if reached via arc_read_done, 4496244781f1SPrakash Surya * the hash lock will protect it if it's found in the 4497244781f1SPrakash Surya * hash table; otherwise no other thread should be 4498244781f1SPrakash Surya * trying to [add|remove]_reference it. 4499244781f1SPrakash Surya */ 4500244781f1SPrakash Surya if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 450189c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4502dcbf3bd6SGeorge Wilson (void) refcount_add_many(&state->arcs_esize[type], 4503dcbf3bd6SGeorge Wilson size, tag); 4504fa9e4066Sahrens } 4505dcbf3bd6SGeorge Wilson 4506641fbdaeSmaybee /* 4507641fbdaeSmaybee * If we are growing the cache, and we are adding anonymous 450844cb6abcSbmc * data, and we have outgrown arc_p, update arc_p 4509641fbdaeSmaybee */ 45103a2d8a1bSPaul Dagnelie if (aggsum_compare(&arc_size, arc_c) < 0 && 45113a2d8a1bSPaul Dagnelie hdr->b_l1hdr.b_state == arc_anon && 45122fd872a7SPrakash Surya (refcount_count(&arc_anon->arcs_size) + 45132fd872a7SPrakash Surya refcount_count(&arc_mru->arcs_size) > arc_p)) 451444cb6abcSbmc arc_p = MIN(arc_c, arc_p + size); 4515fa9e4066Sahrens } 4516770499e1SDan Kimmel } 4517770499e1SDan Kimmel 4518770499e1SDan Kimmel static void 4519770499e1SDan Kimmel arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag) 4520770499e1SDan Kimmel { 4521770499e1SDan Kimmel arc_free_data_impl(hdr, size, tag); 4522770499e1SDan Kimmel abd_free(abd); 4523770499e1SDan Kimmel } 4524770499e1SDan Kimmel 4525770499e1SDan Kimmel static void 4526770499e1SDan Kimmel arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag) 4527770499e1SDan Kimmel { 4528770499e1SDan Kimmel arc_buf_contents_t type = arc_buf_type(hdr); 4529770499e1SDan Kimmel 4530770499e1SDan Kimmel arc_free_data_impl(hdr, size, tag); 4531770499e1SDan Kimmel if (type == ARC_BUFC_METADATA) { 4532770499e1SDan Kimmel zio_buf_free(buf, size); 4533770499e1SDan Kimmel } else { 4534770499e1SDan Kimmel ASSERT(type == ARC_BUFC_DATA); 4535770499e1SDan Kimmel zio_data_buf_free(buf, size); 4536770499e1SDan Kimmel } 4537dcbf3bd6SGeorge Wilson } 4538dcbf3bd6SGeorge Wilson 4539dcbf3bd6SGeorge Wilson /* 4540dcbf3bd6SGeorge Wilson * Free the arc data buffer. 4541dcbf3bd6SGeorge Wilson */ 4542dcbf3bd6SGeorge Wilson static void 4543770499e1SDan Kimmel arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4544dcbf3bd6SGeorge Wilson { 4545dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 4546dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 4547dcbf3bd6SGeorge Wilson 4548dcbf3bd6SGeorge Wilson /* protected by hash lock, if in the hash table */ 4549dcbf3bd6SGeorge Wilson if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4550dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4551dcbf3bd6SGeorge Wilson ASSERT(state != arc_anon && state != arc_l2c_only); 4552dcbf3bd6SGeorge Wilson 4553dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 4554dcbf3bd6SGeorge Wilson size, tag); 4555dcbf3bd6SGeorge Wilson } 4556dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_size, size, tag); 4557dcbf3bd6SGeorge Wilson 4558dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 4559dcbf3bd6SGeorge Wilson if (type == ARC_BUFC_METADATA) { 4560dcbf3bd6SGeorge Wilson arc_space_return(size, ARC_SPACE_META); 4561dcbf3bd6SGeorge Wilson } else { 4562dcbf3bd6SGeorge Wilson ASSERT(type == ARC_BUFC_DATA); 4563dcbf3bd6SGeorge Wilson arc_space_return(size, ARC_SPACE_DATA); 4564dcbf3bd6SGeorge Wilson } 4565fa9e4066Sahrens } 4566fa9e4066Sahrens 4567fa9e4066Sahrens /* 4568fa9e4066Sahrens * This routine is called whenever a buffer is accessed. 4569ea8dc4b6Seschrock * NOTE: the hash lock is dropped in this function. 4570fa9e4066Sahrens */ 4571fa9e4066Sahrens static void 45727adb730bSGeorge Wilson arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 4573fa9e4066Sahrens { 4574d3d50737SRafael Vanoni clock_t now; 4575d3d50737SRafael Vanoni 4576fa9e4066Sahrens ASSERT(MUTEX_HELD(hash_lock)); 457789c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 4578fa9e4066Sahrens 457989c86e32SChris Williamson if (hdr->b_l1hdr.b_state == arc_anon) { 4580fa9e4066Sahrens /* 4581fa9e4066Sahrens * This buffer is not in the cache, and does not 4582fa9e4066Sahrens * appear in our "ghost" list. Add the new buffer 4583fa9e4066Sahrens * to the MRU state. 4584fa9e4066Sahrens */ 4585fa9e4066Sahrens 458689c86e32SChris Williamson ASSERT0(hdr->b_l1hdr.b_arc_access); 458789c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 45887adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 45897adb730bSGeorge Wilson arc_change_state(arc_mru, hdr, hash_lock); 4590fa9e4066Sahrens 459189c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mru) { 4592d3d50737SRafael Vanoni now = ddi_get_lbolt(); 4593d3d50737SRafael Vanoni 4594fa9e4066Sahrens /* 459513506d1eSmaybee * If this buffer is here because of a prefetch, then either: 459613506d1eSmaybee * - clear the flag if this is a "referencing" read 459713506d1eSmaybee * (any subsequent access will bump this into the MFU state). 459813506d1eSmaybee * or 459913506d1eSmaybee * - move the buffer to the head of the list if this is 460013506d1eSmaybee * another prefetch (to make it less likely to be evicted). 4601fa9e4066Sahrens */ 460289c86e32SChris Williamson if (HDR_PREFETCH(hdr)) { 460389c86e32SChris Williamson if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4604244781f1SPrakash Surya /* link protected by hash lock */ 4605244781f1SPrakash Surya ASSERT(multilist_link_active( 460689c86e32SChris Williamson &hdr->b_l1hdr.b_arc_node)); 460713506d1eSmaybee } else { 4608dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 460944cb6abcSbmc ARCSTAT_BUMP(arcstat_mru_hits); 461013506d1eSmaybee } 461189c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = now; 4612fa9e4066Sahrens return; 4613fa9e4066Sahrens } 4614fa9e4066Sahrens 4615fa9e4066Sahrens /* 4616fa9e4066Sahrens * This buffer has been "accessed" only once so far, 4617fa9e4066Sahrens * but it is still in the cache. Move it to the MFU 4618fa9e4066Sahrens * state. 4619fa9e4066Sahrens */ 462089c86e32SChris Williamson if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 4621fa9e4066Sahrens /* 4622fa9e4066Sahrens * More than 125ms have passed since we 4623fa9e4066Sahrens * instantiated this buffer. Move it to the 4624fa9e4066Sahrens * most frequently used state. 4625fa9e4066Sahrens */ 462689c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = now; 46277adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 46287adb730bSGeorge Wilson arc_change_state(arc_mfu, hdr, hash_lock); 4629fa9e4066Sahrens } 463044cb6abcSbmc ARCSTAT_BUMP(arcstat_mru_hits); 463189c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 4632fa9e4066Sahrens arc_state_t *new_state; 4633fa9e4066Sahrens /* 4634fa9e4066Sahrens * This buffer has been "accessed" recently, but 4635fa9e4066Sahrens * was evicted from the cache. Move it to the 4636fa9e4066Sahrens * MFU state. 4637fa9e4066Sahrens */ 4638fa9e4066Sahrens 463989c86e32SChris Williamson if (HDR_PREFETCH(hdr)) { 464044cb6abcSbmc new_state = arc_mru; 464189c86e32SChris Williamson if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 4642dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 46437adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4644fa9e4066Sahrens } else { 464544cb6abcSbmc new_state = arc_mfu; 46467adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4647fa9e4066Sahrens } 4648fa9e4066Sahrens 464989c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 46507adb730bSGeorge Wilson arc_change_state(new_state, hdr, hash_lock); 4651fa9e4066Sahrens 465244cb6abcSbmc ARCSTAT_BUMP(arcstat_mru_ghost_hits); 465389c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mfu) { 4654fa9e4066Sahrens /* 4655fa9e4066Sahrens * This buffer has been accessed more than once and is 4656fa9e4066Sahrens * still in the cache. Keep it in the MFU state. 4657fa9e4066Sahrens * 465813506d1eSmaybee * NOTE: an add_reference() that occurred when we did 465913506d1eSmaybee * the arc_read() will have kicked this off the list. 466013506d1eSmaybee * If it was a prefetch, we will explicitly move it to 466113506d1eSmaybee * the head of the list now. 4662fa9e4066Sahrens */ 466389c86e32SChris Williamson if ((HDR_PREFETCH(hdr)) != 0) { 466489c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4665244781f1SPrakash Surya /* link protected by hash_lock */ 4666244781f1SPrakash Surya ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 466713506d1eSmaybee } 466844cb6abcSbmc ARCSTAT_BUMP(arcstat_mfu_hits); 466989c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 467089c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 467144cb6abcSbmc arc_state_t *new_state = arc_mfu; 4672fa9e4066Sahrens /* 4673fa9e4066Sahrens * This buffer has been accessed more than once but has 4674fa9e4066Sahrens * been evicted from the cache. Move it back to the 4675fa9e4066Sahrens * MFU state. 4676fa9e4066Sahrens */ 4677fa9e4066Sahrens 467889c86e32SChris Williamson if (HDR_PREFETCH(hdr)) { 467913506d1eSmaybee /* 468013506d1eSmaybee * This is a prefetch access... 468113506d1eSmaybee * move this block back to the MRU state. 468213506d1eSmaybee */ 468389c86e32SChris Williamson ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 468444cb6abcSbmc new_state = arc_mru; 468513506d1eSmaybee } 468613506d1eSmaybee 468789c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 46887adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 46897adb730bSGeorge Wilson arc_change_state(new_state, hdr, hash_lock); 4690fa9e4066Sahrens 469144cb6abcSbmc ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 469289c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4693fa94a07fSbrendan /* 4694fa94a07fSbrendan * This buffer is on the 2nd Level ARC. 4695fa94a07fSbrendan */ 4696fa94a07fSbrendan 469789c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 46987adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 46997adb730bSGeorge Wilson arc_change_state(arc_mfu, hdr, hash_lock); 4700fa9e4066Sahrens } else { 4701fa9e4066Sahrens ASSERT(!"invalid arc state"); 4702fa9e4066Sahrens } 4703fa9e4066Sahrens } 4704fa9e4066Sahrens 4705fa9e4066Sahrens /* a generic arc_done_func_t which you can use */ 4706fa9e4066Sahrens /* ARGSUSED */ 4707fa9e4066Sahrens void 4708fa9e4066Sahrens arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4709fa9e4066Sahrens { 47103f9d6ad7SLin Ling if (zio == NULL || zio->io_error == 0) 47115602294fSDan Kimmel bcopy(buf->b_data, arg, arc_buf_size(buf)); 4712dcbf3bd6SGeorge Wilson arc_buf_destroy(buf, arg); 4713fa9e4066Sahrens } 4714fa9e4066Sahrens 47150e8c6158Smaybee /* a generic arc_done_func_t */ 4716fa9e4066Sahrens void 4717fa9e4066Sahrens arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4718fa9e4066Sahrens { 4719fa9e4066Sahrens arc_buf_t **bufp = arg; 4720*fa98e487SMatthew Ahrens if (buf == NULL) { 4721*fa98e487SMatthew Ahrens ASSERT(zio == NULL || zio->io_error != 0); 4722fa9e4066Sahrens *bufp = NULL; 4723fa9e4066Sahrens } else { 4724*fa98e487SMatthew Ahrens ASSERT(zio == NULL || zio->io_error == 0); 4725fa9e4066Sahrens *bufp = buf; 4726*fa98e487SMatthew Ahrens ASSERT(buf->b_data != NULL); 4727fa9e4066Sahrens } 4728fa9e4066Sahrens } 4729fa9e4066Sahrens 4730dcbf3bd6SGeorge Wilson static void 4731dcbf3bd6SGeorge Wilson arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) 4732dcbf3bd6SGeorge Wilson { 4733dcbf3bd6SGeorge Wilson if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { 4734dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); 4735dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 4736dcbf3bd6SGeorge Wilson } else { 4737dcbf3bd6SGeorge Wilson if (HDR_COMPRESSION_ENABLED(hdr)) { 4738dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, 4739dcbf3bd6SGeorge Wilson BP_GET_COMPRESS(bp)); 4740dcbf3bd6SGeorge Wilson } 4741dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); 4742dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); 4743dcbf3bd6SGeorge Wilson } 4744dcbf3bd6SGeorge Wilson } 4745dcbf3bd6SGeorge Wilson 4746fa9e4066Sahrens static void 4747fa9e4066Sahrens arc_read_done(zio_t *zio) 4748fa9e4066Sahrens { 4749dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = zio->io_private; 47505d7b4d43SMatthew Ahrens kmutex_t *hash_lock = NULL; 47515602294fSDan Kimmel arc_callback_t *callback_list; 47525602294fSDan Kimmel arc_callback_t *acb; 47535602294fSDan Kimmel boolean_t freeable = B_FALSE; 47545602294fSDan Kimmel boolean_t no_zio_error = (zio->io_error == 0); 4755fa9e4066Sahrens 4756bbf4a8dfSmaybee /* 4757bbf4a8dfSmaybee * The hdr was inserted into hash-table and removed from lists 4758bbf4a8dfSmaybee * prior to starting I/O. We should find this header, since 4759bbf4a8dfSmaybee * it's in the hash table, and it should be legit since it's 4760bbf4a8dfSmaybee * not possible to evict it during the I/O. The only possible 4761bbf4a8dfSmaybee * reason for it not to be found is if we were freed during the 4762bbf4a8dfSmaybee * read. 4763bbf4a8dfSmaybee */ 47645d7b4d43SMatthew Ahrens if (HDR_IN_HASH_TABLE(hdr)) { 47655d7b4d43SMatthew Ahrens ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 47665d7b4d43SMatthew Ahrens ASSERT3U(hdr->b_dva.dva_word[0], ==, 47675d7b4d43SMatthew Ahrens BP_IDENTITY(zio->io_bp)->dva_word[0]); 47685d7b4d43SMatthew Ahrens ASSERT3U(hdr->b_dva.dva_word[1], ==, 47695d7b4d43SMatthew Ahrens BP_IDENTITY(zio->io_bp)->dva_word[1]); 47705d7b4d43SMatthew Ahrens 47715d7b4d43SMatthew Ahrens arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 47725d7b4d43SMatthew Ahrens &hash_lock); 47735d7b4d43SMatthew Ahrens 4774dcbf3bd6SGeorge Wilson ASSERT((found == hdr && 47755d7b4d43SMatthew Ahrens DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 47765d7b4d43SMatthew Ahrens (found == hdr && HDR_L2_READING(hdr))); 4777dcbf3bd6SGeorge Wilson ASSERT3P(hash_lock, !=, NULL); 4778dcbf3bd6SGeorge Wilson } 4779dcbf3bd6SGeorge Wilson 47805602294fSDan Kimmel if (no_zio_error) { 4781dcbf3bd6SGeorge Wilson /* byteswap if necessary */ 4782dcbf3bd6SGeorge Wilson if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 4783dcbf3bd6SGeorge Wilson if (BP_GET_LEVEL(zio->io_bp) > 0) { 4784dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; 4785dcbf3bd6SGeorge Wilson } else { 4786dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = 4787dcbf3bd6SGeorge Wilson DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4788dcbf3bd6SGeorge Wilson } 4789dcbf3bd6SGeorge Wilson } else { 4790dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 4791dcbf3bd6SGeorge Wilson } 47925d7b4d43SMatthew Ahrens } 4793fa94a07fSbrendan 4794dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); 479589c86e32SChris Williamson if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4796dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); 4797fa9e4066Sahrens 479889c86e32SChris Williamson callback_list = hdr->b_l1hdr.b_acb; 4799dcbf3bd6SGeorge Wilson ASSERT3P(callback_list, !=, NULL); 48006b4acc8bSahrens 48015602294fSDan Kimmel if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) { 4802b24ab676SJeff Bonwick /* 4803b24ab676SJeff Bonwick * Only call arc_access on anonymous buffers. This is because 4804b24ab676SJeff Bonwick * if we've issued an I/O for an evicted buffer, we've already 4805b24ab676SJeff Bonwick * called arc_access (to prevent any simultaneous readers from 4806b24ab676SJeff Bonwick * getting confused). 4807b24ab676SJeff Bonwick */ 4808b24ab676SJeff Bonwick arc_access(hdr, hash_lock); 4809b24ab676SJeff Bonwick } 4810b24ab676SJeff Bonwick 48115602294fSDan Kimmel /* 48125602294fSDan Kimmel * If a read request has a callback (i.e. acb_done is not NULL), then we 48135602294fSDan Kimmel * make a buf containing the data according to the parameters which were 48145602294fSDan Kimmel * passed in. The implementation of arc_buf_alloc_impl() ensures that we 48155602294fSDan Kimmel * aren't needlessly decompressing the data multiple times. 48165602294fSDan Kimmel */ 48175602294fSDan Kimmel int callback_cnt = 0; 48185602294fSDan Kimmel for (acb = callback_list; acb != NULL; acb = acb->acb_next) { 48195602294fSDan Kimmel if (!acb->acb_done) 48205602294fSDan Kimmel continue; 48215602294fSDan Kimmel 48225602294fSDan Kimmel /* This is a demand read since prefetches don't use callbacks */ 48235602294fSDan Kimmel callback_cnt++; 48245602294fSDan Kimmel 48255602294fSDan Kimmel if (no_zio_error) { 4826*fa98e487SMatthew Ahrens int error = arc_buf_alloc_impl(hdr, acb->acb_private, 4827*fa98e487SMatthew Ahrens acb->acb_compressed, zio->io_error == 0, 4828*fa98e487SMatthew Ahrens &acb->acb_buf); 4829*fa98e487SMatthew Ahrens if (error != 0) { 4830*fa98e487SMatthew Ahrens /* 4831*fa98e487SMatthew Ahrens * Decompression failed. Set io_error 4832*fa98e487SMatthew Ahrens * so that when we call acb_done (below), 4833*fa98e487SMatthew Ahrens * we will indicate that the read failed. 4834*fa98e487SMatthew Ahrens * Note that in the unusual case where one 4835*fa98e487SMatthew Ahrens * callback is compressed and another 4836*fa98e487SMatthew Ahrens * uncompressed, we will mark all of them 4837*fa98e487SMatthew Ahrens * as failed, even though the uncompressed 4838*fa98e487SMatthew Ahrens * one can't actually fail. In this case, 4839*fa98e487SMatthew Ahrens * the hdr will not be anonymous, because 4840*fa98e487SMatthew Ahrens * if there are multiple callbacks, it's 4841*fa98e487SMatthew Ahrens * because multiple threads found the same 4842*fa98e487SMatthew Ahrens * arc buf in the hash table. 4843*fa98e487SMatthew Ahrens */ 4844*fa98e487SMatthew Ahrens zio->io_error = error; 4845*fa98e487SMatthew Ahrens } 4846fa9e4066Sahrens } 4847fa9e4066Sahrens } 4848*fa98e487SMatthew Ahrens /* 4849*fa98e487SMatthew Ahrens * If there are multiple callbacks, we must have the hash lock, 4850*fa98e487SMatthew Ahrens * because the only way for multiple threads to find this hdr is 4851*fa98e487SMatthew Ahrens * in the hash table. This ensures that if there are multiple 4852*fa98e487SMatthew Ahrens * callbacks, the hdr is not anonymous. If it were anonymous, 4853*fa98e487SMatthew Ahrens * we couldn't use arc_buf_destroy() in the error case below. 4854*fa98e487SMatthew Ahrens */ 4855*fa98e487SMatthew Ahrens ASSERT(callback_cnt < 2 || hash_lock != NULL); 4856*fa98e487SMatthew Ahrens 485789c86e32SChris Williamson hdr->b_l1hdr.b_acb = NULL; 4858dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 48595602294fSDan Kimmel if (callback_cnt == 0) { 4860dcbf3bd6SGeorge Wilson ASSERT(HDR_PREFETCH(hdr)); 4861dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 4862770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 4863b24ab676SJeff Bonwick } 4864fa9e4066Sahrens 486589c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 486689c86e32SChris Williamson callback_list != NULL); 4867fa9e4066Sahrens 48685602294fSDan Kimmel if (no_zio_error) { 4869dcbf3bd6SGeorge Wilson arc_hdr_verify(hdr, zio->io_bp); 4870dcbf3bd6SGeorge Wilson } else { 4871dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); 487289c86e32SChris Williamson if (hdr->b_l1hdr.b_state != arc_anon) 487344cb6abcSbmc arc_change_state(arc_anon, hdr, hash_lock); 4874ea8dc4b6Seschrock if (HDR_IN_HASH_TABLE(hdr)) 4875ea8dc4b6Seschrock buf_hash_remove(hdr); 487689c86e32SChris Williamson freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4877fa9e4066Sahrens } 4878fa9e4066Sahrens 4879ea8dc4b6Seschrock /* 488013506d1eSmaybee * Broadcast before we drop the hash_lock to avoid the possibility 488113506d1eSmaybee * that the hdr (and hence the cv) might be freed before we get to 488213506d1eSmaybee * the cv_broadcast(). 4883ea8dc4b6Seschrock */ 488489c86e32SChris Williamson cv_broadcast(&hdr->b_l1hdr.b_cv); 4885ea8dc4b6Seschrock 488689c86e32SChris Williamson if (hash_lock != NULL) { 488744eda4d7Smaybee mutex_exit(hash_lock); 4888fa9e4066Sahrens } else { 4889fa9e4066Sahrens /* 4890fa9e4066Sahrens * This block was freed while we waited for the read to 4891fa9e4066Sahrens * complete. It has been removed from the hash table and 4892fa9e4066Sahrens * moved to the anonymous state (so that it won't show up 4893fa9e4066Sahrens * in the cache). 4894fa9e4066Sahrens */ 489589c86e32SChris Williamson ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 489689c86e32SChris Williamson freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4897fa9e4066Sahrens } 4898fa9e4066Sahrens 4899fa9e4066Sahrens /* execute each callback and free its structure */ 4900fa9e4066Sahrens while ((acb = callback_list) != NULL) { 4901*fa98e487SMatthew Ahrens if (acb->acb_done != NULL) { 4902*fa98e487SMatthew Ahrens if (zio->io_error != 0 && acb->acb_buf != NULL) { 4903*fa98e487SMatthew Ahrens /* 4904*fa98e487SMatthew Ahrens * If arc_buf_alloc_impl() fails during 4905*fa98e487SMatthew Ahrens * decompression, the buf will still be 4906*fa98e487SMatthew Ahrens * allocated, and needs to be freed here. 4907*fa98e487SMatthew Ahrens */ 4908*fa98e487SMatthew Ahrens arc_buf_destroy(acb->acb_buf, acb->acb_private); 4909*fa98e487SMatthew Ahrens acb->acb_buf = NULL; 4910*fa98e487SMatthew Ahrens } 4911fa9e4066Sahrens acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4912*fa98e487SMatthew Ahrens } 4913fa9e4066Sahrens 4914fa9e4066Sahrens if (acb->acb_zio_dummy != NULL) { 4915fa9e4066Sahrens acb->acb_zio_dummy->io_error = zio->io_error; 4916fa9e4066Sahrens zio_nowait(acb->acb_zio_dummy); 4917fa9e4066Sahrens } 4918fa9e4066Sahrens 4919fa9e4066Sahrens callback_list = acb->acb_next; 4920fa9e4066Sahrens kmem_free(acb, sizeof (arc_callback_t)); 4921fa9e4066Sahrens } 4922fa9e4066Sahrens 4923fa9e4066Sahrens if (freeable) 4924ea8dc4b6Seschrock arc_hdr_destroy(hdr); 4925fa9e4066Sahrens } 4926fa9e4066Sahrens 4927fa9e4066Sahrens /* 4928fc98fea5SBart Coddens * "Read" the block at the specified DVA (in bp) via the 4929fa9e4066Sahrens * cache. If the block is found in the cache, invoke the provided 4930fa9e4066Sahrens * callback immediately and return. Note that the `zio' parameter 4931fa9e4066Sahrens * in the callback will be NULL in this case, since no IO was 4932fa9e4066Sahrens * required. If the block is not in the cache pass the read request 4933fa9e4066Sahrens * on to the spa with a substitute callback function, so that the 4934fa9e4066Sahrens * requested block will be added to the cache. 4935fa9e4066Sahrens * 4936fa9e4066Sahrens * If a read request arrives for a block that has a read in-progress, 4937fa9e4066Sahrens * either wait for the in-progress read to complete (and return the 4938fa9e4066Sahrens * results); or, if this is a read with a "done" func, add a record 4939fa9e4066Sahrens * to the read to invoke the "done" func when the read completes, 4940fa9e4066Sahrens * and return; or just return. 4941fa9e4066Sahrens * 4942fa9e4066Sahrens * arc_read_done() will invoke all the requested "done" functions 4943fa9e4066Sahrens * for readers of this block. 4944fa9e4066Sahrens */ 4945fa9e4066Sahrens int 49461b912ec7SGeorge Wilson arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 49477adb730bSGeorge Wilson void *private, zio_priority_t priority, int zio_flags, 49487adb730bSGeorge Wilson arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4949fa9e4066Sahrens { 49505d7b4d43SMatthew Ahrens arc_buf_hdr_t *hdr = NULL; 49515d7b4d43SMatthew Ahrens kmutex_t *hash_lock = NULL; 4952fa94a07fSbrendan zio_t *rzio; 4953e9103aaeSGarrett D'Amore uint64_t guid = spa_load_guid(spa); 49545602294fSDan Kimmel boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0; 4955fa9e4066Sahrens 49565d7b4d43SMatthew Ahrens ASSERT(!BP_IS_EMBEDDED(bp) || 49575d7b4d43SMatthew Ahrens BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 49585d7b4d43SMatthew Ahrens 4959fa9e4066Sahrens top: 49605d7b4d43SMatthew Ahrens if (!BP_IS_EMBEDDED(bp)) { 49615d7b4d43SMatthew Ahrens /* 49625d7b4d43SMatthew Ahrens * Embedded BP's have no DVA and require no I/O to "read". 49635d7b4d43SMatthew Ahrens * Create an anonymous arc buf to back it. 49645d7b4d43SMatthew Ahrens */ 49655d7b4d43SMatthew Ahrens hdr = buf_hash_find(guid, bp, &hash_lock); 49665d7b4d43SMatthew Ahrens } 49675d7b4d43SMatthew Ahrens 4968770499e1SDan Kimmel if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) { 4969dcbf3bd6SGeorge Wilson arc_buf_t *buf = NULL; 49707adb730bSGeorge Wilson *arc_flags |= ARC_FLAG_CACHED; 497113506d1eSmaybee 4972fa9e4066Sahrens if (HDR_IO_IN_PROGRESS(hdr)) { 497313506d1eSmaybee 4974cf6106c8SMatthew Ahrens if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 4975cf6106c8SMatthew Ahrens priority == ZIO_PRIORITY_SYNC_READ) { 4976cf6106c8SMatthew Ahrens /* 4977cf6106c8SMatthew Ahrens * This sync read must wait for an 4978cf6106c8SMatthew Ahrens * in-progress async read (e.g. a predictive 4979cf6106c8SMatthew Ahrens * prefetch). Async reads are queued 4980cf6106c8SMatthew Ahrens * separately at the vdev_queue layer, so 4981cf6106c8SMatthew Ahrens * this is a form of priority inversion. 4982cf6106c8SMatthew Ahrens * Ideally, we would "inherit" the demand 4983cf6106c8SMatthew Ahrens * i/o's priority by moving the i/o from 4984cf6106c8SMatthew Ahrens * the async queue to the synchronous queue, 4985cf6106c8SMatthew Ahrens * but there is currently no mechanism to do 4986cf6106c8SMatthew Ahrens * so. Track this so that we can evaluate 4987cf6106c8SMatthew Ahrens * the magnitude of this potential performance 4988cf6106c8SMatthew Ahrens * problem. 4989cf6106c8SMatthew Ahrens * 4990cf6106c8SMatthew Ahrens * Note that if the prefetch i/o is already 4991cf6106c8SMatthew Ahrens * active (has been issued to the device), 4992cf6106c8SMatthew Ahrens * the prefetch improved performance, because 4993cf6106c8SMatthew Ahrens * we issued it sooner than we would have 4994cf6106c8SMatthew Ahrens * without the prefetch. 4995cf6106c8SMatthew Ahrens */ 4996cf6106c8SMatthew Ahrens DTRACE_PROBE1(arc__sync__wait__for__async, 4997cf6106c8SMatthew Ahrens arc_buf_hdr_t *, hdr); 4998cf6106c8SMatthew Ahrens ARCSTAT_BUMP(arcstat_sync_wait_for_async); 4999cf6106c8SMatthew Ahrens } 5000cf6106c8SMatthew Ahrens if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5001dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, 5002dcbf3bd6SGeorge Wilson ARC_FLAG_PREDICTIVE_PREFETCH); 5003cf6106c8SMatthew Ahrens } 5004cf6106c8SMatthew Ahrens 50057adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_WAIT) { 500689c86e32SChris Williamson cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 500713506d1eSmaybee mutex_exit(hash_lock); 500813506d1eSmaybee goto top; 500913506d1eSmaybee } 50107adb730bSGeorge Wilson ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 501113506d1eSmaybee 501213506d1eSmaybee if (done) { 5013cf6106c8SMatthew Ahrens arc_callback_t *acb = NULL; 5014fa9e4066Sahrens 5015fa9e4066Sahrens acb = kmem_zalloc(sizeof (arc_callback_t), 5016fa9e4066Sahrens KM_SLEEP); 5017fa9e4066Sahrens acb->acb_done = done; 5018fa9e4066Sahrens acb->acb_private = private; 50195602294fSDan Kimmel acb->acb_compressed = compressed_read; 5020fa9e4066Sahrens if (pio != NULL) 5021fa9e4066Sahrens acb->acb_zio_dummy = zio_null(pio, 5022a3f829aeSBill Moore spa, NULL, NULL, NULL, zio_flags); 5023fa9e4066Sahrens 5024dcbf3bd6SGeorge Wilson ASSERT3P(acb->acb_done, !=, NULL); 502589c86e32SChris Williamson acb->acb_next = hdr->b_l1hdr.b_acb; 502689c86e32SChris Williamson hdr->b_l1hdr.b_acb = acb; 5027fa9e4066Sahrens mutex_exit(hash_lock); 5028fa9e4066Sahrens return (0); 5029fa9e4066Sahrens } 5030fa9e4066Sahrens mutex_exit(hash_lock); 5031fa9e4066Sahrens return (0); 5032fa9e4066Sahrens } 5033fa9e4066Sahrens 503489c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_state == arc_mru || 503589c86e32SChris Williamson hdr->b_l1hdr.b_state == arc_mfu); 5036fa9e4066Sahrens 5037ea8dc4b6Seschrock if (done) { 5038cf6106c8SMatthew Ahrens if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5039cf6106c8SMatthew Ahrens /* 5040cf6106c8SMatthew Ahrens * This is a demand read which does not have to 5041cf6106c8SMatthew Ahrens * wait for i/o because we did a predictive 5042cf6106c8SMatthew Ahrens * prefetch i/o for it, which has completed. 5043cf6106c8SMatthew Ahrens */ 5044cf6106c8SMatthew Ahrens DTRACE_PROBE1( 5045cf6106c8SMatthew Ahrens arc__demand__hit__predictive__prefetch, 5046cf6106c8SMatthew Ahrens arc_buf_hdr_t *, hdr); 5047cf6106c8SMatthew Ahrens ARCSTAT_BUMP( 5048cf6106c8SMatthew Ahrens arcstat_demand_hit_predictive_prefetch); 5049dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, 5050dcbf3bd6SGeorge Wilson ARC_FLAG_PREDICTIVE_PREFETCH); 5051cf6106c8SMatthew Ahrens } 5052dcbf3bd6SGeorge Wilson ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); 5053dcbf3bd6SGeorge Wilson 50545602294fSDan Kimmel /* Get a buf with the desired data in it. */ 50555602294fSDan Kimmel VERIFY0(arc_buf_alloc_impl(hdr, private, 50565602294fSDan Kimmel compressed_read, B_TRUE, &buf)); 50577adb730bSGeorge Wilson } else if (*arc_flags & ARC_FLAG_PREFETCH && 505889c86e32SChris Williamson refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 5059dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5060fa9e4066Sahrens } 5061fa9e4066Sahrens DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 506244eda4d7Smaybee arc_access(hdr, hash_lock); 50637adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_L2CACHE) 5064dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 506544eda4d7Smaybee mutex_exit(hash_lock); 506644cb6abcSbmc ARCSTAT_BUMP(arcstat_hits); 506789c86e32SChris Williamson ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 506889c86e32SChris Williamson demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 506944cb6abcSbmc data, metadata, hits); 507044cb6abcSbmc 5071fa9e4066Sahrens if (done) 5072fa9e4066Sahrens done(NULL, buf, private); 5073fa9e4066Sahrens } else { 5074dcbf3bd6SGeorge Wilson uint64_t lsize = BP_GET_LSIZE(bp); 5075dcbf3bd6SGeorge Wilson uint64_t psize = BP_GET_PSIZE(bp); 50765d7b4d43SMatthew Ahrens arc_callback_t *acb; 50773a737e0dSbrendan vdev_t *vd = NULL; 5078d5285caeSGeorge Wilson uint64_t addr = 0; 50795a98e54bSBrendan Gregg - Sun Microsystems boolean_t devw = B_FALSE; 5080dcbf3bd6SGeorge Wilson uint64_t size; 5081fa9e4066Sahrens 5082fa9e4066Sahrens if (hdr == NULL) { 5083fa9e4066Sahrens /* this block is not in the cache */ 50845d7b4d43SMatthew Ahrens arc_buf_hdr_t *exists = NULL; 5085ad23a2dbSjohansen arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 5086dcbf3bd6SGeorge Wilson hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 5087dcbf3bd6SGeorge Wilson BP_GET_COMPRESS(bp), type); 5088dcbf3bd6SGeorge Wilson 50895d7b4d43SMatthew Ahrens if (!BP_IS_EMBEDDED(bp)) { 50905d7b4d43SMatthew Ahrens hdr->b_dva = *BP_IDENTITY(bp); 50915d7b4d43SMatthew Ahrens hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 50925d7b4d43SMatthew Ahrens exists = buf_hash_insert(hdr, &hash_lock); 50935d7b4d43SMatthew Ahrens } 50945d7b4d43SMatthew Ahrens if (exists != NULL) { 5095fa9e4066Sahrens /* somebody beat us to the hash insert */ 5096fa9e4066Sahrens mutex_exit(hash_lock); 50973f9d6ad7SLin Ling buf_discard_identity(hdr); 5098dcbf3bd6SGeorge Wilson arc_hdr_destroy(hdr); 5099fa9e4066Sahrens goto top; /* restart the IO request */ 5100fa9e4066Sahrens } 5101fa9e4066Sahrens } else { 510289c86e32SChris Williamson /* 510389c86e32SChris Williamson * This block is in the ghost cache. If it was L2-only 510489c86e32SChris Williamson * (and thus didn't have an L1 hdr), we realloc the 510589c86e32SChris Williamson * header to add an L1 hdr. 510689c86e32SChris Williamson */ 510789c86e32SChris Williamson if (!HDR_HAS_L1HDR(hdr)) { 510889c86e32SChris Williamson hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 510989c86e32SChris Williamson hdr_full_cache); 511089c86e32SChris Williamson } 5111770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 511289c86e32SChris Williamson ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 5113ea8dc4b6Seschrock ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 511489c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5115244781f1SPrakash Surya ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 51165602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 511713506d1eSmaybee 5118cf6106c8SMatthew Ahrens /* 5119dcbf3bd6SGeorge Wilson * This is a delicate dance that we play here. 5120dcbf3bd6SGeorge Wilson * This hdr is in the ghost list so we access it 5121dcbf3bd6SGeorge Wilson * to move it out of the ghost list before we 5122dcbf3bd6SGeorge Wilson * initiate the read. If it's a prefetch then 5123dcbf3bd6SGeorge Wilson * it won't have a callback so we'll remove the 5124dcbf3bd6SGeorge Wilson * reference that arc_buf_alloc_impl() created. We 5125dcbf3bd6SGeorge Wilson * do this after we've called arc_access() to 5126dcbf3bd6SGeorge Wilson * avoid hitting an assert in remove_reference(). 5127cf6106c8SMatthew Ahrens */ 51287e453561SWilliam Gorrell arc_access(hdr, hash_lock); 5129770499e1SDan Kimmel arc_hdr_alloc_pabd(hdr); 5130dcbf3bd6SGeorge Wilson } 5131770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5132dcbf3bd6SGeorge Wilson size = arc_hdr_size(hdr); 5133dcbf3bd6SGeorge Wilson 5134dcbf3bd6SGeorge Wilson /* 5135dcbf3bd6SGeorge Wilson * If compression is enabled on the hdr, then will do 5136dcbf3bd6SGeorge Wilson * RAW I/O and will store the compressed data in the hdr's 5137dcbf3bd6SGeorge Wilson * data block. Otherwise, the hdr's data block will contain 5138dcbf3bd6SGeorge Wilson * the uncompressed data. 5139dcbf3bd6SGeorge Wilson */ 5140dcbf3bd6SGeorge Wilson if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5141dcbf3bd6SGeorge Wilson zio_flags |= ZIO_FLAG_RAW; 5142fa9e4066Sahrens } 5143fa9e4066Sahrens 5144dcbf3bd6SGeorge Wilson if (*arc_flags & ARC_FLAG_PREFETCH) 5145dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5146dcbf3bd6SGeorge Wilson if (*arc_flags & ARC_FLAG_L2CACHE) 5147dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5148dcbf3bd6SGeorge Wilson if (BP_GET_LEVEL(bp) > 0) 5149dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); 5150cf6106c8SMatthew Ahrens if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 5151dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); 515289c86e32SChris Williamson ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 51535614b00aSWilliam Gorrell 5154fa9e4066Sahrens acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 5155fa9e4066Sahrens acb->acb_done = done; 5156fa9e4066Sahrens acb->acb_private = private; 51575602294fSDan Kimmel acb->acb_compressed = compressed_read; 5158fa9e4066Sahrens 5159dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 516089c86e32SChris Williamson hdr->b_l1hdr.b_acb = acb; 5161dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5162fa9e4066Sahrens 516389c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr) && 516489c86e32SChris Williamson (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 516589c86e32SChris Williamson devw = hdr->b_l2hdr.b_dev->l2ad_writing; 516689c86e32SChris Williamson addr = hdr->b_l2hdr.b_daddr; 5167e14bb325SJeff Bonwick /* 51685cabbc6bSPrashanth Sreenivasa * Lock out L2ARC device removal. 5169e14bb325SJeff Bonwick */ 5170e14bb325SJeff Bonwick if (vdev_is_dead(vd) || 5171e14bb325SJeff Bonwick !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 5172e14bb325SJeff Bonwick vd = NULL; 51733a737e0dSbrendan } 51743a737e0dSbrendan 5175dcbf3bd6SGeorge Wilson if (priority == ZIO_PRIORITY_ASYNC_READ) 5176dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5177dcbf3bd6SGeorge Wilson else 5178dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5179dcbf3bd6SGeorge Wilson 51805d7b4d43SMatthew Ahrens if (hash_lock != NULL) 51815d7b4d43SMatthew Ahrens mutex_exit(hash_lock); 51823a737e0dSbrendan 51833e30c24aSWill Andrews /* 51843e30c24aSWill Andrews * At this point, we have a level 1 cache miss. Try again in 51853e30c24aSWill Andrews * L2ARC if possible. 51863e30c24aSWill Andrews */ 5187dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); 5188dcbf3bd6SGeorge Wilson 51895c28183bSBrendan Gregg - Sun Microsystems DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 5190dcbf3bd6SGeorge Wilson uint64_t, lsize, zbookmark_phys_t *, zb); 519144cb6abcSbmc ARCSTAT_BUMP(arcstat_misses); 519289c86e32SChris Williamson ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 519389c86e32SChris Williamson demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 519444cb6abcSbmc data, metadata, misses); 5195ea8dc4b6Seschrock 51965a98e54bSBrendan Gregg - Sun Microsystems if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 5197fa94a07fSbrendan /* 5198fa94a07fSbrendan * Read from the L2ARC if the following are true: 51993a737e0dSbrendan * 1. The L2ARC vdev was previously cached. 52003a737e0dSbrendan * 2. This buffer still has L2ARC metadata. 52013a737e0dSbrendan * 3. This buffer isn't currently writing to the L2ARC. 52023a737e0dSbrendan * 4. The L2ARC entry wasn't evicted, which may 52033a737e0dSbrendan * also have invalidated the vdev. 52045a98e54bSBrendan Gregg - Sun Microsystems * 5. This isn't prefetch and l2arc_noprefetch is set. 5205fa94a07fSbrendan */ 520689c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr) && 52075a98e54bSBrendan Gregg - Sun Microsystems !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 52085a98e54bSBrendan Gregg - Sun Microsystems !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 5209fa94a07fSbrendan l2arc_read_callback_t *cb; 5210403a8da7SAndriy Gapon abd_t *abd; 5211403a8da7SAndriy Gapon uint64_t asize; 5212fa94a07fSbrendan 5213c5904d13Seschrock DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 5214c5904d13Seschrock ARCSTAT_BUMP(arcstat_l2_hits); 5215c5904d13Seschrock 5216fa94a07fSbrendan cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 5217fa94a07fSbrendan KM_SLEEP); 5218dcbf3bd6SGeorge Wilson cb->l2rcb_hdr = hdr; 5219fa94a07fSbrendan cb->l2rcb_bp = *bp; 5220fa94a07fSbrendan cb->l2rcb_zb = *zb; 52213baa08fcSek cb->l2rcb_flags = zio_flags; 5222fa94a07fSbrendan 5223403a8da7SAndriy Gapon asize = vdev_psize_to_asize(vd, size); 5224403a8da7SAndriy Gapon if (asize != size) { 5225403a8da7SAndriy Gapon abd = abd_alloc_for_io(asize, 5226403a8da7SAndriy Gapon HDR_ISTYPE_METADATA(hdr)); 5227403a8da7SAndriy Gapon cb->l2rcb_abd = abd; 5228403a8da7SAndriy Gapon } else { 5229403a8da7SAndriy Gapon abd = hdr->b_l1hdr.b_pabd; 5230403a8da7SAndriy Gapon } 5231403a8da7SAndriy Gapon 5232d5285caeSGeorge Wilson ASSERT(addr >= VDEV_LABEL_START_SIZE && 5233403a8da7SAndriy Gapon addr + asize <= vd->vdev_psize - 5234d5285caeSGeorge Wilson VDEV_LABEL_END_SIZE); 5235d5285caeSGeorge Wilson 5236fa94a07fSbrendan /* 5237e14bb325SJeff Bonwick * l2arc read. The SCL_L2ARC lock will be 5238e14bb325SJeff Bonwick * released by l2arc_read_done(). 5239aad02571SSaso Kiselkov * Issue a null zio if the underlying buffer 5240aad02571SSaso Kiselkov * was squashed to zero size by compression. 5241fa94a07fSbrendan */ 5242dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), !=, 5243dcbf3bd6SGeorge Wilson ZIO_COMPRESS_EMPTY); 5244dcbf3bd6SGeorge Wilson rzio = zio_read_phys(pio, vd, addr, 5245403a8da7SAndriy Gapon asize, abd, 5246dcbf3bd6SGeorge Wilson ZIO_CHECKSUM_OFF, 5247dcbf3bd6SGeorge Wilson l2arc_read_done, cb, priority, 5248dcbf3bd6SGeorge Wilson zio_flags | ZIO_FLAG_DONT_CACHE | 5249dcbf3bd6SGeorge Wilson ZIO_FLAG_CANFAIL | 5250dcbf3bd6SGeorge Wilson ZIO_FLAG_DONT_PROPAGATE | 5251dcbf3bd6SGeorge Wilson ZIO_FLAG_DONT_RETRY, B_FALSE); 5252fa94a07fSbrendan DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 5253fa94a07fSbrendan zio_t *, rzio); 5254dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_l2_read_bytes, size); 5255fa94a07fSbrendan 52567adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_NOWAIT) { 52573a737e0dSbrendan zio_nowait(rzio); 52583a737e0dSbrendan return (0); 52593a737e0dSbrendan } 5260fa94a07fSbrendan 52617adb730bSGeorge Wilson ASSERT(*arc_flags & ARC_FLAG_WAIT); 52623a737e0dSbrendan if (zio_wait(rzio) == 0) 52633a737e0dSbrendan return (0); 52643a737e0dSbrendan 52653a737e0dSbrendan /* l2arc read error; goto zio_read() */ 5266fa94a07fSbrendan } else { 5267fa94a07fSbrendan DTRACE_PROBE1(l2arc__miss, 5268fa94a07fSbrendan arc_buf_hdr_t *, hdr); 5269fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_misses); 5270fa94a07fSbrendan if (HDR_L2_WRITING(hdr)) 5271fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_rw_clash); 5272e14bb325SJeff Bonwick spa_config_exit(spa, SCL_L2ARC, vd); 5273fa94a07fSbrendan } 52745a98e54bSBrendan Gregg - Sun Microsystems } else { 527576a25fafSBill Moore if (vd != NULL) 527676a25fafSBill Moore spa_config_exit(spa, SCL_L2ARC, vd); 52775a98e54bSBrendan Gregg - Sun Microsystems if (l2arc_ndev != 0) { 52785a98e54bSBrendan Gregg - Sun Microsystems DTRACE_PROBE1(l2arc__miss, 52795a98e54bSBrendan Gregg - Sun Microsystems arc_buf_hdr_t *, hdr); 52805a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_BUMP(arcstat_l2_misses); 52815a98e54bSBrendan Gregg - Sun Microsystems } 5282fa94a07fSbrendan } 5283c5904d13Seschrock 5284770499e1SDan Kimmel rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size, 5285dcbf3bd6SGeorge Wilson arc_read_done, hdr, priority, zio_flags, zb); 5286fa9e4066Sahrens 52877adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_WAIT) 5288fa9e4066Sahrens return (zio_wait(rzio)); 5289fa9e4066Sahrens 52907adb730bSGeorge Wilson ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5291fa9e4066Sahrens zio_nowait(rzio); 5292fa9e4066Sahrens } 5293fa9e4066Sahrens return (0); 5294fa9e4066Sahrens } 5295fa9e4066Sahrens 52966e6d5868SMatthew Ahrens /* 52976e6d5868SMatthew Ahrens * Notify the arc that a block was freed, and thus will never be used again. 52986e6d5868SMatthew Ahrens */ 52996e6d5868SMatthew Ahrens void 53006e6d5868SMatthew Ahrens arc_freed(spa_t *spa, const blkptr_t *bp) 53016e6d5868SMatthew Ahrens { 53026e6d5868SMatthew Ahrens arc_buf_hdr_t *hdr; 53036e6d5868SMatthew Ahrens kmutex_t *hash_lock; 53046e6d5868SMatthew Ahrens uint64_t guid = spa_load_guid(spa); 53056e6d5868SMatthew Ahrens 53065d7b4d43SMatthew Ahrens ASSERT(!BP_IS_EMBEDDED(bp)); 53075d7b4d43SMatthew Ahrens 53085d7b4d43SMatthew Ahrens hdr = buf_hash_find(guid, bp, &hash_lock); 53096e6d5868SMatthew Ahrens if (hdr == NULL) 53106e6d5868SMatthew Ahrens return; 53116e6d5868SMatthew Ahrens 5312dcbf3bd6SGeorge Wilson /* 5313dcbf3bd6SGeorge Wilson * We might be trying to free a block that is still doing I/O 5314dcbf3bd6SGeorge Wilson * (i.e. prefetch) or has a reference (i.e. a dedup-ed, 5315dcbf3bd6SGeorge Wilson * dmu_sync-ed block). If this block is being prefetched, then it 5316dcbf3bd6SGeorge Wilson * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr 5317dcbf3bd6SGeorge Wilson * until the I/O completes. A block may also have a reference if it is 5318dcbf3bd6SGeorge Wilson * part of a dedup-ed, dmu_synced write. The dmu_sync() function would 5319dcbf3bd6SGeorge Wilson * have written the new block to its final resting place on disk but 5320dcbf3bd6SGeorge Wilson * without the dedup flag set. This would have left the hdr in the MRU 5321dcbf3bd6SGeorge Wilson * state and discoverable. When the txg finally syncs it detects that 5322dcbf3bd6SGeorge Wilson * the block was overridden in open context and issues an override I/O. 5323dcbf3bd6SGeorge Wilson * Since this is a dedup block, the override I/O will determine if the 5324dcbf3bd6SGeorge Wilson * block is already in the DDT. If so, then it will replace the io_bp 5325dcbf3bd6SGeorge Wilson * with the bp from the DDT and allow the I/O to finish. When the I/O 5326dcbf3bd6SGeorge Wilson * reaches the done callback, dbuf_write_override_done, it will 5327dcbf3bd6SGeorge Wilson * check to see if the io_bp and io_bp_override are identical. 5328dcbf3bd6SGeorge Wilson * If they are not, then it indicates that the bp was replaced with 5329dcbf3bd6SGeorge Wilson * the bp in the DDT and the override bp is freed. This allows 5330dcbf3bd6SGeorge Wilson * us to arrive here with a reference on a block that is being 5331dcbf3bd6SGeorge Wilson * freed. So if we have an I/O in progress, or a reference to 5332dcbf3bd6SGeorge Wilson * this hdr, then we don't destroy the hdr. 5333dcbf3bd6SGeorge Wilson */ 5334dcbf3bd6SGeorge Wilson if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && 5335dcbf3bd6SGeorge Wilson refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { 5336dcbf3bd6SGeorge Wilson arc_change_state(arc_anon, hdr, hash_lock); 5337dcbf3bd6SGeorge Wilson arc_hdr_destroy(hdr); 53386e6d5868SMatthew Ahrens mutex_exit(hash_lock); 5339bbfa8ea8SMatthew Ahrens } else { 5340dcbf3bd6SGeorge Wilson mutex_exit(hash_lock); 5341ea8dc4b6Seschrock } 5342dd6ef538Smaybee 5343ea8dc4b6Seschrock } 5344ea8dc4b6Seschrock 5345fa9e4066Sahrens /* 53463e30c24aSWill Andrews * Release this buffer from the cache, making it an anonymous buffer. This 53473e30c24aSWill Andrews * must be done after a read and prior to modifying the buffer contents. 5348fa9e4066Sahrens * If the buffer has more than one reference, we must make 5349088f3894Sahrens * a new hdr for the buffer. 5350fa9e4066Sahrens */ 5351fa9e4066Sahrens void 5352fa9e4066Sahrens arc_release(arc_buf_t *buf, void *tag) 5353fa9e4066Sahrens { 535489c86e32SChris Williamson arc_buf_hdr_t *hdr = buf->b_hdr; 5355fa9e4066Sahrens 53563f9d6ad7SLin Ling /* 53573f9d6ad7SLin Ling * It would be nice to assert that if it's DMU metadata (level > 53583f9d6ad7SLin Ling * 0 || it's the dnode file), then it must be syncing context. 53593f9d6ad7SLin Ling * But we don't know that information at this level. 53603f9d6ad7SLin Ling */ 53613f9d6ad7SLin Ling 53623f9d6ad7SLin Ling mutex_enter(&buf->b_evict_lock); 53636f83844dSMark Maybee 5364244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 5365244781f1SPrakash Surya 536689c86e32SChris Williamson /* 536789c86e32SChris Williamson * We don't grab the hash lock prior to this check, because if 536889c86e32SChris Williamson * the buffer's header is in the arc_anon state, it won't be 536989c86e32SChris Williamson * linked into the hash table. 537089c86e32SChris Williamson */ 537189c86e32SChris Williamson if (hdr->b_l1hdr.b_state == arc_anon) { 537289c86e32SChris Williamson mutex_exit(&buf->b_evict_lock); 537389c86e32SChris Williamson ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 537489c86e32SChris Williamson ASSERT(!HDR_IN_HASH_TABLE(hdr)); 537589c86e32SChris Williamson ASSERT(!HDR_HAS_L2HDR(hdr)); 5376dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 5377fa9e4066Sahrens 5378dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 537989c86e32SChris Williamson ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 538089c86e32SChris Williamson ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 538189c86e32SChris Williamson 538289c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = 0; 5383dcbf3bd6SGeorge Wilson 5384dcbf3bd6SGeorge Wilson /* 5385dcbf3bd6SGeorge Wilson * If the buf is being overridden then it may already 5386dcbf3bd6SGeorge Wilson * have a hdr that is not empty. 5387dcbf3bd6SGeorge Wilson */ 5388dcbf3bd6SGeorge Wilson buf_discard_identity(hdr); 538989c86e32SChris Williamson arc_buf_thaw(buf); 539089c86e32SChris Williamson 539189c86e32SChris Williamson return; 5392fa9e4066Sahrens } 5393fa9e4066Sahrens 539489c86e32SChris Williamson kmutex_t *hash_lock = HDR_LOCK(hdr); 539589c86e32SChris Williamson mutex_enter(hash_lock); 539689c86e32SChris Williamson 539789c86e32SChris Williamson /* 539889c86e32SChris Williamson * This assignment is only valid as long as the hash_lock is 539989c86e32SChris Williamson * held, we must be careful not to reference state or the 540089c86e32SChris Williamson * b_state field after dropping the lock. 540189c86e32SChris Williamson */ 540289c86e32SChris Williamson arc_state_t *state = hdr->b_l1hdr.b_state; 540389c86e32SChris Williamson ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 540489c86e32SChris Williamson ASSERT3P(state, !=, arc_anon); 540589c86e32SChris Williamson 540689c86e32SChris Williamson /* this buffer is not on any list */ 54075602294fSDan Kimmel ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0); 540889c86e32SChris Williamson 540989c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr)) { 541089c86e32SChris Williamson mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5411244781f1SPrakash Surya 5412244781f1SPrakash Surya /* 5413a52fc310SPrakash Surya * We have to recheck this conditional again now that 5414a52fc310SPrakash Surya * we're holding the l2ad_mtx to prevent a race with 5415a52fc310SPrakash Surya * another thread which might be concurrently calling 5416a52fc310SPrakash Surya * l2arc_evict(). In that case, l2arc_evict() might have 5417a52fc310SPrakash Surya * destroyed the header's L2 portion as we were waiting 5418a52fc310SPrakash Surya * to acquire the l2ad_mtx. 5419244781f1SPrakash Surya */ 5420a52fc310SPrakash Surya if (HDR_HAS_L2HDR(hdr)) 5421a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(hdr); 5422244781f1SPrakash Surya 542389c86e32SChris Williamson mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 54246f83844dSMark Maybee } 54256f83844dSMark Maybee 5426ea8dc4b6Seschrock /* 5427ea8dc4b6Seschrock * Do we have more than one buf? 5428ea8dc4b6Seschrock */ 5429dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_bufcnt > 1) { 5430fa9e4066Sahrens arc_buf_hdr_t *nhdr; 5431ac05c741SMark Maybee uint64_t spa = hdr->b_spa; 5432dcbf3bd6SGeorge Wilson uint64_t psize = HDR_GET_PSIZE(hdr); 5433dcbf3bd6SGeorge Wilson uint64_t lsize = HDR_GET_LSIZE(hdr); 5434dcbf3bd6SGeorge Wilson enum zio_compress compress = HDR_GET_COMPRESS(hdr); 543589c86e32SChris Williamson arc_buf_contents_t type = arc_buf_type(hdr); 5436dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 5437fa9e4066Sahrens 543889c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 5439dcbf3bd6SGeorge Wilson (void) remove_reference(hdr, hash_lock, tag); 5440dcbf3bd6SGeorge Wilson 54415602294fSDan Kimmel if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) { 5442dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5443dcbf3bd6SGeorge Wilson ASSERT(ARC_BUF_LAST(buf)); 5444dcbf3bd6SGeorge Wilson } 5445dcbf3bd6SGeorge Wilson 5446fa9e4066Sahrens /* 54473f9d6ad7SLin Ling * Pull the data off of this hdr and attach it to 5448dcbf3bd6SGeorge Wilson * a new anonymous hdr. Also find the last buffer 5449dcbf3bd6SGeorge Wilson * in the hdr's buffer list. 5450fa9e4066Sahrens */ 54515602294fSDan Kimmel arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 5452dcbf3bd6SGeorge Wilson ASSERT3P(lastbuf, !=, NULL); 5453dcbf3bd6SGeorge Wilson 5454dcbf3bd6SGeorge Wilson /* 5455dcbf3bd6SGeorge Wilson * If the current arc_buf_t and the hdr are sharing their data 54565602294fSDan Kimmel * buffer, then we must stop sharing that block. 5457dcbf3bd6SGeorge Wilson */ 5458dcbf3bd6SGeorge Wilson if (arc_buf_is_shared(buf)) { 5459dcbf3bd6SGeorge Wilson VERIFY(!arc_buf_is_shared(lastbuf)); 5460ea8dc4b6Seschrock 5461dcbf3bd6SGeorge Wilson /* 5462dcbf3bd6SGeorge Wilson * First, sever the block sharing relationship between 54635602294fSDan Kimmel * buf and the arc_buf_hdr_t. 5464dcbf3bd6SGeorge Wilson */ 5465dcbf3bd6SGeorge Wilson arc_unshare_buf(hdr, buf); 54665602294fSDan Kimmel 54675602294fSDan Kimmel /* 5468770499e1SDan Kimmel * Now we need to recreate the hdr's b_pabd. Since we 54695602294fSDan Kimmel * have lastbuf handy, we try to share with it, but if 5470770499e1SDan Kimmel * we can't then we allocate a new b_pabd and copy the 54715602294fSDan Kimmel * data from buf into it. 54725602294fSDan Kimmel */ 54735602294fSDan Kimmel if (arc_can_share(hdr, lastbuf)) { 54745602294fSDan Kimmel arc_share_buf(hdr, lastbuf); 54755602294fSDan Kimmel } else { 5476770499e1SDan Kimmel arc_hdr_alloc_pabd(hdr); 5477770499e1SDan Kimmel abd_copy_from_buf(hdr->b_l1hdr.b_pabd, 5478770499e1SDan Kimmel buf->b_data, psize); 54795602294fSDan Kimmel } 5480dcbf3bd6SGeorge Wilson VERIFY3P(lastbuf->b_data, !=, NULL); 5481dcbf3bd6SGeorge Wilson } else if (HDR_SHARED_DATA(hdr)) { 54825602294fSDan Kimmel /* 54835602294fSDan Kimmel * Uncompressed shared buffers are always at the end 54845602294fSDan Kimmel * of the list. Compressed buffers don't have the 54855602294fSDan Kimmel * same requirements. This makes it hard to 54865602294fSDan Kimmel * simply assert that the lastbuf is shared so 54875602294fSDan Kimmel * we rely on the hdr's compression flags to determine 54885602294fSDan Kimmel * if we have a compressed, shared buffer. 54895602294fSDan Kimmel */ 54905602294fSDan Kimmel ASSERT(arc_buf_is_shared(lastbuf) || 54915602294fSDan Kimmel HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 54925602294fSDan Kimmel ASSERT(!ARC_BUF_SHARED(buf)); 5493dcbf3bd6SGeorge Wilson } 5494770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 549589c86e32SChris Williamson ASSERT3P(state, !=, arc_l2c_only); 54962fd872a7SPrakash Surya 5497dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_size, 54985602294fSDan Kimmel arc_buf_size(buf), buf); 54992fd872a7SPrakash Surya 550089c86e32SChris Williamson if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 550189c86e32SChris Williamson ASSERT3P(state, !=, arc_l2c_only); 5502dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 55035602294fSDan Kimmel arc_buf_size(buf), buf); 5504ea8dc4b6Seschrock } 55059253d63dSGeorge Wilson 5506dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt -= 1; 5507c717a561Smaybee arc_cksum_verify(buf); 5508cd1c8b85SMatthew Ahrens arc_buf_unwatch(buf); 5509ea8dc4b6Seschrock 5510fa9e4066Sahrens mutex_exit(hash_lock); 5511fa9e4066Sahrens 5512dcbf3bd6SGeorge Wilson /* 5513770499e1SDan Kimmel * Allocate a new hdr. The new hdr will contain a b_pabd 5514dcbf3bd6SGeorge Wilson * buffer which will be freed in arc_write(). 5515dcbf3bd6SGeorge Wilson */ 5516dcbf3bd6SGeorge Wilson nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type); 5517dcbf3bd6SGeorge Wilson ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); 5518dcbf3bd6SGeorge Wilson ASSERT0(nhdr->b_l1hdr.b_bufcnt); 5519dcbf3bd6SGeorge Wilson ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt)); 5520dcbf3bd6SGeorge Wilson VERIFY3U(nhdr->b_type, ==, type); 5521dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(nhdr)); 552289c86e32SChris Williamson 552389c86e32SChris Williamson nhdr->b_l1hdr.b_buf = buf; 5524dcbf3bd6SGeorge Wilson nhdr->b_l1hdr.b_bufcnt = 1; 552589c86e32SChris Williamson (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 5526af2c4821Smaybee buf->b_hdr = nhdr; 5527dcbf3bd6SGeorge Wilson 55283f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 5529dcbf3bd6SGeorge Wilson (void) refcount_add_many(&arc_anon->arcs_size, 55305602294fSDan Kimmel arc_buf_size(buf), buf); 5531fa9e4066Sahrens } else { 55323f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 553389c86e32SChris Williamson ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 5534244781f1SPrakash Surya /* protected by hash lock, or hdr is on arc_anon */ 5535244781f1SPrakash Surya ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5536fa9e4066Sahrens ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 553789c86e32SChris Williamson arc_change_state(arc_anon, hdr, hash_lock); 553889c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = 0; 553989c86e32SChris Williamson mutex_exit(hash_lock); 5540fa94a07fSbrendan 55413f9d6ad7SLin Ling buf_discard_identity(hdr); 5542c717a561Smaybee arc_buf_thaw(buf); 5543fa9e4066Sahrens } 5544fa9e4066Sahrens } 5545fa9e4066Sahrens 5546fa9e4066Sahrens int 5547fa9e4066Sahrens arc_released(arc_buf_t *buf) 5548fa9e4066Sahrens { 55496f83844dSMark Maybee int released; 55506f83844dSMark Maybee 55513f9d6ad7SLin Ling mutex_enter(&buf->b_evict_lock); 555289c86e32SChris Williamson released = (buf->b_data != NULL && 555389c86e32SChris Williamson buf->b_hdr->b_l1hdr.b_state == arc_anon); 55543f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 55556f83844dSMark Maybee return (released); 5556ea8dc4b6Seschrock } 5557ea8dc4b6Seschrock 5558ea8dc4b6Seschrock #ifdef ZFS_DEBUG 5559ea8dc4b6Seschrock int 5560ea8dc4b6Seschrock arc_referenced(arc_buf_t *buf) 5561ea8dc4b6Seschrock { 55626f83844dSMark Maybee int referenced; 55636f83844dSMark Maybee 55643f9d6ad7SLin Ling mutex_enter(&buf->b_evict_lock); 556589c86e32SChris Williamson referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 55663f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 55676f83844dSMark Maybee return (referenced); 5568ea8dc4b6Seschrock } 5569ea8dc4b6Seschrock #endif 5570ea8dc4b6Seschrock 5571c717a561Smaybee static void 5572c717a561Smaybee arc_write_ready(zio_t *zio) 5573c717a561Smaybee { 5574c717a561Smaybee arc_write_callback_t *callback = zio->io_private; 5575c717a561Smaybee arc_buf_t *buf = callback->awcb_buf; 55760a4e9518Sgw arc_buf_hdr_t *hdr = buf->b_hdr; 5577dcbf3bd6SGeorge Wilson uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp); 5578c717a561Smaybee 557989c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 558089c86e32SChris Williamson ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 5581dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 5582e14bb325SJeff Bonwick 55830a4e9518Sgw /* 5584dcbf3bd6SGeorge Wilson * If we're reexecuting this zio because the pool suspended, then 5585dcbf3bd6SGeorge Wilson * cleanup any state that was previously set the first time the 55865602294fSDan Kimmel * callback was invoked. 55870a4e9518Sgw */ 5588dcbf3bd6SGeorge Wilson if (zio->io_flags & ZIO_FLAG_REEXECUTED) { 5589dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 5590dcbf3bd6SGeorge Wilson arc_buf_unwatch(buf); 5591770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 5592dcbf3bd6SGeorge Wilson if (arc_buf_is_shared(buf)) { 5593dcbf3bd6SGeorge Wilson arc_unshare_buf(hdr, buf); 5594dcbf3bd6SGeorge Wilson } else { 5595770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 5596dcbf3bd6SGeorge Wilson } 55970a4e9518Sgw } 55980a4e9518Sgw } 5599770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5600dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(hdr)); 5601dcbf3bd6SGeorge Wilson ASSERT(!arc_buf_is_shared(buf)); 5602dcbf3bd6SGeorge Wilson 5603dcbf3bd6SGeorge Wilson callback->awcb_ready(zio, buf, callback->awcb_private); 5604dcbf3bd6SGeorge Wilson 5605dcbf3bd6SGeorge Wilson if (HDR_IO_IN_PROGRESS(hdr)) 5606dcbf3bd6SGeorge Wilson ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); 5607dcbf3bd6SGeorge Wilson 5608dcbf3bd6SGeorge Wilson arc_cksum_compute(buf); 5609dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5610dcbf3bd6SGeorge Wilson 5611dcbf3bd6SGeorge Wilson enum zio_compress compress; 5612dcbf3bd6SGeorge Wilson if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5613dcbf3bd6SGeorge Wilson compress = ZIO_COMPRESS_OFF; 5614dcbf3bd6SGeorge Wilson } else { 5615dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp)); 5616dcbf3bd6SGeorge Wilson compress = BP_GET_COMPRESS(zio->io_bp); 5617dcbf3bd6SGeorge Wilson } 5618dcbf3bd6SGeorge Wilson HDR_SET_PSIZE(hdr, psize); 5619dcbf3bd6SGeorge Wilson arc_hdr_set_compress(hdr, compress); 5620dcbf3bd6SGeorge Wilson 5621770499e1SDan Kimmel 5622dcbf3bd6SGeorge Wilson /* 5623770499e1SDan Kimmel * Fill the hdr with data. If the hdr is compressed, the data we want 5624770499e1SDan Kimmel * is available from the zio, otherwise we can take it from the buf. 5625770499e1SDan Kimmel * 5626770499e1SDan Kimmel * We might be able to share the buf's data with the hdr here. However, 5627770499e1SDan Kimmel * doing so would cause the ARC to be full of linear ABDs if we write a 5628770499e1SDan Kimmel * lot of shareable data. As a compromise, we check whether scattered 5629770499e1SDan Kimmel * ABDs are allowed, and assume that if they are then the user wants 5630770499e1SDan Kimmel * the ARC to be primarily filled with them regardless of the data being 5631770499e1SDan Kimmel * written. Therefore, if they're allowed then we allocate one and copy 5632770499e1SDan Kimmel * the data into it; otherwise, we share the data directly if we can. 5633dcbf3bd6SGeorge Wilson */ 5634770499e1SDan Kimmel if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) { 5635770499e1SDan Kimmel arc_hdr_alloc_pabd(hdr); 5636770499e1SDan Kimmel 5637770499e1SDan Kimmel /* 5638770499e1SDan Kimmel * Ideally, we would always copy the io_abd into b_pabd, but the 5639770499e1SDan Kimmel * user may have disabled compressed ARC, thus we must check the 5640770499e1SDan Kimmel * hdr's compression setting rather than the io_bp's. 5641770499e1SDan Kimmel */ 5642770499e1SDan Kimmel if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5643770499e1SDan Kimmel ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=, 5644770499e1SDan Kimmel ZIO_COMPRESS_OFF); 5645770499e1SDan Kimmel ASSERT3U(psize, >, 0); 5646770499e1SDan Kimmel 5647770499e1SDan Kimmel abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize); 5648770499e1SDan Kimmel } else { 5649770499e1SDan Kimmel ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr)); 5650770499e1SDan Kimmel 5651770499e1SDan Kimmel abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data, 5652770499e1SDan Kimmel arc_buf_size(buf)); 5653770499e1SDan Kimmel } 5654dcbf3bd6SGeorge Wilson } else { 5655770499e1SDan Kimmel ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd)); 56565602294fSDan Kimmel ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf)); 5657dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5658dcbf3bd6SGeorge Wilson 5659dcbf3bd6SGeorge Wilson arc_share_buf(hdr, buf); 5660dcbf3bd6SGeorge Wilson } 5661770499e1SDan Kimmel 5662dcbf3bd6SGeorge Wilson arc_hdr_verify(hdr, zio->io_bp); 5663c717a561Smaybee } 5664c717a561Smaybee 56658df0bcf0SPaul Dagnelie static void 56668df0bcf0SPaul Dagnelie arc_write_children_ready(zio_t *zio) 56678df0bcf0SPaul Dagnelie { 56688df0bcf0SPaul Dagnelie arc_write_callback_t *callback = zio->io_private; 56698df0bcf0SPaul Dagnelie arc_buf_t *buf = callback->awcb_buf; 56708df0bcf0SPaul Dagnelie 56718df0bcf0SPaul Dagnelie callback->awcb_children_ready(zio, buf, callback->awcb_private); 56728df0bcf0SPaul Dagnelie } 56738df0bcf0SPaul Dagnelie 567469962b56SMatthew Ahrens /* 567569962b56SMatthew Ahrens * The SPA calls this callback for each physical write that happens on behalf 567669962b56SMatthew Ahrens * of a logical write. See the comment in dbuf_write_physdone() for details. 567769962b56SMatthew Ahrens */ 567869962b56SMatthew Ahrens static void 567969962b56SMatthew Ahrens arc_write_physdone(zio_t *zio) 568069962b56SMatthew Ahrens { 568169962b56SMatthew Ahrens arc_write_callback_t *cb = zio->io_private; 568269962b56SMatthew Ahrens if (cb->awcb_physdone != NULL) 568369962b56SMatthew Ahrens cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 568469962b56SMatthew Ahrens } 568569962b56SMatthew Ahrens 5686fa9e4066Sahrens static void 5687fa9e4066Sahrens arc_write_done(zio_t *zio) 5688fa9e4066Sahrens { 5689c717a561Smaybee arc_write_callback_t *callback = zio->io_private; 5690c717a561Smaybee arc_buf_t *buf = callback->awcb_buf; 5691c717a561Smaybee arc_buf_hdr_t *hdr = buf->b_hdr; 5692fa9e4066Sahrens 5693dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5694b24ab676SJeff Bonwick 5695b24ab676SJeff Bonwick if (zio->io_error == 0) { 5696dcbf3bd6SGeorge Wilson arc_hdr_verify(hdr, zio->io_bp); 5697dcbf3bd6SGeorge Wilson 56985d7b4d43SMatthew Ahrens if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 569943466aaeSMax Grossman buf_discard_identity(hdr); 570043466aaeSMax Grossman } else { 570143466aaeSMax Grossman hdr->b_dva = *BP_IDENTITY(zio->io_bp); 570243466aaeSMax Grossman hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 570343466aaeSMax Grossman } 5704b24ab676SJeff Bonwick } else { 5705dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 5706b24ab676SJeff Bonwick } 5707fa9e4066Sahrens 5708ea8dc4b6Seschrock /* 57095d7b4d43SMatthew Ahrens * If the block to be written was all-zero or compressed enough to be 57105d7b4d43SMatthew Ahrens * embedded in the BP, no write was performed so there will be no 57115d7b4d43SMatthew Ahrens * dva/birth/checksum. The buffer must therefore remain anonymous 57125d7b4d43SMatthew Ahrens * (and uncached). 5713ea8dc4b6Seschrock */ 5714dcbf3bd6SGeorge Wilson if (!HDR_EMPTY(hdr)) { 5715fa9e4066Sahrens arc_buf_hdr_t *exists; 5716fa9e4066Sahrens kmutex_t *hash_lock; 5717fa9e4066Sahrens 57185602294fSDan Kimmel ASSERT3U(zio->io_error, ==, 0); 5719b24ab676SJeff Bonwick 57206b4acc8bSahrens arc_cksum_verify(buf); 57216b4acc8bSahrens 5722fa9e4066Sahrens exists = buf_hash_insert(hdr, &hash_lock); 572389c86e32SChris Williamson if (exists != NULL) { 5724fa9e4066Sahrens /* 5725fa9e4066Sahrens * This can only happen if we overwrite for 5726fa9e4066Sahrens * sync-to-convergence, because we remove 5727fa9e4066Sahrens * buffers from the hash table when we arc_free(). 5728fa9e4066Sahrens */ 5729b24ab676SJeff Bonwick if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 5730b24ab676SJeff Bonwick if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5731b24ab676SJeff Bonwick panic("bad overwrite, hdr=%p exists=%p", 5732b24ab676SJeff Bonwick (void *)hdr, (void *)exists); 573389c86e32SChris Williamson ASSERT(refcount_is_zero( 573489c86e32SChris Williamson &exists->b_l1hdr.b_refcnt)); 5735b24ab676SJeff Bonwick arc_change_state(arc_anon, exists, hash_lock); 5736b24ab676SJeff Bonwick mutex_exit(hash_lock); 5737b24ab676SJeff Bonwick arc_hdr_destroy(exists); 5738b24ab676SJeff Bonwick exists = buf_hash_insert(hdr, &hash_lock); 5739b24ab676SJeff Bonwick ASSERT3P(exists, ==, NULL); 574080901aeaSGeorge Wilson } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 574180901aeaSGeorge Wilson /* nopwrite */ 574280901aeaSGeorge Wilson ASSERT(zio->io_prop.zp_nopwrite); 574380901aeaSGeorge Wilson if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 574480901aeaSGeorge Wilson panic("bad nopwrite, hdr=%p exists=%p", 574580901aeaSGeorge Wilson (void *)hdr, (void *)exists); 5746b24ab676SJeff Bonwick } else { 5747b24ab676SJeff Bonwick /* Dedup */ 5748dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt == 1); 574989c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5750b24ab676SJeff Bonwick ASSERT(BP_GET_DEDUP(zio->io_bp)); 5751b24ab676SJeff Bonwick ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5752ae46e4c7SMatthew Ahrens } 5753fa9e4066Sahrens } 5754dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5755088f3894Sahrens /* if it's not anon, we are doing a scrub */ 575689c86e32SChris Williamson if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5757088f3894Sahrens arc_access(hdr, hash_lock); 575844eda4d7Smaybee mutex_exit(hash_lock); 5759ea8dc4b6Seschrock } else { 5760dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5761fa9e4066Sahrens } 5762ea8dc4b6Seschrock 576389c86e32SChris Williamson ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5764b24ab676SJeff Bonwick callback->awcb_done(zio, buf, callback->awcb_private); 5765fa9e4066Sahrens 5766770499e1SDan Kimmel abd_put(zio->io_abd); 5767c717a561Smaybee kmem_free(callback, sizeof (arc_write_callback_t)); 5768fa9e4066Sahrens } 5769fa9e4066Sahrens 5770c717a561Smaybee zio_t * 5771dcbf3bd6SGeorge Wilson arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 5772dcbf3bd6SGeorge Wilson boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready, 57738df0bcf0SPaul Dagnelie arc_done_func_t *children_ready, arc_done_func_t *physdone, 577469962b56SMatthew Ahrens arc_done_func_t *done, void *private, zio_priority_t priority, 57757802d7bfSMatthew Ahrens int zio_flags, const zbookmark_phys_t *zb) 5776fa9e4066Sahrens { 5777fa9e4066Sahrens arc_buf_hdr_t *hdr = buf->b_hdr; 5778c717a561Smaybee arc_write_callback_t *callback; 5779e14bb325SJeff Bonwick zio_t *zio; 5780adaec86aSMatthew Ahrens zio_prop_t localprop = *zp; 5781fa9e4066Sahrens 5782dcbf3bd6SGeorge Wilson ASSERT3P(ready, !=, NULL); 5783dcbf3bd6SGeorge Wilson ASSERT3P(done, !=, NULL); 5784fa9e4066Sahrens ASSERT(!HDR_IO_ERROR(hdr)); 578589c86e32SChris Williamson ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5786dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5787dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 57883baa08fcSek if (l2arc) 5789dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 57905602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 5791adaec86aSMatthew Ahrens /* 5792adaec86aSMatthew Ahrens * We're writing a pre-compressed buffer. Make the 5793adaec86aSMatthew Ahrens * compression algorithm requested by the zio_prop_t match 5794adaec86aSMatthew Ahrens * the pre-compressed buffer's compression algorithm. 5795adaec86aSMatthew Ahrens */ 5796adaec86aSMatthew Ahrens localprop.zp_compress = HDR_GET_COMPRESS(hdr); 5797adaec86aSMatthew Ahrens 57985602294fSDan Kimmel ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf)); 57995602294fSDan Kimmel zio_flags |= ZIO_FLAG_RAW; 58005602294fSDan Kimmel } 5801c717a561Smaybee callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5802c717a561Smaybee callback->awcb_ready = ready; 58038df0bcf0SPaul Dagnelie callback->awcb_children_ready = children_ready; 580469962b56SMatthew Ahrens callback->awcb_physdone = physdone; 5805c717a561Smaybee callback->awcb_done = done; 5806c717a561Smaybee callback->awcb_private = private; 5807c717a561Smaybee callback->awcb_buf = buf; 5808088f3894Sahrens 5809dcbf3bd6SGeorge Wilson /* 5810770499e1SDan Kimmel * The hdr's b_pabd is now stale, free it now. A new data block 5811dcbf3bd6SGeorge Wilson * will be allocated when the zio pipeline calls arc_write_ready(). 5812dcbf3bd6SGeorge Wilson */ 5813770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 5814dcbf3bd6SGeorge Wilson /* 5815dcbf3bd6SGeorge Wilson * If the buf is currently sharing the data block with 5816dcbf3bd6SGeorge Wilson * the hdr then we need to break that relationship here. 5817dcbf3bd6SGeorge Wilson * The hdr will remain with a NULL data pointer and the 5818dcbf3bd6SGeorge Wilson * buf will take sole ownership of the block. 5819dcbf3bd6SGeorge Wilson */ 5820dcbf3bd6SGeorge Wilson if (arc_buf_is_shared(buf)) { 5821dcbf3bd6SGeorge Wilson arc_unshare_buf(hdr, buf); 5822dcbf3bd6SGeorge Wilson } else { 5823770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 5824dcbf3bd6SGeorge Wilson } 5825dcbf3bd6SGeorge Wilson VERIFY3P(buf->b_data, !=, NULL); 5826dcbf3bd6SGeorge Wilson arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); 5827dcbf3bd6SGeorge Wilson } 5828dcbf3bd6SGeorge Wilson ASSERT(!arc_buf_is_shared(buf)); 5829770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5830dcbf3bd6SGeorge Wilson 5831770499e1SDan Kimmel zio = zio_write(pio, spa, txg, bp, 5832770499e1SDan Kimmel abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)), 5833adaec86aSMatthew Ahrens HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready, 58348df0bcf0SPaul Dagnelie (children_ready != NULL) ? arc_write_children_ready : NULL, 58358df0bcf0SPaul Dagnelie arc_write_physdone, arc_write_done, callback, 583669962b56SMatthew Ahrens priority, zio_flags, zb); 5837fa9e4066Sahrens 5838c717a561Smaybee return (zio); 5839fa9e4066Sahrens } 5840fa9e4066Sahrens 58411ab7f2deSmaybee static int 584269962b56SMatthew Ahrens arc_memory_throttle(uint64_t reserve, uint64_t txg) 58431ab7f2deSmaybee { 58441ab7f2deSmaybee #ifdef _KERNEL 58451ab7f2deSmaybee uint64_t available_memory = ptob(freemem); 58461ab7f2deSmaybee static uint64_t page_load = 0; 58471ab7f2deSmaybee static uint64_t last_txg = 0; 58481ab7f2deSmaybee 58491ab7f2deSmaybee #if defined(__i386) 58501ab7f2deSmaybee available_memory = 58511ab7f2deSmaybee MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 58521ab7f2deSmaybee #endif 585369962b56SMatthew Ahrens 585469962b56SMatthew Ahrens if (freemem > physmem * arc_lotsfree_percent / 100) 58551ab7f2deSmaybee return (0); 58561ab7f2deSmaybee 58571ab7f2deSmaybee if (txg > last_txg) { 58581ab7f2deSmaybee last_txg = txg; 58591ab7f2deSmaybee page_load = 0; 58601ab7f2deSmaybee } 58611ab7f2deSmaybee /* 58621ab7f2deSmaybee * If we are in pageout, we know that memory is already tight, 58631ab7f2deSmaybee * the arc is already going to be evicting, so we just want to 58641ab7f2deSmaybee * continue to let page writes occur as quickly as possible. 58651ab7f2deSmaybee */ 58661ab7f2deSmaybee if (curproc == proc_pageout) { 58671ab7f2deSmaybee if (page_load > MAX(ptob(minfree), available_memory) / 4) 5868be6fd75aSMatthew Ahrens return (SET_ERROR(ERESTART)); 58691ab7f2deSmaybee /* Note: reserve is inflated, so we deflate */ 58701ab7f2deSmaybee page_load += reserve / 8; 58711ab7f2deSmaybee return (0); 58721ab7f2deSmaybee } else if (page_load > 0 && arc_reclaim_needed()) { 58731ab7f2deSmaybee /* memory is low, delay before restarting */ 58741ab7f2deSmaybee ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5875be6fd75aSMatthew Ahrens return (SET_ERROR(EAGAIN)); 58761ab7f2deSmaybee } 58771ab7f2deSmaybee page_load = 0; 58781ab7f2deSmaybee #endif 58791ab7f2deSmaybee return (0); 58801ab7f2deSmaybee } 58811ab7f2deSmaybee 5882fa9e4066Sahrens void 58831ab7f2deSmaybee arc_tempreserve_clear(uint64_t reserve) 5884fa9e4066Sahrens { 58851ab7f2deSmaybee atomic_add_64(&arc_tempreserve, -reserve); 5886fa9e4066Sahrens ASSERT((int64_t)arc_tempreserve >= 0); 5887fa9e4066Sahrens } 5888fa9e4066Sahrens 5889fa9e4066Sahrens int 58901ab7f2deSmaybee arc_tempreserve_space(uint64_t reserve, uint64_t txg) 5891fa9e4066Sahrens { 58921ab7f2deSmaybee int error; 58932fdbea25SAleksandr Guzovskiy uint64_t anon_size; 58941ab7f2deSmaybee 58951ab7f2deSmaybee if (reserve > arc_c/4 && !arc_no_grow) 58961ab7f2deSmaybee arc_c = MIN(arc_c_max, reserve * 4); 58971ab7f2deSmaybee if (reserve > arc_c) 5898be6fd75aSMatthew Ahrens return (SET_ERROR(ENOMEM)); 5899112fe045Smaybee 59002fdbea25SAleksandr Guzovskiy /* 59012fdbea25SAleksandr Guzovskiy * Don't count loaned bufs as in flight dirty data to prevent long 59022fdbea25SAleksandr Guzovskiy * network delays from blocking transactions that are ready to be 59032fdbea25SAleksandr Guzovskiy * assigned to a txg. 59042fdbea25SAleksandr Guzovskiy */ 59055602294fSDan Kimmel 59065602294fSDan Kimmel /* assert that it has not wrapped around */ 59075602294fSDan Kimmel ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 59085602294fSDan Kimmel 59092fd872a7SPrakash Surya anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 59102fd872a7SPrakash Surya arc_loaned_bytes), 0); 59112fdbea25SAleksandr Guzovskiy 59121ab7f2deSmaybee /* 59131ab7f2deSmaybee * Writes will, almost always, require additional memory allocations 5914f7170741SWill Andrews * in order to compress/encrypt/etc the data. We therefore need to 59151ab7f2deSmaybee * make sure that there is sufficient available memory for this. 59161ab7f2deSmaybee */ 591769962b56SMatthew Ahrens error = arc_memory_throttle(reserve, txg); 591869962b56SMatthew Ahrens if (error != 0) 59191ab7f2deSmaybee return (error); 59201ab7f2deSmaybee 5921fa9e4066Sahrens /* 5922112fe045Smaybee * Throttle writes when the amount of dirty data in the cache 5923112fe045Smaybee * gets too large. We try to keep the cache less than half full 5924112fe045Smaybee * of dirty blocks so that our sync times don't grow too large. 5925112fe045Smaybee * Note: if two requests come in concurrently, we might let them 5926112fe045Smaybee * both succeed, when one of them should fail. Not a huge deal. 5927fa9e4066Sahrens */ 59282fdbea25SAleksandr Guzovskiy 59292fdbea25SAleksandr Guzovskiy if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 59302fdbea25SAleksandr Guzovskiy anon_size > arc_c / 4) { 5931dcbf3bd6SGeorge Wilson uint64_t meta_esize = 5932dcbf3bd6SGeorge Wilson refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5933dcbf3bd6SGeorge Wilson uint64_t data_esize = 5934dcbf3bd6SGeorge Wilson refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 59350e8c6158Smaybee dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 59360e8c6158Smaybee "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5937dcbf3bd6SGeorge Wilson arc_tempreserve >> 10, meta_esize >> 10, 5938dcbf3bd6SGeorge Wilson data_esize >> 10, reserve >> 10, arc_c >> 10); 5939be6fd75aSMatthew Ahrens return (SET_ERROR(ERESTART)); 5940fa9e4066Sahrens } 59411ab7f2deSmaybee atomic_add_64(&arc_tempreserve, reserve); 5942fa9e4066Sahrens return (0); 5943fa9e4066Sahrens } 5944fa9e4066Sahrens 59454076b1bfSPrakash Surya static void 59464076b1bfSPrakash Surya arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 59474076b1bfSPrakash Surya kstat_named_t *evict_data, kstat_named_t *evict_metadata) 59484076b1bfSPrakash Surya { 59492fd872a7SPrakash Surya size->value.ui64 = refcount_count(&state->arcs_size); 5950dcbf3bd6SGeorge Wilson evict_data->value.ui64 = 5951dcbf3bd6SGeorge Wilson refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); 5952dcbf3bd6SGeorge Wilson evict_metadata->value.ui64 = 5953dcbf3bd6SGeorge Wilson refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); 59544076b1bfSPrakash Surya } 59554076b1bfSPrakash Surya 59564076b1bfSPrakash Surya static int 59574076b1bfSPrakash Surya arc_kstat_update(kstat_t *ksp, int rw) 59584076b1bfSPrakash Surya { 59594076b1bfSPrakash Surya arc_stats_t *as = ksp->ks_data; 59604076b1bfSPrakash Surya 59614076b1bfSPrakash Surya if (rw == KSTAT_WRITE) { 59624076b1bfSPrakash Surya return (EACCES); 59634076b1bfSPrakash Surya } else { 59644076b1bfSPrakash Surya arc_kstat_update_state(arc_anon, 59654076b1bfSPrakash Surya &as->arcstat_anon_size, 59664076b1bfSPrakash Surya &as->arcstat_anon_evictable_data, 59674076b1bfSPrakash Surya &as->arcstat_anon_evictable_metadata); 59684076b1bfSPrakash Surya arc_kstat_update_state(arc_mru, 59694076b1bfSPrakash Surya &as->arcstat_mru_size, 59704076b1bfSPrakash Surya &as->arcstat_mru_evictable_data, 59714076b1bfSPrakash Surya &as->arcstat_mru_evictable_metadata); 59724076b1bfSPrakash Surya arc_kstat_update_state(arc_mru_ghost, 59734076b1bfSPrakash Surya &as->arcstat_mru_ghost_size, 59744076b1bfSPrakash Surya &as->arcstat_mru_ghost_evictable_data, 59754076b1bfSPrakash Surya &as->arcstat_mru_ghost_evictable_metadata); 59764076b1bfSPrakash Surya arc_kstat_update_state(arc_mfu, 59774076b1bfSPrakash Surya &as->arcstat_mfu_size, 59784076b1bfSPrakash Surya &as->arcstat_mfu_evictable_data, 59794076b1bfSPrakash Surya &as->arcstat_mfu_evictable_metadata); 59804076b1bfSPrakash Surya arc_kstat_update_state(arc_mfu_ghost, 59814076b1bfSPrakash Surya &as->arcstat_mfu_ghost_size, 59824076b1bfSPrakash Surya &as->arcstat_mfu_ghost_evictable_data, 59834076b1bfSPrakash Surya &as->arcstat_mfu_ghost_evictable_metadata); 59843a2d8a1bSPaul Dagnelie 59853a2d8a1bSPaul Dagnelie ARCSTAT(arcstat_size) = aggsum_value(&arc_size); 59863a2d8a1bSPaul Dagnelie ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used); 59873a2d8a1bSPaul Dagnelie ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size); 59883a2d8a1bSPaul Dagnelie ARCSTAT(arcstat_metadata_size) = 59893a2d8a1bSPaul Dagnelie aggsum_value(&astat_metadata_size); 59903a2d8a1bSPaul Dagnelie ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size); 59913a2d8a1bSPaul Dagnelie ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size); 59923a2d8a1bSPaul Dagnelie ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size); 59934076b1bfSPrakash Surya } 59944076b1bfSPrakash Surya 59954076b1bfSPrakash Surya return (0); 59964076b1bfSPrakash Surya } 59974076b1bfSPrakash Surya 5998dcbf3bd6SGeorge Wilson /* 5999dcbf3bd6SGeorge Wilson * This function *must* return indices evenly distributed between all 6000dcbf3bd6SGeorge Wilson * sublists of the multilist. This is needed due to how the ARC eviction 6001dcbf3bd6SGeorge Wilson * code is laid out; arc_evict_state() assumes ARC buffers are evenly 6002dcbf3bd6SGeorge Wilson * distributed between all sublists and uses this assumption when 6003dcbf3bd6SGeorge Wilson * deciding which sublist to evict from and how much to evict from it. 6004dcbf3bd6SGeorge Wilson */ 6005dcbf3bd6SGeorge Wilson unsigned int 6006dcbf3bd6SGeorge Wilson arc_state_multilist_index_func(multilist_t *ml, void *obj) 6007dcbf3bd6SGeorge Wilson { 6008dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = obj; 6009dcbf3bd6SGeorge Wilson 6010dcbf3bd6SGeorge Wilson /* 6011dcbf3bd6SGeorge Wilson * We rely on b_dva to generate evenly distributed index 6012dcbf3bd6SGeorge Wilson * numbers using buf_hash below. So, as an added precaution, 6013dcbf3bd6SGeorge Wilson * let's make sure we never add empty buffers to the arc lists. 6014dcbf3bd6SGeorge Wilson */ 6015dcbf3bd6SGeorge Wilson ASSERT(!HDR_EMPTY(hdr)); 6016dcbf3bd6SGeorge Wilson 6017dcbf3bd6SGeorge Wilson /* 6018dcbf3bd6SGeorge Wilson * The assumption here, is the hash value for a given 6019dcbf3bd6SGeorge Wilson * arc_buf_hdr_t will remain constant throughout it's lifetime 6020dcbf3bd6SGeorge Wilson * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 6021dcbf3bd6SGeorge Wilson * Thus, we don't need to store the header's sublist index 6022dcbf3bd6SGeorge Wilson * on insertion, as this index can be recalculated on removal. 6023dcbf3bd6SGeorge Wilson * 6024dcbf3bd6SGeorge Wilson * Also, the low order bits of the hash value are thought to be 6025dcbf3bd6SGeorge Wilson * distributed evenly. Otherwise, in the case that the multilist 6026dcbf3bd6SGeorge Wilson * has a power of two number of sublists, each sublists' usage 6027dcbf3bd6SGeorge Wilson * would not be evenly distributed. 6028dcbf3bd6SGeorge Wilson */ 6029dcbf3bd6SGeorge Wilson return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 6030dcbf3bd6SGeorge Wilson multilist_get_num_sublists(ml)); 6031dcbf3bd6SGeorge Wilson } 6032dcbf3bd6SGeorge Wilson 6033dcbf3bd6SGeorge Wilson static void 6034dcbf3bd6SGeorge Wilson arc_state_init(void) 6035dcbf3bd6SGeorge Wilson { 6036dcbf3bd6SGeorge Wilson arc_anon = &ARC_anon; 6037dcbf3bd6SGeorge Wilson arc_mru = &ARC_mru; 6038dcbf3bd6SGeorge Wilson arc_mru_ghost = &ARC_mru_ghost; 6039dcbf3bd6SGeorge Wilson arc_mfu = &ARC_mfu; 6040dcbf3bd6SGeorge Wilson arc_mfu_ghost = &ARC_mfu_ghost; 6041dcbf3bd6SGeorge Wilson arc_l2c_only = &ARC_l2c_only; 6042dcbf3bd6SGeorge Wilson 604394c2d0ebSMatthew Ahrens arc_mru->arcs_list[ARC_BUFC_METADATA] = 604494c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 6045dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 604610fbdecbSMatthew Ahrens arc_state_multilist_index_func); 604794c2d0ebSMatthew Ahrens arc_mru->arcs_list[ARC_BUFC_DATA] = 604894c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 6049dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 605010fbdecbSMatthew Ahrens arc_state_multilist_index_func); 605194c2d0ebSMatthew Ahrens arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] = 605294c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 6053dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 605410fbdecbSMatthew Ahrens arc_state_multilist_index_func); 605594c2d0ebSMatthew Ahrens arc_mru_ghost->arcs_list[ARC_BUFC_DATA] = 605694c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 6057dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 605810fbdecbSMatthew Ahrens arc_state_multilist_index_func); 605994c2d0ebSMatthew Ahrens arc_mfu->arcs_list[ARC_BUFC_METADATA] = 606094c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 6061dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 606210fbdecbSMatthew Ahrens arc_state_multilist_index_func); 606394c2d0ebSMatthew Ahrens arc_mfu->arcs_list[ARC_BUFC_DATA] = 606494c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 6065dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 606610fbdecbSMatthew Ahrens arc_state_multilist_index_func); 606794c2d0ebSMatthew Ahrens arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] = 606894c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 6069dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 607010fbdecbSMatthew Ahrens arc_state_multilist_index_func); 607194c2d0ebSMatthew Ahrens arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] = 607294c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 6073dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 607410fbdecbSMatthew Ahrens arc_state_multilist_index_func); 607594c2d0ebSMatthew Ahrens arc_l2c_only->arcs_list[ARC_BUFC_METADATA] = 607694c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 6077dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 607810fbdecbSMatthew Ahrens arc_state_multilist_index_func); 607994c2d0ebSMatthew Ahrens arc_l2c_only->arcs_list[ARC_BUFC_DATA] = 608094c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 6081dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 608210fbdecbSMatthew Ahrens arc_state_multilist_index_func); 6083dcbf3bd6SGeorge Wilson 6084dcbf3bd6SGeorge Wilson refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6085dcbf3bd6SGeorge Wilson refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6086dcbf3bd6SGeorge Wilson refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6087dcbf3bd6SGeorge Wilson refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6088dcbf3bd6SGeorge Wilson refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6089dcbf3bd6SGeorge Wilson refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6090dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6091dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6092dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6093dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6094dcbf3bd6SGeorge Wilson refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6095dcbf3bd6SGeorge Wilson refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6096dcbf3bd6SGeorge Wilson 6097dcbf3bd6SGeorge Wilson refcount_create(&arc_anon->arcs_size); 6098dcbf3bd6SGeorge Wilson refcount_create(&arc_mru->arcs_size); 6099dcbf3bd6SGeorge Wilson refcount_create(&arc_mru_ghost->arcs_size); 6100dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu->arcs_size); 6101dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu_ghost->arcs_size); 6102dcbf3bd6SGeorge Wilson refcount_create(&arc_l2c_only->arcs_size); 61033a2d8a1bSPaul Dagnelie 61043a2d8a1bSPaul Dagnelie aggsum_init(&arc_meta_used, 0); 61053a2d8a1bSPaul Dagnelie aggsum_init(&arc_size, 0); 61063a2d8a1bSPaul Dagnelie aggsum_init(&astat_data_size, 0); 61073a2d8a1bSPaul Dagnelie aggsum_init(&astat_metadata_size, 0); 61083a2d8a1bSPaul Dagnelie aggsum_init(&astat_hdr_size, 0); 61093a2d8a1bSPaul Dagnelie aggsum_init(&astat_other_size, 0); 61103a2d8a1bSPaul Dagnelie aggsum_init(&astat_l2_hdr_size, 0); 6111dcbf3bd6SGeorge Wilson } 6112dcbf3bd6SGeorge Wilson 6113dcbf3bd6SGeorge Wilson static void 6114dcbf3bd6SGeorge Wilson arc_state_fini(void) 6115dcbf3bd6SGeorge Wilson { 6116dcbf3bd6SGeorge Wilson refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6117dcbf3bd6SGeorge Wilson refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6118dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6119dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6120dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6121dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6122dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6123dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6124dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6125dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6126dcbf3bd6SGeorge Wilson refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6127dcbf3bd6SGeorge Wilson refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6128dcbf3bd6SGeorge Wilson 6129dcbf3bd6SGeorge Wilson refcount_destroy(&arc_anon->arcs_size); 6130dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru->arcs_size); 6131dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru_ghost->arcs_size); 6132dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu->arcs_size); 6133dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu_ghost->arcs_size); 6134dcbf3bd6SGeorge Wilson refcount_destroy(&arc_l2c_only->arcs_size); 6135dcbf3bd6SGeorge Wilson 613694c2d0ebSMatthew Ahrens multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]); 613794c2d0ebSMatthew Ahrens multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 613894c2d0ebSMatthew Ahrens multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]); 613994c2d0ebSMatthew Ahrens multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 614094c2d0ebSMatthew Ahrens multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]); 614194c2d0ebSMatthew Ahrens multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 614294c2d0ebSMatthew Ahrens multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]); 614394c2d0ebSMatthew Ahrens multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 6144dcbf3bd6SGeorge Wilson } 6145dcbf3bd6SGeorge Wilson 6146dcbf3bd6SGeorge Wilson uint64_t 6147dcbf3bd6SGeorge Wilson arc_max_bytes(void) 6148244781f1SPrakash Surya { 6149dcbf3bd6SGeorge Wilson return (arc_c_max); 6150244781f1SPrakash Surya } 6151244781f1SPrakash Surya 6152fa9e4066Sahrens void 6153fa9e4066Sahrens arc_init(void) 6154fa9e4066Sahrens { 61552ec99e3eSMatthew Ahrens /* 61562ec99e3eSMatthew Ahrens * allmem is "all memory that we could possibly use". 61572ec99e3eSMatthew Ahrens */ 61582ec99e3eSMatthew Ahrens #ifdef _KERNEL 61592ec99e3eSMatthew Ahrens uint64_t allmem = ptob(physmem - swapfs_minfree); 61602ec99e3eSMatthew Ahrens #else 61612ec99e3eSMatthew Ahrens uint64_t allmem = (physmem * PAGESIZE) / 2; 61622ec99e3eSMatthew Ahrens #endif 6163de753e34SBrad Lewis mutex_init(&arc_adjust_lock, NULL, MUTEX_DEFAULT, NULL); 6164de753e34SBrad Lewis cv_init(&arc_adjust_waiters_cv, NULL, CV_DEFAULT, NULL); 6165244781f1SPrakash Surya 616613506d1eSmaybee /* Convert seconds to clock ticks */ 6167b19a79ecSperrin arc_min_prefetch_lifespan = 1 * hz; 616813506d1eSmaybee 6169112fe045Smaybee /* set min cache to 1/32 of all memory, or 64MB, whichever is more */ 61702ec99e3eSMatthew Ahrens arc_c_min = MAX(allmem / 32, 64 << 20); 6171112fe045Smaybee /* set max to 3/4 of all memory, or all but 1GB, whichever is more */ 61722ec99e3eSMatthew Ahrens if (allmem >= 1 << 30) 61732ec99e3eSMatthew Ahrens arc_c_max = allmem - (1 << 30); 6174fa9e4066Sahrens else 617544cb6abcSbmc arc_c_max = arc_c_min; 61762ec99e3eSMatthew Ahrens arc_c_max = MAX(allmem * 3 / 4, arc_c_max); 6177a2eea2e1Sahrens 61788fe00bfbSMatthew Ahrens /* 61798fe00bfbSMatthew Ahrens * In userland, there's only the memory pressure that we artificially 61808fe00bfbSMatthew Ahrens * create (see arc_available_memory()). Don't let arc_c get too 61818fe00bfbSMatthew Ahrens * small, because it can cause transactions to be larger than 61828fe00bfbSMatthew Ahrens * arc_c, causing arc_tempreserve_space() to fail. 61838fe00bfbSMatthew Ahrens */ 61848fe00bfbSMatthew Ahrens #ifndef _KERNEL 61858fe00bfbSMatthew Ahrens arc_c_min = arc_c_max / 2; 61868fe00bfbSMatthew Ahrens #endif 61878fe00bfbSMatthew Ahrens 6188a2eea2e1Sahrens /* 6189a2eea2e1Sahrens * Allow the tunables to override our calculations if they are 6190a2eea2e1Sahrens * reasonable (ie. over 64MB) 6191a2eea2e1Sahrens */ 6192e5961710SMatthew Ahrens if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) { 619344cb6abcSbmc arc_c_max = zfs_arc_max; 6194e5961710SMatthew Ahrens arc_c_min = MIN(arc_c_min, arc_c_max); 6195e5961710SMatthew Ahrens } 61962ec99e3eSMatthew Ahrens if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max) 619744cb6abcSbmc arc_c_min = zfs_arc_min; 6198a2eea2e1Sahrens 619944cb6abcSbmc arc_c = arc_c_max; 620044cb6abcSbmc arc_p = (arc_c >> 1); 6201fa9e4066Sahrens 62020e8c6158Smaybee /* limit meta-data to 1/4 of the arc capacity */ 62030e8c6158Smaybee arc_meta_limit = arc_c_max / 4; 62041116048bSek 6205af868f46SMatthew Ahrens #ifdef _KERNEL 6206af868f46SMatthew Ahrens /* 6207af868f46SMatthew Ahrens * Metadata is stored in the kernel's heap. Don't let us 6208af868f46SMatthew Ahrens * use more than half the heap for the ARC. 6209af868f46SMatthew Ahrens */ 6210af868f46SMatthew Ahrens arc_meta_limit = MIN(arc_meta_limit, 6211af868f46SMatthew Ahrens vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2); 6212af868f46SMatthew Ahrens #endif 6213af868f46SMatthew Ahrens 62141116048bSek /* Allow the tunable to override if it is reasonable */ 62151116048bSek if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 62161116048bSek arc_meta_limit = zfs_arc_meta_limit; 62171116048bSek 62180e8c6158Smaybee if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 62190e8c6158Smaybee arc_c_min = arc_meta_limit / 2; 62200e8c6158Smaybee 62213a5286a1SMatthew Ahrens if (zfs_arc_meta_min > 0) { 62223a5286a1SMatthew Ahrens arc_meta_min = zfs_arc_meta_min; 62233a5286a1SMatthew Ahrens } else { 62243a5286a1SMatthew Ahrens arc_meta_min = arc_c_min / 2; 62253a5286a1SMatthew Ahrens } 62263a5286a1SMatthew Ahrens 62275a98e54bSBrendan Gregg - Sun Microsystems if (zfs_arc_grow_retry > 0) 62285a98e54bSBrendan Gregg - Sun Microsystems arc_grow_retry = zfs_arc_grow_retry; 62295a98e54bSBrendan Gregg - Sun Microsystems 62305a98e54bSBrendan Gregg - Sun Microsystems if (zfs_arc_shrink_shift > 0) 62315a98e54bSBrendan Gregg - Sun Microsystems arc_shrink_shift = zfs_arc_shrink_shift; 62325a98e54bSBrendan Gregg - Sun Microsystems 62332ec99e3eSMatthew Ahrens /* 62342ec99e3eSMatthew Ahrens * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 62352ec99e3eSMatthew Ahrens */ 62362ec99e3eSMatthew Ahrens if (arc_no_grow_shift >= arc_shrink_shift) 62372ec99e3eSMatthew Ahrens arc_no_grow_shift = arc_shrink_shift - 1; 62382ec99e3eSMatthew Ahrens 62395a98e54bSBrendan Gregg - Sun Microsystems if (zfs_arc_p_min_shift > 0) 62405a98e54bSBrendan Gregg - Sun Microsystems arc_p_min_shift = zfs_arc_p_min_shift; 62415a98e54bSBrendan Gregg - Sun Microsystems 6242fa9e4066Sahrens /* if kmem_flags are set, lets try to use less memory */ 6243fa9e4066Sahrens if (kmem_debugging()) 624444cb6abcSbmc arc_c = arc_c / 2; 624544cb6abcSbmc if (arc_c < arc_c_min) 624644cb6abcSbmc arc_c = arc_c_min; 624744cb6abcSbmc 6248dcbf3bd6SGeorge Wilson arc_state_init(); 6249fa9e4066Sahrens 6250de753e34SBrad Lewis /* 6251de753e34SBrad Lewis * The arc must be "uninitialized", so that hdr_recl() (which is 6252de753e34SBrad Lewis * registered by buf_init()) will not access arc_reap_zthr before 6253de753e34SBrad Lewis * it is created. 6254de753e34SBrad Lewis */ 6255de753e34SBrad Lewis ASSERT(!arc_initialized); 6256de753e34SBrad Lewis buf_init(); 6257fa9e4066Sahrens 625844cb6abcSbmc arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 625944cb6abcSbmc sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 626044cb6abcSbmc 626144cb6abcSbmc if (arc_ksp != NULL) { 626244cb6abcSbmc arc_ksp->ks_data = &arc_stats; 62634076b1bfSPrakash Surya arc_ksp->ks_update = arc_kstat_update; 626444cb6abcSbmc kstat_install(arc_ksp); 626544cb6abcSbmc } 626644cb6abcSbmc 6267de753e34SBrad Lewis arc_adjust_zthr = zthr_create(arc_adjust_cb_check, 6268de753e34SBrad Lewis arc_adjust_cb, NULL); 6269de753e34SBrad Lewis arc_reap_zthr = zthr_create_timer(arc_reap_cb_check, 6270de753e34SBrad Lewis arc_reap_cb, NULL, SEC2NSEC(1)); 627149e3519aSmaybee 6272de753e34SBrad Lewis arc_initialized = B_TRUE; 62733a737e0dSbrendan arc_warm = B_FALSE; 62741ab7f2deSmaybee 627569962b56SMatthew Ahrens /* 627669962b56SMatthew Ahrens * Calculate maximum amount of dirty data per pool. 627769962b56SMatthew Ahrens * 627869962b56SMatthew Ahrens * If it has been set by /etc/system, take that. 627969962b56SMatthew Ahrens * Otherwise, use a percentage of physical memory defined by 628069962b56SMatthew Ahrens * zfs_dirty_data_max_percent (default 10%) with a cap at 628169962b56SMatthew Ahrens * zfs_dirty_data_max_max (default 4GB). 628269962b56SMatthew Ahrens */ 628369962b56SMatthew Ahrens if (zfs_dirty_data_max == 0) { 628469962b56SMatthew Ahrens zfs_dirty_data_max = physmem * PAGESIZE * 628569962b56SMatthew Ahrens zfs_dirty_data_max_percent / 100; 628669962b56SMatthew Ahrens zfs_dirty_data_max = MIN(zfs_dirty_data_max, 628769962b56SMatthew Ahrens zfs_dirty_data_max_max); 628869962b56SMatthew Ahrens } 6289fa9e4066Sahrens } 6290fa9e4066Sahrens 6291fa9e4066Sahrens void 6292fa9e4066Sahrens arc_fini(void) 6293fa9e4066Sahrens { 6294dcbf3bd6SGeorge Wilson /* Use B_TRUE to ensure *all* buffers are evicted */ 6295dcbf3bd6SGeorge Wilson arc_flush(NULL, B_TRUE); 6296fa9e4066Sahrens 6297de753e34SBrad Lewis arc_initialized = B_FALSE; 6298fa9e4066Sahrens 629944cb6abcSbmc if (arc_ksp != NULL) { 630044cb6abcSbmc kstat_delete(arc_ksp); 630144cb6abcSbmc arc_ksp = NULL; 630244cb6abcSbmc } 630344cb6abcSbmc 6304de753e34SBrad Lewis (void) zthr_cancel(arc_adjust_zthr); 6305de753e34SBrad Lewis zthr_destroy(arc_adjust_zthr); 6306de753e34SBrad Lewis 6307de753e34SBrad Lewis (void) zthr_cancel(arc_reap_zthr); 6308de753e34SBrad Lewis zthr_destroy(arc_reap_zthr); 6309de753e34SBrad Lewis 6310de753e34SBrad Lewis mutex_destroy(&arc_adjust_lock); 6311de753e34SBrad Lewis cv_destroy(&arc_adjust_waiters_cv); 6312244781f1SPrakash Surya 6313dcbf3bd6SGeorge Wilson arc_state_fini(); 6314fa9e4066Sahrens buf_fini(); 63152fdbea25SAleksandr Guzovskiy 631689c86e32SChris Williamson ASSERT0(arc_loaned_bytes); 6317fa9e4066Sahrens } 6318fa94a07fSbrendan 6319fa94a07fSbrendan /* 6320fa94a07fSbrendan * Level 2 ARC 6321fa94a07fSbrendan * 6322fa94a07fSbrendan * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 6323fa94a07fSbrendan * It uses dedicated storage devices to hold cached data, which are populated 6324fa94a07fSbrendan * using large infrequent writes. The main role of this cache is to boost 6325fa94a07fSbrendan * the performance of random read workloads. The intended L2ARC devices 6326fa94a07fSbrendan * include short-stroked disks, solid state disks, and other media with 6327fa94a07fSbrendan * substantially faster read latency than disk. 6328fa94a07fSbrendan * 6329fa94a07fSbrendan * +-----------------------+ 6330fa94a07fSbrendan * | ARC | 6331fa94a07fSbrendan * +-----------------------+ 6332fa94a07fSbrendan * | ^ ^ 6333fa94a07fSbrendan * | | | 6334fa94a07fSbrendan * l2arc_feed_thread() arc_read() 6335fa94a07fSbrendan * | | | 6336fa94a07fSbrendan * | l2arc read | 6337fa94a07fSbrendan * V | | 6338fa94a07fSbrendan * +---------------+ | 6339fa94a07fSbrendan * | L2ARC | | 6340fa94a07fSbrendan * +---------------+ | 6341fa94a07fSbrendan * | ^ | 6342fa94a07fSbrendan * l2arc_write() | | 6343fa94a07fSbrendan * | | | 6344fa94a07fSbrendan * V | | 6345fa94a07fSbrendan * +-------+ +-------+ 6346fa94a07fSbrendan * | vdev | | vdev | 6347fa94a07fSbrendan * | cache | | cache | 6348fa94a07fSbrendan * +-------+ +-------+ 6349fa94a07fSbrendan * +=========+ .-----. 6350fa94a07fSbrendan * : L2ARC : |-_____-| 6351fa94a07fSbrendan * : devices : | Disks | 6352fa94a07fSbrendan * +=========+ `-_____-' 6353fa94a07fSbrendan * 6354fa94a07fSbrendan * Read requests are satisfied from the following sources, in order: 6355fa94a07fSbrendan * 6356fa94a07fSbrendan * 1) ARC 6357fa94a07fSbrendan * 2) vdev cache of L2ARC devices 6358fa94a07fSbrendan * 3) L2ARC devices 6359fa94a07fSbrendan * 4) vdev cache of disks 6360fa94a07fSbrendan * 5) disks 6361fa94a07fSbrendan * 6362fa94a07fSbrendan * Some L2ARC device types exhibit extremely slow write performance. 6363fa94a07fSbrendan * To accommodate for this there are some significant differences between 6364fa94a07fSbrendan * the L2ARC and traditional cache design: 6365fa94a07fSbrendan * 6366fa94a07fSbrendan * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 6367fa94a07fSbrendan * the ARC behave as usual, freeing buffers and placing headers on ghost 6368fa94a07fSbrendan * lists. The ARC does not send buffers to the L2ARC during eviction as 6369fa94a07fSbrendan * this would add inflated write latencies for all ARC memory pressure. 6370fa94a07fSbrendan * 6371fa94a07fSbrendan * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 6372fa94a07fSbrendan * It does this by periodically scanning buffers from the eviction-end of 6373fa94a07fSbrendan * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 6374aad02571SSaso Kiselkov * not already there. It scans until a headroom of buffers is satisfied, 6375aad02571SSaso Kiselkov * which itself is a buffer for ARC eviction. If a compressible buffer is 6376aad02571SSaso Kiselkov * found during scanning and selected for writing to an L2ARC device, we 6377aad02571SSaso Kiselkov * temporarily boost scanning headroom during the next scan cycle to make 6378aad02571SSaso Kiselkov * sure we adapt to compression effects (which might significantly reduce 6379aad02571SSaso Kiselkov * the data volume we write to L2ARC). The thread that does this is 6380fa94a07fSbrendan * l2arc_feed_thread(), illustrated below; example sizes are included to 6381fa94a07fSbrendan * provide a better sense of ratio than this diagram: 6382fa94a07fSbrendan * 6383fa94a07fSbrendan * head --> tail 6384fa94a07fSbrendan * +---------------------+----------+ 6385fa94a07fSbrendan * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 6386fa94a07fSbrendan * +---------------------+----------+ | o L2ARC eligible 6387fa94a07fSbrendan * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 6388fa94a07fSbrendan * +---------------------+----------+ | 6389fa94a07fSbrendan * 15.9 Gbytes ^ 32 Mbytes | 6390fa94a07fSbrendan * headroom | 6391fa94a07fSbrendan * l2arc_feed_thread() 6392fa94a07fSbrendan * | 6393fa94a07fSbrendan * l2arc write hand <--[oooo]--' 6394fa94a07fSbrendan * | 8 Mbyte 6395fa94a07fSbrendan * | write max 6396fa94a07fSbrendan * V 6397fa94a07fSbrendan * +==============================+ 6398fa94a07fSbrendan * L2ARC dev |####|#|###|###| |####| ... | 6399fa94a07fSbrendan * +==============================+ 6400fa94a07fSbrendan * 32 Gbytes 6401fa94a07fSbrendan * 6402fa94a07fSbrendan * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 6403fa94a07fSbrendan * evicted, then the L2ARC has cached a buffer much sooner than it probably 6404fa94a07fSbrendan * needed to, potentially wasting L2ARC device bandwidth and storage. It is 6405fa94a07fSbrendan * safe to say that this is an uncommon case, since buffers at the end of 6406fa94a07fSbrendan * the ARC lists have moved there due to inactivity. 6407fa94a07fSbrendan * 6408fa94a07fSbrendan * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 6409fa94a07fSbrendan * then the L2ARC simply misses copying some buffers. This serves as a 6410fa94a07fSbrendan * pressure valve to prevent heavy read workloads from both stalling the ARC 6411fa94a07fSbrendan * with waits and clogging the L2ARC with writes. This also helps prevent 6412fa94a07fSbrendan * the potential for the L2ARC to churn if it attempts to cache content too 6413fa94a07fSbrendan * quickly, such as during backups of the entire pool. 6414fa94a07fSbrendan * 64153a737e0dSbrendan * 5. After system boot and before the ARC has filled main memory, there are 64163a737e0dSbrendan * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 64173a737e0dSbrendan * lists can remain mostly static. Instead of searching from tail of these 64183a737e0dSbrendan * lists as pictured, the l2arc_feed_thread() will search from the list heads 64193a737e0dSbrendan * for eligible buffers, greatly increasing its chance of finding them. 64203a737e0dSbrendan * 64213a737e0dSbrendan * The L2ARC device write speed is also boosted during this time so that 64223a737e0dSbrendan * the L2ARC warms up faster. Since there have been no ARC evictions yet, 64233a737e0dSbrendan * there are no L2ARC reads, and no fear of degrading read performance 64243a737e0dSbrendan * through increased writes. 64253a737e0dSbrendan * 64263a737e0dSbrendan * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 6427fa94a07fSbrendan * the vdev queue can aggregate them into larger and fewer writes. Each 6428fa94a07fSbrendan * device is written to in a rotor fashion, sweeping writes through 6429fa94a07fSbrendan * available space then repeating. 6430fa94a07fSbrendan * 64313a737e0dSbrendan * 7. The L2ARC does not store dirty content. It never needs to flush 6432fa94a07fSbrendan * write buffers back to disk based storage. 6433fa94a07fSbrendan * 64343a737e0dSbrendan * 8. If an ARC buffer is written (and dirtied) which also exists in the 6435fa94a07fSbrendan * L2ARC, the now stale L2ARC buffer is immediately dropped. 6436fa94a07fSbrendan * 6437fa94a07fSbrendan * The performance of the L2ARC can be tweaked by a number of tunables, which 6438fa94a07fSbrendan * may be necessary for different workloads: 6439fa94a07fSbrendan * 6440fa94a07fSbrendan * l2arc_write_max max write bytes per interval 64413a737e0dSbrendan * l2arc_write_boost extra write bytes during device warmup 6442fa94a07fSbrendan * l2arc_noprefetch skip caching prefetched buffers 6443fa94a07fSbrendan * l2arc_headroom number of max device writes to precache 6444aad02571SSaso Kiselkov * l2arc_headroom_boost when we find compressed buffers during ARC 6445aad02571SSaso Kiselkov * scanning, we multiply headroom by this 6446aad02571SSaso Kiselkov * percentage factor for the next scan cycle, 6447aad02571SSaso Kiselkov * since more compressed buffers are likely to 6448aad02571SSaso Kiselkov * be present 6449fa94a07fSbrendan * l2arc_feed_secs seconds between L2ARC writing 6450fa94a07fSbrendan * 6451fa94a07fSbrendan * Tunables may be removed or added as future performance improvements are 6452fa94a07fSbrendan * integrated, and also may become zpool properties. 64535a98e54bSBrendan Gregg - Sun Microsystems * 64545a98e54bSBrendan Gregg - Sun Microsystems * There are three key functions that control how the L2ARC warms up: 64555a98e54bSBrendan Gregg - Sun Microsystems * 64565a98e54bSBrendan Gregg - Sun Microsystems * l2arc_write_eligible() check if a buffer is eligible to cache 64575a98e54bSBrendan Gregg - Sun Microsystems * l2arc_write_size() calculate how much to write 64585a98e54bSBrendan Gregg - Sun Microsystems * l2arc_write_interval() calculate sleep delay between writes 64595a98e54bSBrendan Gregg - Sun Microsystems * 64605a98e54bSBrendan Gregg - Sun Microsystems * These three functions determine what to write, how much, and how quickly 64615a98e54bSBrendan Gregg - Sun Microsystems * to send writes. 6462fa94a07fSbrendan */ 6463fa94a07fSbrendan 64645a98e54bSBrendan Gregg - Sun Microsystems static boolean_t 64657adb730bSGeorge Wilson l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 64665a98e54bSBrendan Gregg - Sun Microsystems { 64675a98e54bSBrendan Gregg - Sun Microsystems /* 64685a98e54bSBrendan Gregg - Sun Microsystems * A buffer is *not* eligible for the L2ARC if it: 64695a98e54bSBrendan Gregg - Sun Microsystems * 1. belongs to a different spa. 64705ea40c06SBrendan Gregg - Sun Microsystems * 2. is already cached on the L2ARC. 64715ea40c06SBrendan Gregg - Sun Microsystems * 3. has an I/O in progress (it may be an incomplete read). 64725ea40c06SBrendan Gregg - Sun Microsystems * 4. is flagged not eligible (zfs property). 64735a98e54bSBrendan Gregg - Sun Microsystems */ 647489c86e32SChris Williamson if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) || 64757adb730bSGeorge Wilson HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr)) 64765a98e54bSBrendan Gregg - Sun Microsystems return (B_FALSE); 64775a98e54bSBrendan Gregg - Sun Microsystems 64785a98e54bSBrendan Gregg - Sun Microsystems return (B_TRUE); 64795a98e54bSBrendan Gregg - Sun Microsystems } 64805a98e54bSBrendan Gregg - Sun Microsystems 64815a98e54bSBrendan Gregg - Sun Microsystems static uint64_t 6482aad02571SSaso Kiselkov l2arc_write_size(void) 64835a98e54bSBrendan Gregg - Sun Microsystems { 64845a98e54bSBrendan Gregg - Sun Microsystems uint64_t size; 64855a98e54bSBrendan Gregg - Sun Microsystems 6486aad02571SSaso Kiselkov /* 6487aad02571SSaso Kiselkov * Make sure our globals have meaningful values in case the user 6488aad02571SSaso Kiselkov * altered them. 6489aad02571SSaso Kiselkov */ 6490aad02571SSaso Kiselkov size = l2arc_write_max; 6491aad02571SSaso Kiselkov if (size == 0) { 6492aad02571SSaso Kiselkov cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 6493aad02571SSaso Kiselkov "be greater than zero, resetting it to the default (%d)", 6494aad02571SSaso Kiselkov L2ARC_WRITE_SIZE); 6495aad02571SSaso Kiselkov size = l2arc_write_max = L2ARC_WRITE_SIZE; 6496aad02571SSaso Kiselkov } 64975a98e54bSBrendan Gregg - Sun Microsystems 64985a98e54bSBrendan Gregg - Sun Microsystems if (arc_warm == B_FALSE) 6499aad02571SSaso Kiselkov size += l2arc_write_boost; 65005a98e54bSBrendan Gregg - Sun Microsystems 65015a98e54bSBrendan Gregg - Sun Microsystems return (size); 65025a98e54bSBrendan Gregg - Sun Microsystems 65035a98e54bSBrendan Gregg - Sun Microsystems } 65045a98e54bSBrendan Gregg - Sun Microsystems 65055a98e54bSBrendan Gregg - Sun Microsystems static clock_t 65065a98e54bSBrendan Gregg - Sun Microsystems l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 65075a98e54bSBrendan Gregg - Sun Microsystems { 6508d3d50737SRafael Vanoni clock_t interval, next, now; 65095a98e54bSBrendan Gregg - Sun Microsystems 65105a98e54bSBrendan Gregg - Sun Microsystems /* 65115a98e54bSBrendan Gregg - Sun Microsystems * If the ARC lists are busy, increase our write rate; if the 65125a98e54bSBrendan Gregg - Sun Microsystems * lists are stale, idle back. This is achieved by checking 65135a98e54bSBrendan Gregg - Sun Microsystems * how much we previously wrote - if it was more than half of 65145a98e54bSBrendan Gregg - Sun Microsystems * what we wanted, schedule the next write much sooner. 65155a98e54bSBrendan Gregg - Sun Microsystems */ 65165a98e54bSBrendan Gregg - Sun Microsystems if (l2arc_feed_again && wrote > (wanted / 2)) 65175a98e54bSBrendan Gregg - Sun Microsystems interval = (hz * l2arc_feed_min_ms) / 1000; 65185a98e54bSBrendan Gregg - Sun Microsystems else 65195a98e54bSBrendan Gregg - Sun Microsystems interval = hz * l2arc_feed_secs; 65205a98e54bSBrendan Gregg - Sun Microsystems 6521d3d50737SRafael Vanoni now = ddi_get_lbolt(); 6522d3d50737SRafael Vanoni next = MAX(now, MIN(now + interval, began + interval)); 65235a98e54bSBrendan Gregg - Sun Microsystems 65245a98e54bSBrendan Gregg - Sun Microsystems return (next); 65255a98e54bSBrendan Gregg - Sun Microsystems } 65265a98e54bSBrendan Gregg - Sun Microsystems 6527fa94a07fSbrendan /* 6528fa94a07fSbrendan * Cycle through L2ARC devices. This is how L2ARC load balances. 65293a737e0dSbrendan * If a device is returned, this also returns holding the spa config lock. 6530fa94a07fSbrendan */ 6531fa94a07fSbrendan static l2arc_dev_t * 6532fa94a07fSbrendan l2arc_dev_get_next(void) 6533fa94a07fSbrendan { 65343a737e0dSbrendan l2arc_dev_t *first, *next = NULL; 65353a737e0dSbrendan 65363a737e0dSbrendan /* 65373a737e0dSbrendan * Lock out the removal of spas (spa_namespace_lock), then removal 65383a737e0dSbrendan * of cache devices (l2arc_dev_mtx). Once a device has been selected, 65393a737e0dSbrendan * both locks will be dropped and a spa config lock held instead. 65403a737e0dSbrendan */ 65413a737e0dSbrendan mutex_enter(&spa_namespace_lock); 65423a737e0dSbrendan mutex_enter(&l2arc_dev_mtx); 6543fa94a07fSbrendan 6544c5904d13Seschrock /* if there are no vdevs, there is nothing to do */ 6545c5904d13Seschrock if (l2arc_ndev == 0) 65463a737e0dSbrendan goto out; 6547c5904d13Seschrock 6548c5904d13Seschrock first = NULL; 6549c5904d13Seschrock next = l2arc_dev_last; 6550c5904d13Seschrock do { 6551c5904d13Seschrock /* loop around the list looking for a non-faulted vdev */ 6552c5904d13Seschrock if (next == NULL) { 6553fa94a07fSbrendan next = list_head(l2arc_dev_list); 6554c5904d13Seschrock } else { 6555c5904d13Seschrock next = list_next(l2arc_dev_list, next); 6556c5904d13Seschrock if (next == NULL) 6557c5904d13Seschrock next = list_head(l2arc_dev_list); 6558c5904d13Seschrock } 6559c5904d13Seschrock 6560c5904d13Seschrock /* if we have come back to the start, bail out */ 6561c5904d13Seschrock if (first == NULL) 6562c5904d13Seschrock first = next; 6563c5904d13Seschrock else if (next == first) 6564c5904d13Seschrock break; 6565c5904d13Seschrock 6566c5904d13Seschrock } while (vdev_is_dead(next->l2ad_vdev)); 6567c5904d13Seschrock 6568c5904d13Seschrock /* if we were unable to find any usable vdevs, return NULL */ 6569c5904d13Seschrock if (vdev_is_dead(next->l2ad_vdev)) 65703a737e0dSbrendan next = NULL; 6571fa94a07fSbrendan 6572fa94a07fSbrendan l2arc_dev_last = next; 6573fa94a07fSbrendan 65743a737e0dSbrendan out: 65753a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 65763a737e0dSbrendan 65773a737e0dSbrendan /* 65783a737e0dSbrendan * Grab the config lock to prevent the 'next' device from being 65793a737e0dSbrendan * removed while we are writing to it. 65803a737e0dSbrendan */ 65813a737e0dSbrendan if (next != NULL) 6582e14bb325SJeff Bonwick spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 65833a737e0dSbrendan mutex_exit(&spa_namespace_lock); 65843a737e0dSbrendan 6585fa94a07fSbrendan return (next); 6586fa94a07fSbrendan } 6587fa94a07fSbrendan 65883a737e0dSbrendan /* 65893a737e0dSbrendan * Free buffers that were tagged for destruction. 65903a737e0dSbrendan */ 65913a737e0dSbrendan static void 65923a737e0dSbrendan l2arc_do_free_on_write() 65933a737e0dSbrendan { 65943a737e0dSbrendan list_t *buflist; 65953a737e0dSbrendan l2arc_data_free_t *df, *df_prev; 65963a737e0dSbrendan 65973a737e0dSbrendan mutex_enter(&l2arc_free_on_write_mtx); 65983a737e0dSbrendan buflist = l2arc_free_on_write; 65993a737e0dSbrendan 66003a737e0dSbrendan for (df = list_tail(buflist); df; df = df_prev) { 66013a737e0dSbrendan df_prev = list_prev(buflist, df); 6602770499e1SDan Kimmel ASSERT3P(df->l2df_abd, !=, NULL); 6603770499e1SDan Kimmel abd_free(df->l2df_abd); 66043a737e0dSbrendan list_remove(buflist, df); 66053a737e0dSbrendan kmem_free(df, sizeof (l2arc_data_free_t)); 66063a737e0dSbrendan } 66073a737e0dSbrendan 66083a737e0dSbrendan mutex_exit(&l2arc_free_on_write_mtx); 66093a737e0dSbrendan } 66103a737e0dSbrendan 6611fa94a07fSbrendan /* 6612fa94a07fSbrendan * A write to a cache device has completed. Update all headers to allow 6613fa94a07fSbrendan * reads from these buffers to begin. 6614fa94a07fSbrendan */ 6615fa94a07fSbrendan static void 6616fa94a07fSbrendan l2arc_write_done(zio_t *zio) 6617fa94a07fSbrendan { 6618fa94a07fSbrendan l2arc_write_callback_t *cb; 6619fa94a07fSbrendan l2arc_dev_t *dev; 6620fa94a07fSbrendan list_t *buflist; 66217adb730bSGeorge Wilson arc_buf_hdr_t *head, *hdr, *hdr_prev; 6622fa94a07fSbrendan kmutex_t *hash_lock; 66233038a2b4SSaso Kiselkov int64_t bytes_dropped = 0; 6624fa94a07fSbrendan 6625fa94a07fSbrendan cb = zio->io_private; 6626dcbf3bd6SGeorge Wilson ASSERT3P(cb, !=, NULL); 6627fa94a07fSbrendan dev = cb->l2wcb_dev; 6628dcbf3bd6SGeorge Wilson ASSERT3P(dev, !=, NULL); 6629fa94a07fSbrendan head = cb->l2wcb_head; 6630dcbf3bd6SGeorge Wilson ASSERT3P(head, !=, NULL); 663189c86e32SChris Williamson buflist = &dev->l2ad_buflist; 6632dcbf3bd6SGeorge Wilson ASSERT3P(buflist, !=, NULL); 6633fa94a07fSbrendan DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 6634fa94a07fSbrendan l2arc_write_callback_t *, cb); 6635fa94a07fSbrendan 6636fa94a07fSbrendan if (zio->io_error != 0) 6637fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_writes_error); 6638fa94a07fSbrendan 6639fa94a07fSbrendan /* 6640fa94a07fSbrendan * All writes completed, or an error was hit. 6641fa94a07fSbrendan */ 6642244781f1SPrakash Surya top: 6643244781f1SPrakash Surya mutex_enter(&dev->l2ad_mtx); 66447adb730bSGeorge Wilson for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 66457adb730bSGeorge Wilson hdr_prev = list_prev(buflist, hdr); 6646fa94a07fSbrendan 66477adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 6648244781f1SPrakash Surya 6649244781f1SPrakash Surya /* 6650244781f1SPrakash Surya * We cannot use mutex_enter or else we can deadlock 6651244781f1SPrakash Surya * with l2arc_write_buffers (due to swapping the order 6652244781f1SPrakash Surya * the hash lock and l2ad_mtx are taken). 6653244781f1SPrakash Surya */ 6654fa94a07fSbrendan if (!mutex_tryenter(hash_lock)) { 6655fa94a07fSbrendan /* 6656244781f1SPrakash Surya * Missed the hash lock. We must retry so we 6657244781f1SPrakash Surya * don't leave the ARC_FLAG_L2_WRITING bit set. 6658fa94a07fSbrendan */ 6659244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 6660244781f1SPrakash Surya 6661244781f1SPrakash Surya /* 6662244781f1SPrakash Surya * We don't want to rescan the headers we've 6663244781f1SPrakash Surya * already marked as having been written out, so 6664244781f1SPrakash Surya * we reinsert the head node so we can pick up 6665244781f1SPrakash Surya * where we left off. 6666244781f1SPrakash Surya */ 6667244781f1SPrakash Surya list_remove(buflist, head); 6668244781f1SPrakash Surya list_insert_after(buflist, hdr, head); 6669244781f1SPrakash Surya 6670244781f1SPrakash Surya mutex_exit(&dev->l2ad_mtx); 6671244781f1SPrakash Surya 6672244781f1SPrakash Surya /* 6673244781f1SPrakash Surya * We wait for the hash lock to become available 6674244781f1SPrakash Surya * to try and prevent busy waiting, and increase 6675244781f1SPrakash Surya * the chance we'll be able to acquire the lock 6676244781f1SPrakash Surya * the next time around. 6677244781f1SPrakash Surya */ 6678244781f1SPrakash Surya mutex_enter(hash_lock); 6679244781f1SPrakash Surya mutex_exit(hash_lock); 6680244781f1SPrakash Surya goto top; 6681fa94a07fSbrendan } 6682fa94a07fSbrendan 668389c86e32SChris Williamson /* 6684244781f1SPrakash Surya * We could not have been moved into the arc_l2c_only 6685244781f1SPrakash Surya * state while in-flight due to our ARC_FLAG_L2_WRITING 6686244781f1SPrakash Surya * bit being set. Let's just ensure that's being enforced. 6687244781f1SPrakash Surya */ 6688244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 6689244781f1SPrakash Surya 6690fa94a07fSbrendan if (zio->io_error != 0) { 6691fa94a07fSbrendan /* 66923a737e0dSbrendan * Error - drop L2ARC entry. 6693fa94a07fSbrendan */ 66947adb730bSGeorge Wilson list_remove(buflist, hdr); 6695dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 669689c86e32SChris Williamson 669716a7e5acSAndriy Gapon ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr)); 669816a7e5acSAndriy Gapon ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); 6699a52fc310SPrakash Surya 6700dcbf3bd6SGeorge Wilson bytes_dropped += arc_hdr_size(hdr); 6701a52fc310SPrakash Surya (void) refcount_remove_many(&dev->l2ad_alloc, 6702dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 6703fa94a07fSbrendan } 6704fa94a07fSbrendan 6705fa94a07fSbrendan /* 6706244781f1SPrakash Surya * Allow ARC to begin reads and ghost list evictions to 6707244781f1SPrakash Surya * this L2ARC entry. 6708fa94a07fSbrendan */ 6709dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); 6710fa94a07fSbrendan 6711fa94a07fSbrendan mutex_exit(hash_lock); 6712fa94a07fSbrendan } 6713fa94a07fSbrendan 6714fa94a07fSbrendan atomic_inc_64(&l2arc_writes_done); 6715fa94a07fSbrendan list_remove(buflist, head); 671689c86e32SChris Williamson ASSERT(!HDR_HAS_L1HDR(head)); 671789c86e32SChris Williamson kmem_cache_free(hdr_l2only_cache, head); 671889c86e32SChris Williamson mutex_exit(&dev->l2ad_mtx); 6719fa94a07fSbrendan 67203038a2b4SSaso Kiselkov vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 67213038a2b4SSaso Kiselkov 67223a737e0dSbrendan l2arc_do_free_on_write(); 6723fa94a07fSbrendan 6724fa94a07fSbrendan kmem_free(cb, sizeof (l2arc_write_callback_t)); 6725fa94a07fSbrendan } 6726fa94a07fSbrendan 6727fa94a07fSbrendan /* 6728fa94a07fSbrendan * A read to a cache device completed. Validate buffer contents before 6729fa94a07fSbrendan * handing over to the regular ARC routines. 6730fa94a07fSbrendan */ 6731fa94a07fSbrendan static void 6732fa94a07fSbrendan l2arc_read_done(zio_t *zio) 6733fa94a07fSbrendan { 6734fa94a07fSbrendan l2arc_read_callback_t *cb; 6735fa94a07fSbrendan arc_buf_hdr_t *hdr; 6736fa94a07fSbrendan kmutex_t *hash_lock; 6737dcbf3bd6SGeorge Wilson boolean_t valid_cksum; 6738fa94a07fSbrendan 6739dcbf3bd6SGeorge Wilson ASSERT3P(zio->io_vd, !=, NULL); 6740e14bb325SJeff Bonwick ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 6741e14bb325SJeff Bonwick 6742e14bb325SJeff Bonwick spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 6743e14bb325SJeff Bonwick 6744fa94a07fSbrendan cb = zio->io_private; 6745dcbf3bd6SGeorge Wilson ASSERT3P(cb, !=, NULL); 6746dcbf3bd6SGeorge Wilson hdr = cb->l2rcb_hdr; 6747dcbf3bd6SGeorge Wilson ASSERT3P(hdr, !=, NULL); 6748fa94a07fSbrendan 6749dcbf3bd6SGeorge Wilson hash_lock = HDR_LOCK(hdr); 6750fa94a07fSbrendan mutex_enter(hash_lock); 67513f9d6ad7SLin Ling ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 6752fa94a07fSbrendan 6753403a8da7SAndriy Gapon /* 6754403a8da7SAndriy Gapon * If the data was read into a temporary buffer, 6755403a8da7SAndriy Gapon * move it and free the buffer. 6756403a8da7SAndriy Gapon */ 6757403a8da7SAndriy Gapon if (cb->l2rcb_abd != NULL) { 6758403a8da7SAndriy Gapon ASSERT3U(arc_hdr_size(hdr), <, zio->io_size); 6759403a8da7SAndriy Gapon if (zio->io_error == 0) { 6760403a8da7SAndriy Gapon abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd, 6761403a8da7SAndriy Gapon arc_hdr_size(hdr)); 6762403a8da7SAndriy Gapon } 6763403a8da7SAndriy Gapon 6764403a8da7SAndriy Gapon /* 6765403a8da7SAndriy Gapon * The following must be done regardless of whether 6766403a8da7SAndriy Gapon * there was an error: 6767403a8da7SAndriy Gapon * - free the temporary buffer 6768403a8da7SAndriy Gapon * - point zio to the real ARC buffer 6769403a8da7SAndriy Gapon * - set zio size accordingly 6770403a8da7SAndriy Gapon * These are required because zio is either re-used for 6771403a8da7SAndriy Gapon * an I/O of the block in the case of the error 6772403a8da7SAndriy Gapon * or the zio is passed to arc_read_done() and it 6773403a8da7SAndriy Gapon * needs real data. 6774403a8da7SAndriy Gapon */ 6775403a8da7SAndriy Gapon abd_free(cb->l2rcb_abd); 6776403a8da7SAndriy Gapon zio->io_size = zio->io_orig_size = arc_hdr_size(hdr); 6777403a8da7SAndriy Gapon zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd; 6778403a8da7SAndriy Gapon } 6779403a8da7SAndriy Gapon 6780770499e1SDan Kimmel ASSERT3P(zio->io_abd, !=, NULL); 6781aad02571SSaso Kiselkov 6782fa94a07fSbrendan /* 6783fa94a07fSbrendan * Check this survived the L2ARC journey. 6784fa94a07fSbrendan */ 6785770499e1SDan Kimmel ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd); 6786dcbf3bd6SGeorge Wilson zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 6787dcbf3bd6SGeorge Wilson zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 6788dcbf3bd6SGeorge Wilson 6789dcbf3bd6SGeorge Wilson valid_cksum = arc_cksum_is_equal(hdr, zio); 6790dcbf3bd6SGeorge Wilson if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 6791fa94a07fSbrendan mutex_exit(hash_lock); 6792dcbf3bd6SGeorge Wilson zio->io_private = hdr; 6793fa94a07fSbrendan arc_read_done(zio); 6794fa94a07fSbrendan } else { 6795fa94a07fSbrendan mutex_exit(hash_lock); 6796fa94a07fSbrendan /* 6797fa94a07fSbrendan * Buffer didn't survive caching. Increment stats and 6798fa94a07fSbrendan * reissue to the original storage device. 6799fa94a07fSbrendan */ 68003a737e0dSbrendan if (zio->io_error != 0) { 6801fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_io_error); 68023a737e0dSbrendan } else { 6803be6fd75aSMatthew Ahrens zio->io_error = SET_ERROR(EIO); 68043a737e0dSbrendan } 6805dcbf3bd6SGeorge Wilson if (!valid_cksum) 6806fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_cksum_bad); 6807fa94a07fSbrendan 6808e14bb325SJeff Bonwick /* 6809e14bb325SJeff Bonwick * If there's no waiter, issue an async i/o to the primary 6810e14bb325SJeff Bonwick * storage now. If there *is* a waiter, the caller must 6811e14bb325SJeff Bonwick * issue the i/o in a context where it's OK to block. 6812e14bb325SJeff Bonwick */ 6813a3f829aeSBill Moore if (zio->io_waiter == NULL) { 6814a3f829aeSBill Moore zio_t *pio = zio_unique_parent(zio); 6815a3f829aeSBill Moore 6816a3f829aeSBill Moore ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6817a3f829aeSBill Moore 6818dcbf3bd6SGeorge Wilson zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, 6819770499e1SDan Kimmel hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done, 6820dcbf3bd6SGeorge Wilson hdr, zio->io_priority, cb->l2rcb_flags, 6821dcbf3bd6SGeorge Wilson &cb->l2rcb_zb)); 6822a3f829aeSBill Moore } 6823fa94a07fSbrendan } 6824fa94a07fSbrendan 6825fa94a07fSbrendan kmem_free(cb, sizeof (l2arc_read_callback_t)); 6826fa94a07fSbrendan } 6827fa94a07fSbrendan 6828fa94a07fSbrendan /* 6829fa94a07fSbrendan * This is the list priority from which the L2ARC will search for pages to 6830fa94a07fSbrendan * cache. This is used within loops (0..3) to cycle through lists in the 6831fa94a07fSbrendan * desired order. This order can have a significant effect on cache 6832fa94a07fSbrendan * performance. 6833fa94a07fSbrendan * 6834fa94a07fSbrendan * Currently the metadata lists are hit first, MFU then MRU, followed by 6835fa94a07fSbrendan * the data lists. This function returns a locked list, and also returns 6836fa94a07fSbrendan * the lock pointer. 6837fa94a07fSbrendan */ 6838244781f1SPrakash Surya static multilist_sublist_t * 6839244781f1SPrakash Surya l2arc_sublist_lock(int list_num) 6840fa94a07fSbrendan { 6841244781f1SPrakash Surya multilist_t *ml = NULL; 6842244781f1SPrakash Surya unsigned int idx; 6843fa94a07fSbrendan 6844fa94a07fSbrendan ASSERT(list_num >= 0 && list_num <= 3); 6845fa94a07fSbrendan 6846fa94a07fSbrendan switch (list_num) { 6847fa94a07fSbrendan case 0: 684894c2d0ebSMatthew Ahrens ml = arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6849fa94a07fSbrendan break; 6850fa94a07fSbrendan case 1: 685194c2d0ebSMatthew Ahrens ml = arc_mru->arcs_list[ARC_BUFC_METADATA]; 6852fa94a07fSbrendan break; 6853fa94a07fSbrendan case 2: 685494c2d0ebSMatthew Ahrens ml = arc_mfu->arcs_list[ARC_BUFC_DATA]; 6855fa94a07fSbrendan break; 6856fa94a07fSbrendan case 3: 685794c2d0ebSMatthew Ahrens ml = arc_mru->arcs_list[ARC_BUFC_DATA]; 6858fa94a07fSbrendan break; 6859fa94a07fSbrendan } 6860fa94a07fSbrendan 6861244781f1SPrakash Surya /* 6862244781f1SPrakash Surya * Return a randomly-selected sublist. This is acceptable 6863244781f1SPrakash Surya * because the caller feeds only a little bit of data for each 6864244781f1SPrakash Surya * call (8MB). Subsequent calls will result in different 6865244781f1SPrakash Surya * sublists being selected. 6866244781f1SPrakash Surya */ 6867244781f1SPrakash Surya idx = multilist_get_random_index(ml); 6868244781f1SPrakash Surya return (multilist_sublist_lock(ml, idx)); 6869fa94a07fSbrendan } 6870fa94a07fSbrendan 6871fa94a07fSbrendan /* 6872fa94a07fSbrendan * Evict buffers from the device write hand to the distance specified in 6873fa94a07fSbrendan * bytes. This distance may span populated buffers, it may span nothing. 6874fa94a07fSbrendan * This is clearing a region on the L2ARC device ready for writing. 6875fa94a07fSbrendan * If the 'all' boolean is set, every buffer is evicted. 6876fa94a07fSbrendan */ 6877fa94a07fSbrendan static void 6878fa94a07fSbrendan l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6879fa94a07fSbrendan { 6880fa94a07fSbrendan list_t *buflist; 68817adb730bSGeorge Wilson arc_buf_hdr_t *hdr, *hdr_prev; 6882fa94a07fSbrendan kmutex_t *hash_lock; 6883fa94a07fSbrendan uint64_t taddr; 6884fa94a07fSbrendan 688589c86e32SChris Williamson buflist = &dev->l2ad_buflist; 6886fa94a07fSbrendan 6887fa94a07fSbrendan if (!all && dev->l2ad_first) { 6888fa94a07fSbrendan /* 6889fa94a07fSbrendan * This is the first sweep through the device. There is 6890fa94a07fSbrendan * nothing to evict. 6891fa94a07fSbrendan */ 6892fa94a07fSbrendan return; 6893fa94a07fSbrendan } 6894fa94a07fSbrendan 68953a737e0dSbrendan if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6896fa94a07fSbrendan /* 6897fa94a07fSbrendan * When nearing the end of the device, evict to the end 6898fa94a07fSbrendan * before the device write hand jumps to the start. 6899fa94a07fSbrendan */ 6900fa94a07fSbrendan taddr = dev->l2ad_end; 6901fa94a07fSbrendan } else { 6902fa94a07fSbrendan taddr = dev->l2ad_hand + distance; 6903fa94a07fSbrendan } 6904fa94a07fSbrendan DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6905fa94a07fSbrendan uint64_t, taddr, boolean_t, all); 6906fa94a07fSbrendan 6907fa94a07fSbrendan top: 690889c86e32SChris Williamson mutex_enter(&dev->l2ad_mtx); 69097adb730bSGeorge Wilson for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 69107adb730bSGeorge Wilson hdr_prev = list_prev(buflist, hdr); 6911fa94a07fSbrendan 69127adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 6913244781f1SPrakash Surya 6914244781f1SPrakash Surya /* 6915244781f1SPrakash Surya * We cannot use mutex_enter or else we can deadlock 6916244781f1SPrakash Surya * with l2arc_write_buffers (due to swapping the order 6917244781f1SPrakash Surya * the hash lock and l2ad_mtx are taken). 6918244781f1SPrakash Surya */ 6919fa94a07fSbrendan if (!mutex_tryenter(hash_lock)) { 6920fa94a07fSbrendan /* 6921fa94a07fSbrendan * Missed the hash lock. Retry. 6922fa94a07fSbrendan */ 6923fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 692489c86e32SChris Williamson mutex_exit(&dev->l2ad_mtx); 6925fa94a07fSbrendan mutex_enter(hash_lock); 6926fa94a07fSbrendan mutex_exit(hash_lock); 6927fa94a07fSbrendan goto top; 6928fa94a07fSbrendan } 6929fa94a07fSbrendan 6930267ae6c3SAndriy Gapon /* 6931267ae6c3SAndriy Gapon * A header can't be on this list if it doesn't have L2 header. 6932267ae6c3SAndriy Gapon */ 6933267ae6c3SAndriy Gapon ASSERT(HDR_HAS_L2HDR(hdr)); 6934fa94a07fSbrendan 6935267ae6c3SAndriy Gapon /* Ensure this header has finished being written. */ 6936267ae6c3SAndriy Gapon ASSERT(!HDR_L2_WRITING(hdr)); 6937267ae6c3SAndriy Gapon ASSERT(!HDR_L2_WRITE_HEAD(hdr)); 6938267ae6c3SAndriy Gapon 6939267ae6c3SAndriy Gapon if (!all && (hdr->b_l2hdr.b_daddr >= taddr || 694089c86e32SChris Williamson hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6941fa94a07fSbrendan /* 6942fa94a07fSbrendan * We've evicted to the target address, 6943fa94a07fSbrendan * or the end of the device. 6944fa94a07fSbrendan */ 6945fa94a07fSbrendan mutex_exit(hash_lock); 6946fa94a07fSbrendan break; 6947fa94a07fSbrendan } 6948fa94a07fSbrendan 694989c86e32SChris Williamson if (!HDR_HAS_L1HDR(hdr)) { 69507adb730bSGeorge Wilson ASSERT(!HDR_L2_READING(hdr)); 6951fa94a07fSbrendan /* 6952fa94a07fSbrendan * This doesn't exist in the ARC. Destroy. 6953fa94a07fSbrendan * arc_hdr_destroy() will call list_remove() 695416a7e5acSAndriy Gapon * and decrement arcstat_l2_lsize. 6955fa94a07fSbrendan */ 69567adb730bSGeorge Wilson arc_change_state(arc_anon, hdr, hash_lock); 69577adb730bSGeorge Wilson arc_hdr_destroy(hdr); 6958fa94a07fSbrendan } else { 695989c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 696089c86e32SChris Williamson ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 69613a737e0dSbrendan /* 69623a737e0dSbrendan * Invalidate issued or about to be issued 69633a737e0dSbrendan * reads, since we may be about to write 69643a737e0dSbrendan * over this location. 69653a737e0dSbrendan */ 69667adb730bSGeorge Wilson if (HDR_L2_READING(hdr)) { 69673a737e0dSbrendan ARCSTAT_BUMP(arcstat_l2_evict_reading); 6968dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); 69693a737e0dSbrendan } 69703a737e0dSbrendan 6971a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(hdr); 6972fa94a07fSbrendan } 6973fa94a07fSbrendan mutex_exit(hash_lock); 6974fa94a07fSbrendan } 697589c86e32SChris Williamson mutex_exit(&dev->l2ad_mtx); 6976fa94a07fSbrendan } 6977fa94a07fSbrendan 6978fa94a07fSbrendan /* 6979fa94a07fSbrendan * Find and write ARC buffers to the L2ARC device. 6980fa94a07fSbrendan * 69817adb730bSGeorge Wilson * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6982fa94a07fSbrendan * for reading until they have completed writing. 6983aad02571SSaso Kiselkov * The headroom_boost is an in-out parameter used to maintain headroom boost 6984aad02571SSaso Kiselkov * state between calls to this function. 6985aad02571SSaso Kiselkov * 6986aad02571SSaso Kiselkov * Returns the number of bytes actually written (which may be smaller than 6987aad02571SSaso Kiselkov * the delta by which the device hand has changed due to alignment). 6988fa94a07fSbrendan */ 69895a98e54bSBrendan Gregg - Sun Microsystems static uint64_t 6990dcbf3bd6SGeorge Wilson l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 6991fa94a07fSbrendan { 69927adb730bSGeorge Wilson arc_buf_hdr_t *hdr, *hdr_prev, *head; 699316a7e5acSAndriy Gapon uint64_t write_asize, write_psize, write_lsize, headroom; 6994aad02571SSaso Kiselkov boolean_t full; 6995fa94a07fSbrendan l2arc_write_callback_t *cb; 6996fa94a07fSbrendan zio_t *pio, *wzio; 6997e9103aaeSGarrett D'Amore uint64_t guid = spa_load_guid(spa); 6998fa94a07fSbrendan 6999dcbf3bd6SGeorge Wilson ASSERT3P(dev->l2ad_vdev, !=, NULL); 7000aad02571SSaso Kiselkov 7001fa94a07fSbrendan pio = NULL; 700216a7e5acSAndriy Gapon write_lsize = write_asize = write_psize = 0; 7003fa94a07fSbrendan full = B_FALSE; 700489c86e32SChris Williamson head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 7005dcbf3bd6SGeorge Wilson arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); 7006aad02571SSaso Kiselkov 7007fa94a07fSbrendan /* 7008fa94a07fSbrendan * Copy buffers for L2ARC writing. 7009fa94a07fSbrendan */ 7010fa94a07fSbrendan for (int try = 0; try <= 3; try++) { 7011244781f1SPrakash Surya multilist_sublist_t *mls = l2arc_sublist_lock(try); 7012aad02571SSaso Kiselkov uint64_t passed_sz = 0; 7013aad02571SSaso Kiselkov 70143a737e0dSbrendan /* 70153a737e0dSbrendan * L2ARC fast warmup. 70163a737e0dSbrendan * 70173a737e0dSbrendan * Until the ARC is warm and starts to evict, read from the 70183a737e0dSbrendan * head of the ARC lists rather than the tail. 70193a737e0dSbrendan */ 70203a737e0dSbrendan if (arc_warm == B_FALSE) 7021244781f1SPrakash Surya hdr = multilist_sublist_head(mls); 70223a737e0dSbrendan else 7023244781f1SPrakash Surya hdr = multilist_sublist_tail(mls); 70243a737e0dSbrendan 7025aad02571SSaso Kiselkov headroom = target_sz * l2arc_headroom; 7026dcbf3bd6SGeorge Wilson if (zfs_compressed_arc_enabled) 7027aad02571SSaso Kiselkov headroom = (headroom * l2arc_headroom_boost) / 100; 7028aad02571SSaso Kiselkov 70297adb730bSGeorge Wilson for (; hdr; hdr = hdr_prev) { 7030aad02571SSaso Kiselkov kmutex_t *hash_lock; 7031aad02571SSaso Kiselkov 70323a737e0dSbrendan if (arc_warm == B_FALSE) 7033244781f1SPrakash Surya hdr_prev = multilist_sublist_next(mls, hdr); 70343a737e0dSbrendan else 7035244781f1SPrakash Surya hdr_prev = multilist_sublist_prev(mls, hdr); 7036fa94a07fSbrendan 70377adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 7038aad02571SSaso Kiselkov if (!mutex_tryenter(hash_lock)) { 7039fa94a07fSbrendan /* 7040fa94a07fSbrendan * Skip this buffer rather than waiting. 7041fa94a07fSbrendan */ 7042fa94a07fSbrendan continue; 7043fa94a07fSbrendan } 7044fa94a07fSbrendan 7045dcbf3bd6SGeorge Wilson passed_sz += HDR_GET_LSIZE(hdr); 7046fa94a07fSbrendan if (passed_sz > headroom) { 7047fa94a07fSbrendan /* 7048fa94a07fSbrendan * Searched too far. 7049fa94a07fSbrendan */ 7050fa94a07fSbrendan mutex_exit(hash_lock); 7051fa94a07fSbrendan break; 7052fa94a07fSbrendan } 7053fa94a07fSbrendan 70547adb730bSGeorge Wilson if (!l2arc_write_eligible(guid, hdr)) { 7055fa94a07fSbrendan mutex_exit(hash_lock); 7056fa94a07fSbrendan continue; 7057fa94a07fSbrendan } 7058fa94a07fSbrendan 705916a7e5acSAndriy Gapon /* 706016a7e5acSAndriy Gapon * We rely on the L1 portion of the header below, so 706116a7e5acSAndriy Gapon * it's invalid for this header to have been evicted out 706216a7e5acSAndriy Gapon * of the ghost cache, prior to being written out. The 706316a7e5acSAndriy Gapon * ARC_FLAG_L2_WRITING bit ensures this won't happen. 706416a7e5acSAndriy Gapon */ 706516a7e5acSAndriy Gapon ASSERT(HDR_HAS_L1HDR(hdr)); 706616a7e5acSAndriy Gapon 706716a7e5acSAndriy Gapon ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); 706816a7e5acSAndriy Gapon ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 706916a7e5acSAndriy Gapon ASSERT3U(arc_hdr_size(hdr), >, 0); 707016a7e5acSAndriy Gapon uint64_t psize = arc_hdr_size(hdr); 707116a7e5acSAndriy Gapon uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, 707216a7e5acSAndriy Gapon psize); 707316a7e5acSAndriy Gapon 707416a7e5acSAndriy Gapon if ((write_asize + asize) > target_sz) { 7075fa94a07fSbrendan full = B_TRUE; 7076fa94a07fSbrendan mutex_exit(hash_lock); 7077fa94a07fSbrendan break; 7078fa94a07fSbrendan } 7079fa94a07fSbrendan 7080fa94a07fSbrendan if (pio == NULL) { 7081fa94a07fSbrendan /* 7082fa94a07fSbrendan * Insert a dummy header on the buflist so 7083fa94a07fSbrendan * l2arc_write_done() can find where the 7084fa94a07fSbrendan * write buffers begin without searching. 7085fa94a07fSbrendan */ 7086244781f1SPrakash Surya mutex_enter(&dev->l2ad_mtx); 708789c86e32SChris Williamson list_insert_head(&dev->l2ad_buflist, head); 7088244781f1SPrakash Surya mutex_exit(&dev->l2ad_mtx); 7089fa94a07fSbrendan 7090fa94a07fSbrendan cb = kmem_alloc( 7091fa94a07fSbrendan sizeof (l2arc_write_callback_t), KM_SLEEP); 7092fa94a07fSbrendan cb->l2wcb_dev = dev; 7093fa94a07fSbrendan cb->l2wcb_head = head; 7094fa94a07fSbrendan pio = zio_root(spa, l2arc_write_done, cb, 7095fa94a07fSbrendan ZIO_FLAG_CANFAIL); 7096fa94a07fSbrendan } 7097fa94a07fSbrendan 709889c86e32SChris Williamson hdr->b_l2hdr.b_dev = dev; 7099dcbf3bd6SGeorge Wilson hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 7100dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, 7101dcbf3bd6SGeorge Wilson ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR); 7102dcbf3bd6SGeorge Wilson 7103dcbf3bd6SGeorge Wilson mutex_enter(&dev->l2ad_mtx); 7104dcbf3bd6SGeorge Wilson list_insert_head(&dev->l2ad_buflist, hdr); 7105dcbf3bd6SGeorge Wilson mutex_exit(&dev->l2ad_mtx); 7106dcbf3bd6SGeorge Wilson 710716a7e5acSAndriy Gapon (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr); 7108aad02571SSaso Kiselkov 7109a52fc310SPrakash Surya /* 7110dcbf3bd6SGeorge Wilson * Normally the L2ARC can use the hdr's data, but if 7111dcbf3bd6SGeorge Wilson * we're sharing data between the hdr and one of its 7112dcbf3bd6SGeorge Wilson * bufs, L2ARC needs its own copy of the data so that 7113403a8da7SAndriy Gapon * the ZIO below can't race with the buf consumer. 7114403a8da7SAndriy Gapon * Another case where we need to create a copy of the 7115403a8da7SAndriy Gapon * data is when the buffer size is not device-aligned 7116403a8da7SAndriy Gapon * and we need to pad the block to make it such. 7117403a8da7SAndriy Gapon * That also keeps the clock hand suitably aligned. 7118403a8da7SAndriy Gapon * 7119403a8da7SAndriy Gapon * To ensure that the copy will be available for the 7120dcbf3bd6SGeorge Wilson * lifetime of the ZIO and be cleaned up afterwards, we 7121dcbf3bd6SGeorge Wilson * add it to the l2arc_free_on_write queue. 7122a52fc310SPrakash Surya */ 7123770499e1SDan Kimmel abd_t *to_write; 712416a7e5acSAndriy Gapon if (!HDR_SHARED_DATA(hdr) && psize == asize) { 7125770499e1SDan Kimmel to_write = hdr->b_l1hdr.b_pabd; 7126dcbf3bd6SGeorge Wilson } else { 7127403a8da7SAndriy Gapon to_write = abd_alloc_for_io(asize, 7128770499e1SDan Kimmel HDR_ISTYPE_METADATA(hdr)); 712916a7e5acSAndriy Gapon abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize); 713016a7e5acSAndriy Gapon if (asize != psize) { 713116a7e5acSAndriy Gapon abd_zero_off(to_write, psize, 713216a7e5acSAndriy Gapon asize - psize); 7133403a8da7SAndriy Gapon } 713416a7e5acSAndriy Gapon l2arc_free_abd_on_write(to_write, asize, 7135770499e1SDan Kimmel arc_buf_type(hdr)); 7136dcbf3bd6SGeorge Wilson } 7137dcbf3bd6SGeorge Wilson wzio = zio_write_phys(pio, dev->l2ad_vdev, 7138403a8da7SAndriy Gapon hdr->b_l2hdr.b_daddr, asize, to_write, 7139dcbf3bd6SGeorge Wilson ZIO_CHECKSUM_OFF, NULL, hdr, 7140dcbf3bd6SGeorge Wilson ZIO_PRIORITY_ASYNC_WRITE, 7141dcbf3bd6SGeorge Wilson ZIO_FLAG_CANFAIL, B_FALSE); 7142aad02571SSaso Kiselkov 714316a7e5acSAndriy Gapon write_lsize += HDR_GET_LSIZE(hdr); 7144dcbf3bd6SGeorge Wilson DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 7145dcbf3bd6SGeorge Wilson zio_t *, wzio); 7146fa94a07fSbrendan 714716a7e5acSAndriy Gapon write_psize += psize; 714816a7e5acSAndriy Gapon write_asize += asize; 7149dcbf3bd6SGeorge Wilson dev->l2ad_hand += asize; 7150fa94a07fSbrendan 7151fa94a07fSbrendan mutex_exit(hash_lock); 7152fa94a07fSbrendan 7153dcbf3bd6SGeorge Wilson (void) zio_nowait(wzio); 7154aad02571SSaso Kiselkov } 7155aad02571SSaso Kiselkov 7156244781f1SPrakash Surya multilist_sublist_unlock(mls); 7157aad02571SSaso Kiselkov 7158aad02571SSaso Kiselkov if (full == B_TRUE) 7159aad02571SSaso Kiselkov break; 7160aad02571SSaso Kiselkov } 7161aad02571SSaso Kiselkov 7162aad02571SSaso Kiselkov /* No buffers selected for writing? */ 7163aad02571SSaso Kiselkov if (pio == NULL) { 716416a7e5acSAndriy Gapon ASSERT0(write_lsize); 716589c86e32SChris Williamson ASSERT(!HDR_HAS_L1HDR(head)); 716689c86e32SChris Williamson kmem_cache_free(hdr_l2only_cache, head); 7167aad02571SSaso Kiselkov return (0); 7168aad02571SSaso Kiselkov } 7169aad02571SSaso Kiselkov 7170aad02571SSaso Kiselkov ASSERT3U(write_asize, <=, target_sz); 7171fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_writes_sent); 717216a7e5acSAndriy Gapon ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize); 717316a7e5acSAndriy Gapon ARCSTAT_INCR(arcstat_l2_lsize, write_lsize); 717416a7e5acSAndriy Gapon ARCSTAT_INCR(arcstat_l2_psize, write_psize); 717516a7e5acSAndriy Gapon vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 7176fa94a07fSbrendan 7177fa94a07fSbrendan /* 7178fa94a07fSbrendan * Bump device hand to the device start if it is approaching the end. 7179fa94a07fSbrendan * l2arc_evict() will already have evicted ahead for this case. 7180fa94a07fSbrendan */ 71813a737e0dSbrendan if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 7182fa94a07fSbrendan dev->l2ad_hand = dev->l2ad_start; 7183fa94a07fSbrendan dev->l2ad_first = B_FALSE; 7184fa94a07fSbrendan } 7185fa94a07fSbrendan 71865a98e54bSBrendan Gregg - Sun Microsystems dev->l2ad_writing = B_TRUE; 7187fa94a07fSbrendan (void) zio_wait(pio); 71885a98e54bSBrendan Gregg - Sun Microsystems dev->l2ad_writing = B_FALSE; 71895a98e54bSBrendan Gregg - Sun Microsystems 7190aad02571SSaso Kiselkov return (write_asize); 7191aad02571SSaso Kiselkov } 7192aad02571SSaso Kiselkov 7193fa94a07fSbrendan /* 7194fa94a07fSbrendan * This thread feeds the L2ARC at regular intervals. This is the beating 7195fa94a07fSbrendan * heart of the L2ARC. 7196fa94a07fSbrendan */ 71973f7978d0SAlan Somers /* ARGSUSED */ 7198fa94a07fSbrendan static void 71993f7978d0SAlan Somers l2arc_feed_thread(void *unused) 7200fa94a07fSbrendan { 7201fa94a07fSbrendan callb_cpr_t cpr; 7202fa94a07fSbrendan l2arc_dev_t *dev; 7203fa94a07fSbrendan spa_t *spa; 72045a98e54bSBrendan Gregg - Sun Microsystems uint64_t size, wrote; 7205d3d50737SRafael Vanoni clock_t begin, next = ddi_get_lbolt(); 7206fa94a07fSbrendan 7207fa94a07fSbrendan CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 7208fa94a07fSbrendan 7209fa94a07fSbrendan mutex_enter(&l2arc_feed_thr_lock); 7210fa94a07fSbrendan 7211fa94a07fSbrendan while (l2arc_thread_exit == 0) { 7212fa94a07fSbrendan CALLB_CPR_SAFE_BEGIN(&cpr); 7213fa94a07fSbrendan (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 72145a98e54bSBrendan Gregg - Sun Microsystems next); 7215fa94a07fSbrendan CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 7216d3d50737SRafael Vanoni next = ddi_get_lbolt() + hz; 7217fa94a07fSbrendan 72183a737e0dSbrendan /* 72193a737e0dSbrendan * Quick check for L2ARC devices. 72203a737e0dSbrendan */ 7221c5904d13Seschrock mutex_enter(&l2arc_dev_mtx); 72223a737e0dSbrendan if (l2arc_ndev == 0) { 72233a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 72243a737e0dSbrendan continue; 72253a737e0dSbrendan } 72263a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 7227d3d50737SRafael Vanoni begin = ddi_get_lbolt(); 7228c5904d13Seschrock 7229fa94a07fSbrendan /* 7230c5904d13Seschrock * This selects the next l2arc device to write to, and in 7231c5904d13Seschrock * doing so the next spa to feed from: dev->l2ad_spa. This 72323a737e0dSbrendan * will return NULL if there are now no l2arc devices or if 72333a737e0dSbrendan * they are all faulted. 72343a737e0dSbrendan * 72353a737e0dSbrendan * If a device is returned, its spa's config lock is also 72363a737e0dSbrendan * held to prevent device removal. l2arc_dev_get_next() 72373a737e0dSbrendan * will grab and release l2arc_dev_mtx. 7238fa94a07fSbrendan */ 72393a737e0dSbrendan if ((dev = l2arc_dev_get_next()) == NULL) 7240fa94a07fSbrendan continue; 72413a737e0dSbrendan 72423a737e0dSbrendan spa = dev->l2ad_spa; 7243dcbf3bd6SGeorge Wilson ASSERT3P(spa, !=, NULL); 7244fa94a07fSbrendan 7245f9af39baSGeorge Wilson /* 7246f9af39baSGeorge Wilson * If the pool is read-only then force the feed thread to 7247f9af39baSGeorge Wilson * sleep a little longer. 7248f9af39baSGeorge Wilson */ 7249f9af39baSGeorge Wilson if (!spa_writeable(spa)) { 7250f9af39baSGeorge Wilson next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 7251f9af39baSGeorge Wilson spa_config_exit(spa, SCL_L2ARC, dev); 7252f9af39baSGeorge Wilson continue; 7253f9af39baSGeorge Wilson } 7254f9af39baSGeorge Wilson 7255fa94a07fSbrendan /* 7256fa94a07fSbrendan * Avoid contributing to memory pressure. 7257fa94a07fSbrendan */ 7258fa94a07fSbrendan if (arc_reclaim_needed()) { 7259fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 7260e14bb325SJeff Bonwick spa_config_exit(spa, SCL_L2ARC, dev); 7261fa94a07fSbrendan continue; 7262fa94a07fSbrendan } 7263fa94a07fSbrendan 7264fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_feeds); 7265fa94a07fSbrendan 7266aad02571SSaso Kiselkov size = l2arc_write_size(); 72673a737e0dSbrendan 7268fa94a07fSbrendan /* 7269fa94a07fSbrendan * Evict L2ARC buffers that will be overwritten. 7270fa94a07fSbrendan */ 72713a737e0dSbrendan l2arc_evict(dev, size, B_FALSE); 7272fa94a07fSbrendan 7273fa94a07fSbrendan /* 7274fa94a07fSbrendan * Write ARC buffers. 7275fa94a07fSbrendan */ 7276dcbf3bd6SGeorge Wilson wrote = l2arc_write_buffers(spa, dev, size); 72775a98e54bSBrendan Gregg - Sun Microsystems 72785a98e54bSBrendan Gregg - Sun Microsystems /* 72795a98e54bSBrendan Gregg - Sun Microsystems * Calculate interval between writes. 72805a98e54bSBrendan Gregg - Sun Microsystems */ 72815a98e54bSBrendan Gregg - Sun Microsystems next = l2arc_write_interval(begin, size, wrote); 7282e14bb325SJeff Bonwick spa_config_exit(spa, SCL_L2ARC, dev); 7283fa94a07fSbrendan } 7284fa94a07fSbrendan 7285fa94a07fSbrendan l2arc_thread_exit = 0; 7286fa94a07fSbrendan cv_broadcast(&l2arc_feed_thr_cv); 7287fa94a07fSbrendan CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 7288fa94a07fSbrendan thread_exit(); 7289fa94a07fSbrendan } 7290fa94a07fSbrendan 7291c5904d13Seschrock boolean_t 7292c5904d13Seschrock l2arc_vdev_present(vdev_t *vd) 7293c5904d13Seschrock { 7294c5904d13Seschrock l2arc_dev_t *dev; 7295c5904d13Seschrock 7296c5904d13Seschrock mutex_enter(&l2arc_dev_mtx); 7297c5904d13Seschrock for (dev = list_head(l2arc_dev_list); dev != NULL; 7298c5904d13Seschrock dev = list_next(l2arc_dev_list, dev)) { 7299c5904d13Seschrock if (dev->l2ad_vdev == vd) 7300c5904d13Seschrock break; 7301c5904d13Seschrock } 7302c5904d13Seschrock mutex_exit(&l2arc_dev_mtx); 7303c5904d13Seschrock 7304c5904d13Seschrock return (dev != NULL); 7305c5904d13Seschrock } 7306c5904d13Seschrock 7307fa94a07fSbrendan /* 7308fa94a07fSbrendan * Add a vdev for use by the L2ARC. By this point the spa has already 7309fa94a07fSbrendan * validated the vdev and opened it. 7310fa94a07fSbrendan */ 7311fa94a07fSbrendan void 7312573ca77eSGeorge Wilson l2arc_add_vdev(spa_t *spa, vdev_t *vd) 7313fa94a07fSbrendan { 7314fa94a07fSbrendan l2arc_dev_t *adddev; 7315fa94a07fSbrendan 7316c5904d13Seschrock ASSERT(!l2arc_vdev_present(vd)); 7317c5904d13Seschrock 7318fa94a07fSbrendan /* 7319fa94a07fSbrendan * Create a new l2arc device entry. 7320fa94a07fSbrendan */ 7321fa94a07fSbrendan adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 7322fa94a07fSbrendan adddev->l2ad_spa = spa; 7323fa94a07fSbrendan adddev->l2ad_vdev = vd; 7324573ca77eSGeorge Wilson adddev->l2ad_start = VDEV_LABEL_START_SIZE; 7325573ca77eSGeorge Wilson adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 7326fa94a07fSbrendan adddev->l2ad_hand = adddev->l2ad_start; 7327fa94a07fSbrendan adddev->l2ad_first = B_TRUE; 73285a98e54bSBrendan Gregg - Sun Microsystems adddev->l2ad_writing = B_FALSE; 7329fa94a07fSbrendan 733089c86e32SChris Williamson mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 7331fa94a07fSbrendan /* 7332fa94a07fSbrendan * This is a list of all ARC buffers that are still valid on the 7333fa94a07fSbrendan * device. 7334fa94a07fSbrendan */ 733589c86e32SChris Williamson list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 733689c86e32SChris Williamson offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 7337fa94a07fSbrendan 7338b24ab676SJeff Bonwick vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 7339a52fc310SPrakash Surya refcount_create(&adddev->l2ad_alloc); 7340fa94a07fSbrendan 7341fa94a07fSbrendan /* 7342fa94a07fSbrendan * Add device to global list 7343fa94a07fSbrendan */ 7344fa94a07fSbrendan mutex_enter(&l2arc_dev_mtx); 7345fa94a07fSbrendan list_insert_head(l2arc_dev_list, adddev); 7346fa94a07fSbrendan atomic_inc_64(&l2arc_ndev); 7347fa94a07fSbrendan mutex_exit(&l2arc_dev_mtx); 7348fa94a07fSbrendan } 7349fa94a07fSbrendan 7350fa94a07fSbrendan /* 7351fa94a07fSbrendan * Remove a vdev from the L2ARC. 7352fa94a07fSbrendan */ 7353fa94a07fSbrendan void 7354fa94a07fSbrendan l2arc_remove_vdev(vdev_t *vd) 7355fa94a07fSbrendan { 7356fa94a07fSbrendan l2arc_dev_t *dev, *nextdev, *remdev = NULL; 7357fa94a07fSbrendan 7358fa94a07fSbrendan /* 7359fa94a07fSbrendan * Find the device by vdev 7360fa94a07fSbrendan */ 7361fa94a07fSbrendan mutex_enter(&l2arc_dev_mtx); 7362fa94a07fSbrendan for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 7363fa94a07fSbrendan nextdev = list_next(l2arc_dev_list, dev); 7364fa94a07fSbrendan if (vd == dev->l2ad_vdev) { 7365fa94a07fSbrendan remdev = dev; 7366fa94a07fSbrendan break; 7367fa94a07fSbrendan } 7368fa94a07fSbrendan } 7369dcbf3bd6SGeorge Wilson ASSERT3P(remdev, !=, NULL); 7370fa94a07fSbrendan 7371fa94a07fSbrendan /* 7372fa94a07fSbrendan * Remove device from global list 7373fa94a07fSbrendan */ 7374fa94a07fSbrendan list_remove(l2arc_dev_list, remdev); 7375fa94a07fSbrendan l2arc_dev_last = NULL; /* may have been invalidated */ 73763a737e0dSbrendan atomic_dec_64(&l2arc_ndev); 73773a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 7378fa94a07fSbrendan 7379fa94a07fSbrendan /* 7380fa94a07fSbrendan * Clear all buflists and ARC references. L2ARC device flush. 7381fa94a07fSbrendan */ 7382fa94a07fSbrendan l2arc_evict(remdev, 0, B_TRUE); 738389c86e32SChris Williamson list_destroy(&remdev->l2ad_buflist); 738489c86e32SChris Williamson mutex_destroy(&remdev->l2ad_mtx); 7385a52fc310SPrakash Surya refcount_destroy(&remdev->l2ad_alloc); 7386fa94a07fSbrendan kmem_free(remdev, sizeof (l2arc_dev_t)); 7387fa94a07fSbrendan } 7388fa94a07fSbrendan 7389fa94a07fSbrendan void 7390e14bb325SJeff Bonwick l2arc_init(void) 7391fa94a07fSbrendan { 7392fa94a07fSbrendan l2arc_thread_exit = 0; 7393fa94a07fSbrendan l2arc_ndev = 0; 7394fa94a07fSbrendan l2arc_writes_sent = 0; 7395fa94a07fSbrendan l2arc_writes_done = 0; 7396fa94a07fSbrendan 7397fa94a07fSbrendan mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7398fa94a07fSbrendan cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7399fa94a07fSbrendan mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7400fa94a07fSbrendan mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7401fa94a07fSbrendan 7402fa94a07fSbrendan l2arc_dev_list = &L2ARC_dev_list; 7403fa94a07fSbrendan l2arc_free_on_write = &L2ARC_free_on_write; 7404fa94a07fSbrendan list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7405fa94a07fSbrendan offsetof(l2arc_dev_t, l2ad_node)); 7406fa94a07fSbrendan list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7407fa94a07fSbrendan offsetof(l2arc_data_free_t, l2df_list_node)); 7408fa94a07fSbrendan } 7409fa94a07fSbrendan 7410fa94a07fSbrendan void 7411e14bb325SJeff Bonwick l2arc_fini(void) 7412fa94a07fSbrendan { 74133a737e0dSbrendan /* 74143a737e0dSbrendan * This is called from dmu_fini(), which is called from spa_fini(); 74153a737e0dSbrendan * Because of this, we can assume that all l2arc devices have 74163a737e0dSbrendan * already been removed when the pools themselves were removed. 74173a737e0dSbrendan */ 74183a737e0dSbrendan 74193a737e0dSbrendan l2arc_do_free_on_write(); 74203a737e0dSbrendan 7421fa94a07fSbrendan mutex_destroy(&l2arc_feed_thr_lock); 7422fa94a07fSbrendan cv_destroy(&l2arc_feed_thr_cv); 7423fa94a07fSbrendan mutex_destroy(&l2arc_dev_mtx); 7424fa94a07fSbrendan mutex_destroy(&l2arc_free_on_write_mtx); 7425fa94a07fSbrendan 7426fa94a07fSbrendan list_destroy(l2arc_dev_list); 7427fa94a07fSbrendan list_destroy(l2arc_free_on_write); 7428fa94a07fSbrendan } 7429e14bb325SJeff Bonwick 7430e14bb325SJeff Bonwick void 7431e14bb325SJeff Bonwick l2arc_start(void) 7432e14bb325SJeff Bonwick { 74338ad4d6ddSJeff Bonwick if (!(spa_mode_global & FWRITE)) 7434e14bb325SJeff Bonwick return; 7435e14bb325SJeff Bonwick 7436e14bb325SJeff Bonwick (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7437e14bb325SJeff Bonwick TS_RUN, minclsyspri); 7438e14bb325SJeff Bonwick } 7439e14bb325SJeff Bonwick 7440e14bb325SJeff Bonwick void 7441e14bb325SJeff Bonwick l2arc_stop(void) 7442e14bb325SJeff Bonwick { 74438ad4d6ddSJeff Bonwick if (!(spa_mode_global & FWRITE)) 7444e14bb325SJeff Bonwick return; 7445e14bb325SJeff Bonwick 7446e14bb325SJeff Bonwick mutex_enter(&l2arc_feed_thr_lock); 7447e14bb325SJeff Bonwick cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7448e14bb325SJeff Bonwick l2arc_thread_exit = 1; 7449e14bb325SJeff Bonwick while (l2arc_thread_exit != 0) 7450e14bb325SJeff Bonwick cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7451e14bb325SJeff Bonwick mutex_exit(&l2arc_feed_thr_lock); 7452e14bb325SJeff Bonwick } 7453