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. 23cf746768SBryan Cantrill * Copyright (c) 2012, Joyent, Inc. All rights reserved. 2410fbdecbSMatthew Ahrens * Copyright (c) 2011, 2017 by Delphix. All rights reserved. 2571cb1b74SSaso Kiselkov * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 26*01a059eeSRoman 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> 277b24ab676SJeff Bonwick #include <zfs_fletcher.h> 278fa9e4066Sahrens 279cd1c8b85SMatthew Ahrens #ifndef _KERNEL 280cd1c8b85SMatthew Ahrens /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 281cd1c8b85SMatthew Ahrens boolean_t arc_watch = B_FALSE; 282cd1c8b85SMatthew Ahrens int arc_procfd; 283cd1c8b85SMatthew Ahrens #endif 284cd1c8b85SMatthew Ahrens 285244781f1SPrakash Surya static kmutex_t arc_reclaim_lock; 286244781f1SPrakash Surya static kcondvar_t arc_reclaim_thread_cv; 287244781f1SPrakash Surya static boolean_t arc_reclaim_thread_exit; 288244781f1SPrakash Surya static kcondvar_t arc_reclaim_waiters_cv; 289244781f1SPrakash Surya 2902ec99e3eSMatthew Ahrens uint_t arc_reduce_dnlc_percent = 3; 291fa9e4066Sahrens 29269962b56SMatthew Ahrens /* 293244781f1SPrakash Surya * The number of headers to evict in arc_evict_state_impl() before 294244781f1SPrakash Surya * dropping the sublist lock and evicting from another sublist. A lower 295244781f1SPrakash Surya * value means we're more likely to evict the "correct" header (i.e. the 296244781f1SPrakash Surya * oldest header in the arc state), but comes with higher overhead 297244781f1SPrakash Surya * (i.e. more invocations of arc_evict_state_impl()). 298244781f1SPrakash Surya */ 299244781f1SPrakash Surya int zfs_arc_evict_batch_limit = 10; 300244781f1SPrakash Surya 301fa9e4066Sahrens /* number of seconds before growing cache again */ 302fa9e4066Sahrens static int arc_grow_retry = 60; 303fa9e4066Sahrens 304770499e1SDan Kimmel /* shift of arc_c for calculating overflow limit in arc_get_data_impl */ 305244781f1SPrakash Surya int zfs_arc_overflow_shift = 8; 306244781f1SPrakash Surya 3075a98e54bSBrendan Gregg - Sun Microsystems /* shift of arc_c for calculating both min and max arc_p */ 3085a98e54bSBrendan Gregg - Sun Microsystems static int arc_p_min_shift = 4; 3095a98e54bSBrendan Gregg - Sun Microsystems 3105a98e54bSBrendan Gregg - Sun Microsystems /* log2(fraction of arc to reclaim) */ 3112ec99e3eSMatthew Ahrens static int arc_shrink_shift = 7; 3122ec99e3eSMatthew Ahrens 3132ec99e3eSMatthew Ahrens /* 3142ec99e3eSMatthew Ahrens * log2(fraction of ARC which must be free to allow growing). 3152ec99e3eSMatthew Ahrens * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 3162ec99e3eSMatthew Ahrens * when reading a new block into the ARC, we will evict an equal-sized block 3172ec99e3eSMatthew Ahrens * from the ARC. 3182ec99e3eSMatthew Ahrens * 3192ec99e3eSMatthew Ahrens * This must be less than arc_shrink_shift, so that when we shrink the ARC, 3202ec99e3eSMatthew Ahrens * we will still not allow it to grow. 3212ec99e3eSMatthew Ahrens */ 3222ec99e3eSMatthew Ahrens int arc_no_grow_shift = 5; 3232ec99e3eSMatthew Ahrens 3245a98e54bSBrendan Gregg - Sun Microsystems 32513506d1eSmaybee /* 326b19a79ecSperrin * minimum lifespan of a prefetch block in clock ticks 327b19a79ecSperrin * (initialized in arc_init()) 32813506d1eSmaybee */ 329b19a79ecSperrin static int arc_min_prefetch_lifespan; 33013506d1eSmaybee 33169962b56SMatthew Ahrens /* 33269962b56SMatthew Ahrens * If this percent of memory is free, don't throttle. 33369962b56SMatthew Ahrens */ 33469962b56SMatthew Ahrens int arc_lotsfree_percent = 10; 33569962b56SMatthew Ahrens 336fa9e4066Sahrens static int arc_dead; 337fa9e4066Sahrens 3383a737e0dSbrendan /* 3393a737e0dSbrendan * The arc has filled available memory and has now warmed up. 3403a737e0dSbrendan */ 3413a737e0dSbrendan static boolean_t arc_warm; 3423a737e0dSbrendan 3430dd053d7SPrakash Surya /* 3440dd053d7SPrakash Surya * log2 fraction of the zio arena to keep free. 3450dd053d7SPrakash Surya */ 3460dd053d7SPrakash Surya int arc_zio_arena_free_shift = 2; 3470dd053d7SPrakash Surya 348a2eea2e1Sahrens /* 349a2eea2e1Sahrens * These tunables are for performance analysis. 350a2eea2e1Sahrens */ 351a2eea2e1Sahrens uint64_t zfs_arc_max; 352a2eea2e1Sahrens uint64_t zfs_arc_min; 3531116048bSek uint64_t zfs_arc_meta_limit = 0; 3543a5286a1SMatthew Ahrens uint64_t zfs_arc_meta_min = 0; 3555a98e54bSBrendan Gregg - Sun Microsystems int zfs_arc_grow_retry = 0; 3565a98e54bSBrendan Gregg - Sun Microsystems int zfs_arc_shrink_shift = 0; 3575a98e54bSBrendan Gregg - Sun Microsystems int zfs_arc_p_min_shift = 0; 35863e911b6SMatthew Ahrens int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 359a2eea2e1Sahrens 360dcbf3bd6SGeorge Wilson boolean_t zfs_compressed_arc_enabled = B_TRUE; 361dcbf3bd6SGeorge Wilson 362fa9e4066Sahrens /* 363fa94a07fSbrendan * Note that buffers can be in one of 6 states: 364fa9e4066Sahrens * ARC_anon - anonymous (discussed below) 365ea8dc4b6Seschrock * ARC_mru - recently used, currently cached 366ea8dc4b6Seschrock * ARC_mru_ghost - recentely used, no longer in cache 367ea8dc4b6Seschrock * ARC_mfu - frequently used, currently cached 368ea8dc4b6Seschrock * ARC_mfu_ghost - frequently used, no longer in cache 369fa94a07fSbrendan * ARC_l2c_only - exists in L2ARC but not other states 3700e8c6158Smaybee * When there are no active references to the buffer, they are 3710e8c6158Smaybee * are linked onto a list in one of these arc states. These are 3720e8c6158Smaybee * the only buffers that can be evicted or deleted. Within each 3730e8c6158Smaybee * state there are multiple lists, one for meta-data and one for 3740e8c6158Smaybee * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 3750e8c6158Smaybee * etc.) is tracked separately so that it can be managed more 376fa94a07fSbrendan * explicitly: favored over data, limited explicitly. 377fa9e4066Sahrens * 378fa9e4066Sahrens * Anonymous buffers are buffers that are not associated with 379fa9e4066Sahrens * a DVA. These are buffers that hold dirty block copies 380fa9e4066Sahrens * before they are written to stable storage. By definition, 381ea8dc4b6Seschrock * they are "ref'd" and are considered part of arc_mru 382fa9e4066Sahrens * that cannot be freed. Generally, they will aquire a DVA 383ea8dc4b6Seschrock * as they are written and migrate onto the arc_mru list. 384fa94a07fSbrendan * 385fa94a07fSbrendan * The ARC_l2c_only state is for buffers that are in the second 386fa94a07fSbrendan * level ARC but no longer in any of the ARC_m* lists. The second 387fa94a07fSbrendan * level ARC itself may also contain buffers that are in any of 388fa94a07fSbrendan * the ARC_m* states - meaning that a buffer can exist in two 389fa94a07fSbrendan * places. The reason for the ARC_l2c_only state is to keep the 390fa94a07fSbrendan * buffer header in the hash table, so that reads that hit the 391fa94a07fSbrendan * second level ARC benefit from these fast lookups. 392fa9e4066Sahrens */ 393fa9e4066Sahrens 394fa9e4066Sahrens typedef struct arc_state { 395244781f1SPrakash Surya /* 396244781f1SPrakash Surya * list of evictable buffers 397244781f1SPrakash Surya */ 39894c2d0ebSMatthew Ahrens multilist_t *arcs_list[ARC_BUFC_NUMTYPES]; 399244781f1SPrakash Surya /* 400244781f1SPrakash Surya * total amount of evictable data in this state 401244781f1SPrakash Surya */ 402dcbf3bd6SGeorge Wilson refcount_t arcs_esize[ARC_BUFC_NUMTYPES]; 403244781f1SPrakash Surya /* 404244781f1SPrakash Surya * total amount of data in this state; this includes: evictable, 405244781f1SPrakash Surya * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 406244781f1SPrakash Surya */ 4072fd872a7SPrakash Surya refcount_t arcs_size; 408fa9e4066Sahrens } arc_state_t; 409fa9e4066Sahrens 410fa94a07fSbrendan /* The 6 states: */ 411fa9e4066Sahrens static arc_state_t ARC_anon; 412ea8dc4b6Seschrock static arc_state_t ARC_mru; 413ea8dc4b6Seschrock static arc_state_t ARC_mru_ghost; 414ea8dc4b6Seschrock static arc_state_t ARC_mfu; 415ea8dc4b6Seschrock static arc_state_t ARC_mfu_ghost; 416fa94a07fSbrendan static arc_state_t ARC_l2c_only; 417fa9e4066Sahrens 41844cb6abcSbmc typedef struct arc_stats { 41944cb6abcSbmc kstat_named_t arcstat_hits; 42044cb6abcSbmc kstat_named_t arcstat_misses; 42144cb6abcSbmc kstat_named_t arcstat_demand_data_hits; 42244cb6abcSbmc kstat_named_t arcstat_demand_data_misses; 42344cb6abcSbmc kstat_named_t arcstat_demand_metadata_hits; 42444cb6abcSbmc kstat_named_t arcstat_demand_metadata_misses; 42544cb6abcSbmc kstat_named_t arcstat_prefetch_data_hits; 42644cb6abcSbmc kstat_named_t arcstat_prefetch_data_misses; 42744cb6abcSbmc kstat_named_t arcstat_prefetch_metadata_hits; 42844cb6abcSbmc kstat_named_t arcstat_prefetch_metadata_misses; 42944cb6abcSbmc kstat_named_t arcstat_mru_hits; 43044cb6abcSbmc kstat_named_t arcstat_mru_ghost_hits; 43144cb6abcSbmc kstat_named_t arcstat_mfu_hits; 43244cb6abcSbmc kstat_named_t arcstat_mfu_ghost_hits; 43344cb6abcSbmc kstat_named_t arcstat_deleted; 4343e30c24aSWill Andrews /* 4353e30c24aSWill Andrews * Number of buffers that could not be evicted because the hash lock 4363e30c24aSWill Andrews * was held by another thread. The lock may not necessarily be held 4373e30c24aSWill Andrews * by something using the same buffer, since hash locks are shared 4383e30c24aSWill Andrews * by multiple buffers. 4393e30c24aSWill Andrews */ 44044cb6abcSbmc kstat_named_t arcstat_mutex_miss; 4413e30c24aSWill Andrews /* 4423e30c24aSWill Andrews * Number of buffers skipped because they have I/O in progress, are 4433e30c24aSWill Andrews * indrect prefetch buffers that have not lived long enough, or are 4443e30c24aSWill Andrews * not from the spa we're trying to evict from. 4453e30c24aSWill Andrews */ 44644cb6abcSbmc kstat_named_t arcstat_evict_skip; 447244781f1SPrakash Surya /* 448244781f1SPrakash Surya * Number of times arc_evict_state() was unable to evict enough 449244781f1SPrakash Surya * buffers to reach it's target amount. 450244781f1SPrakash Surya */ 451244781f1SPrakash Surya kstat_named_t arcstat_evict_not_enough; 4525ea40c06SBrendan Gregg - Sun Microsystems kstat_named_t arcstat_evict_l2_cached; 4535ea40c06SBrendan Gregg - Sun Microsystems kstat_named_t arcstat_evict_l2_eligible; 4545ea40c06SBrendan Gregg - Sun Microsystems kstat_named_t arcstat_evict_l2_ineligible; 455244781f1SPrakash Surya kstat_named_t arcstat_evict_l2_skip; 45644cb6abcSbmc kstat_named_t arcstat_hash_elements; 45744cb6abcSbmc kstat_named_t arcstat_hash_elements_max; 45844cb6abcSbmc kstat_named_t arcstat_hash_collisions; 45944cb6abcSbmc kstat_named_t arcstat_hash_chains; 46044cb6abcSbmc kstat_named_t arcstat_hash_chain_max; 46144cb6abcSbmc kstat_named_t arcstat_p; 46244cb6abcSbmc kstat_named_t arcstat_c; 46344cb6abcSbmc kstat_named_t arcstat_c_min; 46444cb6abcSbmc kstat_named_t arcstat_c_max; 46544cb6abcSbmc kstat_named_t arcstat_size; 466dcbf3bd6SGeorge Wilson /* 467770499e1SDan Kimmel * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd. 468dcbf3bd6SGeorge Wilson * Note that the compressed bytes may match the uncompressed bytes 469dcbf3bd6SGeorge Wilson * if the block is either not compressed or compressed arc is disabled. 470dcbf3bd6SGeorge Wilson */ 471dcbf3bd6SGeorge Wilson kstat_named_t arcstat_compressed_size; 472dcbf3bd6SGeorge Wilson /* 473770499e1SDan Kimmel * Uncompressed size of the data stored in b_pabd. If compressed 474dcbf3bd6SGeorge Wilson * arc is disabled then this value will be identical to the stat 475dcbf3bd6SGeorge Wilson * above. 476dcbf3bd6SGeorge Wilson */ 477dcbf3bd6SGeorge Wilson kstat_named_t arcstat_uncompressed_size; 478dcbf3bd6SGeorge Wilson /* 479dcbf3bd6SGeorge Wilson * Number of bytes stored in all the arc_buf_t's. This is classified 480dcbf3bd6SGeorge Wilson * as "overhead" since this data is typically short-lived and will 481dcbf3bd6SGeorge Wilson * be evicted from the arc when it becomes unreferenced unless the 482dcbf3bd6SGeorge Wilson * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level 483dcbf3bd6SGeorge Wilson * values have been set (see comment in dbuf.c for more information). 484dcbf3bd6SGeorge Wilson */ 485dcbf3bd6SGeorge Wilson kstat_named_t arcstat_overhead_size; 4864076b1bfSPrakash Surya /* 4874076b1bfSPrakash Surya * Number of bytes consumed by internal ARC structures necessary 4884076b1bfSPrakash Surya * for tracking purposes; these structures are not actually 4894076b1bfSPrakash Surya * backed by ARC buffers. This includes arc_buf_hdr_t structures 4904076b1bfSPrakash Surya * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 4914076b1bfSPrakash Surya * caches), and arc_buf_t structures (allocated via arc_buf_t 4924076b1bfSPrakash Surya * cache). 4934076b1bfSPrakash Surya */ 494fa94a07fSbrendan kstat_named_t arcstat_hdr_size; 4954076b1bfSPrakash Surya /* 4964076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers of type equal to 4974076b1bfSPrakash Surya * ARC_BUFC_DATA. This is generally consumed by buffers backing 4984076b1bfSPrakash Surya * on disk user data (e.g. plain file contents). 4994076b1bfSPrakash Surya */ 5005a98e54bSBrendan Gregg - Sun Microsystems kstat_named_t arcstat_data_size; 5014076b1bfSPrakash Surya /* 5024076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers of type equal to 5034076b1bfSPrakash Surya * ARC_BUFC_METADATA. This is generally consumed by buffers 5044076b1bfSPrakash Surya * backing on disk data that is used for internal ZFS 5054076b1bfSPrakash Surya * structures (e.g. ZAP, dnode, indirect blocks, etc). 5064076b1bfSPrakash Surya */ 5074076b1bfSPrakash Surya kstat_named_t arcstat_metadata_size; 5084076b1bfSPrakash Surya /* 5094076b1bfSPrakash Surya * Number of bytes consumed by various buffers and structures 5104076b1bfSPrakash Surya * not actually backed with ARC buffers. This includes bonus 5114076b1bfSPrakash Surya * buffers (allocated directly via zio_buf_* functions), 5124076b1bfSPrakash Surya * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 5134076b1bfSPrakash Surya * cache), and dnode_t structures (allocated via dnode_t cache). 5144076b1bfSPrakash Surya */ 5155a98e54bSBrendan Gregg - Sun Microsystems kstat_named_t arcstat_other_size; 5164076b1bfSPrakash Surya /* 5174076b1bfSPrakash Surya * Total number of bytes consumed by ARC buffers residing in the 5184076b1bfSPrakash Surya * arc_anon state. This includes *all* buffers in the arc_anon 5194076b1bfSPrakash Surya * state; e.g. data, metadata, evictable, and unevictable buffers 5204076b1bfSPrakash Surya * are all included in this value. 5214076b1bfSPrakash Surya */ 5224076b1bfSPrakash Surya kstat_named_t arcstat_anon_size; 5234076b1bfSPrakash Surya /* 5244076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that meet the 5254076b1bfSPrakash Surya * following criteria: backing buffers of type ARC_BUFC_DATA, 5264076b1bfSPrakash Surya * residing in the arc_anon state, and are eligible for eviction 5274076b1bfSPrakash Surya * (e.g. have no outstanding holds on the buffer). 5284076b1bfSPrakash Surya */ 5294076b1bfSPrakash Surya kstat_named_t arcstat_anon_evictable_data; 5304076b1bfSPrakash Surya /* 5314076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that meet the 5324076b1bfSPrakash Surya * following criteria: backing buffers of type ARC_BUFC_METADATA, 5334076b1bfSPrakash Surya * residing in the arc_anon state, and are eligible for eviction 5344076b1bfSPrakash Surya * (e.g. have no outstanding holds on the buffer). 5354076b1bfSPrakash Surya */ 5364076b1bfSPrakash Surya kstat_named_t arcstat_anon_evictable_metadata; 5374076b1bfSPrakash Surya /* 5384076b1bfSPrakash Surya * Total number of bytes consumed by ARC buffers residing in the 5394076b1bfSPrakash Surya * arc_mru state. This includes *all* buffers in the arc_mru 5404076b1bfSPrakash Surya * state; e.g. data, metadata, evictable, and unevictable buffers 5414076b1bfSPrakash Surya * are all included in this value. 5424076b1bfSPrakash Surya */ 5434076b1bfSPrakash Surya kstat_named_t arcstat_mru_size; 5444076b1bfSPrakash Surya /* 5454076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that meet the 5464076b1bfSPrakash Surya * following criteria: backing buffers of type ARC_BUFC_DATA, 5474076b1bfSPrakash Surya * residing in the arc_mru state, and are eligible for eviction 5484076b1bfSPrakash Surya * (e.g. have no outstanding holds on the buffer). 5494076b1bfSPrakash Surya */ 5504076b1bfSPrakash Surya kstat_named_t arcstat_mru_evictable_data; 5514076b1bfSPrakash Surya /* 5524076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that meet the 5534076b1bfSPrakash Surya * following criteria: backing buffers of type ARC_BUFC_METADATA, 5544076b1bfSPrakash Surya * residing in the arc_mru state, and are eligible for eviction 5554076b1bfSPrakash Surya * (e.g. have no outstanding holds on the buffer). 5564076b1bfSPrakash Surya */ 5574076b1bfSPrakash Surya kstat_named_t arcstat_mru_evictable_metadata; 5584076b1bfSPrakash Surya /* 5594076b1bfSPrakash Surya * Total number of bytes that *would have been* consumed by ARC 5604076b1bfSPrakash Surya * buffers in the arc_mru_ghost state. The key thing to note 5614076b1bfSPrakash Surya * here, is the fact that this size doesn't actually indicate 5624076b1bfSPrakash Surya * RAM consumption. The ghost lists only consist of headers and 5634076b1bfSPrakash Surya * don't actually have ARC buffers linked off of these headers. 5644076b1bfSPrakash Surya * Thus, *if* the headers had associated ARC buffers, these 5654076b1bfSPrakash Surya * buffers *would have* consumed this number of bytes. 5664076b1bfSPrakash Surya */ 5674076b1bfSPrakash Surya kstat_named_t arcstat_mru_ghost_size; 5684076b1bfSPrakash Surya /* 5694076b1bfSPrakash Surya * Number of bytes that *would have been* consumed by ARC 5704076b1bfSPrakash Surya * buffers that are eligible for eviction, of type 5714076b1bfSPrakash Surya * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 5724076b1bfSPrakash Surya */ 5734076b1bfSPrakash Surya kstat_named_t arcstat_mru_ghost_evictable_data; 5744076b1bfSPrakash Surya /* 5754076b1bfSPrakash Surya * Number of bytes that *would have been* consumed by ARC 5764076b1bfSPrakash Surya * buffers that are eligible for eviction, of type 5774076b1bfSPrakash Surya * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 5784076b1bfSPrakash Surya */ 5794076b1bfSPrakash Surya kstat_named_t arcstat_mru_ghost_evictable_metadata; 5804076b1bfSPrakash Surya /* 5814076b1bfSPrakash Surya * Total number of bytes consumed by ARC buffers residing in the 5824076b1bfSPrakash Surya * arc_mfu state. This includes *all* buffers in the arc_mfu 5834076b1bfSPrakash Surya * state; e.g. data, metadata, evictable, and unevictable buffers 5844076b1bfSPrakash Surya * are all included in this value. 5854076b1bfSPrakash Surya */ 5864076b1bfSPrakash Surya kstat_named_t arcstat_mfu_size; 5874076b1bfSPrakash Surya /* 5884076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that are eligible for 5894076b1bfSPrakash Surya * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 5904076b1bfSPrakash Surya * state. 5914076b1bfSPrakash Surya */ 5924076b1bfSPrakash Surya kstat_named_t arcstat_mfu_evictable_data; 5934076b1bfSPrakash Surya /* 5944076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that are eligible for 5954076b1bfSPrakash Surya * eviction, of type ARC_BUFC_METADATA, and reside in the 5964076b1bfSPrakash Surya * arc_mfu state. 5974076b1bfSPrakash Surya */ 5984076b1bfSPrakash Surya kstat_named_t arcstat_mfu_evictable_metadata; 5994076b1bfSPrakash Surya /* 6004076b1bfSPrakash Surya * Total number of bytes that *would have been* consumed by ARC 6014076b1bfSPrakash Surya * buffers in the arc_mfu_ghost state. See the comment above 6024076b1bfSPrakash Surya * arcstat_mru_ghost_size for more details. 6034076b1bfSPrakash Surya */ 6044076b1bfSPrakash Surya kstat_named_t arcstat_mfu_ghost_size; 6054076b1bfSPrakash Surya /* 6064076b1bfSPrakash Surya * Number of bytes that *would have been* consumed by ARC 6074076b1bfSPrakash Surya * buffers that are eligible for eviction, of type 6084076b1bfSPrakash Surya * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 6094076b1bfSPrakash Surya */ 6104076b1bfSPrakash Surya kstat_named_t arcstat_mfu_ghost_evictable_data; 6114076b1bfSPrakash Surya /* 6124076b1bfSPrakash Surya * Number of bytes that *would have been* consumed by ARC 6134076b1bfSPrakash Surya * buffers that are eligible for eviction, of type 6144076b1bfSPrakash Surya * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 6154076b1bfSPrakash Surya */ 6164076b1bfSPrakash Surya kstat_named_t arcstat_mfu_ghost_evictable_metadata; 617fa94a07fSbrendan kstat_named_t arcstat_l2_hits; 618fa94a07fSbrendan kstat_named_t arcstat_l2_misses; 619fa94a07fSbrendan kstat_named_t arcstat_l2_feeds; 620fa94a07fSbrendan kstat_named_t arcstat_l2_rw_clash; 6215a98e54bSBrendan Gregg - Sun Microsystems kstat_named_t arcstat_l2_read_bytes; 6225a98e54bSBrendan Gregg - Sun Microsystems kstat_named_t arcstat_l2_write_bytes; 623fa94a07fSbrendan kstat_named_t arcstat_l2_writes_sent; 624fa94a07fSbrendan kstat_named_t arcstat_l2_writes_done; 625fa94a07fSbrendan kstat_named_t arcstat_l2_writes_error; 626244781f1SPrakash Surya kstat_named_t arcstat_l2_writes_lock_retry; 627fa94a07fSbrendan kstat_named_t arcstat_l2_evict_lock_retry; 628fa94a07fSbrendan kstat_named_t arcstat_l2_evict_reading; 62989c86e32SChris Williamson kstat_named_t arcstat_l2_evict_l1cached; 630fa94a07fSbrendan kstat_named_t arcstat_l2_free_on_write; 631fa94a07fSbrendan kstat_named_t arcstat_l2_abort_lowmem; 632fa94a07fSbrendan kstat_named_t arcstat_l2_cksum_bad; 633fa94a07fSbrendan kstat_named_t arcstat_l2_io_error; 63416a7e5acSAndriy Gapon kstat_named_t arcstat_l2_lsize; 63516a7e5acSAndriy Gapon kstat_named_t arcstat_l2_psize; 636fa94a07fSbrendan kstat_named_t arcstat_l2_hdr_size; 6371ab7f2deSmaybee kstat_named_t arcstat_memory_throttle_count; 63820128a08SGeorge Wilson kstat_named_t arcstat_meta_used; 63920128a08SGeorge Wilson kstat_named_t arcstat_meta_limit; 64020128a08SGeorge Wilson kstat_named_t arcstat_meta_max; 6413a5286a1SMatthew Ahrens kstat_named_t arcstat_meta_min; 642cf6106c8SMatthew Ahrens kstat_named_t arcstat_sync_wait_for_async; 643cf6106c8SMatthew Ahrens kstat_named_t arcstat_demand_hit_predictive_prefetch; 64444cb6abcSbmc } arc_stats_t; 64544cb6abcSbmc 64644cb6abcSbmc static arc_stats_t arc_stats = { 64744cb6abcSbmc { "hits", KSTAT_DATA_UINT64 }, 64844cb6abcSbmc { "misses", KSTAT_DATA_UINT64 }, 64944cb6abcSbmc { "demand_data_hits", KSTAT_DATA_UINT64 }, 65044cb6abcSbmc { "demand_data_misses", KSTAT_DATA_UINT64 }, 65144cb6abcSbmc { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 65244cb6abcSbmc { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 65344cb6abcSbmc { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 65444cb6abcSbmc { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 65544cb6abcSbmc { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 65644cb6abcSbmc { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 65744cb6abcSbmc { "mru_hits", KSTAT_DATA_UINT64 }, 65844cb6abcSbmc { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 65944cb6abcSbmc { "mfu_hits", KSTAT_DATA_UINT64 }, 66044cb6abcSbmc { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 66144cb6abcSbmc { "deleted", KSTAT_DATA_UINT64 }, 66244cb6abcSbmc { "mutex_miss", KSTAT_DATA_UINT64 }, 66344cb6abcSbmc { "evict_skip", KSTAT_DATA_UINT64 }, 664244781f1SPrakash Surya { "evict_not_enough", KSTAT_DATA_UINT64 }, 6655ea40c06SBrendan Gregg - Sun Microsystems { "evict_l2_cached", KSTAT_DATA_UINT64 }, 6665ea40c06SBrendan Gregg - Sun Microsystems { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 6675ea40c06SBrendan Gregg - Sun Microsystems { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 668244781f1SPrakash Surya { "evict_l2_skip", KSTAT_DATA_UINT64 }, 66944cb6abcSbmc { "hash_elements", KSTAT_DATA_UINT64 }, 67044cb6abcSbmc { "hash_elements_max", KSTAT_DATA_UINT64 }, 67144cb6abcSbmc { "hash_collisions", KSTAT_DATA_UINT64 }, 67244cb6abcSbmc { "hash_chains", KSTAT_DATA_UINT64 }, 67344cb6abcSbmc { "hash_chain_max", KSTAT_DATA_UINT64 }, 67444cb6abcSbmc { "p", KSTAT_DATA_UINT64 }, 67544cb6abcSbmc { "c", KSTAT_DATA_UINT64 }, 67644cb6abcSbmc { "c_min", KSTAT_DATA_UINT64 }, 67744cb6abcSbmc { "c_max", KSTAT_DATA_UINT64 }, 678fa94a07fSbrendan { "size", KSTAT_DATA_UINT64 }, 679dcbf3bd6SGeorge Wilson { "compressed_size", KSTAT_DATA_UINT64 }, 680dcbf3bd6SGeorge Wilson { "uncompressed_size", KSTAT_DATA_UINT64 }, 681dcbf3bd6SGeorge Wilson { "overhead_size", KSTAT_DATA_UINT64 }, 682fa94a07fSbrendan { "hdr_size", KSTAT_DATA_UINT64 }, 6835a98e54bSBrendan Gregg - Sun Microsystems { "data_size", KSTAT_DATA_UINT64 }, 6844076b1bfSPrakash Surya { "metadata_size", KSTAT_DATA_UINT64 }, 6855a98e54bSBrendan Gregg - Sun Microsystems { "other_size", KSTAT_DATA_UINT64 }, 6864076b1bfSPrakash Surya { "anon_size", KSTAT_DATA_UINT64 }, 6874076b1bfSPrakash Surya { "anon_evictable_data", KSTAT_DATA_UINT64 }, 6884076b1bfSPrakash Surya { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 6894076b1bfSPrakash Surya { "mru_size", KSTAT_DATA_UINT64 }, 6904076b1bfSPrakash Surya { "mru_evictable_data", KSTAT_DATA_UINT64 }, 6914076b1bfSPrakash Surya { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 6924076b1bfSPrakash Surya { "mru_ghost_size", KSTAT_DATA_UINT64 }, 6934076b1bfSPrakash Surya { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 6944076b1bfSPrakash Surya { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 6954076b1bfSPrakash Surya { "mfu_size", KSTAT_DATA_UINT64 }, 6964076b1bfSPrakash Surya { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 6974076b1bfSPrakash Surya { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 6984076b1bfSPrakash Surya { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 6994076b1bfSPrakash Surya { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 7004076b1bfSPrakash Surya { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 701fa94a07fSbrendan { "l2_hits", KSTAT_DATA_UINT64 }, 702fa94a07fSbrendan { "l2_misses", KSTAT_DATA_UINT64 }, 703fa94a07fSbrendan { "l2_feeds", KSTAT_DATA_UINT64 }, 704fa94a07fSbrendan { "l2_rw_clash", KSTAT_DATA_UINT64 }, 7055a98e54bSBrendan Gregg - Sun Microsystems { "l2_read_bytes", KSTAT_DATA_UINT64 }, 7065a98e54bSBrendan Gregg - Sun Microsystems { "l2_write_bytes", KSTAT_DATA_UINT64 }, 707fa94a07fSbrendan { "l2_writes_sent", KSTAT_DATA_UINT64 }, 708fa94a07fSbrendan { "l2_writes_done", KSTAT_DATA_UINT64 }, 709fa94a07fSbrendan { "l2_writes_error", KSTAT_DATA_UINT64 }, 710244781f1SPrakash Surya { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 711fa94a07fSbrendan { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 712fa94a07fSbrendan { "l2_evict_reading", KSTAT_DATA_UINT64 }, 71389c86e32SChris Williamson { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 714fa94a07fSbrendan { "l2_free_on_write", KSTAT_DATA_UINT64 }, 715fa94a07fSbrendan { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 716fa94a07fSbrendan { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 717fa94a07fSbrendan { "l2_io_error", KSTAT_DATA_UINT64 }, 718fa94a07fSbrendan { "l2_size", KSTAT_DATA_UINT64 }, 719aad02571SSaso Kiselkov { "l2_asize", KSTAT_DATA_UINT64 }, 7201ab7f2deSmaybee { "l2_hdr_size", KSTAT_DATA_UINT64 }, 7219253d63dSGeorge Wilson { "memory_throttle_count", KSTAT_DATA_UINT64 }, 72220128a08SGeorge Wilson { "arc_meta_used", KSTAT_DATA_UINT64 }, 72320128a08SGeorge Wilson { "arc_meta_limit", KSTAT_DATA_UINT64 }, 7243a5286a1SMatthew Ahrens { "arc_meta_max", KSTAT_DATA_UINT64 }, 725cf6106c8SMatthew Ahrens { "arc_meta_min", KSTAT_DATA_UINT64 }, 726cf6106c8SMatthew Ahrens { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 727cf6106c8SMatthew Ahrens { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 72844cb6abcSbmc }; 72944cb6abcSbmc 73044cb6abcSbmc #define ARCSTAT(stat) (arc_stats.stat.value.ui64) 73144cb6abcSbmc 73244cb6abcSbmc #define ARCSTAT_INCR(stat, val) \ 733f7170741SWill Andrews atomic_add_64(&arc_stats.stat.value.ui64, (val)) 73444cb6abcSbmc 735b24ab676SJeff Bonwick #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 73644cb6abcSbmc #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 73744cb6abcSbmc 73844cb6abcSbmc #define ARCSTAT_MAX(stat, val) { \ 73944cb6abcSbmc uint64_t m; \ 74044cb6abcSbmc while ((val) > (m = arc_stats.stat.value.ui64) && \ 74144cb6abcSbmc (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 74244cb6abcSbmc continue; \ 74344cb6abcSbmc } 74444cb6abcSbmc 74544cb6abcSbmc #define ARCSTAT_MAXSTAT(stat) \ 74644cb6abcSbmc ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 74744cb6abcSbmc 74844cb6abcSbmc /* 74944cb6abcSbmc * We define a macro to allow ARC hits/misses to be easily broken down by 75044cb6abcSbmc * two separate conditions, giving a total of four different subtypes for 75144cb6abcSbmc * each of hits and misses (so eight statistics total). 75244cb6abcSbmc */ 75344cb6abcSbmc #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 75444cb6abcSbmc if (cond1) { \ 75544cb6abcSbmc if (cond2) { \ 75644cb6abcSbmc ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 75744cb6abcSbmc } else { \ 75844cb6abcSbmc ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 75944cb6abcSbmc } \ 76044cb6abcSbmc } else { \ 76144cb6abcSbmc if (cond2) { \ 76244cb6abcSbmc ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 76344cb6abcSbmc } else { \ 76444cb6abcSbmc ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 76544cb6abcSbmc } \ 76644cb6abcSbmc } 76744cb6abcSbmc 76844cb6abcSbmc kstat_t *arc_ksp; 769b24ab676SJeff Bonwick static arc_state_t *arc_anon; 77044cb6abcSbmc static arc_state_t *arc_mru; 77144cb6abcSbmc static arc_state_t *arc_mru_ghost; 77244cb6abcSbmc static arc_state_t *arc_mfu; 77344cb6abcSbmc static arc_state_t *arc_mfu_ghost; 774fa94a07fSbrendan static arc_state_t *arc_l2c_only; 77544cb6abcSbmc 77644cb6abcSbmc /* 77744cb6abcSbmc * There are several ARC variables that are critical to export as kstats -- 77844cb6abcSbmc * but we don't want to have to grovel around in the kstat whenever we wish to 77944cb6abcSbmc * manipulate them. For these variables, we therefore define them to be in 78044cb6abcSbmc * terms of the statistic variable. This assures that we are not introducing 78144cb6abcSbmc * the possibility of inconsistency by having shadow copies of the variables, 78244cb6abcSbmc * while still allowing the code to be readable. 78344cb6abcSbmc */ 78444cb6abcSbmc #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 78544cb6abcSbmc #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 78644cb6abcSbmc #define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 78744cb6abcSbmc #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 78844cb6abcSbmc #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 78920128a08SGeorge Wilson #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 7903a5286a1SMatthew Ahrens #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 79120128a08SGeorge Wilson #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 79220128a08SGeorge Wilson #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 79344cb6abcSbmc 794dcbf3bd6SGeorge Wilson /* compressed size of entire arc */ 795dcbf3bd6SGeorge Wilson #define arc_compressed_size ARCSTAT(arcstat_compressed_size) 796dcbf3bd6SGeorge Wilson /* uncompressed size of entire arc */ 797dcbf3bd6SGeorge Wilson #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size) 798dcbf3bd6SGeorge Wilson /* number of bytes in the arc from arc_buf_t's */ 799dcbf3bd6SGeorge Wilson #define arc_overhead_size ARCSTAT(arcstat_overhead_size) 800aad02571SSaso Kiselkov 80144cb6abcSbmc static int arc_no_grow; /* Don't try to grow cache size */ 80244cb6abcSbmc static uint64_t arc_tempreserve; 8032fdbea25SAleksandr Guzovskiy static uint64_t arc_loaned_bytes; 804fa9e4066Sahrens 805fa9e4066Sahrens typedef struct arc_callback arc_callback_t; 806fa9e4066Sahrens 807fa9e4066Sahrens struct arc_callback { 808fa9e4066Sahrens void *acb_private; 809c717a561Smaybee arc_done_func_t *acb_done; 810fa9e4066Sahrens arc_buf_t *acb_buf; 8115602294fSDan Kimmel boolean_t acb_compressed; 812fa9e4066Sahrens zio_t *acb_zio_dummy; 813fa9e4066Sahrens arc_callback_t *acb_next; 814fa9e4066Sahrens }; 815fa9e4066Sahrens 816c717a561Smaybee typedef struct arc_write_callback arc_write_callback_t; 817c717a561Smaybee 818c717a561Smaybee struct arc_write_callback { 819c717a561Smaybee void *awcb_private; 820c717a561Smaybee arc_done_func_t *awcb_ready; 8218df0bcf0SPaul Dagnelie arc_done_func_t *awcb_children_ready; 82269962b56SMatthew Ahrens arc_done_func_t *awcb_physdone; 823c717a561Smaybee arc_done_func_t *awcb_done; 824c717a561Smaybee arc_buf_t *awcb_buf; 825c717a561Smaybee }; 826c717a561Smaybee 82789c86e32SChris Williamson /* 82889c86e32SChris Williamson * ARC buffers are separated into multiple structs as a memory saving measure: 82989c86e32SChris Williamson * - Common fields struct, always defined, and embedded within it: 83089c86e32SChris Williamson * - L2-only fields, always allocated but undefined when not in L2ARC 83189c86e32SChris Williamson * - L1-only fields, only allocated when in L1ARC 83289c86e32SChris Williamson * 83389c86e32SChris Williamson * Buffer in L1 Buffer only in L2 83489c86e32SChris Williamson * +------------------------+ +------------------------+ 83589c86e32SChris Williamson * | arc_buf_hdr_t | | arc_buf_hdr_t | 83689c86e32SChris Williamson * | | | | 83789c86e32SChris Williamson * | | | | 83889c86e32SChris Williamson * | | | | 83989c86e32SChris Williamson * +------------------------+ +------------------------+ 84089c86e32SChris Williamson * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 84189c86e32SChris Williamson * | (undefined if L1-only) | | | 84289c86e32SChris Williamson * +------------------------+ +------------------------+ 84389c86e32SChris Williamson * | l1arc_buf_hdr_t | 84489c86e32SChris Williamson * | | 84589c86e32SChris Williamson * | | 84689c86e32SChris Williamson * | | 84789c86e32SChris Williamson * | | 84889c86e32SChris Williamson * +------------------------+ 84989c86e32SChris Williamson * 85089c86e32SChris Williamson * Because it's possible for the L2ARC to become extremely large, we can wind 85189c86e32SChris Williamson * up eating a lot of memory in L2ARC buffer headers, so the size of a header 85289c86e32SChris Williamson * is minimized by only allocating the fields necessary for an L1-cached buffer 85389c86e32SChris Williamson * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 85489c86e32SChris Williamson * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 85589c86e32SChris Williamson * words in pointers. arc_hdr_realloc() is used to switch a header between 85689c86e32SChris Williamson * these two allocation states. 85789c86e32SChris Williamson */ 85889c86e32SChris Williamson typedef struct l1arc_buf_hdr { 8596b4acc8bSahrens kmutex_t b_freeze_lock; 860dcbf3bd6SGeorge Wilson zio_cksum_t *b_freeze_cksum; 86189c86e32SChris Williamson #ifdef ZFS_DEBUG 86289c86e32SChris Williamson /* 8635602294fSDan Kimmel * Used for debugging with kmem_flags - by allocating and freeing 86489c86e32SChris Williamson * b_thawed when the buffer is thawed, we get a record of the stack 86589c86e32SChris Williamson * trace that thawed it. 86689c86e32SChris Williamson */ 8673f9d6ad7SLin Ling void *b_thawed; 86889c86e32SChris Williamson #endif 8696b4acc8bSahrens 870fa9e4066Sahrens arc_buf_t *b_buf; 871dcbf3bd6SGeorge Wilson uint32_t b_bufcnt; 87289c86e32SChris Williamson /* for waiting on writes to complete */ 873ad23a2dbSjohansen kcondvar_t b_cv; 874dcbf3bd6SGeorge Wilson uint8_t b_byteswap; 875ad23a2dbSjohansen 876fa9e4066Sahrens /* protected by arc state mutex */ 877fa9e4066Sahrens arc_state_t *b_state; 878244781f1SPrakash Surya multilist_node_t b_arc_node; 879fa9e4066Sahrens 880fa9e4066Sahrens /* updated atomically */ 881fa9e4066Sahrens clock_t b_arc_access; 882fa9e4066Sahrens 883fa9e4066Sahrens /* self protecting */ 884fa9e4066Sahrens refcount_t b_refcnt; 885fa94a07fSbrendan 88689c86e32SChris Williamson arc_callback_t *b_acb; 887770499e1SDan Kimmel abd_t *b_pabd; 88889c86e32SChris Williamson } l1arc_buf_hdr_t; 88989c86e32SChris Williamson 89089c86e32SChris Williamson typedef struct l2arc_dev l2arc_dev_t; 89189c86e32SChris Williamson 89289c86e32SChris Williamson typedef struct l2arc_buf_hdr { 89389c86e32SChris Williamson /* protected by arc_buf_hdr mutex */ 89489c86e32SChris Williamson l2arc_dev_t *b_dev; /* L2ARC device */ 89589c86e32SChris Williamson uint64_t b_daddr; /* disk address, offset byte */ 89689c86e32SChris Williamson 897fa94a07fSbrendan list_node_t b_l2node; 89889c86e32SChris Williamson } l2arc_buf_hdr_t; 89989c86e32SChris Williamson 90089c86e32SChris Williamson struct arc_buf_hdr { 90189c86e32SChris Williamson /* protected by hash lock */ 90289c86e32SChris Williamson dva_t b_dva; 90389c86e32SChris Williamson uint64_t b_birth; 90489c86e32SChris Williamson 905dcbf3bd6SGeorge Wilson arc_buf_contents_t b_type; 90689c86e32SChris Williamson arc_buf_hdr_t *b_hash_next; 90789c86e32SChris Williamson arc_flags_t b_flags; 90889c86e32SChris Williamson 909dcbf3bd6SGeorge Wilson /* 910dcbf3bd6SGeorge Wilson * This field stores the size of the data buffer after 911dcbf3bd6SGeorge Wilson * compression, and is set in the arc's zio completion handlers. 912dcbf3bd6SGeorge Wilson * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes). 913dcbf3bd6SGeorge Wilson * 914dcbf3bd6SGeorge Wilson * While the block pointers can store up to 32MB in their psize 915dcbf3bd6SGeorge Wilson * field, we can only store up to 32MB minus 512B. This is due 916dcbf3bd6SGeorge Wilson * to the bp using a bias of 1, whereas we use a bias of 0 (i.e. 917dcbf3bd6SGeorge Wilson * a field of zeros represents 512B in the bp). We can't use a 918dcbf3bd6SGeorge Wilson * bias of 1 since we need to reserve a psize of zero, here, to 919dcbf3bd6SGeorge Wilson * represent holes and embedded blocks. 920dcbf3bd6SGeorge Wilson * 921dcbf3bd6SGeorge Wilson * This isn't a problem in practice, since the maximum size of a 922dcbf3bd6SGeorge Wilson * buffer is limited to 16MB, so we never need to store 32MB in 923dcbf3bd6SGeorge Wilson * this field. Even in the upstream illumos code base, the 924dcbf3bd6SGeorge Wilson * maximum size of a buffer is limited to 16MB. 925dcbf3bd6SGeorge Wilson */ 926dcbf3bd6SGeorge Wilson uint16_t b_psize; 927dcbf3bd6SGeorge Wilson 928dcbf3bd6SGeorge Wilson /* 929dcbf3bd6SGeorge Wilson * This field stores the size of the data buffer before 930dcbf3bd6SGeorge Wilson * compression, and cannot change once set. It is in units 931dcbf3bd6SGeorge Wilson * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes) 932dcbf3bd6SGeorge Wilson */ 933dcbf3bd6SGeorge Wilson uint16_t b_lsize; /* immutable */ 934dcbf3bd6SGeorge Wilson uint64_t b_spa; /* immutable */ 93589c86e32SChris Williamson 93689c86e32SChris Williamson /* L2ARC fields. Undefined when not in L2ARC. */ 93789c86e32SChris Williamson l2arc_buf_hdr_t b_l2hdr; 93889c86e32SChris Williamson /* L1ARC fields. Undefined when in l2arc_only state */ 93989c86e32SChris Williamson l1arc_buf_hdr_t b_l1hdr; 940fa9e4066Sahrens }; 941fa9e4066Sahrens 942ea8dc4b6Seschrock #define GHOST_STATE(state) \ 943fa94a07fSbrendan ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 944fa94a07fSbrendan (state) == arc_l2c_only) 945ea8dc4b6Seschrock 9467adb730bSGeorge Wilson #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 9477adb730bSGeorge Wilson #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 9487adb730bSGeorge Wilson #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 9497adb730bSGeorge Wilson #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 950dcbf3bd6SGeorge Wilson #define HDR_COMPRESSION_ENABLED(hdr) \ 951dcbf3bd6SGeorge Wilson ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC) 95289c86e32SChris Williamson 9537adb730bSGeorge Wilson #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 9547adb730bSGeorge Wilson #define HDR_L2_READING(hdr) \ 955dcbf3bd6SGeorge Wilson (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 956dcbf3bd6SGeorge Wilson ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 9577adb730bSGeorge Wilson #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 9587adb730bSGeorge Wilson #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 9597adb730bSGeorge Wilson #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 960dcbf3bd6SGeorge Wilson #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA) 961fa9e4066Sahrens 96289c86e32SChris Williamson #define HDR_ISTYPE_METADATA(hdr) \ 963dcbf3bd6SGeorge Wilson ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 96489c86e32SChris Williamson #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 96589c86e32SChris Williamson 96689c86e32SChris Williamson #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 96789c86e32SChris Williamson #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 96889c86e32SChris Williamson 969dcbf3bd6SGeorge Wilson /* For storing compression mode in b_flags */ 970dcbf3bd6SGeorge Wilson #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1) 971dcbf3bd6SGeorge Wilson 972dcbf3bd6SGeorge Wilson #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \ 973dcbf3bd6SGeorge Wilson HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS)) 974dcbf3bd6SGeorge Wilson #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \ 975dcbf3bd6SGeorge Wilson HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp)); 976dcbf3bd6SGeorge Wilson 977dcbf3bd6SGeorge Wilson #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL) 9785602294fSDan Kimmel #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED) 9795602294fSDan Kimmel #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED) 980dcbf3bd6SGeorge Wilson 981e6c728e1Sbrendan /* 982e6c728e1Sbrendan * Other sizes 983e6c728e1Sbrendan */ 984e6c728e1Sbrendan 98589c86e32SChris Williamson #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 98689c86e32SChris Williamson #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 987e6c728e1Sbrendan 988fa9e4066Sahrens /* 989fa9e4066Sahrens * Hash table routines 990fa9e4066Sahrens */ 991fa9e4066Sahrens 992fa9e4066Sahrens #define HT_LOCK_PAD 64 993fa9e4066Sahrens 994fa9e4066Sahrens struct ht_lock { 995fa9e4066Sahrens kmutex_t ht_lock; 996fa9e4066Sahrens #ifdef _KERNEL 997fa9e4066Sahrens unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 998fa9e4066Sahrens #endif 999fa9e4066Sahrens }; 1000fa9e4066Sahrens 1001fa9e4066Sahrens #define BUF_LOCKS 256 1002fa9e4066Sahrens typedef struct buf_hash_table { 1003fa9e4066Sahrens uint64_t ht_mask; 1004fa9e4066Sahrens arc_buf_hdr_t **ht_table; 1005fa9e4066Sahrens struct ht_lock ht_locks[BUF_LOCKS]; 1006fa9e4066Sahrens } buf_hash_table_t; 1007fa9e4066Sahrens 1008fa9e4066Sahrens static buf_hash_table_t buf_hash_table; 1009fa9e4066Sahrens 1010fa9e4066Sahrens #define BUF_HASH_INDEX(spa, dva, birth) \ 1011fa9e4066Sahrens (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 1012fa9e4066Sahrens #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 1013fa9e4066Sahrens #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 10143f9d6ad7SLin Ling #define HDR_LOCK(hdr) \ 10153f9d6ad7SLin Ling (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 1016fa9e4066Sahrens 1017fa9e4066Sahrens uint64_t zfs_crc64_table[256]; 1018fa9e4066Sahrens 1019fa94a07fSbrendan /* 1020fa94a07fSbrendan * Level 2 ARC 1021fa94a07fSbrendan */ 1022fa94a07fSbrendan 1023fa94a07fSbrendan #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 1024aad02571SSaso Kiselkov #define L2ARC_HEADROOM 2 /* num of writes */ 1025aad02571SSaso Kiselkov /* 1026aad02571SSaso Kiselkov * If we discover during ARC scan any buffers to be compressed, we boost 1027aad02571SSaso Kiselkov * our headroom for the next scanning cycle by this percentage multiple. 1028aad02571SSaso Kiselkov */ 1029aad02571SSaso Kiselkov #define L2ARC_HEADROOM_BOOST 200 10305a98e54bSBrendan Gregg - Sun Microsystems #define L2ARC_FEED_SECS 1 /* caching interval secs */ 10315a98e54bSBrendan Gregg - Sun Microsystems #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 1032fa94a07fSbrendan 1033fa94a07fSbrendan #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 1034fa94a07fSbrendan #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 1035fa94a07fSbrendan 1036f7170741SWill Andrews /* L2ARC Performance Tunables */ 1037fa94a07fSbrendan uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 10383a737e0dSbrendan uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1039fa94a07fSbrendan uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1040aad02571SSaso Kiselkov uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1041fa94a07fSbrendan uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 10425a98e54bSBrendan Gregg - Sun Microsystems uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1043fa94a07fSbrendan boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 10445a98e54bSBrendan Gregg - Sun Microsystems boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 10455a98e54bSBrendan Gregg - Sun Microsystems boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1046fa94a07fSbrendan 1047fa94a07fSbrendan /* 1048fa94a07fSbrendan * L2ARC Internals 1049fa94a07fSbrendan */ 105089c86e32SChris Williamson struct l2arc_dev { 1051fa94a07fSbrendan vdev_t *l2ad_vdev; /* vdev */ 1052fa94a07fSbrendan spa_t *l2ad_spa; /* spa */ 1053fa94a07fSbrendan uint64_t l2ad_hand; /* next write location */ 1054fa94a07fSbrendan uint64_t l2ad_start; /* first addr on device */ 1055fa94a07fSbrendan uint64_t l2ad_end; /* last addr on device */ 1056fa94a07fSbrendan boolean_t l2ad_first; /* first sweep through */ 10575a98e54bSBrendan Gregg - Sun Microsystems boolean_t l2ad_writing; /* currently writing */ 105889c86e32SChris Williamson kmutex_t l2ad_mtx; /* lock for buffer list */ 105989c86e32SChris Williamson list_t l2ad_buflist; /* buffer list */ 1060fa94a07fSbrendan list_node_t l2ad_node; /* device list node */ 1061a52fc310SPrakash Surya refcount_t l2ad_alloc; /* allocated bytes */ 106289c86e32SChris Williamson }; 1063fa94a07fSbrendan 1064fa94a07fSbrendan static list_t L2ARC_dev_list; /* device list */ 1065fa94a07fSbrendan static list_t *l2arc_dev_list; /* device list pointer */ 1066fa94a07fSbrendan static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1067fa94a07fSbrendan static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1068fa94a07fSbrendan static list_t L2ARC_free_on_write; /* free after write buf list */ 1069fa94a07fSbrendan static list_t *l2arc_free_on_write; /* free after write list ptr */ 1070fa94a07fSbrendan static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1071fa94a07fSbrendan static uint64_t l2arc_ndev; /* number of devices */ 1072fa94a07fSbrendan 1073fa94a07fSbrendan typedef struct l2arc_read_callback { 10745602294fSDan Kimmel arc_buf_hdr_t *l2rcb_hdr; /* read header */ 1075aad02571SSaso Kiselkov blkptr_t l2rcb_bp; /* original blkptr */ 10767802d7bfSMatthew Ahrens zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1077aad02571SSaso Kiselkov int l2rcb_flags; /* original flags */ 1078403a8da7SAndriy Gapon abd_t *l2rcb_abd; /* temporary buffer */ 1079fa94a07fSbrendan } l2arc_read_callback_t; 1080fa94a07fSbrendan 1081fa94a07fSbrendan typedef struct l2arc_write_callback { 1082fa94a07fSbrendan l2arc_dev_t *l2wcb_dev; /* device info */ 1083fa94a07fSbrendan arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1084fa94a07fSbrendan } l2arc_write_callback_t; 1085fa94a07fSbrendan 1086fa94a07fSbrendan typedef struct l2arc_data_free { 1087fa94a07fSbrendan /* protected by l2arc_free_on_write_mtx */ 1088770499e1SDan Kimmel abd_t *l2df_abd; 1089fa94a07fSbrendan size_t l2df_size; 1090dcbf3bd6SGeorge Wilson arc_buf_contents_t l2df_type; 1091fa94a07fSbrendan list_node_t l2df_list_node; 1092fa94a07fSbrendan } l2arc_data_free_t; 1093fa94a07fSbrendan 1094fa94a07fSbrendan static kmutex_t l2arc_feed_thr_lock; 1095fa94a07fSbrendan static kcondvar_t l2arc_feed_thr_cv; 1096fa94a07fSbrendan static uint8_t l2arc_thread_exit; 1097fa94a07fSbrendan 1098770499e1SDan Kimmel static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *); 1099dcbf3bd6SGeorge Wilson static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); 1100770499e1SDan Kimmel static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *); 1101770499e1SDan Kimmel static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *); 1102dcbf3bd6SGeorge Wilson static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); 1103770499e1SDan Kimmel static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag); 1104770499e1SDan Kimmel static void arc_hdr_free_pabd(arc_buf_hdr_t *); 1105770499e1SDan Kimmel static void arc_hdr_alloc_pabd(arc_buf_hdr_t *); 11067adb730bSGeorge Wilson static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1107244781f1SPrakash Surya static boolean_t arc_is_overflowing(); 11087adb730bSGeorge Wilson static void arc_buf_watch(arc_buf_t *); 11097adb730bSGeorge Wilson 111089c86e32SChris Williamson static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 111189c86e32SChris Williamson static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1112dcbf3bd6SGeorge Wilson static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1113dcbf3bd6SGeorge Wilson static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 111489c86e32SChris Williamson 11157adb730bSGeorge Wilson static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 11167adb730bSGeorge Wilson static void l2arc_read_done(zio_t *); 1117fa94a07fSbrendan 1118fa9e4066Sahrens static uint64_t 1119ac05c741SMark Maybee buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1120fa9e4066Sahrens { 1121fa9e4066Sahrens uint8_t *vdva = (uint8_t *)dva; 1122fa9e4066Sahrens uint64_t crc = -1ULL; 1123fa9e4066Sahrens int i; 1124fa9e4066Sahrens 1125fa9e4066Sahrens ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1126fa9e4066Sahrens 1127fa9e4066Sahrens for (i = 0; i < sizeof (dva_t); i++) 1128fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1129fa9e4066Sahrens 1130ac05c741SMark Maybee crc ^= (spa>>8) ^ birth; 1131fa9e4066Sahrens 1132fa9e4066Sahrens return (crc); 1133fa9e4066Sahrens } 1134fa9e4066Sahrens 1135dcbf3bd6SGeorge Wilson #define HDR_EMPTY(hdr) \ 1136dcbf3bd6SGeorge Wilson ((hdr)->b_dva.dva_word[0] == 0 && \ 1137dcbf3bd6SGeorge Wilson (hdr)->b_dva.dva_word[1] == 0) 1138fa9e4066Sahrens 1139dcbf3bd6SGeorge Wilson #define HDR_EQUAL(spa, dva, birth, hdr) \ 1140dcbf3bd6SGeorge Wilson ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1141dcbf3bd6SGeorge Wilson ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1142dcbf3bd6SGeorge Wilson ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) 1143fa9e4066Sahrens 11443f9d6ad7SLin Ling static void 11453f9d6ad7SLin Ling buf_discard_identity(arc_buf_hdr_t *hdr) 11463f9d6ad7SLin Ling { 11473f9d6ad7SLin Ling hdr->b_dva.dva_word[0] = 0; 11483f9d6ad7SLin Ling hdr->b_dva.dva_word[1] = 0; 11493f9d6ad7SLin Ling hdr->b_birth = 0; 11503f9d6ad7SLin Ling } 11513f9d6ad7SLin Ling 1152fa9e4066Sahrens static arc_buf_hdr_t * 11535d7b4d43SMatthew Ahrens buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1154fa9e4066Sahrens { 11555d7b4d43SMatthew Ahrens const dva_t *dva = BP_IDENTITY(bp); 11565d7b4d43SMatthew Ahrens uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1157fa9e4066Sahrens uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1158fa9e4066Sahrens kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 11597adb730bSGeorge Wilson arc_buf_hdr_t *hdr; 1160fa9e4066Sahrens 1161fa9e4066Sahrens mutex_enter(hash_lock); 11627adb730bSGeorge Wilson for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 11637adb730bSGeorge Wilson hdr = hdr->b_hash_next) { 1164dcbf3bd6SGeorge Wilson if (HDR_EQUAL(spa, dva, birth, hdr)) { 1165fa9e4066Sahrens *lockp = hash_lock; 11667adb730bSGeorge Wilson return (hdr); 1167fa9e4066Sahrens } 1168fa9e4066Sahrens } 1169fa9e4066Sahrens mutex_exit(hash_lock); 1170fa9e4066Sahrens *lockp = NULL; 1171fa9e4066Sahrens return (NULL); 1172fa9e4066Sahrens } 1173fa9e4066Sahrens 1174fa9e4066Sahrens /* 1175fa9e4066Sahrens * Insert an entry into the hash table. If there is already an element 1176fa9e4066Sahrens * equal to elem in the hash table, then the already existing element 1177fa9e4066Sahrens * will be returned and the new element will not be inserted. 1178fa9e4066Sahrens * Otherwise returns NULL. 117989c86e32SChris Williamson * If lockp == NULL, the caller is assumed to already hold the hash lock. 1180fa9e4066Sahrens */ 1181fa9e4066Sahrens static arc_buf_hdr_t * 11827adb730bSGeorge Wilson buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1183fa9e4066Sahrens { 11847adb730bSGeorge Wilson uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1185fa9e4066Sahrens kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 11867adb730bSGeorge Wilson arc_buf_hdr_t *fhdr; 118744cb6abcSbmc uint32_t i; 1188fa9e4066Sahrens 11897adb730bSGeorge Wilson ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 11907adb730bSGeorge Wilson ASSERT(hdr->b_birth != 0); 11917adb730bSGeorge Wilson ASSERT(!HDR_IN_HASH_TABLE(hdr)); 119289c86e32SChris Williamson 119389c86e32SChris Williamson if (lockp != NULL) { 119489c86e32SChris Williamson *lockp = hash_lock; 119589c86e32SChris Williamson mutex_enter(hash_lock); 119689c86e32SChris Williamson } else { 119789c86e32SChris Williamson ASSERT(MUTEX_HELD(hash_lock)); 119889c86e32SChris Williamson } 119989c86e32SChris Williamson 12007adb730bSGeorge Wilson for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 12017adb730bSGeorge Wilson fhdr = fhdr->b_hash_next, i++) { 1202dcbf3bd6SGeorge Wilson if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 12037adb730bSGeorge Wilson return (fhdr); 1204fa9e4066Sahrens } 1205fa9e4066Sahrens 12067adb730bSGeorge Wilson hdr->b_hash_next = buf_hash_table.ht_table[idx]; 12077adb730bSGeorge Wilson buf_hash_table.ht_table[idx] = hdr; 1208dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1209fa9e4066Sahrens 1210fa9e4066Sahrens /* collect some hash table performance data */ 1211fa9e4066Sahrens if (i > 0) { 121244cb6abcSbmc ARCSTAT_BUMP(arcstat_hash_collisions); 1213fa9e4066Sahrens if (i == 1) 121444cb6abcSbmc ARCSTAT_BUMP(arcstat_hash_chains); 121544cb6abcSbmc 121644cb6abcSbmc ARCSTAT_MAX(arcstat_hash_chain_max, i); 1217fa9e4066Sahrens } 121844cb6abcSbmc 121944cb6abcSbmc ARCSTAT_BUMP(arcstat_hash_elements); 122044cb6abcSbmc ARCSTAT_MAXSTAT(arcstat_hash_elements); 1221fa9e4066Sahrens 1222fa9e4066Sahrens return (NULL); 1223fa9e4066Sahrens } 1224fa9e4066Sahrens 1225fa9e4066Sahrens static void 12267adb730bSGeorge Wilson buf_hash_remove(arc_buf_hdr_t *hdr) 1227fa9e4066Sahrens { 12287adb730bSGeorge Wilson arc_buf_hdr_t *fhdr, **hdrp; 12297adb730bSGeorge Wilson uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1230fa9e4066Sahrens 1231fa9e4066Sahrens ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 12327adb730bSGeorge Wilson ASSERT(HDR_IN_HASH_TABLE(hdr)); 1233fa9e4066Sahrens 12347adb730bSGeorge Wilson hdrp = &buf_hash_table.ht_table[idx]; 12357adb730bSGeorge Wilson while ((fhdr = *hdrp) != hdr) { 1236dcbf3bd6SGeorge Wilson ASSERT3P(fhdr, !=, NULL); 12377adb730bSGeorge Wilson hdrp = &fhdr->b_hash_next; 1238fa9e4066Sahrens } 12397adb730bSGeorge Wilson *hdrp = hdr->b_hash_next; 12407adb730bSGeorge Wilson hdr->b_hash_next = NULL; 1241dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1242fa9e4066Sahrens 1243fa9e4066Sahrens /* collect some hash table performance data */ 124444cb6abcSbmc ARCSTAT_BUMPDOWN(arcstat_hash_elements); 124544cb6abcSbmc 1246fa9e4066Sahrens if (buf_hash_table.ht_table[idx] && 1247fa9e4066Sahrens buf_hash_table.ht_table[idx]->b_hash_next == NULL) 124844cb6abcSbmc ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1249fa9e4066Sahrens } 1250fa9e4066Sahrens 1251fa9e4066Sahrens /* 1252fa9e4066Sahrens * Global data structures and functions for the buf kmem cache. 1253fa9e4066Sahrens */ 125489c86e32SChris Williamson static kmem_cache_t *hdr_full_cache; 125589c86e32SChris Williamson static kmem_cache_t *hdr_l2only_cache; 1256fa9e4066Sahrens static kmem_cache_t *buf_cache; 1257fa9e4066Sahrens 1258fa9e4066Sahrens static void 1259fa9e4066Sahrens buf_fini(void) 1260fa9e4066Sahrens { 1261fa9e4066Sahrens int i; 1262fa9e4066Sahrens 1263fa9e4066Sahrens kmem_free(buf_hash_table.ht_table, 1264fa9e4066Sahrens (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1265fa9e4066Sahrens for (i = 0; i < BUF_LOCKS; i++) 1266fa9e4066Sahrens mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 126789c86e32SChris Williamson kmem_cache_destroy(hdr_full_cache); 126889c86e32SChris Williamson kmem_cache_destroy(hdr_l2only_cache); 1269fa9e4066Sahrens kmem_cache_destroy(buf_cache); 1270fa9e4066Sahrens } 1271fa9e4066Sahrens 1272fa9e4066Sahrens /* 1273fa9e4066Sahrens * Constructor callback - called when the cache is empty 1274fa9e4066Sahrens * and a new buf is requested. 1275fa9e4066Sahrens */ 1276fa9e4066Sahrens /* ARGSUSED */ 1277fa9e4066Sahrens static int 127889c86e32SChris Williamson hdr_full_cons(void *vbuf, void *unused, int kmflag) 1279fa9e4066Sahrens { 12807adb730bSGeorge Wilson arc_buf_hdr_t *hdr = vbuf; 1281fa9e4066Sahrens 128289c86e32SChris Williamson bzero(hdr, HDR_FULL_SIZE); 128389c86e32SChris Williamson cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 128489c86e32SChris Williamson refcount_create(&hdr->b_l1hdr.b_refcnt); 128589c86e32SChris Williamson mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1286244781f1SPrakash Surya multilist_link_init(&hdr->b_l1hdr.b_arc_node); 128789c86e32SChris Williamson arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 128889c86e32SChris Williamson 128989c86e32SChris Williamson return (0); 129089c86e32SChris Williamson } 129189c86e32SChris Williamson 129289c86e32SChris Williamson /* ARGSUSED */ 129389c86e32SChris Williamson static int 129489c86e32SChris Williamson hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 129589c86e32SChris Williamson { 129689c86e32SChris Williamson arc_buf_hdr_t *hdr = vbuf; 129789c86e32SChris Williamson 129889c86e32SChris Williamson bzero(hdr, HDR_L2ONLY_SIZE); 129989c86e32SChris Williamson arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1300fa94a07fSbrendan 1301fa9e4066Sahrens return (0); 1302fa9e4066Sahrens } 1303fa9e4066Sahrens 13046f83844dSMark Maybee /* ARGSUSED */ 13056f83844dSMark Maybee static int 13066f83844dSMark Maybee buf_cons(void *vbuf, void *unused, int kmflag) 13076f83844dSMark Maybee { 13086f83844dSMark Maybee arc_buf_t *buf = vbuf; 13096f83844dSMark Maybee 13106f83844dSMark Maybee bzero(buf, sizeof (arc_buf_t)); 13113f9d6ad7SLin Ling mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 13125a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 13135a98e54bSBrendan Gregg - Sun Microsystems 13146f83844dSMark Maybee return (0); 13156f83844dSMark Maybee } 13166f83844dSMark Maybee 1317fa9e4066Sahrens /* 1318fa9e4066Sahrens * Destructor callback - called when a cached buf is 1319fa9e4066Sahrens * no longer required. 1320fa9e4066Sahrens */ 1321fa9e4066Sahrens /* ARGSUSED */ 1322fa9e4066Sahrens static void 132389c86e32SChris Williamson hdr_full_dest(void *vbuf, void *unused) 132489c86e32SChris Williamson { 132589c86e32SChris Williamson arc_buf_hdr_t *hdr = vbuf; 132689c86e32SChris Williamson 1327dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 132889c86e32SChris Williamson cv_destroy(&hdr->b_l1hdr.b_cv); 132989c86e32SChris Williamson refcount_destroy(&hdr->b_l1hdr.b_refcnt); 133089c86e32SChris Williamson mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1331244781f1SPrakash Surya ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 133289c86e32SChris Williamson arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 133389c86e32SChris Williamson } 133489c86e32SChris Williamson 133589c86e32SChris Williamson /* ARGSUSED */ 133689c86e32SChris Williamson static void 133789c86e32SChris Williamson hdr_l2only_dest(void *vbuf, void *unused) 1338fa9e4066Sahrens { 13397adb730bSGeorge Wilson arc_buf_hdr_t *hdr = vbuf; 1340fa9e4066Sahrens 1341dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 134289c86e32SChris Williamson arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1343fa9e4066Sahrens } 1344fa9e4066Sahrens 13456f83844dSMark Maybee /* ARGSUSED */ 13466f83844dSMark Maybee static void 13476f83844dSMark Maybee buf_dest(void *vbuf, void *unused) 13486f83844dSMark Maybee { 13496f83844dSMark Maybee arc_buf_t *buf = vbuf; 13506f83844dSMark Maybee 13513f9d6ad7SLin Ling mutex_destroy(&buf->b_evict_lock); 13525a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 13536f83844dSMark Maybee } 13546f83844dSMark Maybee 1355fa9e4066Sahrens /* 1356fa9e4066Sahrens * Reclaim callback -- invoked when memory is low. 1357fa9e4066Sahrens */ 1358fa9e4066Sahrens /* ARGSUSED */ 1359fa9e4066Sahrens static void 1360fa9e4066Sahrens hdr_recl(void *unused) 1361fa9e4066Sahrens { 1362fa9e4066Sahrens dprintf("hdr_recl called\n"); 136349e3519aSmaybee /* 136449e3519aSmaybee * umem calls the reclaim func when we destroy the buf cache, 136549e3519aSmaybee * which is after we do arc_fini(). 136649e3519aSmaybee */ 136749e3519aSmaybee if (!arc_dead) 1368244781f1SPrakash Surya cv_signal(&arc_reclaim_thread_cv); 1369fa9e4066Sahrens } 1370fa9e4066Sahrens 1371fa9e4066Sahrens static void 1372fa9e4066Sahrens buf_init(void) 1373fa9e4066Sahrens { 1374fa9e4066Sahrens uint64_t *ct; 1375ea8dc4b6Seschrock uint64_t hsize = 1ULL << 12; 1376fa9e4066Sahrens int i, j; 1377fa9e4066Sahrens 1378fa9e4066Sahrens /* 1379fa9e4066Sahrens * The hash table is big enough to fill all of physical memory 138063e911b6SMatthew Ahrens * with an average block size of zfs_arc_average_blocksize (default 8K). 138163e911b6SMatthew Ahrens * By default, the table will take up 138263e911b6SMatthew Ahrens * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1383fa9e4066Sahrens */ 138463e911b6SMatthew Ahrens while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE) 1385fa9e4066Sahrens hsize <<= 1; 1386ea8dc4b6Seschrock retry: 1387fa9e4066Sahrens buf_hash_table.ht_mask = hsize - 1; 1388ea8dc4b6Seschrock buf_hash_table.ht_table = 1389ea8dc4b6Seschrock kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1390ea8dc4b6Seschrock if (buf_hash_table.ht_table == NULL) { 1391ea8dc4b6Seschrock ASSERT(hsize > (1ULL << 8)); 1392ea8dc4b6Seschrock hsize >>= 1; 1393ea8dc4b6Seschrock goto retry; 1394ea8dc4b6Seschrock } 1395fa9e4066Sahrens 139689c86e32SChris Williamson hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 139789c86e32SChris Williamson 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 139889c86e32SChris Williamson hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 139989c86e32SChris Williamson HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 140089c86e32SChris Williamson NULL, NULL, 0); 1401fa9e4066Sahrens buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 14026f83844dSMark Maybee 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1403fa9e4066Sahrens 1404fa9e4066Sahrens for (i = 0; i < 256; i++) 1405fa9e4066Sahrens for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1406fa9e4066Sahrens *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1407fa9e4066Sahrens 1408fa9e4066Sahrens for (i = 0; i < BUF_LOCKS; i++) { 1409fa9e4066Sahrens mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1410fa9e4066Sahrens NULL, MUTEX_DEFAULT, NULL); 1411fa9e4066Sahrens } 1412fa9e4066Sahrens } 1413fa9e4066Sahrens 14145602294fSDan Kimmel /* 14155602294fSDan Kimmel * This is the size that the buf occupies in memory. If the buf is compressed, 14165602294fSDan Kimmel * it will correspond to the compressed size. You should use this method of 14175602294fSDan Kimmel * getting the buf size unless you explicitly need the logical size. 14185602294fSDan Kimmel */ 14195602294fSDan Kimmel int32_t 14205602294fSDan Kimmel arc_buf_size(arc_buf_t *buf) 14215602294fSDan Kimmel { 14225602294fSDan Kimmel return (ARC_BUF_COMPRESSED(buf) ? 14235602294fSDan Kimmel HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr)); 14245602294fSDan Kimmel } 14255602294fSDan Kimmel 14265602294fSDan Kimmel int32_t 14275602294fSDan Kimmel arc_buf_lsize(arc_buf_t *buf) 14285602294fSDan Kimmel { 14295602294fSDan Kimmel return (HDR_GET_LSIZE(buf->b_hdr)); 14305602294fSDan Kimmel } 14315602294fSDan Kimmel 14325602294fSDan Kimmel enum zio_compress 14335602294fSDan Kimmel arc_get_compression(arc_buf_t *buf) 14345602294fSDan Kimmel { 14355602294fSDan Kimmel return (ARC_BUF_COMPRESSED(buf) ? 14365602294fSDan Kimmel HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF); 14375602294fSDan Kimmel } 14385602294fSDan Kimmel 1439dcbf3bd6SGeorge Wilson #define ARC_MINTIME (hz>>4) /* 62 ms */ 1440244781f1SPrakash Surya 1441dcbf3bd6SGeorge Wilson static inline boolean_t 1442dcbf3bd6SGeorge Wilson arc_buf_is_shared(arc_buf_t *buf) 1443dcbf3bd6SGeorge Wilson { 1444dcbf3bd6SGeorge Wilson boolean_t shared = (buf->b_data != NULL && 1445770499e1SDan Kimmel buf->b_hdr->b_l1hdr.b_pabd != NULL && 1446770499e1SDan Kimmel abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) && 1447770499e1SDan Kimmel buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd)); 1448dcbf3bd6SGeorge Wilson IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); 14495602294fSDan Kimmel IMPLY(shared, ARC_BUF_SHARED(buf)); 14505602294fSDan Kimmel IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf)); 14515602294fSDan Kimmel 14525602294fSDan Kimmel /* 14535602294fSDan Kimmel * It would be nice to assert arc_can_share() too, but the "hdr isn't 14545602294fSDan Kimmel * already being shared" requirement prevents us from doing that. 14555602294fSDan Kimmel */ 14565602294fSDan Kimmel 1457dcbf3bd6SGeorge Wilson return (shared); 1458dcbf3bd6SGeorge Wilson } 1459c546f36aSArne Jansen 14605602294fSDan Kimmel /* 14615602294fSDan Kimmel * Free the checksum associated with this header. If there is no checksum, this 14625602294fSDan Kimmel * is a no-op. 14635602294fSDan Kimmel */ 1464dcbf3bd6SGeorge Wilson static inline void 1465dcbf3bd6SGeorge Wilson arc_cksum_free(arc_buf_hdr_t *hdr) 1466dcbf3bd6SGeorge Wilson { 1467dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1468dcbf3bd6SGeorge Wilson mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1469dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1470dcbf3bd6SGeorge Wilson kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); 1471dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_freeze_cksum = NULL; 147289c86e32SChris Williamson } 1473dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 147489c86e32SChris Williamson } 147589c86e32SChris Williamson 14765602294fSDan Kimmel /* 14775602294fSDan Kimmel * Return true iff at least one of the bufs on hdr is not compressed. 14785602294fSDan Kimmel */ 14795602294fSDan Kimmel static boolean_t 14805602294fSDan Kimmel arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr) 14815602294fSDan Kimmel { 14825602294fSDan Kimmel for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) { 14835602294fSDan Kimmel if (!ARC_BUF_COMPRESSED(b)) { 14845602294fSDan Kimmel return (B_TRUE); 14855602294fSDan Kimmel } 14865602294fSDan Kimmel } 14875602294fSDan Kimmel return (B_FALSE); 14885602294fSDan Kimmel } 14895602294fSDan Kimmel 14905602294fSDan Kimmel /* 14915602294fSDan Kimmel * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data 14925602294fSDan Kimmel * matches the checksum that is stored in the hdr. If there is no checksum, 14935602294fSDan Kimmel * or if the buf is compressed, this is a no-op. 14945602294fSDan Kimmel */ 14956b4acc8bSahrens static void 14966b4acc8bSahrens arc_cksum_verify(arc_buf_t *buf) 14976b4acc8bSahrens { 1498dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 14996b4acc8bSahrens zio_cksum_t zc; 15006b4acc8bSahrens 1501cc60fd72Sahrens if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 15026b4acc8bSahrens return; 15036b4acc8bSahrens 15045602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 15055602294fSDan Kimmel ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 15065602294fSDan Kimmel arc_hdr_has_uncompressed_buf(hdr)); 15075602294fSDan Kimmel return; 15085602294fSDan Kimmel } 15095602294fSDan Kimmel 1510dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1511dcbf3bd6SGeorge Wilson 1512dcbf3bd6SGeorge Wilson mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1513dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { 1514dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 15156b4acc8bSahrens return; 15166b4acc8bSahrens } 15175602294fSDan Kimmel 15185602294fSDan Kimmel fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc); 1519dcbf3bd6SGeorge Wilson if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) 15206b4acc8bSahrens panic("buffer modified while frozen!"); 1521dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 15226b4acc8bSahrens } 15236b4acc8bSahrens 1524dcbf3bd6SGeorge Wilson static boolean_t 1525dcbf3bd6SGeorge Wilson arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) 1526fa94a07fSbrendan { 1527dcbf3bd6SGeorge Wilson enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp); 1528dcbf3bd6SGeorge Wilson boolean_t valid_cksum; 1529fa94a07fSbrendan 1530dcbf3bd6SGeorge Wilson ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); 1531dcbf3bd6SGeorge Wilson VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); 1532dcbf3bd6SGeorge Wilson 1533dcbf3bd6SGeorge Wilson /* 1534dcbf3bd6SGeorge Wilson * We rely on the blkptr's checksum to determine if the block 1535dcbf3bd6SGeorge Wilson * is valid or not. When compressed arc is enabled, the l2arc 1536dcbf3bd6SGeorge Wilson * writes the block to the l2arc just as it appears in the pool. 1537dcbf3bd6SGeorge Wilson * This allows us to use the blkptr's checksum to validate the 1538dcbf3bd6SGeorge Wilson * data that we just read off of the l2arc without having to store 1539dcbf3bd6SGeorge Wilson * a separate checksum in the arc_buf_hdr_t. However, if compressed 1540dcbf3bd6SGeorge Wilson * arc is disabled, then the data written to the l2arc is always 1541dcbf3bd6SGeorge Wilson * uncompressed and won't match the block as it exists in the main 1542dcbf3bd6SGeorge Wilson * pool. When this is the case, we must first compress it if it is 1543dcbf3bd6SGeorge Wilson * compressed on the main pool before we can validate the checksum. 1544dcbf3bd6SGeorge Wilson */ 1545dcbf3bd6SGeorge Wilson if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) { 1546dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1547dcbf3bd6SGeorge Wilson uint64_t lsize = HDR_GET_LSIZE(hdr); 1548dcbf3bd6SGeorge Wilson uint64_t csize; 1549dcbf3bd6SGeorge Wilson 1550*01a059eeSRoman Strashkin abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE); 1551*01a059eeSRoman Strashkin csize = zio_compress_data(compress, zio->io_abd, 1552*01a059eeSRoman Strashkin abd_to_buf(cdata), lsize); 1553770499e1SDan Kimmel 1554dcbf3bd6SGeorge Wilson ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr)); 1555dcbf3bd6SGeorge Wilson if (csize < HDR_GET_PSIZE(hdr)) { 1556dcbf3bd6SGeorge Wilson /* 1557dcbf3bd6SGeorge Wilson * Compressed blocks are always a multiple of the 1558dcbf3bd6SGeorge Wilson * smallest ashift in the pool. Ideally, we would 1559dcbf3bd6SGeorge Wilson * like to round up the csize to the next 1560dcbf3bd6SGeorge Wilson * spa_min_ashift but that value may have changed 1561dcbf3bd6SGeorge Wilson * since the block was last written. Instead, 1562dcbf3bd6SGeorge Wilson * we rely on the fact that the hdr's psize 1563dcbf3bd6SGeorge Wilson * was set to the psize of the block when it was 1564dcbf3bd6SGeorge Wilson * last written. We set the csize to that value 1565dcbf3bd6SGeorge Wilson * and zero out any part that should not contain 1566dcbf3bd6SGeorge Wilson * data. 1567dcbf3bd6SGeorge Wilson */ 1568*01a059eeSRoman Strashkin abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize); 1569dcbf3bd6SGeorge Wilson csize = HDR_GET_PSIZE(hdr); 1570dcbf3bd6SGeorge Wilson } 1571*01a059eeSRoman Strashkin zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL); 1572dcbf3bd6SGeorge Wilson } 1573fa94a07fSbrendan 1574dcbf3bd6SGeorge Wilson /* 1575dcbf3bd6SGeorge Wilson * Block pointers always store the checksum for the logical data. 1576dcbf3bd6SGeorge Wilson * If the block pointer has the gang bit set, then the checksum 1577dcbf3bd6SGeorge Wilson * it represents is for the reconstituted data and not for an 1578dcbf3bd6SGeorge Wilson * individual gang member. The zio pipeline, however, must be able to 1579dcbf3bd6SGeorge Wilson * determine the checksum of each of the gang constituents so it 1580dcbf3bd6SGeorge Wilson * treats the checksum comparison differently than what we need 1581dcbf3bd6SGeorge Wilson * for l2arc blocks. This prevents us from using the 1582dcbf3bd6SGeorge Wilson * zio_checksum_error() interface directly. Instead we must call the 1583dcbf3bd6SGeorge Wilson * zio_checksum_error_impl() so that we can ensure the checksum is 1584dcbf3bd6SGeorge Wilson * generated using the correct checksum algorithm and accounts for the 1585dcbf3bd6SGeorge Wilson * logical I/O size and not just a gang fragment. 1586dcbf3bd6SGeorge Wilson */ 1587dcbf3bd6SGeorge Wilson valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp, 1588770499e1SDan Kimmel BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size, 1589dcbf3bd6SGeorge Wilson zio->io_offset, NULL) == 0); 1590dcbf3bd6SGeorge Wilson zio_pop_transforms(zio); 1591dcbf3bd6SGeorge Wilson return (valid_cksum); 1592fa94a07fSbrendan } 1593fa94a07fSbrendan 15945602294fSDan Kimmel /* 15955602294fSDan Kimmel * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a 15965602294fSDan Kimmel * checksum and attaches it to the buf's hdr so that we can ensure that the buf 15975602294fSDan Kimmel * isn't modified later on. If buf is compressed or there is already a checksum 15985602294fSDan Kimmel * on the hdr, this is a no-op (we only checksum uncompressed bufs). 15995602294fSDan Kimmel */ 16006b4acc8bSahrens static void 1601dcbf3bd6SGeorge Wilson arc_cksum_compute(arc_buf_t *buf) 16026b4acc8bSahrens { 1603dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 1604dcbf3bd6SGeorge Wilson 1605dcbf3bd6SGeorge Wilson if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 16066b4acc8bSahrens return; 16076b4acc8bSahrens 1608dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 16095602294fSDan Kimmel 161089c86e32SChris Williamson mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1611dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 16125602294fSDan Kimmel ASSERT(arc_hdr_has_uncompressed_buf(hdr)); 16135602294fSDan Kimmel mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 16145602294fSDan Kimmel return; 16155602294fSDan Kimmel } else if (ARC_BUF_COMPRESSED(buf)) { 1616dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 16176b4acc8bSahrens return; 16186b4acc8bSahrens } 16195602294fSDan Kimmel 16205602294fSDan Kimmel ASSERT(!ARC_BUF_COMPRESSED(buf)); 1621dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), 1622dcbf3bd6SGeorge Wilson KM_SLEEP); 16235602294fSDan Kimmel fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, 1624dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_freeze_cksum); 1625dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1626cd1c8b85SMatthew Ahrens arc_buf_watch(buf); 1627cd1c8b85SMatthew Ahrens } 1628cd1c8b85SMatthew Ahrens 1629cd1c8b85SMatthew Ahrens #ifndef _KERNEL 1630cd1c8b85SMatthew Ahrens typedef struct procctl { 1631cd1c8b85SMatthew Ahrens long cmd; 1632cd1c8b85SMatthew Ahrens prwatch_t prwatch; 1633cd1c8b85SMatthew Ahrens } procctl_t; 1634cd1c8b85SMatthew Ahrens #endif 1635cd1c8b85SMatthew Ahrens 1636cd1c8b85SMatthew Ahrens /* ARGSUSED */ 1637cd1c8b85SMatthew Ahrens static void 1638cd1c8b85SMatthew Ahrens arc_buf_unwatch(arc_buf_t *buf) 1639cd1c8b85SMatthew Ahrens { 1640cd1c8b85SMatthew Ahrens #ifndef _KERNEL 1641cd1c8b85SMatthew Ahrens if (arc_watch) { 1642cd1c8b85SMatthew Ahrens int result; 1643cd1c8b85SMatthew Ahrens procctl_t ctl; 1644cd1c8b85SMatthew Ahrens ctl.cmd = PCWATCH; 1645cd1c8b85SMatthew Ahrens ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1646cd1c8b85SMatthew Ahrens ctl.prwatch.pr_size = 0; 1647cd1c8b85SMatthew Ahrens ctl.prwatch.pr_wflags = 0; 1648cd1c8b85SMatthew Ahrens result = write(arc_procfd, &ctl, sizeof (ctl)); 1649cd1c8b85SMatthew Ahrens ASSERT3U(result, ==, sizeof (ctl)); 1650cd1c8b85SMatthew Ahrens } 1651cd1c8b85SMatthew Ahrens #endif 1652cd1c8b85SMatthew Ahrens } 1653cd1c8b85SMatthew Ahrens 1654cd1c8b85SMatthew Ahrens /* ARGSUSED */ 1655cd1c8b85SMatthew Ahrens static void 1656cd1c8b85SMatthew Ahrens arc_buf_watch(arc_buf_t *buf) 1657cd1c8b85SMatthew Ahrens { 1658cd1c8b85SMatthew Ahrens #ifndef _KERNEL 1659cd1c8b85SMatthew Ahrens if (arc_watch) { 1660cd1c8b85SMatthew Ahrens int result; 1661cd1c8b85SMatthew Ahrens procctl_t ctl; 1662cd1c8b85SMatthew Ahrens ctl.cmd = PCWATCH; 1663cd1c8b85SMatthew Ahrens ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 16645602294fSDan Kimmel ctl.prwatch.pr_size = arc_buf_size(buf); 1665cd1c8b85SMatthew Ahrens ctl.prwatch.pr_wflags = WA_WRITE; 1666cd1c8b85SMatthew Ahrens result = write(arc_procfd, &ctl, sizeof (ctl)); 1667cd1c8b85SMatthew Ahrens ASSERT3U(result, ==, sizeof (ctl)); 1668cd1c8b85SMatthew Ahrens } 1669cd1c8b85SMatthew Ahrens #endif 16706b4acc8bSahrens } 16716b4acc8bSahrens 167289c86e32SChris Williamson static arc_buf_contents_t 167389c86e32SChris Williamson arc_buf_type(arc_buf_hdr_t *hdr) 167489c86e32SChris Williamson { 1675dcbf3bd6SGeorge Wilson arc_buf_contents_t type; 167689c86e32SChris Williamson if (HDR_ISTYPE_METADATA(hdr)) { 1677dcbf3bd6SGeorge Wilson type = ARC_BUFC_METADATA; 167889c86e32SChris Williamson } else { 1679dcbf3bd6SGeorge Wilson type = ARC_BUFC_DATA; 168089c86e32SChris Williamson } 1681dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 1682dcbf3bd6SGeorge Wilson return (type); 168389c86e32SChris Williamson } 168489c86e32SChris Williamson 16855602294fSDan Kimmel boolean_t 16865602294fSDan Kimmel arc_is_metadata(arc_buf_t *buf) 16875602294fSDan Kimmel { 16885602294fSDan Kimmel return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0); 16895602294fSDan Kimmel } 16905602294fSDan Kimmel 169189c86e32SChris Williamson static uint32_t 169289c86e32SChris Williamson arc_bufc_to_flags(arc_buf_contents_t type) 169389c86e32SChris Williamson { 169489c86e32SChris Williamson switch (type) { 169589c86e32SChris Williamson case ARC_BUFC_DATA: 169689c86e32SChris Williamson /* metadata field is 0 if buffer contains normal data */ 169789c86e32SChris Williamson return (0); 169889c86e32SChris Williamson case ARC_BUFC_METADATA: 169989c86e32SChris Williamson return (ARC_FLAG_BUFC_METADATA); 170089c86e32SChris Williamson default: 170189c86e32SChris Williamson break; 170289c86e32SChris Williamson } 170389c86e32SChris Williamson panic("undefined ARC buffer type!"); 170489c86e32SChris Williamson return ((uint32_t)-1); 170589c86e32SChris Williamson } 170689c86e32SChris Williamson 17076b4acc8bSahrens void 17086b4acc8bSahrens arc_buf_thaw(arc_buf_t *buf) 17096b4acc8bSahrens { 1710dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 1711dcbf3bd6SGeorge Wilson 17125602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 17135602294fSDan Kimmel ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 17145602294fSDan Kimmel 17155602294fSDan Kimmel arc_cksum_verify(buf); 17165602294fSDan Kimmel 17175602294fSDan Kimmel /* 17185602294fSDan Kimmel * Compressed buffers do not manipulate the b_freeze_cksum or 17195602294fSDan Kimmel * allocate b_thawed. 17205602294fSDan Kimmel */ 17215602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 17225602294fSDan Kimmel ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 17235602294fSDan Kimmel arc_hdr_has_uncompressed_buf(hdr)); 17245602294fSDan Kimmel return; 1725fa94a07fSbrendan } 17266b4acc8bSahrens 1727dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1728dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 17293f9d6ad7SLin Ling 1730dcbf3bd6SGeorge Wilson mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 173189c86e32SChris Williamson #ifdef ZFS_DEBUG 17323f9d6ad7SLin Ling if (zfs_flags & ZFS_DEBUG_MODIFY) { 1733dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_thawed != NULL) 1734dcbf3bd6SGeorge Wilson kmem_free(hdr->b_l1hdr.b_thawed, 1); 1735dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 17363f9d6ad7SLin Ling } 173789c86e32SChris Williamson #endif 17383f9d6ad7SLin Ling 1739dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1740cd1c8b85SMatthew Ahrens 1741cd1c8b85SMatthew Ahrens arc_buf_unwatch(buf); 17426b4acc8bSahrens } 17436b4acc8bSahrens 17446b4acc8bSahrens void 17456b4acc8bSahrens arc_buf_freeze(arc_buf_t *buf) 17466b4acc8bSahrens { 1747dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 17483f9d6ad7SLin Ling kmutex_t *hash_lock; 17493f9d6ad7SLin Ling 1750cc60fd72Sahrens if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1751cc60fd72Sahrens return; 1752cc60fd72Sahrens 17535602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 17545602294fSDan Kimmel ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 17555602294fSDan Kimmel arc_hdr_has_uncompressed_buf(hdr)); 17565602294fSDan Kimmel return; 17575602294fSDan Kimmel } 17585602294fSDan Kimmel 1759dcbf3bd6SGeorge Wilson hash_lock = HDR_LOCK(hdr); 17603f9d6ad7SLin Ling mutex_enter(hash_lock); 17613f9d6ad7SLin Ling 1762dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1763dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL || 1764dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_state == arc_anon); 1765dcbf3bd6SGeorge Wilson arc_cksum_compute(buf); 17663f9d6ad7SLin Ling mutex_exit(hash_lock); 17676b4acc8bSahrens } 17686b4acc8bSahrens 1769dcbf3bd6SGeorge Wilson /* 1770dcbf3bd6SGeorge Wilson * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, 1771dcbf3bd6SGeorge Wilson * the following functions should be used to ensure that the flags are 1772dcbf3bd6SGeorge Wilson * updated in a thread-safe way. When manipulating the flags either 1773dcbf3bd6SGeorge Wilson * the hash_lock must be held or the hdr must be undiscoverable. This 1774dcbf3bd6SGeorge Wilson * ensures that we're not racing with any other threads when updating 1775dcbf3bd6SGeorge Wilson * the flags. 1776dcbf3bd6SGeorge Wilson */ 1777dcbf3bd6SGeorge Wilson static inline void 1778dcbf3bd6SGeorge Wilson arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1779dcbf3bd6SGeorge Wilson { 1780dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1781dcbf3bd6SGeorge Wilson hdr->b_flags |= flags; 1782dcbf3bd6SGeorge Wilson } 1783dcbf3bd6SGeorge Wilson 1784dcbf3bd6SGeorge Wilson static inline void 1785dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1786dcbf3bd6SGeorge Wilson { 1787dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1788dcbf3bd6SGeorge Wilson hdr->b_flags &= ~flags; 1789dcbf3bd6SGeorge Wilson } 1790dcbf3bd6SGeorge Wilson 1791dcbf3bd6SGeorge Wilson /* 1792dcbf3bd6SGeorge Wilson * Setting the compression bits in the arc_buf_hdr_t's b_flags is 1793dcbf3bd6SGeorge Wilson * done in a special way since we have to clear and set bits 1794dcbf3bd6SGeorge Wilson * at the same time. Consumers that wish to set the compression bits 1795dcbf3bd6SGeorge Wilson * must use this function to ensure that the flags are updated in 1796dcbf3bd6SGeorge Wilson * thread-safe manner. 1797dcbf3bd6SGeorge Wilson */ 1798fa9e4066Sahrens static void 1799dcbf3bd6SGeorge Wilson arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) 1800fa9e4066Sahrens { 1801dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1802fa9e4066Sahrens 1803dcbf3bd6SGeorge Wilson /* 1804dcbf3bd6SGeorge Wilson * Holes and embedded blocks will always have a psize = 0 so 1805dcbf3bd6SGeorge Wilson * we ignore the compression of the blkptr and set the 1806dcbf3bd6SGeorge Wilson * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF. 1807dcbf3bd6SGeorge Wilson * Holes and embedded blocks remain anonymous so we don't 1808dcbf3bd6SGeorge Wilson * want to uncompress them. Mark them as uncompressed. 1809dcbf3bd6SGeorge Wilson */ 1810dcbf3bd6SGeorge Wilson if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { 1811dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1812dcbf3bd6SGeorge Wilson HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 1813dcbf3bd6SGeorge Wilson ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); 1814dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1815dcbf3bd6SGeorge Wilson } else { 1816dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1817dcbf3bd6SGeorge Wilson HDR_SET_COMPRESS(hdr, cmp); 1818dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); 1819dcbf3bd6SGeorge Wilson ASSERT(HDR_COMPRESSION_ENABLED(hdr)); 1820dcbf3bd6SGeorge Wilson } 1821dcbf3bd6SGeorge Wilson } 1822244781f1SPrakash Surya 18235602294fSDan Kimmel /* 18245602294fSDan Kimmel * Looks for another buf on the same hdr which has the data decompressed, copies 18255602294fSDan Kimmel * from it, and returns true. If no such buf exists, returns false. 18265602294fSDan Kimmel */ 18275602294fSDan Kimmel static boolean_t 18285602294fSDan Kimmel arc_buf_try_copy_decompressed_data(arc_buf_t *buf) 18295602294fSDan Kimmel { 18305602294fSDan Kimmel arc_buf_hdr_t *hdr = buf->b_hdr; 18315602294fSDan Kimmel boolean_t copied = B_FALSE; 18325602294fSDan Kimmel 18335602294fSDan Kimmel ASSERT(HDR_HAS_L1HDR(hdr)); 18345602294fSDan Kimmel ASSERT3P(buf->b_data, !=, NULL); 18355602294fSDan Kimmel ASSERT(!ARC_BUF_COMPRESSED(buf)); 18365602294fSDan Kimmel 18375602294fSDan Kimmel for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL; 18385602294fSDan Kimmel from = from->b_next) { 18395602294fSDan Kimmel /* can't use our own data buffer */ 18405602294fSDan Kimmel if (from == buf) { 18415602294fSDan Kimmel continue; 18425602294fSDan Kimmel } 18435602294fSDan Kimmel 18445602294fSDan Kimmel if (!ARC_BUF_COMPRESSED(from)) { 18455602294fSDan Kimmel bcopy(from->b_data, buf->b_data, arc_buf_size(buf)); 18465602294fSDan Kimmel copied = B_TRUE; 18475602294fSDan Kimmel break; 18485602294fSDan Kimmel } 18495602294fSDan Kimmel } 18505602294fSDan Kimmel 18515602294fSDan Kimmel /* 18525602294fSDan Kimmel * There were no decompressed bufs, so there should not be a 18535602294fSDan Kimmel * checksum on the hdr either. 18545602294fSDan Kimmel */ 18555602294fSDan Kimmel EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL); 18565602294fSDan Kimmel 18575602294fSDan Kimmel return (copied); 18585602294fSDan Kimmel } 18595602294fSDan Kimmel 18605602294fSDan Kimmel /* 18615602294fSDan Kimmel * Given a buf that has a data buffer attached to it, this function will 18625602294fSDan Kimmel * efficiently fill the buf with data of the specified compression setting from 18635602294fSDan Kimmel * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr 18645602294fSDan Kimmel * are already sharing a data buf, no copy is performed. 18655602294fSDan Kimmel * 18665602294fSDan Kimmel * If the buf is marked as compressed but uncompressed data was requested, this 18675602294fSDan Kimmel * will allocate a new data buffer for the buf, remove that flag, and fill the 18685602294fSDan Kimmel * buf with uncompressed data. You can't request a compressed buf on a hdr with 18695602294fSDan Kimmel * uncompressed data, and (since we haven't added support for it yet) if you 18705602294fSDan Kimmel * want compressed data your buf must already be marked as compressed and have 18715602294fSDan Kimmel * the correct-sized data buffer. 18725602294fSDan Kimmel */ 1873dcbf3bd6SGeorge Wilson static int 18745602294fSDan Kimmel arc_buf_fill(arc_buf_t *buf, boolean_t compressed) 1875dcbf3bd6SGeorge Wilson { 1876dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 18775602294fSDan Kimmel boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 1878dcbf3bd6SGeorge Wilson dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; 187989c86e32SChris Williamson 18805602294fSDan Kimmel ASSERT3P(buf->b_data, !=, NULL); 18815602294fSDan Kimmel IMPLY(compressed, hdr_compressed); 18825602294fSDan Kimmel IMPLY(compressed, ARC_BUF_COMPRESSED(buf)); 18835602294fSDan Kimmel 18845602294fSDan Kimmel if (hdr_compressed == compressed) { 18855602294fSDan Kimmel if (!arc_buf_is_shared(buf)) { 1886770499e1SDan Kimmel abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd, 18875602294fSDan Kimmel arc_buf_size(buf)); 18885602294fSDan Kimmel } 1889dcbf3bd6SGeorge Wilson } else { 18905602294fSDan Kimmel ASSERT(hdr_compressed); 18915602294fSDan Kimmel ASSERT(!compressed); 1892dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); 18935602294fSDan Kimmel 18945602294fSDan Kimmel /* 18955602294fSDan Kimmel * If the buf is sharing its data with the hdr, unlink it and 18965602294fSDan Kimmel * allocate a new data buffer for the buf. 18975602294fSDan Kimmel */ 18985602294fSDan Kimmel if (arc_buf_is_shared(buf)) { 18995602294fSDan Kimmel ASSERT(ARC_BUF_COMPRESSED(buf)); 19005602294fSDan Kimmel 19015602294fSDan Kimmel /* We need to give the buf it's own b_data */ 19025602294fSDan Kimmel buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 19035602294fSDan Kimmel buf->b_data = 19045602294fSDan Kimmel arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 19055602294fSDan Kimmel arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 19065602294fSDan Kimmel 19075602294fSDan Kimmel /* Previously overhead was 0; just add new overhead */ 19085602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 19095602294fSDan Kimmel } else if (ARC_BUF_COMPRESSED(buf)) { 19105602294fSDan Kimmel /* We need to reallocate the buf's b_data */ 19115602294fSDan Kimmel arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr), 19125602294fSDan Kimmel buf); 19135602294fSDan Kimmel buf->b_data = 19145602294fSDan Kimmel arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 19155602294fSDan Kimmel 19165602294fSDan Kimmel /* We increased the size of b_data; update overhead */ 19175602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, 19185602294fSDan Kimmel HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr)); 19195602294fSDan Kimmel } 19205602294fSDan Kimmel 19215602294fSDan Kimmel /* 19225602294fSDan Kimmel * Regardless of the buf's previous compression settings, it 19235602294fSDan Kimmel * should not be compressed at the end of this function. 19245602294fSDan Kimmel */ 19255602294fSDan Kimmel buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; 19265602294fSDan Kimmel 19275602294fSDan Kimmel /* 19285602294fSDan Kimmel * Try copying the data from another buf which already has a 19295602294fSDan Kimmel * decompressed version. If that's not possible, it's time to 19305602294fSDan Kimmel * bite the bullet and decompress the data from the hdr. 19315602294fSDan Kimmel */ 19325602294fSDan Kimmel if (arc_buf_try_copy_decompressed_data(buf)) { 19335602294fSDan Kimmel /* Skip byteswapping and checksumming (already done) */ 19345602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL); 19355602294fSDan Kimmel return (0); 19365602294fSDan Kimmel } else { 19375602294fSDan Kimmel int error = zio_decompress_data(HDR_GET_COMPRESS(hdr), 1938770499e1SDan Kimmel hdr->b_l1hdr.b_pabd, buf->b_data, 19395602294fSDan Kimmel HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 19405602294fSDan Kimmel 19415602294fSDan Kimmel /* 19425602294fSDan Kimmel * Absent hardware errors or software bugs, this should 19435602294fSDan Kimmel * be impossible, but log it anyway so we can debug it. 19445602294fSDan Kimmel */ 19455602294fSDan Kimmel if (error != 0) { 19465602294fSDan Kimmel zfs_dbgmsg( 19475602294fSDan Kimmel "hdr %p, compress %d, psize %d, lsize %d", 19485602294fSDan Kimmel hdr, HDR_GET_COMPRESS(hdr), 19495602294fSDan Kimmel HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 19505602294fSDan Kimmel return (SET_ERROR(EIO)); 19515602294fSDan Kimmel } 1952ea8dc4b6Seschrock } 1953fa9e4066Sahrens } 19545602294fSDan Kimmel 19555602294fSDan Kimmel /* Byteswap the buf's data if necessary */ 1956dcbf3bd6SGeorge Wilson if (bswap != DMU_BSWAP_NUMFUNCS) { 1957dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(hdr)); 1958dcbf3bd6SGeorge Wilson ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); 1959dcbf3bd6SGeorge Wilson dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); 1960dcbf3bd6SGeorge Wilson } 19615602294fSDan Kimmel 19625602294fSDan Kimmel /* Compute the hdr's checksum if necessary */ 1963dcbf3bd6SGeorge Wilson arc_cksum_compute(buf); 19645602294fSDan Kimmel 1965dcbf3bd6SGeorge Wilson return (0); 1966fa9e4066Sahrens } 1967fa9e4066Sahrens 19685602294fSDan Kimmel int 19695602294fSDan Kimmel arc_decompress(arc_buf_t *buf) 19705602294fSDan Kimmel { 19715602294fSDan Kimmel return (arc_buf_fill(buf, B_FALSE)); 19725602294fSDan Kimmel } 19735602294fSDan Kimmel 1974dcbf3bd6SGeorge Wilson /* 1975770499e1SDan Kimmel * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t. 1976dcbf3bd6SGeorge Wilson */ 1977dcbf3bd6SGeorge Wilson static uint64_t 1978dcbf3bd6SGeorge Wilson arc_hdr_size(arc_buf_hdr_t *hdr) 1979fa9e4066Sahrens { 1980dcbf3bd6SGeorge Wilson uint64_t size; 1981fa9e4066Sahrens 1982dcbf3bd6SGeorge Wilson if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 1983dcbf3bd6SGeorge Wilson HDR_GET_PSIZE(hdr) > 0) { 1984dcbf3bd6SGeorge Wilson size = HDR_GET_PSIZE(hdr); 1985dcbf3bd6SGeorge Wilson } else { 1986dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); 1987dcbf3bd6SGeorge Wilson size = HDR_GET_LSIZE(hdr); 1988dcbf3bd6SGeorge Wilson } 1989dcbf3bd6SGeorge Wilson return (size); 1990dcbf3bd6SGeorge Wilson } 1991fa9e4066Sahrens 1992dcbf3bd6SGeorge Wilson /* 1993dcbf3bd6SGeorge Wilson * Increment the amount of evictable space in the arc_state_t's refcount. 1994dcbf3bd6SGeorge Wilson * We account for the space used by the hdr and the arc buf individually 1995dcbf3bd6SGeorge Wilson * so that we can add and remove them from the refcount individually. 1996dcbf3bd6SGeorge Wilson */ 1997dcbf3bd6SGeorge Wilson static void 1998dcbf3bd6SGeorge Wilson arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) 1999dcbf3bd6SGeorge Wilson { 2000dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 2001244781f1SPrakash Surya 2002dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 20030e8c6158Smaybee 2004dcbf3bd6SGeorge Wilson if (GHOST_STATE(state)) { 2005dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 2006dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2007770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 20085602294fSDan Kimmel (void) refcount_add_many(&state->arcs_esize[type], 20095602294fSDan Kimmel HDR_GET_LSIZE(hdr), hdr); 2010dcbf3bd6SGeorge Wilson return; 2011dcbf3bd6SGeorge Wilson } 2012dcbf3bd6SGeorge Wilson 2013dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(state)); 2014770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 2015dcbf3bd6SGeorge Wilson (void) refcount_add_many(&state->arcs_esize[type], 2016dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2017dcbf3bd6SGeorge Wilson } 2018dcbf3bd6SGeorge Wilson for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2019dcbf3bd6SGeorge Wilson buf = buf->b_next) { 20205602294fSDan Kimmel if (arc_buf_is_shared(buf)) 2021dcbf3bd6SGeorge Wilson continue; 20225602294fSDan Kimmel (void) refcount_add_many(&state->arcs_esize[type], 20235602294fSDan Kimmel arc_buf_size(buf), buf); 2024fa9e4066Sahrens } 2025fa9e4066Sahrens } 2026fa9e4066Sahrens 2027fa9e4066Sahrens /* 2028dcbf3bd6SGeorge Wilson * Decrement the amount of evictable space in the arc_state_t's refcount. 2029dcbf3bd6SGeorge Wilson * We account for the space used by the hdr and the arc buf individually 2030dcbf3bd6SGeorge Wilson * so that we can add and remove them from the refcount individually. 2031fa9e4066Sahrens */ 2032fa9e4066Sahrens static void 20335602294fSDan Kimmel arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) 2034fa9e4066Sahrens { 2035dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 2036dcbf3bd6SGeorge Wilson 2037dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2038dcbf3bd6SGeorge Wilson 2039dcbf3bd6SGeorge Wilson if (GHOST_STATE(state)) { 2040dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 2041dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2042770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2043dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 20445602294fSDan Kimmel HDR_GET_LSIZE(hdr), hdr); 2045dcbf3bd6SGeorge Wilson return; 2046dcbf3bd6SGeorge Wilson } 2047dcbf3bd6SGeorge Wilson 2048dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(state)); 2049770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 2050dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 2051dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2052dcbf3bd6SGeorge Wilson } 2053dcbf3bd6SGeorge Wilson for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2054dcbf3bd6SGeorge Wilson buf = buf->b_next) { 20555602294fSDan Kimmel if (arc_buf_is_shared(buf)) 2056dcbf3bd6SGeorge Wilson continue; 2057dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 20585602294fSDan Kimmel arc_buf_size(buf), buf); 2059dcbf3bd6SGeorge Wilson } 2060dcbf3bd6SGeorge Wilson } 2061dcbf3bd6SGeorge Wilson 2062dcbf3bd6SGeorge Wilson /* 2063dcbf3bd6SGeorge Wilson * Add a reference to this hdr indicating that someone is actively 2064dcbf3bd6SGeorge Wilson * referencing that memory. When the refcount transitions from 0 to 1, 2065dcbf3bd6SGeorge Wilson * we remove it from the respective arc_state_t list to indicate that 2066dcbf3bd6SGeorge Wilson * it is not evictable. 2067dcbf3bd6SGeorge Wilson */ 2068dcbf3bd6SGeorge Wilson static void 2069dcbf3bd6SGeorge Wilson add_reference(arc_buf_hdr_t *hdr, void *tag) 2070dcbf3bd6SGeorge Wilson { 2071dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2072dcbf3bd6SGeorge Wilson if (!MUTEX_HELD(HDR_LOCK(hdr))) { 2073dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_state == arc_anon); 2074dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2075dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2076dcbf3bd6SGeorge Wilson } 2077dcbf3bd6SGeorge Wilson 2078dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2079dcbf3bd6SGeorge Wilson 2080dcbf3bd6SGeorge Wilson if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 2081dcbf3bd6SGeorge Wilson (state != arc_anon)) { 2082dcbf3bd6SGeorge Wilson /* We don't use the L2-only state list. */ 2083dcbf3bd6SGeorge Wilson if (state != arc_l2c_only) { 208494c2d0ebSMatthew Ahrens multilist_remove(state->arcs_list[arc_buf_type(hdr)], 2085dcbf3bd6SGeorge Wilson hdr); 20865602294fSDan Kimmel arc_evictable_space_decrement(hdr, state); 2087dcbf3bd6SGeorge Wilson } 2088dcbf3bd6SGeorge Wilson /* remove the prefetch flag if we get a reference */ 2089dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 2090dcbf3bd6SGeorge Wilson } 2091dcbf3bd6SGeorge Wilson } 2092dcbf3bd6SGeorge Wilson 2093dcbf3bd6SGeorge Wilson /* 2094dcbf3bd6SGeorge Wilson * Remove a reference from this hdr. When the reference transitions from 2095dcbf3bd6SGeorge Wilson * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's 2096dcbf3bd6SGeorge Wilson * list making it eligible for eviction. 2097dcbf3bd6SGeorge Wilson */ 2098dcbf3bd6SGeorge Wilson static int 2099dcbf3bd6SGeorge Wilson remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 2100dcbf3bd6SGeorge Wilson { 2101dcbf3bd6SGeorge Wilson int cnt; 2102dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2103dcbf3bd6SGeorge Wilson 2104dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2105dcbf3bd6SGeorge Wilson ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 2106dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(state)); 2107dcbf3bd6SGeorge Wilson 2108dcbf3bd6SGeorge Wilson /* 2109dcbf3bd6SGeorge Wilson * arc_l2c_only counts as a ghost state so we don't need to explicitly 2110dcbf3bd6SGeorge Wilson * check to prevent usage of the arc_l2c_only list. 2111dcbf3bd6SGeorge Wilson */ 2112dcbf3bd6SGeorge Wilson if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 2113dcbf3bd6SGeorge Wilson (state != arc_anon)) { 211494c2d0ebSMatthew Ahrens multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr); 2115dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 2116dcbf3bd6SGeorge Wilson arc_evictable_space_increment(hdr, state); 2117dcbf3bd6SGeorge Wilson } 2118dcbf3bd6SGeorge Wilson return (cnt); 2119dcbf3bd6SGeorge Wilson } 2120dcbf3bd6SGeorge Wilson 2121dcbf3bd6SGeorge Wilson /* 2122dcbf3bd6SGeorge Wilson * Move the supplied buffer to the indicated state. The hash lock 2123dcbf3bd6SGeorge Wilson * for the buffer must be held by the caller. 2124dcbf3bd6SGeorge Wilson */ 2125dcbf3bd6SGeorge Wilson static void 2126dcbf3bd6SGeorge Wilson arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 2127dcbf3bd6SGeorge Wilson kmutex_t *hash_lock) 2128dcbf3bd6SGeorge Wilson { 2129dcbf3bd6SGeorge Wilson arc_state_t *old_state; 2130dcbf3bd6SGeorge Wilson int64_t refcnt; 2131dcbf3bd6SGeorge Wilson uint32_t bufcnt; 2132dcbf3bd6SGeorge Wilson boolean_t update_old, update_new; 2133dcbf3bd6SGeorge Wilson arc_buf_contents_t buftype = arc_buf_type(hdr); 213489c86e32SChris Williamson 213589c86e32SChris Williamson /* 213689c86e32SChris Williamson * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 213789c86e32SChris Williamson * in arc_read() when bringing a buffer out of the L2ARC. However, the 213889c86e32SChris Williamson * L1 hdr doesn't always exist when we change state to arc_anon before 213989c86e32SChris Williamson * destroying a header, in which case reallocating to add the L1 hdr is 214089c86e32SChris Williamson * pointless. 214189c86e32SChris Williamson */ 214289c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) { 214389c86e32SChris Williamson old_state = hdr->b_l1hdr.b_state; 214489c86e32SChris Williamson refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 2145dcbf3bd6SGeorge Wilson bufcnt = hdr->b_l1hdr.b_bufcnt; 2146770499e1SDan Kimmel update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL); 214789c86e32SChris Williamson } else { 214889c86e32SChris Williamson old_state = arc_l2c_only; 214989c86e32SChris Williamson refcnt = 0; 2150dcbf3bd6SGeorge Wilson bufcnt = 0; 2151dcbf3bd6SGeorge Wilson update_old = B_FALSE; 215289c86e32SChris Williamson } 2153dcbf3bd6SGeorge Wilson update_new = update_old; 2154fa9e4066Sahrens 2155fa9e4066Sahrens ASSERT(MUTEX_HELD(hash_lock)); 215669962b56SMatthew Ahrens ASSERT3P(new_state, !=, old_state); 2157dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); 2158dcbf3bd6SGeorge Wilson ASSERT(old_state != arc_anon || bufcnt <= 1); 2159fa9e4066Sahrens 2160fa9e4066Sahrens /* 2161fa9e4066Sahrens * If this buffer is evictable, transfer it from the 2162fa9e4066Sahrens * old state list to the new state list. 2163fa9e4066Sahrens */ 2164ea8dc4b6Seschrock if (refcnt == 0) { 216589c86e32SChris Williamson if (old_state != arc_anon && old_state != arc_l2c_only) { 216689c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 216794c2d0ebSMatthew Ahrens multilist_remove(old_state->arcs_list[buftype], hdr); 2168ea8dc4b6Seschrock 2169dcbf3bd6SGeorge Wilson if (GHOST_STATE(old_state)) { 2170dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 2171dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2172dcbf3bd6SGeorge Wilson update_old = B_TRUE; 2173ea8dc4b6Seschrock } 21745602294fSDan Kimmel arc_evictable_space_decrement(hdr, old_state); 2175fa9e4066Sahrens } 217689c86e32SChris Williamson if (new_state != arc_anon && new_state != arc_l2c_only) { 2177fa9e4066Sahrens 217889c86e32SChris Williamson /* 217989c86e32SChris Williamson * An L1 header always exists here, since if we're 218089c86e32SChris Williamson * moving to some L1-cached state (i.e. not l2c_only or 218189c86e32SChris Williamson * anonymous), we realloc the header to add an L1hdr 218289c86e32SChris Williamson * beforehand. 218389c86e32SChris Williamson */ 218489c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 218594c2d0ebSMatthew Ahrens multilist_insert(new_state->arcs_list[buftype], hdr); 2186ea8dc4b6Seschrock 2187ea8dc4b6Seschrock if (GHOST_STATE(new_state)) { 2188dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 2189dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2190dcbf3bd6SGeorge Wilson update_new = B_TRUE; 2191ea8dc4b6Seschrock } 2192dcbf3bd6SGeorge Wilson arc_evictable_space_increment(hdr, new_state); 2193fa9e4066Sahrens } 2194fa9e4066Sahrens } 2195fa9e4066Sahrens 2196dcbf3bd6SGeorge Wilson ASSERT(!HDR_EMPTY(hdr)); 21977adb730bSGeorge Wilson if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 21987adb730bSGeorge Wilson buf_hash_remove(hdr); 2199fa9e4066Sahrens 220089c86e32SChris Williamson /* adjust state sizes (ignore arc_l2c_only) */ 22012fd872a7SPrakash Surya 2202dcbf3bd6SGeorge Wilson if (update_new && new_state != arc_l2c_only) { 22032fd872a7SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 22042fd872a7SPrakash Surya if (GHOST_STATE(new_state)) { 2205dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 22062fd872a7SPrakash Surya 22072fd872a7SPrakash Surya /* 2208dcbf3bd6SGeorge Wilson * When moving a header to a ghost state, we first 22092fd872a7SPrakash Surya * remove all arc buffers. Thus, we'll have a 2210dcbf3bd6SGeorge Wilson * bufcnt of zero, and no arc buffer to use for 22112fd872a7SPrakash Surya * the reference. As a result, we use the arc 22122fd872a7SPrakash Surya * header pointer for the reference. 22132fd872a7SPrakash Surya */ 22142fd872a7SPrakash Surya (void) refcount_add_many(&new_state->arcs_size, 2215dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr), hdr); 2216770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 22172fd872a7SPrakash Surya } else { 2218dcbf3bd6SGeorge Wilson uint32_t buffers = 0; 22192fd872a7SPrakash Surya 22202fd872a7SPrakash Surya /* 22212fd872a7SPrakash Surya * Each individual buffer holds a unique reference, 22222fd872a7SPrakash Surya * thus we must remove each of these references one 22232fd872a7SPrakash Surya * at a time. 22242fd872a7SPrakash Surya */ 22252fd872a7SPrakash Surya for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 22262fd872a7SPrakash Surya buf = buf->b_next) { 2227dcbf3bd6SGeorge Wilson ASSERT3U(bufcnt, !=, 0); 2228dcbf3bd6SGeorge Wilson buffers++; 2229dcbf3bd6SGeorge Wilson 2230dcbf3bd6SGeorge Wilson /* 2231dcbf3bd6SGeorge Wilson * When the arc_buf_t is sharing the data 2232dcbf3bd6SGeorge Wilson * block with the hdr, the owner of the 2233dcbf3bd6SGeorge Wilson * reference belongs to the hdr. Only 2234dcbf3bd6SGeorge Wilson * add to the refcount if the arc_buf_t is 2235dcbf3bd6SGeorge Wilson * not shared. 2236dcbf3bd6SGeorge Wilson */ 22375602294fSDan Kimmel if (arc_buf_is_shared(buf)) 2238dcbf3bd6SGeorge Wilson continue; 2239dcbf3bd6SGeorge Wilson 22402fd872a7SPrakash Surya (void) refcount_add_many(&new_state->arcs_size, 22415602294fSDan Kimmel arc_buf_size(buf), buf); 2242dcbf3bd6SGeorge Wilson } 2243dcbf3bd6SGeorge Wilson ASSERT3U(bufcnt, ==, buffers); 2244dcbf3bd6SGeorge Wilson 2245770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 2246dcbf3bd6SGeorge Wilson (void) refcount_add_many(&new_state->arcs_size, 2247dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2248dcbf3bd6SGeorge Wilson } else { 2249dcbf3bd6SGeorge Wilson ASSERT(GHOST_STATE(old_state)); 22502fd872a7SPrakash Surya } 22512fd872a7SPrakash Surya } 22522fd872a7SPrakash Surya } 22532fd872a7SPrakash Surya 2254dcbf3bd6SGeorge Wilson if (update_old && old_state != arc_l2c_only) { 22552fd872a7SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 22562fd872a7SPrakash Surya if (GHOST_STATE(old_state)) { 2257dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 2258770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2259dcbf3bd6SGeorge Wilson 22602fd872a7SPrakash Surya /* 22612fd872a7SPrakash Surya * When moving a header off of a ghost state, 2262dcbf3bd6SGeorge Wilson * the header will not contain any arc buffers. 2263dcbf3bd6SGeorge Wilson * We use the arc header pointer for the reference 2264dcbf3bd6SGeorge Wilson * which is exactly what we did when we put the 2265dcbf3bd6SGeorge Wilson * header on the ghost state. 22662fd872a7SPrakash Surya */ 22672fd872a7SPrakash Surya 22682fd872a7SPrakash Surya (void) refcount_remove_many(&old_state->arcs_size, 2269dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr), hdr); 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) { 22805602294fSDan Kimmel 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_remove_many( 22945602294fSDan Kimmel &old_state->arcs_size, arc_buf_size(buf), 2295dcbf3bd6SGeorge Wilson buf); 22962fd872a7SPrakash Surya } 2297dcbf3bd6SGeorge Wilson ASSERT3U(bufcnt, ==, buffers); 2298770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2299dcbf3bd6SGeorge Wilson (void) refcount_remove_many( 2300dcbf3bd6SGeorge Wilson &old_state->arcs_size, arc_hdr_size(hdr), hdr); 23012fd872a7SPrakash Surya } 2302fa9e4066Sahrens } 23032fd872a7SPrakash Surya 230489c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) 230589c86e32SChris Williamson hdr->b_l1hdr.b_state = new_state; 2306fa94a07fSbrendan 230789c86e32SChris Williamson /* 230889c86e32SChris Williamson * L2 headers should never be on the L2 state list since they don't 230989c86e32SChris Williamson * have L1 headers allocated. 231089c86e32SChris Williamson */ 231194c2d0ebSMatthew Ahrens ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 231294c2d0ebSMatthew Ahrens multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 2313fa9e4066Sahrens } 2314fa9e4066Sahrens 23150e8c6158Smaybee void 23165a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(uint64_t space, arc_space_type_t type) 23170e8c6158Smaybee { 23185a98e54bSBrendan Gregg - Sun Microsystems ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 23195a98e54bSBrendan Gregg - Sun Microsystems 23205a98e54bSBrendan Gregg - Sun Microsystems switch (type) { 23215a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_DATA: 23225a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_data_size, space); 23235a98e54bSBrendan Gregg - Sun Microsystems break; 23244076b1bfSPrakash Surya case ARC_SPACE_META: 23254076b1bfSPrakash Surya ARCSTAT_INCR(arcstat_metadata_size, space); 23264076b1bfSPrakash Surya break; 23275a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_OTHER: 23285a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_other_size, space); 23295a98e54bSBrendan Gregg - Sun Microsystems break; 23305a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_HDRS: 23315a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_hdr_size, space); 23325a98e54bSBrendan Gregg - Sun Microsystems break; 23335a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_L2HDRS: 23345a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_l2_hdr_size, space); 23355a98e54bSBrendan Gregg - Sun Microsystems break; 23365a98e54bSBrendan Gregg - Sun Microsystems } 23375a98e54bSBrendan Gregg - Sun Microsystems 23384076b1bfSPrakash Surya if (type != ARC_SPACE_DATA) 23394076b1bfSPrakash Surya ARCSTAT_INCR(arcstat_meta_used, space); 23404076b1bfSPrakash Surya 23410e8c6158Smaybee atomic_add_64(&arc_size, space); 23420e8c6158Smaybee } 23430e8c6158Smaybee 23440e8c6158Smaybee void 23455a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(uint64_t space, arc_space_type_t type) 23460e8c6158Smaybee { 23475a98e54bSBrendan Gregg - Sun Microsystems ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 23485a98e54bSBrendan Gregg - Sun Microsystems 23495a98e54bSBrendan Gregg - Sun Microsystems switch (type) { 23505a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_DATA: 23515a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_data_size, -space); 23525a98e54bSBrendan Gregg - Sun Microsystems break; 23534076b1bfSPrakash Surya case ARC_SPACE_META: 23544076b1bfSPrakash Surya ARCSTAT_INCR(arcstat_metadata_size, -space); 23554076b1bfSPrakash Surya break; 23565a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_OTHER: 23575a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_other_size, -space); 23585a98e54bSBrendan Gregg - Sun Microsystems break; 23595a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_HDRS: 23605a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_hdr_size, -space); 23615a98e54bSBrendan Gregg - Sun Microsystems break; 23625a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_L2HDRS: 23635a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 23645a98e54bSBrendan Gregg - Sun Microsystems break; 23655a98e54bSBrendan Gregg - Sun Microsystems } 23665a98e54bSBrendan Gregg - Sun Microsystems 23674076b1bfSPrakash Surya if (type != ARC_SPACE_DATA) { 23684076b1bfSPrakash Surya ASSERT(arc_meta_used >= space); 23694076b1bfSPrakash Surya if (arc_meta_max < arc_meta_used) 23704076b1bfSPrakash Surya arc_meta_max = arc_meta_used; 23714076b1bfSPrakash Surya ARCSTAT_INCR(arcstat_meta_used, -space); 23724076b1bfSPrakash Surya } 23734076b1bfSPrakash Surya 23740e8c6158Smaybee ASSERT(arc_size >= space); 23750e8c6158Smaybee atomic_add_64(&arc_size, -space); 23760e8c6158Smaybee } 23770e8c6158Smaybee 2378dcbf3bd6SGeorge Wilson /* 23795602294fSDan Kimmel * Given a hdr and a buf, returns whether that buf can share its b_data buffer 2380770499e1SDan Kimmel * with the hdr's b_pabd. 2381dcbf3bd6SGeorge Wilson */ 23825602294fSDan Kimmel static boolean_t 23835602294fSDan Kimmel arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf) 23845602294fSDan Kimmel { 23855602294fSDan Kimmel /* 23865602294fSDan Kimmel * The criteria for sharing a hdr's data are: 23875602294fSDan Kimmel * 1. the hdr's compression matches the buf's compression 23885602294fSDan Kimmel * 2. the hdr doesn't need to be byteswapped 23895602294fSDan Kimmel * 3. the hdr isn't already being shared 23905602294fSDan Kimmel * 4. the buf is either compressed or it is the last buf in the hdr list 23915602294fSDan Kimmel * 23925602294fSDan Kimmel * Criterion #4 maintains the invariant that shared uncompressed 23935602294fSDan Kimmel * bufs must be the final buf in the hdr's b_buf list. Reading this, you 23945602294fSDan Kimmel * might ask, "if a compressed buf is allocated first, won't that be the 23955602294fSDan Kimmel * last thing in the list?", but in that case it's impossible to create 23965602294fSDan Kimmel * a shared uncompressed buf anyway (because the hdr must be compressed 23975602294fSDan Kimmel * to have the compressed buf). You might also think that #3 is 23985602294fSDan Kimmel * sufficient to make this guarantee, however it's possible 23995602294fSDan Kimmel * (specifically in the rare L2ARC write race mentioned in 24005602294fSDan Kimmel * arc_buf_alloc_impl()) there will be an existing uncompressed buf that 24015602294fSDan Kimmel * is sharable, but wasn't at the time of its allocation. Rather than 24025602294fSDan Kimmel * allow a new shared uncompressed buf to be created and then shuffle 24035602294fSDan Kimmel * the list around to make it the last element, this simply disallows 24045602294fSDan Kimmel * sharing if the new buf isn't the first to be added. 24055602294fSDan Kimmel */ 24065602294fSDan Kimmel ASSERT3P(buf->b_hdr, ==, hdr); 24075602294fSDan Kimmel boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF; 24085602294fSDan Kimmel boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0; 24095602294fSDan Kimmel return (buf_compressed == hdr_compressed && 24105602294fSDan Kimmel hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && 24115602294fSDan Kimmel !HDR_SHARED_DATA(hdr) && 24125602294fSDan Kimmel (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf))); 24135602294fSDan Kimmel } 24145602294fSDan Kimmel 24155602294fSDan Kimmel /* 24165602294fSDan Kimmel * Allocate a buf for this hdr. If you care about the data that's in the hdr, 24175602294fSDan Kimmel * or if you want a compressed buffer, pass those flags in. Returns 0 if the 24185602294fSDan Kimmel * copy was made successfully, or an error code otherwise. 24195602294fSDan Kimmel */ 24205602294fSDan Kimmel static int 24215602294fSDan Kimmel arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed, 24225602294fSDan Kimmel boolean_t fill, arc_buf_t **ret) 2423fa9e4066Sahrens { 2424fa9e4066Sahrens arc_buf_t *buf; 2425fa9e4066Sahrens 2426dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2427dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2428dcbf3bd6SGeorge Wilson VERIFY(hdr->b_type == ARC_BUFC_DATA || 2429dcbf3bd6SGeorge Wilson hdr->b_type == ARC_BUFC_METADATA); 24305602294fSDan Kimmel ASSERT3P(ret, !=, NULL); 24315602294fSDan Kimmel ASSERT3P(*ret, ==, NULL); 2432dcbf3bd6SGeorge Wilson 24335602294fSDan Kimmel buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2434fa9e4066Sahrens buf->b_hdr = hdr; 243544eda4d7Smaybee buf->b_data = NULL; 24365602294fSDan Kimmel buf->b_next = hdr->b_l1hdr.b_buf; 24375602294fSDan Kimmel buf->b_flags = 0; 243889c86e32SChris Williamson 2439dcbf3bd6SGeorge Wilson add_reference(hdr, tag); 2440dcbf3bd6SGeorge Wilson 2441dcbf3bd6SGeorge Wilson /* 2442dcbf3bd6SGeorge Wilson * We're about to change the hdr's b_flags. We must either 2443dcbf3bd6SGeorge Wilson * hold the hash_lock or be undiscoverable. 2444dcbf3bd6SGeorge Wilson */ 2445dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2446dcbf3bd6SGeorge Wilson 2447dcbf3bd6SGeorge Wilson /* 24485602294fSDan Kimmel * Only honor requests for compressed bufs if the hdr is actually 24495602294fSDan Kimmel * compressed. 24505602294fSDan Kimmel */ 24515602294fSDan Kimmel if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 24525602294fSDan Kimmel buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; 24535602294fSDan Kimmel 24545602294fSDan Kimmel /* 24555602294fSDan Kimmel * If the hdr's data can be shared then we share the data buffer and 24565602294fSDan Kimmel * set the appropriate bit in the hdr's b_flags to indicate the hdr is 2457770499e1SDan Kimmel * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new 24585602294fSDan Kimmel * buffer to store the buf's data. 24595602294fSDan Kimmel * 2460770499e1SDan Kimmel * There are two additional restrictions here because we're sharing 2461770499e1SDan Kimmel * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be 2462770499e1SDan Kimmel * actively involved in an L2ARC write, because if this buf is used by 2463770499e1SDan Kimmel * an arc_write() then the hdr's data buffer will be released when the 24645602294fSDan Kimmel * write completes, even though the L2ARC write might still be using it. 2465770499e1SDan Kimmel * Second, the hdr's ABD must be linear so that the buf's user doesn't 2466770499e1SDan Kimmel * need to be ABD-aware. 2467dcbf3bd6SGeorge Wilson */ 2468770499e1SDan Kimmel boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) && 2469770499e1SDan Kimmel abd_is_linear(hdr->b_l1hdr.b_pabd); 24705602294fSDan Kimmel 24715602294fSDan Kimmel /* Set up b_data and sharing */ 24725602294fSDan Kimmel if (can_share) { 2473770499e1SDan Kimmel buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd); 24745602294fSDan Kimmel buf->b_flags |= ARC_BUF_FLAG_SHARED; 2475dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2476dcbf3bd6SGeorge Wilson } else { 24775602294fSDan Kimmel buf->b_data = 24785602294fSDan Kimmel arc_get_data_buf(hdr, arc_buf_size(buf), buf); 24795602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2480dcbf3bd6SGeorge Wilson } 2481dcbf3bd6SGeorge Wilson VERIFY3P(buf->b_data, !=, NULL); 248289c86e32SChris Williamson 248389c86e32SChris Williamson hdr->b_l1hdr.b_buf = buf; 2484dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt += 1; 248589c86e32SChris Williamson 24865602294fSDan Kimmel /* 24875602294fSDan Kimmel * If the user wants the data from the hdr, we need to either copy or 24885602294fSDan Kimmel * decompress the data. 24895602294fSDan Kimmel */ 24905602294fSDan Kimmel if (fill) { 24915602294fSDan Kimmel return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0)); 24925602294fSDan Kimmel } 2493dcbf3bd6SGeorge Wilson 24945602294fSDan Kimmel return (0); 24955602294fSDan Kimmel } 2496dcbf3bd6SGeorge Wilson 24975602294fSDan Kimmel static char *arc_onloan_tag = "onloan"; 2498fa9e4066Sahrens 24995602294fSDan Kimmel static inline void 25005602294fSDan Kimmel arc_loaned_bytes_update(int64_t delta) 25015602294fSDan Kimmel { 25025602294fSDan Kimmel atomic_add_64(&arc_loaned_bytes, delta); 2503dcbf3bd6SGeorge Wilson 25045602294fSDan Kimmel /* assert that it did not wrap around */ 25055602294fSDan Kimmel ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 2506fa9e4066Sahrens } 2507fa9e4066Sahrens 25082fdbea25SAleksandr Guzovskiy /* 25092fdbea25SAleksandr Guzovskiy * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 25102fdbea25SAleksandr Guzovskiy * flight data by arc_tempreserve_space() until they are "returned". Loaned 25112fdbea25SAleksandr Guzovskiy * buffers must be returned to the arc before they can be used by the DMU or 25122fdbea25SAleksandr Guzovskiy * freed. 25132fdbea25SAleksandr Guzovskiy */ 25142fdbea25SAleksandr Guzovskiy arc_buf_t * 25155602294fSDan Kimmel arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size) 25162fdbea25SAleksandr Guzovskiy { 25175602294fSDan Kimmel arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag, 25185602294fSDan Kimmel is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size); 25195602294fSDan Kimmel 25205602294fSDan Kimmel arc_loaned_bytes_update(size); 25215602294fSDan Kimmel 25225602294fSDan Kimmel return (buf); 25235602294fSDan Kimmel } 25242fdbea25SAleksandr Guzovskiy 25255602294fSDan Kimmel arc_buf_t * 25265602294fSDan Kimmel arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize, 25275602294fSDan Kimmel enum zio_compress compression_type) 25285602294fSDan Kimmel { 25295602294fSDan Kimmel arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag, 25305602294fSDan Kimmel psize, lsize, compression_type); 25315602294fSDan Kimmel 25325602294fSDan Kimmel arc_loaned_bytes_update(psize); 25332fdbea25SAleksandr Guzovskiy 25342fdbea25SAleksandr Guzovskiy return (buf); 25352fdbea25SAleksandr Guzovskiy } 25362fdbea25SAleksandr Guzovskiy 25375602294fSDan Kimmel 25382fdbea25SAleksandr Guzovskiy /* 25392fdbea25SAleksandr Guzovskiy * Return a loaned arc buffer to the arc. 25402fdbea25SAleksandr Guzovskiy */ 25412fdbea25SAleksandr Guzovskiy void 25422fdbea25SAleksandr Guzovskiy arc_return_buf(arc_buf_t *buf, void *tag) 25432fdbea25SAleksandr Guzovskiy { 25442fdbea25SAleksandr Guzovskiy arc_buf_hdr_t *hdr = buf->b_hdr; 25452fdbea25SAleksandr Guzovskiy 2546dcbf3bd6SGeorge Wilson ASSERT3P(buf->b_data, !=, NULL); 254789c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 254889c86e32SChris Williamson (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 254989c86e32SChris Williamson (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 25502fdbea25SAleksandr Guzovskiy 25515602294fSDan Kimmel arc_loaned_bytes_update(-arc_buf_size(buf)); 25522fdbea25SAleksandr Guzovskiy } 25532fdbea25SAleksandr Guzovskiy 2554c242f9a0Schunli zhang - Sun Microsystems - Irvine United States /* Detach an arc_buf from a dbuf (tag) */ 2555c242f9a0Schunli zhang - Sun Microsystems - Irvine United States void 2556c242f9a0Schunli zhang - Sun Microsystems - Irvine United States arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2557c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 255889c86e32SChris Williamson arc_buf_hdr_t *hdr = buf->b_hdr; 2559c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 2560dcbf3bd6SGeorge Wilson ASSERT3P(buf->b_data, !=, NULL); 256189c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 256289c86e32SChris Williamson (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 256389c86e32SChris Williamson (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2564c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 25655602294fSDan Kimmel arc_loaned_bytes_update(arc_buf_size(buf)); 2566c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 2567c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 2568dcbf3bd6SGeorge Wilson static void 2569770499e1SDan Kimmel l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type) 2570ea8dc4b6Seschrock { 2571dcbf3bd6SGeorge Wilson l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); 2572ea8dc4b6Seschrock 2573770499e1SDan Kimmel df->l2df_abd = abd; 2574dcbf3bd6SGeorge Wilson df->l2df_size = size; 2575dcbf3bd6SGeorge Wilson df->l2df_type = type; 2576dcbf3bd6SGeorge Wilson mutex_enter(&l2arc_free_on_write_mtx); 2577dcbf3bd6SGeorge Wilson list_insert_head(l2arc_free_on_write, df); 2578dcbf3bd6SGeorge Wilson mutex_exit(&l2arc_free_on_write_mtx); 2579dcbf3bd6SGeorge Wilson } 2580b24ab676SJeff Bonwick 2581dcbf3bd6SGeorge Wilson static void 2582dcbf3bd6SGeorge Wilson arc_hdr_free_on_write(arc_buf_hdr_t *hdr) 2583dcbf3bd6SGeorge Wilson { 2584dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2585dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 2586dcbf3bd6SGeorge Wilson uint64_t size = arc_hdr_size(hdr); 2587dcbf3bd6SGeorge Wilson 2588dcbf3bd6SGeorge Wilson /* protected by hash lock, if in the hash table */ 2589dcbf3bd6SGeorge Wilson if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 2590dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2591dcbf3bd6SGeorge Wilson ASSERT(state != arc_anon && state != arc_l2c_only); 2592dcbf3bd6SGeorge Wilson 2593dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 2594dcbf3bd6SGeorge Wilson size, hdr); 2595dcbf3bd6SGeorge Wilson } 2596dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_size, size, hdr); 25976de76ce2SAndriy Gapon if (type == ARC_BUFC_METADATA) { 25986de76ce2SAndriy Gapon arc_space_return(size, ARC_SPACE_META); 25996de76ce2SAndriy Gapon } else { 26006de76ce2SAndriy Gapon ASSERT(type == ARC_BUFC_DATA); 26016de76ce2SAndriy Gapon arc_space_return(size, ARC_SPACE_DATA); 26026de76ce2SAndriy Gapon } 2603dcbf3bd6SGeorge Wilson 2604770499e1SDan Kimmel l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type); 2605dcbf3bd6SGeorge Wilson } 2606dcbf3bd6SGeorge Wilson 2607dcbf3bd6SGeorge Wilson /* 2608dcbf3bd6SGeorge Wilson * Share the arc_buf_t's data with the hdr. Whenever we are sharing the 2609dcbf3bd6SGeorge Wilson * data buffer, we transfer the refcount ownership to the hdr and update 2610dcbf3bd6SGeorge Wilson * the appropriate kstats. 2611dcbf3bd6SGeorge Wilson */ 2612dcbf3bd6SGeorge Wilson static void 2613dcbf3bd6SGeorge Wilson arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2614dcbf3bd6SGeorge Wilson { 2615dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2616dcbf3bd6SGeorge Wilson 26175602294fSDan Kimmel ASSERT(arc_can_share(hdr, buf)); 2618770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2619dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 26209253d63dSGeorge Wilson 26219253d63dSGeorge Wilson /* 2622dcbf3bd6SGeorge Wilson * Start sharing the data buffer. We transfer the 2623dcbf3bd6SGeorge Wilson * refcount ownership to the hdr since it always owns 2624dcbf3bd6SGeorge Wilson * the refcount whenever an arc_buf_t is shared. 26259253d63dSGeorge Wilson */ 2626dcbf3bd6SGeorge Wilson refcount_transfer_ownership(&state->arcs_size, buf, hdr); 2627770499e1SDan Kimmel hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf)); 2628770499e1SDan Kimmel abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd, 2629770499e1SDan Kimmel HDR_ISTYPE_METADATA(hdr)); 2630dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 26315602294fSDan Kimmel buf->b_flags |= ARC_BUF_FLAG_SHARED; 2632dcbf3bd6SGeorge Wilson 2633dcbf3bd6SGeorge Wilson /* 2634dcbf3bd6SGeorge Wilson * Since we've transferred ownership to the hdr we need 2635dcbf3bd6SGeorge Wilson * to increment its compressed and uncompressed kstats and 2636dcbf3bd6SGeorge Wilson * decrement the overhead size. 2637dcbf3bd6SGeorge Wilson */ 2638dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2639dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 26405602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf)); 2641ea8dc4b6Seschrock } 2642ea8dc4b6Seschrock 2643dcbf3bd6SGeorge Wilson static void 2644dcbf3bd6SGeorge Wilson arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2645ea8dc4b6Seschrock { 2646dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2647dcbf3bd6SGeorge Wilson 2648dcbf3bd6SGeorge Wilson ASSERT(arc_buf_is_shared(buf)); 2649770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2650dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2651ea8dc4b6Seschrock 26529b23f181Smaybee /* 2653dcbf3bd6SGeorge Wilson * We are no longer sharing this buffer so we need 2654dcbf3bd6SGeorge Wilson * to transfer its ownership to the rightful owner. 26559b23f181Smaybee */ 2656dcbf3bd6SGeorge Wilson refcount_transfer_ownership(&state->arcs_size, hdr, buf); 2657dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2658770499e1SDan Kimmel abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd); 2659770499e1SDan Kimmel abd_put(hdr->b_l1hdr.b_pabd); 2660770499e1SDan Kimmel hdr->b_l1hdr.b_pabd = NULL; 26615602294fSDan Kimmel buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 2662dcbf3bd6SGeorge Wilson 2663dcbf3bd6SGeorge Wilson /* 2664dcbf3bd6SGeorge Wilson * Since the buffer is no longer shared between 2665dcbf3bd6SGeorge Wilson * the arc buf and the hdr, count it as overhead. 2666dcbf3bd6SGeorge Wilson */ 2667dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2668dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 26695602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2670dcbf3bd6SGeorge Wilson } 2671dcbf3bd6SGeorge Wilson 2672dcbf3bd6SGeorge Wilson /* 26735602294fSDan Kimmel * Remove an arc_buf_t from the hdr's buf list and return the last 26745602294fSDan Kimmel * arc_buf_t on the list. If no buffers remain on the list then return 26755602294fSDan Kimmel * NULL. 26765602294fSDan Kimmel */ 26775602294fSDan Kimmel static arc_buf_t * 26785602294fSDan Kimmel arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf) 26795602294fSDan Kimmel { 26805602294fSDan Kimmel ASSERT(HDR_HAS_L1HDR(hdr)); 26815602294fSDan Kimmel ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 26825602294fSDan Kimmel 26835602294fSDan Kimmel arc_buf_t **bufp = &hdr->b_l1hdr.b_buf; 26845602294fSDan Kimmel arc_buf_t *lastbuf = NULL; 26855602294fSDan Kimmel 26865602294fSDan Kimmel /* 26875602294fSDan Kimmel * Remove the buf from the hdr list and locate the last 26885602294fSDan Kimmel * remaining buffer on the list. 26895602294fSDan Kimmel */ 26905602294fSDan Kimmel while (*bufp != NULL) { 26915602294fSDan Kimmel if (*bufp == buf) 26925602294fSDan Kimmel *bufp = buf->b_next; 26935602294fSDan Kimmel 26945602294fSDan Kimmel /* 26955602294fSDan Kimmel * If we've removed a buffer in the middle of 26965602294fSDan Kimmel * the list then update the lastbuf and update 26975602294fSDan Kimmel * bufp. 26985602294fSDan Kimmel */ 26995602294fSDan Kimmel if (*bufp != NULL) { 27005602294fSDan Kimmel lastbuf = *bufp; 27015602294fSDan Kimmel bufp = &(*bufp)->b_next; 27025602294fSDan Kimmel } 27035602294fSDan Kimmel } 27045602294fSDan Kimmel buf->b_next = NULL; 27055602294fSDan Kimmel ASSERT3P(lastbuf, !=, buf); 27065602294fSDan Kimmel IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL); 27075602294fSDan Kimmel IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL); 27085602294fSDan Kimmel IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf)); 27095602294fSDan Kimmel 27105602294fSDan Kimmel return (lastbuf); 27115602294fSDan Kimmel } 27125602294fSDan Kimmel 27135602294fSDan Kimmel /* 27145602294fSDan Kimmel * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's 27155602294fSDan Kimmel * list and free it. 2716dcbf3bd6SGeorge Wilson */ 2717dcbf3bd6SGeorge Wilson static void 27185602294fSDan Kimmel arc_buf_destroy_impl(arc_buf_t *buf) 2719dcbf3bd6SGeorge Wilson { 2720dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 2721dcbf3bd6SGeorge Wilson 2722dcbf3bd6SGeorge Wilson /* 27235602294fSDan Kimmel * Free up the data associated with the buf but only if we're not 27245602294fSDan Kimmel * sharing this with the hdr. If we are sharing it with the hdr, the 27255602294fSDan Kimmel * hdr is responsible for doing the free. 2726dcbf3bd6SGeorge Wilson */ 2727dcbf3bd6SGeorge Wilson if (buf->b_data != NULL) { 2728dcbf3bd6SGeorge Wilson /* 2729dcbf3bd6SGeorge Wilson * We're about to change the hdr's b_flags. We must either 2730dcbf3bd6SGeorge Wilson * hold the hash_lock or be undiscoverable. 2731dcbf3bd6SGeorge Wilson */ 2732dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2733dcbf3bd6SGeorge Wilson 2734dcbf3bd6SGeorge Wilson arc_cksum_verify(buf); 2735dcbf3bd6SGeorge Wilson arc_buf_unwatch(buf); 2736dcbf3bd6SGeorge Wilson 27375602294fSDan Kimmel if (arc_buf_is_shared(buf)) { 2738dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2739dcbf3bd6SGeorge Wilson } else { 27405602294fSDan Kimmel uint64_t size = arc_buf_size(buf); 2741dcbf3bd6SGeorge Wilson arc_free_data_buf(hdr, buf->b_data, size, buf); 2742dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_overhead_size, -size); 2743dcbf3bd6SGeorge Wilson } 2744dcbf3bd6SGeorge Wilson buf->b_data = NULL; 2745dcbf3bd6SGeorge Wilson 2746dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 2747dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt -= 1; 2748dcbf3bd6SGeorge Wilson } 2749dcbf3bd6SGeorge Wilson 27505602294fSDan Kimmel arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 2751dcbf3bd6SGeorge Wilson 27525602294fSDan Kimmel if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) { 2753dcbf3bd6SGeorge Wilson /* 27545602294fSDan Kimmel * If the current arc_buf_t is sharing its data buffer with the 2755770499e1SDan Kimmel * hdr, then reassign the hdr's b_pabd to share it with the new 27565602294fSDan Kimmel * buffer at the end of the list. The shared buffer is always 27575602294fSDan Kimmel * the last one on the hdr's buffer list. 27585602294fSDan Kimmel * 27595602294fSDan Kimmel * There is an equivalent case for compressed bufs, but since 27605602294fSDan Kimmel * they aren't guaranteed to be the last buf in the list and 27615602294fSDan Kimmel * that is an exceedingly rare case, we just allow that space be 27625602294fSDan Kimmel * wasted temporarily. 2763dcbf3bd6SGeorge Wilson */ 27645602294fSDan Kimmel if (lastbuf != NULL) { 27655602294fSDan Kimmel /* Only one buf can be shared at once */ 27665602294fSDan Kimmel VERIFY(!arc_buf_is_shared(lastbuf)); 27675602294fSDan Kimmel /* hdr is uncompressed so can't have compressed buf */ 27685602294fSDan Kimmel VERIFY(!ARC_BUF_COMPRESSED(lastbuf)); 2769dcbf3bd6SGeorge Wilson 2770770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2771770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 2772dcbf3bd6SGeorge Wilson 27735602294fSDan Kimmel /* 27745602294fSDan Kimmel * We must setup a new shared block between the 27755602294fSDan Kimmel * last buffer and the hdr. The data would have 27765602294fSDan Kimmel * been allocated by the arc buf so we need to transfer 27775602294fSDan Kimmel * ownership to the hdr since it's now being shared. 27785602294fSDan Kimmel */ 27795602294fSDan Kimmel arc_share_buf(hdr, lastbuf); 27805602294fSDan Kimmel } 27815602294fSDan Kimmel } else if (HDR_SHARED_DATA(hdr)) { 2782dcbf3bd6SGeorge Wilson /* 27835602294fSDan Kimmel * Uncompressed shared buffers are always at the end 27845602294fSDan Kimmel * of the list. Compressed buffers don't have the 27855602294fSDan Kimmel * same requirements. This makes it hard to 27865602294fSDan Kimmel * simply assert that the lastbuf is shared so 27875602294fSDan Kimmel * we rely on the hdr's compression flags to determine 27885602294fSDan Kimmel * if we have a compressed, shared buffer. 2789dcbf3bd6SGeorge Wilson */ 27905602294fSDan Kimmel ASSERT3P(lastbuf, !=, NULL); 27915602294fSDan Kimmel ASSERT(arc_buf_is_shared(lastbuf) || 27925602294fSDan Kimmel HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 2793dcbf3bd6SGeorge Wilson } 2794dcbf3bd6SGeorge Wilson 27955602294fSDan Kimmel /* 27965602294fSDan Kimmel * Free the checksum if we're removing the last uncompressed buf from 27975602294fSDan Kimmel * this hdr. 27985602294fSDan Kimmel */ 27995602294fSDan Kimmel if (!arc_hdr_has_uncompressed_buf(hdr)) { 2800dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 28015602294fSDan Kimmel } 2802dcbf3bd6SGeorge Wilson 2803dcbf3bd6SGeorge Wilson /* clean up the buf */ 2804dcbf3bd6SGeorge Wilson buf->b_hdr = NULL; 2805dcbf3bd6SGeorge Wilson kmem_cache_free(buf_cache, buf); 2806dcbf3bd6SGeorge Wilson } 2807dcbf3bd6SGeorge Wilson 2808dcbf3bd6SGeorge Wilson static void 2809770499e1SDan Kimmel arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr) 2810dcbf3bd6SGeorge Wilson { 2811dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 281289c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 2813dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(hdr)); 2814ea8dc4b6Seschrock 2815770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2816770499e1SDan Kimmel hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr); 2817dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2818770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 281989c86e32SChris Williamson 2820dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2821dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2822ea8dc4b6Seschrock } 2823ea8dc4b6Seschrock 2824244781f1SPrakash Surya static void 2825770499e1SDan Kimmel arc_hdr_free_pabd(arc_buf_hdr_t *hdr) 2826244781f1SPrakash Surya { 2827dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2828770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2829244781f1SPrakash Surya 2830dcbf3bd6SGeorge Wilson /* 2831dcbf3bd6SGeorge Wilson * If the hdr is currently being written to the l2arc then 2832dcbf3bd6SGeorge Wilson * we defer freeing the data by adding it to the l2arc_free_on_write 2833dcbf3bd6SGeorge Wilson * list. The l2arc will free the data once it's finished 2834dcbf3bd6SGeorge Wilson * writing it to the l2arc device. 2835dcbf3bd6SGeorge Wilson */ 2836dcbf3bd6SGeorge Wilson if (HDR_L2_WRITING(hdr)) { 2837dcbf3bd6SGeorge Wilson arc_hdr_free_on_write(hdr); 2838dcbf3bd6SGeorge Wilson ARCSTAT_BUMP(arcstat_l2_free_on_write); 2839dcbf3bd6SGeorge Wilson } else { 2840770499e1SDan Kimmel arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, 2841dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2842dcbf3bd6SGeorge Wilson } 2843770499e1SDan Kimmel hdr->b_l1hdr.b_pabd = NULL; 2844dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2845dcbf3bd6SGeorge Wilson 2846dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2847dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2848dcbf3bd6SGeorge Wilson } 2849dcbf3bd6SGeorge Wilson 2850dcbf3bd6SGeorge Wilson static arc_buf_hdr_t * 2851dcbf3bd6SGeorge Wilson arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, 28525602294fSDan Kimmel enum zio_compress compression_type, arc_buf_contents_t type) 2853dcbf3bd6SGeorge Wilson { 2854dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr; 2855dcbf3bd6SGeorge Wilson 2856dcbf3bd6SGeorge Wilson VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); 2857dcbf3bd6SGeorge Wilson 2858dcbf3bd6SGeorge Wilson hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2859dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 2860dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2861dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL); 2862dcbf3bd6SGeorge Wilson HDR_SET_PSIZE(hdr, psize); 2863dcbf3bd6SGeorge Wilson HDR_SET_LSIZE(hdr, lsize); 2864dcbf3bd6SGeorge Wilson hdr->b_spa = spa; 2865dcbf3bd6SGeorge Wilson hdr->b_type = type; 2866dcbf3bd6SGeorge Wilson hdr->b_flags = 0; 2867dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); 28685602294fSDan Kimmel arc_hdr_set_compress(hdr, compression_type); 2869dcbf3bd6SGeorge Wilson 2870dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_state = arc_anon; 2871dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_arc_access = 0; 2872dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt = 0; 2873dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_buf = NULL; 2874dcbf3bd6SGeorge Wilson 2875dcbf3bd6SGeorge Wilson /* 2876dcbf3bd6SGeorge Wilson * Allocate the hdr's buffer. This will contain either 2877dcbf3bd6SGeorge Wilson * the compressed or uncompressed data depending on the block 2878dcbf3bd6SGeorge Wilson * it references and compressed arc enablement. 2879dcbf3bd6SGeorge Wilson */ 2880770499e1SDan Kimmel arc_hdr_alloc_pabd(hdr); 2881dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2882dcbf3bd6SGeorge Wilson 2883dcbf3bd6SGeorge Wilson return (hdr); 2884244781f1SPrakash Surya } 2885244781f1SPrakash Surya 2886fa94a07fSbrendan /* 2887dcbf3bd6SGeorge Wilson * Transition between the two allocation states for the arc_buf_hdr struct. 2888dcbf3bd6SGeorge Wilson * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 2889dcbf3bd6SGeorge Wilson * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 2890dcbf3bd6SGeorge Wilson * version is used when a cache buffer is only in the L2ARC in order to reduce 2891dcbf3bd6SGeorge Wilson * memory usage. 2892fa94a07fSbrendan */ 2893dcbf3bd6SGeorge Wilson static arc_buf_hdr_t * 2894dcbf3bd6SGeorge Wilson arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 2895fa94a07fSbrendan { 2896dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L2HDR(hdr)); 2897dcbf3bd6SGeorge Wilson 2898dcbf3bd6SGeorge Wilson arc_buf_hdr_t *nhdr; 2899dcbf3bd6SGeorge Wilson l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2900dcbf3bd6SGeorge Wilson 2901dcbf3bd6SGeorge Wilson ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 2902dcbf3bd6SGeorge Wilson (old == hdr_l2only_cache && new == hdr_full_cache)); 2903dcbf3bd6SGeorge Wilson 2904dcbf3bd6SGeorge Wilson nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 2905dcbf3bd6SGeorge Wilson 2906dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 2907dcbf3bd6SGeorge Wilson buf_hash_remove(hdr); 2908dcbf3bd6SGeorge Wilson 2909dcbf3bd6SGeorge Wilson bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 2910dcbf3bd6SGeorge Wilson 2911dcbf3bd6SGeorge Wilson if (new == hdr_full_cache) { 2912dcbf3bd6SGeorge Wilson arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2913dcbf3bd6SGeorge Wilson /* 2914dcbf3bd6SGeorge Wilson * arc_access and arc_change_state need to be aware that a 2915dcbf3bd6SGeorge Wilson * header has just come out of L2ARC, so we set its state to 2916dcbf3bd6SGeorge Wilson * l2c_only even though it's about to change. 2917dcbf3bd6SGeorge Wilson */ 2918dcbf3bd6SGeorge Wilson nhdr->b_l1hdr.b_state = arc_l2c_only; 2919dcbf3bd6SGeorge Wilson 2920dcbf3bd6SGeorge Wilson /* Verify previous threads set to NULL before freeing */ 2921770499e1SDan Kimmel ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL); 2922dcbf3bd6SGeorge Wilson } else { 2923dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2924dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 2925dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2926dcbf3bd6SGeorge Wilson 2927dcbf3bd6SGeorge Wilson /* 2928dcbf3bd6SGeorge Wilson * If we've reached here, We must have been called from 2929dcbf3bd6SGeorge Wilson * arc_evict_hdr(), as such we should have already been 2930dcbf3bd6SGeorge Wilson * removed from any ghost list we were previously on 2931dcbf3bd6SGeorge Wilson * (which protects us from racing with arc_evict_state), 2932dcbf3bd6SGeorge Wilson * thus no locking is needed during this check. 2933dcbf3bd6SGeorge Wilson */ 2934dcbf3bd6SGeorge Wilson ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2935cd1c8b85SMatthew Ahrens 2936dcbf3bd6SGeorge Wilson /* 2937dcbf3bd6SGeorge Wilson * A buffer must not be moved into the arc_l2c_only 2938dcbf3bd6SGeorge Wilson * state if it's not finished being written out to the 2939770499e1SDan Kimmel * l2arc device. Otherwise, the b_l1hdr.b_pabd field 2940dcbf3bd6SGeorge Wilson * might try to be accessed, even though it was removed. 2941dcbf3bd6SGeorge Wilson */ 2942dcbf3bd6SGeorge Wilson VERIFY(!HDR_L2_WRITING(hdr)); 2943770499e1SDan Kimmel VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2944fa94a07fSbrendan 2945dcbf3bd6SGeorge Wilson #ifdef ZFS_DEBUG 2946dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_thawed != NULL) { 2947dcbf3bd6SGeorge Wilson kmem_free(hdr->b_l1hdr.b_thawed, 1); 2948dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_thawed = NULL; 2949dcbf3bd6SGeorge Wilson } 2950dcbf3bd6SGeorge Wilson #endif 2951244781f1SPrakash Surya 2952dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2953dcbf3bd6SGeorge Wilson } 2954244781f1SPrakash Surya /* 2955dcbf3bd6SGeorge Wilson * The header has been reallocated so we need to re-insert it into any 2956dcbf3bd6SGeorge Wilson * lists it was on. 2957244781f1SPrakash Surya */ 2958dcbf3bd6SGeorge Wilson (void) buf_hash_insert(nhdr, NULL); 2959244781f1SPrakash Surya 2960dcbf3bd6SGeorge Wilson ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 2961dcbf3bd6SGeorge Wilson 2962dcbf3bd6SGeorge Wilson mutex_enter(&dev->l2ad_mtx); 2963244781f1SPrakash Surya 2964244781f1SPrakash Surya /* 2965dcbf3bd6SGeorge Wilson * We must place the realloc'ed header back into the list at 2966dcbf3bd6SGeorge Wilson * the same spot. Otherwise, if it's placed earlier in the list, 2967dcbf3bd6SGeorge Wilson * l2arc_write_buffers() could find it during the function's 2968dcbf3bd6SGeorge Wilson * write phase, and try to write it out to the l2arc. 2969244781f1SPrakash Surya */ 2970dcbf3bd6SGeorge Wilson list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 2971dcbf3bd6SGeorge Wilson list_remove(&dev->l2ad_buflist, hdr); 2972dcbf3bd6SGeorge Wilson 2973dcbf3bd6SGeorge Wilson mutex_exit(&dev->l2ad_mtx); 2974244781f1SPrakash Surya 2975244781f1SPrakash Surya /* 2976dcbf3bd6SGeorge Wilson * Since we're using the pointer address as the tag when 2977dcbf3bd6SGeorge Wilson * incrementing and decrementing the l2ad_alloc refcount, we 2978dcbf3bd6SGeorge Wilson * must remove the old pointer (that we're about to destroy) and 2979dcbf3bd6SGeorge Wilson * add the new pointer to the refcount. Otherwise we'd remove 2980dcbf3bd6SGeorge Wilson * the wrong pointer address when calling arc_hdr_destroy() later. 2981244781f1SPrakash Surya */ 2982244781f1SPrakash Surya 2983dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); 2984dcbf3bd6SGeorge Wilson (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr); 2985244781f1SPrakash Surya 2986dcbf3bd6SGeorge Wilson buf_discard_identity(hdr); 2987dcbf3bd6SGeorge Wilson kmem_cache_free(old, hdr); 2988244781f1SPrakash Surya 2989dcbf3bd6SGeorge Wilson return (nhdr); 2990244781f1SPrakash Surya } 2991244781f1SPrakash Surya 2992bbfa8ea8SMatthew Ahrens /* 2993dcbf3bd6SGeorge Wilson * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. 2994dcbf3bd6SGeorge Wilson * The buf is returned thawed since we expect the consumer to modify it. 2995bbfa8ea8SMatthew Ahrens */ 2996dcbf3bd6SGeorge Wilson arc_buf_t * 29975602294fSDan Kimmel arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size) 2998ea8dc4b6Seschrock { 2999dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, 3000dcbf3bd6SGeorge Wilson ZIO_COMPRESS_OFF, type); 3001dcbf3bd6SGeorge Wilson ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 30025602294fSDan Kimmel 30035602294fSDan Kimmel arc_buf_t *buf = NULL; 30045602294fSDan Kimmel VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf)); 3005dcbf3bd6SGeorge Wilson arc_buf_thaw(buf); 30065602294fSDan Kimmel 30075602294fSDan Kimmel return (buf); 30085602294fSDan Kimmel } 30095602294fSDan Kimmel 30105602294fSDan Kimmel /* 30115602294fSDan Kimmel * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this 30125602294fSDan Kimmel * for bufs containing metadata. 30135602294fSDan Kimmel */ 30145602294fSDan Kimmel arc_buf_t * 30155602294fSDan Kimmel arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize, 30165602294fSDan Kimmel enum zio_compress compression_type) 30175602294fSDan Kimmel { 30185602294fSDan Kimmel ASSERT3U(lsize, >, 0); 30195602294fSDan Kimmel ASSERT3U(lsize, >=, psize); 30205602294fSDan Kimmel ASSERT(compression_type > ZIO_COMPRESS_OFF); 30215602294fSDan Kimmel ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS); 30225602294fSDan Kimmel 30235602294fSDan Kimmel arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 30245602294fSDan Kimmel compression_type, ARC_BUFC_DATA); 30255602294fSDan Kimmel ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 30265602294fSDan Kimmel 30275602294fSDan Kimmel arc_buf_t *buf = NULL; 30285602294fSDan Kimmel VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf)); 30295602294fSDan Kimmel arc_buf_thaw(buf); 30305602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 30315602294fSDan Kimmel 3032770499e1SDan Kimmel if (!arc_buf_is_shared(buf)) { 3033770499e1SDan Kimmel /* 3034770499e1SDan Kimmel * To ensure that the hdr has the correct data in it if we call 3035770499e1SDan Kimmel * arc_decompress() on this buf before it's been written to 3036770499e1SDan Kimmel * disk, it's easiest if we just set up sharing between the 3037770499e1SDan Kimmel * buf and the hdr. 3038770499e1SDan Kimmel */ 3039770499e1SDan Kimmel ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd)); 3040770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 3041770499e1SDan Kimmel arc_share_buf(hdr, buf); 3042770499e1SDan Kimmel } 3043770499e1SDan Kimmel 3044dcbf3bd6SGeorge Wilson return (buf); 3045ea8dc4b6Seschrock } 3046ea8dc4b6Seschrock 3047a52fc310SPrakash Surya static void 3048a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 3049a52fc310SPrakash Surya { 3050a52fc310SPrakash Surya l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 3051a52fc310SPrakash Surya l2arc_dev_t *dev = l2hdr->b_dev; 305216a7e5acSAndriy Gapon uint64_t psize = arc_hdr_size(hdr); 3053a52fc310SPrakash Surya 3054a52fc310SPrakash Surya ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 3055a52fc310SPrakash Surya ASSERT(HDR_HAS_L2HDR(hdr)); 3056a52fc310SPrakash Surya 3057a52fc310SPrakash Surya list_remove(&dev->l2ad_buflist, hdr); 3058a52fc310SPrakash Surya 305916a7e5acSAndriy Gapon ARCSTAT_INCR(arcstat_l2_psize, -psize); 306016a7e5acSAndriy Gapon ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); 3061a52fc310SPrakash Surya 306216a7e5acSAndriy Gapon vdev_space_update(dev->l2ad_vdev, -psize, 0, 0); 3063a52fc310SPrakash Surya 306416a7e5acSAndriy Gapon (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr); 3065dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 3066a52fc310SPrakash Surya } 3067a52fc310SPrakash Surya 3068fa9e4066Sahrens static void 3069ea8dc4b6Seschrock arc_hdr_destroy(arc_buf_hdr_t *hdr) 3070fa9e4066Sahrens { 307189c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) { 307289c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_buf == NULL || 3073dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt > 0); 307489c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 307589c86e32SChris Williamson ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 307689c86e32SChris Williamson } 3077ea8dc4b6Seschrock ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 307889c86e32SChris Williamson ASSERT(!HDR_IN_HASH_TABLE(hdr)); 307989c86e32SChris Williamson 3080dcbf3bd6SGeorge Wilson if (!HDR_EMPTY(hdr)) 3081dcbf3bd6SGeorge Wilson buf_discard_identity(hdr); 3082dcbf3bd6SGeorge Wilson 308389c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr)) { 3084a52fc310SPrakash Surya l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3085a52fc310SPrakash Surya boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 3086b24ab676SJeff Bonwick 3087a52fc310SPrakash Surya if (!buflist_held) 3088a52fc310SPrakash Surya mutex_enter(&dev->l2ad_mtx); 308989c86e32SChris Williamson 3090244781f1SPrakash Surya /* 3091a52fc310SPrakash Surya * Even though we checked this conditional above, we 3092a52fc310SPrakash Surya * need to check this again now that we have the 3093a52fc310SPrakash Surya * l2ad_mtx. This is because we could be racing with 3094a52fc310SPrakash Surya * another thread calling l2arc_evict() which might have 3095a52fc310SPrakash Surya * destroyed this header's L2 portion as we were waiting 3096a52fc310SPrakash Surya * to acquire the l2ad_mtx. If that happens, we don't 3097a52fc310SPrakash Surya * want to re-destroy the header's L2 portion. 3098244781f1SPrakash Surya */ 3099a52fc310SPrakash Surya if (HDR_HAS_L2HDR(hdr)) 3100a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(hdr); 3101b24ab676SJeff Bonwick 3102b24ab676SJeff Bonwick if (!buflist_held) 3103a52fc310SPrakash Surya mutex_exit(&dev->l2ad_mtx); 3104fa94a07fSbrendan } 3105fa94a07fSbrendan 3106dcbf3bd6SGeorge Wilson if (HDR_HAS_L1HDR(hdr)) { 3107dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 310889c86e32SChris Williamson 3109dcbf3bd6SGeorge Wilson while (hdr->b_l1hdr.b_buf != NULL) 31105602294fSDan Kimmel arc_buf_destroy_impl(hdr->b_l1hdr.b_buf); 311189c86e32SChris Williamson 311289c86e32SChris Williamson #ifdef ZFS_DEBUG 311389c86e32SChris Williamson if (hdr->b_l1hdr.b_thawed != NULL) { 311489c86e32SChris Williamson kmem_free(hdr->b_l1hdr.b_thawed, 1); 311589c86e32SChris Williamson hdr->b_l1hdr.b_thawed = NULL; 311689c86e32SChris Williamson } 311789c86e32SChris Williamson #endif 3118dcbf3bd6SGeorge Wilson 3119770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 3120770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 3121dcbf3bd6SGeorge Wilson } 31223f9d6ad7SLin Ling } 3123ea8dc4b6Seschrock 3124fa9e4066Sahrens ASSERT3P(hdr->b_hash_next, ==, NULL); 312589c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) { 3126244781f1SPrakash Surya ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 312789c86e32SChris Williamson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 312889c86e32SChris Williamson kmem_cache_free(hdr_full_cache, hdr); 312989c86e32SChris Williamson } else { 313089c86e32SChris Williamson kmem_cache_free(hdr_l2only_cache, hdr); 313189c86e32SChris Williamson } 3132fa9e4066Sahrens } 3133fa9e4066Sahrens 3134fa9e4066Sahrens void 3135dcbf3bd6SGeorge Wilson arc_buf_destroy(arc_buf_t *buf, void* tag) 3136ea8dc4b6Seschrock { 3137ea8dc4b6Seschrock arc_buf_hdr_t *hdr = buf->b_hdr; 3138ea8dc4b6Seschrock kmutex_t *hash_lock = HDR_LOCK(hdr); 3139fa9e4066Sahrens 314089c86e32SChris Williamson if (hdr->b_l1hdr.b_state == arc_anon) { 3141dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 3142dcbf3bd6SGeorge Wilson ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3143dcbf3bd6SGeorge Wilson VERIFY0(remove_reference(hdr, NULL, tag)); 3144dcbf3bd6SGeorge Wilson arc_hdr_destroy(hdr); 3145dcbf3bd6SGeorge Wilson return; 3146ea8dc4b6Seschrock } 3147ea8dc4b6Seschrock 3148ea8dc4b6Seschrock mutex_enter(hash_lock); 3149dcbf3bd6SGeorge Wilson ASSERT3P(hdr, ==, buf->b_hdr); 3150dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 31513f9d6ad7SLin Ling ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3152dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); 3153dcbf3bd6SGeorge Wilson ASSERT3P(buf->b_data, !=, NULL); 3154ea8dc4b6Seschrock 3155ea8dc4b6Seschrock (void) remove_reference(hdr, hash_lock, tag); 31565602294fSDan Kimmel arc_buf_destroy_impl(buf); 3157ea8dc4b6Seschrock mutex_exit(hash_lock); 3158fa9e4066Sahrens } 3159fa9e4066Sahrens 3160fa9e4066Sahrens /* 3161244781f1SPrakash Surya * Evict the arc_buf_hdr that is provided as a parameter. The resultant 3162244781f1SPrakash Surya * state of the header is dependent on it's state prior to entering this 3163244781f1SPrakash Surya * function. The following transitions are possible: 3164874395d5Smaybee * 3165244781f1SPrakash Surya * - arc_mru -> arc_mru_ghost 3166244781f1SPrakash Surya * - arc_mfu -> arc_mfu_ghost 3167244781f1SPrakash Surya * - arc_mru_ghost -> arc_l2c_only 3168244781f1SPrakash Surya * - arc_mru_ghost -> deleted 3169244781f1SPrakash Surya * - arc_mfu_ghost -> arc_l2c_only 3170244781f1SPrakash Surya * - arc_mfu_ghost -> deleted 3171fa9e4066Sahrens */ 3172244781f1SPrakash Surya static int64_t 3173244781f1SPrakash Surya arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3174fa9e4066Sahrens { 3175244781f1SPrakash Surya arc_state_t *evicted_state, *state; 3176244781f1SPrakash Surya int64_t bytes_evicted = 0; 3177fa9e4066Sahrens 3178244781f1SPrakash Surya ASSERT(MUTEX_HELD(hash_lock)); 3179244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 3180fa9e4066Sahrens 3181244781f1SPrakash Surya state = hdr->b_l1hdr.b_state; 3182244781f1SPrakash Surya if (GHOST_STATE(state)) { 3183244781f1SPrakash Surya ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3184dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3185fa9e4066Sahrens 3186244781f1SPrakash Surya /* 3187244781f1SPrakash Surya * l2arc_write_buffers() relies on a header's L1 portion 3188770499e1SDan Kimmel * (i.e. its b_pabd field) during it's write phase. 3189244781f1SPrakash Surya * Thus, we cannot push a header onto the arc_l2c_only 3190244781f1SPrakash Surya * state (removing it's L1 piece) until the header is 3191244781f1SPrakash Surya * done being written to the l2arc. 3192244781f1SPrakash Surya */ 3193244781f1SPrakash Surya if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 3194244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_l2_skip); 3195244781f1SPrakash Surya return (bytes_evicted); 31963a5286a1SMatthew Ahrens } 3197244781f1SPrakash Surya 3198244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_deleted); 3199dcbf3bd6SGeorge Wilson bytes_evicted += HDR_GET_LSIZE(hdr); 3200244781f1SPrakash Surya 3201244781f1SPrakash Surya DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 3202244781f1SPrakash Surya 3203770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3204244781f1SPrakash Surya if (HDR_HAS_L2HDR(hdr)) { 3205244781f1SPrakash Surya /* 3206244781f1SPrakash Surya * This buffer is cached on the 2nd Level ARC; 3207244781f1SPrakash Surya * don't destroy the header. 3208244781f1SPrakash Surya */ 3209244781f1SPrakash Surya arc_change_state(arc_l2c_only, hdr, hash_lock); 32103a5286a1SMatthew Ahrens /* 3211244781f1SPrakash Surya * dropping from L1+L2 cached to L2-only, 3212244781f1SPrakash Surya * realloc to remove the L1 header. 32133a5286a1SMatthew Ahrens */ 3214244781f1SPrakash Surya hdr = arc_hdr_realloc(hdr, hdr_full_cache, 3215244781f1SPrakash Surya hdr_l2only_cache); 3216244781f1SPrakash Surya } else { 3217244781f1SPrakash Surya arc_change_state(arc_anon, hdr, hash_lock); 3218244781f1SPrakash Surya arc_hdr_destroy(hdr); 32193a5286a1SMatthew Ahrens } 3220244781f1SPrakash Surya return (bytes_evicted); 32213a5286a1SMatthew Ahrens } 32223a5286a1SMatthew Ahrens 3223244781f1SPrakash Surya ASSERT(state == arc_mru || state == arc_mfu); 3224244781f1SPrakash Surya evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3225244781f1SPrakash Surya 3226244781f1SPrakash Surya /* prefetch buffers have a minimum lifespan */ 3227244781f1SPrakash Surya if (HDR_IO_IN_PROGRESS(hdr) || 3228244781f1SPrakash Surya ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 3229244781f1SPrakash Surya ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 3230244781f1SPrakash Surya arc_min_prefetch_lifespan)) { 3231244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_skip); 3232244781f1SPrakash Surya return (bytes_evicted); 3233244781f1SPrakash Surya } 3234244781f1SPrakash Surya 3235244781f1SPrakash Surya ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3236244781f1SPrakash Surya while (hdr->b_l1hdr.b_buf) { 3237244781f1SPrakash Surya arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3238244781f1SPrakash Surya if (!mutex_tryenter(&buf->b_evict_lock)) { 3239244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_mutex_miss); 3240244781f1SPrakash Surya break; 324113506d1eSmaybee } 3242244781f1SPrakash Surya if (buf->b_data != NULL) 3243dcbf3bd6SGeorge Wilson bytes_evicted += HDR_GET_LSIZE(hdr); 3244dcbf3bd6SGeorge Wilson mutex_exit(&buf->b_evict_lock); 32455602294fSDan Kimmel arc_buf_destroy_impl(buf); 3246244781f1SPrakash Surya } 324769962b56SMatthew Ahrens 3248244781f1SPrakash Surya if (HDR_HAS_L2HDR(hdr)) { 3249dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); 3250244781f1SPrakash Surya } else { 3251dcbf3bd6SGeorge Wilson if (l2arc_write_eligible(hdr->b_spa, hdr)) { 3252dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_evict_l2_eligible, 3253dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr)); 3254dcbf3bd6SGeorge Wilson } else { 3255dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_evict_l2_ineligible, 3256dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr)); 3257dcbf3bd6SGeorge Wilson } 3258244781f1SPrakash Surya } 3259244781f1SPrakash Surya 3260dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_bufcnt == 0) { 3261dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 3262dcbf3bd6SGeorge Wilson 3263dcbf3bd6SGeorge Wilson bytes_evicted += arc_hdr_size(hdr); 3264dcbf3bd6SGeorge Wilson 3265dcbf3bd6SGeorge Wilson /* 3266dcbf3bd6SGeorge Wilson * If this hdr is being evicted and has a compressed 3267dcbf3bd6SGeorge Wilson * buffer then we discard it here before we change states. 3268dcbf3bd6SGeorge Wilson * This ensures that the accounting is updated correctly 3269770499e1SDan Kimmel * in arc_free_data_impl(). 3270dcbf3bd6SGeorge Wilson */ 3271770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 3272dcbf3bd6SGeorge Wilson 3273244781f1SPrakash Surya arc_change_state(evicted_state, hdr, hash_lock); 3274244781f1SPrakash Surya ASSERT(HDR_IN_HASH_TABLE(hdr)); 3275dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 3276244781f1SPrakash Surya DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 3277244781f1SPrakash Surya } 3278244781f1SPrakash Surya 3279244781f1SPrakash Surya return (bytes_evicted); 3280244781f1SPrakash Surya } 3281244781f1SPrakash Surya 3282244781f1SPrakash Surya static uint64_t 3283244781f1SPrakash Surya arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 3284244781f1SPrakash Surya uint64_t spa, int64_t bytes) 3285244781f1SPrakash Surya { 3286244781f1SPrakash Surya multilist_sublist_t *mls; 3287244781f1SPrakash Surya uint64_t bytes_evicted = 0; 3288244781f1SPrakash Surya arc_buf_hdr_t *hdr; 3289244781f1SPrakash Surya kmutex_t *hash_lock; 3290244781f1SPrakash Surya int evict_count = 0; 3291244781f1SPrakash Surya 3292244781f1SPrakash Surya ASSERT3P(marker, !=, NULL); 3293244781f1SPrakash Surya IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3294244781f1SPrakash Surya 3295244781f1SPrakash Surya mls = multilist_sublist_lock(ml, idx); 3296244781f1SPrakash Surya 3297244781f1SPrakash Surya for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 3298244781f1SPrakash Surya hdr = multilist_sublist_prev(mls, marker)) { 3299244781f1SPrakash Surya if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 3300244781f1SPrakash Surya (evict_count >= zfs_arc_evict_batch_limit)) 3301244781f1SPrakash Surya break; 330269962b56SMatthew Ahrens 330369962b56SMatthew Ahrens /* 3304244781f1SPrakash Surya * To keep our iteration location, move the marker 3305244781f1SPrakash Surya * forward. Since we're not holding hdr's hash lock, we 3306244781f1SPrakash Surya * must be very careful and not remove 'hdr' from the 3307244781f1SPrakash Surya * sublist. Otherwise, other consumers might mistake the 3308244781f1SPrakash Surya * 'hdr' as not being on a sublist when they call the 3309244781f1SPrakash Surya * multilist_link_active() function (they all rely on 3310244781f1SPrakash Surya * the hash lock protecting concurrent insertions and 3311244781f1SPrakash Surya * removals). multilist_sublist_move_forward() was 3312244781f1SPrakash Surya * specifically implemented to ensure this is the case 3313244781f1SPrakash Surya * (only 'marker' will be removed and re-inserted). 3314244781f1SPrakash Surya */ 3315244781f1SPrakash Surya multilist_sublist_move_forward(mls, marker); 3316244781f1SPrakash Surya 3317244781f1SPrakash Surya /* 3318244781f1SPrakash Surya * The only case where the b_spa field should ever be 3319244781f1SPrakash Surya * zero, is the marker headers inserted by 3320244781f1SPrakash Surya * arc_evict_state(). It's possible for multiple threads 3321244781f1SPrakash Surya * to be calling arc_evict_state() concurrently (e.g. 3322244781f1SPrakash Surya * dsl_pool_close() and zio_inject_fault()), so we must 3323244781f1SPrakash Surya * skip any markers we see from these other threads. 332469962b56SMatthew Ahrens */ 3325244781f1SPrakash Surya if (hdr->b_spa == 0) 3326244781f1SPrakash Surya continue; 3327244781f1SPrakash Surya 3328244781f1SPrakash Surya /* we're only interested in evicting buffers of a certain spa */ 3329244781f1SPrakash Surya if (spa != 0 && hdr->b_spa != spa) { 3330244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_skip); 333169962b56SMatthew Ahrens continue; 333269962b56SMatthew Ahrens } 333369962b56SMatthew Ahrens 33347adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 33355ea40c06SBrendan Gregg - Sun Microsystems 3336244781f1SPrakash Surya /* 3337244781f1SPrakash Surya * We aren't calling this function from any code path 3338244781f1SPrakash Surya * that would already be holding a hash lock, so we're 3339244781f1SPrakash Surya * asserting on this assumption to be defensive in case 3340244781f1SPrakash Surya * this ever changes. Without this check, it would be 3341244781f1SPrakash Surya * possible to incorrectly increment arcstat_mutex_miss 3342244781f1SPrakash Surya * below (e.g. if the code changed such that we called 3343244781f1SPrakash Surya * this function with a hash lock held). 3344244781f1SPrakash Surya */ 3345244781f1SPrakash Surya ASSERT(!MUTEX_HELD(hash_lock)); 334644cb6abcSbmc 3347244781f1SPrakash Surya if (mutex_tryenter(hash_lock)) { 3348244781f1SPrakash Surya uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 3349244781f1SPrakash Surya mutex_exit(hash_lock); 3350fa9e4066Sahrens 3351244781f1SPrakash Surya bytes_evicted += evicted; 3352fa9e4066Sahrens 3353244781f1SPrakash Surya /* 3354244781f1SPrakash Surya * If evicted is zero, arc_evict_hdr() must have 3355244781f1SPrakash Surya * decided to skip this header, don't increment 3356244781f1SPrakash Surya * evict_count in this case. 3357244781f1SPrakash Surya */ 3358244781f1SPrakash Surya if (evicted != 0) 3359244781f1SPrakash Surya evict_count++; 336044cb6abcSbmc 3361244781f1SPrakash Surya /* 3362244781f1SPrakash Surya * If arc_size isn't overflowing, signal any 3363244781f1SPrakash Surya * threads that might happen to be waiting. 3364244781f1SPrakash Surya * 3365244781f1SPrakash Surya * For each header evicted, we wake up a single 3366244781f1SPrakash Surya * thread. If we used cv_broadcast, we could 3367244781f1SPrakash Surya * wake up "too many" threads causing arc_size 3368244781f1SPrakash Surya * to significantly overflow arc_c; since 3369770499e1SDan Kimmel * arc_get_data_impl() doesn't check for overflow 3370244781f1SPrakash Surya * when it's woken up (it doesn't because it's 3371244781f1SPrakash Surya * possible for the ARC to be overflowing while 3372244781f1SPrakash Surya * full of un-evictable buffers, and the 3373244781f1SPrakash Surya * function should proceed in this case). 3374244781f1SPrakash Surya * 3375244781f1SPrakash Surya * If threads are left sleeping, due to not 3376244781f1SPrakash Surya * using cv_broadcast, they will be woken up 3377244781f1SPrakash Surya * just before arc_reclaim_thread() sleeps. 3378244781f1SPrakash Surya */ 3379244781f1SPrakash Surya mutex_enter(&arc_reclaim_lock); 3380244781f1SPrakash Surya if (!arc_is_overflowing()) 3381244781f1SPrakash Surya cv_signal(&arc_reclaim_waiters_cv); 3382244781f1SPrakash Surya mutex_exit(&arc_reclaim_lock); 3383244781f1SPrakash Surya } else { 3384244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_mutex_miss); 3385244781f1SPrakash Surya } 3386244781f1SPrakash Surya } 3387f4d2e9e6Smaybee 3388244781f1SPrakash Surya multilist_sublist_unlock(mls); 338944cb6abcSbmc 3390244781f1SPrakash Surya return (bytes_evicted); 3391fa9e4066Sahrens } 3392fa9e4066Sahrens 3393fa9e4066Sahrens /* 3394244781f1SPrakash Surya * Evict buffers from the given arc state, until we've removed the 3395244781f1SPrakash Surya * specified number of bytes. Move the removed buffers to the 3396244781f1SPrakash Surya * appropriate evict state. 3397244781f1SPrakash Surya * 3398244781f1SPrakash Surya * This function makes a "best effort". It skips over any buffers 3399244781f1SPrakash Surya * it can't get a hash_lock on, and so, may not catch all candidates. 3400244781f1SPrakash Surya * It may also return without evicting as much space as requested. 3401244781f1SPrakash Surya * 3402244781f1SPrakash Surya * If bytes is specified using the special value ARC_EVICT_ALL, this 3403244781f1SPrakash Surya * will evict all available (i.e. unlocked and evictable) buffers from 3404244781f1SPrakash Surya * the given arc state; which is used by arc_flush(). 3405fa9e4066Sahrens */ 3406244781f1SPrakash Surya static uint64_t 3407244781f1SPrakash Surya arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 3408244781f1SPrakash Surya arc_buf_contents_t type) 3409fa9e4066Sahrens { 3410244781f1SPrakash Surya uint64_t total_evicted = 0; 341194c2d0ebSMatthew Ahrens multilist_t *ml = state->arcs_list[type]; 3412244781f1SPrakash Surya int num_sublists; 3413244781f1SPrakash Surya arc_buf_hdr_t **markers; 3414fa9e4066Sahrens 3415244781f1SPrakash Surya IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3416b802aa8cSSanjeev Bagewadi 3417244781f1SPrakash Surya num_sublists = multilist_get_num_sublists(ml); 3418b802aa8cSSanjeev Bagewadi 3419244781f1SPrakash Surya /* 3420244781f1SPrakash Surya * If we've tried to evict from each sublist, made some 3421244781f1SPrakash Surya * progress, but still have not hit the target number of bytes 3422244781f1SPrakash Surya * to evict, we want to keep trying. The markers allow us to 3423244781f1SPrakash Surya * pick up where we left off for each individual sublist, rather 3424244781f1SPrakash Surya * than starting from the tail each time. 3425244781f1SPrakash Surya */ 3426244781f1SPrakash Surya markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 3427244781f1SPrakash Surya for (int i = 0; i < num_sublists; i++) { 3428244781f1SPrakash Surya markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 342969962b56SMatthew Ahrens 343069962b56SMatthew Ahrens /* 3431244781f1SPrakash Surya * A b_spa of 0 is used to indicate that this header is 3432244781f1SPrakash Surya * a marker. This fact is used in arc_adjust_type() and 3433244781f1SPrakash Surya * arc_evict_state_impl(). 343469962b56SMatthew Ahrens */ 3435244781f1SPrakash Surya markers[i]->b_spa = 0; 3436fa94a07fSbrendan 3437244781f1SPrakash Surya multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3438244781f1SPrakash Surya multilist_sublist_insert_tail(mls, markers[i]); 3439244781f1SPrakash Surya multilist_sublist_unlock(mls); 3440244781f1SPrakash Surya } 3441fa94a07fSbrendan 3442244781f1SPrakash Surya /* 3443244781f1SPrakash Surya * While we haven't hit our target number of bytes to evict, or 3444244781f1SPrakash Surya * we're evicting all available buffers. 3445244781f1SPrakash Surya */ 3446244781f1SPrakash Surya while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 3447244781f1SPrakash Surya /* 3448244781f1SPrakash Surya * Start eviction using a randomly selected sublist, 3449244781f1SPrakash Surya * this is to try and evenly balance eviction across all 3450244781f1SPrakash Surya * sublists. Always starting at the same sublist 3451244781f1SPrakash Surya * (e.g. index 0) would cause evictions to favor certain 3452244781f1SPrakash Surya * sublists over others. 3453244781f1SPrakash Surya */ 3454244781f1SPrakash Surya int sublist_idx = multilist_get_random_index(ml); 3455244781f1SPrakash Surya uint64_t scan_evicted = 0; 3456244781f1SPrakash Surya 3457244781f1SPrakash Surya for (int i = 0; i < num_sublists; i++) { 3458244781f1SPrakash Surya uint64_t bytes_remaining; 3459244781f1SPrakash Surya uint64_t bytes_evicted; 3460244781f1SPrakash Surya 3461244781f1SPrakash Surya if (bytes == ARC_EVICT_ALL) 3462244781f1SPrakash Surya bytes_remaining = ARC_EVICT_ALL; 3463244781f1SPrakash Surya else if (total_evicted < bytes) 3464244781f1SPrakash Surya bytes_remaining = bytes - total_evicted; 3465244781f1SPrakash Surya else 3466fa9e4066Sahrens break; 3467244781f1SPrakash Surya 3468244781f1SPrakash Surya bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 3469244781f1SPrakash Surya markers[sublist_idx], spa, bytes_remaining); 3470244781f1SPrakash Surya 3471244781f1SPrakash Surya scan_evicted += bytes_evicted; 3472244781f1SPrakash Surya total_evicted += bytes_evicted; 3473244781f1SPrakash Surya 3474244781f1SPrakash Surya /* we've reached the end, wrap to the beginning */ 3475244781f1SPrakash Surya if (++sublist_idx >= num_sublists) 3476244781f1SPrakash Surya sublist_idx = 0; 3477244781f1SPrakash Surya } 3478244781f1SPrakash Surya 3479244781f1SPrakash Surya /* 3480244781f1SPrakash Surya * If we didn't evict anything during this scan, we have 3481244781f1SPrakash Surya * no reason to believe we'll evict more during another 3482244781f1SPrakash Surya * scan, so break the loop. 3483244781f1SPrakash Surya */ 3484244781f1SPrakash Surya if (scan_evicted == 0) { 3485244781f1SPrakash Surya /* This isn't possible, let's make that obvious */ 3486244781f1SPrakash Surya ASSERT3S(bytes, !=, 0); 3487244781f1SPrakash Surya 3488b802aa8cSSanjeev Bagewadi /* 3489244781f1SPrakash Surya * When bytes is ARC_EVICT_ALL, the only way to 3490244781f1SPrakash Surya * break the loop is when scan_evicted is zero. 3491244781f1SPrakash Surya * In that case, we actually have evicted enough, 3492244781f1SPrakash Surya * so we don't want to increment the kstat. 3493b802aa8cSSanjeev Bagewadi */ 3494244781f1SPrakash Surya if (bytes != ARC_EVICT_ALL) { 3495244781f1SPrakash Surya ASSERT3S(total_evicted, <, bytes); 3496244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_not_enough); 3497244781f1SPrakash Surya } 3498244781f1SPrakash Surya 3499244781f1SPrakash Surya break; 350069962b56SMatthew Ahrens } 3501244781f1SPrakash Surya } 3502244781f1SPrakash Surya 3503244781f1SPrakash Surya for (int i = 0; i < num_sublists; i++) { 3504244781f1SPrakash Surya multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3505244781f1SPrakash Surya multilist_sublist_remove(mls, markers[i]); 3506244781f1SPrakash Surya multilist_sublist_unlock(mls); 350769962b56SMatthew Ahrens 3508244781f1SPrakash Surya kmem_cache_free(hdr_full_cache, markers[i]); 3509fa9e4066Sahrens } 3510244781f1SPrakash Surya kmem_free(markers, sizeof (*markers) * num_sublists); 3511fa9e4066Sahrens 3512244781f1SPrakash Surya return (total_evicted); 3513244781f1SPrakash Surya } 3514244781f1SPrakash Surya 3515244781f1SPrakash Surya /* 3516244781f1SPrakash Surya * Flush all "evictable" data of the given type from the arc state 3517244781f1SPrakash Surya * specified. This will not evict any "active" buffers (i.e. referenced). 3518244781f1SPrakash Surya * 3519dcbf3bd6SGeorge Wilson * When 'retry' is set to B_FALSE, the function will make a single pass 3520244781f1SPrakash Surya * over the state and evict any buffers that it can. Since it doesn't 3521244781f1SPrakash Surya * continually retry the eviction, it might end up leaving some buffers 3522244781f1SPrakash Surya * in the ARC due to lock misses. 3523244781f1SPrakash Surya * 3524dcbf3bd6SGeorge Wilson * When 'retry' is set to B_TRUE, the function will continually retry the 3525244781f1SPrakash Surya * eviction until *all* evictable buffers have been removed from the 3526244781f1SPrakash Surya * state. As a result, if concurrent insertions into the state are 3527244781f1SPrakash Surya * allowed (e.g. if the ARC isn't shutting down), this function might 3528244781f1SPrakash Surya * wind up in an infinite loop, continually trying to evict buffers. 3529244781f1SPrakash Surya */ 3530244781f1SPrakash Surya static uint64_t 3531244781f1SPrakash Surya arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 3532244781f1SPrakash Surya boolean_t retry) 3533244781f1SPrakash Surya { 3534244781f1SPrakash Surya uint64_t evicted = 0; 3535244781f1SPrakash Surya 3536dcbf3bd6SGeorge Wilson while (refcount_count(&state->arcs_esize[type]) != 0) { 3537244781f1SPrakash Surya evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 3538244781f1SPrakash Surya 3539244781f1SPrakash Surya if (!retry) 3540244781f1SPrakash Surya break; 35410e8c6158Smaybee } 35420e8c6158Smaybee 3543244781f1SPrakash Surya return (evicted); 3544244781f1SPrakash Surya } 3545244781f1SPrakash Surya 3546244781f1SPrakash Surya /* 3547244781f1SPrakash Surya * Evict the specified number of bytes from the state specified, 3548244781f1SPrakash Surya * restricting eviction to the spa and type given. This function 3549244781f1SPrakash Surya * prevents us from trying to evict more from a state's list than 3550244781f1SPrakash Surya * is "evictable", and to skip evicting altogether when passed a 3551244781f1SPrakash Surya * negative value for "bytes". In contrast, arc_evict_state() will 3552244781f1SPrakash Surya * evict everything it can, when passed a negative value for "bytes". 3553244781f1SPrakash Surya */ 3554244781f1SPrakash Surya static uint64_t 3555244781f1SPrakash Surya arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3556244781f1SPrakash Surya arc_buf_contents_t type) 3557244781f1SPrakash Surya { 3558244781f1SPrakash Surya int64_t delta; 3559244781f1SPrakash Surya 3560dcbf3bd6SGeorge Wilson if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) { 3561dcbf3bd6SGeorge Wilson delta = MIN(refcount_count(&state->arcs_esize[type]), bytes); 3562244781f1SPrakash Surya return (arc_evict_state(state, spa, delta, type)); 3563fa9e4066Sahrens } 3564fa9e4066Sahrens 3565244781f1SPrakash Surya return (0); 3566fa9e4066Sahrens } 3567fa9e4066Sahrens 3568244781f1SPrakash Surya /* 3569244781f1SPrakash Surya * Evict metadata buffers from the cache, such that arc_meta_used is 3570244781f1SPrakash Surya * capped by the arc_meta_limit tunable. 3571244781f1SPrakash Surya */ 3572244781f1SPrakash Surya static uint64_t 3573244781f1SPrakash Surya arc_adjust_meta(void) 3574244781f1SPrakash Surya { 3575244781f1SPrakash Surya uint64_t total_evicted = 0; 3576244781f1SPrakash Surya int64_t target; 3577244781f1SPrakash Surya 3578244781f1SPrakash Surya /* 3579244781f1SPrakash Surya * If we're over the meta limit, we want to evict enough 3580244781f1SPrakash Surya * metadata to get back under the meta limit. We don't want to 3581244781f1SPrakash Surya * evict so much that we drop the MRU below arc_p, though. If 3582244781f1SPrakash Surya * we're over the meta limit more than we're over arc_p, we 3583244781f1SPrakash Surya * evict some from the MRU here, and some from the MFU below. 3584244781f1SPrakash Surya */ 3585244781f1SPrakash Surya target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 35862fd872a7SPrakash Surya (int64_t)(refcount_count(&arc_anon->arcs_size) + 35872fd872a7SPrakash Surya refcount_count(&arc_mru->arcs_size) - arc_p)); 3588244781f1SPrakash Surya 3589244781f1SPrakash Surya total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3590244781f1SPrakash Surya 3591244781f1SPrakash Surya /* 3592244781f1SPrakash Surya * Similar to the above, we want to evict enough bytes to get us 3593244781f1SPrakash Surya * below the meta limit, but not so much as to drop us below the 35945602294fSDan Kimmel * space allotted to the MFU (which is defined as arc_c - arc_p). 3595244781f1SPrakash Surya */ 3596244781f1SPrakash Surya target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 35972fd872a7SPrakash Surya (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3598244781f1SPrakash Surya 3599244781f1SPrakash Surya total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3600244781f1SPrakash Surya 3601244781f1SPrakash Surya return (total_evicted); 3602244781f1SPrakash Surya } 3603244781f1SPrakash Surya 3604244781f1SPrakash Surya /* 3605244781f1SPrakash Surya * Return the type of the oldest buffer in the given arc state 3606244781f1SPrakash Surya * 3607244781f1SPrakash Surya * This function will select a random sublist of type ARC_BUFC_DATA and 3608244781f1SPrakash Surya * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3609244781f1SPrakash Surya * is compared, and the type which contains the "older" buffer will be 3610244781f1SPrakash Surya * returned. 3611244781f1SPrakash Surya */ 3612244781f1SPrakash Surya static arc_buf_contents_t 3613244781f1SPrakash Surya arc_adjust_type(arc_state_t *state) 3614244781f1SPrakash Surya { 361594c2d0ebSMatthew Ahrens multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA]; 361694c2d0ebSMatthew Ahrens multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA]; 3617244781f1SPrakash Surya int data_idx = multilist_get_random_index(data_ml); 3618244781f1SPrakash Surya int meta_idx = multilist_get_random_index(meta_ml); 3619244781f1SPrakash Surya multilist_sublist_t *data_mls; 3620244781f1SPrakash Surya multilist_sublist_t *meta_mls; 3621244781f1SPrakash Surya arc_buf_contents_t type; 3622244781f1SPrakash Surya arc_buf_hdr_t *data_hdr; 3623244781f1SPrakash Surya arc_buf_hdr_t *meta_hdr; 3624244781f1SPrakash Surya 3625244781f1SPrakash Surya /* 3626244781f1SPrakash Surya * We keep the sublist lock until we're finished, to prevent 3627244781f1SPrakash Surya * the headers from being destroyed via arc_evict_state(). 3628244781f1SPrakash Surya */ 3629244781f1SPrakash Surya data_mls = multilist_sublist_lock(data_ml, data_idx); 3630244781f1SPrakash Surya meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3631244781f1SPrakash Surya 3632244781f1SPrakash Surya /* 3633244781f1SPrakash Surya * These two loops are to ensure we skip any markers that 3634244781f1SPrakash Surya * might be at the tail of the lists due to arc_evict_state(). 3635244781f1SPrakash Surya */ 3636244781f1SPrakash Surya 3637244781f1SPrakash Surya for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3638244781f1SPrakash Surya data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3639244781f1SPrakash Surya if (data_hdr->b_spa != 0) 3640244781f1SPrakash Surya break; 3641244781f1SPrakash Surya } 3642244781f1SPrakash Surya 3643244781f1SPrakash Surya for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3644244781f1SPrakash Surya meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3645244781f1SPrakash Surya if (meta_hdr->b_spa != 0) 3646244781f1SPrakash Surya break; 3647244781f1SPrakash Surya } 3648244781f1SPrakash Surya 3649244781f1SPrakash Surya if (data_hdr == NULL && meta_hdr == NULL) { 3650244781f1SPrakash Surya type = ARC_BUFC_DATA; 3651244781f1SPrakash Surya } else if (data_hdr == NULL) { 3652244781f1SPrakash Surya ASSERT3P(meta_hdr, !=, NULL); 3653244781f1SPrakash Surya type = ARC_BUFC_METADATA; 3654244781f1SPrakash Surya } else if (meta_hdr == NULL) { 3655244781f1SPrakash Surya ASSERT3P(data_hdr, !=, NULL); 3656244781f1SPrakash Surya type = ARC_BUFC_DATA; 3657244781f1SPrakash Surya } else { 3658244781f1SPrakash Surya ASSERT3P(data_hdr, !=, NULL); 3659244781f1SPrakash Surya ASSERT3P(meta_hdr, !=, NULL); 3660244781f1SPrakash Surya 3661244781f1SPrakash Surya /* The headers can't be on the sublist without an L1 header */ 3662244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(data_hdr)); 3663244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3664244781f1SPrakash Surya 3665244781f1SPrakash Surya if (data_hdr->b_l1hdr.b_arc_access < 3666244781f1SPrakash Surya meta_hdr->b_l1hdr.b_arc_access) { 3667244781f1SPrakash Surya type = ARC_BUFC_DATA; 3668244781f1SPrakash Surya } else { 3669244781f1SPrakash Surya type = ARC_BUFC_METADATA; 3670244781f1SPrakash Surya } 3671244781f1SPrakash Surya } 3672244781f1SPrakash Surya 3673244781f1SPrakash Surya multilist_sublist_unlock(meta_mls); 3674244781f1SPrakash Surya multilist_sublist_unlock(data_mls); 3675244781f1SPrakash Surya 3676244781f1SPrakash Surya return (type); 3677244781f1SPrakash Surya } 3678244781f1SPrakash Surya 3679244781f1SPrakash Surya /* 3680244781f1SPrakash Surya * Evict buffers from the cache, such that arc_size is capped by arc_c. 3681244781f1SPrakash Surya */ 3682244781f1SPrakash Surya static uint64_t 3683fa9e4066Sahrens arc_adjust(void) 3684fa9e4066Sahrens { 3685244781f1SPrakash Surya uint64_t total_evicted = 0; 3686244781f1SPrakash Surya uint64_t bytes; 3687244781f1SPrakash Surya int64_t target; 3688fa9e4066Sahrens 36895a98e54bSBrendan Gregg - Sun Microsystems /* 3690244781f1SPrakash Surya * If we're over arc_meta_limit, we want to correct that before 3691244781f1SPrakash Surya * potentially evicting data buffers below. 36925a98e54bSBrendan Gregg - Sun Microsystems */ 3693244781f1SPrakash Surya total_evicted += arc_adjust_meta(); 36945a98e54bSBrendan Gregg - Sun Microsystems 3695244781f1SPrakash Surya /* 3696244781f1SPrakash Surya * Adjust MRU size 3697244781f1SPrakash Surya * 3698244781f1SPrakash Surya * If we're over the target cache size, we want to evict enough 3699244781f1SPrakash Surya * from the list to get back to our target size. We don't want 3700244781f1SPrakash Surya * to evict too much from the MRU, such that it drops below 3701244781f1SPrakash Surya * arc_p. So, if we're over our target cache size more than 3702244781f1SPrakash Surya * the MRU is over arc_p, we'll evict enough to get back to 3703244781f1SPrakash Surya * arc_p here, and then evict more from the MFU below. 3704244781f1SPrakash Surya */ 3705244781f1SPrakash Surya target = MIN((int64_t)(arc_size - arc_c), 37062fd872a7SPrakash Surya (int64_t)(refcount_count(&arc_anon->arcs_size) + 37072fd872a7SPrakash Surya refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 3708fa9e4066Sahrens 3709244781f1SPrakash Surya /* 3710244781f1SPrakash Surya * If we're below arc_meta_min, always prefer to evict data. 3711244781f1SPrakash Surya * Otherwise, try to satisfy the requested number of bytes to 3712244781f1SPrakash Surya * evict from the type which contains older buffers; in an 3713244781f1SPrakash Surya * effort to keep newer buffers in the cache regardless of their 3714244781f1SPrakash Surya * type. If we cannot satisfy the number of bytes from this 3715244781f1SPrakash Surya * type, spill over into the next type. 3716244781f1SPrakash Surya */ 3717244781f1SPrakash Surya if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3718244781f1SPrakash Surya arc_meta_used > arc_meta_min) { 3719244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3720244781f1SPrakash Surya total_evicted += bytes; 37210e8c6158Smaybee 3722244781f1SPrakash Surya /* 3723244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3724244781f1SPrakash Surya * metadata, we try to get the rest from data. 3725244781f1SPrakash Surya */ 3726244781f1SPrakash Surya target -= bytes; 3727244781f1SPrakash Surya 3728244781f1SPrakash Surya total_evicted += 3729244781f1SPrakash Surya arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3730244781f1SPrakash Surya } else { 3731244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3732244781f1SPrakash Surya total_evicted += bytes; 3733244781f1SPrakash Surya 3734244781f1SPrakash Surya /* 3735244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3736244781f1SPrakash Surya * data, we try to get the rest from metadata. 3737244781f1SPrakash Surya */ 3738244781f1SPrakash Surya target -= bytes; 3739244781f1SPrakash Surya 3740244781f1SPrakash Surya total_evicted += 3741244781f1SPrakash Surya arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3742fa9e4066Sahrens } 3743fa9e4066Sahrens 37445a98e54bSBrendan Gregg - Sun Microsystems /* 37455a98e54bSBrendan Gregg - Sun Microsystems * Adjust MFU size 3746244781f1SPrakash Surya * 3747244781f1SPrakash Surya * Now that we've tried to evict enough from the MRU to get its 3748244781f1SPrakash Surya * size back to arc_p, if we're still above the target cache 3749244781f1SPrakash Surya * size, we evict the rest from the MFU. 37505a98e54bSBrendan Gregg - Sun Microsystems */ 3751244781f1SPrakash Surya target = arc_size - arc_c; 3752fa9e4066Sahrens 375331c46cf2SAlek Pinchuk if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3754244781f1SPrakash Surya arc_meta_used > arc_meta_min) { 3755244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3756244781f1SPrakash Surya total_evicted += bytes; 37575a98e54bSBrendan Gregg - Sun Microsystems 3758244781f1SPrakash Surya /* 3759244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3760244781f1SPrakash Surya * metadata, we try to get the rest from data. 3761244781f1SPrakash Surya */ 3762244781f1SPrakash Surya target -= bytes; 3763244781f1SPrakash Surya 3764244781f1SPrakash Surya total_evicted += 3765244781f1SPrakash Surya arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3766244781f1SPrakash Surya } else { 3767244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3768244781f1SPrakash Surya total_evicted += bytes; 3769244781f1SPrakash Surya 3770244781f1SPrakash Surya /* 3771244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3772244781f1SPrakash Surya * data, we try to get the rest from data. 3773244781f1SPrakash Surya */ 3774244781f1SPrakash Surya target -= bytes; 3775fa9e4066Sahrens 3776244781f1SPrakash Surya total_evicted += 3777244781f1SPrakash Surya arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 37785a98e54bSBrendan Gregg - Sun Microsystems } 3779fa9e4066Sahrens 37805a98e54bSBrendan Gregg - Sun Microsystems /* 37815a98e54bSBrendan Gregg - Sun Microsystems * Adjust ghost lists 3782244781f1SPrakash Surya * 3783244781f1SPrakash Surya * In addition to the above, the ARC also defines target values 3784244781f1SPrakash Surya * for the ghost lists. The sum of the mru list and mru ghost 3785244781f1SPrakash Surya * list should never exceed the target size of the cache, and 3786244781f1SPrakash Surya * the sum of the mru list, mfu list, mru ghost list, and mfu 3787244781f1SPrakash Surya * ghost list should never exceed twice the target size of the 3788244781f1SPrakash Surya * cache. The following logic enforces these limits on the ghost 3789244781f1SPrakash Surya * caches, and evicts from them as needed. 37905a98e54bSBrendan Gregg - Sun Microsystems */ 37912fd872a7SPrakash Surya target = refcount_count(&arc_mru->arcs_size) + 37922fd872a7SPrakash Surya refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3793fa9e4066Sahrens 3794244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3795244781f1SPrakash Surya total_evicted += bytes; 3796fa9e4066Sahrens 3797244781f1SPrakash Surya target -= bytes; 37980e8c6158Smaybee 3799244781f1SPrakash Surya total_evicted += 3800244781f1SPrakash Surya arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 38015a98e54bSBrendan Gregg - Sun Microsystems 3802244781f1SPrakash Surya /* 3803244781f1SPrakash Surya * We assume the sum of the mru list and mfu list is less than 3804244781f1SPrakash Surya * or equal to arc_c (we enforced this above), which means we 3805244781f1SPrakash Surya * can use the simpler of the two equations below: 3806244781f1SPrakash Surya * 3807244781f1SPrakash Surya * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3808244781f1SPrakash Surya * mru ghost + mfu ghost <= arc_c 3809244781f1SPrakash Surya */ 38102fd872a7SPrakash Surya target = refcount_count(&arc_mru_ghost->arcs_size) + 38112fd872a7SPrakash Surya refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3812244781f1SPrakash Surya 3813244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3814244781f1SPrakash Surya total_evicted += bytes; 3815244781f1SPrakash Surya 3816244781f1SPrakash Surya target -= bytes; 3817244781f1SPrakash Surya 3818244781f1SPrakash Surya total_evicted += 3819244781f1SPrakash Surya arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3820244781f1SPrakash Surya 3821244781f1SPrakash Surya return (total_evicted); 3822fa9e4066Sahrens } 3823fa9e4066Sahrens 3824fa9e4066Sahrens void 3825244781f1SPrakash Surya arc_flush(spa_t *spa, boolean_t retry) 3826fa9e4066Sahrens { 3827ac05c741SMark Maybee uint64_t guid = 0; 3828ac05c741SMark Maybee 3829244781f1SPrakash Surya /* 3830dcbf3bd6SGeorge Wilson * If retry is B_TRUE, a spa must not be specified since we have 3831244781f1SPrakash Surya * no good way to determine if all of a spa's buffers have been 3832244781f1SPrakash Surya * evicted from an arc state. 3833244781f1SPrakash Surya */ 3834244781f1SPrakash Surya ASSERT(!retry || spa == 0); 3835244781f1SPrakash Surya 383689c86e32SChris Williamson if (spa != NULL) 3837e9103aaeSGarrett D'Amore guid = spa_load_guid(spa); 3838ac05c741SMark Maybee 3839244781f1SPrakash Surya (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3840244781f1SPrakash Surya (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3841244781f1SPrakash Surya 3842244781f1SPrakash Surya (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3843244781f1SPrakash Surya (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3844874395d5Smaybee 3845244781f1SPrakash Surya (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3846244781f1SPrakash Surya (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3847244781f1SPrakash Surya 3848244781f1SPrakash Surya (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3849244781f1SPrakash Surya (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3850fa9e4066Sahrens } 3851fa9e4066Sahrens 3852fa9e4066Sahrens void 38532ec99e3eSMatthew Ahrens arc_shrink(int64_t to_free) 3854fa9e4066Sahrens { 385544cb6abcSbmc if (arc_c > arc_c_min) { 3856fa9e4066Sahrens 385744cb6abcSbmc if (arc_c > arc_c_min + to_free) 385844cb6abcSbmc atomic_add_64(&arc_c, -to_free); 385949e3519aSmaybee else 386044cb6abcSbmc arc_c = arc_c_min; 386144cb6abcSbmc 386244cb6abcSbmc atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 386344cb6abcSbmc if (arc_c > arc_size) 386444cb6abcSbmc arc_c = MAX(arc_size, arc_c_min); 386544cb6abcSbmc if (arc_p > arc_c) 386644cb6abcSbmc arc_p = (arc_c >> 1); 386744cb6abcSbmc ASSERT(arc_c >= arc_c_min); 386844cb6abcSbmc ASSERT((int64_t)arc_p >= 0); 386949e3519aSmaybee } 3870fa9e4066Sahrens 387144cb6abcSbmc if (arc_size > arc_c) 3872244781f1SPrakash Surya (void) arc_adjust(); 3873fa9e4066Sahrens } 3874fa9e4066Sahrens 38752ec99e3eSMatthew Ahrens typedef enum free_memory_reason_t { 38762ec99e3eSMatthew Ahrens FMR_UNKNOWN, 38772ec99e3eSMatthew Ahrens FMR_NEEDFREE, 38782ec99e3eSMatthew Ahrens FMR_LOTSFREE, 38792ec99e3eSMatthew Ahrens FMR_SWAPFS_MINFREE, 38802ec99e3eSMatthew Ahrens FMR_PAGES_PP_MAXIMUM, 38812ec99e3eSMatthew Ahrens FMR_HEAP_ARENA, 38822ec99e3eSMatthew Ahrens FMR_ZIO_ARENA, 38832ec99e3eSMatthew Ahrens } free_memory_reason_t; 38842ec99e3eSMatthew Ahrens 38852ec99e3eSMatthew Ahrens int64_t last_free_memory; 38862ec99e3eSMatthew Ahrens free_memory_reason_t last_free_reason; 38872ec99e3eSMatthew Ahrens 388894dd93aeSGeorge Wilson /* 38892ec99e3eSMatthew Ahrens * Additional reserve of pages for pp_reserve. 389094dd93aeSGeorge Wilson */ 38912ec99e3eSMatthew Ahrens int64_t arc_pages_pp_reserve = 64; 3892fa9e4066Sahrens 38932ec99e3eSMatthew Ahrens /* 38942ec99e3eSMatthew Ahrens * Additional reserve of pages for swapfs. 38952ec99e3eSMatthew Ahrens */ 38962ec99e3eSMatthew Ahrens int64_t arc_swapfs_reserve = 64; 38973cff2f43Sstans 38982ec99e3eSMatthew Ahrens /* 38992ec99e3eSMatthew Ahrens * Return the amount of memory that can be consumed before reclaim will be 39002ec99e3eSMatthew Ahrens * needed. Positive if there is sufficient free memory, negative indicates 39012ec99e3eSMatthew Ahrens * the amount of memory that needs to be freed up. 39022ec99e3eSMatthew Ahrens */ 39032ec99e3eSMatthew Ahrens static int64_t 39042ec99e3eSMatthew Ahrens arc_available_memory(void) 39052ec99e3eSMatthew Ahrens { 39062ec99e3eSMatthew Ahrens int64_t lowest = INT64_MAX; 39072ec99e3eSMatthew Ahrens int64_t n; 39082ec99e3eSMatthew Ahrens free_memory_reason_t r = FMR_UNKNOWN; 39093cff2f43Sstans 39102ec99e3eSMatthew Ahrens #ifdef _KERNEL 39112ec99e3eSMatthew Ahrens if (needfree > 0) { 39122ec99e3eSMatthew Ahrens n = PAGESIZE * (-needfree); 39132ec99e3eSMatthew Ahrens if (n < lowest) { 39142ec99e3eSMatthew Ahrens lowest = n; 39152ec99e3eSMatthew Ahrens r = FMR_NEEDFREE; 39162ec99e3eSMatthew Ahrens } 39172ec99e3eSMatthew Ahrens } 3918fa9e4066Sahrens 3919fa9e4066Sahrens /* 3920fa9e4066Sahrens * check that we're out of range of the pageout scanner. It starts to 3921fa9e4066Sahrens * schedule paging if freemem is less than lotsfree and needfree. 3922fa9e4066Sahrens * lotsfree is the high-water mark for pageout, and needfree is the 3923fa9e4066Sahrens * number of needed free pages. We add extra pages here to make sure 3924fa9e4066Sahrens * the scanner doesn't start up while we're freeing memory. 3925fa9e4066Sahrens */ 39262ec99e3eSMatthew Ahrens n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 39272ec99e3eSMatthew Ahrens if (n < lowest) { 39282ec99e3eSMatthew Ahrens lowest = n; 39292ec99e3eSMatthew Ahrens r = FMR_LOTSFREE; 39302ec99e3eSMatthew Ahrens } 3931fa9e4066Sahrens 3932fa9e4066Sahrens /* 3933fa9e4066Sahrens * check to make sure that swapfs has enough space so that anon 3934fa94a07fSbrendan * reservations can still succeed. anon_resvmem() checks that the 3935fa9e4066Sahrens * availrmem is greater than swapfs_minfree, and the number of reserved 3936fa9e4066Sahrens * swap pages. We also add a bit of extra here just to prevent 3937fa9e4066Sahrens * circumstances from getting really dire. 3938fa9e4066Sahrens */ 39392ec99e3eSMatthew Ahrens n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 39402ec99e3eSMatthew Ahrens desfree - arc_swapfs_reserve); 39412ec99e3eSMatthew Ahrens if (n < lowest) { 39422ec99e3eSMatthew Ahrens lowest = n; 39432ec99e3eSMatthew Ahrens r = FMR_SWAPFS_MINFREE; 39442ec99e3eSMatthew Ahrens } 39452ec99e3eSMatthew Ahrens 3946fa9e4066Sahrens 3947cf746768SBryan Cantrill /* 3948cf746768SBryan Cantrill * Check that we have enough availrmem that memory locking (e.g., via 3949cf746768SBryan Cantrill * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3950cf746768SBryan Cantrill * stores the number of pages that cannot be locked; when availrmem 3951cf746768SBryan Cantrill * drops below pages_pp_maximum, page locking mechanisms such as 3952cf746768SBryan Cantrill * page_pp_lock() will fail.) 3953cf746768SBryan Cantrill */ 39542ec99e3eSMatthew Ahrens n = PAGESIZE * (availrmem - pages_pp_maximum - 39552ec99e3eSMatthew Ahrens arc_pages_pp_reserve); 39562ec99e3eSMatthew Ahrens if (n < lowest) { 39572ec99e3eSMatthew Ahrens lowest = n; 39582ec99e3eSMatthew Ahrens r = FMR_PAGES_PP_MAXIMUM; 39592ec99e3eSMatthew Ahrens } 3960cf746768SBryan Cantrill 39615dc8af33Smaybee #if defined(__i386) 3962fa9e4066Sahrens /* 3963fa9e4066Sahrens * If we're on an i386 platform, it's possible that we'll exhaust the 3964fa9e4066Sahrens * kernel heap space before we ever run out of available physical 3965fa9e4066Sahrens * memory. Most checks of the size of the heap_area compare against 3966fa9e4066Sahrens * tune.t_minarmem, which is the minimum available real memory that we 3967fa9e4066Sahrens * can have in the system. However, this is generally fixed at 25 pages 3968fa9e4066Sahrens * which is so low that it's useless. In this comparison, we seek to 3969fa9e4066Sahrens * calculate the total heap-size, and reclaim if more than 3/4ths of the 3970fa94a07fSbrendan * heap is allocated. (Or, in the calculation, if less than 1/4th is 3971fa9e4066Sahrens * free) 3972fa9e4066Sahrens */ 39730dd053d7SPrakash Surya n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 39742ec99e3eSMatthew Ahrens (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 39752ec99e3eSMatthew Ahrens if (n < lowest) { 39762ec99e3eSMatthew Ahrens lowest = n; 39772ec99e3eSMatthew Ahrens r = FMR_HEAP_ARENA; 39782ec99e3eSMatthew Ahrens } 3979fa9e4066Sahrens #endif 3980fa9e4066Sahrens 398194dd93aeSGeorge Wilson /* 398294dd93aeSGeorge Wilson * If zio data pages are being allocated out of a separate heap segment, 398394dd93aeSGeorge Wilson * then enforce that the size of available vmem for this arena remains 39840dd053d7SPrakash Surya * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free. 398594dd93aeSGeorge Wilson * 39860dd053d7SPrakash Surya * Note that reducing the arc_zio_arena_free_shift keeps more virtual 39870dd053d7SPrakash Surya * memory (in the zio_arena) free, which can avoid memory 39880dd053d7SPrakash Surya * fragmentation issues. 398994dd93aeSGeorge Wilson */ 39902ec99e3eSMatthew Ahrens if (zio_arena != NULL) { 39910dd053d7SPrakash Surya n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 39920dd053d7SPrakash Surya (vmem_size(zio_arena, VMEM_ALLOC) >> 39930dd053d7SPrakash Surya arc_zio_arena_free_shift); 39942ec99e3eSMatthew Ahrens if (n < lowest) { 39952ec99e3eSMatthew Ahrens lowest = n; 39962ec99e3eSMatthew Ahrens r = FMR_ZIO_ARENA; 39972ec99e3eSMatthew Ahrens } 39982ec99e3eSMatthew Ahrens } 3999fa9e4066Sahrens #else 40002ec99e3eSMatthew Ahrens /* Every 100 calls, free a small amount */ 4001fa9e4066Sahrens if (spa_get_random(100) == 0) 40022ec99e3eSMatthew Ahrens lowest = -1024; 4003fa9e4066Sahrens #endif 40042ec99e3eSMatthew Ahrens 40052ec99e3eSMatthew Ahrens last_free_memory = lowest; 40062ec99e3eSMatthew Ahrens last_free_reason = r; 40072ec99e3eSMatthew Ahrens 40082ec99e3eSMatthew Ahrens return (lowest); 40092ec99e3eSMatthew Ahrens } 40102ec99e3eSMatthew Ahrens 40112ec99e3eSMatthew Ahrens 40122ec99e3eSMatthew Ahrens /* 40132ec99e3eSMatthew Ahrens * Determine if the system is under memory pressure and is asking 4014dcbf3bd6SGeorge Wilson * to reclaim memory. A return value of B_TRUE indicates that the system 40152ec99e3eSMatthew Ahrens * is under memory pressure and that the arc should adjust accordingly. 40162ec99e3eSMatthew Ahrens */ 40172ec99e3eSMatthew Ahrens static boolean_t 40182ec99e3eSMatthew Ahrens arc_reclaim_needed(void) 40192ec99e3eSMatthew Ahrens { 40202ec99e3eSMatthew Ahrens return (arc_available_memory() < 0); 4021fa9e4066Sahrens } 4022fa9e4066Sahrens 4023fa9e4066Sahrens static void 40242ec99e3eSMatthew Ahrens arc_kmem_reap_now(void) 4025fa9e4066Sahrens { 4026fa9e4066Sahrens size_t i; 4027fa9e4066Sahrens kmem_cache_t *prev_cache = NULL; 4028ad23a2dbSjohansen kmem_cache_t *prev_data_cache = NULL; 4029fa9e4066Sahrens extern kmem_cache_t *zio_buf_cache[]; 4030ad23a2dbSjohansen extern kmem_cache_t *zio_data_buf_cache[]; 403183803b51SGeorge Wilson extern kmem_cache_t *range_seg_cache; 4032770499e1SDan Kimmel extern kmem_cache_t *abd_chunk_cache; 4033fa9e4066Sahrens 4034033f9833Sek #ifdef _KERNEL 40350e8c6158Smaybee if (arc_meta_used >= arc_meta_limit) { 40360e8c6158Smaybee /* 40370e8c6158Smaybee * We are exceeding our meta-data cache limit. 40380e8c6158Smaybee * Purge some DNLC entries to release holds on meta-data. 40390e8c6158Smaybee */ 40400e8c6158Smaybee dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 40410e8c6158Smaybee } 40425dc8af33Smaybee #if defined(__i386) 40435dc8af33Smaybee /* 40445dc8af33Smaybee * Reclaim unused memory from all kmem caches. 40455dc8af33Smaybee */ 40465dc8af33Smaybee kmem_reap(); 40475dc8af33Smaybee #endif 4048033f9833Sek #endif 4049033f9833Sek 4050fa9e4066Sahrens for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 4051fa9e4066Sahrens if (zio_buf_cache[i] != prev_cache) { 4052fa9e4066Sahrens prev_cache = zio_buf_cache[i]; 4053fa9e4066Sahrens kmem_cache_reap_now(zio_buf_cache[i]); 4054fa9e4066Sahrens } 4055ad23a2dbSjohansen if (zio_data_buf_cache[i] != prev_data_cache) { 4056ad23a2dbSjohansen prev_data_cache = zio_data_buf_cache[i]; 4057ad23a2dbSjohansen kmem_cache_reap_now(zio_data_buf_cache[i]); 4058ad23a2dbSjohansen } 4059fa9e4066Sahrens } 4060770499e1SDan Kimmel kmem_cache_reap_now(abd_chunk_cache); 4061ea8dc4b6Seschrock kmem_cache_reap_now(buf_cache); 406289c86e32SChris Williamson kmem_cache_reap_now(hdr_full_cache); 406389c86e32SChris Williamson kmem_cache_reap_now(hdr_l2only_cache); 406483803b51SGeorge Wilson kmem_cache_reap_now(range_seg_cache); 406594dd93aeSGeorge Wilson 40662ec99e3eSMatthew Ahrens if (zio_arena != NULL) { 40672ec99e3eSMatthew Ahrens /* 40682ec99e3eSMatthew Ahrens * Ask the vmem arena to reclaim unused memory from its 40692ec99e3eSMatthew Ahrens * quantum caches. 40702ec99e3eSMatthew Ahrens */ 407194dd93aeSGeorge Wilson vmem_qcache_reap(zio_arena); 40722ec99e3eSMatthew Ahrens } 4073fa9e4066Sahrens } 4074fa9e4066Sahrens 4075244781f1SPrakash Surya /* 4076770499e1SDan Kimmel * Threads can block in arc_get_data_impl() waiting for this thread to evict 4077244781f1SPrakash Surya * enough data and signal them to proceed. When this happens, the threads in 4078770499e1SDan Kimmel * arc_get_data_impl() are sleeping while holding the hash lock for their 4079244781f1SPrakash Surya * particular arc header. Thus, we must be careful to never sleep on a 4080244781f1SPrakash Surya * hash lock in this thread. This is to prevent the following deadlock: 4081244781f1SPrakash Surya * 4082770499e1SDan Kimmel * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L", 4083244781f1SPrakash Surya * waiting for the reclaim thread to signal it. 4084244781f1SPrakash Surya * 4085244781f1SPrakash Surya * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 4086244781f1SPrakash Surya * fails, and goes to sleep forever. 4087244781f1SPrakash Surya * 4088244781f1SPrakash Surya * This possible deadlock is avoided by always acquiring a hash lock 4089244781f1SPrakash Surya * using mutex_tryenter() from arc_reclaim_thread(). 4090244781f1SPrakash Surya */ 40913f7978d0SAlan Somers /* ARGSUSED */ 4092fa9e4066Sahrens static void 40933f7978d0SAlan Somers arc_reclaim_thread(void *unused) 4094fa9e4066Sahrens { 4095a8f6344fSEli Rosenthal hrtime_t growtime = 0; 4096fa9e4066Sahrens callb_cpr_t cpr; 4097fa9e4066Sahrens 4098244781f1SPrakash Surya CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 4099fa9e4066Sahrens 4100244781f1SPrakash Surya mutex_enter(&arc_reclaim_lock); 4101244781f1SPrakash Surya while (!arc_reclaim_thread_exit) { 4102244781f1SPrakash Surya uint64_t evicted = 0; 4103244781f1SPrakash Surya 4104dcbf3bd6SGeorge Wilson /* 4105dcbf3bd6SGeorge Wilson * This is necessary in order for the mdb ::arc dcmd to 4106dcbf3bd6SGeorge Wilson * show up to date information. Since the ::arc command 4107dcbf3bd6SGeorge Wilson * does not call the kstat's update function, without 4108dcbf3bd6SGeorge Wilson * this call, the command may show stale stats for the 4109dcbf3bd6SGeorge Wilson * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 4110dcbf3bd6SGeorge Wilson * with this change, the data might be up to 1 second 4111dcbf3bd6SGeorge Wilson * out of date; but that should suffice. The arc_state_t 4112dcbf3bd6SGeorge Wilson * structures can be queried directly if more accurate 4113dcbf3bd6SGeorge Wilson * information is needed. 4114dcbf3bd6SGeorge Wilson */ 4115dcbf3bd6SGeorge Wilson if (arc_ksp != NULL) 4116dcbf3bd6SGeorge Wilson arc_ksp->ks_update(arc_ksp, KSTAT_READ); 4117dcbf3bd6SGeorge Wilson 4118244781f1SPrakash Surya mutex_exit(&arc_reclaim_lock); 4119244781f1SPrakash Surya 4120405a5a0fSMatthew Ahrens /* 4121405a5a0fSMatthew Ahrens * We call arc_adjust() before (possibly) calling 4122405a5a0fSMatthew Ahrens * arc_kmem_reap_now(), so that we can wake up 4123770499e1SDan Kimmel * arc_get_data_impl() sooner. 4124405a5a0fSMatthew Ahrens */ 4125405a5a0fSMatthew Ahrens evicted = arc_adjust(); 4126405a5a0fSMatthew Ahrens 4127405a5a0fSMatthew Ahrens int64_t free_memory = arc_available_memory(); 41282ec99e3eSMatthew Ahrens if (free_memory < 0) { 4129fa9e4066Sahrens 41302ec99e3eSMatthew Ahrens arc_no_grow = B_TRUE; 41312ec99e3eSMatthew Ahrens arc_warm = B_TRUE; 4132fa9e4066Sahrens 41332ec99e3eSMatthew Ahrens /* 41342ec99e3eSMatthew Ahrens * Wait at least zfs_grow_retry (default 60) seconds 41352ec99e3eSMatthew Ahrens * before considering growing. 41362ec99e3eSMatthew Ahrens */ 4137a8f6344fSEli Rosenthal growtime = gethrtime() + SEC2NSEC(arc_grow_retry); 4138fa9e4066Sahrens 41392ec99e3eSMatthew Ahrens arc_kmem_reap_now(); 4140fa9e4066Sahrens 41412ec99e3eSMatthew Ahrens /* 41422ec99e3eSMatthew Ahrens * If we are still low on memory, shrink the ARC 41432ec99e3eSMatthew Ahrens * so that we have arc_shrink_min free space. 41442ec99e3eSMatthew Ahrens */ 41452ec99e3eSMatthew Ahrens free_memory = arc_available_memory(); 41462ec99e3eSMatthew Ahrens 41472ec99e3eSMatthew Ahrens int64_t to_free = 41482ec99e3eSMatthew Ahrens (arc_c >> arc_shrink_shift) - free_memory; 41492ec99e3eSMatthew Ahrens if (to_free > 0) { 41502ec99e3eSMatthew Ahrens #ifdef _KERNEL 41512ec99e3eSMatthew Ahrens to_free = MAX(to_free, ptob(needfree)); 41522ec99e3eSMatthew Ahrens #endif 41532ec99e3eSMatthew Ahrens arc_shrink(to_free); 41542ec99e3eSMatthew Ahrens } 41552ec99e3eSMatthew Ahrens } else if (free_memory < arc_c >> arc_no_grow_shift) { 41562ec99e3eSMatthew Ahrens arc_no_grow = B_TRUE; 4157a8f6344fSEli Rosenthal } else if (gethrtime() >= growtime) { 41582ec99e3eSMatthew Ahrens arc_no_grow = B_FALSE; 4159fa9e4066Sahrens } 4160fa9e4066Sahrens 4161244781f1SPrakash Surya mutex_enter(&arc_reclaim_lock); 4162244781f1SPrakash Surya 4163244781f1SPrakash Surya /* 4164244781f1SPrakash Surya * If evicted is zero, we couldn't evict anything via 4165244781f1SPrakash Surya * arc_adjust(). This could be due to hash lock 4166244781f1SPrakash Surya * collisions, but more likely due to the majority of 4167244781f1SPrakash Surya * arc buffers being unevictable. Therefore, even if 4168244781f1SPrakash Surya * arc_size is above arc_c, another pass is unlikely to 4169244781f1SPrakash Surya * be helpful and could potentially cause us to enter an 4170244781f1SPrakash Surya * infinite loop. 4171244781f1SPrakash Surya */ 4172244781f1SPrakash Surya if (arc_size <= arc_c || evicted == 0) { 4173244781f1SPrakash Surya /* 4174244781f1SPrakash Surya * We're either no longer overflowing, or we 4175244781f1SPrakash Surya * can't evict anything more, so we should wake 4176244781f1SPrakash Surya * up any threads before we go to sleep. 4177244781f1SPrakash Surya */ 4178244781f1SPrakash Surya cv_broadcast(&arc_reclaim_waiters_cv); 4179244781f1SPrakash Surya 4180244781f1SPrakash Surya /* 4181244781f1SPrakash Surya * Block until signaled, or after one second (we 4182244781f1SPrakash Surya * might need to perform arc_kmem_reap_now() 4183244781f1SPrakash Surya * even if we aren't being signalled) 4184244781f1SPrakash Surya */ 4185244781f1SPrakash Surya CALLB_CPR_SAFE_BEGIN(&cpr); 4186a8f6344fSEli Rosenthal (void) cv_timedwait_hires(&arc_reclaim_thread_cv, 4187a8f6344fSEli Rosenthal &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 4188244781f1SPrakash Surya CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 4189244781f1SPrakash Surya } 4190244781f1SPrakash Surya } 4191244781f1SPrakash Surya 4192dcbf3bd6SGeorge Wilson arc_reclaim_thread_exit = B_FALSE; 4193244781f1SPrakash Surya cv_broadcast(&arc_reclaim_thread_cv); 4194244781f1SPrakash Surya CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 4195244781f1SPrakash Surya thread_exit(); 4196244781f1SPrakash Surya } 4197244781f1SPrakash Surya 4198ea8dc4b6Seschrock /* 4199ea8dc4b6Seschrock * Adapt arc info given the number of bytes we are trying to add and 4200ea8dc4b6Seschrock * the state that we are comming from. This function is only called 4201ea8dc4b6Seschrock * when we are adding new content to the cache. 4202ea8dc4b6Seschrock */ 4203fa9e4066Sahrens static void 4204ea8dc4b6Seschrock arc_adapt(int bytes, arc_state_t *state) 4205fa9e4066Sahrens { 4206ea8dc4b6Seschrock int mult; 42075a98e54bSBrendan Gregg - Sun Microsystems uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 42082fd872a7SPrakash Surya int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 42092fd872a7SPrakash Surya int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 4210ea8dc4b6Seschrock 4211fa94a07fSbrendan if (state == arc_l2c_only) 4212fa94a07fSbrendan return; 4213fa94a07fSbrendan 4214ea8dc4b6Seschrock ASSERT(bytes > 0); 4215fa9e4066Sahrens /* 4216ea8dc4b6Seschrock * Adapt the target size of the MRU list: 4217ea8dc4b6Seschrock * - if we just hit in the MRU ghost list, then increase 4218ea8dc4b6Seschrock * the target size of the MRU list. 4219ea8dc4b6Seschrock * - if we just hit in the MFU ghost list, then increase 4220ea8dc4b6Seschrock * the target size of the MFU list by decreasing the 4221ea8dc4b6Seschrock * target size of the MRU list. 4222fa9e4066Sahrens */ 422344cb6abcSbmc if (state == arc_mru_ghost) { 42242fd872a7SPrakash Surya mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 42253e4e8481STom Erickson mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 4226ea8dc4b6Seschrock 42275a98e54bSBrendan Gregg - Sun Microsystems arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 422844cb6abcSbmc } else if (state == arc_mfu_ghost) { 42295a98e54bSBrendan Gregg - Sun Microsystems uint64_t delta; 42305a98e54bSBrendan Gregg - Sun Microsystems 42312fd872a7SPrakash Surya mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 42323e4e8481STom Erickson mult = MIN(mult, 10); 4233ea8dc4b6Seschrock 42345a98e54bSBrendan Gregg - Sun Microsystems delta = MIN(bytes * mult, arc_p); 42355a98e54bSBrendan Gregg - Sun Microsystems arc_p = MAX(arc_p_min, arc_p - delta); 4236ea8dc4b6Seschrock } 423744cb6abcSbmc ASSERT((int64_t)arc_p >= 0); 4238fa9e4066Sahrens 4239fa9e4066Sahrens if (arc_reclaim_needed()) { 4240244781f1SPrakash Surya cv_signal(&arc_reclaim_thread_cv); 4241fa9e4066Sahrens return; 4242fa9e4066Sahrens } 4243fa9e4066Sahrens 424444cb6abcSbmc if (arc_no_grow) 4245fa9e4066Sahrens return; 4246fa9e4066Sahrens 424744cb6abcSbmc if (arc_c >= arc_c_max) 4248ea8dc4b6Seschrock return; 4249ea8dc4b6Seschrock 4250fa9e4066Sahrens /* 4251ea8dc4b6Seschrock * If we're within (2 * maxblocksize) bytes of the target 4252ea8dc4b6Seschrock * cache size, increment the target cache size 4253fa9e4066Sahrens */ 425444cb6abcSbmc if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 425544cb6abcSbmc atomic_add_64(&arc_c, (int64_t)bytes); 425644cb6abcSbmc if (arc_c > arc_c_max) 425744cb6abcSbmc arc_c = arc_c_max; 425844cb6abcSbmc else if (state == arc_anon) 425944cb6abcSbmc atomic_add_64(&arc_p, (int64_t)bytes); 426044cb6abcSbmc if (arc_p > arc_c) 426144cb6abcSbmc arc_p = arc_c; 4262fa9e4066Sahrens } 426344cb6abcSbmc ASSERT((int64_t)arc_p >= 0); 4264fa9e4066Sahrens } 4265fa9e4066Sahrens 4266fa9e4066Sahrens /* 4267244781f1SPrakash Surya * Check if arc_size has grown past our upper threshold, determined by 4268244781f1SPrakash Surya * zfs_arc_overflow_shift. 4269fa9e4066Sahrens */ 4270244781f1SPrakash Surya static boolean_t 4271244781f1SPrakash Surya arc_is_overflowing(void) 4272fa9e4066Sahrens { 4273244781f1SPrakash Surya /* Always allow at least one block of overflow */ 4274244781f1SPrakash Surya uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 4275244781f1SPrakash Surya arc_c >> zfs_arc_overflow_shift); 4276fa9e4066Sahrens 4277244781f1SPrakash Surya return (arc_size >= arc_c + overflow); 4278fa9e4066Sahrens } 4279fa9e4066Sahrens 4280770499e1SDan Kimmel static abd_t * 4281770499e1SDan Kimmel arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4282770499e1SDan Kimmel { 4283770499e1SDan Kimmel arc_buf_contents_t type = arc_buf_type(hdr); 4284770499e1SDan Kimmel 4285770499e1SDan Kimmel arc_get_data_impl(hdr, size, tag); 4286770499e1SDan Kimmel if (type == ARC_BUFC_METADATA) { 4287770499e1SDan Kimmel return (abd_alloc(size, B_TRUE)); 4288770499e1SDan Kimmel } else { 4289770499e1SDan Kimmel ASSERT(type == ARC_BUFC_DATA); 4290770499e1SDan Kimmel return (abd_alloc(size, B_FALSE)); 4291770499e1SDan Kimmel } 4292770499e1SDan Kimmel } 4293770499e1SDan Kimmel 4294770499e1SDan Kimmel static void * 4295770499e1SDan Kimmel arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4296770499e1SDan Kimmel { 4297770499e1SDan Kimmel arc_buf_contents_t type = arc_buf_type(hdr); 4298770499e1SDan Kimmel 4299770499e1SDan Kimmel arc_get_data_impl(hdr, size, tag); 4300770499e1SDan Kimmel if (type == ARC_BUFC_METADATA) { 4301770499e1SDan Kimmel return (zio_buf_alloc(size)); 4302770499e1SDan Kimmel } else { 4303770499e1SDan Kimmel ASSERT(type == ARC_BUFC_DATA); 4304770499e1SDan Kimmel return (zio_data_buf_alloc(size)); 4305770499e1SDan Kimmel } 4306770499e1SDan Kimmel } 4307770499e1SDan Kimmel 4308fa9e4066Sahrens /* 4309dcbf3bd6SGeorge Wilson * Allocate a block and return it to the caller. If we are hitting the 4310dcbf3bd6SGeorge Wilson * hard limit for the cache size, we must sleep, waiting for the eviction 4311dcbf3bd6SGeorge Wilson * thread to catch up. If we're past the target size but below the hard 4312dcbf3bd6SGeorge Wilson * limit, we'll only signal the reclaim thread and continue on. 4313fa9e4066Sahrens */ 4314770499e1SDan Kimmel static void 4315770499e1SDan Kimmel arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4316fa9e4066Sahrens { 4317770499e1SDan Kimmel arc_state_t *state = hdr->b_l1hdr.b_state; 4318770499e1SDan Kimmel arc_buf_contents_t type = arc_buf_type(hdr); 4319fa9e4066Sahrens 432044eda4d7Smaybee arc_adapt(size, state); 4321fa9e4066Sahrens 432244eda4d7Smaybee /* 4323244781f1SPrakash Surya * If arc_size is currently overflowing, and has grown past our 4324244781f1SPrakash Surya * upper limit, we must be adding data faster than the evict 4325244781f1SPrakash Surya * thread can evict. Thus, to ensure we don't compound the 4326244781f1SPrakash Surya * problem by adding more data and forcing arc_size to grow even 4327244781f1SPrakash Surya * further past it's target size, we halt and wait for the 4328244781f1SPrakash Surya * eviction thread to catch up. 4329244781f1SPrakash Surya * 4330244781f1SPrakash Surya * It's also possible that the reclaim thread is unable to evict 4331244781f1SPrakash Surya * enough buffers to get arc_size below the overflow limit (e.g. 4332244781f1SPrakash Surya * due to buffers being un-evictable, or hash lock collisions). 4333244781f1SPrakash Surya * In this case, we want to proceed regardless if we're 4334244781f1SPrakash Surya * overflowing; thus we don't use a while loop here. 433544eda4d7Smaybee */ 4336244781f1SPrakash Surya if (arc_is_overflowing()) { 4337244781f1SPrakash Surya mutex_enter(&arc_reclaim_lock); 4338244781f1SPrakash Surya 4339244781f1SPrakash Surya /* 4340244781f1SPrakash Surya * Now that we've acquired the lock, we may no longer be 4341244781f1SPrakash Surya * over the overflow limit, lets check. 4342244781f1SPrakash Surya * 4343244781f1SPrakash Surya * We're ignoring the case of spurious wake ups. If that 4344244781f1SPrakash Surya * were to happen, it'd let this thread consume an ARC 4345244781f1SPrakash Surya * buffer before it should have (i.e. before we're under 4346244781f1SPrakash Surya * the overflow limit and were signalled by the reclaim 4347244781f1SPrakash Surya * thread). As long as that is a rare occurrence, it 4348244781f1SPrakash Surya * shouldn't cause any harm. 4349244781f1SPrakash Surya */ 4350244781f1SPrakash Surya if (arc_is_overflowing()) { 4351244781f1SPrakash Surya cv_signal(&arc_reclaim_thread_cv); 4352244781f1SPrakash Surya cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 4353ad23a2dbSjohansen } 4354244781f1SPrakash Surya 4355244781f1SPrakash Surya mutex_exit(&arc_reclaim_lock); 435644eda4d7Smaybee } 435744eda4d7Smaybee 4358dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 4359244781f1SPrakash Surya if (type == ARC_BUFC_METADATA) { 4360244781f1SPrakash Surya arc_space_consume(size, ARC_SPACE_META); 4361fa9e4066Sahrens } else { 4362244781f1SPrakash Surya arc_space_consume(size, ARC_SPACE_DATA); 436344eda4d7Smaybee } 4364244781f1SPrakash Surya 436544eda4d7Smaybee /* 436644eda4d7Smaybee * Update the state size. Note that ghost states have a 436744eda4d7Smaybee * "ghost size" and so don't need to be updated. 436844eda4d7Smaybee */ 4369dcbf3bd6SGeorge Wilson if (!GHOST_STATE(state)) { 437044eda4d7Smaybee 4371dcbf3bd6SGeorge Wilson (void) refcount_add_many(&state->arcs_size, size, tag); 4372244781f1SPrakash Surya 4373244781f1SPrakash Surya /* 4374244781f1SPrakash Surya * If this is reached via arc_read, the link is 4375244781f1SPrakash Surya * protected by the hash lock. If reached via 4376244781f1SPrakash Surya * arc_buf_alloc, the header should not be accessed by 4377244781f1SPrakash Surya * any other thread. And, if reached via arc_read_done, 4378244781f1SPrakash Surya * the hash lock will protect it if it's found in the 4379244781f1SPrakash Surya * hash table; otherwise no other thread should be 4380244781f1SPrakash Surya * trying to [add|remove]_reference it. 4381244781f1SPrakash Surya */ 4382244781f1SPrakash Surya if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 438389c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4384dcbf3bd6SGeorge Wilson (void) refcount_add_many(&state->arcs_esize[type], 4385dcbf3bd6SGeorge Wilson size, tag); 4386fa9e4066Sahrens } 4387dcbf3bd6SGeorge Wilson 4388641fbdaeSmaybee /* 4389641fbdaeSmaybee * If we are growing the cache, and we are adding anonymous 439044cb6abcSbmc * data, and we have outgrown arc_p, update arc_p 4391641fbdaeSmaybee */ 439289c86e32SChris Williamson if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 43932fd872a7SPrakash Surya (refcount_count(&arc_anon->arcs_size) + 43942fd872a7SPrakash Surya refcount_count(&arc_mru->arcs_size) > arc_p)) 439544cb6abcSbmc arc_p = MIN(arc_c, arc_p + size); 4396fa9e4066Sahrens } 4397770499e1SDan Kimmel } 4398770499e1SDan Kimmel 4399770499e1SDan Kimmel static void 4400770499e1SDan Kimmel arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag) 4401770499e1SDan Kimmel { 4402770499e1SDan Kimmel arc_free_data_impl(hdr, size, tag); 4403770499e1SDan Kimmel abd_free(abd); 4404770499e1SDan Kimmel } 4405770499e1SDan Kimmel 4406770499e1SDan Kimmel static void 4407770499e1SDan Kimmel arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag) 4408770499e1SDan Kimmel { 4409770499e1SDan Kimmel arc_buf_contents_t type = arc_buf_type(hdr); 4410770499e1SDan Kimmel 4411770499e1SDan Kimmel arc_free_data_impl(hdr, size, tag); 4412770499e1SDan Kimmel if (type == ARC_BUFC_METADATA) { 4413770499e1SDan Kimmel zio_buf_free(buf, size); 4414770499e1SDan Kimmel } else { 4415770499e1SDan Kimmel ASSERT(type == ARC_BUFC_DATA); 4416770499e1SDan Kimmel zio_data_buf_free(buf, size); 4417770499e1SDan Kimmel } 4418dcbf3bd6SGeorge Wilson } 4419dcbf3bd6SGeorge Wilson 4420dcbf3bd6SGeorge Wilson /* 4421dcbf3bd6SGeorge Wilson * Free the arc data buffer. 4422dcbf3bd6SGeorge Wilson */ 4423dcbf3bd6SGeorge Wilson static void 4424770499e1SDan Kimmel arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4425dcbf3bd6SGeorge Wilson { 4426dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 4427dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 4428dcbf3bd6SGeorge Wilson 4429dcbf3bd6SGeorge Wilson /* protected by hash lock, if in the hash table */ 4430dcbf3bd6SGeorge Wilson if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4431dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4432dcbf3bd6SGeorge Wilson ASSERT(state != arc_anon && state != arc_l2c_only); 4433dcbf3bd6SGeorge Wilson 4434dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 4435dcbf3bd6SGeorge Wilson size, tag); 4436dcbf3bd6SGeorge Wilson } 4437dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_size, size, tag); 4438dcbf3bd6SGeorge Wilson 4439dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 4440dcbf3bd6SGeorge Wilson if (type == ARC_BUFC_METADATA) { 4441dcbf3bd6SGeorge Wilson arc_space_return(size, ARC_SPACE_META); 4442dcbf3bd6SGeorge Wilson } else { 4443dcbf3bd6SGeorge Wilson ASSERT(type == ARC_BUFC_DATA); 4444dcbf3bd6SGeorge Wilson arc_space_return(size, ARC_SPACE_DATA); 4445dcbf3bd6SGeorge Wilson } 4446fa9e4066Sahrens } 4447fa9e4066Sahrens 4448fa9e4066Sahrens /* 4449fa9e4066Sahrens * This routine is called whenever a buffer is accessed. 4450ea8dc4b6Seschrock * NOTE: the hash lock is dropped in this function. 4451fa9e4066Sahrens */ 4452fa9e4066Sahrens static void 44537adb730bSGeorge Wilson arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 4454fa9e4066Sahrens { 4455d3d50737SRafael Vanoni clock_t now; 4456d3d50737SRafael Vanoni 4457fa9e4066Sahrens ASSERT(MUTEX_HELD(hash_lock)); 445889c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 4459fa9e4066Sahrens 446089c86e32SChris Williamson if (hdr->b_l1hdr.b_state == arc_anon) { 4461fa9e4066Sahrens /* 4462fa9e4066Sahrens * This buffer is not in the cache, and does not 4463fa9e4066Sahrens * appear in our "ghost" list. Add the new buffer 4464fa9e4066Sahrens * to the MRU state. 4465fa9e4066Sahrens */ 4466fa9e4066Sahrens 446789c86e32SChris Williamson ASSERT0(hdr->b_l1hdr.b_arc_access); 446889c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 44697adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 44707adb730bSGeorge Wilson arc_change_state(arc_mru, hdr, hash_lock); 4471fa9e4066Sahrens 447289c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mru) { 4473d3d50737SRafael Vanoni now = ddi_get_lbolt(); 4474d3d50737SRafael Vanoni 4475fa9e4066Sahrens /* 447613506d1eSmaybee * If this buffer is here because of a prefetch, then either: 447713506d1eSmaybee * - clear the flag if this is a "referencing" read 447813506d1eSmaybee * (any subsequent access will bump this into the MFU state). 447913506d1eSmaybee * or 448013506d1eSmaybee * - move the buffer to the head of the list if this is 448113506d1eSmaybee * another prefetch (to make it less likely to be evicted). 4482fa9e4066Sahrens */ 448389c86e32SChris Williamson if (HDR_PREFETCH(hdr)) { 448489c86e32SChris Williamson if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4485244781f1SPrakash Surya /* link protected by hash lock */ 4486244781f1SPrakash Surya ASSERT(multilist_link_active( 448789c86e32SChris Williamson &hdr->b_l1hdr.b_arc_node)); 448813506d1eSmaybee } else { 4489dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 449044cb6abcSbmc ARCSTAT_BUMP(arcstat_mru_hits); 449113506d1eSmaybee } 449289c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = now; 4493fa9e4066Sahrens return; 4494fa9e4066Sahrens } 4495fa9e4066Sahrens 4496fa9e4066Sahrens /* 4497fa9e4066Sahrens * This buffer has been "accessed" only once so far, 4498fa9e4066Sahrens * but it is still in the cache. Move it to the MFU 4499fa9e4066Sahrens * state. 4500fa9e4066Sahrens */ 450189c86e32SChris Williamson if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 4502fa9e4066Sahrens /* 4503fa9e4066Sahrens * More than 125ms have passed since we 4504fa9e4066Sahrens * instantiated this buffer. Move it to the 4505fa9e4066Sahrens * most frequently used state. 4506fa9e4066Sahrens */ 450789c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = now; 45087adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 45097adb730bSGeorge Wilson arc_change_state(arc_mfu, hdr, hash_lock); 4510fa9e4066Sahrens } 451144cb6abcSbmc ARCSTAT_BUMP(arcstat_mru_hits); 451289c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 4513fa9e4066Sahrens arc_state_t *new_state; 4514fa9e4066Sahrens /* 4515fa9e4066Sahrens * This buffer has been "accessed" recently, but 4516fa9e4066Sahrens * was evicted from the cache. Move it to the 4517fa9e4066Sahrens * MFU state. 4518fa9e4066Sahrens */ 4519fa9e4066Sahrens 452089c86e32SChris Williamson if (HDR_PREFETCH(hdr)) { 452144cb6abcSbmc new_state = arc_mru; 452289c86e32SChris Williamson if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 4523dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 45247adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4525fa9e4066Sahrens } else { 452644cb6abcSbmc new_state = arc_mfu; 45277adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4528fa9e4066Sahrens } 4529fa9e4066Sahrens 453089c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 45317adb730bSGeorge Wilson arc_change_state(new_state, hdr, hash_lock); 4532fa9e4066Sahrens 453344cb6abcSbmc ARCSTAT_BUMP(arcstat_mru_ghost_hits); 453489c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mfu) { 4535fa9e4066Sahrens /* 4536fa9e4066Sahrens * This buffer has been accessed more than once and is 4537fa9e4066Sahrens * still in the cache. Keep it in the MFU state. 4538fa9e4066Sahrens * 453913506d1eSmaybee * NOTE: an add_reference() that occurred when we did 454013506d1eSmaybee * the arc_read() will have kicked this off the list. 454113506d1eSmaybee * If it was a prefetch, we will explicitly move it to 454213506d1eSmaybee * the head of the list now. 4543fa9e4066Sahrens */ 454489c86e32SChris Williamson if ((HDR_PREFETCH(hdr)) != 0) { 454589c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4546244781f1SPrakash Surya /* link protected by hash_lock */ 4547244781f1SPrakash Surya ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 454813506d1eSmaybee } 454944cb6abcSbmc ARCSTAT_BUMP(arcstat_mfu_hits); 455089c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 455189c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 455244cb6abcSbmc arc_state_t *new_state = arc_mfu; 4553fa9e4066Sahrens /* 4554fa9e4066Sahrens * This buffer has been accessed more than once but has 4555fa9e4066Sahrens * been evicted from the cache. Move it back to the 4556fa9e4066Sahrens * MFU state. 4557fa9e4066Sahrens */ 4558fa9e4066Sahrens 455989c86e32SChris Williamson if (HDR_PREFETCH(hdr)) { 456013506d1eSmaybee /* 456113506d1eSmaybee * This is a prefetch access... 456213506d1eSmaybee * move this block back to the MRU state. 456313506d1eSmaybee */ 456489c86e32SChris Williamson ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 456544cb6abcSbmc new_state = arc_mru; 456613506d1eSmaybee } 456713506d1eSmaybee 456889c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 45697adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 45707adb730bSGeorge Wilson arc_change_state(new_state, hdr, hash_lock); 4571fa9e4066Sahrens 457244cb6abcSbmc ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 457389c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4574fa94a07fSbrendan /* 4575fa94a07fSbrendan * This buffer is on the 2nd Level ARC. 4576fa94a07fSbrendan */ 4577fa94a07fSbrendan 457889c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 45797adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 45807adb730bSGeorge Wilson arc_change_state(arc_mfu, hdr, hash_lock); 4581fa9e4066Sahrens } else { 4582fa9e4066Sahrens ASSERT(!"invalid arc state"); 4583fa9e4066Sahrens } 4584fa9e4066Sahrens } 4585fa9e4066Sahrens 4586fa9e4066Sahrens /* a generic arc_done_func_t which you can use */ 4587fa9e4066Sahrens /* ARGSUSED */ 4588fa9e4066Sahrens void 4589fa9e4066Sahrens arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4590fa9e4066Sahrens { 45913f9d6ad7SLin Ling if (zio == NULL || zio->io_error == 0) 45925602294fSDan Kimmel bcopy(buf->b_data, arg, arc_buf_size(buf)); 4593dcbf3bd6SGeorge Wilson arc_buf_destroy(buf, arg); 4594fa9e4066Sahrens } 4595fa9e4066Sahrens 45960e8c6158Smaybee /* a generic arc_done_func_t */ 4597fa9e4066Sahrens void 4598fa9e4066Sahrens arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4599fa9e4066Sahrens { 4600fa9e4066Sahrens arc_buf_t **bufp = arg; 4601fa9e4066Sahrens if (zio && zio->io_error) { 4602dcbf3bd6SGeorge Wilson arc_buf_destroy(buf, arg); 4603fa9e4066Sahrens *bufp = NULL; 4604fa9e4066Sahrens } else { 4605fa9e4066Sahrens *bufp = buf; 46063f9d6ad7SLin Ling ASSERT(buf->b_data); 4607fa9e4066Sahrens } 4608fa9e4066Sahrens } 4609fa9e4066Sahrens 4610dcbf3bd6SGeorge Wilson static void 4611dcbf3bd6SGeorge Wilson arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) 4612dcbf3bd6SGeorge Wilson { 4613dcbf3bd6SGeorge Wilson if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { 4614dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); 4615dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 4616dcbf3bd6SGeorge Wilson } else { 4617dcbf3bd6SGeorge Wilson if (HDR_COMPRESSION_ENABLED(hdr)) { 4618dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, 4619dcbf3bd6SGeorge Wilson BP_GET_COMPRESS(bp)); 4620dcbf3bd6SGeorge Wilson } 4621dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); 4622dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); 4623dcbf3bd6SGeorge Wilson } 4624dcbf3bd6SGeorge Wilson } 4625dcbf3bd6SGeorge Wilson 4626fa9e4066Sahrens static void 4627fa9e4066Sahrens arc_read_done(zio_t *zio) 4628fa9e4066Sahrens { 4629dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = zio->io_private; 46305d7b4d43SMatthew Ahrens kmutex_t *hash_lock = NULL; 46315602294fSDan Kimmel arc_callback_t *callback_list; 46325602294fSDan Kimmel arc_callback_t *acb; 46335602294fSDan Kimmel boolean_t freeable = B_FALSE; 46345602294fSDan Kimmel boolean_t no_zio_error = (zio->io_error == 0); 4635fa9e4066Sahrens 4636bbf4a8dfSmaybee /* 4637bbf4a8dfSmaybee * The hdr was inserted into hash-table and removed from lists 4638bbf4a8dfSmaybee * prior to starting I/O. We should find this header, since 4639bbf4a8dfSmaybee * it's in the hash table, and it should be legit since it's 4640bbf4a8dfSmaybee * not possible to evict it during the I/O. The only possible 4641bbf4a8dfSmaybee * reason for it not to be found is if we were freed during the 4642bbf4a8dfSmaybee * read. 4643bbf4a8dfSmaybee */ 46445d7b4d43SMatthew Ahrens if (HDR_IN_HASH_TABLE(hdr)) { 46455d7b4d43SMatthew Ahrens ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 46465d7b4d43SMatthew Ahrens ASSERT3U(hdr->b_dva.dva_word[0], ==, 46475d7b4d43SMatthew Ahrens BP_IDENTITY(zio->io_bp)->dva_word[0]); 46485d7b4d43SMatthew Ahrens ASSERT3U(hdr->b_dva.dva_word[1], ==, 46495d7b4d43SMatthew Ahrens BP_IDENTITY(zio->io_bp)->dva_word[1]); 46505d7b4d43SMatthew Ahrens 46515d7b4d43SMatthew Ahrens arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 46525d7b4d43SMatthew Ahrens &hash_lock); 46535d7b4d43SMatthew Ahrens 4654dcbf3bd6SGeorge Wilson ASSERT((found == hdr && 46555d7b4d43SMatthew Ahrens DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 46565d7b4d43SMatthew Ahrens (found == hdr && HDR_L2_READING(hdr))); 4657dcbf3bd6SGeorge Wilson ASSERT3P(hash_lock, !=, NULL); 4658dcbf3bd6SGeorge Wilson } 4659dcbf3bd6SGeorge Wilson 46605602294fSDan Kimmel if (no_zio_error) { 4661dcbf3bd6SGeorge Wilson /* byteswap if necessary */ 4662dcbf3bd6SGeorge Wilson if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 4663dcbf3bd6SGeorge Wilson if (BP_GET_LEVEL(zio->io_bp) > 0) { 4664dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; 4665dcbf3bd6SGeorge Wilson } else { 4666dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = 4667dcbf3bd6SGeorge Wilson DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4668dcbf3bd6SGeorge Wilson } 4669dcbf3bd6SGeorge Wilson } else { 4670dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 4671dcbf3bd6SGeorge Wilson } 46725d7b4d43SMatthew Ahrens } 4673fa94a07fSbrendan 4674dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); 467589c86e32SChris Williamson if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4676dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); 4677fa9e4066Sahrens 467889c86e32SChris Williamson callback_list = hdr->b_l1hdr.b_acb; 4679dcbf3bd6SGeorge Wilson ASSERT3P(callback_list, !=, NULL); 46806b4acc8bSahrens 46815602294fSDan Kimmel if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) { 4682b24ab676SJeff Bonwick /* 4683b24ab676SJeff Bonwick * Only call arc_access on anonymous buffers. This is because 4684b24ab676SJeff Bonwick * if we've issued an I/O for an evicted buffer, we've already 4685b24ab676SJeff Bonwick * called arc_access (to prevent any simultaneous readers from 4686b24ab676SJeff Bonwick * getting confused). 4687b24ab676SJeff Bonwick */ 4688b24ab676SJeff Bonwick arc_access(hdr, hash_lock); 4689b24ab676SJeff Bonwick } 4690b24ab676SJeff Bonwick 46915602294fSDan Kimmel /* 46925602294fSDan Kimmel * If a read request has a callback (i.e. acb_done is not NULL), then we 46935602294fSDan Kimmel * make a buf containing the data according to the parameters which were 46945602294fSDan Kimmel * passed in. The implementation of arc_buf_alloc_impl() ensures that we 46955602294fSDan Kimmel * aren't needlessly decompressing the data multiple times. 46965602294fSDan Kimmel */ 46975602294fSDan Kimmel int callback_cnt = 0; 46985602294fSDan Kimmel for (acb = callback_list; acb != NULL; acb = acb->acb_next) { 46995602294fSDan Kimmel if (!acb->acb_done) 47005602294fSDan Kimmel continue; 47015602294fSDan Kimmel 47025602294fSDan Kimmel /* This is a demand read since prefetches don't use callbacks */ 47035602294fSDan Kimmel callback_cnt++; 47045602294fSDan Kimmel 47055602294fSDan Kimmel int error = arc_buf_alloc_impl(hdr, acb->acb_private, 47065602294fSDan Kimmel acb->acb_compressed, no_zio_error, &acb->acb_buf); 47075602294fSDan Kimmel if (no_zio_error) { 47085602294fSDan Kimmel zio->io_error = error; 4709fa9e4066Sahrens } 4710fa9e4066Sahrens } 471189c86e32SChris Williamson hdr->b_l1hdr.b_acb = NULL; 4712dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 47135602294fSDan Kimmel if (callback_cnt == 0) { 4714dcbf3bd6SGeorge Wilson ASSERT(HDR_PREFETCH(hdr)); 4715dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 4716770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 4717b24ab676SJeff Bonwick } 4718fa9e4066Sahrens 471989c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 472089c86e32SChris Williamson callback_list != NULL); 4721fa9e4066Sahrens 47225602294fSDan Kimmel if (no_zio_error) { 4723dcbf3bd6SGeorge Wilson arc_hdr_verify(hdr, zio->io_bp); 4724dcbf3bd6SGeorge Wilson } else { 4725dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); 472689c86e32SChris Williamson if (hdr->b_l1hdr.b_state != arc_anon) 472744cb6abcSbmc arc_change_state(arc_anon, hdr, hash_lock); 4728ea8dc4b6Seschrock if (HDR_IN_HASH_TABLE(hdr)) 4729ea8dc4b6Seschrock buf_hash_remove(hdr); 473089c86e32SChris Williamson freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4731fa9e4066Sahrens } 4732fa9e4066Sahrens 4733ea8dc4b6Seschrock /* 473413506d1eSmaybee * Broadcast before we drop the hash_lock to avoid the possibility 473513506d1eSmaybee * that the hdr (and hence the cv) might be freed before we get to 473613506d1eSmaybee * the cv_broadcast(). 4737ea8dc4b6Seschrock */ 473889c86e32SChris Williamson cv_broadcast(&hdr->b_l1hdr.b_cv); 4739ea8dc4b6Seschrock 474089c86e32SChris Williamson if (hash_lock != NULL) { 474144eda4d7Smaybee mutex_exit(hash_lock); 4742fa9e4066Sahrens } else { 4743fa9e4066Sahrens /* 4744fa9e4066Sahrens * This block was freed while we waited for the read to 4745fa9e4066Sahrens * complete. It has been removed from the hash table and 4746fa9e4066Sahrens * moved to the anonymous state (so that it won't show up 4747fa9e4066Sahrens * in the cache). 4748fa9e4066Sahrens */ 474989c86e32SChris Williamson ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 475089c86e32SChris Williamson freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4751fa9e4066Sahrens } 4752fa9e4066Sahrens 4753fa9e4066Sahrens /* execute each callback and free its structure */ 4754fa9e4066Sahrens while ((acb = callback_list) != NULL) { 4755fa9e4066Sahrens if (acb->acb_done) 4756fa9e4066Sahrens acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4757fa9e4066Sahrens 4758fa9e4066Sahrens if (acb->acb_zio_dummy != NULL) { 4759fa9e4066Sahrens acb->acb_zio_dummy->io_error = zio->io_error; 4760fa9e4066Sahrens zio_nowait(acb->acb_zio_dummy); 4761fa9e4066Sahrens } 4762fa9e4066Sahrens 4763fa9e4066Sahrens callback_list = acb->acb_next; 4764fa9e4066Sahrens kmem_free(acb, sizeof (arc_callback_t)); 4765fa9e4066Sahrens } 4766fa9e4066Sahrens 4767fa9e4066Sahrens if (freeable) 4768ea8dc4b6Seschrock arc_hdr_destroy(hdr); 4769fa9e4066Sahrens } 4770fa9e4066Sahrens 4771fa9e4066Sahrens /* 4772fc98fea5SBart Coddens * "Read" the block at the specified DVA (in bp) via the 4773fa9e4066Sahrens * cache. If the block is found in the cache, invoke the provided 4774fa9e4066Sahrens * callback immediately and return. Note that the `zio' parameter 4775fa9e4066Sahrens * in the callback will be NULL in this case, since no IO was 4776fa9e4066Sahrens * required. If the block is not in the cache pass the read request 4777fa9e4066Sahrens * on to the spa with a substitute callback function, so that the 4778fa9e4066Sahrens * requested block will be added to the cache. 4779fa9e4066Sahrens * 4780fa9e4066Sahrens * If a read request arrives for a block that has a read in-progress, 4781fa9e4066Sahrens * either wait for the in-progress read to complete (and return the 4782fa9e4066Sahrens * results); or, if this is a read with a "done" func, add a record 4783fa9e4066Sahrens * to the read to invoke the "done" func when the read completes, 4784fa9e4066Sahrens * and return; or just return. 4785fa9e4066Sahrens * 4786fa9e4066Sahrens * arc_read_done() will invoke all the requested "done" functions 4787fa9e4066Sahrens * for readers of this block. 4788fa9e4066Sahrens */ 4789fa9e4066Sahrens int 47901b912ec7SGeorge Wilson arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 47917adb730bSGeorge Wilson void *private, zio_priority_t priority, int zio_flags, 47927adb730bSGeorge Wilson arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4793fa9e4066Sahrens { 47945d7b4d43SMatthew Ahrens arc_buf_hdr_t *hdr = NULL; 47955d7b4d43SMatthew Ahrens kmutex_t *hash_lock = NULL; 4796fa94a07fSbrendan zio_t *rzio; 4797e9103aaeSGarrett D'Amore uint64_t guid = spa_load_guid(spa); 47985602294fSDan Kimmel boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0; 4799fa9e4066Sahrens 48005d7b4d43SMatthew Ahrens ASSERT(!BP_IS_EMBEDDED(bp) || 48015d7b4d43SMatthew Ahrens BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 48025d7b4d43SMatthew Ahrens 4803fa9e4066Sahrens top: 48045d7b4d43SMatthew Ahrens if (!BP_IS_EMBEDDED(bp)) { 48055d7b4d43SMatthew Ahrens /* 48065d7b4d43SMatthew Ahrens * Embedded BP's have no DVA and require no I/O to "read". 48075d7b4d43SMatthew Ahrens * Create an anonymous arc buf to back it. 48085d7b4d43SMatthew Ahrens */ 48095d7b4d43SMatthew Ahrens hdr = buf_hash_find(guid, bp, &hash_lock); 48105d7b4d43SMatthew Ahrens } 48115d7b4d43SMatthew Ahrens 4812770499e1SDan Kimmel if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) { 4813dcbf3bd6SGeorge Wilson arc_buf_t *buf = NULL; 48147adb730bSGeorge Wilson *arc_flags |= ARC_FLAG_CACHED; 481513506d1eSmaybee 4816fa9e4066Sahrens if (HDR_IO_IN_PROGRESS(hdr)) { 481713506d1eSmaybee 4818cf6106c8SMatthew Ahrens if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 4819cf6106c8SMatthew Ahrens priority == ZIO_PRIORITY_SYNC_READ) { 4820cf6106c8SMatthew Ahrens /* 4821cf6106c8SMatthew Ahrens * This sync read must wait for an 4822cf6106c8SMatthew Ahrens * in-progress async read (e.g. a predictive 4823cf6106c8SMatthew Ahrens * prefetch). Async reads are queued 4824cf6106c8SMatthew Ahrens * separately at the vdev_queue layer, so 4825cf6106c8SMatthew Ahrens * this is a form of priority inversion. 4826cf6106c8SMatthew Ahrens * Ideally, we would "inherit" the demand 4827cf6106c8SMatthew Ahrens * i/o's priority by moving the i/o from 4828cf6106c8SMatthew Ahrens * the async queue to the synchronous queue, 4829cf6106c8SMatthew Ahrens * but there is currently no mechanism to do 4830cf6106c8SMatthew Ahrens * so. Track this so that we can evaluate 4831cf6106c8SMatthew Ahrens * the magnitude of this potential performance 4832cf6106c8SMatthew Ahrens * problem. 4833cf6106c8SMatthew Ahrens * 4834cf6106c8SMatthew Ahrens * Note that if the prefetch i/o is already 4835cf6106c8SMatthew Ahrens * active (has been issued to the device), 4836cf6106c8SMatthew Ahrens * the prefetch improved performance, because 4837cf6106c8SMatthew Ahrens * we issued it sooner than we would have 4838cf6106c8SMatthew Ahrens * without the prefetch. 4839cf6106c8SMatthew Ahrens */ 4840cf6106c8SMatthew Ahrens DTRACE_PROBE1(arc__sync__wait__for__async, 4841cf6106c8SMatthew Ahrens arc_buf_hdr_t *, hdr); 4842cf6106c8SMatthew Ahrens ARCSTAT_BUMP(arcstat_sync_wait_for_async); 4843cf6106c8SMatthew Ahrens } 4844cf6106c8SMatthew Ahrens if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4845dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, 4846dcbf3bd6SGeorge Wilson ARC_FLAG_PREDICTIVE_PREFETCH); 4847cf6106c8SMatthew Ahrens } 4848cf6106c8SMatthew Ahrens 48497adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_WAIT) { 485089c86e32SChris Williamson cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 485113506d1eSmaybee mutex_exit(hash_lock); 485213506d1eSmaybee goto top; 485313506d1eSmaybee } 48547adb730bSGeorge Wilson ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 485513506d1eSmaybee 485613506d1eSmaybee if (done) { 4857cf6106c8SMatthew Ahrens arc_callback_t *acb = NULL; 4858fa9e4066Sahrens 4859fa9e4066Sahrens acb = kmem_zalloc(sizeof (arc_callback_t), 4860fa9e4066Sahrens KM_SLEEP); 4861fa9e4066Sahrens acb->acb_done = done; 4862fa9e4066Sahrens acb->acb_private = private; 48635602294fSDan Kimmel acb->acb_compressed = compressed_read; 4864fa9e4066Sahrens if (pio != NULL) 4865fa9e4066Sahrens acb->acb_zio_dummy = zio_null(pio, 4866a3f829aeSBill Moore spa, NULL, NULL, NULL, zio_flags); 4867fa9e4066Sahrens 4868dcbf3bd6SGeorge Wilson ASSERT3P(acb->acb_done, !=, NULL); 486989c86e32SChris Williamson acb->acb_next = hdr->b_l1hdr.b_acb; 487089c86e32SChris Williamson hdr->b_l1hdr.b_acb = acb; 4871fa9e4066Sahrens mutex_exit(hash_lock); 4872fa9e4066Sahrens return (0); 4873fa9e4066Sahrens } 4874fa9e4066Sahrens mutex_exit(hash_lock); 4875fa9e4066Sahrens return (0); 4876fa9e4066Sahrens } 4877fa9e4066Sahrens 487889c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_state == arc_mru || 487989c86e32SChris Williamson hdr->b_l1hdr.b_state == arc_mfu); 4880fa9e4066Sahrens 4881ea8dc4b6Seschrock if (done) { 4882cf6106c8SMatthew Ahrens if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4883cf6106c8SMatthew Ahrens /* 4884cf6106c8SMatthew Ahrens * This is a demand read which does not have to 4885cf6106c8SMatthew Ahrens * wait for i/o because we did a predictive 4886cf6106c8SMatthew Ahrens * prefetch i/o for it, which has completed. 4887cf6106c8SMatthew Ahrens */ 4888cf6106c8SMatthew Ahrens DTRACE_PROBE1( 4889cf6106c8SMatthew Ahrens arc__demand__hit__predictive__prefetch, 4890cf6106c8SMatthew Ahrens arc_buf_hdr_t *, hdr); 4891cf6106c8SMatthew Ahrens ARCSTAT_BUMP( 4892cf6106c8SMatthew Ahrens arcstat_demand_hit_predictive_prefetch); 4893dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, 4894dcbf3bd6SGeorge Wilson ARC_FLAG_PREDICTIVE_PREFETCH); 4895cf6106c8SMatthew Ahrens } 4896dcbf3bd6SGeorge Wilson ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); 4897dcbf3bd6SGeorge Wilson 48985602294fSDan Kimmel /* Get a buf with the desired data in it. */ 48995602294fSDan Kimmel VERIFY0(arc_buf_alloc_impl(hdr, private, 49005602294fSDan Kimmel compressed_read, B_TRUE, &buf)); 49017adb730bSGeorge Wilson } else if (*arc_flags & ARC_FLAG_PREFETCH && 490289c86e32SChris Williamson refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4903dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 4904fa9e4066Sahrens } 4905fa9e4066Sahrens DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 490644eda4d7Smaybee arc_access(hdr, hash_lock); 49077adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_L2CACHE) 4908dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 490944eda4d7Smaybee mutex_exit(hash_lock); 491044cb6abcSbmc ARCSTAT_BUMP(arcstat_hits); 491189c86e32SChris Williamson ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 491289c86e32SChris Williamson demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 491344cb6abcSbmc data, metadata, hits); 491444cb6abcSbmc 4915fa9e4066Sahrens if (done) 4916fa9e4066Sahrens done(NULL, buf, private); 4917fa9e4066Sahrens } else { 4918dcbf3bd6SGeorge Wilson uint64_t lsize = BP_GET_LSIZE(bp); 4919dcbf3bd6SGeorge Wilson uint64_t psize = BP_GET_PSIZE(bp); 49205d7b4d43SMatthew Ahrens arc_callback_t *acb; 49213a737e0dSbrendan vdev_t *vd = NULL; 4922d5285caeSGeorge Wilson uint64_t addr = 0; 49235a98e54bSBrendan Gregg - Sun Microsystems boolean_t devw = B_FALSE; 4924dcbf3bd6SGeorge Wilson uint64_t size; 4925fa9e4066Sahrens 4926fa9e4066Sahrens if (hdr == NULL) { 4927fa9e4066Sahrens /* this block is not in the cache */ 49285d7b4d43SMatthew Ahrens arc_buf_hdr_t *exists = NULL; 4929ad23a2dbSjohansen arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4930dcbf3bd6SGeorge Wilson hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 4931dcbf3bd6SGeorge Wilson BP_GET_COMPRESS(bp), type); 4932dcbf3bd6SGeorge Wilson 49335d7b4d43SMatthew Ahrens if (!BP_IS_EMBEDDED(bp)) { 49345d7b4d43SMatthew Ahrens hdr->b_dva = *BP_IDENTITY(bp); 49355d7b4d43SMatthew Ahrens hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 49365d7b4d43SMatthew Ahrens exists = buf_hash_insert(hdr, &hash_lock); 49375d7b4d43SMatthew Ahrens } 49385d7b4d43SMatthew Ahrens if (exists != NULL) { 4939fa9e4066Sahrens /* somebody beat us to the hash insert */ 4940fa9e4066Sahrens mutex_exit(hash_lock); 49413f9d6ad7SLin Ling buf_discard_identity(hdr); 4942dcbf3bd6SGeorge Wilson arc_hdr_destroy(hdr); 4943fa9e4066Sahrens goto top; /* restart the IO request */ 4944fa9e4066Sahrens } 4945fa9e4066Sahrens } else { 494689c86e32SChris Williamson /* 494789c86e32SChris Williamson * This block is in the ghost cache. If it was L2-only 494889c86e32SChris Williamson * (and thus didn't have an L1 hdr), we realloc the 494989c86e32SChris Williamson * header to add an L1 hdr. 495089c86e32SChris Williamson */ 495189c86e32SChris Williamson if (!HDR_HAS_L1HDR(hdr)) { 495289c86e32SChris Williamson hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 495389c86e32SChris Williamson hdr_full_cache); 495489c86e32SChris Williamson } 4955770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 495689c86e32SChris Williamson ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4957ea8dc4b6Seschrock ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 495889c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4959244781f1SPrakash Surya ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 49605602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 496113506d1eSmaybee 4962cf6106c8SMatthew Ahrens /* 4963dcbf3bd6SGeorge Wilson * This is a delicate dance that we play here. 4964dcbf3bd6SGeorge Wilson * This hdr is in the ghost list so we access it 4965dcbf3bd6SGeorge Wilson * to move it out of the ghost list before we 4966dcbf3bd6SGeorge Wilson * initiate the read. If it's a prefetch then 4967dcbf3bd6SGeorge Wilson * it won't have a callback so we'll remove the 4968dcbf3bd6SGeorge Wilson * reference that arc_buf_alloc_impl() created. We 4969dcbf3bd6SGeorge Wilson * do this after we've called arc_access() to 4970dcbf3bd6SGeorge Wilson * avoid hitting an assert in remove_reference(). 4971cf6106c8SMatthew Ahrens */ 49727e453561SWilliam Gorrell arc_access(hdr, hash_lock); 4973770499e1SDan Kimmel arc_hdr_alloc_pabd(hdr); 4974dcbf3bd6SGeorge Wilson } 4975770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 4976dcbf3bd6SGeorge Wilson size = arc_hdr_size(hdr); 4977dcbf3bd6SGeorge Wilson 4978dcbf3bd6SGeorge Wilson /* 4979dcbf3bd6SGeorge Wilson * If compression is enabled on the hdr, then will do 4980dcbf3bd6SGeorge Wilson * RAW I/O and will store the compressed data in the hdr's 4981dcbf3bd6SGeorge Wilson * data block. Otherwise, the hdr's data block will contain 4982dcbf3bd6SGeorge Wilson * the uncompressed data. 4983dcbf3bd6SGeorge Wilson */ 4984dcbf3bd6SGeorge Wilson if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 4985dcbf3bd6SGeorge Wilson zio_flags |= ZIO_FLAG_RAW; 4986fa9e4066Sahrens } 4987fa9e4066Sahrens 4988dcbf3bd6SGeorge Wilson if (*arc_flags & ARC_FLAG_PREFETCH) 4989dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 4990dcbf3bd6SGeorge Wilson if (*arc_flags & ARC_FLAG_L2CACHE) 4991dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 4992dcbf3bd6SGeorge Wilson if (BP_GET_LEVEL(bp) > 0) 4993dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); 4994cf6106c8SMatthew Ahrens if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 4995dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); 499689c86e32SChris Williamson ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 49975614b00aSWilliam Gorrell 4998fa9e4066Sahrens acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4999fa9e4066Sahrens acb->acb_done = done; 5000fa9e4066Sahrens acb->acb_private = private; 50015602294fSDan Kimmel acb->acb_compressed = compressed_read; 5002fa9e4066Sahrens 5003dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 500489c86e32SChris Williamson hdr->b_l1hdr.b_acb = acb; 5005dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5006fa9e4066Sahrens 500789c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr) && 500889c86e32SChris Williamson (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 500989c86e32SChris Williamson devw = hdr->b_l2hdr.b_dev->l2ad_writing; 501089c86e32SChris Williamson addr = hdr->b_l2hdr.b_daddr; 5011e14bb325SJeff Bonwick /* 50125cabbc6bSPrashanth Sreenivasa * Lock out L2ARC device removal. 5013e14bb325SJeff Bonwick */ 5014e14bb325SJeff Bonwick if (vdev_is_dead(vd) || 5015e14bb325SJeff Bonwick !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 5016e14bb325SJeff Bonwick vd = NULL; 50173a737e0dSbrendan } 50183a737e0dSbrendan 5019dcbf3bd6SGeorge Wilson if (priority == ZIO_PRIORITY_ASYNC_READ) 5020dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5021dcbf3bd6SGeorge Wilson else 5022dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5023dcbf3bd6SGeorge Wilson 50245d7b4d43SMatthew Ahrens if (hash_lock != NULL) 50255d7b4d43SMatthew Ahrens mutex_exit(hash_lock); 50263a737e0dSbrendan 50273e30c24aSWill Andrews /* 50283e30c24aSWill Andrews * At this point, we have a level 1 cache miss. Try again in 50293e30c24aSWill Andrews * L2ARC if possible. 50303e30c24aSWill Andrews */ 5031dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); 5032dcbf3bd6SGeorge Wilson 50335c28183bSBrendan Gregg - Sun Microsystems DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 5034dcbf3bd6SGeorge Wilson uint64_t, lsize, zbookmark_phys_t *, zb); 503544cb6abcSbmc ARCSTAT_BUMP(arcstat_misses); 503689c86e32SChris Williamson ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 503789c86e32SChris Williamson demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 503844cb6abcSbmc data, metadata, misses); 5039ea8dc4b6Seschrock 50405a98e54bSBrendan Gregg - Sun Microsystems if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 5041fa94a07fSbrendan /* 5042fa94a07fSbrendan * Read from the L2ARC if the following are true: 50433a737e0dSbrendan * 1. The L2ARC vdev was previously cached. 50443a737e0dSbrendan * 2. This buffer still has L2ARC metadata. 50453a737e0dSbrendan * 3. This buffer isn't currently writing to the L2ARC. 50463a737e0dSbrendan * 4. The L2ARC entry wasn't evicted, which may 50473a737e0dSbrendan * also have invalidated the vdev. 50485a98e54bSBrendan Gregg - Sun Microsystems * 5. This isn't prefetch and l2arc_noprefetch is set. 5049fa94a07fSbrendan */ 505089c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr) && 50515a98e54bSBrendan Gregg - Sun Microsystems !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 50525a98e54bSBrendan Gregg - Sun Microsystems !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 5053fa94a07fSbrendan l2arc_read_callback_t *cb; 5054403a8da7SAndriy Gapon abd_t *abd; 5055403a8da7SAndriy Gapon uint64_t asize; 5056fa94a07fSbrendan 5057c5904d13Seschrock DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 5058c5904d13Seschrock ARCSTAT_BUMP(arcstat_l2_hits); 5059c5904d13Seschrock 5060fa94a07fSbrendan cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 5061fa94a07fSbrendan KM_SLEEP); 5062dcbf3bd6SGeorge Wilson cb->l2rcb_hdr = hdr; 5063fa94a07fSbrendan cb->l2rcb_bp = *bp; 5064fa94a07fSbrendan cb->l2rcb_zb = *zb; 50653baa08fcSek cb->l2rcb_flags = zio_flags; 5066fa94a07fSbrendan 5067403a8da7SAndriy Gapon asize = vdev_psize_to_asize(vd, size); 5068403a8da7SAndriy Gapon if (asize != size) { 5069403a8da7SAndriy Gapon abd = abd_alloc_for_io(asize, 5070403a8da7SAndriy Gapon HDR_ISTYPE_METADATA(hdr)); 5071403a8da7SAndriy Gapon cb->l2rcb_abd = abd; 5072403a8da7SAndriy Gapon } else { 5073403a8da7SAndriy Gapon abd = hdr->b_l1hdr.b_pabd; 5074403a8da7SAndriy Gapon } 5075403a8da7SAndriy Gapon 5076d5285caeSGeorge Wilson ASSERT(addr >= VDEV_LABEL_START_SIZE && 5077403a8da7SAndriy Gapon addr + asize <= vd->vdev_psize - 5078d5285caeSGeorge Wilson VDEV_LABEL_END_SIZE); 5079d5285caeSGeorge Wilson 5080fa94a07fSbrendan /* 5081e14bb325SJeff Bonwick * l2arc read. The SCL_L2ARC lock will be 5082e14bb325SJeff Bonwick * released by l2arc_read_done(). 5083aad02571SSaso Kiselkov * Issue a null zio if the underlying buffer 5084aad02571SSaso Kiselkov * was squashed to zero size by compression. 5085fa94a07fSbrendan */ 5086dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), !=, 5087dcbf3bd6SGeorge Wilson ZIO_COMPRESS_EMPTY); 5088dcbf3bd6SGeorge Wilson rzio = zio_read_phys(pio, vd, addr, 5089403a8da7SAndriy Gapon asize, abd, 5090dcbf3bd6SGeorge Wilson ZIO_CHECKSUM_OFF, 5091dcbf3bd6SGeorge Wilson l2arc_read_done, cb, priority, 5092dcbf3bd6SGeorge Wilson zio_flags | ZIO_FLAG_DONT_CACHE | 5093dcbf3bd6SGeorge Wilson ZIO_FLAG_CANFAIL | 5094dcbf3bd6SGeorge Wilson ZIO_FLAG_DONT_PROPAGATE | 5095dcbf3bd6SGeorge Wilson ZIO_FLAG_DONT_RETRY, B_FALSE); 5096fa94a07fSbrendan DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 5097fa94a07fSbrendan zio_t *, rzio); 5098dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_l2_read_bytes, size); 5099fa94a07fSbrendan 51007adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_NOWAIT) { 51013a737e0dSbrendan zio_nowait(rzio); 51023a737e0dSbrendan return (0); 51033a737e0dSbrendan } 5104fa94a07fSbrendan 51057adb730bSGeorge Wilson ASSERT(*arc_flags & ARC_FLAG_WAIT); 51063a737e0dSbrendan if (zio_wait(rzio) == 0) 51073a737e0dSbrendan return (0); 51083a737e0dSbrendan 51093a737e0dSbrendan /* l2arc read error; goto zio_read() */ 5110fa94a07fSbrendan } else { 5111fa94a07fSbrendan DTRACE_PROBE1(l2arc__miss, 5112fa94a07fSbrendan arc_buf_hdr_t *, hdr); 5113fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_misses); 5114fa94a07fSbrendan if (HDR_L2_WRITING(hdr)) 5115fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_rw_clash); 5116e14bb325SJeff Bonwick spa_config_exit(spa, SCL_L2ARC, vd); 5117fa94a07fSbrendan } 51185a98e54bSBrendan Gregg - Sun Microsystems } else { 511976a25fafSBill Moore if (vd != NULL) 512076a25fafSBill Moore spa_config_exit(spa, SCL_L2ARC, vd); 51215a98e54bSBrendan Gregg - Sun Microsystems if (l2arc_ndev != 0) { 51225a98e54bSBrendan Gregg - Sun Microsystems DTRACE_PROBE1(l2arc__miss, 51235a98e54bSBrendan Gregg - Sun Microsystems arc_buf_hdr_t *, hdr); 51245a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_BUMP(arcstat_l2_misses); 51255a98e54bSBrendan Gregg - Sun Microsystems } 5126fa94a07fSbrendan } 5127c5904d13Seschrock 5128770499e1SDan Kimmel rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size, 5129dcbf3bd6SGeorge Wilson arc_read_done, hdr, priority, zio_flags, zb); 5130fa9e4066Sahrens 51317adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_WAIT) 5132fa9e4066Sahrens return (zio_wait(rzio)); 5133fa9e4066Sahrens 51347adb730bSGeorge Wilson ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5135fa9e4066Sahrens zio_nowait(rzio); 5136fa9e4066Sahrens } 5137fa9e4066Sahrens return (0); 5138fa9e4066Sahrens } 5139fa9e4066Sahrens 51406e6d5868SMatthew Ahrens /* 51416e6d5868SMatthew Ahrens * Notify the arc that a block was freed, and thus will never be used again. 51426e6d5868SMatthew Ahrens */ 51436e6d5868SMatthew Ahrens void 51446e6d5868SMatthew Ahrens arc_freed(spa_t *spa, const blkptr_t *bp) 51456e6d5868SMatthew Ahrens { 51466e6d5868SMatthew Ahrens arc_buf_hdr_t *hdr; 51476e6d5868SMatthew Ahrens kmutex_t *hash_lock; 51486e6d5868SMatthew Ahrens uint64_t guid = spa_load_guid(spa); 51496e6d5868SMatthew Ahrens 51505d7b4d43SMatthew Ahrens ASSERT(!BP_IS_EMBEDDED(bp)); 51515d7b4d43SMatthew Ahrens 51525d7b4d43SMatthew Ahrens hdr = buf_hash_find(guid, bp, &hash_lock); 51536e6d5868SMatthew Ahrens if (hdr == NULL) 51546e6d5868SMatthew Ahrens return; 51556e6d5868SMatthew Ahrens 5156dcbf3bd6SGeorge Wilson /* 5157dcbf3bd6SGeorge Wilson * We might be trying to free a block that is still doing I/O 5158dcbf3bd6SGeorge Wilson * (i.e. prefetch) or has a reference (i.e. a dedup-ed, 5159dcbf3bd6SGeorge Wilson * dmu_sync-ed block). If this block is being prefetched, then it 5160dcbf3bd6SGeorge Wilson * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr 5161dcbf3bd6SGeorge Wilson * until the I/O completes. A block may also have a reference if it is 5162dcbf3bd6SGeorge Wilson * part of a dedup-ed, dmu_synced write. The dmu_sync() function would 5163dcbf3bd6SGeorge Wilson * have written the new block to its final resting place on disk but 5164dcbf3bd6SGeorge Wilson * without the dedup flag set. This would have left the hdr in the MRU 5165dcbf3bd6SGeorge Wilson * state and discoverable. When the txg finally syncs it detects that 5166dcbf3bd6SGeorge Wilson * the block was overridden in open context and issues an override I/O. 5167dcbf3bd6SGeorge Wilson * Since this is a dedup block, the override I/O will determine if the 5168dcbf3bd6SGeorge Wilson * block is already in the DDT. If so, then it will replace the io_bp 5169dcbf3bd6SGeorge Wilson * with the bp from the DDT and allow the I/O to finish. When the I/O 5170dcbf3bd6SGeorge Wilson * reaches the done callback, dbuf_write_override_done, it will 5171dcbf3bd6SGeorge Wilson * check to see if the io_bp and io_bp_override are identical. 5172dcbf3bd6SGeorge Wilson * If they are not, then it indicates that the bp was replaced with 5173dcbf3bd6SGeorge Wilson * the bp in the DDT and the override bp is freed. This allows 5174dcbf3bd6SGeorge Wilson * us to arrive here with a reference on a block that is being 5175dcbf3bd6SGeorge Wilson * freed. So if we have an I/O in progress, or a reference to 5176dcbf3bd6SGeorge Wilson * this hdr, then we don't destroy the hdr. 5177dcbf3bd6SGeorge Wilson */ 5178dcbf3bd6SGeorge Wilson if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && 5179dcbf3bd6SGeorge Wilson refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { 5180dcbf3bd6SGeorge Wilson arc_change_state(arc_anon, hdr, hash_lock); 5181dcbf3bd6SGeorge Wilson arc_hdr_destroy(hdr); 51826e6d5868SMatthew Ahrens mutex_exit(hash_lock); 5183bbfa8ea8SMatthew Ahrens } else { 5184dcbf3bd6SGeorge Wilson mutex_exit(hash_lock); 5185ea8dc4b6Seschrock } 5186dd6ef538Smaybee 5187ea8dc4b6Seschrock } 5188ea8dc4b6Seschrock 5189fa9e4066Sahrens /* 51903e30c24aSWill Andrews * Release this buffer from the cache, making it an anonymous buffer. This 51913e30c24aSWill Andrews * must be done after a read and prior to modifying the buffer contents. 5192fa9e4066Sahrens * If the buffer has more than one reference, we must make 5193088f3894Sahrens * a new hdr for the buffer. 5194fa9e4066Sahrens */ 5195fa9e4066Sahrens void 5196fa9e4066Sahrens arc_release(arc_buf_t *buf, void *tag) 5197fa9e4066Sahrens { 519889c86e32SChris Williamson arc_buf_hdr_t *hdr = buf->b_hdr; 5199fa9e4066Sahrens 52003f9d6ad7SLin Ling /* 52013f9d6ad7SLin Ling * It would be nice to assert that if it's DMU metadata (level > 52023f9d6ad7SLin Ling * 0 || it's the dnode file), then it must be syncing context. 52033f9d6ad7SLin Ling * But we don't know that information at this level. 52043f9d6ad7SLin Ling */ 52053f9d6ad7SLin Ling 52063f9d6ad7SLin Ling mutex_enter(&buf->b_evict_lock); 52076f83844dSMark Maybee 5208244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 5209244781f1SPrakash Surya 521089c86e32SChris Williamson /* 521189c86e32SChris Williamson * We don't grab the hash lock prior to this check, because if 521289c86e32SChris Williamson * the buffer's header is in the arc_anon state, it won't be 521389c86e32SChris Williamson * linked into the hash table. 521489c86e32SChris Williamson */ 521589c86e32SChris Williamson if (hdr->b_l1hdr.b_state == arc_anon) { 521689c86e32SChris Williamson mutex_exit(&buf->b_evict_lock); 521789c86e32SChris Williamson ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 521889c86e32SChris Williamson ASSERT(!HDR_IN_HASH_TABLE(hdr)); 521989c86e32SChris Williamson ASSERT(!HDR_HAS_L2HDR(hdr)); 5220dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 5221fa9e4066Sahrens 5222dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 522389c86e32SChris Williamson ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 522489c86e32SChris Williamson ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 522589c86e32SChris Williamson 522689c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = 0; 5227dcbf3bd6SGeorge Wilson 5228dcbf3bd6SGeorge Wilson /* 5229dcbf3bd6SGeorge Wilson * If the buf is being overridden then it may already 5230dcbf3bd6SGeorge Wilson * have a hdr that is not empty. 5231dcbf3bd6SGeorge Wilson */ 5232dcbf3bd6SGeorge Wilson buf_discard_identity(hdr); 523389c86e32SChris Williamson arc_buf_thaw(buf); 523489c86e32SChris Williamson 523589c86e32SChris Williamson return; 5236fa9e4066Sahrens } 5237fa9e4066Sahrens 523889c86e32SChris Williamson kmutex_t *hash_lock = HDR_LOCK(hdr); 523989c86e32SChris Williamson mutex_enter(hash_lock); 524089c86e32SChris Williamson 524189c86e32SChris Williamson /* 524289c86e32SChris Williamson * This assignment is only valid as long as the hash_lock is 524389c86e32SChris Williamson * held, we must be careful not to reference state or the 524489c86e32SChris Williamson * b_state field after dropping the lock. 524589c86e32SChris Williamson */ 524689c86e32SChris Williamson arc_state_t *state = hdr->b_l1hdr.b_state; 524789c86e32SChris Williamson ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 524889c86e32SChris Williamson ASSERT3P(state, !=, arc_anon); 524989c86e32SChris Williamson 525089c86e32SChris Williamson /* this buffer is not on any list */ 52515602294fSDan Kimmel ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0); 525289c86e32SChris Williamson 525389c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr)) { 525489c86e32SChris Williamson mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5255244781f1SPrakash Surya 5256244781f1SPrakash Surya /* 5257a52fc310SPrakash Surya * We have to recheck this conditional again now that 5258a52fc310SPrakash Surya * we're holding the l2ad_mtx to prevent a race with 5259a52fc310SPrakash Surya * another thread which might be concurrently calling 5260a52fc310SPrakash Surya * l2arc_evict(). In that case, l2arc_evict() might have 5261a52fc310SPrakash Surya * destroyed the header's L2 portion as we were waiting 5262a52fc310SPrakash Surya * to acquire the l2ad_mtx. 5263244781f1SPrakash Surya */ 5264a52fc310SPrakash Surya if (HDR_HAS_L2HDR(hdr)) 5265a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(hdr); 5266244781f1SPrakash Surya 526789c86e32SChris Williamson mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 52686f83844dSMark Maybee } 52696f83844dSMark Maybee 5270ea8dc4b6Seschrock /* 5271ea8dc4b6Seschrock * Do we have more than one buf? 5272ea8dc4b6Seschrock */ 5273dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_bufcnt > 1) { 5274fa9e4066Sahrens arc_buf_hdr_t *nhdr; 5275ac05c741SMark Maybee uint64_t spa = hdr->b_spa; 5276dcbf3bd6SGeorge Wilson uint64_t psize = HDR_GET_PSIZE(hdr); 5277dcbf3bd6SGeorge Wilson uint64_t lsize = HDR_GET_LSIZE(hdr); 5278dcbf3bd6SGeorge Wilson enum zio_compress compress = HDR_GET_COMPRESS(hdr); 527989c86e32SChris Williamson arc_buf_contents_t type = arc_buf_type(hdr); 5280dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 5281fa9e4066Sahrens 528289c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 5283dcbf3bd6SGeorge Wilson (void) remove_reference(hdr, hash_lock, tag); 5284dcbf3bd6SGeorge Wilson 52855602294fSDan Kimmel if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) { 5286dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5287dcbf3bd6SGeorge Wilson ASSERT(ARC_BUF_LAST(buf)); 5288dcbf3bd6SGeorge Wilson } 5289dcbf3bd6SGeorge Wilson 5290fa9e4066Sahrens /* 52913f9d6ad7SLin Ling * Pull the data off of this hdr and attach it to 5292dcbf3bd6SGeorge Wilson * a new anonymous hdr. Also find the last buffer 5293dcbf3bd6SGeorge Wilson * in the hdr's buffer list. 5294fa9e4066Sahrens */ 52955602294fSDan Kimmel arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 5296dcbf3bd6SGeorge Wilson ASSERT3P(lastbuf, !=, NULL); 5297dcbf3bd6SGeorge Wilson 5298dcbf3bd6SGeorge Wilson /* 5299dcbf3bd6SGeorge Wilson * If the current arc_buf_t and the hdr are sharing their data 53005602294fSDan Kimmel * buffer, then we must stop sharing that block. 5301dcbf3bd6SGeorge Wilson */ 5302dcbf3bd6SGeorge Wilson if (arc_buf_is_shared(buf)) { 5303dcbf3bd6SGeorge Wilson VERIFY(!arc_buf_is_shared(lastbuf)); 5304ea8dc4b6Seschrock 5305dcbf3bd6SGeorge Wilson /* 5306dcbf3bd6SGeorge Wilson * First, sever the block sharing relationship between 53075602294fSDan Kimmel * buf and the arc_buf_hdr_t. 5308dcbf3bd6SGeorge Wilson */ 5309dcbf3bd6SGeorge Wilson arc_unshare_buf(hdr, buf); 53105602294fSDan Kimmel 53115602294fSDan Kimmel /* 5312770499e1SDan Kimmel * Now we need to recreate the hdr's b_pabd. Since we 53135602294fSDan Kimmel * have lastbuf handy, we try to share with it, but if 5314770499e1SDan Kimmel * we can't then we allocate a new b_pabd and copy the 53155602294fSDan Kimmel * data from buf into it. 53165602294fSDan Kimmel */ 53175602294fSDan Kimmel if (arc_can_share(hdr, lastbuf)) { 53185602294fSDan Kimmel arc_share_buf(hdr, lastbuf); 53195602294fSDan Kimmel } else { 5320770499e1SDan Kimmel arc_hdr_alloc_pabd(hdr); 5321770499e1SDan Kimmel abd_copy_from_buf(hdr->b_l1hdr.b_pabd, 5322770499e1SDan Kimmel buf->b_data, psize); 53235602294fSDan Kimmel } 5324dcbf3bd6SGeorge Wilson VERIFY3P(lastbuf->b_data, !=, NULL); 5325dcbf3bd6SGeorge Wilson } else if (HDR_SHARED_DATA(hdr)) { 53265602294fSDan Kimmel /* 53275602294fSDan Kimmel * Uncompressed shared buffers are always at the end 53285602294fSDan Kimmel * of the list. Compressed buffers don't have the 53295602294fSDan Kimmel * same requirements. This makes it hard to 53305602294fSDan Kimmel * simply assert that the lastbuf is shared so 53315602294fSDan Kimmel * we rely on the hdr's compression flags to determine 53325602294fSDan Kimmel * if we have a compressed, shared buffer. 53335602294fSDan Kimmel */ 53345602294fSDan Kimmel ASSERT(arc_buf_is_shared(lastbuf) || 53355602294fSDan Kimmel HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 53365602294fSDan Kimmel ASSERT(!ARC_BUF_SHARED(buf)); 5337dcbf3bd6SGeorge Wilson } 5338770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 533989c86e32SChris Williamson ASSERT3P(state, !=, arc_l2c_only); 53402fd872a7SPrakash Surya 5341dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_size, 53425602294fSDan Kimmel arc_buf_size(buf), buf); 53432fd872a7SPrakash Surya 534489c86e32SChris Williamson if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 534589c86e32SChris Williamson ASSERT3P(state, !=, arc_l2c_only); 5346dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 53475602294fSDan Kimmel arc_buf_size(buf), buf); 5348ea8dc4b6Seschrock } 53499253d63dSGeorge Wilson 5350dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt -= 1; 5351c717a561Smaybee arc_cksum_verify(buf); 5352cd1c8b85SMatthew Ahrens arc_buf_unwatch(buf); 5353ea8dc4b6Seschrock 5354fa9e4066Sahrens mutex_exit(hash_lock); 5355fa9e4066Sahrens 5356dcbf3bd6SGeorge Wilson /* 5357770499e1SDan Kimmel * Allocate a new hdr. The new hdr will contain a b_pabd 5358dcbf3bd6SGeorge Wilson * buffer which will be freed in arc_write(). 5359dcbf3bd6SGeorge Wilson */ 5360dcbf3bd6SGeorge Wilson nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type); 5361dcbf3bd6SGeorge Wilson ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); 5362dcbf3bd6SGeorge Wilson ASSERT0(nhdr->b_l1hdr.b_bufcnt); 5363dcbf3bd6SGeorge Wilson ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt)); 5364dcbf3bd6SGeorge Wilson VERIFY3U(nhdr->b_type, ==, type); 5365dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(nhdr)); 536689c86e32SChris Williamson 536789c86e32SChris Williamson nhdr->b_l1hdr.b_buf = buf; 5368dcbf3bd6SGeorge Wilson nhdr->b_l1hdr.b_bufcnt = 1; 536989c86e32SChris Williamson (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 5370af2c4821Smaybee buf->b_hdr = nhdr; 5371dcbf3bd6SGeorge Wilson 53723f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 5373dcbf3bd6SGeorge Wilson (void) refcount_add_many(&arc_anon->arcs_size, 53745602294fSDan Kimmel arc_buf_size(buf), buf); 5375fa9e4066Sahrens } else { 53763f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 537789c86e32SChris Williamson ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 5378244781f1SPrakash Surya /* protected by hash lock, or hdr is on arc_anon */ 5379244781f1SPrakash Surya ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5380fa9e4066Sahrens ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 538189c86e32SChris Williamson arc_change_state(arc_anon, hdr, hash_lock); 538289c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = 0; 538389c86e32SChris Williamson mutex_exit(hash_lock); 5384fa94a07fSbrendan 53853f9d6ad7SLin Ling buf_discard_identity(hdr); 5386c717a561Smaybee arc_buf_thaw(buf); 5387fa9e4066Sahrens } 5388fa9e4066Sahrens } 5389fa9e4066Sahrens 5390fa9e4066Sahrens int 5391fa9e4066Sahrens arc_released(arc_buf_t *buf) 5392fa9e4066Sahrens { 53936f83844dSMark Maybee int released; 53946f83844dSMark Maybee 53953f9d6ad7SLin Ling mutex_enter(&buf->b_evict_lock); 539689c86e32SChris Williamson released = (buf->b_data != NULL && 539789c86e32SChris Williamson buf->b_hdr->b_l1hdr.b_state == arc_anon); 53983f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 53996f83844dSMark Maybee return (released); 5400ea8dc4b6Seschrock } 5401ea8dc4b6Seschrock 5402ea8dc4b6Seschrock #ifdef ZFS_DEBUG 5403ea8dc4b6Seschrock int 5404ea8dc4b6Seschrock arc_referenced(arc_buf_t *buf) 5405ea8dc4b6Seschrock { 54066f83844dSMark Maybee int referenced; 54076f83844dSMark Maybee 54083f9d6ad7SLin Ling mutex_enter(&buf->b_evict_lock); 540989c86e32SChris Williamson referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 54103f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 54116f83844dSMark Maybee return (referenced); 5412ea8dc4b6Seschrock } 5413ea8dc4b6Seschrock #endif 5414ea8dc4b6Seschrock 5415c717a561Smaybee static void 5416c717a561Smaybee arc_write_ready(zio_t *zio) 5417c717a561Smaybee { 5418c717a561Smaybee arc_write_callback_t *callback = zio->io_private; 5419c717a561Smaybee arc_buf_t *buf = callback->awcb_buf; 54200a4e9518Sgw arc_buf_hdr_t *hdr = buf->b_hdr; 5421dcbf3bd6SGeorge Wilson uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp); 5422c717a561Smaybee 542389c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 542489c86e32SChris Williamson ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 5425dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 5426e14bb325SJeff Bonwick 54270a4e9518Sgw /* 5428dcbf3bd6SGeorge Wilson * If we're reexecuting this zio because the pool suspended, then 5429dcbf3bd6SGeorge Wilson * cleanup any state that was previously set the first time the 54305602294fSDan Kimmel * callback was invoked. 54310a4e9518Sgw */ 5432dcbf3bd6SGeorge Wilson if (zio->io_flags & ZIO_FLAG_REEXECUTED) { 5433dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 5434dcbf3bd6SGeorge Wilson arc_buf_unwatch(buf); 5435770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 5436dcbf3bd6SGeorge Wilson if (arc_buf_is_shared(buf)) { 5437dcbf3bd6SGeorge Wilson arc_unshare_buf(hdr, buf); 5438dcbf3bd6SGeorge Wilson } else { 5439770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 5440dcbf3bd6SGeorge Wilson } 54410a4e9518Sgw } 54420a4e9518Sgw } 5443770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5444dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(hdr)); 5445dcbf3bd6SGeorge Wilson ASSERT(!arc_buf_is_shared(buf)); 5446dcbf3bd6SGeorge Wilson 5447dcbf3bd6SGeorge Wilson callback->awcb_ready(zio, buf, callback->awcb_private); 5448dcbf3bd6SGeorge Wilson 5449dcbf3bd6SGeorge Wilson if (HDR_IO_IN_PROGRESS(hdr)) 5450dcbf3bd6SGeorge Wilson ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); 5451dcbf3bd6SGeorge Wilson 5452dcbf3bd6SGeorge Wilson arc_cksum_compute(buf); 5453dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5454dcbf3bd6SGeorge Wilson 5455dcbf3bd6SGeorge Wilson enum zio_compress compress; 5456dcbf3bd6SGeorge Wilson if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5457dcbf3bd6SGeorge Wilson compress = ZIO_COMPRESS_OFF; 5458dcbf3bd6SGeorge Wilson } else { 5459dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp)); 5460dcbf3bd6SGeorge Wilson compress = BP_GET_COMPRESS(zio->io_bp); 5461dcbf3bd6SGeorge Wilson } 5462dcbf3bd6SGeorge Wilson HDR_SET_PSIZE(hdr, psize); 5463dcbf3bd6SGeorge Wilson arc_hdr_set_compress(hdr, compress); 5464dcbf3bd6SGeorge Wilson 5465770499e1SDan Kimmel 5466dcbf3bd6SGeorge Wilson /* 5467770499e1SDan Kimmel * Fill the hdr with data. If the hdr is compressed, the data we want 5468770499e1SDan Kimmel * is available from the zio, otherwise we can take it from the buf. 5469770499e1SDan Kimmel * 5470770499e1SDan Kimmel * We might be able to share the buf's data with the hdr here. However, 5471770499e1SDan Kimmel * doing so would cause the ARC to be full of linear ABDs if we write a 5472770499e1SDan Kimmel * lot of shareable data. As a compromise, we check whether scattered 5473770499e1SDan Kimmel * ABDs are allowed, and assume that if they are then the user wants 5474770499e1SDan Kimmel * the ARC to be primarily filled with them regardless of the data being 5475770499e1SDan Kimmel * written. Therefore, if they're allowed then we allocate one and copy 5476770499e1SDan Kimmel * the data into it; otherwise, we share the data directly if we can. 5477dcbf3bd6SGeorge Wilson */ 5478770499e1SDan Kimmel if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) { 5479770499e1SDan Kimmel arc_hdr_alloc_pabd(hdr); 5480770499e1SDan Kimmel 5481770499e1SDan Kimmel /* 5482770499e1SDan Kimmel * Ideally, we would always copy the io_abd into b_pabd, but the 5483770499e1SDan Kimmel * user may have disabled compressed ARC, thus we must check the 5484770499e1SDan Kimmel * hdr's compression setting rather than the io_bp's. 5485770499e1SDan Kimmel */ 5486770499e1SDan Kimmel if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5487770499e1SDan Kimmel ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=, 5488770499e1SDan Kimmel ZIO_COMPRESS_OFF); 5489770499e1SDan Kimmel ASSERT3U(psize, >, 0); 5490770499e1SDan Kimmel 5491770499e1SDan Kimmel abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize); 5492770499e1SDan Kimmel } else { 5493770499e1SDan Kimmel ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr)); 5494770499e1SDan Kimmel 5495770499e1SDan Kimmel abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data, 5496770499e1SDan Kimmel arc_buf_size(buf)); 5497770499e1SDan Kimmel } 5498dcbf3bd6SGeorge Wilson } else { 5499770499e1SDan Kimmel ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd)); 55005602294fSDan Kimmel ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf)); 5501dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5502dcbf3bd6SGeorge Wilson 5503dcbf3bd6SGeorge Wilson arc_share_buf(hdr, buf); 5504dcbf3bd6SGeorge Wilson } 5505770499e1SDan Kimmel 5506dcbf3bd6SGeorge Wilson arc_hdr_verify(hdr, zio->io_bp); 5507c717a561Smaybee } 5508c717a561Smaybee 55098df0bcf0SPaul Dagnelie static void 55108df0bcf0SPaul Dagnelie arc_write_children_ready(zio_t *zio) 55118df0bcf0SPaul Dagnelie { 55128df0bcf0SPaul Dagnelie arc_write_callback_t *callback = zio->io_private; 55138df0bcf0SPaul Dagnelie arc_buf_t *buf = callback->awcb_buf; 55148df0bcf0SPaul Dagnelie 55158df0bcf0SPaul Dagnelie callback->awcb_children_ready(zio, buf, callback->awcb_private); 55168df0bcf0SPaul Dagnelie } 55178df0bcf0SPaul Dagnelie 551869962b56SMatthew Ahrens /* 551969962b56SMatthew Ahrens * The SPA calls this callback for each physical write that happens on behalf 552069962b56SMatthew Ahrens * of a logical write. See the comment in dbuf_write_physdone() for details. 552169962b56SMatthew Ahrens */ 552269962b56SMatthew Ahrens static void 552369962b56SMatthew Ahrens arc_write_physdone(zio_t *zio) 552469962b56SMatthew Ahrens { 552569962b56SMatthew Ahrens arc_write_callback_t *cb = zio->io_private; 552669962b56SMatthew Ahrens if (cb->awcb_physdone != NULL) 552769962b56SMatthew Ahrens cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 552869962b56SMatthew Ahrens } 552969962b56SMatthew Ahrens 5530fa9e4066Sahrens static void 5531fa9e4066Sahrens arc_write_done(zio_t *zio) 5532fa9e4066Sahrens { 5533c717a561Smaybee arc_write_callback_t *callback = zio->io_private; 5534c717a561Smaybee arc_buf_t *buf = callback->awcb_buf; 5535c717a561Smaybee arc_buf_hdr_t *hdr = buf->b_hdr; 5536fa9e4066Sahrens 5537dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5538b24ab676SJeff Bonwick 5539b24ab676SJeff Bonwick if (zio->io_error == 0) { 5540dcbf3bd6SGeorge Wilson arc_hdr_verify(hdr, zio->io_bp); 5541dcbf3bd6SGeorge Wilson 55425d7b4d43SMatthew Ahrens if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 554343466aaeSMax Grossman buf_discard_identity(hdr); 554443466aaeSMax Grossman } else { 554543466aaeSMax Grossman hdr->b_dva = *BP_IDENTITY(zio->io_bp); 554643466aaeSMax Grossman hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 554743466aaeSMax Grossman } 5548b24ab676SJeff Bonwick } else { 5549dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 5550b24ab676SJeff Bonwick } 5551fa9e4066Sahrens 5552ea8dc4b6Seschrock /* 55535d7b4d43SMatthew Ahrens * If the block to be written was all-zero or compressed enough to be 55545d7b4d43SMatthew Ahrens * embedded in the BP, no write was performed so there will be no 55555d7b4d43SMatthew Ahrens * dva/birth/checksum. The buffer must therefore remain anonymous 55565d7b4d43SMatthew Ahrens * (and uncached). 5557ea8dc4b6Seschrock */ 5558dcbf3bd6SGeorge Wilson if (!HDR_EMPTY(hdr)) { 5559fa9e4066Sahrens arc_buf_hdr_t *exists; 5560fa9e4066Sahrens kmutex_t *hash_lock; 5561fa9e4066Sahrens 55625602294fSDan Kimmel ASSERT3U(zio->io_error, ==, 0); 5563b24ab676SJeff Bonwick 55646b4acc8bSahrens arc_cksum_verify(buf); 55656b4acc8bSahrens 5566fa9e4066Sahrens exists = buf_hash_insert(hdr, &hash_lock); 556789c86e32SChris Williamson if (exists != NULL) { 5568fa9e4066Sahrens /* 5569fa9e4066Sahrens * This can only happen if we overwrite for 5570fa9e4066Sahrens * sync-to-convergence, because we remove 5571fa9e4066Sahrens * buffers from the hash table when we arc_free(). 5572fa9e4066Sahrens */ 5573b24ab676SJeff Bonwick if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 5574b24ab676SJeff Bonwick if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5575b24ab676SJeff Bonwick panic("bad overwrite, hdr=%p exists=%p", 5576b24ab676SJeff Bonwick (void *)hdr, (void *)exists); 557789c86e32SChris Williamson ASSERT(refcount_is_zero( 557889c86e32SChris Williamson &exists->b_l1hdr.b_refcnt)); 5579b24ab676SJeff Bonwick arc_change_state(arc_anon, exists, hash_lock); 5580b24ab676SJeff Bonwick mutex_exit(hash_lock); 5581b24ab676SJeff Bonwick arc_hdr_destroy(exists); 5582b24ab676SJeff Bonwick exists = buf_hash_insert(hdr, &hash_lock); 5583b24ab676SJeff Bonwick ASSERT3P(exists, ==, NULL); 558480901aeaSGeorge Wilson } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 558580901aeaSGeorge Wilson /* nopwrite */ 558680901aeaSGeorge Wilson ASSERT(zio->io_prop.zp_nopwrite); 558780901aeaSGeorge Wilson if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 558880901aeaSGeorge Wilson panic("bad nopwrite, hdr=%p exists=%p", 558980901aeaSGeorge Wilson (void *)hdr, (void *)exists); 5590b24ab676SJeff Bonwick } else { 5591b24ab676SJeff Bonwick /* Dedup */ 5592dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt == 1); 559389c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5594b24ab676SJeff Bonwick ASSERT(BP_GET_DEDUP(zio->io_bp)); 5595b24ab676SJeff Bonwick ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5596ae46e4c7SMatthew Ahrens } 5597fa9e4066Sahrens } 5598dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5599088f3894Sahrens /* if it's not anon, we are doing a scrub */ 560089c86e32SChris Williamson if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5601088f3894Sahrens arc_access(hdr, hash_lock); 560244eda4d7Smaybee mutex_exit(hash_lock); 5603ea8dc4b6Seschrock } else { 5604dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5605fa9e4066Sahrens } 5606ea8dc4b6Seschrock 560789c86e32SChris Williamson ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5608b24ab676SJeff Bonwick callback->awcb_done(zio, buf, callback->awcb_private); 5609fa9e4066Sahrens 5610770499e1SDan Kimmel abd_put(zio->io_abd); 5611c717a561Smaybee kmem_free(callback, sizeof (arc_write_callback_t)); 5612fa9e4066Sahrens } 5613fa9e4066Sahrens 5614c717a561Smaybee zio_t * 5615dcbf3bd6SGeorge Wilson arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 5616dcbf3bd6SGeorge Wilson boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready, 56178df0bcf0SPaul Dagnelie arc_done_func_t *children_ready, arc_done_func_t *physdone, 561869962b56SMatthew Ahrens arc_done_func_t *done, void *private, zio_priority_t priority, 56197802d7bfSMatthew Ahrens int zio_flags, const zbookmark_phys_t *zb) 5620fa9e4066Sahrens { 5621fa9e4066Sahrens arc_buf_hdr_t *hdr = buf->b_hdr; 5622c717a561Smaybee arc_write_callback_t *callback; 5623e14bb325SJeff Bonwick zio_t *zio; 5624adaec86aSMatthew Ahrens zio_prop_t localprop = *zp; 5625fa9e4066Sahrens 5626dcbf3bd6SGeorge Wilson ASSERT3P(ready, !=, NULL); 5627dcbf3bd6SGeorge Wilson ASSERT3P(done, !=, NULL); 5628fa9e4066Sahrens ASSERT(!HDR_IO_ERROR(hdr)); 562989c86e32SChris Williamson ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5630dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5631dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 56323baa08fcSek if (l2arc) 5633dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 56345602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 5635adaec86aSMatthew Ahrens /* 5636adaec86aSMatthew Ahrens * We're writing a pre-compressed buffer. Make the 5637adaec86aSMatthew Ahrens * compression algorithm requested by the zio_prop_t match 5638adaec86aSMatthew Ahrens * the pre-compressed buffer's compression algorithm. 5639adaec86aSMatthew Ahrens */ 5640adaec86aSMatthew Ahrens localprop.zp_compress = HDR_GET_COMPRESS(hdr); 5641adaec86aSMatthew Ahrens 56425602294fSDan Kimmel ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf)); 56435602294fSDan Kimmel zio_flags |= ZIO_FLAG_RAW; 56445602294fSDan Kimmel } 5645c717a561Smaybee callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5646c717a561Smaybee callback->awcb_ready = ready; 56478df0bcf0SPaul Dagnelie callback->awcb_children_ready = children_ready; 564869962b56SMatthew Ahrens callback->awcb_physdone = physdone; 5649c717a561Smaybee callback->awcb_done = done; 5650c717a561Smaybee callback->awcb_private = private; 5651c717a561Smaybee callback->awcb_buf = buf; 5652088f3894Sahrens 5653dcbf3bd6SGeorge Wilson /* 5654770499e1SDan Kimmel * The hdr's b_pabd is now stale, free it now. A new data block 5655dcbf3bd6SGeorge Wilson * will be allocated when the zio pipeline calls arc_write_ready(). 5656dcbf3bd6SGeorge Wilson */ 5657770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 5658dcbf3bd6SGeorge Wilson /* 5659dcbf3bd6SGeorge Wilson * If the buf is currently sharing the data block with 5660dcbf3bd6SGeorge Wilson * the hdr then we need to break that relationship here. 5661dcbf3bd6SGeorge Wilson * The hdr will remain with a NULL data pointer and the 5662dcbf3bd6SGeorge Wilson * buf will take sole ownership of the block. 5663dcbf3bd6SGeorge Wilson */ 5664dcbf3bd6SGeorge Wilson if (arc_buf_is_shared(buf)) { 5665dcbf3bd6SGeorge Wilson arc_unshare_buf(hdr, buf); 5666dcbf3bd6SGeorge Wilson } else { 5667770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 5668dcbf3bd6SGeorge Wilson } 5669dcbf3bd6SGeorge Wilson VERIFY3P(buf->b_data, !=, NULL); 5670dcbf3bd6SGeorge Wilson arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); 5671dcbf3bd6SGeorge Wilson } 5672dcbf3bd6SGeorge Wilson ASSERT(!arc_buf_is_shared(buf)); 5673770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5674dcbf3bd6SGeorge Wilson 5675770499e1SDan Kimmel zio = zio_write(pio, spa, txg, bp, 5676770499e1SDan Kimmel abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)), 5677adaec86aSMatthew Ahrens HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready, 56788df0bcf0SPaul Dagnelie (children_ready != NULL) ? arc_write_children_ready : NULL, 56798df0bcf0SPaul Dagnelie arc_write_physdone, arc_write_done, callback, 568069962b56SMatthew Ahrens priority, zio_flags, zb); 5681fa9e4066Sahrens 5682c717a561Smaybee return (zio); 5683fa9e4066Sahrens } 5684fa9e4066Sahrens 56851ab7f2deSmaybee static int 568669962b56SMatthew Ahrens arc_memory_throttle(uint64_t reserve, uint64_t txg) 56871ab7f2deSmaybee { 56881ab7f2deSmaybee #ifdef _KERNEL 56891ab7f2deSmaybee uint64_t available_memory = ptob(freemem); 56901ab7f2deSmaybee static uint64_t page_load = 0; 56911ab7f2deSmaybee static uint64_t last_txg = 0; 56921ab7f2deSmaybee 56931ab7f2deSmaybee #if defined(__i386) 56941ab7f2deSmaybee available_memory = 56951ab7f2deSmaybee MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 56961ab7f2deSmaybee #endif 569769962b56SMatthew Ahrens 569869962b56SMatthew Ahrens if (freemem > physmem * arc_lotsfree_percent / 100) 56991ab7f2deSmaybee return (0); 57001ab7f2deSmaybee 57011ab7f2deSmaybee if (txg > last_txg) { 57021ab7f2deSmaybee last_txg = txg; 57031ab7f2deSmaybee page_load = 0; 57041ab7f2deSmaybee } 57051ab7f2deSmaybee /* 57061ab7f2deSmaybee * If we are in pageout, we know that memory is already tight, 57071ab7f2deSmaybee * the arc is already going to be evicting, so we just want to 57081ab7f2deSmaybee * continue to let page writes occur as quickly as possible. 57091ab7f2deSmaybee */ 57101ab7f2deSmaybee if (curproc == proc_pageout) { 57111ab7f2deSmaybee if (page_load > MAX(ptob(minfree), available_memory) / 4) 5712be6fd75aSMatthew Ahrens return (SET_ERROR(ERESTART)); 57131ab7f2deSmaybee /* Note: reserve is inflated, so we deflate */ 57141ab7f2deSmaybee page_load += reserve / 8; 57151ab7f2deSmaybee return (0); 57161ab7f2deSmaybee } else if (page_load > 0 && arc_reclaim_needed()) { 57171ab7f2deSmaybee /* memory is low, delay before restarting */ 57181ab7f2deSmaybee ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5719be6fd75aSMatthew Ahrens return (SET_ERROR(EAGAIN)); 57201ab7f2deSmaybee } 57211ab7f2deSmaybee page_load = 0; 57221ab7f2deSmaybee #endif 57231ab7f2deSmaybee return (0); 57241ab7f2deSmaybee } 57251ab7f2deSmaybee 5726fa9e4066Sahrens void 57271ab7f2deSmaybee arc_tempreserve_clear(uint64_t reserve) 5728fa9e4066Sahrens { 57291ab7f2deSmaybee atomic_add_64(&arc_tempreserve, -reserve); 5730fa9e4066Sahrens ASSERT((int64_t)arc_tempreserve >= 0); 5731fa9e4066Sahrens } 5732fa9e4066Sahrens 5733fa9e4066Sahrens int 57341ab7f2deSmaybee arc_tempreserve_space(uint64_t reserve, uint64_t txg) 5735fa9e4066Sahrens { 57361ab7f2deSmaybee int error; 57372fdbea25SAleksandr Guzovskiy uint64_t anon_size; 57381ab7f2deSmaybee 57391ab7f2deSmaybee if (reserve > arc_c/4 && !arc_no_grow) 57401ab7f2deSmaybee arc_c = MIN(arc_c_max, reserve * 4); 57411ab7f2deSmaybee if (reserve > arc_c) 5742be6fd75aSMatthew Ahrens return (SET_ERROR(ENOMEM)); 5743112fe045Smaybee 57442fdbea25SAleksandr Guzovskiy /* 57452fdbea25SAleksandr Guzovskiy * Don't count loaned bufs as in flight dirty data to prevent long 57462fdbea25SAleksandr Guzovskiy * network delays from blocking transactions that are ready to be 57472fdbea25SAleksandr Guzovskiy * assigned to a txg. 57482fdbea25SAleksandr Guzovskiy */ 57495602294fSDan Kimmel 57505602294fSDan Kimmel /* assert that it has not wrapped around */ 57515602294fSDan Kimmel ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 57525602294fSDan Kimmel 57532fd872a7SPrakash Surya anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 57542fd872a7SPrakash Surya arc_loaned_bytes), 0); 57552fdbea25SAleksandr Guzovskiy 57561ab7f2deSmaybee /* 57571ab7f2deSmaybee * Writes will, almost always, require additional memory allocations 5758f7170741SWill Andrews * in order to compress/encrypt/etc the data. We therefore need to 57591ab7f2deSmaybee * make sure that there is sufficient available memory for this. 57601ab7f2deSmaybee */ 576169962b56SMatthew Ahrens error = arc_memory_throttle(reserve, txg); 576269962b56SMatthew Ahrens if (error != 0) 57631ab7f2deSmaybee return (error); 57641ab7f2deSmaybee 5765fa9e4066Sahrens /* 5766112fe045Smaybee * Throttle writes when the amount of dirty data in the cache 5767112fe045Smaybee * gets too large. We try to keep the cache less than half full 5768112fe045Smaybee * of dirty blocks so that our sync times don't grow too large. 5769112fe045Smaybee * Note: if two requests come in concurrently, we might let them 5770112fe045Smaybee * both succeed, when one of them should fail. Not a huge deal. 5771fa9e4066Sahrens */ 57722fdbea25SAleksandr Guzovskiy 57732fdbea25SAleksandr Guzovskiy if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 57742fdbea25SAleksandr Guzovskiy anon_size > arc_c / 4) { 5775dcbf3bd6SGeorge Wilson uint64_t meta_esize = 5776dcbf3bd6SGeorge Wilson refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5777dcbf3bd6SGeorge Wilson uint64_t data_esize = 5778dcbf3bd6SGeorge Wilson refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 57790e8c6158Smaybee dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 57800e8c6158Smaybee "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5781dcbf3bd6SGeorge Wilson arc_tempreserve >> 10, meta_esize >> 10, 5782dcbf3bd6SGeorge Wilson data_esize >> 10, reserve >> 10, arc_c >> 10); 5783be6fd75aSMatthew Ahrens return (SET_ERROR(ERESTART)); 5784fa9e4066Sahrens } 57851ab7f2deSmaybee atomic_add_64(&arc_tempreserve, reserve); 5786fa9e4066Sahrens return (0); 5787fa9e4066Sahrens } 5788fa9e4066Sahrens 57894076b1bfSPrakash Surya static void 57904076b1bfSPrakash Surya arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 57914076b1bfSPrakash Surya kstat_named_t *evict_data, kstat_named_t *evict_metadata) 57924076b1bfSPrakash Surya { 57932fd872a7SPrakash Surya size->value.ui64 = refcount_count(&state->arcs_size); 5794dcbf3bd6SGeorge Wilson evict_data->value.ui64 = 5795dcbf3bd6SGeorge Wilson refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); 5796dcbf3bd6SGeorge Wilson evict_metadata->value.ui64 = 5797dcbf3bd6SGeorge Wilson refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); 57984076b1bfSPrakash Surya } 57994076b1bfSPrakash Surya 58004076b1bfSPrakash Surya static int 58014076b1bfSPrakash Surya arc_kstat_update(kstat_t *ksp, int rw) 58024076b1bfSPrakash Surya { 58034076b1bfSPrakash Surya arc_stats_t *as = ksp->ks_data; 58044076b1bfSPrakash Surya 58054076b1bfSPrakash Surya if (rw == KSTAT_WRITE) { 58064076b1bfSPrakash Surya return (EACCES); 58074076b1bfSPrakash Surya } else { 58084076b1bfSPrakash Surya arc_kstat_update_state(arc_anon, 58094076b1bfSPrakash Surya &as->arcstat_anon_size, 58104076b1bfSPrakash Surya &as->arcstat_anon_evictable_data, 58114076b1bfSPrakash Surya &as->arcstat_anon_evictable_metadata); 58124076b1bfSPrakash Surya arc_kstat_update_state(arc_mru, 58134076b1bfSPrakash Surya &as->arcstat_mru_size, 58144076b1bfSPrakash Surya &as->arcstat_mru_evictable_data, 58154076b1bfSPrakash Surya &as->arcstat_mru_evictable_metadata); 58164076b1bfSPrakash Surya arc_kstat_update_state(arc_mru_ghost, 58174076b1bfSPrakash Surya &as->arcstat_mru_ghost_size, 58184076b1bfSPrakash Surya &as->arcstat_mru_ghost_evictable_data, 58194076b1bfSPrakash Surya &as->arcstat_mru_ghost_evictable_metadata); 58204076b1bfSPrakash Surya arc_kstat_update_state(arc_mfu, 58214076b1bfSPrakash Surya &as->arcstat_mfu_size, 58224076b1bfSPrakash Surya &as->arcstat_mfu_evictable_data, 58234076b1bfSPrakash Surya &as->arcstat_mfu_evictable_metadata); 58244076b1bfSPrakash Surya arc_kstat_update_state(arc_mfu_ghost, 58254076b1bfSPrakash Surya &as->arcstat_mfu_ghost_size, 58264076b1bfSPrakash Surya &as->arcstat_mfu_ghost_evictable_data, 58274076b1bfSPrakash Surya &as->arcstat_mfu_ghost_evictable_metadata); 58284076b1bfSPrakash Surya } 58294076b1bfSPrakash Surya 58304076b1bfSPrakash Surya return (0); 58314076b1bfSPrakash Surya } 58324076b1bfSPrakash Surya 5833dcbf3bd6SGeorge Wilson /* 5834dcbf3bd6SGeorge Wilson * This function *must* return indices evenly distributed between all 5835dcbf3bd6SGeorge Wilson * sublists of the multilist. This is needed due to how the ARC eviction 5836dcbf3bd6SGeorge Wilson * code is laid out; arc_evict_state() assumes ARC buffers are evenly 5837dcbf3bd6SGeorge Wilson * distributed between all sublists and uses this assumption when 5838dcbf3bd6SGeorge Wilson * deciding which sublist to evict from and how much to evict from it. 5839dcbf3bd6SGeorge Wilson */ 5840dcbf3bd6SGeorge Wilson unsigned int 5841dcbf3bd6SGeorge Wilson arc_state_multilist_index_func(multilist_t *ml, void *obj) 5842dcbf3bd6SGeorge Wilson { 5843dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = obj; 5844dcbf3bd6SGeorge Wilson 5845dcbf3bd6SGeorge Wilson /* 5846dcbf3bd6SGeorge Wilson * We rely on b_dva to generate evenly distributed index 5847dcbf3bd6SGeorge Wilson * numbers using buf_hash below. So, as an added precaution, 5848dcbf3bd6SGeorge Wilson * let's make sure we never add empty buffers to the arc lists. 5849dcbf3bd6SGeorge Wilson */ 5850dcbf3bd6SGeorge Wilson ASSERT(!HDR_EMPTY(hdr)); 5851dcbf3bd6SGeorge Wilson 5852dcbf3bd6SGeorge Wilson /* 5853dcbf3bd6SGeorge Wilson * The assumption here, is the hash value for a given 5854dcbf3bd6SGeorge Wilson * arc_buf_hdr_t will remain constant throughout it's lifetime 5855dcbf3bd6SGeorge Wilson * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 5856dcbf3bd6SGeorge Wilson * Thus, we don't need to store the header's sublist index 5857dcbf3bd6SGeorge Wilson * on insertion, as this index can be recalculated on removal. 5858dcbf3bd6SGeorge Wilson * 5859dcbf3bd6SGeorge Wilson * Also, the low order bits of the hash value are thought to be 5860dcbf3bd6SGeorge Wilson * distributed evenly. Otherwise, in the case that the multilist 5861dcbf3bd6SGeorge Wilson * has a power of two number of sublists, each sublists' usage 5862dcbf3bd6SGeorge Wilson * would not be evenly distributed. 5863dcbf3bd6SGeorge Wilson */ 5864dcbf3bd6SGeorge Wilson return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 5865dcbf3bd6SGeorge Wilson multilist_get_num_sublists(ml)); 5866dcbf3bd6SGeorge Wilson } 5867dcbf3bd6SGeorge Wilson 5868dcbf3bd6SGeorge Wilson static void 5869dcbf3bd6SGeorge Wilson arc_state_init(void) 5870dcbf3bd6SGeorge Wilson { 5871dcbf3bd6SGeorge Wilson arc_anon = &ARC_anon; 5872dcbf3bd6SGeorge Wilson arc_mru = &ARC_mru; 5873dcbf3bd6SGeorge Wilson arc_mru_ghost = &ARC_mru_ghost; 5874dcbf3bd6SGeorge Wilson arc_mfu = &ARC_mfu; 5875dcbf3bd6SGeorge Wilson arc_mfu_ghost = &ARC_mfu_ghost; 5876dcbf3bd6SGeorge Wilson arc_l2c_only = &ARC_l2c_only; 5877dcbf3bd6SGeorge Wilson 587894c2d0ebSMatthew Ahrens arc_mru->arcs_list[ARC_BUFC_METADATA] = 587994c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5880dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 588110fbdecbSMatthew Ahrens arc_state_multilist_index_func); 588294c2d0ebSMatthew Ahrens arc_mru->arcs_list[ARC_BUFC_DATA] = 588394c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5884dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 588510fbdecbSMatthew Ahrens arc_state_multilist_index_func); 588694c2d0ebSMatthew Ahrens arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] = 588794c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5888dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 588910fbdecbSMatthew Ahrens arc_state_multilist_index_func); 589094c2d0ebSMatthew Ahrens arc_mru_ghost->arcs_list[ARC_BUFC_DATA] = 589194c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5892dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 589310fbdecbSMatthew Ahrens arc_state_multilist_index_func); 589494c2d0ebSMatthew Ahrens arc_mfu->arcs_list[ARC_BUFC_METADATA] = 589594c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5896dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 589710fbdecbSMatthew Ahrens arc_state_multilist_index_func); 589894c2d0ebSMatthew Ahrens arc_mfu->arcs_list[ARC_BUFC_DATA] = 589994c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5900dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 590110fbdecbSMatthew Ahrens arc_state_multilist_index_func); 590294c2d0ebSMatthew Ahrens arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] = 590394c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5904dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 590510fbdecbSMatthew Ahrens arc_state_multilist_index_func); 590694c2d0ebSMatthew Ahrens arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] = 590794c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5908dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 590910fbdecbSMatthew Ahrens arc_state_multilist_index_func); 591094c2d0ebSMatthew Ahrens arc_l2c_only->arcs_list[ARC_BUFC_METADATA] = 591194c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5912dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 591310fbdecbSMatthew Ahrens arc_state_multilist_index_func); 591494c2d0ebSMatthew Ahrens arc_l2c_only->arcs_list[ARC_BUFC_DATA] = 591594c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5916dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 591710fbdecbSMatthew Ahrens arc_state_multilist_index_func); 5918dcbf3bd6SGeorge Wilson 5919dcbf3bd6SGeorge Wilson refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5920dcbf3bd6SGeorge Wilson refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5921dcbf3bd6SGeorge Wilson refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 5922dcbf3bd6SGeorge Wilson refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 5923dcbf3bd6SGeorge Wilson refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 5924dcbf3bd6SGeorge Wilson refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 5925dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 5926dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 5927dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 5928dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 5929dcbf3bd6SGeorge Wilson refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 5930dcbf3bd6SGeorge Wilson refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 5931dcbf3bd6SGeorge Wilson 5932dcbf3bd6SGeorge Wilson refcount_create(&arc_anon->arcs_size); 5933dcbf3bd6SGeorge Wilson refcount_create(&arc_mru->arcs_size); 5934dcbf3bd6SGeorge Wilson refcount_create(&arc_mru_ghost->arcs_size); 5935dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu->arcs_size); 5936dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu_ghost->arcs_size); 5937dcbf3bd6SGeorge Wilson refcount_create(&arc_l2c_only->arcs_size); 5938dcbf3bd6SGeorge Wilson } 5939dcbf3bd6SGeorge Wilson 5940dcbf3bd6SGeorge Wilson static void 5941dcbf3bd6SGeorge Wilson arc_state_fini(void) 5942dcbf3bd6SGeorge Wilson { 5943dcbf3bd6SGeorge Wilson refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5944dcbf3bd6SGeorge Wilson refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5945dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 5946dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 5947dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 5948dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 5949dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 5950dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 5951dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 5952dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 5953dcbf3bd6SGeorge Wilson refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 5954dcbf3bd6SGeorge Wilson refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 5955dcbf3bd6SGeorge Wilson 5956dcbf3bd6SGeorge Wilson refcount_destroy(&arc_anon->arcs_size); 5957dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru->arcs_size); 5958dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru_ghost->arcs_size); 5959dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu->arcs_size); 5960dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu_ghost->arcs_size); 5961dcbf3bd6SGeorge Wilson refcount_destroy(&arc_l2c_only->arcs_size); 5962dcbf3bd6SGeorge Wilson 596394c2d0ebSMatthew Ahrens multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]); 596494c2d0ebSMatthew Ahrens multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 596594c2d0ebSMatthew Ahrens multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]); 596694c2d0ebSMatthew Ahrens multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 596794c2d0ebSMatthew Ahrens multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]); 596894c2d0ebSMatthew Ahrens multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 596994c2d0ebSMatthew Ahrens multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]); 597094c2d0ebSMatthew Ahrens multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 5971dcbf3bd6SGeorge Wilson } 5972dcbf3bd6SGeorge Wilson 5973dcbf3bd6SGeorge Wilson uint64_t 5974dcbf3bd6SGeorge Wilson arc_max_bytes(void) 5975244781f1SPrakash Surya { 5976dcbf3bd6SGeorge Wilson return (arc_c_max); 5977244781f1SPrakash Surya } 5978244781f1SPrakash Surya 5979fa9e4066Sahrens void 5980fa9e4066Sahrens arc_init(void) 5981fa9e4066Sahrens { 59822ec99e3eSMatthew Ahrens /* 59832ec99e3eSMatthew Ahrens * allmem is "all memory that we could possibly use". 59842ec99e3eSMatthew Ahrens */ 59852ec99e3eSMatthew Ahrens #ifdef _KERNEL 59862ec99e3eSMatthew Ahrens uint64_t allmem = ptob(physmem - swapfs_minfree); 59872ec99e3eSMatthew Ahrens #else 59882ec99e3eSMatthew Ahrens uint64_t allmem = (physmem * PAGESIZE) / 2; 59892ec99e3eSMatthew Ahrens #endif 59902ec99e3eSMatthew Ahrens 5991244781f1SPrakash Surya mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 5992244781f1SPrakash Surya cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 5993244781f1SPrakash Surya cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 5994244781f1SPrakash Surya 599513506d1eSmaybee /* Convert seconds to clock ticks */ 5996b19a79ecSperrin arc_min_prefetch_lifespan = 1 * hz; 599713506d1eSmaybee 5998112fe045Smaybee /* set min cache to 1/32 of all memory, or 64MB, whichever is more */ 59992ec99e3eSMatthew Ahrens arc_c_min = MAX(allmem / 32, 64 << 20); 6000112fe045Smaybee /* set max to 3/4 of all memory, or all but 1GB, whichever is more */ 60012ec99e3eSMatthew Ahrens if (allmem >= 1 << 30) 60022ec99e3eSMatthew Ahrens arc_c_max = allmem - (1 << 30); 6003fa9e4066Sahrens else 600444cb6abcSbmc arc_c_max = arc_c_min; 60052ec99e3eSMatthew Ahrens arc_c_max = MAX(allmem * 3 / 4, arc_c_max); 6006a2eea2e1Sahrens 60078fe00bfbSMatthew Ahrens /* 60088fe00bfbSMatthew Ahrens * In userland, there's only the memory pressure that we artificially 60098fe00bfbSMatthew Ahrens * create (see arc_available_memory()). Don't let arc_c get too 60108fe00bfbSMatthew Ahrens * small, because it can cause transactions to be larger than 60118fe00bfbSMatthew Ahrens * arc_c, causing arc_tempreserve_space() to fail. 60128fe00bfbSMatthew Ahrens */ 60138fe00bfbSMatthew Ahrens #ifndef _KERNEL 60148fe00bfbSMatthew Ahrens arc_c_min = arc_c_max / 2; 60158fe00bfbSMatthew Ahrens #endif 60168fe00bfbSMatthew Ahrens 6017a2eea2e1Sahrens /* 6018a2eea2e1Sahrens * Allow the tunables to override our calculations if they are 6019a2eea2e1Sahrens * reasonable (ie. over 64MB) 6020a2eea2e1Sahrens */ 6021e5961710SMatthew Ahrens if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) { 602244cb6abcSbmc arc_c_max = zfs_arc_max; 6023e5961710SMatthew Ahrens arc_c_min = MIN(arc_c_min, arc_c_max); 6024e5961710SMatthew Ahrens } 60252ec99e3eSMatthew Ahrens if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max) 602644cb6abcSbmc arc_c_min = zfs_arc_min; 6027a2eea2e1Sahrens 602844cb6abcSbmc arc_c = arc_c_max; 602944cb6abcSbmc arc_p = (arc_c >> 1); 6030dcbf3bd6SGeorge Wilson arc_size = 0; 6031fa9e4066Sahrens 60320e8c6158Smaybee /* limit meta-data to 1/4 of the arc capacity */ 60330e8c6158Smaybee arc_meta_limit = arc_c_max / 4; 60341116048bSek 6035af868f46SMatthew Ahrens #ifdef _KERNEL 6036af868f46SMatthew Ahrens /* 6037af868f46SMatthew Ahrens * Metadata is stored in the kernel's heap. Don't let us 6038af868f46SMatthew Ahrens * use more than half the heap for the ARC. 6039af868f46SMatthew Ahrens */ 6040af868f46SMatthew Ahrens arc_meta_limit = MIN(arc_meta_limit, 6041af868f46SMatthew Ahrens vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2); 6042af868f46SMatthew Ahrens #endif 6043af868f46SMatthew Ahrens 60441116048bSek /* Allow the tunable to override if it is reasonable */ 60451116048bSek if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 60461116048bSek arc_meta_limit = zfs_arc_meta_limit; 60471116048bSek 60480e8c6158Smaybee if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 60490e8c6158Smaybee arc_c_min = arc_meta_limit / 2; 60500e8c6158Smaybee 60513a5286a1SMatthew Ahrens if (zfs_arc_meta_min > 0) { 60523a5286a1SMatthew Ahrens arc_meta_min = zfs_arc_meta_min; 60533a5286a1SMatthew Ahrens } else { 60543a5286a1SMatthew Ahrens arc_meta_min = arc_c_min / 2; 60553a5286a1SMatthew Ahrens } 60563a5286a1SMatthew Ahrens 60575a98e54bSBrendan Gregg - Sun Microsystems if (zfs_arc_grow_retry > 0) 60585a98e54bSBrendan Gregg - Sun Microsystems arc_grow_retry = zfs_arc_grow_retry; 60595a98e54bSBrendan Gregg - Sun Microsystems 60605a98e54bSBrendan Gregg - Sun Microsystems if (zfs_arc_shrink_shift > 0) 60615a98e54bSBrendan Gregg - Sun Microsystems arc_shrink_shift = zfs_arc_shrink_shift; 60625a98e54bSBrendan Gregg - Sun Microsystems 60632ec99e3eSMatthew Ahrens /* 60642ec99e3eSMatthew Ahrens * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 60652ec99e3eSMatthew Ahrens */ 60662ec99e3eSMatthew Ahrens if (arc_no_grow_shift >= arc_shrink_shift) 60672ec99e3eSMatthew Ahrens arc_no_grow_shift = arc_shrink_shift - 1; 60682ec99e3eSMatthew Ahrens 60695a98e54bSBrendan Gregg - Sun Microsystems if (zfs_arc_p_min_shift > 0) 60705a98e54bSBrendan Gregg - Sun Microsystems arc_p_min_shift = zfs_arc_p_min_shift; 60715a98e54bSBrendan Gregg - Sun Microsystems 6072fa9e4066Sahrens /* if kmem_flags are set, lets try to use less memory */ 6073fa9e4066Sahrens if (kmem_debugging()) 607444cb6abcSbmc arc_c = arc_c / 2; 607544cb6abcSbmc if (arc_c < arc_c_min) 607644cb6abcSbmc arc_c = arc_c_min; 607744cb6abcSbmc 6078dcbf3bd6SGeorge Wilson arc_state_init(); 6079fa9e4066Sahrens buf_init(); 6080fa9e4066Sahrens 6081dcbf3bd6SGeorge Wilson arc_reclaim_thread_exit = B_FALSE; 6082fa9e4066Sahrens 608344cb6abcSbmc arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 608444cb6abcSbmc sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 608544cb6abcSbmc 608644cb6abcSbmc if (arc_ksp != NULL) { 608744cb6abcSbmc arc_ksp->ks_data = &arc_stats; 60884076b1bfSPrakash Surya arc_ksp->ks_update = arc_kstat_update; 608944cb6abcSbmc kstat_install(arc_ksp); 609044cb6abcSbmc } 609144cb6abcSbmc 6092fa9e4066Sahrens (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 6093fa9e4066Sahrens TS_RUN, minclsyspri); 609449e3519aSmaybee 6095dcbf3bd6SGeorge Wilson arc_dead = B_FALSE; 60963a737e0dSbrendan arc_warm = B_FALSE; 60971ab7f2deSmaybee 609869962b56SMatthew Ahrens /* 609969962b56SMatthew Ahrens * Calculate maximum amount of dirty data per pool. 610069962b56SMatthew Ahrens * 610169962b56SMatthew Ahrens * If it has been set by /etc/system, take that. 610269962b56SMatthew Ahrens * Otherwise, use a percentage of physical memory defined by 610369962b56SMatthew Ahrens * zfs_dirty_data_max_percent (default 10%) with a cap at 610469962b56SMatthew Ahrens * zfs_dirty_data_max_max (default 4GB). 610569962b56SMatthew Ahrens */ 610669962b56SMatthew Ahrens if (zfs_dirty_data_max == 0) { 610769962b56SMatthew Ahrens zfs_dirty_data_max = physmem * PAGESIZE * 610869962b56SMatthew Ahrens zfs_dirty_data_max_percent / 100; 610969962b56SMatthew Ahrens zfs_dirty_data_max = MIN(zfs_dirty_data_max, 611069962b56SMatthew Ahrens zfs_dirty_data_max_max); 611169962b56SMatthew Ahrens } 6112fa9e4066Sahrens } 6113fa9e4066Sahrens 6114fa9e4066Sahrens void 6115fa9e4066Sahrens arc_fini(void) 6116fa9e4066Sahrens { 6117244781f1SPrakash Surya mutex_enter(&arc_reclaim_lock); 6118dcbf3bd6SGeorge Wilson arc_reclaim_thread_exit = B_TRUE; 6119244781f1SPrakash Surya /* 6120244781f1SPrakash Surya * The reclaim thread will set arc_reclaim_thread_exit back to 6121dcbf3bd6SGeorge Wilson * B_FALSE when it is finished exiting; we're waiting for that. 6122244781f1SPrakash Surya */ 6123244781f1SPrakash Surya while (arc_reclaim_thread_exit) { 6124244781f1SPrakash Surya cv_signal(&arc_reclaim_thread_cv); 6125244781f1SPrakash Surya cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 6126244781f1SPrakash Surya } 6127244781f1SPrakash Surya mutex_exit(&arc_reclaim_lock); 6128fa9e4066Sahrens 6129dcbf3bd6SGeorge Wilson /* Use B_TRUE to ensure *all* buffers are evicted */ 6130dcbf3bd6SGeorge Wilson arc_flush(NULL, B_TRUE); 6131fa9e4066Sahrens 6132dcbf3bd6SGeorge Wilson arc_dead = B_TRUE; 6133fa9e4066Sahrens 613444cb6abcSbmc if (arc_ksp != NULL) { 613544cb6abcSbmc kstat_delete(arc_ksp); 613644cb6abcSbmc arc_ksp = NULL; 613744cb6abcSbmc } 613844cb6abcSbmc 6139244781f1SPrakash Surya mutex_destroy(&arc_reclaim_lock); 6140244781f1SPrakash Surya cv_destroy(&arc_reclaim_thread_cv); 6141244781f1SPrakash Surya cv_destroy(&arc_reclaim_waiters_cv); 6142244781f1SPrakash Surya 6143dcbf3bd6SGeorge Wilson arc_state_fini(); 6144fa9e4066Sahrens buf_fini(); 61452fdbea25SAleksandr Guzovskiy 614689c86e32SChris Williamson ASSERT0(arc_loaned_bytes); 6147fa9e4066Sahrens } 6148fa94a07fSbrendan 6149fa94a07fSbrendan /* 6150fa94a07fSbrendan * Level 2 ARC 6151fa94a07fSbrendan * 6152fa94a07fSbrendan * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 6153fa94a07fSbrendan * It uses dedicated storage devices to hold cached data, which are populated 6154fa94a07fSbrendan * using large infrequent writes. The main role of this cache is to boost 6155fa94a07fSbrendan * the performance of random read workloads. The intended L2ARC devices 6156fa94a07fSbrendan * include short-stroked disks, solid state disks, and other media with 6157fa94a07fSbrendan * substantially faster read latency than disk. 6158fa94a07fSbrendan * 6159fa94a07fSbrendan * +-----------------------+ 6160fa94a07fSbrendan * | ARC | 6161fa94a07fSbrendan * +-----------------------+ 6162fa94a07fSbrendan * | ^ ^ 6163fa94a07fSbrendan * | | | 6164fa94a07fSbrendan * l2arc_feed_thread() arc_read() 6165fa94a07fSbrendan * | | | 6166fa94a07fSbrendan * | l2arc read | 6167fa94a07fSbrendan * V | | 6168fa94a07fSbrendan * +---------------+ | 6169fa94a07fSbrendan * | L2ARC | | 6170fa94a07fSbrendan * +---------------+ | 6171fa94a07fSbrendan * | ^ | 6172fa94a07fSbrendan * l2arc_write() | | 6173fa94a07fSbrendan * | | | 6174fa94a07fSbrendan * V | | 6175fa94a07fSbrendan * +-------+ +-------+ 6176fa94a07fSbrendan * | vdev | | vdev | 6177fa94a07fSbrendan * | cache | | cache | 6178fa94a07fSbrendan * +-------+ +-------+ 6179fa94a07fSbrendan * +=========+ .-----. 6180fa94a07fSbrendan * : L2ARC : |-_____-| 6181fa94a07fSbrendan * : devices : | Disks | 6182fa94a07fSbrendan * +=========+ `-_____-' 6183fa94a07fSbrendan * 6184fa94a07fSbrendan * Read requests are satisfied from the following sources, in order: 6185fa94a07fSbrendan * 6186fa94a07fSbrendan * 1) ARC 6187fa94a07fSbrendan * 2) vdev cache of L2ARC devices 6188fa94a07fSbrendan * 3) L2ARC devices 6189fa94a07fSbrendan * 4) vdev cache of disks 6190fa94a07fSbrendan * 5) disks 6191fa94a07fSbrendan * 6192fa94a07fSbrendan * Some L2ARC device types exhibit extremely slow write performance. 6193fa94a07fSbrendan * To accommodate for this there are some significant differences between 6194fa94a07fSbrendan * the L2ARC and traditional cache design: 6195fa94a07fSbrendan * 6196fa94a07fSbrendan * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 6197fa94a07fSbrendan * the ARC behave as usual, freeing buffers and placing headers on ghost 6198fa94a07fSbrendan * lists. The ARC does not send buffers to the L2ARC during eviction as 6199fa94a07fSbrendan * this would add inflated write latencies for all ARC memory pressure. 6200fa94a07fSbrendan * 6201fa94a07fSbrendan * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 6202fa94a07fSbrendan * It does this by periodically scanning buffers from the eviction-end of 6203fa94a07fSbrendan * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 6204aad02571SSaso Kiselkov * not already there. It scans until a headroom of buffers is satisfied, 6205aad02571SSaso Kiselkov * which itself is a buffer for ARC eviction. If a compressible buffer is 6206aad02571SSaso Kiselkov * found during scanning and selected for writing to an L2ARC device, we 6207aad02571SSaso Kiselkov * temporarily boost scanning headroom during the next scan cycle to make 6208aad02571SSaso Kiselkov * sure we adapt to compression effects (which might significantly reduce 6209aad02571SSaso Kiselkov * the data volume we write to L2ARC). The thread that does this is 6210fa94a07fSbrendan * l2arc_feed_thread(), illustrated below; example sizes are included to 6211fa94a07fSbrendan * provide a better sense of ratio than this diagram: 6212fa94a07fSbrendan * 6213fa94a07fSbrendan * head --> tail 6214fa94a07fSbrendan * +---------------------+----------+ 6215fa94a07fSbrendan * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 6216fa94a07fSbrendan * +---------------------+----------+ | o L2ARC eligible 6217fa94a07fSbrendan * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 6218fa94a07fSbrendan * +---------------------+----------+ | 6219fa94a07fSbrendan * 15.9 Gbytes ^ 32 Mbytes | 6220fa94a07fSbrendan * headroom | 6221fa94a07fSbrendan * l2arc_feed_thread() 6222fa94a07fSbrendan * | 6223fa94a07fSbrendan * l2arc write hand <--[oooo]--' 6224fa94a07fSbrendan * | 8 Mbyte 6225fa94a07fSbrendan * | write max 6226fa94a07fSbrendan * V 6227fa94a07fSbrendan * +==============================+ 6228fa94a07fSbrendan * L2ARC dev |####|#|###|###| |####| ... | 6229fa94a07fSbrendan * +==============================+ 6230fa94a07fSbrendan * 32 Gbytes 6231fa94a07fSbrendan * 6232fa94a07fSbrendan * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 6233fa94a07fSbrendan * evicted, then the L2ARC has cached a buffer much sooner than it probably 6234fa94a07fSbrendan * needed to, potentially wasting L2ARC device bandwidth and storage. It is 6235fa94a07fSbrendan * safe to say that this is an uncommon case, since buffers at the end of 6236fa94a07fSbrendan * the ARC lists have moved there due to inactivity. 6237fa94a07fSbrendan * 6238fa94a07fSbrendan * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 6239fa94a07fSbrendan * then the L2ARC simply misses copying some buffers. This serves as a 6240fa94a07fSbrendan * pressure valve to prevent heavy read workloads from both stalling the ARC 6241fa94a07fSbrendan * with waits and clogging the L2ARC with writes. This also helps prevent 6242fa94a07fSbrendan * the potential for the L2ARC to churn if it attempts to cache content too 6243fa94a07fSbrendan * quickly, such as during backups of the entire pool. 6244fa94a07fSbrendan * 62453a737e0dSbrendan * 5. After system boot and before the ARC has filled main memory, there are 62463a737e0dSbrendan * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 62473a737e0dSbrendan * lists can remain mostly static. Instead of searching from tail of these 62483a737e0dSbrendan * lists as pictured, the l2arc_feed_thread() will search from the list heads 62493a737e0dSbrendan * for eligible buffers, greatly increasing its chance of finding them. 62503a737e0dSbrendan * 62513a737e0dSbrendan * The L2ARC device write speed is also boosted during this time so that 62523a737e0dSbrendan * the L2ARC warms up faster. Since there have been no ARC evictions yet, 62533a737e0dSbrendan * there are no L2ARC reads, and no fear of degrading read performance 62543a737e0dSbrendan * through increased writes. 62553a737e0dSbrendan * 62563a737e0dSbrendan * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 6257fa94a07fSbrendan * the vdev queue can aggregate them into larger and fewer writes. Each 6258fa94a07fSbrendan * device is written to in a rotor fashion, sweeping writes through 6259fa94a07fSbrendan * available space then repeating. 6260fa94a07fSbrendan * 62613a737e0dSbrendan * 7. The L2ARC does not store dirty content. It never needs to flush 6262fa94a07fSbrendan * write buffers back to disk based storage. 6263fa94a07fSbrendan * 62643a737e0dSbrendan * 8. If an ARC buffer is written (and dirtied) which also exists in the 6265fa94a07fSbrendan * L2ARC, the now stale L2ARC buffer is immediately dropped. 6266fa94a07fSbrendan * 6267fa94a07fSbrendan * The performance of the L2ARC can be tweaked by a number of tunables, which 6268fa94a07fSbrendan * may be necessary for different workloads: 6269fa94a07fSbrendan * 6270fa94a07fSbrendan * l2arc_write_max max write bytes per interval 62713a737e0dSbrendan * l2arc_write_boost extra write bytes during device warmup 6272fa94a07fSbrendan * l2arc_noprefetch skip caching prefetched buffers 6273fa94a07fSbrendan * l2arc_headroom number of max device writes to precache 6274aad02571SSaso Kiselkov * l2arc_headroom_boost when we find compressed buffers during ARC 6275aad02571SSaso Kiselkov * scanning, we multiply headroom by this 6276aad02571SSaso Kiselkov * percentage factor for the next scan cycle, 6277aad02571SSaso Kiselkov * since more compressed buffers are likely to 6278aad02571SSaso Kiselkov * be present 6279fa94a07fSbrendan * l2arc_feed_secs seconds between L2ARC writing 6280fa94a07fSbrendan * 6281fa94a07fSbrendan * Tunables may be removed or added as future performance improvements are 6282fa94a07fSbrendan * integrated, and also may become zpool properties. 62835a98e54bSBrendan Gregg - Sun Microsystems * 62845a98e54bSBrendan Gregg - Sun Microsystems * There are three key functions that control how the L2ARC warms up: 62855a98e54bSBrendan Gregg - Sun Microsystems * 62865a98e54bSBrendan Gregg - Sun Microsystems * l2arc_write_eligible() check if a buffer is eligible to cache 62875a98e54bSBrendan Gregg - Sun Microsystems * l2arc_write_size() calculate how much to write 62885a98e54bSBrendan Gregg - Sun Microsystems * l2arc_write_interval() calculate sleep delay between writes 62895a98e54bSBrendan Gregg - Sun Microsystems * 62905a98e54bSBrendan Gregg - Sun Microsystems * These three functions determine what to write, how much, and how quickly 62915a98e54bSBrendan Gregg - Sun Microsystems * to send writes. 6292fa94a07fSbrendan */ 6293fa94a07fSbrendan 62945a98e54bSBrendan Gregg - Sun Microsystems static boolean_t 62957adb730bSGeorge Wilson l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 62965a98e54bSBrendan Gregg - Sun Microsystems { 62975a98e54bSBrendan Gregg - Sun Microsystems /* 62985a98e54bSBrendan Gregg - Sun Microsystems * A buffer is *not* eligible for the L2ARC if it: 62995a98e54bSBrendan Gregg - Sun Microsystems * 1. belongs to a different spa. 63005ea40c06SBrendan Gregg - Sun Microsystems * 2. is already cached on the L2ARC. 63015ea40c06SBrendan Gregg - Sun Microsystems * 3. has an I/O in progress (it may be an incomplete read). 63025ea40c06SBrendan Gregg - Sun Microsystems * 4. is flagged not eligible (zfs property). 63035a98e54bSBrendan Gregg - Sun Microsystems */ 630489c86e32SChris Williamson if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) || 63057adb730bSGeorge Wilson HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr)) 63065a98e54bSBrendan Gregg - Sun Microsystems return (B_FALSE); 63075a98e54bSBrendan Gregg - Sun Microsystems 63085a98e54bSBrendan Gregg - Sun Microsystems return (B_TRUE); 63095a98e54bSBrendan Gregg - Sun Microsystems } 63105a98e54bSBrendan Gregg - Sun Microsystems 63115a98e54bSBrendan Gregg - Sun Microsystems static uint64_t 6312aad02571SSaso Kiselkov l2arc_write_size(void) 63135a98e54bSBrendan Gregg - Sun Microsystems { 63145a98e54bSBrendan Gregg - Sun Microsystems uint64_t size; 63155a98e54bSBrendan Gregg - Sun Microsystems 6316aad02571SSaso Kiselkov /* 6317aad02571SSaso Kiselkov * Make sure our globals have meaningful values in case the user 6318aad02571SSaso Kiselkov * altered them. 6319aad02571SSaso Kiselkov */ 6320aad02571SSaso Kiselkov size = l2arc_write_max; 6321aad02571SSaso Kiselkov if (size == 0) { 6322aad02571SSaso Kiselkov cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 6323aad02571SSaso Kiselkov "be greater than zero, resetting it to the default (%d)", 6324aad02571SSaso Kiselkov L2ARC_WRITE_SIZE); 6325aad02571SSaso Kiselkov size = l2arc_write_max = L2ARC_WRITE_SIZE; 6326aad02571SSaso Kiselkov } 63275a98e54bSBrendan Gregg - Sun Microsystems 63285a98e54bSBrendan Gregg - Sun Microsystems if (arc_warm == B_FALSE) 6329aad02571SSaso Kiselkov size += l2arc_write_boost; 63305a98e54bSBrendan Gregg - Sun Microsystems 63315a98e54bSBrendan Gregg - Sun Microsystems return (size); 63325a98e54bSBrendan Gregg - Sun Microsystems 63335a98e54bSBrendan Gregg - Sun Microsystems } 63345a98e54bSBrendan Gregg - Sun Microsystems 63355a98e54bSBrendan Gregg - Sun Microsystems static clock_t 63365a98e54bSBrendan Gregg - Sun Microsystems l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 63375a98e54bSBrendan Gregg - Sun Microsystems { 6338d3d50737SRafael Vanoni clock_t interval, next, now; 63395a98e54bSBrendan Gregg - Sun Microsystems 63405a98e54bSBrendan Gregg - Sun Microsystems /* 63415a98e54bSBrendan Gregg - Sun Microsystems * If the ARC lists are busy, increase our write rate; if the 63425a98e54bSBrendan Gregg - Sun Microsystems * lists are stale, idle back. This is achieved by checking 63435a98e54bSBrendan Gregg - Sun Microsystems * how much we previously wrote - if it was more than half of 63445a98e54bSBrendan Gregg - Sun Microsystems * what we wanted, schedule the next write much sooner. 63455a98e54bSBrendan Gregg - Sun Microsystems */ 63465a98e54bSBrendan Gregg - Sun Microsystems if (l2arc_feed_again && wrote > (wanted / 2)) 63475a98e54bSBrendan Gregg - Sun Microsystems interval = (hz * l2arc_feed_min_ms) / 1000; 63485a98e54bSBrendan Gregg - Sun Microsystems else 63495a98e54bSBrendan Gregg - Sun Microsystems interval = hz * l2arc_feed_secs; 63505a98e54bSBrendan Gregg - Sun Microsystems 6351d3d50737SRafael Vanoni now = ddi_get_lbolt(); 6352d3d50737SRafael Vanoni next = MAX(now, MIN(now + interval, began + interval)); 63535a98e54bSBrendan Gregg - Sun Microsystems 63545a98e54bSBrendan Gregg - Sun Microsystems return (next); 63555a98e54bSBrendan Gregg - Sun Microsystems } 63565a98e54bSBrendan Gregg - Sun Microsystems 6357fa94a07fSbrendan /* 6358fa94a07fSbrendan * Cycle through L2ARC devices. This is how L2ARC load balances. 63593a737e0dSbrendan * If a device is returned, this also returns holding the spa config lock. 6360fa94a07fSbrendan */ 6361fa94a07fSbrendan static l2arc_dev_t * 6362fa94a07fSbrendan l2arc_dev_get_next(void) 6363fa94a07fSbrendan { 63643a737e0dSbrendan l2arc_dev_t *first, *next = NULL; 63653a737e0dSbrendan 63663a737e0dSbrendan /* 63673a737e0dSbrendan * Lock out the removal of spas (spa_namespace_lock), then removal 63683a737e0dSbrendan * of cache devices (l2arc_dev_mtx). Once a device has been selected, 63693a737e0dSbrendan * both locks will be dropped and a spa config lock held instead. 63703a737e0dSbrendan */ 63713a737e0dSbrendan mutex_enter(&spa_namespace_lock); 63723a737e0dSbrendan mutex_enter(&l2arc_dev_mtx); 6373fa94a07fSbrendan 6374c5904d13Seschrock /* if there are no vdevs, there is nothing to do */ 6375c5904d13Seschrock if (l2arc_ndev == 0) 63763a737e0dSbrendan goto out; 6377c5904d13Seschrock 6378c5904d13Seschrock first = NULL; 6379c5904d13Seschrock next = l2arc_dev_last; 6380c5904d13Seschrock do { 6381c5904d13Seschrock /* loop around the list looking for a non-faulted vdev */ 6382c5904d13Seschrock if (next == NULL) { 6383fa94a07fSbrendan next = list_head(l2arc_dev_list); 6384c5904d13Seschrock } else { 6385c5904d13Seschrock next = list_next(l2arc_dev_list, next); 6386c5904d13Seschrock if (next == NULL) 6387c5904d13Seschrock next = list_head(l2arc_dev_list); 6388c5904d13Seschrock } 6389c5904d13Seschrock 6390c5904d13Seschrock /* if we have come back to the start, bail out */ 6391c5904d13Seschrock if (first == NULL) 6392c5904d13Seschrock first = next; 6393c5904d13Seschrock else if (next == first) 6394c5904d13Seschrock break; 6395c5904d13Seschrock 6396c5904d13Seschrock } while (vdev_is_dead(next->l2ad_vdev)); 6397c5904d13Seschrock 6398c5904d13Seschrock /* if we were unable to find any usable vdevs, return NULL */ 6399c5904d13Seschrock if (vdev_is_dead(next->l2ad_vdev)) 64003a737e0dSbrendan next = NULL; 6401fa94a07fSbrendan 6402fa94a07fSbrendan l2arc_dev_last = next; 6403fa94a07fSbrendan 64043a737e0dSbrendan out: 64053a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 64063a737e0dSbrendan 64073a737e0dSbrendan /* 64083a737e0dSbrendan * Grab the config lock to prevent the 'next' device from being 64093a737e0dSbrendan * removed while we are writing to it. 64103a737e0dSbrendan */ 64113a737e0dSbrendan if (next != NULL) 6412e14bb325SJeff Bonwick spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 64133a737e0dSbrendan mutex_exit(&spa_namespace_lock); 64143a737e0dSbrendan 6415fa94a07fSbrendan return (next); 6416fa94a07fSbrendan } 6417fa94a07fSbrendan 64183a737e0dSbrendan /* 64193a737e0dSbrendan * Free buffers that were tagged for destruction. 64203a737e0dSbrendan */ 64213a737e0dSbrendan static void 64223a737e0dSbrendan l2arc_do_free_on_write() 64233a737e0dSbrendan { 64243a737e0dSbrendan list_t *buflist; 64253a737e0dSbrendan l2arc_data_free_t *df, *df_prev; 64263a737e0dSbrendan 64273a737e0dSbrendan mutex_enter(&l2arc_free_on_write_mtx); 64283a737e0dSbrendan buflist = l2arc_free_on_write; 64293a737e0dSbrendan 64303a737e0dSbrendan for (df = list_tail(buflist); df; df = df_prev) { 64313a737e0dSbrendan df_prev = list_prev(buflist, df); 6432770499e1SDan Kimmel ASSERT3P(df->l2df_abd, !=, NULL); 6433770499e1SDan Kimmel abd_free(df->l2df_abd); 64343a737e0dSbrendan list_remove(buflist, df); 64353a737e0dSbrendan kmem_free(df, sizeof (l2arc_data_free_t)); 64363a737e0dSbrendan } 64373a737e0dSbrendan 64383a737e0dSbrendan mutex_exit(&l2arc_free_on_write_mtx); 64393a737e0dSbrendan } 64403a737e0dSbrendan 6441fa94a07fSbrendan /* 6442fa94a07fSbrendan * A write to a cache device has completed. Update all headers to allow 6443fa94a07fSbrendan * reads from these buffers to begin. 6444fa94a07fSbrendan */ 6445fa94a07fSbrendan static void 6446fa94a07fSbrendan l2arc_write_done(zio_t *zio) 6447fa94a07fSbrendan { 6448fa94a07fSbrendan l2arc_write_callback_t *cb; 6449fa94a07fSbrendan l2arc_dev_t *dev; 6450fa94a07fSbrendan list_t *buflist; 64517adb730bSGeorge Wilson arc_buf_hdr_t *head, *hdr, *hdr_prev; 6452fa94a07fSbrendan kmutex_t *hash_lock; 64533038a2b4SSaso Kiselkov int64_t bytes_dropped = 0; 6454fa94a07fSbrendan 6455fa94a07fSbrendan cb = zio->io_private; 6456dcbf3bd6SGeorge Wilson ASSERT3P(cb, !=, NULL); 6457fa94a07fSbrendan dev = cb->l2wcb_dev; 6458dcbf3bd6SGeorge Wilson ASSERT3P(dev, !=, NULL); 6459fa94a07fSbrendan head = cb->l2wcb_head; 6460dcbf3bd6SGeorge Wilson ASSERT3P(head, !=, NULL); 646189c86e32SChris Williamson buflist = &dev->l2ad_buflist; 6462dcbf3bd6SGeorge Wilson ASSERT3P(buflist, !=, NULL); 6463fa94a07fSbrendan DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 6464fa94a07fSbrendan l2arc_write_callback_t *, cb); 6465fa94a07fSbrendan 6466fa94a07fSbrendan if (zio->io_error != 0) 6467fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_writes_error); 6468fa94a07fSbrendan 6469fa94a07fSbrendan /* 6470fa94a07fSbrendan * All writes completed, or an error was hit. 6471fa94a07fSbrendan */ 6472244781f1SPrakash Surya top: 6473244781f1SPrakash Surya mutex_enter(&dev->l2ad_mtx); 64747adb730bSGeorge Wilson for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 64757adb730bSGeorge Wilson hdr_prev = list_prev(buflist, hdr); 6476fa94a07fSbrendan 64777adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 6478244781f1SPrakash Surya 6479244781f1SPrakash Surya /* 6480244781f1SPrakash Surya * We cannot use mutex_enter or else we can deadlock 6481244781f1SPrakash Surya * with l2arc_write_buffers (due to swapping the order 6482244781f1SPrakash Surya * the hash lock and l2ad_mtx are taken). 6483244781f1SPrakash Surya */ 6484fa94a07fSbrendan if (!mutex_tryenter(hash_lock)) { 6485fa94a07fSbrendan /* 6486244781f1SPrakash Surya * Missed the hash lock. We must retry so we 6487244781f1SPrakash Surya * don't leave the ARC_FLAG_L2_WRITING bit set. 6488fa94a07fSbrendan */ 6489244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 6490244781f1SPrakash Surya 6491244781f1SPrakash Surya /* 6492244781f1SPrakash Surya * We don't want to rescan the headers we've 6493244781f1SPrakash Surya * already marked as having been written out, so 6494244781f1SPrakash Surya * we reinsert the head node so we can pick up 6495244781f1SPrakash Surya * where we left off. 6496244781f1SPrakash Surya */ 6497244781f1SPrakash Surya list_remove(buflist, head); 6498244781f1SPrakash Surya list_insert_after(buflist, hdr, head); 6499244781f1SPrakash Surya 6500244781f1SPrakash Surya mutex_exit(&dev->l2ad_mtx); 6501244781f1SPrakash Surya 6502244781f1SPrakash Surya /* 6503244781f1SPrakash Surya * We wait for the hash lock to become available 6504244781f1SPrakash Surya * to try and prevent busy waiting, and increase 6505244781f1SPrakash Surya * the chance we'll be able to acquire the lock 6506244781f1SPrakash Surya * the next time around. 6507244781f1SPrakash Surya */ 6508244781f1SPrakash Surya mutex_enter(hash_lock); 6509244781f1SPrakash Surya mutex_exit(hash_lock); 6510244781f1SPrakash Surya goto top; 6511fa94a07fSbrendan } 6512fa94a07fSbrendan 651389c86e32SChris Williamson /* 6514244781f1SPrakash Surya * We could not have been moved into the arc_l2c_only 6515244781f1SPrakash Surya * state while in-flight due to our ARC_FLAG_L2_WRITING 6516244781f1SPrakash Surya * bit being set. Let's just ensure that's being enforced. 6517244781f1SPrakash Surya */ 6518244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 6519244781f1SPrakash Surya 6520fa94a07fSbrendan if (zio->io_error != 0) { 6521fa94a07fSbrendan /* 65223a737e0dSbrendan * Error - drop L2ARC entry. 6523fa94a07fSbrendan */ 65247adb730bSGeorge Wilson list_remove(buflist, hdr); 6525dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 652689c86e32SChris Williamson 652716a7e5acSAndriy Gapon ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr)); 652816a7e5acSAndriy Gapon ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); 6529a52fc310SPrakash Surya 6530dcbf3bd6SGeorge Wilson bytes_dropped += arc_hdr_size(hdr); 6531a52fc310SPrakash Surya (void) refcount_remove_many(&dev->l2ad_alloc, 6532dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 6533fa94a07fSbrendan } 6534fa94a07fSbrendan 6535fa94a07fSbrendan /* 6536244781f1SPrakash Surya * Allow ARC to begin reads and ghost list evictions to 6537244781f1SPrakash Surya * this L2ARC entry. 6538fa94a07fSbrendan */ 6539dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); 6540fa94a07fSbrendan 6541fa94a07fSbrendan mutex_exit(hash_lock); 6542fa94a07fSbrendan } 6543fa94a07fSbrendan 6544fa94a07fSbrendan atomic_inc_64(&l2arc_writes_done); 6545fa94a07fSbrendan list_remove(buflist, head); 654689c86e32SChris Williamson ASSERT(!HDR_HAS_L1HDR(head)); 654789c86e32SChris Williamson kmem_cache_free(hdr_l2only_cache, head); 654889c86e32SChris Williamson mutex_exit(&dev->l2ad_mtx); 6549fa94a07fSbrendan 65503038a2b4SSaso Kiselkov vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 65513038a2b4SSaso Kiselkov 65523a737e0dSbrendan l2arc_do_free_on_write(); 6553fa94a07fSbrendan 6554fa94a07fSbrendan kmem_free(cb, sizeof (l2arc_write_callback_t)); 6555fa94a07fSbrendan } 6556fa94a07fSbrendan 6557fa94a07fSbrendan /* 6558fa94a07fSbrendan * A read to a cache device completed. Validate buffer contents before 6559fa94a07fSbrendan * handing over to the regular ARC routines. 6560fa94a07fSbrendan */ 6561fa94a07fSbrendan static void 6562fa94a07fSbrendan l2arc_read_done(zio_t *zio) 6563fa94a07fSbrendan { 6564fa94a07fSbrendan l2arc_read_callback_t *cb; 6565fa94a07fSbrendan arc_buf_hdr_t *hdr; 6566fa94a07fSbrendan kmutex_t *hash_lock; 6567dcbf3bd6SGeorge Wilson boolean_t valid_cksum; 6568fa94a07fSbrendan 6569dcbf3bd6SGeorge Wilson ASSERT3P(zio->io_vd, !=, NULL); 6570e14bb325SJeff Bonwick ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 6571e14bb325SJeff Bonwick 6572e14bb325SJeff Bonwick spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 6573e14bb325SJeff Bonwick 6574fa94a07fSbrendan cb = zio->io_private; 6575dcbf3bd6SGeorge Wilson ASSERT3P(cb, !=, NULL); 6576dcbf3bd6SGeorge Wilson hdr = cb->l2rcb_hdr; 6577dcbf3bd6SGeorge Wilson ASSERT3P(hdr, !=, NULL); 6578fa94a07fSbrendan 6579dcbf3bd6SGeorge Wilson hash_lock = HDR_LOCK(hdr); 6580fa94a07fSbrendan mutex_enter(hash_lock); 65813f9d6ad7SLin Ling ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 6582fa94a07fSbrendan 6583403a8da7SAndriy Gapon /* 6584403a8da7SAndriy Gapon * If the data was read into a temporary buffer, 6585403a8da7SAndriy Gapon * move it and free the buffer. 6586403a8da7SAndriy Gapon */ 6587403a8da7SAndriy Gapon if (cb->l2rcb_abd != NULL) { 6588403a8da7SAndriy Gapon ASSERT3U(arc_hdr_size(hdr), <, zio->io_size); 6589403a8da7SAndriy Gapon if (zio->io_error == 0) { 6590403a8da7SAndriy Gapon abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd, 6591403a8da7SAndriy Gapon arc_hdr_size(hdr)); 6592403a8da7SAndriy Gapon } 6593403a8da7SAndriy Gapon 6594403a8da7SAndriy Gapon /* 6595403a8da7SAndriy Gapon * The following must be done regardless of whether 6596403a8da7SAndriy Gapon * there was an error: 6597403a8da7SAndriy Gapon * - free the temporary buffer 6598403a8da7SAndriy Gapon * - point zio to the real ARC buffer 6599403a8da7SAndriy Gapon * - set zio size accordingly 6600403a8da7SAndriy Gapon * These are required because zio is either re-used for 6601403a8da7SAndriy Gapon * an I/O of the block in the case of the error 6602403a8da7SAndriy Gapon * or the zio is passed to arc_read_done() and it 6603403a8da7SAndriy Gapon * needs real data. 6604403a8da7SAndriy Gapon */ 6605403a8da7SAndriy Gapon abd_free(cb->l2rcb_abd); 6606403a8da7SAndriy Gapon zio->io_size = zio->io_orig_size = arc_hdr_size(hdr); 6607403a8da7SAndriy Gapon zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd; 6608403a8da7SAndriy Gapon } 6609403a8da7SAndriy Gapon 6610770499e1SDan Kimmel ASSERT3P(zio->io_abd, !=, NULL); 6611aad02571SSaso Kiselkov 6612fa94a07fSbrendan /* 6613fa94a07fSbrendan * Check this survived the L2ARC journey. 6614fa94a07fSbrendan */ 6615770499e1SDan Kimmel ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd); 6616dcbf3bd6SGeorge Wilson zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 6617dcbf3bd6SGeorge Wilson zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 6618dcbf3bd6SGeorge Wilson 6619dcbf3bd6SGeorge Wilson valid_cksum = arc_cksum_is_equal(hdr, zio); 6620dcbf3bd6SGeorge Wilson if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 6621fa94a07fSbrendan mutex_exit(hash_lock); 6622dcbf3bd6SGeorge Wilson zio->io_private = hdr; 6623fa94a07fSbrendan arc_read_done(zio); 6624fa94a07fSbrendan } else { 6625fa94a07fSbrendan mutex_exit(hash_lock); 6626fa94a07fSbrendan /* 6627fa94a07fSbrendan * Buffer didn't survive caching. Increment stats and 6628fa94a07fSbrendan * reissue to the original storage device. 6629fa94a07fSbrendan */ 66303a737e0dSbrendan if (zio->io_error != 0) { 6631fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_io_error); 66323a737e0dSbrendan } else { 6633be6fd75aSMatthew Ahrens zio->io_error = SET_ERROR(EIO); 66343a737e0dSbrendan } 6635dcbf3bd6SGeorge Wilson if (!valid_cksum) 6636fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_cksum_bad); 6637fa94a07fSbrendan 6638e14bb325SJeff Bonwick /* 6639e14bb325SJeff Bonwick * If there's no waiter, issue an async i/o to the primary 6640e14bb325SJeff Bonwick * storage now. If there *is* a waiter, the caller must 6641e14bb325SJeff Bonwick * issue the i/o in a context where it's OK to block. 6642e14bb325SJeff Bonwick */ 6643a3f829aeSBill Moore if (zio->io_waiter == NULL) { 6644a3f829aeSBill Moore zio_t *pio = zio_unique_parent(zio); 6645a3f829aeSBill Moore 6646a3f829aeSBill Moore ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6647a3f829aeSBill Moore 6648dcbf3bd6SGeorge Wilson zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, 6649770499e1SDan Kimmel hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done, 6650dcbf3bd6SGeorge Wilson hdr, zio->io_priority, cb->l2rcb_flags, 6651dcbf3bd6SGeorge Wilson &cb->l2rcb_zb)); 6652a3f829aeSBill Moore } 6653fa94a07fSbrendan } 6654fa94a07fSbrendan 6655fa94a07fSbrendan kmem_free(cb, sizeof (l2arc_read_callback_t)); 6656fa94a07fSbrendan } 6657fa94a07fSbrendan 6658fa94a07fSbrendan /* 6659fa94a07fSbrendan * This is the list priority from which the L2ARC will search for pages to 6660fa94a07fSbrendan * cache. This is used within loops (0..3) to cycle through lists in the 6661fa94a07fSbrendan * desired order. This order can have a significant effect on cache 6662fa94a07fSbrendan * performance. 6663fa94a07fSbrendan * 6664fa94a07fSbrendan * Currently the metadata lists are hit first, MFU then MRU, followed by 6665fa94a07fSbrendan * the data lists. This function returns a locked list, and also returns 6666fa94a07fSbrendan * the lock pointer. 6667fa94a07fSbrendan */ 6668244781f1SPrakash Surya static multilist_sublist_t * 6669244781f1SPrakash Surya l2arc_sublist_lock(int list_num) 6670fa94a07fSbrendan { 6671244781f1SPrakash Surya multilist_t *ml = NULL; 6672244781f1SPrakash Surya unsigned int idx; 6673fa94a07fSbrendan 6674fa94a07fSbrendan ASSERT(list_num >= 0 && list_num <= 3); 6675fa94a07fSbrendan 6676fa94a07fSbrendan switch (list_num) { 6677fa94a07fSbrendan case 0: 667894c2d0ebSMatthew Ahrens ml = arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6679fa94a07fSbrendan break; 6680fa94a07fSbrendan case 1: 668194c2d0ebSMatthew Ahrens ml = arc_mru->arcs_list[ARC_BUFC_METADATA]; 6682fa94a07fSbrendan break; 6683fa94a07fSbrendan case 2: 668494c2d0ebSMatthew Ahrens ml = arc_mfu->arcs_list[ARC_BUFC_DATA]; 6685fa94a07fSbrendan break; 6686fa94a07fSbrendan case 3: 668794c2d0ebSMatthew Ahrens ml = arc_mru->arcs_list[ARC_BUFC_DATA]; 6688fa94a07fSbrendan break; 6689fa94a07fSbrendan } 6690fa94a07fSbrendan 6691244781f1SPrakash Surya /* 6692244781f1SPrakash Surya * Return a randomly-selected sublist. This is acceptable 6693244781f1SPrakash Surya * because the caller feeds only a little bit of data for each 6694244781f1SPrakash Surya * call (8MB). Subsequent calls will result in different 6695244781f1SPrakash Surya * sublists being selected. 6696244781f1SPrakash Surya */ 6697244781f1SPrakash Surya idx = multilist_get_random_index(ml); 6698244781f1SPrakash Surya return (multilist_sublist_lock(ml, idx)); 6699fa94a07fSbrendan } 6700fa94a07fSbrendan 6701fa94a07fSbrendan /* 6702fa94a07fSbrendan * Evict buffers from the device write hand to the distance specified in 6703fa94a07fSbrendan * bytes. This distance may span populated buffers, it may span nothing. 6704fa94a07fSbrendan * This is clearing a region on the L2ARC device ready for writing. 6705fa94a07fSbrendan * If the 'all' boolean is set, every buffer is evicted. 6706fa94a07fSbrendan */ 6707fa94a07fSbrendan static void 6708fa94a07fSbrendan l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6709fa94a07fSbrendan { 6710fa94a07fSbrendan list_t *buflist; 67117adb730bSGeorge Wilson arc_buf_hdr_t *hdr, *hdr_prev; 6712fa94a07fSbrendan kmutex_t *hash_lock; 6713fa94a07fSbrendan uint64_t taddr; 6714fa94a07fSbrendan 671589c86e32SChris Williamson buflist = &dev->l2ad_buflist; 6716fa94a07fSbrendan 6717fa94a07fSbrendan if (!all && dev->l2ad_first) { 6718fa94a07fSbrendan /* 6719fa94a07fSbrendan * This is the first sweep through the device. There is 6720fa94a07fSbrendan * nothing to evict. 6721fa94a07fSbrendan */ 6722fa94a07fSbrendan return; 6723fa94a07fSbrendan } 6724fa94a07fSbrendan 67253a737e0dSbrendan if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6726fa94a07fSbrendan /* 6727fa94a07fSbrendan * When nearing the end of the device, evict to the end 6728fa94a07fSbrendan * before the device write hand jumps to the start. 6729fa94a07fSbrendan */ 6730fa94a07fSbrendan taddr = dev->l2ad_end; 6731fa94a07fSbrendan } else { 6732fa94a07fSbrendan taddr = dev->l2ad_hand + distance; 6733fa94a07fSbrendan } 6734fa94a07fSbrendan DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6735fa94a07fSbrendan uint64_t, taddr, boolean_t, all); 6736fa94a07fSbrendan 6737fa94a07fSbrendan top: 673889c86e32SChris Williamson mutex_enter(&dev->l2ad_mtx); 67397adb730bSGeorge Wilson for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 67407adb730bSGeorge Wilson hdr_prev = list_prev(buflist, hdr); 6741fa94a07fSbrendan 67427adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 6743244781f1SPrakash Surya 6744244781f1SPrakash Surya /* 6745244781f1SPrakash Surya * We cannot use mutex_enter or else we can deadlock 6746244781f1SPrakash Surya * with l2arc_write_buffers (due to swapping the order 6747244781f1SPrakash Surya * the hash lock and l2ad_mtx are taken). 6748244781f1SPrakash Surya */ 6749fa94a07fSbrendan if (!mutex_tryenter(hash_lock)) { 6750fa94a07fSbrendan /* 6751fa94a07fSbrendan * Missed the hash lock. Retry. 6752fa94a07fSbrendan */ 6753fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 675489c86e32SChris Williamson mutex_exit(&dev->l2ad_mtx); 6755fa94a07fSbrendan mutex_enter(hash_lock); 6756fa94a07fSbrendan mutex_exit(hash_lock); 6757fa94a07fSbrendan goto top; 6758fa94a07fSbrendan } 6759fa94a07fSbrendan 6760267ae6c3SAndriy Gapon /* 6761267ae6c3SAndriy Gapon * A header can't be on this list if it doesn't have L2 header. 6762267ae6c3SAndriy Gapon */ 6763267ae6c3SAndriy Gapon ASSERT(HDR_HAS_L2HDR(hdr)); 6764fa94a07fSbrendan 6765267ae6c3SAndriy Gapon /* Ensure this header has finished being written. */ 6766267ae6c3SAndriy Gapon ASSERT(!HDR_L2_WRITING(hdr)); 6767267ae6c3SAndriy Gapon ASSERT(!HDR_L2_WRITE_HEAD(hdr)); 6768267ae6c3SAndriy Gapon 6769267ae6c3SAndriy Gapon if (!all && (hdr->b_l2hdr.b_daddr >= taddr || 677089c86e32SChris Williamson hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6771fa94a07fSbrendan /* 6772fa94a07fSbrendan * We've evicted to the target address, 6773fa94a07fSbrendan * or the end of the device. 6774fa94a07fSbrendan */ 6775fa94a07fSbrendan mutex_exit(hash_lock); 6776fa94a07fSbrendan break; 6777fa94a07fSbrendan } 6778fa94a07fSbrendan 677989c86e32SChris Williamson if (!HDR_HAS_L1HDR(hdr)) { 67807adb730bSGeorge Wilson ASSERT(!HDR_L2_READING(hdr)); 6781fa94a07fSbrendan /* 6782fa94a07fSbrendan * This doesn't exist in the ARC. Destroy. 6783fa94a07fSbrendan * arc_hdr_destroy() will call list_remove() 678416a7e5acSAndriy Gapon * and decrement arcstat_l2_lsize. 6785fa94a07fSbrendan */ 67867adb730bSGeorge Wilson arc_change_state(arc_anon, hdr, hash_lock); 67877adb730bSGeorge Wilson arc_hdr_destroy(hdr); 6788fa94a07fSbrendan } else { 678989c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 679089c86e32SChris Williamson ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 67913a737e0dSbrendan /* 67923a737e0dSbrendan * Invalidate issued or about to be issued 67933a737e0dSbrendan * reads, since we may be about to write 67943a737e0dSbrendan * over this location. 67953a737e0dSbrendan */ 67967adb730bSGeorge Wilson if (HDR_L2_READING(hdr)) { 67973a737e0dSbrendan ARCSTAT_BUMP(arcstat_l2_evict_reading); 6798dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); 67993a737e0dSbrendan } 68003a737e0dSbrendan 6801a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(hdr); 6802fa94a07fSbrendan } 6803fa94a07fSbrendan mutex_exit(hash_lock); 6804fa94a07fSbrendan } 680589c86e32SChris Williamson mutex_exit(&dev->l2ad_mtx); 6806fa94a07fSbrendan } 6807fa94a07fSbrendan 6808fa94a07fSbrendan /* 6809fa94a07fSbrendan * Find and write ARC buffers to the L2ARC device. 6810fa94a07fSbrendan * 68117adb730bSGeorge Wilson * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6812fa94a07fSbrendan * for reading until they have completed writing. 6813aad02571SSaso Kiselkov * The headroom_boost is an in-out parameter used to maintain headroom boost 6814aad02571SSaso Kiselkov * state between calls to this function. 6815aad02571SSaso Kiselkov * 6816aad02571SSaso Kiselkov * Returns the number of bytes actually written (which may be smaller than 6817aad02571SSaso Kiselkov * the delta by which the device hand has changed due to alignment). 6818fa94a07fSbrendan */ 68195a98e54bSBrendan Gregg - Sun Microsystems static uint64_t 6820dcbf3bd6SGeorge Wilson l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 6821fa94a07fSbrendan { 68227adb730bSGeorge Wilson arc_buf_hdr_t *hdr, *hdr_prev, *head; 682316a7e5acSAndriy Gapon uint64_t write_asize, write_psize, write_lsize, headroom; 6824aad02571SSaso Kiselkov boolean_t full; 6825fa94a07fSbrendan l2arc_write_callback_t *cb; 6826fa94a07fSbrendan zio_t *pio, *wzio; 6827e9103aaeSGarrett D'Amore uint64_t guid = spa_load_guid(spa); 6828fa94a07fSbrendan 6829dcbf3bd6SGeorge Wilson ASSERT3P(dev->l2ad_vdev, !=, NULL); 6830aad02571SSaso Kiselkov 6831fa94a07fSbrendan pio = NULL; 683216a7e5acSAndriy Gapon write_lsize = write_asize = write_psize = 0; 6833fa94a07fSbrendan full = B_FALSE; 683489c86e32SChris Williamson head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 6835dcbf3bd6SGeorge Wilson arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); 6836aad02571SSaso Kiselkov 6837fa94a07fSbrendan /* 6838fa94a07fSbrendan * Copy buffers for L2ARC writing. 6839fa94a07fSbrendan */ 6840fa94a07fSbrendan for (int try = 0; try <= 3; try++) { 6841244781f1SPrakash Surya multilist_sublist_t *mls = l2arc_sublist_lock(try); 6842aad02571SSaso Kiselkov uint64_t passed_sz = 0; 6843aad02571SSaso Kiselkov 68443a737e0dSbrendan /* 68453a737e0dSbrendan * L2ARC fast warmup. 68463a737e0dSbrendan * 68473a737e0dSbrendan * Until the ARC is warm and starts to evict, read from the 68483a737e0dSbrendan * head of the ARC lists rather than the tail. 68493a737e0dSbrendan */ 68503a737e0dSbrendan if (arc_warm == B_FALSE) 6851244781f1SPrakash Surya hdr = multilist_sublist_head(mls); 68523a737e0dSbrendan else 6853244781f1SPrakash Surya hdr = multilist_sublist_tail(mls); 68543a737e0dSbrendan 6855aad02571SSaso Kiselkov headroom = target_sz * l2arc_headroom; 6856dcbf3bd6SGeorge Wilson if (zfs_compressed_arc_enabled) 6857aad02571SSaso Kiselkov headroom = (headroom * l2arc_headroom_boost) / 100; 6858aad02571SSaso Kiselkov 68597adb730bSGeorge Wilson for (; hdr; hdr = hdr_prev) { 6860aad02571SSaso Kiselkov kmutex_t *hash_lock; 6861aad02571SSaso Kiselkov 68623a737e0dSbrendan if (arc_warm == B_FALSE) 6863244781f1SPrakash Surya hdr_prev = multilist_sublist_next(mls, hdr); 68643a737e0dSbrendan else 6865244781f1SPrakash Surya hdr_prev = multilist_sublist_prev(mls, hdr); 6866fa94a07fSbrendan 68677adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 6868aad02571SSaso Kiselkov if (!mutex_tryenter(hash_lock)) { 6869fa94a07fSbrendan /* 6870fa94a07fSbrendan * Skip this buffer rather than waiting. 6871fa94a07fSbrendan */ 6872fa94a07fSbrendan continue; 6873fa94a07fSbrendan } 6874fa94a07fSbrendan 6875dcbf3bd6SGeorge Wilson passed_sz += HDR_GET_LSIZE(hdr); 6876fa94a07fSbrendan if (passed_sz > headroom) { 6877fa94a07fSbrendan /* 6878fa94a07fSbrendan * Searched too far. 6879fa94a07fSbrendan */ 6880fa94a07fSbrendan mutex_exit(hash_lock); 6881fa94a07fSbrendan break; 6882fa94a07fSbrendan } 6883fa94a07fSbrendan 68847adb730bSGeorge Wilson if (!l2arc_write_eligible(guid, hdr)) { 6885fa94a07fSbrendan mutex_exit(hash_lock); 6886fa94a07fSbrendan continue; 6887fa94a07fSbrendan } 6888fa94a07fSbrendan 688916a7e5acSAndriy Gapon /* 689016a7e5acSAndriy Gapon * We rely on the L1 portion of the header below, so 689116a7e5acSAndriy Gapon * it's invalid for this header to have been evicted out 689216a7e5acSAndriy Gapon * of the ghost cache, prior to being written out. The 689316a7e5acSAndriy Gapon * ARC_FLAG_L2_WRITING bit ensures this won't happen. 689416a7e5acSAndriy Gapon */ 689516a7e5acSAndriy Gapon ASSERT(HDR_HAS_L1HDR(hdr)); 689616a7e5acSAndriy Gapon 689716a7e5acSAndriy Gapon ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); 689816a7e5acSAndriy Gapon ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 689916a7e5acSAndriy Gapon ASSERT3U(arc_hdr_size(hdr), >, 0); 690016a7e5acSAndriy Gapon uint64_t psize = arc_hdr_size(hdr); 690116a7e5acSAndriy Gapon uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, 690216a7e5acSAndriy Gapon psize); 690316a7e5acSAndriy Gapon 690416a7e5acSAndriy Gapon if ((write_asize + asize) > target_sz) { 6905fa94a07fSbrendan full = B_TRUE; 6906fa94a07fSbrendan mutex_exit(hash_lock); 6907fa94a07fSbrendan break; 6908fa94a07fSbrendan } 6909fa94a07fSbrendan 6910fa94a07fSbrendan if (pio == NULL) { 6911fa94a07fSbrendan /* 6912fa94a07fSbrendan * Insert a dummy header on the buflist so 6913fa94a07fSbrendan * l2arc_write_done() can find where the 6914fa94a07fSbrendan * write buffers begin without searching. 6915fa94a07fSbrendan */ 6916244781f1SPrakash Surya mutex_enter(&dev->l2ad_mtx); 691789c86e32SChris Williamson list_insert_head(&dev->l2ad_buflist, head); 6918244781f1SPrakash Surya mutex_exit(&dev->l2ad_mtx); 6919fa94a07fSbrendan 6920fa94a07fSbrendan cb = kmem_alloc( 6921fa94a07fSbrendan sizeof (l2arc_write_callback_t), KM_SLEEP); 6922fa94a07fSbrendan cb->l2wcb_dev = dev; 6923fa94a07fSbrendan cb->l2wcb_head = head; 6924fa94a07fSbrendan pio = zio_root(spa, l2arc_write_done, cb, 6925fa94a07fSbrendan ZIO_FLAG_CANFAIL); 6926fa94a07fSbrendan } 6927fa94a07fSbrendan 692889c86e32SChris Williamson hdr->b_l2hdr.b_dev = dev; 6929dcbf3bd6SGeorge Wilson hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 6930dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, 6931dcbf3bd6SGeorge Wilson ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR); 6932dcbf3bd6SGeorge Wilson 6933dcbf3bd6SGeorge Wilson mutex_enter(&dev->l2ad_mtx); 6934dcbf3bd6SGeorge Wilson list_insert_head(&dev->l2ad_buflist, hdr); 6935dcbf3bd6SGeorge Wilson mutex_exit(&dev->l2ad_mtx); 6936dcbf3bd6SGeorge Wilson 693716a7e5acSAndriy Gapon (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr); 6938aad02571SSaso Kiselkov 6939a52fc310SPrakash Surya /* 6940dcbf3bd6SGeorge Wilson * Normally the L2ARC can use the hdr's data, but if 6941dcbf3bd6SGeorge Wilson * we're sharing data between the hdr and one of its 6942dcbf3bd6SGeorge Wilson * bufs, L2ARC needs its own copy of the data so that 6943403a8da7SAndriy Gapon * the ZIO below can't race with the buf consumer. 6944403a8da7SAndriy Gapon * Another case where we need to create a copy of the 6945403a8da7SAndriy Gapon * data is when the buffer size is not device-aligned 6946403a8da7SAndriy Gapon * and we need to pad the block to make it such. 6947403a8da7SAndriy Gapon * That also keeps the clock hand suitably aligned. 6948403a8da7SAndriy Gapon * 6949403a8da7SAndriy Gapon * To ensure that the copy will be available for the 6950dcbf3bd6SGeorge Wilson * lifetime of the ZIO and be cleaned up afterwards, we 6951dcbf3bd6SGeorge Wilson * add it to the l2arc_free_on_write queue. 6952a52fc310SPrakash Surya */ 6953770499e1SDan Kimmel abd_t *to_write; 695416a7e5acSAndriy Gapon if (!HDR_SHARED_DATA(hdr) && psize == asize) { 6955770499e1SDan Kimmel to_write = hdr->b_l1hdr.b_pabd; 6956dcbf3bd6SGeorge Wilson } else { 6957403a8da7SAndriy Gapon to_write = abd_alloc_for_io(asize, 6958770499e1SDan Kimmel HDR_ISTYPE_METADATA(hdr)); 695916a7e5acSAndriy Gapon abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize); 696016a7e5acSAndriy Gapon if (asize != psize) { 696116a7e5acSAndriy Gapon abd_zero_off(to_write, psize, 696216a7e5acSAndriy Gapon asize - psize); 6963403a8da7SAndriy Gapon } 696416a7e5acSAndriy Gapon l2arc_free_abd_on_write(to_write, asize, 6965770499e1SDan Kimmel arc_buf_type(hdr)); 6966dcbf3bd6SGeorge Wilson } 6967dcbf3bd6SGeorge Wilson wzio = zio_write_phys(pio, dev->l2ad_vdev, 6968403a8da7SAndriy Gapon hdr->b_l2hdr.b_daddr, asize, to_write, 6969dcbf3bd6SGeorge Wilson ZIO_CHECKSUM_OFF, NULL, hdr, 6970dcbf3bd6SGeorge Wilson ZIO_PRIORITY_ASYNC_WRITE, 6971dcbf3bd6SGeorge Wilson ZIO_FLAG_CANFAIL, B_FALSE); 6972aad02571SSaso Kiselkov 697316a7e5acSAndriy Gapon write_lsize += HDR_GET_LSIZE(hdr); 6974dcbf3bd6SGeorge Wilson DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 6975dcbf3bd6SGeorge Wilson zio_t *, wzio); 6976fa94a07fSbrendan 697716a7e5acSAndriy Gapon write_psize += psize; 697816a7e5acSAndriy Gapon write_asize += asize; 6979dcbf3bd6SGeorge Wilson dev->l2ad_hand += asize; 6980fa94a07fSbrendan 6981fa94a07fSbrendan mutex_exit(hash_lock); 6982fa94a07fSbrendan 6983dcbf3bd6SGeorge Wilson (void) zio_nowait(wzio); 6984aad02571SSaso Kiselkov } 6985aad02571SSaso Kiselkov 6986244781f1SPrakash Surya multilist_sublist_unlock(mls); 6987aad02571SSaso Kiselkov 6988aad02571SSaso Kiselkov if (full == B_TRUE) 6989aad02571SSaso Kiselkov break; 6990aad02571SSaso Kiselkov } 6991aad02571SSaso Kiselkov 6992aad02571SSaso Kiselkov /* No buffers selected for writing? */ 6993aad02571SSaso Kiselkov if (pio == NULL) { 699416a7e5acSAndriy Gapon ASSERT0(write_lsize); 699589c86e32SChris Williamson ASSERT(!HDR_HAS_L1HDR(head)); 699689c86e32SChris Williamson kmem_cache_free(hdr_l2only_cache, head); 6997aad02571SSaso Kiselkov return (0); 6998aad02571SSaso Kiselkov } 6999aad02571SSaso Kiselkov 7000aad02571SSaso Kiselkov ASSERT3U(write_asize, <=, target_sz); 7001fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_writes_sent); 700216a7e5acSAndriy Gapon ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize); 700316a7e5acSAndriy Gapon ARCSTAT_INCR(arcstat_l2_lsize, write_lsize); 700416a7e5acSAndriy Gapon ARCSTAT_INCR(arcstat_l2_psize, write_psize); 700516a7e5acSAndriy Gapon vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 7006fa94a07fSbrendan 7007fa94a07fSbrendan /* 7008fa94a07fSbrendan * Bump device hand to the device start if it is approaching the end. 7009fa94a07fSbrendan * l2arc_evict() will already have evicted ahead for this case. 7010fa94a07fSbrendan */ 70113a737e0dSbrendan if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 7012fa94a07fSbrendan dev->l2ad_hand = dev->l2ad_start; 7013fa94a07fSbrendan dev->l2ad_first = B_FALSE; 7014fa94a07fSbrendan } 7015fa94a07fSbrendan 70165a98e54bSBrendan Gregg - Sun Microsystems dev->l2ad_writing = B_TRUE; 7017fa94a07fSbrendan (void) zio_wait(pio); 70185a98e54bSBrendan Gregg - Sun Microsystems dev->l2ad_writing = B_FALSE; 70195a98e54bSBrendan Gregg - Sun Microsystems 7020aad02571SSaso Kiselkov return (write_asize); 7021aad02571SSaso Kiselkov } 7022aad02571SSaso Kiselkov 7023fa94a07fSbrendan /* 7024fa94a07fSbrendan * This thread feeds the L2ARC at regular intervals. This is the beating 7025fa94a07fSbrendan * heart of the L2ARC. 7026fa94a07fSbrendan */ 70273f7978d0SAlan Somers /* ARGSUSED */ 7028fa94a07fSbrendan static void 70293f7978d0SAlan Somers l2arc_feed_thread(void *unused) 7030fa94a07fSbrendan { 7031fa94a07fSbrendan callb_cpr_t cpr; 7032fa94a07fSbrendan l2arc_dev_t *dev; 7033fa94a07fSbrendan spa_t *spa; 70345a98e54bSBrendan Gregg - Sun Microsystems uint64_t size, wrote; 7035d3d50737SRafael Vanoni clock_t begin, next = ddi_get_lbolt(); 7036fa94a07fSbrendan 7037fa94a07fSbrendan CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 7038fa94a07fSbrendan 7039fa94a07fSbrendan mutex_enter(&l2arc_feed_thr_lock); 7040fa94a07fSbrendan 7041fa94a07fSbrendan while (l2arc_thread_exit == 0) { 7042fa94a07fSbrendan CALLB_CPR_SAFE_BEGIN(&cpr); 7043fa94a07fSbrendan (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 70445a98e54bSBrendan Gregg - Sun Microsystems next); 7045fa94a07fSbrendan CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 7046d3d50737SRafael Vanoni next = ddi_get_lbolt() + hz; 7047fa94a07fSbrendan 70483a737e0dSbrendan /* 70493a737e0dSbrendan * Quick check for L2ARC devices. 70503a737e0dSbrendan */ 7051c5904d13Seschrock mutex_enter(&l2arc_dev_mtx); 70523a737e0dSbrendan if (l2arc_ndev == 0) { 70533a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 70543a737e0dSbrendan continue; 70553a737e0dSbrendan } 70563a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 7057d3d50737SRafael Vanoni begin = ddi_get_lbolt(); 7058c5904d13Seschrock 7059fa94a07fSbrendan /* 7060c5904d13Seschrock * This selects the next l2arc device to write to, and in 7061c5904d13Seschrock * doing so the next spa to feed from: dev->l2ad_spa. This 70623a737e0dSbrendan * will return NULL if there are now no l2arc devices or if 70633a737e0dSbrendan * they are all faulted. 70643a737e0dSbrendan * 70653a737e0dSbrendan * If a device is returned, its spa's config lock is also 70663a737e0dSbrendan * held to prevent device removal. l2arc_dev_get_next() 70673a737e0dSbrendan * will grab and release l2arc_dev_mtx. 7068fa94a07fSbrendan */ 70693a737e0dSbrendan if ((dev = l2arc_dev_get_next()) == NULL) 7070fa94a07fSbrendan continue; 70713a737e0dSbrendan 70723a737e0dSbrendan spa = dev->l2ad_spa; 7073dcbf3bd6SGeorge Wilson ASSERT3P(spa, !=, NULL); 7074fa94a07fSbrendan 7075f9af39baSGeorge Wilson /* 7076f9af39baSGeorge Wilson * If the pool is read-only then force the feed thread to 7077f9af39baSGeorge Wilson * sleep a little longer. 7078f9af39baSGeorge Wilson */ 7079f9af39baSGeorge Wilson if (!spa_writeable(spa)) { 7080f9af39baSGeorge Wilson next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 7081f9af39baSGeorge Wilson spa_config_exit(spa, SCL_L2ARC, dev); 7082f9af39baSGeorge Wilson continue; 7083f9af39baSGeorge Wilson } 7084f9af39baSGeorge Wilson 7085fa94a07fSbrendan /* 7086fa94a07fSbrendan * Avoid contributing to memory pressure. 7087fa94a07fSbrendan */ 7088fa94a07fSbrendan if (arc_reclaim_needed()) { 7089fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 7090e14bb325SJeff Bonwick spa_config_exit(spa, SCL_L2ARC, dev); 7091fa94a07fSbrendan continue; 7092fa94a07fSbrendan } 7093fa94a07fSbrendan 7094fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_feeds); 7095fa94a07fSbrendan 7096aad02571SSaso Kiselkov size = l2arc_write_size(); 70973a737e0dSbrendan 7098fa94a07fSbrendan /* 7099fa94a07fSbrendan * Evict L2ARC buffers that will be overwritten. 7100fa94a07fSbrendan */ 71013a737e0dSbrendan l2arc_evict(dev, size, B_FALSE); 7102fa94a07fSbrendan 7103fa94a07fSbrendan /* 7104fa94a07fSbrendan * Write ARC buffers. 7105fa94a07fSbrendan */ 7106dcbf3bd6SGeorge Wilson wrote = l2arc_write_buffers(spa, dev, size); 71075a98e54bSBrendan Gregg - Sun Microsystems 71085a98e54bSBrendan Gregg - Sun Microsystems /* 71095a98e54bSBrendan Gregg - Sun Microsystems * Calculate interval between writes. 71105a98e54bSBrendan Gregg - Sun Microsystems */ 71115a98e54bSBrendan Gregg - Sun Microsystems next = l2arc_write_interval(begin, size, wrote); 7112e14bb325SJeff Bonwick spa_config_exit(spa, SCL_L2ARC, dev); 7113fa94a07fSbrendan } 7114fa94a07fSbrendan 7115fa94a07fSbrendan l2arc_thread_exit = 0; 7116fa94a07fSbrendan cv_broadcast(&l2arc_feed_thr_cv); 7117fa94a07fSbrendan CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 7118fa94a07fSbrendan thread_exit(); 7119fa94a07fSbrendan } 7120fa94a07fSbrendan 7121c5904d13Seschrock boolean_t 7122c5904d13Seschrock l2arc_vdev_present(vdev_t *vd) 7123c5904d13Seschrock { 7124c5904d13Seschrock l2arc_dev_t *dev; 7125c5904d13Seschrock 7126c5904d13Seschrock mutex_enter(&l2arc_dev_mtx); 7127c5904d13Seschrock for (dev = list_head(l2arc_dev_list); dev != NULL; 7128c5904d13Seschrock dev = list_next(l2arc_dev_list, dev)) { 7129c5904d13Seschrock if (dev->l2ad_vdev == vd) 7130c5904d13Seschrock break; 7131c5904d13Seschrock } 7132c5904d13Seschrock mutex_exit(&l2arc_dev_mtx); 7133c5904d13Seschrock 7134c5904d13Seschrock return (dev != NULL); 7135c5904d13Seschrock } 7136c5904d13Seschrock 7137fa94a07fSbrendan /* 7138fa94a07fSbrendan * Add a vdev for use by the L2ARC. By this point the spa has already 7139fa94a07fSbrendan * validated the vdev and opened it. 7140fa94a07fSbrendan */ 7141fa94a07fSbrendan void 7142573ca77eSGeorge Wilson l2arc_add_vdev(spa_t *spa, vdev_t *vd) 7143fa94a07fSbrendan { 7144fa94a07fSbrendan l2arc_dev_t *adddev; 7145fa94a07fSbrendan 7146c5904d13Seschrock ASSERT(!l2arc_vdev_present(vd)); 7147c5904d13Seschrock 7148fa94a07fSbrendan /* 7149fa94a07fSbrendan * Create a new l2arc device entry. 7150fa94a07fSbrendan */ 7151fa94a07fSbrendan adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 7152fa94a07fSbrendan adddev->l2ad_spa = spa; 7153fa94a07fSbrendan adddev->l2ad_vdev = vd; 7154573ca77eSGeorge Wilson adddev->l2ad_start = VDEV_LABEL_START_SIZE; 7155573ca77eSGeorge Wilson adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 7156fa94a07fSbrendan adddev->l2ad_hand = adddev->l2ad_start; 7157fa94a07fSbrendan adddev->l2ad_first = B_TRUE; 71585a98e54bSBrendan Gregg - Sun Microsystems adddev->l2ad_writing = B_FALSE; 7159fa94a07fSbrendan 716089c86e32SChris Williamson mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 7161fa94a07fSbrendan /* 7162fa94a07fSbrendan * This is a list of all ARC buffers that are still valid on the 7163fa94a07fSbrendan * device. 7164fa94a07fSbrendan */ 716589c86e32SChris Williamson list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 716689c86e32SChris Williamson offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 7167fa94a07fSbrendan 7168b24ab676SJeff Bonwick vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 7169a52fc310SPrakash Surya refcount_create(&adddev->l2ad_alloc); 7170fa94a07fSbrendan 7171fa94a07fSbrendan /* 7172fa94a07fSbrendan * Add device to global list 7173fa94a07fSbrendan */ 7174fa94a07fSbrendan mutex_enter(&l2arc_dev_mtx); 7175fa94a07fSbrendan list_insert_head(l2arc_dev_list, adddev); 7176fa94a07fSbrendan atomic_inc_64(&l2arc_ndev); 7177fa94a07fSbrendan mutex_exit(&l2arc_dev_mtx); 7178fa94a07fSbrendan } 7179fa94a07fSbrendan 7180fa94a07fSbrendan /* 7181fa94a07fSbrendan * Remove a vdev from the L2ARC. 7182fa94a07fSbrendan */ 7183fa94a07fSbrendan void 7184fa94a07fSbrendan l2arc_remove_vdev(vdev_t *vd) 7185fa94a07fSbrendan { 7186fa94a07fSbrendan l2arc_dev_t *dev, *nextdev, *remdev = NULL; 7187fa94a07fSbrendan 7188fa94a07fSbrendan /* 7189fa94a07fSbrendan * Find the device by vdev 7190fa94a07fSbrendan */ 7191fa94a07fSbrendan mutex_enter(&l2arc_dev_mtx); 7192fa94a07fSbrendan for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 7193fa94a07fSbrendan nextdev = list_next(l2arc_dev_list, dev); 7194fa94a07fSbrendan if (vd == dev->l2ad_vdev) { 7195fa94a07fSbrendan remdev = dev; 7196fa94a07fSbrendan break; 7197fa94a07fSbrendan } 7198fa94a07fSbrendan } 7199dcbf3bd6SGeorge Wilson ASSERT3P(remdev, !=, NULL); 7200fa94a07fSbrendan 7201fa94a07fSbrendan /* 7202fa94a07fSbrendan * Remove device from global list 7203fa94a07fSbrendan */ 7204fa94a07fSbrendan list_remove(l2arc_dev_list, remdev); 7205fa94a07fSbrendan l2arc_dev_last = NULL; /* may have been invalidated */ 72063a737e0dSbrendan atomic_dec_64(&l2arc_ndev); 72073a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 7208fa94a07fSbrendan 7209fa94a07fSbrendan /* 7210fa94a07fSbrendan * Clear all buflists and ARC references. L2ARC device flush. 7211fa94a07fSbrendan */ 7212fa94a07fSbrendan l2arc_evict(remdev, 0, B_TRUE); 721389c86e32SChris Williamson list_destroy(&remdev->l2ad_buflist); 721489c86e32SChris Williamson mutex_destroy(&remdev->l2ad_mtx); 7215a52fc310SPrakash Surya refcount_destroy(&remdev->l2ad_alloc); 7216fa94a07fSbrendan kmem_free(remdev, sizeof (l2arc_dev_t)); 7217fa94a07fSbrendan } 7218fa94a07fSbrendan 7219fa94a07fSbrendan void 7220e14bb325SJeff Bonwick l2arc_init(void) 7221fa94a07fSbrendan { 7222fa94a07fSbrendan l2arc_thread_exit = 0; 7223fa94a07fSbrendan l2arc_ndev = 0; 7224fa94a07fSbrendan l2arc_writes_sent = 0; 7225fa94a07fSbrendan l2arc_writes_done = 0; 7226fa94a07fSbrendan 7227fa94a07fSbrendan mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7228fa94a07fSbrendan cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7229fa94a07fSbrendan mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7230fa94a07fSbrendan mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7231fa94a07fSbrendan 7232fa94a07fSbrendan l2arc_dev_list = &L2ARC_dev_list; 7233fa94a07fSbrendan l2arc_free_on_write = &L2ARC_free_on_write; 7234fa94a07fSbrendan list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7235fa94a07fSbrendan offsetof(l2arc_dev_t, l2ad_node)); 7236fa94a07fSbrendan list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7237fa94a07fSbrendan offsetof(l2arc_data_free_t, l2df_list_node)); 7238fa94a07fSbrendan } 7239fa94a07fSbrendan 7240fa94a07fSbrendan void 7241e14bb325SJeff Bonwick l2arc_fini(void) 7242fa94a07fSbrendan { 72433a737e0dSbrendan /* 72443a737e0dSbrendan * This is called from dmu_fini(), which is called from spa_fini(); 72453a737e0dSbrendan * Because of this, we can assume that all l2arc devices have 72463a737e0dSbrendan * already been removed when the pools themselves were removed. 72473a737e0dSbrendan */ 72483a737e0dSbrendan 72493a737e0dSbrendan l2arc_do_free_on_write(); 72503a737e0dSbrendan 7251fa94a07fSbrendan mutex_destroy(&l2arc_feed_thr_lock); 7252fa94a07fSbrendan cv_destroy(&l2arc_feed_thr_cv); 7253fa94a07fSbrendan mutex_destroy(&l2arc_dev_mtx); 7254fa94a07fSbrendan mutex_destroy(&l2arc_free_on_write_mtx); 7255fa94a07fSbrendan 7256fa94a07fSbrendan list_destroy(l2arc_dev_list); 7257fa94a07fSbrendan list_destroy(l2arc_free_on_write); 7258fa94a07fSbrendan } 7259e14bb325SJeff Bonwick 7260e14bb325SJeff Bonwick void 7261e14bb325SJeff Bonwick l2arc_start(void) 7262e14bb325SJeff Bonwick { 72638ad4d6ddSJeff Bonwick if (!(spa_mode_global & FWRITE)) 7264e14bb325SJeff Bonwick return; 7265e14bb325SJeff Bonwick 7266e14bb325SJeff Bonwick (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7267e14bb325SJeff Bonwick TS_RUN, minclsyspri); 7268e14bb325SJeff Bonwick } 7269e14bb325SJeff Bonwick 7270e14bb325SJeff Bonwick void 7271e14bb325SJeff Bonwick l2arc_stop(void) 7272e14bb325SJeff Bonwick { 72738ad4d6ddSJeff Bonwick if (!(spa_mode_global & FWRITE)) 7274e14bb325SJeff Bonwick return; 7275e14bb325SJeff Bonwick 7276e14bb325SJeff Bonwick mutex_enter(&l2arc_feed_thr_lock); 7277e14bb325SJeff Bonwick cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7278e14bb325SJeff Bonwick l2arc_thread_exit = 1; 7279e14bb325SJeff Bonwick while (l2arc_thread_exit != 0) 7280e14bb325SJeff Bonwick cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7281e14bb325SJeff Bonwick mutex_exit(&l2arc_feed_thr_lock); 7282e14bb325SJeff Bonwick } 7283