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. 2631c46cf2SAlek Pinchuk * Copyright 2015 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 131*770499e1SDan 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 134*770499e1SDan Kimmel * ability to store the physical data (b_pabd) associated with the DVA of the 135*770499e1SDan 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 138*770499e1SDan 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 177*770499e1SDan 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 194*770499e1SDan Kimmel * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a 195*770499e1SDan 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 +---->+---------+ 219*770499e1SDan 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 * 233*770499e1SDan 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 238*770499e1SDan 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 * 244*770499e1SDan Kimmel * When the L2ARC is in use, it will also take advantage of the b_pabd. The 245*770499e1SDan 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> 266*770499e1SDan Kimmel #include <sys/zio_checksum.h> 267244781f1SPrakash Surya #include <sys/multilist.h> 268*770499e1SDan 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 304*770499e1SDan 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 /* 467*770499e1SDan 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 /* 473*770499e1SDan 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; 634fa94a07fSbrendan kstat_named_t arcstat_l2_size; 635aad02571SSaso Kiselkov kstat_named_t arcstat_l2_asize; 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; 887*770499e1SDan 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 */ 1078fa94a07fSbrendan } l2arc_read_callback_t; 1079fa94a07fSbrendan 1080fa94a07fSbrendan typedef struct l2arc_write_callback { 1081fa94a07fSbrendan l2arc_dev_t *l2wcb_dev; /* device info */ 1082fa94a07fSbrendan arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1083fa94a07fSbrendan } l2arc_write_callback_t; 1084fa94a07fSbrendan 1085fa94a07fSbrendan typedef struct l2arc_data_free { 1086fa94a07fSbrendan /* protected by l2arc_free_on_write_mtx */ 1087*770499e1SDan Kimmel abd_t *l2df_abd; 1088fa94a07fSbrendan size_t l2df_size; 1089dcbf3bd6SGeorge Wilson arc_buf_contents_t l2df_type; 1090fa94a07fSbrendan list_node_t l2df_list_node; 1091fa94a07fSbrendan } l2arc_data_free_t; 1092fa94a07fSbrendan 1093fa94a07fSbrendan static kmutex_t l2arc_feed_thr_lock; 1094fa94a07fSbrendan static kcondvar_t l2arc_feed_thr_cv; 1095fa94a07fSbrendan static uint8_t l2arc_thread_exit; 1096fa94a07fSbrendan 1097*770499e1SDan Kimmel static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *); 1098dcbf3bd6SGeorge Wilson static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); 1099*770499e1SDan Kimmel static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *); 1100*770499e1SDan Kimmel static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *); 1101dcbf3bd6SGeorge Wilson static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); 1102*770499e1SDan Kimmel static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag); 1103*770499e1SDan Kimmel static void arc_hdr_free_pabd(arc_buf_hdr_t *); 1104*770499e1SDan Kimmel static void arc_hdr_alloc_pabd(arc_buf_hdr_t *); 11057adb730bSGeorge Wilson static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1106244781f1SPrakash Surya static boolean_t arc_is_overflowing(); 11077adb730bSGeorge Wilson static void arc_buf_watch(arc_buf_t *); 11087adb730bSGeorge Wilson 110989c86e32SChris Williamson static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 111089c86e32SChris Williamson static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1111dcbf3bd6SGeorge Wilson static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1112dcbf3bd6SGeorge Wilson static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 111389c86e32SChris Williamson 11147adb730bSGeorge Wilson static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 11157adb730bSGeorge Wilson static void l2arc_read_done(zio_t *); 1116fa94a07fSbrendan 1117fa9e4066Sahrens static uint64_t 1118ac05c741SMark Maybee buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1119fa9e4066Sahrens { 1120fa9e4066Sahrens uint8_t *vdva = (uint8_t *)dva; 1121fa9e4066Sahrens uint64_t crc = -1ULL; 1122fa9e4066Sahrens int i; 1123fa9e4066Sahrens 1124fa9e4066Sahrens ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1125fa9e4066Sahrens 1126fa9e4066Sahrens for (i = 0; i < sizeof (dva_t); i++) 1127fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1128fa9e4066Sahrens 1129ac05c741SMark Maybee crc ^= (spa>>8) ^ birth; 1130fa9e4066Sahrens 1131fa9e4066Sahrens return (crc); 1132fa9e4066Sahrens } 1133fa9e4066Sahrens 1134dcbf3bd6SGeorge Wilson #define HDR_EMPTY(hdr) \ 1135dcbf3bd6SGeorge Wilson ((hdr)->b_dva.dva_word[0] == 0 && \ 1136dcbf3bd6SGeorge Wilson (hdr)->b_dva.dva_word[1] == 0) 1137fa9e4066Sahrens 1138dcbf3bd6SGeorge Wilson #define HDR_EQUAL(spa, dva, birth, hdr) \ 1139dcbf3bd6SGeorge Wilson ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1140dcbf3bd6SGeorge Wilson ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1141dcbf3bd6SGeorge Wilson ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) 1142fa9e4066Sahrens 11433f9d6ad7SLin Ling static void 11443f9d6ad7SLin Ling buf_discard_identity(arc_buf_hdr_t *hdr) 11453f9d6ad7SLin Ling { 11463f9d6ad7SLin Ling hdr->b_dva.dva_word[0] = 0; 11473f9d6ad7SLin Ling hdr->b_dva.dva_word[1] = 0; 11483f9d6ad7SLin Ling hdr->b_birth = 0; 11493f9d6ad7SLin Ling } 11503f9d6ad7SLin Ling 1151fa9e4066Sahrens static arc_buf_hdr_t * 11525d7b4d43SMatthew Ahrens buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1153fa9e4066Sahrens { 11545d7b4d43SMatthew Ahrens const dva_t *dva = BP_IDENTITY(bp); 11555d7b4d43SMatthew Ahrens uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1156fa9e4066Sahrens uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1157fa9e4066Sahrens kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 11587adb730bSGeorge Wilson arc_buf_hdr_t *hdr; 1159fa9e4066Sahrens 1160fa9e4066Sahrens mutex_enter(hash_lock); 11617adb730bSGeorge Wilson for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 11627adb730bSGeorge Wilson hdr = hdr->b_hash_next) { 1163dcbf3bd6SGeorge Wilson if (HDR_EQUAL(spa, dva, birth, hdr)) { 1164fa9e4066Sahrens *lockp = hash_lock; 11657adb730bSGeorge Wilson return (hdr); 1166fa9e4066Sahrens } 1167fa9e4066Sahrens } 1168fa9e4066Sahrens mutex_exit(hash_lock); 1169fa9e4066Sahrens *lockp = NULL; 1170fa9e4066Sahrens return (NULL); 1171fa9e4066Sahrens } 1172fa9e4066Sahrens 1173fa9e4066Sahrens /* 1174fa9e4066Sahrens * Insert an entry into the hash table. If there is already an element 1175fa9e4066Sahrens * equal to elem in the hash table, then the already existing element 1176fa9e4066Sahrens * will be returned and the new element will not be inserted. 1177fa9e4066Sahrens * Otherwise returns NULL. 117889c86e32SChris Williamson * If lockp == NULL, the caller is assumed to already hold the hash lock. 1179fa9e4066Sahrens */ 1180fa9e4066Sahrens static arc_buf_hdr_t * 11817adb730bSGeorge Wilson buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1182fa9e4066Sahrens { 11837adb730bSGeorge Wilson uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1184fa9e4066Sahrens kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 11857adb730bSGeorge Wilson arc_buf_hdr_t *fhdr; 118644cb6abcSbmc uint32_t i; 1187fa9e4066Sahrens 11887adb730bSGeorge Wilson ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 11897adb730bSGeorge Wilson ASSERT(hdr->b_birth != 0); 11907adb730bSGeorge Wilson ASSERT(!HDR_IN_HASH_TABLE(hdr)); 119189c86e32SChris Williamson 119289c86e32SChris Williamson if (lockp != NULL) { 119389c86e32SChris Williamson *lockp = hash_lock; 119489c86e32SChris Williamson mutex_enter(hash_lock); 119589c86e32SChris Williamson } else { 119689c86e32SChris Williamson ASSERT(MUTEX_HELD(hash_lock)); 119789c86e32SChris Williamson } 119889c86e32SChris Williamson 11997adb730bSGeorge Wilson for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 12007adb730bSGeorge Wilson fhdr = fhdr->b_hash_next, i++) { 1201dcbf3bd6SGeorge Wilson if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 12027adb730bSGeorge Wilson return (fhdr); 1203fa9e4066Sahrens } 1204fa9e4066Sahrens 12057adb730bSGeorge Wilson hdr->b_hash_next = buf_hash_table.ht_table[idx]; 12067adb730bSGeorge Wilson buf_hash_table.ht_table[idx] = hdr; 1207dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1208fa9e4066Sahrens 1209fa9e4066Sahrens /* collect some hash table performance data */ 1210fa9e4066Sahrens if (i > 0) { 121144cb6abcSbmc ARCSTAT_BUMP(arcstat_hash_collisions); 1212fa9e4066Sahrens if (i == 1) 121344cb6abcSbmc ARCSTAT_BUMP(arcstat_hash_chains); 121444cb6abcSbmc 121544cb6abcSbmc ARCSTAT_MAX(arcstat_hash_chain_max, i); 1216fa9e4066Sahrens } 121744cb6abcSbmc 121844cb6abcSbmc ARCSTAT_BUMP(arcstat_hash_elements); 121944cb6abcSbmc ARCSTAT_MAXSTAT(arcstat_hash_elements); 1220fa9e4066Sahrens 1221fa9e4066Sahrens return (NULL); 1222fa9e4066Sahrens } 1223fa9e4066Sahrens 1224fa9e4066Sahrens static void 12257adb730bSGeorge Wilson buf_hash_remove(arc_buf_hdr_t *hdr) 1226fa9e4066Sahrens { 12277adb730bSGeorge Wilson arc_buf_hdr_t *fhdr, **hdrp; 12287adb730bSGeorge Wilson uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1229fa9e4066Sahrens 1230fa9e4066Sahrens ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 12317adb730bSGeorge Wilson ASSERT(HDR_IN_HASH_TABLE(hdr)); 1232fa9e4066Sahrens 12337adb730bSGeorge Wilson hdrp = &buf_hash_table.ht_table[idx]; 12347adb730bSGeorge Wilson while ((fhdr = *hdrp) != hdr) { 1235dcbf3bd6SGeorge Wilson ASSERT3P(fhdr, !=, NULL); 12367adb730bSGeorge Wilson hdrp = &fhdr->b_hash_next; 1237fa9e4066Sahrens } 12387adb730bSGeorge Wilson *hdrp = hdr->b_hash_next; 12397adb730bSGeorge Wilson hdr->b_hash_next = NULL; 1240dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1241fa9e4066Sahrens 1242fa9e4066Sahrens /* collect some hash table performance data */ 124344cb6abcSbmc ARCSTAT_BUMPDOWN(arcstat_hash_elements); 124444cb6abcSbmc 1245fa9e4066Sahrens if (buf_hash_table.ht_table[idx] && 1246fa9e4066Sahrens buf_hash_table.ht_table[idx]->b_hash_next == NULL) 124744cb6abcSbmc ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1248fa9e4066Sahrens } 1249fa9e4066Sahrens 1250fa9e4066Sahrens /* 1251fa9e4066Sahrens * Global data structures and functions for the buf kmem cache. 1252fa9e4066Sahrens */ 125389c86e32SChris Williamson static kmem_cache_t *hdr_full_cache; 125489c86e32SChris Williamson static kmem_cache_t *hdr_l2only_cache; 1255fa9e4066Sahrens static kmem_cache_t *buf_cache; 1256fa9e4066Sahrens 1257fa9e4066Sahrens static void 1258fa9e4066Sahrens buf_fini(void) 1259fa9e4066Sahrens { 1260fa9e4066Sahrens int i; 1261fa9e4066Sahrens 1262fa9e4066Sahrens kmem_free(buf_hash_table.ht_table, 1263fa9e4066Sahrens (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1264fa9e4066Sahrens for (i = 0; i < BUF_LOCKS; i++) 1265fa9e4066Sahrens mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 126689c86e32SChris Williamson kmem_cache_destroy(hdr_full_cache); 126789c86e32SChris Williamson kmem_cache_destroy(hdr_l2only_cache); 1268fa9e4066Sahrens kmem_cache_destroy(buf_cache); 1269fa9e4066Sahrens } 1270fa9e4066Sahrens 1271fa9e4066Sahrens /* 1272fa9e4066Sahrens * Constructor callback - called when the cache is empty 1273fa9e4066Sahrens * and a new buf is requested. 1274fa9e4066Sahrens */ 1275fa9e4066Sahrens /* ARGSUSED */ 1276fa9e4066Sahrens static int 127789c86e32SChris Williamson hdr_full_cons(void *vbuf, void *unused, int kmflag) 1278fa9e4066Sahrens { 12797adb730bSGeorge Wilson arc_buf_hdr_t *hdr = vbuf; 1280fa9e4066Sahrens 128189c86e32SChris Williamson bzero(hdr, HDR_FULL_SIZE); 128289c86e32SChris Williamson cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 128389c86e32SChris Williamson refcount_create(&hdr->b_l1hdr.b_refcnt); 128489c86e32SChris Williamson mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1285244781f1SPrakash Surya multilist_link_init(&hdr->b_l1hdr.b_arc_node); 128689c86e32SChris Williamson arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 128789c86e32SChris Williamson 128889c86e32SChris Williamson return (0); 128989c86e32SChris Williamson } 129089c86e32SChris Williamson 129189c86e32SChris Williamson /* ARGSUSED */ 129289c86e32SChris Williamson static int 129389c86e32SChris Williamson hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 129489c86e32SChris Williamson { 129589c86e32SChris Williamson arc_buf_hdr_t *hdr = vbuf; 129689c86e32SChris Williamson 129789c86e32SChris Williamson bzero(hdr, HDR_L2ONLY_SIZE); 129889c86e32SChris Williamson arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1299fa94a07fSbrendan 1300fa9e4066Sahrens return (0); 1301fa9e4066Sahrens } 1302fa9e4066Sahrens 13036f83844dSMark Maybee /* ARGSUSED */ 13046f83844dSMark Maybee static int 13056f83844dSMark Maybee buf_cons(void *vbuf, void *unused, int kmflag) 13066f83844dSMark Maybee { 13076f83844dSMark Maybee arc_buf_t *buf = vbuf; 13086f83844dSMark Maybee 13096f83844dSMark Maybee bzero(buf, sizeof (arc_buf_t)); 13103f9d6ad7SLin Ling mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 13115a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 13125a98e54bSBrendan Gregg - Sun Microsystems 13136f83844dSMark Maybee return (0); 13146f83844dSMark Maybee } 13156f83844dSMark Maybee 1316fa9e4066Sahrens /* 1317fa9e4066Sahrens * Destructor callback - called when a cached buf is 1318fa9e4066Sahrens * no longer required. 1319fa9e4066Sahrens */ 1320fa9e4066Sahrens /* ARGSUSED */ 1321fa9e4066Sahrens static void 132289c86e32SChris Williamson hdr_full_dest(void *vbuf, void *unused) 132389c86e32SChris Williamson { 132489c86e32SChris Williamson arc_buf_hdr_t *hdr = vbuf; 132589c86e32SChris Williamson 1326dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 132789c86e32SChris Williamson cv_destroy(&hdr->b_l1hdr.b_cv); 132889c86e32SChris Williamson refcount_destroy(&hdr->b_l1hdr.b_refcnt); 132989c86e32SChris Williamson mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1330244781f1SPrakash Surya ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 133189c86e32SChris Williamson arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 133289c86e32SChris Williamson } 133389c86e32SChris Williamson 133489c86e32SChris Williamson /* ARGSUSED */ 133589c86e32SChris Williamson static void 133689c86e32SChris Williamson hdr_l2only_dest(void *vbuf, void *unused) 1337fa9e4066Sahrens { 13387adb730bSGeorge Wilson arc_buf_hdr_t *hdr = vbuf; 1339fa9e4066Sahrens 1340dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 134189c86e32SChris Williamson arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1342fa9e4066Sahrens } 1343fa9e4066Sahrens 13446f83844dSMark Maybee /* ARGSUSED */ 13456f83844dSMark Maybee static void 13466f83844dSMark Maybee buf_dest(void *vbuf, void *unused) 13476f83844dSMark Maybee { 13486f83844dSMark Maybee arc_buf_t *buf = vbuf; 13496f83844dSMark Maybee 13503f9d6ad7SLin Ling mutex_destroy(&buf->b_evict_lock); 13515a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 13526f83844dSMark Maybee } 13536f83844dSMark Maybee 1354fa9e4066Sahrens /* 1355fa9e4066Sahrens * Reclaim callback -- invoked when memory is low. 1356fa9e4066Sahrens */ 1357fa9e4066Sahrens /* ARGSUSED */ 1358fa9e4066Sahrens static void 1359fa9e4066Sahrens hdr_recl(void *unused) 1360fa9e4066Sahrens { 1361fa9e4066Sahrens dprintf("hdr_recl called\n"); 136249e3519aSmaybee /* 136349e3519aSmaybee * umem calls the reclaim func when we destroy the buf cache, 136449e3519aSmaybee * which is after we do arc_fini(). 136549e3519aSmaybee */ 136649e3519aSmaybee if (!arc_dead) 1367244781f1SPrakash Surya cv_signal(&arc_reclaim_thread_cv); 1368fa9e4066Sahrens } 1369fa9e4066Sahrens 1370fa9e4066Sahrens static void 1371fa9e4066Sahrens buf_init(void) 1372fa9e4066Sahrens { 1373fa9e4066Sahrens uint64_t *ct; 1374ea8dc4b6Seschrock uint64_t hsize = 1ULL << 12; 1375fa9e4066Sahrens int i, j; 1376fa9e4066Sahrens 1377fa9e4066Sahrens /* 1378fa9e4066Sahrens * The hash table is big enough to fill all of physical memory 137963e911b6SMatthew Ahrens * with an average block size of zfs_arc_average_blocksize (default 8K). 138063e911b6SMatthew Ahrens * By default, the table will take up 138163e911b6SMatthew Ahrens * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1382fa9e4066Sahrens */ 138363e911b6SMatthew Ahrens while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE) 1384fa9e4066Sahrens hsize <<= 1; 1385ea8dc4b6Seschrock retry: 1386fa9e4066Sahrens buf_hash_table.ht_mask = hsize - 1; 1387ea8dc4b6Seschrock buf_hash_table.ht_table = 1388ea8dc4b6Seschrock kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1389ea8dc4b6Seschrock if (buf_hash_table.ht_table == NULL) { 1390ea8dc4b6Seschrock ASSERT(hsize > (1ULL << 8)); 1391ea8dc4b6Seschrock hsize >>= 1; 1392ea8dc4b6Seschrock goto retry; 1393ea8dc4b6Seschrock } 1394fa9e4066Sahrens 139589c86e32SChris Williamson hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 139689c86e32SChris Williamson 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 139789c86e32SChris Williamson hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 139889c86e32SChris Williamson HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 139989c86e32SChris Williamson NULL, NULL, 0); 1400fa9e4066Sahrens buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 14016f83844dSMark Maybee 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1402fa9e4066Sahrens 1403fa9e4066Sahrens for (i = 0; i < 256; i++) 1404fa9e4066Sahrens for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1405fa9e4066Sahrens *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1406fa9e4066Sahrens 1407fa9e4066Sahrens for (i = 0; i < BUF_LOCKS; i++) { 1408fa9e4066Sahrens mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1409fa9e4066Sahrens NULL, MUTEX_DEFAULT, NULL); 1410fa9e4066Sahrens } 1411fa9e4066Sahrens } 1412fa9e4066Sahrens 14135602294fSDan Kimmel /* 14145602294fSDan Kimmel * This is the size that the buf occupies in memory. If the buf is compressed, 14155602294fSDan Kimmel * it will correspond to the compressed size. You should use this method of 14165602294fSDan Kimmel * getting the buf size unless you explicitly need the logical size. 14175602294fSDan Kimmel */ 14185602294fSDan Kimmel int32_t 14195602294fSDan Kimmel arc_buf_size(arc_buf_t *buf) 14205602294fSDan Kimmel { 14215602294fSDan Kimmel return (ARC_BUF_COMPRESSED(buf) ? 14225602294fSDan Kimmel HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr)); 14235602294fSDan Kimmel } 14245602294fSDan Kimmel 14255602294fSDan Kimmel int32_t 14265602294fSDan Kimmel arc_buf_lsize(arc_buf_t *buf) 14275602294fSDan Kimmel { 14285602294fSDan Kimmel return (HDR_GET_LSIZE(buf->b_hdr)); 14295602294fSDan Kimmel } 14305602294fSDan Kimmel 14315602294fSDan Kimmel enum zio_compress 14325602294fSDan Kimmel arc_get_compression(arc_buf_t *buf) 14335602294fSDan Kimmel { 14345602294fSDan Kimmel return (ARC_BUF_COMPRESSED(buf) ? 14355602294fSDan Kimmel HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF); 14365602294fSDan Kimmel } 14375602294fSDan Kimmel 1438dcbf3bd6SGeorge Wilson #define ARC_MINTIME (hz>>4) /* 62 ms */ 1439244781f1SPrakash Surya 1440dcbf3bd6SGeorge Wilson static inline boolean_t 1441dcbf3bd6SGeorge Wilson arc_buf_is_shared(arc_buf_t *buf) 1442dcbf3bd6SGeorge Wilson { 1443dcbf3bd6SGeorge Wilson boolean_t shared = (buf->b_data != NULL && 1444*770499e1SDan Kimmel buf->b_hdr->b_l1hdr.b_pabd != NULL && 1445*770499e1SDan Kimmel abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) && 1446*770499e1SDan Kimmel buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd)); 1447dcbf3bd6SGeorge Wilson IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); 14485602294fSDan Kimmel IMPLY(shared, ARC_BUF_SHARED(buf)); 14495602294fSDan Kimmel IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf)); 14505602294fSDan Kimmel 14515602294fSDan Kimmel /* 14525602294fSDan Kimmel * It would be nice to assert arc_can_share() too, but the "hdr isn't 14535602294fSDan Kimmel * already being shared" requirement prevents us from doing that. 14545602294fSDan Kimmel */ 14555602294fSDan Kimmel 1456dcbf3bd6SGeorge Wilson return (shared); 1457dcbf3bd6SGeorge Wilson } 1458c546f36aSArne Jansen 14595602294fSDan Kimmel /* 14605602294fSDan Kimmel * Free the checksum associated with this header. If there is no checksum, this 14615602294fSDan Kimmel * is a no-op. 14625602294fSDan Kimmel */ 1463dcbf3bd6SGeorge Wilson static inline void 1464dcbf3bd6SGeorge Wilson arc_cksum_free(arc_buf_hdr_t *hdr) 1465dcbf3bd6SGeorge Wilson { 1466dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1467dcbf3bd6SGeorge Wilson mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1468dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1469dcbf3bd6SGeorge Wilson kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); 1470dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_freeze_cksum = NULL; 147189c86e32SChris Williamson } 1472dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 147389c86e32SChris Williamson } 147489c86e32SChris Williamson 14755602294fSDan Kimmel /* 14765602294fSDan Kimmel * Return true iff at least one of the bufs on hdr is not compressed. 14775602294fSDan Kimmel */ 14785602294fSDan Kimmel static boolean_t 14795602294fSDan Kimmel arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr) 14805602294fSDan Kimmel { 14815602294fSDan Kimmel for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) { 14825602294fSDan Kimmel if (!ARC_BUF_COMPRESSED(b)) { 14835602294fSDan Kimmel return (B_TRUE); 14845602294fSDan Kimmel } 14855602294fSDan Kimmel } 14865602294fSDan Kimmel return (B_FALSE); 14875602294fSDan Kimmel } 14885602294fSDan Kimmel 14895602294fSDan Kimmel /* 14905602294fSDan Kimmel * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data 14915602294fSDan Kimmel * matches the checksum that is stored in the hdr. If there is no checksum, 14925602294fSDan Kimmel * or if the buf is compressed, this is a no-op. 14935602294fSDan Kimmel */ 14946b4acc8bSahrens static void 14956b4acc8bSahrens arc_cksum_verify(arc_buf_t *buf) 14966b4acc8bSahrens { 1497dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 14986b4acc8bSahrens zio_cksum_t zc; 14996b4acc8bSahrens 1500cc60fd72Sahrens if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 15016b4acc8bSahrens return; 15026b4acc8bSahrens 15035602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 15045602294fSDan Kimmel ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 15055602294fSDan Kimmel arc_hdr_has_uncompressed_buf(hdr)); 15065602294fSDan Kimmel return; 15075602294fSDan Kimmel } 15085602294fSDan Kimmel 1509dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1510dcbf3bd6SGeorge Wilson 1511dcbf3bd6SGeorge Wilson mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1512dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { 1513dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 15146b4acc8bSahrens return; 15156b4acc8bSahrens } 15165602294fSDan Kimmel 15175602294fSDan Kimmel fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc); 1518dcbf3bd6SGeorge Wilson if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) 15196b4acc8bSahrens panic("buffer modified while frozen!"); 1520dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 15216b4acc8bSahrens } 15226b4acc8bSahrens 1523dcbf3bd6SGeorge Wilson static boolean_t 1524dcbf3bd6SGeorge Wilson arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) 1525fa94a07fSbrendan { 1526dcbf3bd6SGeorge Wilson enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp); 1527dcbf3bd6SGeorge Wilson boolean_t valid_cksum; 1528fa94a07fSbrendan 1529dcbf3bd6SGeorge Wilson ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); 1530dcbf3bd6SGeorge Wilson VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); 1531dcbf3bd6SGeorge Wilson 1532dcbf3bd6SGeorge Wilson /* 1533dcbf3bd6SGeorge Wilson * We rely on the blkptr's checksum to determine if the block 1534dcbf3bd6SGeorge Wilson * is valid or not. When compressed arc is enabled, the l2arc 1535dcbf3bd6SGeorge Wilson * writes the block to the l2arc just as it appears in the pool. 1536dcbf3bd6SGeorge Wilson * This allows us to use the blkptr's checksum to validate the 1537dcbf3bd6SGeorge Wilson * data that we just read off of the l2arc without having to store 1538dcbf3bd6SGeorge Wilson * a separate checksum in the arc_buf_hdr_t. However, if compressed 1539dcbf3bd6SGeorge Wilson * arc is disabled, then the data written to the l2arc is always 1540dcbf3bd6SGeorge Wilson * uncompressed and won't match the block as it exists in the main 1541dcbf3bd6SGeorge Wilson * pool. When this is the case, we must first compress it if it is 1542dcbf3bd6SGeorge Wilson * compressed on the main pool before we can validate the checksum. 1543dcbf3bd6SGeorge Wilson */ 1544dcbf3bd6SGeorge Wilson if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) { 1545dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1546dcbf3bd6SGeorge Wilson uint64_t lsize = HDR_GET_LSIZE(hdr); 1547dcbf3bd6SGeorge Wilson uint64_t csize; 1548dcbf3bd6SGeorge Wilson 1549dcbf3bd6SGeorge Wilson void *cbuf = zio_buf_alloc(HDR_GET_PSIZE(hdr)); 1550*770499e1SDan Kimmel csize = zio_compress_data(compress, zio->io_abd, cbuf, lsize); 1551*770499e1SDan Kimmel 1552dcbf3bd6SGeorge Wilson ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr)); 1553dcbf3bd6SGeorge Wilson if (csize < HDR_GET_PSIZE(hdr)) { 1554dcbf3bd6SGeorge Wilson /* 1555dcbf3bd6SGeorge Wilson * Compressed blocks are always a multiple of the 1556dcbf3bd6SGeorge Wilson * smallest ashift in the pool. Ideally, we would 1557dcbf3bd6SGeorge Wilson * like to round up the csize to the next 1558dcbf3bd6SGeorge Wilson * spa_min_ashift but that value may have changed 1559dcbf3bd6SGeorge Wilson * since the block was last written. Instead, 1560dcbf3bd6SGeorge Wilson * we rely on the fact that the hdr's psize 1561dcbf3bd6SGeorge Wilson * was set to the psize of the block when it was 1562dcbf3bd6SGeorge Wilson * last written. We set the csize to that value 1563dcbf3bd6SGeorge Wilson * and zero out any part that should not contain 1564dcbf3bd6SGeorge Wilson * data. 1565dcbf3bd6SGeorge Wilson */ 1566dcbf3bd6SGeorge Wilson bzero((char *)cbuf + csize, HDR_GET_PSIZE(hdr) - csize); 1567dcbf3bd6SGeorge Wilson csize = HDR_GET_PSIZE(hdr); 1568dcbf3bd6SGeorge Wilson } 1569dcbf3bd6SGeorge Wilson zio_push_transform(zio, cbuf, csize, HDR_GET_PSIZE(hdr), NULL); 1570dcbf3bd6SGeorge Wilson } 1571fa94a07fSbrendan 1572dcbf3bd6SGeorge Wilson /* 1573dcbf3bd6SGeorge Wilson * Block pointers always store the checksum for the logical data. 1574dcbf3bd6SGeorge Wilson * If the block pointer has the gang bit set, then the checksum 1575dcbf3bd6SGeorge Wilson * it represents is for the reconstituted data and not for an 1576dcbf3bd6SGeorge Wilson * individual gang member. The zio pipeline, however, must be able to 1577dcbf3bd6SGeorge Wilson * determine the checksum of each of the gang constituents so it 1578dcbf3bd6SGeorge Wilson * treats the checksum comparison differently than what we need 1579dcbf3bd6SGeorge Wilson * for l2arc blocks. This prevents us from using the 1580dcbf3bd6SGeorge Wilson * zio_checksum_error() interface directly. Instead we must call the 1581dcbf3bd6SGeorge Wilson * zio_checksum_error_impl() so that we can ensure the checksum is 1582dcbf3bd6SGeorge Wilson * generated using the correct checksum algorithm and accounts for the 1583dcbf3bd6SGeorge Wilson * logical I/O size and not just a gang fragment. 1584dcbf3bd6SGeorge Wilson */ 1585dcbf3bd6SGeorge Wilson valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp, 1586*770499e1SDan Kimmel BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size, 1587dcbf3bd6SGeorge Wilson zio->io_offset, NULL) == 0); 1588dcbf3bd6SGeorge Wilson zio_pop_transforms(zio); 1589dcbf3bd6SGeorge Wilson return (valid_cksum); 1590fa94a07fSbrendan } 1591fa94a07fSbrendan 15925602294fSDan Kimmel /* 15935602294fSDan Kimmel * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a 15945602294fSDan Kimmel * checksum and attaches it to the buf's hdr so that we can ensure that the buf 15955602294fSDan Kimmel * isn't modified later on. If buf is compressed or there is already a checksum 15965602294fSDan Kimmel * on the hdr, this is a no-op (we only checksum uncompressed bufs). 15975602294fSDan Kimmel */ 15986b4acc8bSahrens static void 1599dcbf3bd6SGeorge Wilson arc_cksum_compute(arc_buf_t *buf) 16006b4acc8bSahrens { 1601dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 1602dcbf3bd6SGeorge Wilson 1603dcbf3bd6SGeorge Wilson if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 16046b4acc8bSahrens return; 16056b4acc8bSahrens 1606dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 16075602294fSDan Kimmel 160889c86e32SChris Williamson mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1609dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 16105602294fSDan Kimmel ASSERT(arc_hdr_has_uncompressed_buf(hdr)); 16115602294fSDan Kimmel mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 16125602294fSDan Kimmel return; 16135602294fSDan Kimmel } else if (ARC_BUF_COMPRESSED(buf)) { 1614dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 16156b4acc8bSahrens return; 16166b4acc8bSahrens } 16175602294fSDan Kimmel 16185602294fSDan Kimmel ASSERT(!ARC_BUF_COMPRESSED(buf)); 1619dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), 1620dcbf3bd6SGeorge Wilson KM_SLEEP); 16215602294fSDan Kimmel fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, 1622dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_freeze_cksum); 1623dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1624cd1c8b85SMatthew Ahrens arc_buf_watch(buf); 1625cd1c8b85SMatthew Ahrens } 1626cd1c8b85SMatthew Ahrens 1627cd1c8b85SMatthew Ahrens #ifndef _KERNEL 1628cd1c8b85SMatthew Ahrens typedef struct procctl { 1629cd1c8b85SMatthew Ahrens long cmd; 1630cd1c8b85SMatthew Ahrens prwatch_t prwatch; 1631cd1c8b85SMatthew Ahrens } procctl_t; 1632cd1c8b85SMatthew Ahrens #endif 1633cd1c8b85SMatthew Ahrens 1634cd1c8b85SMatthew Ahrens /* ARGSUSED */ 1635cd1c8b85SMatthew Ahrens static void 1636cd1c8b85SMatthew Ahrens arc_buf_unwatch(arc_buf_t *buf) 1637cd1c8b85SMatthew Ahrens { 1638cd1c8b85SMatthew Ahrens #ifndef _KERNEL 1639cd1c8b85SMatthew Ahrens if (arc_watch) { 1640cd1c8b85SMatthew Ahrens int result; 1641cd1c8b85SMatthew Ahrens procctl_t ctl; 1642cd1c8b85SMatthew Ahrens ctl.cmd = PCWATCH; 1643cd1c8b85SMatthew Ahrens ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1644cd1c8b85SMatthew Ahrens ctl.prwatch.pr_size = 0; 1645cd1c8b85SMatthew Ahrens ctl.prwatch.pr_wflags = 0; 1646cd1c8b85SMatthew Ahrens result = write(arc_procfd, &ctl, sizeof (ctl)); 1647cd1c8b85SMatthew Ahrens ASSERT3U(result, ==, sizeof (ctl)); 1648cd1c8b85SMatthew Ahrens } 1649cd1c8b85SMatthew Ahrens #endif 1650cd1c8b85SMatthew Ahrens } 1651cd1c8b85SMatthew Ahrens 1652cd1c8b85SMatthew Ahrens /* ARGSUSED */ 1653cd1c8b85SMatthew Ahrens static void 1654cd1c8b85SMatthew Ahrens arc_buf_watch(arc_buf_t *buf) 1655cd1c8b85SMatthew Ahrens { 1656cd1c8b85SMatthew Ahrens #ifndef _KERNEL 1657cd1c8b85SMatthew Ahrens if (arc_watch) { 1658cd1c8b85SMatthew Ahrens int result; 1659cd1c8b85SMatthew Ahrens procctl_t ctl; 1660cd1c8b85SMatthew Ahrens ctl.cmd = PCWATCH; 1661cd1c8b85SMatthew Ahrens ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 16625602294fSDan Kimmel ctl.prwatch.pr_size = arc_buf_size(buf); 1663cd1c8b85SMatthew Ahrens ctl.prwatch.pr_wflags = WA_WRITE; 1664cd1c8b85SMatthew Ahrens result = write(arc_procfd, &ctl, sizeof (ctl)); 1665cd1c8b85SMatthew Ahrens ASSERT3U(result, ==, sizeof (ctl)); 1666cd1c8b85SMatthew Ahrens } 1667cd1c8b85SMatthew Ahrens #endif 16686b4acc8bSahrens } 16696b4acc8bSahrens 167089c86e32SChris Williamson static arc_buf_contents_t 167189c86e32SChris Williamson arc_buf_type(arc_buf_hdr_t *hdr) 167289c86e32SChris Williamson { 1673dcbf3bd6SGeorge Wilson arc_buf_contents_t type; 167489c86e32SChris Williamson if (HDR_ISTYPE_METADATA(hdr)) { 1675dcbf3bd6SGeorge Wilson type = ARC_BUFC_METADATA; 167689c86e32SChris Williamson } else { 1677dcbf3bd6SGeorge Wilson type = ARC_BUFC_DATA; 167889c86e32SChris Williamson } 1679dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 1680dcbf3bd6SGeorge Wilson return (type); 168189c86e32SChris Williamson } 168289c86e32SChris Williamson 16835602294fSDan Kimmel boolean_t 16845602294fSDan Kimmel arc_is_metadata(arc_buf_t *buf) 16855602294fSDan Kimmel { 16865602294fSDan Kimmel return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0); 16875602294fSDan Kimmel } 16885602294fSDan Kimmel 168989c86e32SChris Williamson static uint32_t 169089c86e32SChris Williamson arc_bufc_to_flags(arc_buf_contents_t type) 169189c86e32SChris Williamson { 169289c86e32SChris Williamson switch (type) { 169389c86e32SChris Williamson case ARC_BUFC_DATA: 169489c86e32SChris Williamson /* metadata field is 0 if buffer contains normal data */ 169589c86e32SChris Williamson return (0); 169689c86e32SChris Williamson case ARC_BUFC_METADATA: 169789c86e32SChris Williamson return (ARC_FLAG_BUFC_METADATA); 169889c86e32SChris Williamson default: 169989c86e32SChris Williamson break; 170089c86e32SChris Williamson } 170189c86e32SChris Williamson panic("undefined ARC buffer type!"); 170289c86e32SChris Williamson return ((uint32_t)-1); 170389c86e32SChris Williamson } 170489c86e32SChris Williamson 17056b4acc8bSahrens void 17066b4acc8bSahrens arc_buf_thaw(arc_buf_t *buf) 17076b4acc8bSahrens { 1708dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 1709dcbf3bd6SGeorge Wilson 17105602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 17115602294fSDan Kimmel ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 17125602294fSDan Kimmel 17135602294fSDan Kimmel arc_cksum_verify(buf); 17145602294fSDan Kimmel 17155602294fSDan Kimmel /* 17165602294fSDan Kimmel * Compressed buffers do not manipulate the b_freeze_cksum or 17175602294fSDan Kimmel * allocate b_thawed. 17185602294fSDan Kimmel */ 17195602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 17205602294fSDan Kimmel ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 17215602294fSDan Kimmel arc_hdr_has_uncompressed_buf(hdr)); 17225602294fSDan Kimmel return; 1723fa94a07fSbrendan } 17246b4acc8bSahrens 1725dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1726dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 17273f9d6ad7SLin Ling 1728dcbf3bd6SGeorge Wilson mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 172989c86e32SChris Williamson #ifdef ZFS_DEBUG 17303f9d6ad7SLin Ling if (zfs_flags & ZFS_DEBUG_MODIFY) { 1731dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_thawed != NULL) 1732dcbf3bd6SGeorge Wilson kmem_free(hdr->b_l1hdr.b_thawed, 1); 1733dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 17343f9d6ad7SLin Ling } 173589c86e32SChris Williamson #endif 17363f9d6ad7SLin Ling 1737dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1738cd1c8b85SMatthew Ahrens 1739cd1c8b85SMatthew Ahrens arc_buf_unwatch(buf); 17406b4acc8bSahrens } 17416b4acc8bSahrens 17426b4acc8bSahrens void 17436b4acc8bSahrens arc_buf_freeze(arc_buf_t *buf) 17446b4acc8bSahrens { 1745dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 17463f9d6ad7SLin Ling kmutex_t *hash_lock; 17473f9d6ad7SLin Ling 1748cc60fd72Sahrens if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1749cc60fd72Sahrens return; 1750cc60fd72Sahrens 17515602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 17525602294fSDan Kimmel ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 17535602294fSDan Kimmel arc_hdr_has_uncompressed_buf(hdr)); 17545602294fSDan Kimmel return; 17555602294fSDan Kimmel } 17565602294fSDan Kimmel 1757dcbf3bd6SGeorge Wilson hash_lock = HDR_LOCK(hdr); 17583f9d6ad7SLin Ling mutex_enter(hash_lock); 17593f9d6ad7SLin Ling 1760dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1761dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL || 1762dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_state == arc_anon); 1763dcbf3bd6SGeorge Wilson arc_cksum_compute(buf); 17643f9d6ad7SLin Ling mutex_exit(hash_lock); 17656b4acc8bSahrens } 17666b4acc8bSahrens 1767dcbf3bd6SGeorge Wilson /* 1768dcbf3bd6SGeorge Wilson * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, 1769dcbf3bd6SGeorge Wilson * the following functions should be used to ensure that the flags are 1770dcbf3bd6SGeorge Wilson * updated in a thread-safe way. When manipulating the flags either 1771dcbf3bd6SGeorge Wilson * the hash_lock must be held or the hdr must be undiscoverable. This 1772dcbf3bd6SGeorge Wilson * ensures that we're not racing with any other threads when updating 1773dcbf3bd6SGeorge Wilson * the flags. 1774dcbf3bd6SGeorge Wilson */ 1775dcbf3bd6SGeorge Wilson static inline void 1776dcbf3bd6SGeorge Wilson arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1777dcbf3bd6SGeorge Wilson { 1778dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1779dcbf3bd6SGeorge Wilson hdr->b_flags |= flags; 1780dcbf3bd6SGeorge Wilson } 1781dcbf3bd6SGeorge Wilson 1782dcbf3bd6SGeorge Wilson static inline void 1783dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1784dcbf3bd6SGeorge Wilson { 1785dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1786dcbf3bd6SGeorge Wilson hdr->b_flags &= ~flags; 1787dcbf3bd6SGeorge Wilson } 1788dcbf3bd6SGeorge Wilson 1789dcbf3bd6SGeorge Wilson /* 1790dcbf3bd6SGeorge Wilson * Setting the compression bits in the arc_buf_hdr_t's b_flags is 1791dcbf3bd6SGeorge Wilson * done in a special way since we have to clear and set bits 1792dcbf3bd6SGeorge Wilson * at the same time. Consumers that wish to set the compression bits 1793dcbf3bd6SGeorge Wilson * must use this function to ensure that the flags are updated in 1794dcbf3bd6SGeorge Wilson * thread-safe manner. 1795dcbf3bd6SGeorge Wilson */ 1796fa9e4066Sahrens static void 1797dcbf3bd6SGeorge Wilson arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) 1798fa9e4066Sahrens { 1799dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1800fa9e4066Sahrens 1801dcbf3bd6SGeorge Wilson /* 1802dcbf3bd6SGeorge Wilson * Holes and embedded blocks will always have a psize = 0 so 1803dcbf3bd6SGeorge Wilson * we ignore the compression of the blkptr and set the 1804dcbf3bd6SGeorge Wilson * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF. 1805dcbf3bd6SGeorge Wilson * Holes and embedded blocks remain anonymous so we don't 1806dcbf3bd6SGeorge Wilson * want to uncompress them. Mark them as uncompressed. 1807dcbf3bd6SGeorge Wilson */ 1808dcbf3bd6SGeorge Wilson if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { 1809dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1810dcbf3bd6SGeorge Wilson HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 1811dcbf3bd6SGeorge Wilson ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); 1812dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1813dcbf3bd6SGeorge Wilson } else { 1814dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1815dcbf3bd6SGeorge Wilson HDR_SET_COMPRESS(hdr, cmp); 1816dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); 1817dcbf3bd6SGeorge Wilson ASSERT(HDR_COMPRESSION_ENABLED(hdr)); 1818dcbf3bd6SGeorge Wilson } 1819dcbf3bd6SGeorge Wilson } 1820244781f1SPrakash Surya 18215602294fSDan Kimmel /* 18225602294fSDan Kimmel * Looks for another buf on the same hdr which has the data decompressed, copies 18235602294fSDan Kimmel * from it, and returns true. If no such buf exists, returns false. 18245602294fSDan Kimmel */ 18255602294fSDan Kimmel static boolean_t 18265602294fSDan Kimmel arc_buf_try_copy_decompressed_data(arc_buf_t *buf) 18275602294fSDan Kimmel { 18285602294fSDan Kimmel arc_buf_hdr_t *hdr = buf->b_hdr; 18295602294fSDan Kimmel boolean_t copied = B_FALSE; 18305602294fSDan Kimmel 18315602294fSDan Kimmel ASSERT(HDR_HAS_L1HDR(hdr)); 18325602294fSDan Kimmel ASSERT3P(buf->b_data, !=, NULL); 18335602294fSDan Kimmel ASSERT(!ARC_BUF_COMPRESSED(buf)); 18345602294fSDan Kimmel 18355602294fSDan Kimmel for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL; 18365602294fSDan Kimmel from = from->b_next) { 18375602294fSDan Kimmel /* can't use our own data buffer */ 18385602294fSDan Kimmel if (from == buf) { 18395602294fSDan Kimmel continue; 18405602294fSDan Kimmel } 18415602294fSDan Kimmel 18425602294fSDan Kimmel if (!ARC_BUF_COMPRESSED(from)) { 18435602294fSDan Kimmel bcopy(from->b_data, buf->b_data, arc_buf_size(buf)); 18445602294fSDan Kimmel copied = B_TRUE; 18455602294fSDan Kimmel break; 18465602294fSDan Kimmel } 18475602294fSDan Kimmel } 18485602294fSDan Kimmel 18495602294fSDan Kimmel /* 18505602294fSDan Kimmel * There were no decompressed bufs, so there should not be a 18515602294fSDan Kimmel * checksum on the hdr either. 18525602294fSDan Kimmel */ 18535602294fSDan Kimmel EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL); 18545602294fSDan Kimmel 18555602294fSDan Kimmel return (copied); 18565602294fSDan Kimmel } 18575602294fSDan Kimmel 18585602294fSDan Kimmel /* 18595602294fSDan Kimmel * Given a buf that has a data buffer attached to it, this function will 18605602294fSDan Kimmel * efficiently fill the buf with data of the specified compression setting from 18615602294fSDan Kimmel * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr 18625602294fSDan Kimmel * are already sharing a data buf, no copy is performed. 18635602294fSDan Kimmel * 18645602294fSDan Kimmel * If the buf is marked as compressed but uncompressed data was requested, this 18655602294fSDan Kimmel * will allocate a new data buffer for the buf, remove that flag, and fill the 18665602294fSDan Kimmel * buf with uncompressed data. You can't request a compressed buf on a hdr with 18675602294fSDan Kimmel * uncompressed data, and (since we haven't added support for it yet) if you 18685602294fSDan Kimmel * want compressed data your buf must already be marked as compressed and have 18695602294fSDan Kimmel * the correct-sized data buffer. 18705602294fSDan Kimmel */ 1871dcbf3bd6SGeorge Wilson static int 18725602294fSDan Kimmel arc_buf_fill(arc_buf_t *buf, boolean_t compressed) 1873dcbf3bd6SGeorge Wilson { 1874dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 18755602294fSDan Kimmel boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 1876dcbf3bd6SGeorge Wilson dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; 187789c86e32SChris Williamson 18785602294fSDan Kimmel ASSERT3P(buf->b_data, !=, NULL); 18795602294fSDan Kimmel IMPLY(compressed, hdr_compressed); 18805602294fSDan Kimmel IMPLY(compressed, ARC_BUF_COMPRESSED(buf)); 18815602294fSDan Kimmel 18825602294fSDan Kimmel if (hdr_compressed == compressed) { 18835602294fSDan Kimmel if (!arc_buf_is_shared(buf)) { 1884*770499e1SDan Kimmel abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd, 18855602294fSDan Kimmel arc_buf_size(buf)); 18865602294fSDan Kimmel } 1887dcbf3bd6SGeorge Wilson } else { 18885602294fSDan Kimmel ASSERT(hdr_compressed); 18895602294fSDan Kimmel ASSERT(!compressed); 1890dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); 18915602294fSDan Kimmel 18925602294fSDan Kimmel /* 18935602294fSDan Kimmel * If the buf is sharing its data with the hdr, unlink it and 18945602294fSDan Kimmel * allocate a new data buffer for the buf. 18955602294fSDan Kimmel */ 18965602294fSDan Kimmel if (arc_buf_is_shared(buf)) { 18975602294fSDan Kimmel ASSERT(ARC_BUF_COMPRESSED(buf)); 18985602294fSDan Kimmel 18995602294fSDan Kimmel /* We need to give the buf it's own b_data */ 19005602294fSDan Kimmel buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 19015602294fSDan Kimmel buf->b_data = 19025602294fSDan Kimmel arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 19035602294fSDan Kimmel arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 19045602294fSDan Kimmel 19055602294fSDan Kimmel /* Previously overhead was 0; just add new overhead */ 19065602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 19075602294fSDan Kimmel } else if (ARC_BUF_COMPRESSED(buf)) { 19085602294fSDan Kimmel /* We need to reallocate the buf's b_data */ 19095602294fSDan Kimmel arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr), 19105602294fSDan Kimmel buf); 19115602294fSDan Kimmel buf->b_data = 19125602294fSDan Kimmel arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 19135602294fSDan Kimmel 19145602294fSDan Kimmel /* We increased the size of b_data; update overhead */ 19155602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, 19165602294fSDan Kimmel HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr)); 19175602294fSDan Kimmel } 19185602294fSDan Kimmel 19195602294fSDan Kimmel /* 19205602294fSDan Kimmel * Regardless of the buf's previous compression settings, it 19215602294fSDan Kimmel * should not be compressed at the end of this function. 19225602294fSDan Kimmel */ 19235602294fSDan Kimmel buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; 19245602294fSDan Kimmel 19255602294fSDan Kimmel /* 19265602294fSDan Kimmel * Try copying the data from another buf which already has a 19275602294fSDan Kimmel * decompressed version. If that's not possible, it's time to 19285602294fSDan Kimmel * bite the bullet and decompress the data from the hdr. 19295602294fSDan Kimmel */ 19305602294fSDan Kimmel if (arc_buf_try_copy_decompressed_data(buf)) { 19315602294fSDan Kimmel /* Skip byteswapping and checksumming (already done) */ 19325602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL); 19335602294fSDan Kimmel return (0); 19345602294fSDan Kimmel } else { 19355602294fSDan Kimmel int error = zio_decompress_data(HDR_GET_COMPRESS(hdr), 1936*770499e1SDan Kimmel hdr->b_l1hdr.b_pabd, buf->b_data, 19375602294fSDan Kimmel HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 19385602294fSDan Kimmel 19395602294fSDan Kimmel /* 19405602294fSDan Kimmel * Absent hardware errors or software bugs, this should 19415602294fSDan Kimmel * be impossible, but log it anyway so we can debug it. 19425602294fSDan Kimmel */ 19435602294fSDan Kimmel if (error != 0) { 19445602294fSDan Kimmel zfs_dbgmsg( 19455602294fSDan Kimmel "hdr %p, compress %d, psize %d, lsize %d", 19465602294fSDan Kimmel hdr, HDR_GET_COMPRESS(hdr), 19475602294fSDan Kimmel HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 19485602294fSDan Kimmel return (SET_ERROR(EIO)); 19495602294fSDan Kimmel } 1950ea8dc4b6Seschrock } 1951fa9e4066Sahrens } 19525602294fSDan Kimmel 19535602294fSDan Kimmel /* Byteswap the buf's data if necessary */ 1954dcbf3bd6SGeorge Wilson if (bswap != DMU_BSWAP_NUMFUNCS) { 1955dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(hdr)); 1956dcbf3bd6SGeorge Wilson ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); 1957dcbf3bd6SGeorge Wilson dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); 1958dcbf3bd6SGeorge Wilson } 19595602294fSDan Kimmel 19605602294fSDan Kimmel /* Compute the hdr's checksum if necessary */ 1961dcbf3bd6SGeorge Wilson arc_cksum_compute(buf); 19625602294fSDan Kimmel 1963dcbf3bd6SGeorge Wilson return (0); 1964fa9e4066Sahrens } 1965fa9e4066Sahrens 19665602294fSDan Kimmel int 19675602294fSDan Kimmel arc_decompress(arc_buf_t *buf) 19685602294fSDan Kimmel { 19695602294fSDan Kimmel return (arc_buf_fill(buf, B_FALSE)); 19705602294fSDan Kimmel } 19715602294fSDan Kimmel 1972dcbf3bd6SGeorge Wilson /* 1973*770499e1SDan Kimmel * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t. 1974dcbf3bd6SGeorge Wilson */ 1975dcbf3bd6SGeorge Wilson static uint64_t 1976dcbf3bd6SGeorge Wilson arc_hdr_size(arc_buf_hdr_t *hdr) 1977fa9e4066Sahrens { 1978dcbf3bd6SGeorge Wilson uint64_t size; 1979fa9e4066Sahrens 1980dcbf3bd6SGeorge Wilson if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 1981dcbf3bd6SGeorge Wilson HDR_GET_PSIZE(hdr) > 0) { 1982dcbf3bd6SGeorge Wilson size = HDR_GET_PSIZE(hdr); 1983dcbf3bd6SGeorge Wilson } else { 1984dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); 1985dcbf3bd6SGeorge Wilson size = HDR_GET_LSIZE(hdr); 1986dcbf3bd6SGeorge Wilson } 1987dcbf3bd6SGeorge Wilson return (size); 1988dcbf3bd6SGeorge Wilson } 1989fa9e4066Sahrens 1990dcbf3bd6SGeorge Wilson /* 1991dcbf3bd6SGeorge Wilson * Increment the amount of evictable space in the arc_state_t's refcount. 1992dcbf3bd6SGeorge Wilson * We account for the space used by the hdr and the arc buf individually 1993dcbf3bd6SGeorge Wilson * so that we can add and remove them from the refcount individually. 1994dcbf3bd6SGeorge Wilson */ 1995dcbf3bd6SGeorge Wilson static void 1996dcbf3bd6SGeorge Wilson arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) 1997dcbf3bd6SGeorge Wilson { 1998dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 1999244781f1SPrakash Surya 2000dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 20010e8c6158Smaybee 2002dcbf3bd6SGeorge Wilson if (GHOST_STATE(state)) { 2003dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 2004dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2005*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 20065602294fSDan Kimmel (void) refcount_add_many(&state->arcs_esize[type], 20075602294fSDan Kimmel HDR_GET_LSIZE(hdr), hdr); 2008dcbf3bd6SGeorge Wilson return; 2009dcbf3bd6SGeorge Wilson } 2010dcbf3bd6SGeorge Wilson 2011dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(state)); 2012*770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 2013dcbf3bd6SGeorge Wilson (void) refcount_add_many(&state->arcs_esize[type], 2014dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2015dcbf3bd6SGeorge Wilson } 2016dcbf3bd6SGeorge Wilson for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2017dcbf3bd6SGeorge Wilson buf = buf->b_next) { 20185602294fSDan Kimmel if (arc_buf_is_shared(buf)) 2019dcbf3bd6SGeorge Wilson continue; 20205602294fSDan Kimmel (void) refcount_add_many(&state->arcs_esize[type], 20215602294fSDan Kimmel arc_buf_size(buf), buf); 2022fa9e4066Sahrens } 2023fa9e4066Sahrens } 2024fa9e4066Sahrens 2025fa9e4066Sahrens /* 2026dcbf3bd6SGeorge Wilson * Decrement the amount of evictable space in the arc_state_t's refcount. 2027dcbf3bd6SGeorge Wilson * We account for the space used by the hdr and the arc buf individually 2028dcbf3bd6SGeorge Wilson * so that we can add and remove them from the refcount individually. 2029fa9e4066Sahrens */ 2030fa9e4066Sahrens static void 20315602294fSDan Kimmel arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) 2032fa9e4066Sahrens { 2033dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 2034dcbf3bd6SGeorge Wilson 2035dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2036dcbf3bd6SGeorge Wilson 2037dcbf3bd6SGeorge Wilson if (GHOST_STATE(state)) { 2038dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 2039dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2040*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2041dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 20425602294fSDan Kimmel HDR_GET_LSIZE(hdr), hdr); 2043dcbf3bd6SGeorge Wilson return; 2044dcbf3bd6SGeorge Wilson } 2045dcbf3bd6SGeorge Wilson 2046dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(state)); 2047*770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 2048dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 2049dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2050dcbf3bd6SGeorge Wilson } 2051dcbf3bd6SGeorge Wilson for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2052dcbf3bd6SGeorge Wilson buf = buf->b_next) { 20535602294fSDan Kimmel if (arc_buf_is_shared(buf)) 2054dcbf3bd6SGeorge Wilson continue; 2055dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 20565602294fSDan Kimmel arc_buf_size(buf), buf); 2057dcbf3bd6SGeorge Wilson } 2058dcbf3bd6SGeorge Wilson } 2059dcbf3bd6SGeorge Wilson 2060dcbf3bd6SGeorge Wilson /* 2061dcbf3bd6SGeorge Wilson * Add a reference to this hdr indicating that someone is actively 2062dcbf3bd6SGeorge Wilson * referencing that memory. When the refcount transitions from 0 to 1, 2063dcbf3bd6SGeorge Wilson * we remove it from the respective arc_state_t list to indicate that 2064dcbf3bd6SGeorge Wilson * it is not evictable. 2065dcbf3bd6SGeorge Wilson */ 2066dcbf3bd6SGeorge Wilson static void 2067dcbf3bd6SGeorge Wilson add_reference(arc_buf_hdr_t *hdr, void *tag) 2068dcbf3bd6SGeorge Wilson { 2069dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2070dcbf3bd6SGeorge Wilson if (!MUTEX_HELD(HDR_LOCK(hdr))) { 2071dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_state == arc_anon); 2072dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2073dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2074dcbf3bd6SGeorge Wilson } 2075dcbf3bd6SGeorge Wilson 2076dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2077dcbf3bd6SGeorge Wilson 2078dcbf3bd6SGeorge Wilson if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 2079dcbf3bd6SGeorge Wilson (state != arc_anon)) { 2080dcbf3bd6SGeorge Wilson /* We don't use the L2-only state list. */ 2081dcbf3bd6SGeorge Wilson if (state != arc_l2c_only) { 208294c2d0ebSMatthew Ahrens multilist_remove(state->arcs_list[arc_buf_type(hdr)], 2083dcbf3bd6SGeorge Wilson hdr); 20845602294fSDan Kimmel arc_evictable_space_decrement(hdr, state); 2085dcbf3bd6SGeorge Wilson } 2086dcbf3bd6SGeorge Wilson /* remove the prefetch flag if we get a reference */ 2087dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 2088dcbf3bd6SGeorge Wilson } 2089dcbf3bd6SGeorge Wilson } 2090dcbf3bd6SGeorge Wilson 2091dcbf3bd6SGeorge Wilson /* 2092dcbf3bd6SGeorge Wilson * Remove a reference from this hdr. When the reference transitions from 2093dcbf3bd6SGeorge Wilson * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's 2094dcbf3bd6SGeorge Wilson * list making it eligible for eviction. 2095dcbf3bd6SGeorge Wilson */ 2096dcbf3bd6SGeorge Wilson static int 2097dcbf3bd6SGeorge Wilson remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 2098dcbf3bd6SGeorge Wilson { 2099dcbf3bd6SGeorge Wilson int cnt; 2100dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2101dcbf3bd6SGeorge Wilson 2102dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2103dcbf3bd6SGeorge Wilson ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 2104dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(state)); 2105dcbf3bd6SGeorge Wilson 2106dcbf3bd6SGeorge Wilson /* 2107dcbf3bd6SGeorge Wilson * arc_l2c_only counts as a ghost state so we don't need to explicitly 2108dcbf3bd6SGeorge Wilson * check to prevent usage of the arc_l2c_only list. 2109dcbf3bd6SGeorge Wilson */ 2110dcbf3bd6SGeorge Wilson if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 2111dcbf3bd6SGeorge Wilson (state != arc_anon)) { 211294c2d0ebSMatthew Ahrens multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr); 2113dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 2114dcbf3bd6SGeorge Wilson arc_evictable_space_increment(hdr, state); 2115dcbf3bd6SGeorge Wilson } 2116dcbf3bd6SGeorge Wilson return (cnt); 2117dcbf3bd6SGeorge Wilson } 2118dcbf3bd6SGeorge Wilson 2119dcbf3bd6SGeorge Wilson /* 2120dcbf3bd6SGeorge Wilson * Move the supplied buffer to the indicated state. The hash lock 2121dcbf3bd6SGeorge Wilson * for the buffer must be held by the caller. 2122dcbf3bd6SGeorge Wilson */ 2123dcbf3bd6SGeorge Wilson static void 2124dcbf3bd6SGeorge Wilson arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 2125dcbf3bd6SGeorge Wilson kmutex_t *hash_lock) 2126dcbf3bd6SGeorge Wilson { 2127dcbf3bd6SGeorge Wilson arc_state_t *old_state; 2128dcbf3bd6SGeorge Wilson int64_t refcnt; 2129dcbf3bd6SGeorge Wilson uint32_t bufcnt; 2130dcbf3bd6SGeorge Wilson boolean_t update_old, update_new; 2131dcbf3bd6SGeorge Wilson arc_buf_contents_t buftype = arc_buf_type(hdr); 213289c86e32SChris Williamson 213389c86e32SChris Williamson /* 213489c86e32SChris Williamson * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 213589c86e32SChris Williamson * in arc_read() when bringing a buffer out of the L2ARC. However, the 213689c86e32SChris Williamson * L1 hdr doesn't always exist when we change state to arc_anon before 213789c86e32SChris Williamson * destroying a header, in which case reallocating to add the L1 hdr is 213889c86e32SChris Williamson * pointless. 213989c86e32SChris Williamson */ 214089c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) { 214189c86e32SChris Williamson old_state = hdr->b_l1hdr.b_state; 214289c86e32SChris Williamson refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 2143dcbf3bd6SGeorge Wilson bufcnt = hdr->b_l1hdr.b_bufcnt; 2144*770499e1SDan Kimmel update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL); 214589c86e32SChris Williamson } else { 214689c86e32SChris Williamson old_state = arc_l2c_only; 214789c86e32SChris Williamson refcnt = 0; 2148dcbf3bd6SGeorge Wilson bufcnt = 0; 2149dcbf3bd6SGeorge Wilson update_old = B_FALSE; 215089c86e32SChris Williamson } 2151dcbf3bd6SGeorge Wilson update_new = update_old; 2152fa9e4066Sahrens 2153fa9e4066Sahrens ASSERT(MUTEX_HELD(hash_lock)); 215469962b56SMatthew Ahrens ASSERT3P(new_state, !=, old_state); 2155dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); 2156dcbf3bd6SGeorge Wilson ASSERT(old_state != arc_anon || bufcnt <= 1); 2157fa9e4066Sahrens 2158fa9e4066Sahrens /* 2159fa9e4066Sahrens * If this buffer is evictable, transfer it from the 2160fa9e4066Sahrens * old state list to the new state list. 2161fa9e4066Sahrens */ 2162ea8dc4b6Seschrock if (refcnt == 0) { 216389c86e32SChris Williamson if (old_state != arc_anon && old_state != arc_l2c_only) { 216489c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 216594c2d0ebSMatthew Ahrens multilist_remove(old_state->arcs_list[buftype], hdr); 2166ea8dc4b6Seschrock 2167dcbf3bd6SGeorge Wilson if (GHOST_STATE(old_state)) { 2168dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 2169dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2170dcbf3bd6SGeorge Wilson update_old = B_TRUE; 2171ea8dc4b6Seschrock } 21725602294fSDan Kimmel arc_evictable_space_decrement(hdr, old_state); 2173fa9e4066Sahrens } 217489c86e32SChris Williamson if (new_state != arc_anon && new_state != arc_l2c_only) { 2175fa9e4066Sahrens 217689c86e32SChris Williamson /* 217789c86e32SChris Williamson * An L1 header always exists here, since if we're 217889c86e32SChris Williamson * moving to some L1-cached state (i.e. not l2c_only or 217989c86e32SChris Williamson * anonymous), we realloc the header to add an L1hdr 218089c86e32SChris Williamson * beforehand. 218189c86e32SChris Williamson */ 218289c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 218394c2d0ebSMatthew Ahrens multilist_insert(new_state->arcs_list[buftype], hdr); 2184ea8dc4b6Seschrock 2185ea8dc4b6Seschrock if (GHOST_STATE(new_state)) { 2186dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 2187dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2188dcbf3bd6SGeorge Wilson update_new = B_TRUE; 2189ea8dc4b6Seschrock } 2190dcbf3bd6SGeorge Wilson arc_evictable_space_increment(hdr, new_state); 2191fa9e4066Sahrens } 2192fa9e4066Sahrens } 2193fa9e4066Sahrens 2194dcbf3bd6SGeorge Wilson ASSERT(!HDR_EMPTY(hdr)); 21957adb730bSGeorge Wilson if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 21967adb730bSGeorge Wilson buf_hash_remove(hdr); 2197fa9e4066Sahrens 219889c86e32SChris Williamson /* adjust state sizes (ignore arc_l2c_only) */ 21992fd872a7SPrakash Surya 2200dcbf3bd6SGeorge Wilson if (update_new && new_state != arc_l2c_only) { 22012fd872a7SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 22022fd872a7SPrakash Surya if (GHOST_STATE(new_state)) { 2203dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 22042fd872a7SPrakash Surya 22052fd872a7SPrakash Surya /* 2206dcbf3bd6SGeorge Wilson * When moving a header to a ghost state, we first 22072fd872a7SPrakash Surya * remove all arc buffers. Thus, we'll have a 2208dcbf3bd6SGeorge Wilson * bufcnt of zero, and no arc buffer to use for 22092fd872a7SPrakash Surya * the reference. As a result, we use the arc 22102fd872a7SPrakash Surya * header pointer for the reference. 22112fd872a7SPrakash Surya */ 22122fd872a7SPrakash Surya (void) refcount_add_many(&new_state->arcs_size, 2213dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr), hdr); 2214*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 22152fd872a7SPrakash Surya } else { 2216dcbf3bd6SGeorge Wilson uint32_t buffers = 0; 22172fd872a7SPrakash Surya 22182fd872a7SPrakash Surya /* 22192fd872a7SPrakash Surya * Each individual buffer holds a unique reference, 22202fd872a7SPrakash Surya * thus we must remove each of these references one 22212fd872a7SPrakash Surya * at a time. 22222fd872a7SPrakash Surya */ 22232fd872a7SPrakash Surya for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 22242fd872a7SPrakash Surya buf = buf->b_next) { 2225dcbf3bd6SGeorge Wilson ASSERT3U(bufcnt, !=, 0); 2226dcbf3bd6SGeorge Wilson buffers++; 2227dcbf3bd6SGeorge Wilson 2228dcbf3bd6SGeorge Wilson /* 2229dcbf3bd6SGeorge Wilson * When the arc_buf_t is sharing the data 2230dcbf3bd6SGeorge Wilson * block with the hdr, the owner of the 2231dcbf3bd6SGeorge Wilson * reference belongs to the hdr. Only 2232dcbf3bd6SGeorge Wilson * add to the refcount if the arc_buf_t is 2233dcbf3bd6SGeorge Wilson * not shared. 2234dcbf3bd6SGeorge Wilson */ 22355602294fSDan Kimmel if (arc_buf_is_shared(buf)) 2236dcbf3bd6SGeorge Wilson continue; 2237dcbf3bd6SGeorge Wilson 22382fd872a7SPrakash Surya (void) refcount_add_many(&new_state->arcs_size, 22395602294fSDan Kimmel arc_buf_size(buf), buf); 2240dcbf3bd6SGeorge Wilson } 2241dcbf3bd6SGeorge Wilson ASSERT3U(bufcnt, ==, buffers); 2242dcbf3bd6SGeorge Wilson 2243*770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 2244dcbf3bd6SGeorge Wilson (void) refcount_add_many(&new_state->arcs_size, 2245dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2246dcbf3bd6SGeorge Wilson } else { 2247dcbf3bd6SGeorge Wilson ASSERT(GHOST_STATE(old_state)); 22482fd872a7SPrakash Surya } 22492fd872a7SPrakash Surya } 22502fd872a7SPrakash Surya } 22512fd872a7SPrakash Surya 2252dcbf3bd6SGeorge Wilson if (update_old && old_state != arc_l2c_only) { 22532fd872a7SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 22542fd872a7SPrakash Surya if (GHOST_STATE(old_state)) { 2255dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 2256*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2257dcbf3bd6SGeorge Wilson 22582fd872a7SPrakash Surya /* 22592fd872a7SPrakash Surya * When moving a header off of a ghost state, 2260dcbf3bd6SGeorge Wilson * the header will not contain any arc buffers. 2261dcbf3bd6SGeorge Wilson * We use the arc header pointer for the reference 2262dcbf3bd6SGeorge Wilson * which is exactly what we did when we put the 2263dcbf3bd6SGeorge Wilson * header on the ghost state. 22642fd872a7SPrakash Surya */ 22652fd872a7SPrakash Surya 22662fd872a7SPrakash Surya (void) refcount_remove_many(&old_state->arcs_size, 2267dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr), hdr); 22682fd872a7SPrakash Surya } else { 2269dcbf3bd6SGeorge Wilson uint32_t buffers = 0; 22702fd872a7SPrakash Surya 22712fd872a7SPrakash Surya /* 22722fd872a7SPrakash Surya * Each individual buffer holds a unique reference, 22732fd872a7SPrakash Surya * thus we must remove each of these references one 22742fd872a7SPrakash Surya * at a time. 22752fd872a7SPrakash Surya */ 22762fd872a7SPrakash Surya for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 22772fd872a7SPrakash Surya buf = buf->b_next) { 22785602294fSDan Kimmel ASSERT3U(bufcnt, !=, 0); 2279dcbf3bd6SGeorge Wilson buffers++; 2280dcbf3bd6SGeorge Wilson 2281dcbf3bd6SGeorge Wilson /* 2282dcbf3bd6SGeorge Wilson * When the arc_buf_t is sharing the data 2283dcbf3bd6SGeorge Wilson * block with the hdr, the owner of the 2284dcbf3bd6SGeorge Wilson * reference belongs to the hdr. Only 2285dcbf3bd6SGeorge Wilson * add to the refcount if the arc_buf_t is 2286dcbf3bd6SGeorge Wilson * not shared. 2287dcbf3bd6SGeorge Wilson */ 22885602294fSDan Kimmel if (arc_buf_is_shared(buf)) 2289dcbf3bd6SGeorge Wilson continue; 2290dcbf3bd6SGeorge Wilson 22912fd872a7SPrakash Surya (void) refcount_remove_many( 22925602294fSDan Kimmel &old_state->arcs_size, arc_buf_size(buf), 2293dcbf3bd6SGeorge Wilson buf); 22942fd872a7SPrakash Surya } 2295dcbf3bd6SGeorge Wilson ASSERT3U(bufcnt, ==, buffers); 2296*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2297dcbf3bd6SGeorge Wilson (void) refcount_remove_many( 2298dcbf3bd6SGeorge Wilson &old_state->arcs_size, arc_hdr_size(hdr), hdr); 22992fd872a7SPrakash Surya } 2300fa9e4066Sahrens } 23012fd872a7SPrakash Surya 230289c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) 230389c86e32SChris Williamson hdr->b_l1hdr.b_state = new_state; 2304fa94a07fSbrendan 230589c86e32SChris Williamson /* 230689c86e32SChris Williamson * L2 headers should never be on the L2 state list since they don't 230789c86e32SChris Williamson * have L1 headers allocated. 230889c86e32SChris Williamson */ 230994c2d0ebSMatthew Ahrens ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 231094c2d0ebSMatthew Ahrens multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 2311fa9e4066Sahrens } 2312fa9e4066Sahrens 23130e8c6158Smaybee void 23145a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(uint64_t space, arc_space_type_t type) 23150e8c6158Smaybee { 23165a98e54bSBrendan Gregg - Sun Microsystems ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 23175a98e54bSBrendan Gregg - Sun Microsystems 23185a98e54bSBrendan Gregg - Sun Microsystems switch (type) { 23195a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_DATA: 23205a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_data_size, space); 23215a98e54bSBrendan Gregg - Sun Microsystems break; 23224076b1bfSPrakash Surya case ARC_SPACE_META: 23234076b1bfSPrakash Surya ARCSTAT_INCR(arcstat_metadata_size, space); 23244076b1bfSPrakash Surya break; 23255a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_OTHER: 23265a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_other_size, space); 23275a98e54bSBrendan Gregg - Sun Microsystems break; 23285a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_HDRS: 23295a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_hdr_size, space); 23305a98e54bSBrendan Gregg - Sun Microsystems break; 23315a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_L2HDRS: 23325a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_l2_hdr_size, space); 23335a98e54bSBrendan Gregg - Sun Microsystems break; 23345a98e54bSBrendan Gregg - Sun Microsystems } 23355a98e54bSBrendan Gregg - Sun Microsystems 23364076b1bfSPrakash Surya if (type != ARC_SPACE_DATA) 23374076b1bfSPrakash Surya ARCSTAT_INCR(arcstat_meta_used, space); 23384076b1bfSPrakash Surya 23390e8c6158Smaybee atomic_add_64(&arc_size, space); 23400e8c6158Smaybee } 23410e8c6158Smaybee 23420e8c6158Smaybee void 23435a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(uint64_t space, arc_space_type_t type) 23440e8c6158Smaybee { 23455a98e54bSBrendan Gregg - Sun Microsystems ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 23465a98e54bSBrendan Gregg - Sun Microsystems 23475a98e54bSBrendan Gregg - Sun Microsystems switch (type) { 23485a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_DATA: 23495a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_data_size, -space); 23505a98e54bSBrendan Gregg - Sun Microsystems break; 23514076b1bfSPrakash Surya case ARC_SPACE_META: 23524076b1bfSPrakash Surya ARCSTAT_INCR(arcstat_metadata_size, -space); 23534076b1bfSPrakash Surya break; 23545a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_OTHER: 23555a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_other_size, -space); 23565a98e54bSBrendan Gregg - Sun Microsystems break; 23575a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_HDRS: 23585a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_hdr_size, -space); 23595a98e54bSBrendan Gregg - Sun Microsystems break; 23605a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_L2HDRS: 23615a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 23625a98e54bSBrendan Gregg - Sun Microsystems break; 23635a98e54bSBrendan Gregg - Sun Microsystems } 23645a98e54bSBrendan Gregg - Sun Microsystems 23654076b1bfSPrakash Surya if (type != ARC_SPACE_DATA) { 23664076b1bfSPrakash Surya ASSERT(arc_meta_used >= space); 23674076b1bfSPrakash Surya if (arc_meta_max < arc_meta_used) 23684076b1bfSPrakash Surya arc_meta_max = arc_meta_used; 23694076b1bfSPrakash Surya ARCSTAT_INCR(arcstat_meta_used, -space); 23704076b1bfSPrakash Surya } 23714076b1bfSPrakash Surya 23720e8c6158Smaybee ASSERT(arc_size >= space); 23730e8c6158Smaybee atomic_add_64(&arc_size, -space); 23740e8c6158Smaybee } 23750e8c6158Smaybee 2376dcbf3bd6SGeorge Wilson /* 23775602294fSDan Kimmel * Given a hdr and a buf, returns whether that buf can share its b_data buffer 2378*770499e1SDan Kimmel * with the hdr's b_pabd. 2379dcbf3bd6SGeorge Wilson */ 23805602294fSDan Kimmel static boolean_t 23815602294fSDan Kimmel arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf) 23825602294fSDan Kimmel { 23835602294fSDan Kimmel /* 23845602294fSDan Kimmel * The criteria for sharing a hdr's data are: 23855602294fSDan Kimmel * 1. the hdr's compression matches the buf's compression 23865602294fSDan Kimmel * 2. the hdr doesn't need to be byteswapped 23875602294fSDan Kimmel * 3. the hdr isn't already being shared 23885602294fSDan Kimmel * 4. the buf is either compressed or it is the last buf in the hdr list 23895602294fSDan Kimmel * 23905602294fSDan Kimmel * Criterion #4 maintains the invariant that shared uncompressed 23915602294fSDan Kimmel * bufs must be the final buf in the hdr's b_buf list. Reading this, you 23925602294fSDan Kimmel * might ask, "if a compressed buf is allocated first, won't that be the 23935602294fSDan Kimmel * last thing in the list?", but in that case it's impossible to create 23945602294fSDan Kimmel * a shared uncompressed buf anyway (because the hdr must be compressed 23955602294fSDan Kimmel * to have the compressed buf). You might also think that #3 is 23965602294fSDan Kimmel * sufficient to make this guarantee, however it's possible 23975602294fSDan Kimmel * (specifically in the rare L2ARC write race mentioned in 23985602294fSDan Kimmel * arc_buf_alloc_impl()) there will be an existing uncompressed buf that 23995602294fSDan Kimmel * is sharable, but wasn't at the time of its allocation. Rather than 24005602294fSDan Kimmel * allow a new shared uncompressed buf to be created and then shuffle 24015602294fSDan Kimmel * the list around to make it the last element, this simply disallows 24025602294fSDan Kimmel * sharing if the new buf isn't the first to be added. 24035602294fSDan Kimmel */ 24045602294fSDan Kimmel ASSERT3P(buf->b_hdr, ==, hdr); 24055602294fSDan Kimmel boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF; 24065602294fSDan Kimmel boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0; 24075602294fSDan Kimmel return (buf_compressed == hdr_compressed && 24085602294fSDan Kimmel hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && 24095602294fSDan Kimmel !HDR_SHARED_DATA(hdr) && 24105602294fSDan Kimmel (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf))); 24115602294fSDan Kimmel } 24125602294fSDan Kimmel 24135602294fSDan Kimmel /* 24145602294fSDan Kimmel * Allocate a buf for this hdr. If you care about the data that's in the hdr, 24155602294fSDan Kimmel * or if you want a compressed buffer, pass those flags in. Returns 0 if the 24165602294fSDan Kimmel * copy was made successfully, or an error code otherwise. 24175602294fSDan Kimmel */ 24185602294fSDan Kimmel static int 24195602294fSDan Kimmel arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed, 24205602294fSDan Kimmel boolean_t fill, arc_buf_t **ret) 2421fa9e4066Sahrens { 2422fa9e4066Sahrens arc_buf_t *buf; 2423fa9e4066Sahrens 2424dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2425dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2426dcbf3bd6SGeorge Wilson VERIFY(hdr->b_type == ARC_BUFC_DATA || 2427dcbf3bd6SGeorge Wilson hdr->b_type == ARC_BUFC_METADATA); 24285602294fSDan Kimmel ASSERT3P(ret, !=, NULL); 24295602294fSDan Kimmel ASSERT3P(*ret, ==, NULL); 2430dcbf3bd6SGeorge Wilson 24315602294fSDan Kimmel buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2432fa9e4066Sahrens buf->b_hdr = hdr; 243344eda4d7Smaybee buf->b_data = NULL; 24345602294fSDan Kimmel buf->b_next = hdr->b_l1hdr.b_buf; 24355602294fSDan Kimmel buf->b_flags = 0; 243689c86e32SChris Williamson 2437dcbf3bd6SGeorge Wilson add_reference(hdr, tag); 2438dcbf3bd6SGeorge Wilson 2439dcbf3bd6SGeorge Wilson /* 2440dcbf3bd6SGeorge Wilson * We're about to change the hdr's b_flags. We must either 2441dcbf3bd6SGeorge Wilson * hold the hash_lock or be undiscoverable. 2442dcbf3bd6SGeorge Wilson */ 2443dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2444dcbf3bd6SGeorge Wilson 2445dcbf3bd6SGeorge Wilson /* 24465602294fSDan Kimmel * Only honor requests for compressed bufs if the hdr is actually 24475602294fSDan Kimmel * compressed. 24485602294fSDan Kimmel */ 24495602294fSDan Kimmel if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 24505602294fSDan Kimmel buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; 24515602294fSDan Kimmel 24525602294fSDan Kimmel /* 24535602294fSDan Kimmel * If the hdr's data can be shared then we share the data buffer and 24545602294fSDan Kimmel * set the appropriate bit in the hdr's b_flags to indicate the hdr is 2455*770499e1SDan Kimmel * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new 24565602294fSDan Kimmel * buffer to store the buf's data. 24575602294fSDan Kimmel * 2458*770499e1SDan Kimmel * There are two additional restrictions here because we're sharing 2459*770499e1SDan Kimmel * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be 2460*770499e1SDan Kimmel * actively involved in an L2ARC write, because if this buf is used by 2461*770499e1SDan Kimmel * an arc_write() then the hdr's data buffer will be released when the 24625602294fSDan Kimmel * write completes, even though the L2ARC write might still be using it. 2463*770499e1SDan Kimmel * Second, the hdr's ABD must be linear so that the buf's user doesn't 2464*770499e1SDan Kimmel * need to be ABD-aware. 2465dcbf3bd6SGeorge Wilson */ 2466*770499e1SDan Kimmel boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) && 2467*770499e1SDan Kimmel abd_is_linear(hdr->b_l1hdr.b_pabd); 24685602294fSDan Kimmel 24695602294fSDan Kimmel /* Set up b_data and sharing */ 24705602294fSDan Kimmel if (can_share) { 2471*770499e1SDan Kimmel buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd); 24725602294fSDan Kimmel buf->b_flags |= ARC_BUF_FLAG_SHARED; 2473dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2474dcbf3bd6SGeorge Wilson } else { 24755602294fSDan Kimmel buf->b_data = 24765602294fSDan Kimmel arc_get_data_buf(hdr, arc_buf_size(buf), buf); 24775602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2478dcbf3bd6SGeorge Wilson } 2479dcbf3bd6SGeorge Wilson VERIFY3P(buf->b_data, !=, NULL); 248089c86e32SChris Williamson 248189c86e32SChris Williamson hdr->b_l1hdr.b_buf = buf; 2482dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt += 1; 248389c86e32SChris Williamson 24845602294fSDan Kimmel /* 24855602294fSDan Kimmel * If the user wants the data from the hdr, we need to either copy or 24865602294fSDan Kimmel * decompress the data. 24875602294fSDan Kimmel */ 24885602294fSDan Kimmel if (fill) { 24895602294fSDan Kimmel return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0)); 24905602294fSDan Kimmel } 2491dcbf3bd6SGeorge Wilson 24925602294fSDan Kimmel return (0); 24935602294fSDan Kimmel } 2494dcbf3bd6SGeorge Wilson 24955602294fSDan Kimmel static char *arc_onloan_tag = "onloan"; 2496fa9e4066Sahrens 24975602294fSDan Kimmel static inline void 24985602294fSDan Kimmel arc_loaned_bytes_update(int64_t delta) 24995602294fSDan Kimmel { 25005602294fSDan Kimmel atomic_add_64(&arc_loaned_bytes, delta); 2501dcbf3bd6SGeorge Wilson 25025602294fSDan Kimmel /* assert that it did not wrap around */ 25035602294fSDan Kimmel ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 2504fa9e4066Sahrens } 2505fa9e4066Sahrens 25062fdbea25SAleksandr Guzovskiy /* 25072fdbea25SAleksandr Guzovskiy * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 25082fdbea25SAleksandr Guzovskiy * flight data by arc_tempreserve_space() until they are "returned". Loaned 25092fdbea25SAleksandr Guzovskiy * buffers must be returned to the arc before they can be used by the DMU or 25102fdbea25SAleksandr Guzovskiy * freed. 25112fdbea25SAleksandr Guzovskiy */ 25122fdbea25SAleksandr Guzovskiy arc_buf_t * 25135602294fSDan Kimmel arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size) 25142fdbea25SAleksandr Guzovskiy { 25155602294fSDan Kimmel arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag, 25165602294fSDan Kimmel is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size); 25175602294fSDan Kimmel 25185602294fSDan Kimmel arc_loaned_bytes_update(size); 25195602294fSDan Kimmel 25205602294fSDan Kimmel return (buf); 25215602294fSDan Kimmel } 25222fdbea25SAleksandr Guzovskiy 25235602294fSDan Kimmel arc_buf_t * 25245602294fSDan Kimmel arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize, 25255602294fSDan Kimmel enum zio_compress compression_type) 25265602294fSDan Kimmel { 25275602294fSDan Kimmel arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag, 25285602294fSDan Kimmel psize, lsize, compression_type); 25295602294fSDan Kimmel 25305602294fSDan Kimmel arc_loaned_bytes_update(psize); 25312fdbea25SAleksandr Guzovskiy 25322fdbea25SAleksandr Guzovskiy return (buf); 25332fdbea25SAleksandr Guzovskiy } 25342fdbea25SAleksandr Guzovskiy 25355602294fSDan Kimmel 25362fdbea25SAleksandr Guzovskiy /* 25372fdbea25SAleksandr Guzovskiy * Return a loaned arc buffer to the arc. 25382fdbea25SAleksandr Guzovskiy */ 25392fdbea25SAleksandr Guzovskiy void 25402fdbea25SAleksandr Guzovskiy arc_return_buf(arc_buf_t *buf, void *tag) 25412fdbea25SAleksandr Guzovskiy { 25422fdbea25SAleksandr Guzovskiy arc_buf_hdr_t *hdr = buf->b_hdr; 25432fdbea25SAleksandr Guzovskiy 2544dcbf3bd6SGeorge Wilson ASSERT3P(buf->b_data, !=, NULL); 254589c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 254689c86e32SChris Williamson (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 254789c86e32SChris Williamson (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 25482fdbea25SAleksandr Guzovskiy 25495602294fSDan Kimmel arc_loaned_bytes_update(-arc_buf_size(buf)); 25502fdbea25SAleksandr Guzovskiy } 25512fdbea25SAleksandr Guzovskiy 2552c242f9a0Schunli zhang - Sun Microsystems - Irvine United States /* Detach an arc_buf from a dbuf (tag) */ 2553c242f9a0Schunli zhang - Sun Microsystems - Irvine United States void 2554c242f9a0Schunli zhang - Sun Microsystems - Irvine United States arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2555c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 255689c86e32SChris Williamson arc_buf_hdr_t *hdr = buf->b_hdr; 2557c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 2558dcbf3bd6SGeorge Wilson ASSERT3P(buf->b_data, !=, NULL); 255989c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 256089c86e32SChris Williamson (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 256189c86e32SChris Williamson (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2562c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 25635602294fSDan Kimmel arc_loaned_bytes_update(arc_buf_size(buf)); 2564c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 2565c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 2566dcbf3bd6SGeorge Wilson static void 2567*770499e1SDan Kimmel l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type) 2568ea8dc4b6Seschrock { 2569dcbf3bd6SGeorge Wilson l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); 2570ea8dc4b6Seschrock 2571*770499e1SDan Kimmel df->l2df_abd = abd; 2572dcbf3bd6SGeorge Wilson df->l2df_size = size; 2573dcbf3bd6SGeorge Wilson df->l2df_type = type; 2574dcbf3bd6SGeorge Wilson mutex_enter(&l2arc_free_on_write_mtx); 2575dcbf3bd6SGeorge Wilson list_insert_head(l2arc_free_on_write, df); 2576dcbf3bd6SGeorge Wilson mutex_exit(&l2arc_free_on_write_mtx); 2577dcbf3bd6SGeorge Wilson } 2578b24ab676SJeff Bonwick 2579dcbf3bd6SGeorge Wilson static void 2580dcbf3bd6SGeorge Wilson arc_hdr_free_on_write(arc_buf_hdr_t *hdr) 2581dcbf3bd6SGeorge Wilson { 2582dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2583dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 2584dcbf3bd6SGeorge Wilson uint64_t size = arc_hdr_size(hdr); 2585dcbf3bd6SGeorge Wilson 2586dcbf3bd6SGeorge Wilson /* protected by hash lock, if in the hash table */ 2587dcbf3bd6SGeorge Wilson if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 2588dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2589dcbf3bd6SGeorge Wilson ASSERT(state != arc_anon && state != arc_l2c_only); 2590dcbf3bd6SGeorge Wilson 2591dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 2592dcbf3bd6SGeorge Wilson size, hdr); 2593dcbf3bd6SGeorge Wilson } 2594dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_size, size, hdr); 25956de76ce2SAndriy Gapon if (type == ARC_BUFC_METADATA) { 25966de76ce2SAndriy Gapon arc_space_return(size, ARC_SPACE_META); 25976de76ce2SAndriy Gapon } else { 25986de76ce2SAndriy Gapon ASSERT(type == ARC_BUFC_DATA); 25996de76ce2SAndriy Gapon arc_space_return(size, ARC_SPACE_DATA); 26006de76ce2SAndriy Gapon } 2601dcbf3bd6SGeorge Wilson 2602*770499e1SDan Kimmel l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type); 2603dcbf3bd6SGeorge Wilson } 2604dcbf3bd6SGeorge Wilson 2605dcbf3bd6SGeorge Wilson /* 2606dcbf3bd6SGeorge Wilson * Share the arc_buf_t's data with the hdr. Whenever we are sharing the 2607dcbf3bd6SGeorge Wilson * data buffer, we transfer the refcount ownership to the hdr and update 2608dcbf3bd6SGeorge Wilson * the appropriate kstats. 2609dcbf3bd6SGeorge Wilson */ 2610dcbf3bd6SGeorge Wilson static void 2611dcbf3bd6SGeorge Wilson arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2612dcbf3bd6SGeorge Wilson { 2613dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2614dcbf3bd6SGeorge Wilson 26155602294fSDan Kimmel ASSERT(arc_can_share(hdr, buf)); 2616*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2617dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 26189253d63dSGeorge Wilson 26199253d63dSGeorge Wilson /* 2620dcbf3bd6SGeorge Wilson * Start sharing the data buffer. We transfer the 2621dcbf3bd6SGeorge Wilson * refcount ownership to the hdr since it always owns 2622dcbf3bd6SGeorge Wilson * the refcount whenever an arc_buf_t is shared. 26239253d63dSGeorge Wilson */ 2624dcbf3bd6SGeorge Wilson refcount_transfer_ownership(&state->arcs_size, buf, hdr); 2625*770499e1SDan Kimmel hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf)); 2626*770499e1SDan Kimmel abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd, 2627*770499e1SDan Kimmel HDR_ISTYPE_METADATA(hdr)); 2628dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 26295602294fSDan Kimmel buf->b_flags |= ARC_BUF_FLAG_SHARED; 2630dcbf3bd6SGeorge Wilson 2631dcbf3bd6SGeorge Wilson /* 2632dcbf3bd6SGeorge Wilson * Since we've transferred ownership to the hdr we need 2633dcbf3bd6SGeorge Wilson * to increment its compressed and uncompressed kstats and 2634dcbf3bd6SGeorge Wilson * decrement the overhead size. 2635dcbf3bd6SGeorge Wilson */ 2636dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2637dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 26385602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf)); 2639ea8dc4b6Seschrock } 2640ea8dc4b6Seschrock 2641dcbf3bd6SGeorge Wilson static void 2642dcbf3bd6SGeorge Wilson arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2643ea8dc4b6Seschrock { 2644dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2645dcbf3bd6SGeorge Wilson 2646dcbf3bd6SGeorge Wilson ASSERT(arc_buf_is_shared(buf)); 2647*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2648dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2649ea8dc4b6Seschrock 26509b23f181Smaybee /* 2651dcbf3bd6SGeorge Wilson * We are no longer sharing this buffer so we need 2652dcbf3bd6SGeorge Wilson * to transfer its ownership to the rightful owner. 26539b23f181Smaybee */ 2654dcbf3bd6SGeorge Wilson refcount_transfer_ownership(&state->arcs_size, hdr, buf); 2655dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2656*770499e1SDan Kimmel abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd); 2657*770499e1SDan Kimmel abd_put(hdr->b_l1hdr.b_pabd); 2658*770499e1SDan Kimmel hdr->b_l1hdr.b_pabd = NULL; 26595602294fSDan Kimmel buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 2660dcbf3bd6SGeorge Wilson 2661dcbf3bd6SGeorge Wilson /* 2662dcbf3bd6SGeorge Wilson * Since the buffer is no longer shared between 2663dcbf3bd6SGeorge Wilson * the arc buf and the hdr, count it as overhead. 2664dcbf3bd6SGeorge Wilson */ 2665dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2666dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 26675602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2668dcbf3bd6SGeorge Wilson } 2669dcbf3bd6SGeorge Wilson 2670dcbf3bd6SGeorge Wilson /* 26715602294fSDan Kimmel * Remove an arc_buf_t from the hdr's buf list and return the last 26725602294fSDan Kimmel * arc_buf_t on the list. If no buffers remain on the list then return 26735602294fSDan Kimmel * NULL. 26745602294fSDan Kimmel */ 26755602294fSDan Kimmel static arc_buf_t * 26765602294fSDan Kimmel arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf) 26775602294fSDan Kimmel { 26785602294fSDan Kimmel ASSERT(HDR_HAS_L1HDR(hdr)); 26795602294fSDan Kimmel ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 26805602294fSDan Kimmel 26815602294fSDan Kimmel arc_buf_t **bufp = &hdr->b_l1hdr.b_buf; 26825602294fSDan Kimmel arc_buf_t *lastbuf = NULL; 26835602294fSDan Kimmel 26845602294fSDan Kimmel /* 26855602294fSDan Kimmel * Remove the buf from the hdr list and locate the last 26865602294fSDan Kimmel * remaining buffer on the list. 26875602294fSDan Kimmel */ 26885602294fSDan Kimmel while (*bufp != NULL) { 26895602294fSDan Kimmel if (*bufp == buf) 26905602294fSDan Kimmel *bufp = buf->b_next; 26915602294fSDan Kimmel 26925602294fSDan Kimmel /* 26935602294fSDan Kimmel * If we've removed a buffer in the middle of 26945602294fSDan Kimmel * the list then update the lastbuf and update 26955602294fSDan Kimmel * bufp. 26965602294fSDan Kimmel */ 26975602294fSDan Kimmel if (*bufp != NULL) { 26985602294fSDan Kimmel lastbuf = *bufp; 26995602294fSDan Kimmel bufp = &(*bufp)->b_next; 27005602294fSDan Kimmel } 27015602294fSDan Kimmel } 27025602294fSDan Kimmel buf->b_next = NULL; 27035602294fSDan Kimmel ASSERT3P(lastbuf, !=, buf); 27045602294fSDan Kimmel IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL); 27055602294fSDan Kimmel IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL); 27065602294fSDan Kimmel IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf)); 27075602294fSDan Kimmel 27085602294fSDan Kimmel return (lastbuf); 27095602294fSDan Kimmel } 27105602294fSDan Kimmel 27115602294fSDan Kimmel /* 27125602294fSDan Kimmel * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's 27135602294fSDan Kimmel * list and free it. 2714dcbf3bd6SGeorge Wilson */ 2715dcbf3bd6SGeorge Wilson static void 27165602294fSDan Kimmel arc_buf_destroy_impl(arc_buf_t *buf) 2717dcbf3bd6SGeorge Wilson { 2718dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 2719dcbf3bd6SGeorge Wilson 2720dcbf3bd6SGeorge Wilson /* 27215602294fSDan Kimmel * Free up the data associated with the buf but only if we're not 27225602294fSDan Kimmel * sharing this with the hdr. If we are sharing it with the hdr, the 27235602294fSDan Kimmel * hdr is responsible for doing the free. 2724dcbf3bd6SGeorge Wilson */ 2725dcbf3bd6SGeorge Wilson if (buf->b_data != NULL) { 2726dcbf3bd6SGeorge Wilson /* 2727dcbf3bd6SGeorge Wilson * We're about to change the hdr's b_flags. We must either 2728dcbf3bd6SGeorge Wilson * hold the hash_lock or be undiscoverable. 2729dcbf3bd6SGeorge Wilson */ 2730dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2731dcbf3bd6SGeorge Wilson 2732dcbf3bd6SGeorge Wilson arc_cksum_verify(buf); 2733dcbf3bd6SGeorge Wilson arc_buf_unwatch(buf); 2734dcbf3bd6SGeorge Wilson 27355602294fSDan Kimmel if (arc_buf_is_shared(buf)) { 2736dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2737dcbf3bd6SGeorge Wilson } else { 27385602294fSDan Kimmel uint64_t size = arc_buf_size(buf); 2739dcbf3bd6SGeorge Wilson arc_free_data_buf(hdr, buf->b_data, size, buf); 2740dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_overhead_size, -size); 2741dcbf3bd6SGeorge Wilson } 2742dcbf3bd6SGeorge Wilson buf->b_data = NULL; 2743dcbf3bd6SGeorge Wilson 2744dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 2745dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt -= 1; 2746dcbf3bd6SGeorge Wilson } 2747dcbf3bd6SGeorge Wilson 27485602294fSDan Kimmel arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 2749dcbf3bd6SGeorge Wilson 27505602294fSDan Kimmel if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) { 2751dcbf3bd6SGeorge Wilson /* 27525602294fSDan Kimmel * If the current arc_buf_t is sharing its data buffer with the 2753*770499e1SDan Kimmel * hdr, then reassign the hdr's b_pabd to share it with the new 27545602294fSDan Kimmel * buffer at the end of the list. The shared buffer is always 27555602294fSDan Kimmel * the last one on the hdr's buffer list. 27565602294fSDan Kimmel * 27575602294fSDan Kimmel * There is an equivalent case for compressed bufs, but since 27585602294fSDan Kimmel * they aren't guaranteed to be the last buf in the list and 27595602294fSDan Kimmel * that is an exceedingly rare case, we just allow that space be 27605602294fSDan Kimmel * wasted temporarily. 2761dcbf3bd6SGeorge Wilson */ 27625602294fSDan Kimmel if (lastbuf != NULL) { 27635602294fSDan Kimmel /* Only one buf can be shared at once */ 27645602294fSDan Kimmel VERIFY(!arc_buf_is_shared(lastbuf)); 27655602294fSDan Kimmel /* hdr is uncompressed so can't have compressed buf */ 27665602294fSDan Kimmel VERIFY(!ARC_BUF_COMPRESSED(lastbuf)); 2767dcbf3bd6SGeorge Wilson 2768*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2769*770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 2770dcbf3bd6SGeorge Wilson 27715602294fSDan Kimmel /* 27725602294fSDan Kimmel * We must setup a new shared block between the 27735602294fSDan Kimmel * last buffer and the hdr. The data would have 27745602294fSDan Kimmel * been allocated by the arc buf so we need to transfer 27755602294fSDan Kimmel * ownership to the hdr since it's now being shared. 27765602294fSDan Kimmel */ 27775602294fSDan Kimmel arc_share_buf(hdr, lastbuf); 27785602294fSDan Kimmel } 27795602294fSDan Kimmel } else if (HDR_SHARED_DATA(hdr)) { 2780dcbf3bd6SGeorge Wilson /* 27815602294fSDan Kimmel * Uncompressed shared buffers are always at the end 27825602294fSDan Kimmel * of the list. Compressed buffers don't have the 27835602294fSDan Kimmel * same requirements. This makes it hard to 27845602294fSDan Kimmel * simply assert that the lastbuf is shared so 27855602294fSDan Kimmel * we rely on the hdr's compression flags to determine 27865602294fSDan Kimmel * if we have a compressed, shared buffer. 2787dcbf3bd6SGeorge Wilson */ 27885602294fSDan Kimmel ASSERT3P(lastbuf, !=, NULL); 27895602294fSDan Kimmel ASSERT(arc_buf_is_shared(lastbuf) || 27905602294fSDan Kimmel HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 2791dcbf3bd6SGeorge Wilson } 2792dcbf3bd6SGeorge Wilson 27935602294fSDan Kimmel /* 27945602294fSDan Kimmel * Free the checksum if we're removing the last uncompressed buf from 27955602294fSDan Kimmel * this hdr. 27965602294fSDan Kimmel */ 27975602294fSDan Kimmel if (!arc_hdr_has_uncompressed_buf(hdr)) { 2798dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 27995602294fSDan Kimmel } 2800dcbf3bd6SGeorge Wilson 2801dcbf3bd6SGeorge Wilson /* clean up the buf */ 2802dcbf3bd6SGeorge Wilson buf->b_hdr = NULL; 2803dcbf3bd6SGeorge Wilson kmem_cache_free(buf_cache, buf); 2804dcbf3bd6SGeorge Wilson } 2805dcbf3bd6SGeorge Wilson 2806dcbf3bd6SGeorge Wilson static void 2807*770499e1SDan Kimmel arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr) 2808dcbf3bd6SGeorge Wilson { 2809dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 281089c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 2811dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(hdr)); 2812ea8dc4b6Seschrock 2813*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2814*770499e1SDan Kimmel hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr); 2815dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2816*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 281789c86e32SChris Williamson 2818dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2819dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2820ea8dc4b6Seschrock } 2821ea8dc4b6Seschrock 2822244781f1SPrakash Surya static void 2823*770499e1SDan Kimmel arc_hdr_free_pabd(arc_buf_hdr_t *hdr) 2824244781f1SPrakash Surya { 2825dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2826*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2827244781f1SPrakash Surya 2828dcbf3bd6SGeorge Wilson /* 2829dcbf3bd6SGeorge Wilson * If the hdr is currently being written to the l2arc then 2830dcbf3bd6SGeorge Wilson * we defer freeing the data by adding it to the l2arc_free_on_write 2831dcbf3bd6SGeorge Wilson * list. The l2arc will free the data once it's finished 2832dcbf3bd6SGeorge Wilson * writing it to the l2arc device. 2833dcbf3bd6SGeorge Wilson */ 2834dcbf3bd6SGeorge Wilson if (HDR_L2_WRITING(hdr)) { 2835dcbf3bd6SGeorge Wilson arc_hdr_free_on_write(hdr); 2836dcbf3bd6SGeorge Wilson ARCSTAT_BUMP(arcstat_l2_free_on_write); 2837dcbf3bd6SGeorge Wilson } else { 2838*770499e1SDan Kimmel arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, 2839dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2840dcbf3bd6SGeorge Wilson } 2841*770499e1SDan Kimmel hdr->b_l1hdr.b_pabd = NULL; 2842dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2843dcbf3bd6SGeorge Wilson 2844dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2845dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2846dcbf3bd6SGeorge Wilson } 2847dcbf3bd6SGeorge Wilson 2848dcbf3bd6SGeorge Wilson static arc_buf_hdr_t * 2849dcbf3bd6SGeorge Wilson arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, 28505602294fSDan Kimmel enum zio_compress compression_type, arc_buf_contents_t type) 2851dcbf3bd6SGeorge Wilson { 2852dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr; 2853dcbf3bd6SGeorge Wilson 2854dcbf3bd6SGeorge Wilson VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); 2855dcbf3bd6SGeorge Wilson 2856dcbf3bd6SGeorge Wilson hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2857dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 2858dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2859dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL); 2860dcbf3bd6SGeorge Wilson HDR_SET_PSIZE(hdr, psize); 2861dcbf3bd6SGeorge Wilson HDR_SET_LSIZE(hdr, lsize); 2862dcbf3bd6SGeorge Wilson hdr->b_spa = spa; 2863dcbf3bd6SGeorge Wilson hdr->b_type = type; 2864dcbf3bd6SGeorge Wilson hdr->b_flags = 0; 2865dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); 28665602294fSDan Kimmel arc_hdr_set_compress(hdr, compression_type); 2867dcbf3bd6SGeorge Wilson 2868dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_state = arc_anon; 2869dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_arc_access = 0; 2870dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt = 0; 2871dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_buf = NULL; 2872dcbf3bd6SGeorge Wilson 2873dcbf3bd6SGeorge Wilson /* 2874dcbf3bd6SGeorge Wilson * Allocate the hdr's buffer. This will contain either 2875dcbf3bd6SGeorge Wilson * the compressed or uncompressed data depending on the block 2876dcbf3bd6SGeorge Wilson * it references and compressed arc enablement. 2877dcbf3bd6SGeorge Wilson */ 2878*770499e1SDan Kimmel arc_hdr_alloc_pabd(hdr); 2879dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2880dcbf3bd6SGeorge Wilson 2881dcbf3bd6SGeorge Wilson return (hdr); 2882244781f1SPrakash Surya } 2883244781f1SPrakash Surya 2884fa94a07fSbrendan /* 2885dcbf3bd6SGeorge Wilson * Transition between the two allocation states for the arc_buf_hdr struct. 2886dcbf3bd6SGeorge Wilson * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 2887dcbf3bd6SGeorge Wilson * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 2888dcbf3bd6SGeorge Wilson * version is used when a cache buffer is only in the L2ARC in order to reduce 2889dcbf3bd6SGeorge Wilson * memory usage. 2890fa94a07fSbrendan */ 2891dcbf3bd6SGeorge Wilson static arc_buf_hdr_t * 2892dcbf3bd6SGeorge Wilson arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 2893fa94a07fSbrendan { 2894dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L2HDR(hdr)); 2895dcbf3bd6SGeorge Wilson 2896dcbf3bd6SGeorge Wilson arc_buf_hdr_t *nhdr; 2897dcbf3bd6SGeorge Wilson l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2898dcbf3bd6SGeorge Wilson 2899dcbf3bd6SGeorge Wilson ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 2900dcbf3bd6SGeorge Wilson (old == hdr_l2only_cache && new == hdr_full_cache)); 2901dcbf3bd6SGeorge Wilson 2902dcbf3bd6SGeorge Wilson nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 2903dcbf3bd6SGeorge Wilson 2904dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 2905dcbf3bd6SGeorge Wilson buf_hash_remove(hdr); 2906dcbf3bd6SGeorge Wilson 2907dcbf3bd6SGeorge Wilson bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 2908dcbf3bd6SGeorge Wilson 2909dcbf3bd6SGeorge Wilson if (new == hdr_full_cache) { 2910dcbf3bd6SGeorge Wilson arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2911dcbf3bd6SGeorge Wilson /* 2912dcbf3bd6SGeorge Wilson * arc_access and arc_change_state need to be aware that a 2913dcbf3bd6SGeorge Wilson * header has just come out of L2ARC, so we set its state to 2914dcbf3bd6SGeorge Wilson * l2c_only even though it's about to change. 2915dcbf3bd6SGeorge Wilson */ 2916dcbf3bd6SGeorge Wilson nhdr->b_l1hdr.b_state = arc_l2c_only; 2917dcbf3bd6SGeorge Wilson 2918dcbf3bd6SGeorge Wilson /* Verify previous threads set to NULL before freeing */ 2919*770499e1SDan Kimmel ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL); 2920dcbf3bd6SGeorge Wilson } else { 2921dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2922dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 2923dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2924dcbf3bd6SGeorge Wilson 2925dcbf3bd6SGeorge Wilson /* 2926dcbf3bd6SGeorge Wilson * If we've reached here, We must have been called from 2927dcbf3bd6SGeorge Wilson * arc_evict_hdr(), as such we should have already been 2928dcbf3bd6SGeorge Wilson * removed from any ghost list we were previously on 2929dcbf3bd6SGeorge Wilson * (which protects us from racing with arc_evict_state), 2930dcbf3bd6SGeorge Wilson * thus no locking is needed during this check. 2931dcbf3bd6SGeorge Wilson */ 2932dcbf3bd6SGeorge Wilson ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2933cd1c8b85SMatthew Ahrens 2934dcbf3bd6SGeorge Wilson /* 2935dcbf3bd6SGeorge Wilson * A buffer must not be moved into the arc_l2c_only 2936dcbf3bd6SGeorge Wilson * state if it's not finished being written out to the 2937*770499e1SDan Kimmel * l2arc device. Otherwise, the b_l1hdr.b_pabd field 2938dcbf3bd6SGeorge Wilson * might try to be accessed, even though it was removed. 2939dcbf3bd6SGeorge Wilson */ 2940dcbf3bd6SGeorge Wilson VERIFY(!HDR_L2_WRITING(hdr)); 2941*770499e1SDan Kimmel VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2942fa94a07fSbrendan 2943dcbf3bd6SGeorge Wilson #ifdef ZFS_DEBUG 2944dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_thawed != NULL) { 2945dcbf3bd6SGeorge Wilson kmem_free(hdr->b_l1hdr.b_thawed, 1); 2946dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_thawed = NULL; 2947dcbf3bd6SGeorge Wilson } 2948dcbf3bd6SGeorge Wilson #endif 2949244781f1SPrakash Surya 2950dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2951dcbf3bd6SGeorge Wilson } 2952244781f1SPrakash Surya /* 2953dcbf3bd6SGeorge Wilson * The header has been reallocated so we need to re-insert it into any 2954dcbf3bd6SGeorge Wilson * lists it was on. 2955244781f1SPrakash Surya */ 2956dcbf3bd6SGeorge Wilson (void) buf_hash_insert(nhdr, NULL); 2957244781f1SPrakash Surya 2958dcbf3bd6SGeorge Wilson ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 2959dcbf3bd6SGeorge Wilson 2960dcbf3bd6SGeorge Wilson mutex_enter(&dev->l2ad_mtx); 2961244781f1SPrakash Surya 2962244781f1SPrakash Surya /* 2963dcbf3bd6SGeorge Wilson * We must place the realloc'ed header back into the list at 2964dcbf3bd6SGeorge Wilson * the same spot. Otherwise, if it's placed earlier in the list, 2965dcbf3bd6SGeorge Wilson * l2arc_write_buffers() could find it during the function's 2966dcbf3bd6SGeorge Wilson * write phase, and try to write it out to the l2arc. 2967244781f1SPrakash Surya */ 2968dcbf3bd6SGeorge Wilson list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 2969dcbf3bd6SGeorge Wilson list_remove(&dev->l2ad_buflist, hdr); 2970dcbf3bd6SGeorge Wilson 2971dcbf3bd6SGeorge Wilson mutex_exit(&dev->l2ad_mtx); 2972244781f1SPrakash Surya 2973244781f1SPrakash Surya /* 2974dcbf3bd6SGeorge Wilson * Since we're using the pointer address as the tag when 2975dcbf3bd6SGeorge Wilson * incrementing and decrementing the l2ad_alloc refcount, we 2976dcbf3bd6SGeorge Wilson * must remove the old pointer (that we're about to destroy) and 2977dcbf3bd6SGeorge Wilson * add the new pointer to the refcount. Otherwise we'd remove 2978dcbf3bd6SGeorge Wilson * the wrong pointer address when calling arc_hdr_destroy() later. 2979244781f1SPrakash Surya */ 2980244781f1SPrakash Surya 2981dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); 2982dcbf3bd6SGeorge Wilson (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr); 2983244781f1SPrakash Surya 2984dcbf3bd6SGeorge Wilson buf_discard_identity(hdr); 2985dcbf3bd6SGeorge Wilson kmem_cache_free(old, hdr); 2986244781f1SPrakash Surya 2987dcbf3bd6SGeorge Wilson return (nhdr); 2988244781f1SPrakash Surya } 2989244781f1SPrakash Surya 2990bbfa8ea8SMatthew Ahrens /* 2991dcbf3bd6SGeorge Wilson * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. 2992dcbf3bd6SGeorge Wilson * The buf is returned thawed since we expect the consumer to modify it. 2993bbfa8ea8SMatthew Ahrens */ 2994dcbf3bd6SGeorge Wilson arc_buf_t * 29955602294fSDan Kimmel arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size) 2996ea8dc4b6Seschrock { 2997dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, 2998dcbf3bd6SGeorge Wilson ZIO_COMPRESS_OFF, type); 2999dcbf3bd6SGeorge Wilson ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 30005602294fSDan Kimmel 30015602294fSDan Kimmel arc_buf_t *buf = NULL; 30025602294fSDan Kimmel VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf)); 3003dcbf3bd6SGeorge Wilson arc_buf_thaw(buf); 30045602294fSDan Kimmel 30055602294fSDan Kimmel return (buf); 30065602294fSDan Kimmel } 30075602294fSDan Kimmel 30085602294fSDan Kimmel /* 30095602294fSDan Kimmel * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this 30105602294fSDan Kimmel * for bufs containing metadata. 30115602294fSDan Kimmel */ 30125602294fSDan Kimmel arc_buf_t * 30135602294fSDan Kimmel arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize, 30145602294fSDan Kimmel enum zio_compress compression_type) 30155602294fSDan Kimmel { 30165602294fSDan Kimmel ASSERT3U(lsize, >, 0); 30175602294fSDan Kimmel ASSERT3U(lsize, >=, psize); 30185602294fSDan Kimmel ASSERT(compression_type > ZIO_COMPRESS_OFF); 30195602294fSDan Kimmel ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS); 30205602294fSDan Kimmel 30215602294fSDan Kimmel arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 30225602294fSDan Kimmel compression_type, ARC_BUFC_DATA); 30235602294fSDan Kimmel ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 30245602294fSDan Kimmel 30255602294fSDan Kimmel arc_buf_t *buf = NULL; 30265602294fSDan Kimmel VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf)); 30275602294fSDan Kimmel arc_buf_thaw(buf); 30285602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 30295602294fSDan Kimmel 3030*770499e1SDan Kimmel if (!arc_buf_is_shared(buf)) { 3031*770499e1SDan Kimmel /* 3032*770499e1SDan Kimmel * To ensure that the hdr has the correct data in it if we call 3033*770499e1SDan Kimmel * arc_decompress() on this buf before it's been written to 3034*770499e1SDan Kimmel * disk, it's easiest if we just set up sharing between the 3035*770499e1SDan Kimmel * buf and the hdr. 3036*770499e1SDan Kimmel */ 3037*770499e1SDan Kimmel ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd)); 3038*770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 3039*770499e1SDan Kimmel arc_share_buf(hdr, buf); 3040*770499e1SDan Kimmel } 3041*770499e1SDan Kimmel 3042dcbf3bd6SGeorge Wilson return (buf); 3043ea8dc4b6Seschrock } 3044ea8dc4b6Seschrock 3045a52fc310SPrakash Surya static void 3046a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 3047a52fc310SPrakash Surya { 3048a52fc310SPrakash Surya l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 3049a52fc310SPrakash Surya l2arc_dev_t *dev = l2hdr->b_dev; 3050dcbf3bd6SGeorge Wilson uint64_t asize = arc_hdr_size(hdr); 3051a52fc310SPrakash Surya 3052a52fc310SPrakash Surya ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 3053a52fc310SPrakash Surya ASSERT(HDR_HAS_L2HDR(hdr)); 3054a52fc310SPrakash Surya 3055a52fc310SPrakash Surya list_remove(&dev->l2ad_buflist, hdr); 3056a52fc310SPrakash Surya 3057dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_l2_asize, -asize); 3058dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr)); 3059a52fc310SPrakash Surya 3060dcbf3bd6SGeorge Wilson vdev_space_update(dev->l2ad_vdev, -asize, 0, 0); 3061a52fc310SPrakash Surya 3062dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&dev->l2ad_alloc, asize, hdr); 3063dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 3064a52fc310SPrakash Surya } 3065a52fc310SPrakash Surya 3066fa9e4066Sahrens static void 3067ea8dc4b6Seschrock arc_hdr_destroy(arc_buf_hdr_t *hdr) 3068fa9e4066Sahrens { 306989c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) { 307089c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_buf == NULL || 3071dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt > 0); 307289c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 307389c86e32SChris Williamson ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 307489c86e32SChris Williamson } 3075ea8dc4b6Seschrock ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 307689c86e32SChris Williamson ASSERT(!HDR_IN_HASH_TABLE(hdr)); 307789c86e32SChris Williamson 3078dcbf3bd6SGeorge Wilson if (!HDR_EMPTY(hdr)) 3079dcbf3bd6SGeorge Wilson buf_discard_identity(hdr); 3080dcbf3bd6SGeorge Wilson 308189c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr)) { 3082a52fc310SPrakash Surya l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3083a52fc310SPrakash Surya boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 3084b24ab676SJeff Bonwick 3085a52fc310SPrakash Surya if (!buflist_held) 3086a52fc310SPrakash Surya mutex_enter(&dev->l2ad_mtx); 308789c86e32SChris Williamson 3088244781f1SPrakash Surya /* 3089a52fc310SPrakash Surya * Even though we checked this conditional above, we 3090a52fc310SPrakash Surya * need to check this again now that we have the 3091a52fc310SPrakash Surya * l2ad_mtx. This is because we could be racing with 3092a52fc310SPrakash Surya * another thread calling l2arc_evict() which might have 3093a52fc310SPrakash Surya * destroyed this header's L2 portion as we were waiting 3094a52fc310SPrakash Surya * to acquire the l2ad_mtx. If that happens, we don't 3095a52fc310SPrakash Surya * want to re-destroy the header's L2 portion. 3096244781f1SPrakash Surya */ 3097a52fc310SPrakash Surya if (HDR_HAS_L2HDR(hdr)) 3098a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(hdr); 3099b24ab676SJeff Bonwick 3100b24ab676SJeff Bonwick if (!buflist_held) 3101a52fc310SPrakash Surya mutex_exit(&dev->l2ad_mtx); 3102fa94a07fSbrendan } 3103fa94a07fSbrendan 3104dcbf3bd6SGeorge Wilson if (HDR_HAS_L1HDR(hdr)) { 3105dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 310689c86e32SChris Williamson 3107dcbf3bd6SGeorge Wilson while (hdr->b_l1hdr.b_buf != NULL) 31085602294fSDan Kimmel arc_buf_destroy_impl(hdr->b_l1hdr.b_buf); 310989c86e32SChris Williamson 311089c86e32SChris Williamson #ifdef ZFS_DEBUG 311189c86e32SChris Williamson if (hdr->b_l1hdr.b_thawed != NULL) { 311289c86e32SChris Williamson kmem_free(hdr->b_l1hdr.b_thawed, 1); 311389c86e32SChris Williamson hdr->b_l1hdr.b_thawed = NULL; 311489c86e32SChris Williamson } 311589c86e32SChris Williamson #endif 3116dcbf3bd6SGeorge Wilson 3117*770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 3118*770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 3119dcbf3bd6SGeorge Wilson } 31203f9d6ad7SLin Ling } 3121ea8dc4b6Seschrock 3122fa9e4066Sahrens ASSERT3P(hdr->b_hash_next, ==, NULL); 312389c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) { 3124244781f1SPrakash Surya ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 312589c86e32SChris Williamson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 312689c86e32SChris Williamson kmem_cache_free(hdr_full_cache, hdr); 312789c86e32SChris Williamson } else { 312889c86e32SChris Williamson kmem_cache_free(hdr_l2only_cache, hdr); 312989c86e32SChris Williamson } 3130fa9e4066Sahrens } 3131fa9e4066Sahrens 3132fa9e4066Sahrens void 3133dcbf3bd6SGeorge Wilson arc_buf_destroy(arc_buf_t *buf, void* tag) 3134ea8dc4b6Seschrock { 3135ea8dc4b6Seschrock arc_buf_hdr_t *hdr = buf->b_hdr; 3136ea8dc4b6Seschrock kmutex_t *hash_lock = HDR_LOCK(hdr); 3137fa9e4066Sahrens 313889c86e32SChris Williamson if (hdr->b_l1hdr.b_state == arc_anon) { 3139dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 3140dcbf3bd6SGeorge Wilson ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3141dcbf3bd6SGeorge Wilson VERIFY0(remove_reference(hdr, NULL, tag)); 3142dcbf3bd6SGeorge Wilson arc_hdr_destroy(hdr); 3143dcbf3bd6SGeorge Wilson return; 3144ea8dc4b6Seschrock } 3145ea8dc4b6Seschrock 3146ea8dc4b6Seschrock mutex_enter(hash_lock); 3147dcbf3bd6SGeorge Wilson ASSERT3P(hdr, ==, buf->b_hdr); 3148dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 31493f9d6ad7SLin Ling ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3150dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); 3151dcbf3bd6SGeorge Wilson ASSERT3P(buf->b_data, !=, NULL); 3152ea8dc4b6Seschrock 3153ea8dc4b6Seschrock (void) remove_reference(hdr, hash_lock, tag); 31545602294fSDan Kimmel arc_buf_destroy_impl(buf); 3155ea8dc4b6Seschrock mutex_exit(hash_lock); 3156fa9e4066Sahrens } 3157fa9e4066Sahrens 3158fa9e4066Sahrens /* 3159244781f1SPrakash Surya * Evict the arc_buf_hdr that is provided as a parameter. The resultant 3160244781f1SPrakash Surya * state of the header is dependent on it's state prior to entering this 3161244781f1SPrakash Surya * function. The following transitions are possible: 3162874395d5Smaybee * 3163244781f1SPrakash Surya * - arc_mru -> arc_mru_ghost 3164244781f1SPrakash Surya * - arc_mfu -> arc_mfu_ghost 3165244781f1SPrakash Surya * - arc_mru_ghost -> arc_l2c_only 3166244781f1SPrakash Surya * - arc_mru_ghost -> deleted 3167244781f1SPrakash Surya * - arc_mfu_ghost -> arc_l2c_only 3168244781f1SPrakash Surya * - arc_mfu_ghost -> deleted 3169fa9e4066Sahrens */ 3170244781f1SPrakash Surya static int64_t 3171244781f1SPrakash Surya arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3172fa9e4066Sahrens { 3173244781f1SPrakash Surya arc_state_t *evicted_state, *state; 3174244781f1SPrakash Surya int64_t bytes_evicted = 0; 3175fa9e4066Sahrens 3176244781f1SPrakash Surya ASSERT(MUTEX_HELD(hash_lock)); 3177244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 3178fa9e4066Sahrens 3179244781f1SPrakash Surya state = hdr->b_l1hdr.b_state; 3180244781f1SPrakash Surya if (GHOST_STATE(state)) { 3181244781f1SPrakash Surya ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3182dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3183fa9e4066Sahrens 3184244781f1SPrakash Surya /* 3185244781f1SPrakash Surya * l2arc_write_buffers() relies on a header's L1 portion 3186*770499e1SDan Kimmel * (i.e. its b_pabd field) during it's write phase. 3187244781f1SPrakash Surya * Thus, we cannot push a header onto the arc_l2c_only 3188244781f1SPrakash Surya * state (removing it's L1 piece) until the header is 3189244781f1SPrakash Surya * done being written to the l2arc. 3190244781f1SPrakash Surya */ 3191244781f1SPrakash Surya if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 3192244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_l2_skip); 3193244781f1SPrakash Surya return (bytes_evicted); 31943a5286a1SMatthew Ahrens } 3195244781f1SPrakash Surya 3196244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_deleted); 3197dcbf3bd6SGeorge Wilson bytes_evicted += HDR_GET_LSIZE(hdr); 3198244781f1SPrakash Surya 3199244781f1SPrakash Surya DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 3200244781f1SPrakash Surya 3201*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3202244781f1SPrakash Surya if (HDR_HAS_L2HDR(hdr)) { 3203244781f1SPrakash Surya /* 3204244781f1SPrakash Surya * This buffer is cached on the 2nd Level ARC; 3205244781f1SPrakash Surya * don't destroy the header. 3206244781f1SPrakash Surya */ 3207244781f1SPrakash Surya arc_change_state(arc_l2c_only, hdr, hash_lock); 32083a5286a1SMatthew Ahrens /* 3209244781f1SPrakash Surya * dropping from L1+L2 cached to L2-only, 3210244781f1SPrakash Surya * realloc to remove the L1 header. 32113a5286a1SMatthew Ahrens */ 3212244781f1SPrakash Surya hdr = arc_hdr_realloc(hdr, hdr_full_cache, 3213244781f1SPrakash Surya hdr_l2only_cache); 3214244781f1SPrakash Surya } else { 3215244781f1SPrakash Surya arc_change_state(arc_anon, hdr, hash_lock); 3216244781f1SPrakash Surya arc_hdr_destroy(hdr); 32173a5286a1SMatthew Ahrens } 3218244781f1SPrakash Surya return (bytes_evicted); 32193a5286a1SMatthew Ahrens } 32203a5286a1SMatthew Ahrens 3221244781f1SPrakash Surya ASSERT(state == arc_mru || state == arc_mfu); 3222244781f1SPrakash Surya evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3223244781f1SPrakash Surya 3224244781f1SPrakash Surya /* prefetch buffers have a minimum lifespan */ 3225244781f1SPrakash Surya if (HDR_IO_IN_PROGRESS(hdr) || 3226244781f1SPrakash Surya ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 3227244781f1SPrakash Surya ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 3228244781f1SPrakash Surya arc_min_prefetch_lifespan)) { 3229244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_skip); 3230244781f1SPrakash Surya return (bytes_evicted); 3231244781f1SPrakash Surya } 3232244781f1SPrakash Surya 3233244781f1SPrakash Surya ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3234244781f1SPrakash Surya while (hdr->b_l1hdr.b_buf) { 3235244781f1SPrakash Surya arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3236244781f1SPrakash Surya if (!mutex_tryenter(&buf->b_evict_lock)) { 3237244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_mutex_miss); 3238244781f1SPrakash Surya break; 323913506d1eSmaybee } 3240244781f1SPrakash Surya if (buf->b_data != NULL) 3241dcbf3bd6SGeorge Wilson bytes_evicted += HDR_GET_LSIZE(hdr); 3242dcbf3bd6SGeorge Wilson mutex_exit(&buf->b_evict_lock); 32435602294fSDan Kimmel arc_buf_destroy_impl(buf); 3244244781f1SPrakash Surya } 324569962b56SMatthew Ahrens 3246244781f1SPrakash Surya if (HDR_HAS_L2HDR(hdr)) { 3247dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); 3248244781f1SPrakash Surya } else { 3249dcbf3bd6SGeorge Wilson if (l2arc_write_eligible(hdr->b_spa, hdr)) { 3250dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_evict_l2_eligible, 3251dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr)); 3252dcbf3bd6SGeorge Wilson } else { 3253dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_evict_l2_ineligible, 3254dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr)); 3255dcbf3bd6SGeorge Wilson } 3256244781f1SPrakash Surya } 3257244781f1SPrakash Surya 3258dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_bufcnt == 0) { 3259dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 3260dcbf3bd6SGeorge Wilson 3261dcbf3bd6SGeorge Wilson bytes_evicted += arc_hdr_size(hdr); 3262dcbf3bd6SGeorge Wilson 3263dcbf3bd6SGeorge Wilson /* 3264dcbf3bd6SGeorge Wilson * If this hdr is being evicted and has a compressed 3265dcbf3bd6SGeorge Wilson * buffer then we discard it here before we change states. 3266dcbf3bd6SGeorge Wilson * This ensures that the accounting is updated correctly 3267*770499e1SDan Kimmel * in arc_free_data_impl(). 3268dcbf3bd6SGeorge Wilson */ 3269*770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 3270dcbf3bd6SGeorge Wilson 3271244781f1SPrakash Surya arc_change_state(evicted_state, hdr, hash_lock); 3272244781f1SPrakash Surya ASSERT(HDR_IN_HASH_TABLE(hdr)); 3273dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 3274244781f1SPrakash Surya DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 3275244781f1SPrakash Surya } 3276244781f1SPrakash Surya 3277244781f1SPrakash Surya return (bytes_evicted); 3278244781f1SPrakash Surya } 3279244781f1SPrakash Surya 3280244781f1SPrakash Surya static uint64_t 3281244781f1SPrakash Surya arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 3282244781f1SPrakash Surya uint64_t spa, int64_t bytes) 3283244781f1SPrakash Surya { 3284244781f1SPrakash Surya multilist_sublist_t *mls; 3285244781f1SPrakash Surya uint64_t bytes_evicted = 0; 3286244781f1SPrakash Surya arc_buf_hdr_t *hdr; 3287244781f1SPrakash Surya kmutex_t *hash_lock; 3288244781f1SPrakash Surya int evict_count = 0; 3289244781f1SPrakash Surya 3290244781f1SPrakash Surya ASSERT3P(marker, !=, NULL); 3291244781f1SPrakash Surya IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3292244781f1SPrakash Surya 3293244781f1SPrakash Surya mls = multilist_sublist_lock(ml, idx); 3294244781f1SPrakash Surya 3295244781f1SPrakash Surya for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 3296244781f1SPrakash Surya hdr = multilist_sublist_prev(mls, marker)) { 3297244781f1SPrakash Surya if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 3298244781f1SPrakash Surya (evict_count >= zfs_arc_evict_batch_limit)) 3299244781f1SPrakash Surya break; 330069962b56SMatthew Ahrens 330169962b56SMatthew Ahrens /* 3302244781f1SPrakash Surya * To keep our iteration location, move the marker 3303244781f1SPrakash Surya * forward. Since we're not holding hdr's hash lock, we 3304244781f1SPrakash Surya * must be very careful and not remove 'hdr' from the 3305244781f1SPrakash Surya * sublist. Otherwise, other consumers might mistake the 3306244781f1SPrakash Surya * 'hdr' as not being on a sublist when they call the 3307244781f1SPrakash Surya * multilist_link_active() function (they all rely on 3308244781f1SPrakash Surya * the hash lock protecting concurrent insertions and 3309244781f1SPrakash Surya * removals). multilist_sublist_move_forward() was 3310244781f1SPrakash Surya * specifically implemented to ensure this is the case 3311244781f1SPrakash Surya * (only 'marker' will be removed and re-inserted). 3312244781f1SPrakash Surya */ 3313244781f1SPrakash Surya multilist_sublist_move_forward(mls, marker); 3314244781f1SPrakash Surya 3315244781f1SPrakash Surya /* 3316244781f1SPrakash Surya * The only case where the b_spa field should ever be 3317244781f1SPrakash Surya * zero, is the marker headers inserted by 3318244781f1SPrakash Surya * arc_evict_state(). It's possible for multiple threads 3319244781f1SPrakash Surya * to be calling arc_evict_state() concurrently (e.g. 3320244781f1SPrakash Surya * dsl_pool_close() and zio_inject_fault()), so we must 3321244781f1SPrakash Surya * skip any markers we see from these other threads. 332269962b56SMatthew Ahrens */ 3323244781f1SPrakash Surya if (hdr->b_spa == 0) 3324244781f1SPrakash Surya continue; 3325244781f1SPrakash Surya 3326244781f1SPrakash Surya /* we're only interested in evicting buffers of a certain spa */ 3327244781f1SPrakash Surya if (spa != 0 && hdr->b_spa != spa) { 3328244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_skip); 332969962b56SMatthew Ahrens continue; 333069962b56SMatthew Ahrens } 333169962b56SMatthew Ahrens 33327adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 33335ea40c06SBrendan Gregg - Sun Microsystems 3334244781f1SPrakash Surya /* 3335244781f1SPrakash Surya * We aren't calling this function from any code path 3336244781f1SPrakash Surya * that would already be holding a hash lock, so we're 3337244781f1SPrakash Surya * asserting on this assumption to be defensive in case 3338244781f1SPrakash Surya * this ever changes. Without this check, it would be 3339244781f1SPrakash Surya * possible to incorrectly increment arcstat_mutex_miss 3340244781f1SPrakash Surya * below (e.g. if the code changed such that we called 3341244781f1SPrakash Surya * this function with a hash lock held). 3342244781f1SPrakash Surya */ 3343244781f1SPrakash Surya ASSERT(!MUTEX_HELD(hash_lock)); 334444cb6abcSbmc 3345244781f1SPrakash Surya if (mutex_tryenter(hash_lock)) { 3346244781f1SPrakash Surya uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 3347244781f1SPrakash Surya mutex_exit(hash_lock); 3348fa9e4066Sahrens 3349244781f1SPrakash Surya bytes_evicted += evicted; 3350fa9e4066Sahrens 3351244781f1SPrakash Surya /* 3352244781f1SPrakash Surya * If evicted is zero, arc_evict_hdr() must have 3353244781f1SPrakash Surya * decided to skip this header, don't increment 3354244781f1SPrakash Surya * evict_count in this case. 3355244781f1SPrakash Surya */ 3356244781f1SPrakash Surya if (evicted != 0) 3357244781f1SPrakash Surya evict_count++; 335844cb6abcSbmc 3359244781f1SPrakash Surya /* 3360244781f1SPrakash Surya * If arc_size isn't overflowing, signal any 3361244781f1SPrakash Surya * threads that might happen to be waiting. 3362244781f1SPrakash Surya * 3363244781f1SPrakash Surya * For each header evicted, we wake up a single 3364244781f1SPrakash Surya * thread. If we used cv_broadcast, we could 3365244781f1SPrakash Surya * wake up "too many" threads causing arc_size 3366244781f1SPrakash Surya * to significantly overflow arc_c; since 3367*770499e1SDan Kimmel * arc_get_data_impl() doesn't check for overflow 3368244781f1SPrakash Surya * when it's woken up (it doesn't because it's 3369244781f1SPrakash Surya * possible for the ARC to be overflowing while 3370244781f1SPrakash Surya * full of un-evictable buffers, and the 3371244781f1SPrakash Surya * function should proceed in this case). 3372244781f1SPrakash Surya * 3373244781f1SPrakash Surya * If threads are left sleeping, due to not 3374244781f1SPrakash Surya * using cv_broadcast, they will be woken up 3375244781f1SPrakash Surya * just before arc_reclaim_thread() sleeps. 3376244781f1SPrakash Surya */ 3377244781f1SPrakash Surya mutex_enter(&arc_reclaim_lock); 3378244781f1SPrakash Surya if (!arc_is_overflowing()) 3379244781f1SPrakash Surya cv_signal(&arc_reclaim_waiters_cv); 3380244781f1SPrakash Surya mutex_exit(&arc_reclaim_lock); 3381244781f1SPrakash Surya } else { 3382244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_mutex_miss); 3383244781f1SPrakash Surya } 3384244781f1SPrakash Surya } 3385f4d2e9e6Smaybee 3386244781f1SPrakash Surya multilist_sublist_unlock(mls); 338744cb6abcSbmc 3388244781f1SPrakash Surya return (bytes_evicted); 3389fa9e4066Sahrens } 3390fa9e4066Sahrens 3391fa9e4066Sahrens /* 3392244781f1SPrakash Surya * Evict buffers from the given arc state, until we've removed the 3393244781f1SPrakash Surya * specified number of bytes. Move the removed buffers to the 3394244781f1SPrakash Surya * appropriate evict state. 3395244781f1SPrakash Surya * 3396244781f1SPrakash Surya * This function makes a "best effort". It skips over any buffers 3397244781f1SPrakash Surya * it can't get a hash_lock on, and so, may not catch all candidates. 3398244781f1SPrakash Surya * It may also return without evicting as much space as requested. 3399244781f1SPrakash Surya * 3400244781f1SPrakash Surya * If bytes is specified using the special value ARC_EVICT_ALL, this 3401244781f1SPrakash Surya * will evict all available (i.e. unlocked and evictable) buffers from 3402244781f1SPrakash Surya * the given arc state; which is used by arc_flush(). 3403fa9e4066Sahrens */ 3404244781f1SPrakash Surya static uint64_t 3405244781f1SPrakash Surya arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 3406244781f1SPrakash Surya arc_buf_contents_t type) 3407fa9e4066Sahrens { 3408244781f1SPrakash Surya uint64_t total_evicted = 0; 340994c2d0ebSMatthew Ahrens multilist_t *ml = state->arcs_list[type]; 3410244781f1SPrakash Surya int num_sublists; 3411244781f1SPrakash Surya arc_buf_hdr_t **markers; 3412fa9e4066Sahrens 3413244781f1SPrakash Surya IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3414b802aa8cSSanjeev Bagewadi 3415244781f1SPrakash Surya num_sublists = multilist_get_num_sublists(ml); 3416b802aa8cSSanjeev Bagewadi 3417244781f1SPrakash Surya /* 3418244781f1SPrakash Surya * If we've tried to evict from each sublist, made some 3419244781f1SPrakash Surya * progress, but still have not hit the target number of bytes 3420244781f1SPrakash Surya * to evict, we want to keep trying. The markers allow us to 3421244781f1SPrakash Surya * pick up where we left off for each individual sublist, rather 3422244781f1SPrakash Surya * than starting from the tail each time. 3423244781f1SPrakash Surya */ 3424244781f1SPrakash Surya markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 3425244781f1SPrakash Surya for (int i = 0; i < num_sublists; i++) { 3426244781f1SPrakash Surya markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 342769962b56SMatthew Ahrens 342869962b56SMatthew Ahrens /* 3429244781f1SPrakash Surya * A b_spa of 0 is used to indicate that this header is 3430244781f1SPrakash Surya * a marker. This fact is used in arc_adjust_type() and 3431244781f1SPrakash Surya * arc_evict_state_impl(). 343269962b56SMatthew Ahrens */ 3433244781f1SPrakash Surya markers[i]->b_spa = 0; 3434fa94a07fSbrendan 3435244781f1SPrakash Surya multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3436244781f1SPrakash Surya multilist_sublist_insert_tail(mls, markers[i]); 3437244781f1SPrakash Surya multilist_sublist_unlock(mls); 3438244781f1SPrakash Surya } 3439fa94a07fSbrendan 3440244781f1SPrakash Surya /* 3441244781f1SPrakash Surya * While we haven't hit our target number of bytes to evict, or 3442244781f1SPrakash Surya * we're evicting all available buffers. 3443244781f1SPrakash Surya */ 3444244781f1SPrakash Surya while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 3445244781f1SPrakash Surya /* 3446244781f1SPrakash Surya * Start eviction using a randomly selected sublist, 3447244781f1SPrakash Surya * this is to try and evenly balance eviction across all 3448244781f1SPrakash Surya * sublists. Always starting at the same sublist 3449244781f1SPrakash Surya * (e.g. index 0) would cause evictions to favor certain 3450244781f1SPrakash Surya * sublists over others. 3451244781f1SPrakash Surya */ 3452244781f1SPrakash Surya int sublist_idx = multilist_get_random_index(ml); 3453244781f1SPrakash Surya uint64_t scan_evicted = 0; 3454244781f1SPrakash Surya 3455244781f1SPrakash Surya for (int i = 0; i < num_sublists; i++) { 3456244781f1SPrakash Surya uint64_t bytes_remaining; 3457244781f1SPrakash Surya uint64_t bytes_evicted; 3458244781f1SPrakash Surya 3459244781f1SPrakash Surya if (bytes == ARC_EVICT_ALL) 3460244781f1SPrakash Surya bytes_remaining = ARC_EVICT_ALL; 3461244781f1SPrakash Surya else if (total_evicted < bytes) 3462244781f1SPrakash Surya bytes_remaining = bytes - total_evicted; 3463244781f1SPrakash Surya else 3464fa9e4066Sahrens break; 3465244781f1SPrakash Surya 3466244781f1SPrakash Surya bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 3467244781f1SPrakash Surya markers[sublist_idx], spa, bytes_remaining); 3468244781f1SPrakash Surya 3469244781f1SPrakash Surya scan_evicted += bytes_evicted; 3470244781f1SPrakash Surya total_evicted += bytes_evicted; 3471244781f1SPrakash Surya 3472244781f1SPrakash Surya /* we've reached the end, wrap to the beginning */ 3473244781f1SPrakash Surya if (++sublist_idx >= num_sublists) 3474244781f1SPrakash Surya sublist_idx = 0; 3475244781f1SPrakash Surya } 3476244781f1SPrakash Surya 3477244781f1SPrakash Surya /* 3478244781f1SPrakash Surya * If we didn't evict anything during this scan, we have 3479244781f1SPrakash Surya * no reason to believe we'll evict more during another 3480244781f1SPrakash Surya * scan, so break the loop. 3481244781f1SPrakash Surya */ 3482244781f1SPrakash Surya if (scan_evicted == 0) { 3483244781f1SPrakash Surya /* This isn't possible, let's make that obvious */ 3484244781f1SPrakash Surya ASSERT3S(bytes, !=, 0); 3485244781f1SPrakash Surya 3486b802aa8cSSanjeev Bagewadi /* 3487244781f1SPrakash Surya * When bytes is ARC_EVICT_ALL, the only way to 3488244781f1SPrakash Surya * break the loop is when scan_evicted is zero. 3489244781f1SPrakash Surya * In that case, we actually have evicted enough, 3490244781f1SPrakash Surya * so we don't want to increment the kstat. 3491b802aa8cSSanjeev Bagewadi */ 3492244781f1SPrakash Surya if (bytes != ARC_EVICT_ALL) { 3493244781f1SPrakash Surya ASSERT3S(total_evicted, <, bytes); 3494244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_not_enough); 3495244781f1SPrakash Surya } 3496244781f1SPrakash Surya 3497244781f1SPrakash Surya break; 349869962b56SMatthew Ahrens } 3499244781f1SPrakash Surya } 3500244781f1SPrakash Surya 3501244781f1SPrakash Surya for (int i = 0; i < num_sublists; i++) { 3502244781f1SPrakash Surya multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3503244781f1SPrakash Surya multilist_sublist_remove(mls, markers[i]); 3504244781f1SPrakash Surya multilist_sublist_unlock(mls); 350569962b56SMatthew Ahrens 3506244781f1SPrakash Surya kmem_cache_free(hdr_full_cache, markers[i]); 3507fa9e4066Sahrens } 3508244781f1SPrakash Surya kmem_free(markers, sizeof (*markers) * num_sublists); 3509fa9e4066Sahrens 3510244781f1SPrakash Surya return (total_evicted); 3511244781f1SPrakash Surya } 3512244781f1SPrakash Surya 3513244781f1SPrakash Surya /* 3514244781f1SPrakash Surya * Flush all "evictable" data of the given type from the arc state 3515244781f1SPrakash Surya * specified. This will not evict any "active" buffers (i.e. referenced). 3516244781f1SPrakash Surya * 3517dcbf3bd6SGeorge Wilson * When 'retry' is set to B_FALSE, the function will make a single pass 3518244781f1SPrakash Surya * over the state and evict any buffers that it can. Since it doesn't 3519244781f1SPrakash Surya * continually retry the eviction, it might end up leaving some buffers 3520244781f1SPrakash Surya * in the ARC due to lock misses. 3521244781f1SPrakash Surya * 3522dcbf3bd6SGeorge Wilson * When 'retry' is set to B_TRUE, the function will continually retry the 3523244781f1SPrakash Surya * eviction until *all* evictable buffers have been removed from the 3524244781f1SPrakash Surya * state. As a result, if concurrent insertions into the state are 3525244781f1SPrakash Surya * allowed (e.g. if the ARC isn't shutting down), this function might 3526244781f1SPrakash Surya * wind up in an infinite loop, continually trying to evict buffers. 3527244781f1SPrakash Surya */ 3528244781f1SPrakash Surya static uint64_t 3529244781f1SPrakash Surya arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 3530244781f1SPrakash Surya boolean_t retry) 3531244781f1SPrakash Surya { 3532244781f1SPrakash Surya uint64_t evicted = 0; 3533244781f1SPrakash Surya 3534dcbf3bd6SGeorge Wilson while (refcount_count(&state->arcs_esize[type]) != 0) { 3535244781f1SPrakash Surya evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 3536244781f1SPrakash Surya 3537244781f1SPrakash Surya if (!retry) 3538244781f1SPrakash Surya break; 35390e8c6158Smaybee } 35400e8c6158Smaybee 3541244781f1SPrakash Surya return (evicted); 3542244781f1SPrakash Surya } 3543244781f1SPrakash Surya 3544244781f1SPrakash Surya /* 3545244781f1SPrakash Surya * Evict the specified number of bytes from the state specified, 3546244781f1SPrakash Surya * restricting eviction to the spa and type given. This function 3547244781f1SPrakash Surya * prevents us from trying to evict more from a state's list than 3548244781f1SPrakash Surya * is "evictable", and to skip evicting altogether when passed a 3549244781f1SPrakash Surya * negative value for "bytes". In contrast, arc_evict_state() will 3550244781f1SPrakash Surya * evict everything it can, when passed a negative value for "bytes". 3551244781f1SPrakash Surya */ 3552244781f1SPrakash Surya static uint64_t 3553244781f1SPrakash Surya arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3554244781f1SPrakash Surya arc_buf_contents_t type) 3555244781f1SPrakash Surya { 3556244781f1SPrakash Surya int64_t delta; 3557244781f1SPrakash Surya 3558dcbf3bd6SGeorge Wilson if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) { 3559dcbf3bd6SGeorge Wilson delta = MIN(refcount_count(&state->arcs_esize[type]), bytes); 3560244781f1SPrakash Surya return (arc_evict_state(state, spa, delta, type)); 3561fa9e4066Sahrens } 3562fa9e4066Sahrens 3563244781f1SPrakash Surya return (0); 3564fa9e4066Sahrens } 3565fa9e4066Sahrens 3566244781f1SPrakash Surya /* 3567244781f1SPrakash Surya * Evict metadata buffers from the cache, such that arc_meta_used is 3568244781f1SPrakash Surya * capped by the arc_meta_limit tunable. 3569244781f1SPrakash Surya */ 3570244781f1SPrakash Surya static uint64_t 3571244781f1SPrakash Surya arc_adjust_meta(void) 3572244781f1SPrakash Surya { 3573244781f1SPrakash Surya uint64_t total_evicted = 0; 3574244781f1SPrakash Surya int64_t target; 3575244781f1SPrakash Surya 3576244781f1SPrakash Surya /* 3577244781f1SPrakash Surya * If we're over the meta limit, we want to evict enough 3578244781f1SPrakash Surya * metadata to get back under the meta limit. We don't want to 3579244781f1SPrakash Surya * evict so much that we drop the MRU below arc_p, though. If 3580244781f1SPrakash Surya * we're over the meta limit more than we're over arc_p, we 3581244781f1SPrakash Surya * evict some from the MRU here, and some from the MFU below. 3582244781f1SPrakash Surya */ 3583244781f1SPrakash Surya target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 35842fd872a7SPrakash Surya (int64_t)(refcount_count(&arc_anon->arcs_size) + 35852fd872a7SPrakash Surya refcount_count(&arc_mru->arcs_size) - arc_p)); 3586244781f1SPrakash Surya 3587244781f1SPrakash Surya total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3588244781f1SPrakash Surya 3589244781f1SPrakash Surya /* 3590244781f1SPrakash Surya * Similar to the above, we want to evict enough bytes to get us 3591244781f1SPrakash Surya * below the meta limit, but not so much as to drop us below the 35925602294fSDan Kimmel * space allotted to the MFU (which is defined as arc_c - arc_p). 3593244781f1SPrakash Surya */ 3594244781f1SPrakash Surya target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 35952fd872a7SPrakash Surya (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3596244781f1SPrakash Surya 3597244781f1SPrakash Surya total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3598244781f1SPrakash Surya 3599244781f1SPrakash Surya return (total_evicted); 3600244781f1SPrakash Surya } 3601244781f1SPrakash Surya 3602244781f1SPrakash Surya /* 3603244781f1SPrakash Surya * Return the type of the oldest buffer in the given arc state 3604244781f1SPrakash Surya * 3605244781f1SPrakash Surya * This function will select a random sublist of type ARC_BUFC_DATA and 3606244781f1SPrakash Surya * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3607244781f1SPrakash Surya * is compared, and the type which contains the "older" buffer will be 3608244781f1SPrakash Surya * returned. 3609244781f1SPrakash Surya */ 3610244781f1SPrakash Surya static arc_buf_contents_t 3611244781f1SPrakash Surya arc_adjust_type(arc_state_t *state) 3612244781f1SPrakash Surya { 361394c2d0ebSMatthew Ahrens multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA]; 361494c2d0ebSMatthew Ahrens multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA]; 3615244781f1SPrakash Surya int data_idx = multilist_get_random_index(data_ml); 3616244781f1SPrakash Surya int meta_idx = multilist_get_random_index(meta_ml); 3617244781f1SPrakash Surya multilist_sublist_t *data_mls; 3618244781f1SPrakash Surya multilist_sublist_t *meta_mls; 3619244781f1SPrakash Surya arc_buf_contents_t type; 3620244781f1SPrakash Surya arc_buf_hdr_t *data_hdr; 3621244781f1SPrakash Surya arc_buf_hdr_t *meta_hdr; 3622244781f1SPrakash Surya 3623244781f1SPrakash Surya /* 3624244781f1SPrakash Surya * We keep the sublist lock until we're finished, to prevent 3625244781f1SPrakash Surya * the headers from being destroyed via arc_evict_state(). 3626244781f1SPrakash Surya */ 3627244781f1SPrakash Surya data_mls = multilist_sublist_lock(data_ml, data_idx); 3628244781f1SPrakash Surya meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3629244781f1SPrakash Surya 3630244781f1SPrakash Surya /* 3631244781f1SPrakash Surya * These two loops are to ensure we skip any markers that 3632244781f1SPrakash Surya * might be at the tail of the lists due to arc_evict_state(). 3633244781f1SPrakash Surya */ 3634244781f1SPrakash Surya 3635244781f1SPrakash Surya for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3636244781f1SPrakash Surya data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3637244781f1SPrakash Surya if (data_hdr->b_spa != 0) 3638244781f1SPrakash Surya break; 3639244781f1SPrakash Surya } 3640244781f1SPrakash Surya 3641244781f1SPrakash Surya for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3642244781f1SPrakash Surya meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3643244781f1SPrakash Surya if (meta_hdr->b_spa != 0) 3644244781f1SPrakash Surya break; 3645244781f1SPrakash Surya } 3646244781f1SPrakash Surya 3647244781f1SPrakash Surya if (data_hdr == NULL && meta_hdr == NULL) { 3648244781f1SPrakash Surya type = ARC_BUFC_DATA; 3649244781f1SPrakash Surya } else if (data_hdr == NULL) { 3650244781f1SPrakash Surya ASSERT3P(meta_hdr, !=, NULL); 3651244781f1SPrakash Surya type = ARC_BUFC_METADATA; 3652244781f1SPrakash Surya } else if (meta_hdr == NULL) { 3653244781f1SPrakash Surya ASSERT3P(data_hdr, !=, NULL); 3654244781f1SPrakash Surya type = ARC_BUFC_DATA; 3655244781f1SPrakash Surya } else { 3656244781f1SPrakash Surya ASSERT3P(data_hdr, !=, NULL); 3657244781f1SPrakash Surya ASSERT3P(meta_hdr, !=, NULL); 3658244781f1SPrakash Surya 3659244781f1SPrakash Surya /* The headers can't be on the sublist without an L1 header */ 3660244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(data_hdr)); 3661244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3662244781f1SPrakash Surya 3663244781f1SPrakash Surya if (data_hdr->b_l1hdr.b_arc_access < 3664244781f1SPrakash Surya meta_hdr->b_l1hdr.b_arc_access) { 3665244781f1SPrakash Surya type = ARC_BUFC_DATA; 3666244781f1SPrakash Surya } else { 3667244781f1SPrakash Surya type = ARC_BUFC_METADATA; 3668244781f1SPrakash Surya } 3669244781f1SPrakash Surya } 3670244781f1SPrakash Surya 3671244781f1SPrakash Surya multilist_sublist_unlock(meta_mls); 3672244781f1SPrakash Surya multilist_sublist_unlock(data_mls); 3673244781f1SPrakash Surya 3674244781f1SPrakash Surya return (type); 3675244781f1SPrakash Surya } 3676244781f1SPrakash Surya 3677244781f1SPrakash Surya /* 3678244781f1SPrakash Surya * Evict buffers from the cache, such that arc_size is capped by arc_c. 3679244781f1SPrakash Surya */ 3680244781f1SPrakash Surya static uint64_t 3681fa9e4066Sahrens arc_adjust(void) 3682fa9e4066Sahrens { 3683244781f1SPrakash Surya uint64_t total_evicted = 0; 3684244781f1SPrakash Surya uint64_t bytes; 3685244781f1SPrakash Surya int64_t target; 3686fa9e4066Sahrens 36875a98e54bSBrendan Gregg - Sun Microsystems /* 3688244781f1SPrakash Surya * If we're over arc_meta_limit, we want to correct that before 3689244781f1SPrakash Surya * potentially evicting data buffers below. 36905a98e54bSBrendan Gregg - Sun Microsystems */ 3691244781f1SPrakash Surya total_evicted += arc_adjust_meta(); 36925a98e54bSBrendan Gregg - Sun Microsystems 3693244781f1SPrakash Surya /* 3694244781f1SPrakash Surya * Adjust MRU size 3695244781f1SPrakash Surya * 3696244781f1SPrakash Surya * If we're over the target cache size, we want to evict enough 3697244781f1SPrakash Surya * from the list to get back to our target size. We don't want 3698244781f1SPrakash Surya * to evict too much from the MRU, such that it drops below 3699244781f1SPrakash Surya * arc_p. So, if we're over our target cache size more than 3700244781f1SPrakash Surya * the MRU is over arc_p, we'll evict enough to get back to 3701244781f1SPrakash Surya * arc_p here, and then evict more from the MFU below. 3702244781f1SPrakash Surya */ 3703244781f1SPrakash Surya target = MIN((int64_t)(arc_size - arc_c), 37042fd872a7SPrakash Surya (int64_t)(refcount_count(&arc_anon->arcs_size) + 37052fd872a7SPrakash Surya refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 3706fa9e4066Sahrens 3707244781f1SPrakash Surya /* 3708244781f1SPrakash Surya * If we're below arc_meta_min, always prefer to evict data. 3709244781f1SPrakash Surya * Otherwise, try to satisfy the requested number of bytes to 3710244781f1SPrakash Surya * evict from the type which contains older buffers; in an 3711244781f1SPrakash Surya * effort to keep newer buffers in the cache regardless of their 3712244781f1SPrakash Surya * type. If we cannot satisfy the number of bytes from this 3713244781f1SPrakash Surya * type, spill over into the next type. 3714244781f1SPrakash Surya */ 3715244781f1SPrakash Surya if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3716244781f1SPrakash Surya arc_meta_used > arc_meta_min) { 3717244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3718244781f1SPrakash Surya total_evicted += bytes; 37190e8c6158Smaybee 3720244781f1SPrakash Surya /* 3721244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3722244781f1SPrakash Surya * metadata, we try to get the rest from data. 3723244781f1SPrakash Surya */ 3724244781f1SPrakash Surya target -= bytes; 3725244781f1SPrakash Surya 3726244781f1SPrakash Surya total_evicted += 3727244781f1SPrakash Surya arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3728244781f1SPrakash Surya } else { 3729244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3730244781f1SPrakash Surya total_evicted += bytes; 3731244781f1SPrakash Surya 3732244781f1SPrakash Surya /* 3733244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3734244781f1SPrakash Surya * data, we try to get the rest from metadata. 3735244781f1SPrakash Surya */ 3736244781f1SPrakash Surya target -= bytes; 3737244781f1SPrakash Surya 3738244781f1SPrakash Surya total_evicted += 3739244781f1SPrakash Surya arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3740fa9e4066Sahrens } 3741fa9e4066Sahrens 37425a98e54bSBrendan Gregg - Sun Microsystems /* 37435a98e54bSBrendan Gregg - Sun Microsystems * Adjust MFU size 3744244781f1SPrakash Surya * 3745244781f1SPrakash Surya * Now that we've tried to evict enough from the MRU to get its 3746244781f1SPrakash Surya * size back to arc_p, if we're still above the target cache 3747244781f1SPrakash Surya * size, we evict the rest from the MFU. 37485a98e54bSBrendan Gregg - Sun Microsystems */ 3749244781f1SPrakash Surya target = arc_size - arc_c; 3750fa9e4066Sahrens 375131c46cf2SAlek Pinchuk if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3752244781f1SPrakash Surya arc_meta_used > arc_meta_min) { 3753244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3754244781f1SPrakash Surya total_evicted += bytes; 37555a98e54bSBrendan Gregg - Sun Microsystems 3756244781f1SPrakash Surya /* 3757244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3758244781f1SPrakash Surya * metadata, we try to get the rest from data. 3759244781f1SPrakash Surya */ 3760244781f1SPrakash Surya target -= bytes; 3761244781f1SPrakash Surya 3762244781f1SPrakash Surya total_evicted += 3763244781f1SPrakash Surya arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3764244781f1SPrakash Surya } else { 3765244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3766244781f1SPrakash Surya total_evicted += bytes; 3767244781f1SPrakash Surya 3768244781f1SPrakash Surya /* 3769244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3770244781f1SPrakash Surya * data, we try to get the rest from data. 3771244781f1SPrakash Surya */ 3772244781f1SPrakash Surya target -= bytes; 3773fa9e4066Sahrens 3774244781f1SPrakash Surya total_evicted += 3775244781f1SPrakash Surya arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 37765a98e54bSBrendan Gregg - Sun Microsystems } 3777fa9e4066Sahrens 37785a98e54bSBrendan Gregg - Sun Microsystems /* 37795a98e54bSBrendan Gregg - Sun Microsystems * Adjust ghost lists 3780244781f1SPrakash Surya * 3781244781f1SPrakash Surya * In addition to the above, the ARC also defines target values 3782244781f1SPrakash Surya * for the ghost lists. The sum of the mru list and mru ghost 3783244781f1SPrakash Surya * list should never exceed the target size of the cache, and 3784244781f1SPrakash Surya * the sum of the mru list, mfu list, mru ghost list, and mfu 3785244781f1SPrakash Surya * ghost list should never exceed twice the target size of the 3786244781f1SPrakash Surya * cache. The following logic enforces these limits on the ghost 3787244781f1SPrakash Surya * caches, and evicts from them as needed. 37885a98e54bSBrendan Gregg - Sun Microsystems */ 37892fd872a7SPrakash Surya target = refcount_count(&arc_mru->arcs_size) + 37902fd872a7SPrakash Surya refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3791fa9e4066Sahrens 3792244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3793244781f1SPrakash Surya total_evicted += bytes; 3794fa9e4066Sahrens 3795244781f1SPrakash Surya target -= bytes; 37960e8c6158Smaybee 3797244781f1SPrakash Surya total_evicted += 3798244781f1SPrakash Surya arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 37995a98e54bSBrendan Gregg - Sun Microsystems 3800244781f1SPrakash Surya /* 3801244781f1SPrakash Surya * We assume the sum of the mru list and mfu list is less than 3802244781f1SPrakash Surya * or equal to arc_c (we enforced this above), which means we 3803244781f1SPrakash Surya * can use the simpler of the two equations below: 3804244781f1SPrakash Surya * 3805244781f1SPrakash Surya * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3806244781f1SPrakash Surya * mru ghost + mfu ghost <= arc_c 3807244781f1SPrakash Surya */ 38082fd872a7SPrakash Surya target = refcount_count(&arc_mru_ghost->arcs_size) + 38092fd872a7SPrakash Surya refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3810244781f1SPrakash Surya 3811244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3812244781f1SPrakash Surya total_evicted += bytes; 3813244781f1SPrakash Surya 3814244781f1SPrakash Surya target -= bytes; 3815244781f1SPrakash Surya 3816244781f1SPrakash Surya total_evicted += 3817244781f1SPrakash Surya arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3818244781f1SPrakash Surya 3819244781f1SPrakash Surya return (total_evicted); 3820fa9e4066Sahrens } 3821fa9e4066Sahrens 3822fa9e4066Sahrens void 3823244781f1SPrakash Surya arc_flush(spa_t *spa, boolean_t retry) 3824fa9e4066Sahrens { 3825ac05c741SMark Maybee uint64_t guid = 0; 3826ac05c741SMark Maybee 3827244781f1SPrakash Surya /* 3828dcbf3bd6SGeorge Wilson * If retry is B_TRUE, a spa must not be specified since we have 3829244781f1SPrakash Surya * no good way to determine if all of a spa's buffers have been 3830244781f1SPrakash Surya * evicted from an arc state. 3831244781f1SPrakash Surya */ 3832244781f1SPrakash Surya ASSERT(!retry || spa == 0); 3833244781f1SPrakash Surya 383489c86e32SChris Williamson if (spa != NULL) 3835e9103aaeSGarrett D'Amore guid = spa_load_guid(spa); 3836ac05c741SMark Maybee 3837244781f1SPrakash Surya (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3838244781f1SPrakash Surya (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3839244781f1SPrakash Surya 3840244781f1SPrakash Surya (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3841244781f1SPrakash Surya (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3842874395d5Smaybee 3843244781f1SPrakash Surya (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3844244781f1SPrakash Surya (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3845244781f1SPrakash Surya 3846244781f1SPrakash Surya (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3847244781f1SPrakash Surya (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3848fa9e4066Sahrens } 3849fa9e4066Sahrens 3850fa9e4066Sahrens void 38512ec99e3eSMatthew Ahrens arc_shrink(int64_t to_free) 3852fa9e4066Sahrens { 385344cb6abcSbmc if (arc_c > arc_c_min) { 3854fa9e4066Sahrens 385544cb6abcSbmc if (arc_c > arc_c_min + to_free) 385644cb6abcSbmc atomic_add_64(&arc_c, -to_free); 385749e3519aSmaybee else 385844cb6abcSbmc arc_c = arc_c_min; 385944cb6abcSbmc 386044cb6abcSbmc atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 386144cb6abcSbmc if (arc_c > arc_size) 386244cb6abcSbmc arc_c = MAX(arc_size, arc_c_min); 386344cb6abcSbmc if (arc_p > arc_c) 386444cb6abcSbmc arc_p = (arc_c >> 1); 386544cb6abcSbmc ASSERT(arc_c >= arc_c_min); 386644cb6abcSbmc ASSERT((int64_t)arc_p >= 0); 386749e3519aSmaybee } 3868fa9e4066Sahrens 386944cb6abcSbmc if (arc_size > arc_c) 3870244781f1SPrakash Surya (void) arc_adjust(); 3871fa9e4066Sahrens } 3872fa9e4066Sahrens 38732ec99e3eSMatthew Ahrens typedef enum free_memory_reason_t { 38742ec99e3eSMatthew Ahrens FMR_UNKNOWN, 38752ec99e3eSMatthew Ahrens FMR_NEEDFREE, 38762ec99e3eSMatthew Ahrens FMR_LOTSFREE, 38772ec99e3eSMatthew Ahrens FMR_SWAPFS_MINFREE, 38782ec99e3eSMatthew Ahrens FMR_PAGES_PP_MAXIMUM, 38792ec99e3eSMatthew Ahrens FMR_HEAP_ARENA, 38802ec99e3eSMatthew Ahrens FMR_ZIO_ARENA, 38812ec99e3eSMatthew Ahrens } free_memory_reason_t; 38822ec99e3eSMatthew Ahrens 38832ec99e3eSMatthew Ahrens int64_t last_free_memory; 38842ec99e3eSMatthew Ahrens free_memory_reason_t last_free_reason; 38852ec99e3eSMatthew Ahrens 388694dd93aeSGeorge Wilson /* 38872ec99e3eSMatthew Ahrens * Additional reserve of pages for pp_reserve. 388894dd93aeSGeorge Wilson */ 38892ec99e3eSMatthew Ahrens int64_t arc_pages_pp_reserve = 64; 3890fa9e4066Sahrens 38912ec99e3eSMatthew Ahrens /* 38922ec99e3eSMatthew Ahrens * Additional reserve of pages for swapfs. 38932ec99e3eSMatthew Ahrens */ 38942ec99e3eSMatthew Ahrens int64_t arc_swapfs_reserve = 64; 38953cff2f43Sstans 38962ec99e3eSMatthew Ahrens /* 38972ec99e3eSMatthew Ahrens * Return the amount of memory that can be consumed before reclaim will be 38982ec99e3eSMatthew Ahrens * needed. Positive if there is sufficient free memory, negative indicates 38992ec99e3eSMatthew Ahrens * the amount of memory that needs to be freed up. 39002ec99e3eSMatthew Ahrens */ 39012ec99e3eSMatthew Ahrens static int64_t 39022ec99e3eSMatthew Ahrens arc_available_memory(void) 39032ec99e3eSMatthew Ahrens { 39042ec99e3eSMatthew Ahrens int64_t lowest = INT64_MAX; 39052ec99e3eSMatthew Ahrens int64_t n; 39062ec99e3eSMatthew Ahrens free_memory_reason_t r = FMR_UNKNOWN; 39073cff2f43Sstans 39082ec99e3eSMatthew Ahrens #ifdef _KERNEL 39092ec99e3eSMatthew Ahrens if (needfree > 0) { 39102ec99e3eSMatthew Ahrens n = PAGESIZE * (-needfree); 39112ec99e3eSMatthew Ahrens if (n < lowest) { 39122ec99e3eSMatthew Ahrens lowest = n; 39132ec99e3eSMatthew Ahrens r = FMR_NEEDFREE; 39142ec99e3eSMatthew Ahrens } 39152ec99e3eSMatthew Ahrens } 3916fa9e4066Sahrens 3917fa9e4066Sahrens /* 3918fa9e4066Sahrens * check that we're out of range of the pageout scanner. It starts to 3919fa9e4066Sahrens * schedule paging if freemem is less than lotsfree and needfree. 3920fa9e4066Sahrens * lotsfree is the high-water mark for pageout, and needfree is the 3921fa9e4066Sahrens * number of needed free pages. We add extra pages here to make sure 3922fa9e4066Sahrens * the scanner doesn't start up while we're freeing memory. 3923fa9e4066Sahrens */ 39242ec99e3eSMatthew Ahrens n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 39252ec99e3eSMatthew Ahrens if (n < lowest) { 39262ec99e3eSMatthew Ahrens lowest = n; 39272ec99e3eSMatthew Ahrens r = FMR_LOTSFREE; 39282ec99e3eSMatthew Ahrens } 3929fa9e4066Sahrens 3930fa9e4066Sahrens /* 3931fa9e4066Sahrens * check to make sure that swapfs has enough space so that anon 3932fa94a07fSbrendan * reservations can still succeed. anon_resvmem() checks that the 3933fa9e4066Sahrens * availrmem is greater than swapfs_minfree, and the number of reserved 3934fa9e4066Sahrens * swap pages. We also add a bit of extra here just to prevent 3935fa9e4066Sahrens * circumstances from getting really dire. 3936fa9e4066Sahrens */ 39372ec99e3eSMatthew Ahrens n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 39382ec99e3eSMatthew Ahrens desfree - arc_swapfs_reserve); 39392ec99e3eSMatthew Ahrens if (n < lowest) { 39402ec99e3eSMatthew Ahrens lowest = n; 39412ec99e3eSMatthew Ahrens r = FMR_SWAPFS_MINFREE; 39422ec99e3eSMatthew Ahrens } 39432ec99e3eSMatthew Ahrens 3944fa9e4066Sahrens 3945cf746768SBryan Cantrill /* 3946cf746768SBryan Cantrill * Check that we have enough availrmem that memory locking (e.g., via 3947cf746768SBryan Cantrill * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3948cf746768SBryan Cantrill * stores the number of pages that cannot be locked; when availrmem 3949cf746768SBryan Cantrill * drops below pages_pp_maximum, page locking mechanisms such as 3950cf746768SBryan Cantrill * page_pp_lock() will fail.) 3951cf746768SBryan Cantrill */ 39522ec99e3eSMatthew Ahrens n = PAGESIZE * (availrmem - pages_pp_maximum - 39532ec99e3eSMatthew Ahrens arc_pages_pp_reserve); 39542ec99e3eSMatthew Ahrens if (n < lowest) { 39552ec99e3eSMatthew Ahrens lowest = n; 39562ec99e3eSMatthew Ahrens r = FMR_PAGES_PP_MAXIMUM; 39572ec99e3eSMatthew Ahrens } 3958cf746768SBryan Cantrill 39595dc8af33Smaybee #if defined(__i386) 3960fa9e4066Sahrens /* 3961fa9e4066Sahrens * If we're on an i386 platform, it's possible that we'll exhaust the 3962fa9e4066Sahrens * kernel heap space before we ever run out of available physical 3963fa9e4066Sahrens * memory. Most checks of the size of the heap_area compare against 3964fa9e4066Sahrens * tune.t_minarmem, which is the minimum available real memory that we 3965fa9e4066Sahrens * can have in the system. However, this is generally fixed at 25 pages 3966fa9e4066Sahrens * which is so low that it's useless. In this comparison, we seek to 3967fa9e4066Sahrens * calculate the total heap-size, and reclaim if more than 3/4ths of the 3968fa94a07fSbrendan * heap is allocated. (Or, in the calculation, if less than 1/4th is 3969fa9e4066Sahrens * free) 3970fa9e4066Sahrens */ 39710dd053d7SPrakash Surya n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 39722ec99e3eSMatthew Ahrens (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 39732ec99e3eSMatthew Ahrens if (n < lowest) { 39742ec99e3eSMatthew Ahrens lowest = n; 39752ec99e3eSMatthew Ahrens r = FMR_HEAP_ARENA; 39762ec99e3eSMatthew Ahrens } 3977fa9e4066Sahrens #endif 3978fa9e4066Sahrens 397994dd93aeSGeorge Wilson /* 398094dd93aeSGeorge Wilson * If zio data pages are being allocated out of a separate heap segment, 398194dd93aeSGeorge Wilson * then enforce that the size of available vmem for this arena remains 39820dd053d7SPrakash Surya * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free. 398394dd93aeSGeorge Wilson * 39840dd053d7SPrakash Surya * Note that reducing the arc_zio_arena_free_shift keeps more virtual 39850dd053d7SPrakash Surya * memory (in the zio_arena) free, which can avoid memory 39860dd053d7SPrakash Surya * fragmentation issues. 398794dd93aeSGeorge Wilson */ 39882ec99e3eSMatthew Ahrens if (zio_arena != NULL) { 39890dd053d7SPrakash Surya n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 39900dd053d7SPrakash Surya (vmem_size(zio_arena, VMEM_ALLOC) >> 39910dd053d7SPrakash Surya arc_zio_arena_free_shift); 39922ec99e3eSMatthew Ahrens if (n < lowest) { 39932ec99e3eSMatthew Ahrens lowest = n; 39942ec99e3eSMatthew Ahrens r = FMR_ZIO_ARENA; 39952ec99e3eSMatthew Ahrens } 39962ec99e3eSMatthew Ahrens } 3997fa9e4066Sahrens #else 39982ec99e3eSMatthew Ahrens /* Every 100 calls, free a small amount */ 3999fa9e4066Sahrens if (spa_get_random(100) == 0) 40002ec99e3eSMatthew Ahrens lowest = -1024; 4001fa9e4066Sahrens #endif 40022ec99e3eSMatthew Ahrens 40032ec99e3eSMatthew Ahrens last_free_memory = lowest; 40042ec99e3eSMatthew Ahrens last_free_reason = r; 40052ec99e3eSMatthew Ahrens 40062ec99e3eSMatthew Ahrens return (lowest); 40072ec99e3eSMatthew Ahrens } 40082ec99e3eSMatthew Ahrens 40092ec99e3eSMatthew Ahrens 40102ec99e3eSMatthew Ahrens /* 40112ec99e3eSMatthew Ahrens * Determine if the system is under memory pressure and is asking 4012dcbf3bd6SGeorge Wilson * to reclaim memory. A return value of B_TRUE indicates that the system 40132ec99e3eSMatthew Ahrens * is under memory pressure and that the arc should adjust accordingly. 40142ec99e3eSMatthew Ahrens */ 40152ec99e3eSMatthew Ahrens static boolean_t 40162ec99e3eSMatthew Ahrens arc_reclaim_needed(void) 40172ec99e3eSMatthew Ahrens { 40182ec99e3eSMatthew Ahrens return (arc_available_memory() < 0); 4019fa9e4066Sahrens } 4020fa9e4066Sahrens 4021fa9e4066Sahrens static void 40222ec99e3eSMatthew Ahrens arc_kmem_reap_now(void) 4023fa9e4066Sahrens { 4024fa9e4066Sahrens size_t i; 4025fa9e4066Sahrens kmem_cache_t *prev_cache = NULL; 4026ad23a2dbSjohansen kmem_cache_t *prev_data_cache = NULL; 4027fa9e4066Sahrens extern kmem_cache_t *zio_buf_cache[]; 4028ad23a2dbSjohansen extern kmem_cache_t *zio_data_buf_cache[]; 402983803b51SGeorge Wilson extern kmem_cache_t *range_seg_cache; 4030*770499e1SDan Kimmel extern kmem_cache_t *abd_chunk_cache; 4031fa9e4066Sahrens 4032033f9833Sek #ifdef _KERNEL 40330e8c6158Smaybee if (arc_meta_used >= arc_meta_limit) { 40340e8c6158Smaybee /* 40350e8c6158Smaybee * We are exceeding our meta-data cache limit. 40360e8c6158Smaybee * Purge some DNLC entries to release holds on meta-data. 40370e8c6158Smaybee */ 40380e8c6158Smaybee dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 40390e8c6158Smaybee } 40405dc8af33Smaybee #if defined(__i386) 40415dc8af33Smaybee /* 40425dc8af33Smaybee * Reclaim unused memory from all kmem caches. 40435dc8af33Smaybee */ 40445dc8af33Smaybee kmem_reap(); 40455dc8af33Smaybee #endif 4046033f9833Sek #endif 4047033f9833Sek 4048fa9e4066Sahrens for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 4049fa9e4066Sahrens if (zio_buf_cache[i] != prev_cache) { 4050fa9e4066Sahrens prev_cache = zio_buf_cache[i]; 4051fa9e4066Sahrens kmem_cache_reap_now(zio_buf_cache[i]); 4052fa9e4066Sahrens } 4053ad23a2dbSjohansen if (zio_data_buf_cache[i] != prev_data_cache) { 4054ad23a2dbSjohansen prev_data_cache = zio_data_buf_cache[i]; 4055ad23a2dbSjohansen kmem_cache_reap_now(zio_data_buf_cache[i]); 4056ad23a2dbSjohansen } 4057fa9e4066Sahrens } 4058*770499e1SDan Kimmel kmem_cache_reap_now(abd_chunk_cache); 4059ea8dc4b6Seschrock kmem_cache_reap_now(buf_cache); 406089c86e32SChris Williamson kmem_cache_reap_now(hdr_full_cache); 406189c86e32SChris Williamson kmem_cache_reap_now(hdr_l2only_cache); 406283803b51SGeorge Wilson kmem_cache_reap_now(range_seg_cache); 406394dd93aeSGeorge Wilson 40642ec99e3eSMatthew Ahrens if (zio_arena != NULL) { 40652ec99e3eSMatthew Ahrens /* 40662ec99e3eSMatthew Ahrens * Ask the vmem arena to reclaim unused memory from its 40672ec99e3eSMatthew Ahrens * quantum caches. 40682ec99e3eSMatthew Ahrens */ 406994dd93aeSGeorge Wilson vmem_qcache_reap(zio_arena); 40702ec99e3eSMatthew Ahrens } 4071fa9e4066Sahrens } 4072fa9e4066Sahrens 4073244781f1SPrakash Surya /* 4074*770499e1SDan Kimmel * Threads can block in arc_get_data_impl() waiting for this thread to evict 4075244781f1SPrakash Surya * enough data and signal them to proceed. When this happens, the threads in 4076*770499e1SDan Kimmel * arc_get_data_impl() are sleeping while holding the hash lock for their 4077244781f1SPrakash Surya * particular arc header. Thus, we must be careful to never sleep on a 4078244781f1SPrakash Surya * hash lock in this thread. This is to prevent the following deadlock: 4079244781f1SPrakash Surya * 4080*770499e1SDan Kimmel * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L", 4081244781f1SPrakash Surya * waiting for the reclaim thread to signal it. 4082244781f1SPrakash Surya * 4083244781f1SPrakash Surya * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 4084244781f1SPrakash Surya * fails, and goes to sleep forever. 4085244781f1SPrakash Surya * 4086244781f1SPrakash Surya * This possible deadlock is avoided by always acquiring a hash lock 4087244781f1SPrakash Surya * using mutex_tryenter() from arc_reclaim_thread(). 4088244781f1SPrakash Surya */ 4089fa9e4066Sahrens static void 4090fa9e4066Sahrens arc_reclaim_thread(void) 4091fa9e4066Sahrens { 4092a8f6344fSEli Rosenthal hrtime_t growtime = 0; 4093fa9e4066Sahrens callb_cpr_t cpr; 4094fa9e4066Sahrens 4095244781f1SPrakash Surya CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 4096fa9e4066Sahrens 4097244781f1SPrakash Surya mutex_enter(&arc_reclaim_lock); 4098244781f1SPrakash Surya while (!arc_reclaim_thread_exit) { 4099244781f1SPrakash Surya uint64_t evicted = 0; 4100244781f1SPrakash Surya 4101dcbf3bd6SGeorge Wilson /* 4102dcbf3bd6SGeorge Wilson * This is necessary in order for the mdb ::arc dcmd to 4103dcbf3bd6SGeorge Wilson * show up to date information. Since the ::arc command 4104dcbf3bd6SGeorge Wilson * does not call the kstat's update function, without 4105dcbf3bd6SGeorge Wilson * this call, the command may show stale stats for the 4106dcbf3bd6SGeorge Wilson * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 4107dcbf3bd6SGeorge Wilson * with this change, the data might be up to 1 second 4108dcbf3bd6SGeorge Wilson * out of date; but that should suffice. The arc_state_t 4109dcbf3bd6SGeorge Wilson * structures can be queried directly if more accurate 4110dcbf3bd6SGeorge Wilson * information is needed. 4111dcbf3bd6SGeorge Wilson */ 4112dcbf3bd6SGeorge Wilson if (arc_ksp != NULL) 4113dcbf3bd6SGeorge Wilson arc_ksp->ks_update(arc_ksp, KSTAT_READ); 4114dcbf3bd6SGeorge Wilson 4115244781f1SPrakash Surya mutex_exit(&arc_reclaim_lock); 4116244781f1SPrakash Surya 4117405a5a0fSMatthew Ahrens /* 4118405a5a0fSMatthew Ahrens * We call arc_adjust() before (possibly) calling 4119405a5a0fSMatthew Ahrens * arc_kmem_reap_now(), so that we can wake up 4120*770499e1SDan Kimmel * arc_get_data_impl() sooner. 4121405a5a0fSMatthew Ahrens */ 4122405a5a0fSMatthew Ahrens evicted = arc_adjust(); 4123405a5a0fSMatthew Ahrens 4124405a5a0fSMatthew Ahrens int64_t free_memory = arc_available_memory(); 41252ec99e3eSMatthew Ahrens if (free_memory < 0) { 4126fa9e4066Sahrens 41272ec99e3eSMatthew Ahrens arc_no_grow = B_TRUE; 41282ec99e3eSMatthew Ahrens arc_warm = B_TRUE; 4129fa9e4066Sahrens 41302ec99e3eSMatthew Ahrens /* 41312ec99e3eSMatthew Ahrens * Wait at least zfs_grow_retry (default 60) seconds 41322ec99e3eSMatthew Ahrens * before considering growing. 41332ec99e3eSMatthew Ahrens */ 4134a8f6344fSEli Rosenthal growtime = gethrtime() + SEC2NSEC(arc_grow_retry); 4135fa9e4066Sahrens 41362ec99e3eSMatthew Ahrens arc_kmem_reap_now(); 4137fa9e4066Sahrens 41382ec99e3eSMatthew Ahrens /* 41392ec99e3eSMatthew Ahrens * If we are still low on memory, shrink the ARC 41402ec99e3eSMatthew Ahrens * so that we have arc_shrink_min free space. 41412ec99e3eSMatthew Ahrens */ 41422ec99e3eSMatthew Ahrens free_memory = arc_available_memory(); 41432ec99e3eSMatthew Ahrens 41442ec99e3eSMatthew Ahrens int64_t to_free = 41452ec99e3eSMatthew Ahrens (arc_c >> arc_shrink_shift) - free_memory; 41462ec99e3eSMatthew Ahrens if (to_free > 0) { 41472ec99e3eSMatthew Ahrens #ifdef _KERNEL 41482ec99e3eSMatthew Ahrens to_free = MAX(to_free, ptob(needfree)); 41492ec99e3eSMatthew Ahrens #endif 41502ec99e3eSMatthew Ahrens arc_shrink(to_free); 41512ec99e3eSMatthew Ahrens } 41522ec99e3eSMatthew Ahrens } else if (free_memory < arc_c >> arc_no_grow_shift) { 41532ec99e3eSMatthew Ahrens arc_no_grow = B_TRUE; 4154a8f6344fSEli Rosenthal } else if (gethrtime() >= growtime) { 41552ec99e3eSMatthew Ahrens arc_no_grow = B_FALSE; 4156fa9e4066Sahrens } 4157fa9e4066Sahrens 4158244781f1SPrakash Surya mutex_enter(&arc_reclaim_lock); 4159244781f1SPrakash Surya 4160244781f1SPrakash Surya /* 4161244781f1SPrakash Surya * If evicted is zero, we couldn't evict anything via 4162244781f1SPrakash Surya * arc_adjust(). This could be due to hash lock 4163244781f1SPrakash Surya * collisions, but more likely due to the majority of 4164244781f1SPrakash Surya * arc buffers being unevictable. Therefore, even if 4165244781f1SPrakash Surya * arc_size is above arc_c, another pass is unlikely to 4166244781f1SPrakash Surya * be helpful and could potentially cause us to enter an 4167244781f1SPrakash Surya * infinite loop. 4168244781f1SPrakash Surya */ 4169244781f1SPrakash Surya if (arc_size <= arc_c || evicted == 0) { 4170244781f1SPrakash Surya /* 4171244781f1SPrakash Surya * We're either no longer overflowing, or we 4172244781f1SPrakash Surya * can't evict anything more, so we should wake 4173244781f1SPrakash Surya * up any threads before we go to sleep. 4174244781f1SPrakash Surya */ 4175244781f1SPrakash Surya cv_broadcast(&arc_reclaim_waiters_cv); 4176244781f1SPrakash Surya 4177244781f1SPrakash Surya /* 4178244781f1SPrakash Surya * Block until signaled, or after one second (we 4179244781f1SPrakash Surya * might need to perform arc_kmem_reap_now() 4180244781f1SPrakash Surya * even if we aren't being signalled) 4181244781f1SPrakash Surya */ 4182244781f1SPrakash Surya CALLB_CPR_SAFE_BEGIN(&cpr); 4183a8f6344fSEli Rosenthal (void) cv_timedwait_hires(&arc_reclaim_thread_cv, 4184a8f6344fSEli Rosenthal &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 4185244781f1SPrakash Surya CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 4186244781f1SPrakash Surya } 4187244781f1SPrakash Surya } 4188244781f1SPrakash Surya 4189dcbf3bd6SGeorge Wilson arc_reclaim_thread_exit = B_FALSE; 4190244781f1SPrakash Surya cv_broadcast(&arc_reclaim_thread_cv); 4191244781f1SPrakash Surya CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 4192244781f1SPrakash Surya thread_exit(); 4193244781f1SPrakash Surya } 4194244781f1SPrakash Surya 4195ea8dc4b6Seschrock /* 4196ea8dc4b6Seschrock * Adapt arc info given the number of bytes we are trying to add and 4197ea8dc4b6Seschrock * the state that we are comming from. This function is only called 4198ea8dc4b6Seschrock * when we are adding new content to the cache. 4199ea8dc4b6Seschrock */ 4200fa9e4066Sahrens static void 4201ea8dc4b6Seschrock arc_adapt(int bytes, arc_state_t *state) 4202fa9e4066Sahrens { 4203ea8dc4b6Seschrock int mult; 42045a98e54bSBrendan Gregg - Sun Microsystems uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 42052fd872a7SPrakash Surya int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 42062fd872a7SPrakash Surya int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 4207ea8dc4b6Seschrock 4208fa94a07fSbrendan if (state == arc_l2c_only) 4209fa94a07fSbrendan return; 4210fa94a07fSbrendan 4211ea8dc4b6Seschrock ASSERT(bytes > 0); 4212fa9e4066Sahrens /* 4213ea8dc4b6Seschrock * Adapt the target size of the MRU list: 4214ea8dc4b6Seschrock * - if we just hit in the MRU ghost list, then increase 4215ea8dc4b6Seschrock * the target size of the MRU list. 4216ea8dc4b6Seschrock * - if we just hit in the MFU ghost list, then increase 4217ea8dc4b6Seschrock * the target size of the MFU list by decreasing the 4218ea8dc4b6Seschrock * target size of the MRU list. 4219fa9e4066Sahrens */ 422044cb6abcSbmc if (state == arc_mru_ghost) { 42212fd872a7SPrakash Surya mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 42223e4e8481STom Erickson mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 4223ea8dc4b6Seschrock 42245a98e54bSBrendan Gregg - Sun Microsystems arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 422544cb6abcSbmc } else if (state == arc_mfu_ghost) { 42265a98e54bSBrendan Gregg - Sun Microsystems uint64_t delta; 42275a98e54bSBrendan Gregg - Sun Microsystems 42282fd872a7SPrakash Surya mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 42293e4e8481STom Erickson mult = MIN(mult, 10); 4230ea8dc4b6Seschrock 42315a98e54bSBrendan Gregg - Sun Microsystems delta = MIN(bytes * mult, arc_p); 42325a98e54bSBrendan Gregg - Sun Microsystems arc_p = MAX(arc_p_min, arc_p - delta); 4233ea8dc4b6Seschrock } 423444cb6abcSbmc ASSERT((int64_t)arc_p >= 0); 4235fa9e4066Sahrens 4236fa9e4066Sahrens if (arc_reclaim_needed()) { 4237244781f1SPrakash Surya cv_signal(&arc_reclaim_thread_cv); 4238fa9e4066Sahrens return; 4239fa9e4066Sahrens } 4240fa9e4066Sahrens 424144cb6abcSbmc if (arc_no_grow) 4242fa9e4066Sahrens return; 4243fa9e4066Sahrens 424444cb6abcSbmc if (arc_c >= arc_c_max) 4245ea8dc4b6Seschrock return; 4246ea8dc4b6Seschrock 4247fa9e4066Sahrens /* 4248ea8dc4b6Seschrock * If we're within (2 * maxblocksize) bytes of the target 4249ea8dc4b6Seschrock * cache size, increment the target cache size 4250fa9e4066Sahrens */ 425144cb6abcSbmc if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 425244cb6abcSbmc atomic_add_64(&arc_c, (int64_t)bytes); 425344cb6abcSbmc if (arc_c > arc_c_max) 425444cb6abcSbmc arc_c = arc_c_max; 425544cb6abcSbmc else if (state == arc_anon) 425644cb6abcSbmc atomic_add_64(&arc_p, (int64_t)bytes); 425744cb6abcSbmc if (arc_p > arc_c) 425844cb6abcSbmc arc_p = arc_c; 4259fa9e4066Sahrens } 426044cb6abcSbmc ASSERT((int64_t)arc_p >= 0); 4261fa9e4066Sahrens } 4262fa9e4066Sahrens 4263fa9e4066Sahrens /* 4264244781f1SPrakash Surya * Check if arc_size has grown past our upper threshold, determined by 4265244781f1SPrakash Surya * zfs_arc_overflow_shift. 4266fa9e4066Sahrens */ 4267244781f1SPrakash Surya static boolean_t 4268244781f1SPrakash Surya arc_is_overflowing(void) 4269fa9e4066Sahrens { 4270244781f1SPrakash Surya /* Always allow at least one block of overflow */ 4271244781f1SPrakash Surya uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 4272244781f1SPrakash Surya arc_c >> zfs_arc_overflow_shift); 4273fa9e4066Sahrens 4274244781f1SPrakash Surya return (arc_size >= arc_c + overflow); 4275fa9e4066Sahrens } 4276fa9e4066Sahrens 4277*770499e1SDan Kimmel static abd_t * 4278*770499e1SDan Kimmel arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4279*770499e1SDan Kimmel { 4280*770499e1SDan Kimmel arc_buf_contents_t type = arc_buf_type(hdr); 4281*770499e1SDan Kimmel 4282*770499e1SDan Kimmel arc_get_data_impl(hdr, size, tag); 4283*770499e1SDan Kimmel if (type == ARC_BUFC_METADATA) { 4284*770499e1SDan Kimmel return (abd_alloc(size, B_TRUE)); 4285*770499e1SDan Kimmel } else { 4286*770499e1SDan Kimmel ASSERT(type == ARC_BUFC_DATA); 4287*770499e1SDan Kimmel return (abd_alloc(size, B_FALSE)); 4288*770499e1SDan Kimmel } 4289*770499e1SDan Kimmel } 4290*770499e1SDan Kimmel 4291*770499e1SDan Kimmel static void * 4292*770499e1SDan Kimmel arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4293*770499e1SDan Kimmel { 4294*770499e1SDan Kimmel arc_buf_contents_t type = arc_buf_type(hdr); 4295*770499e1SDan Kimmel 4296*770499e1SDan Kimmel arc_get_data_impl(hdr, size, tag); 4297*770499e1SDan Kimmel if (type == ARC_BUFC_METADATA) { 4298*770499e1SDan Kimmel return (zio_buf_alloc(size)); 4299*770499e1SDan Kimmel } else { 4300*770499e1SDan Kimmel ASSERT(type == ARC_BUFC_DATA); 4301*770499e1SDan Kimmel return (zio_data_buf_alloc(size)); 4302*770499e1SDan Kimmel } 4303*770499e1SDan Kimmel } 4304*770499e1SDan Kimmel 4305fa9e4066Sahrens /* 4306dcbf3bd6SGeorge Wilson * Allocate a block and return it to the caller. If we are hitting the 4307dcbf3bd6SGeorge Wilson * hard limit for the cache size, we must sleep, waiting for the eviction 4308dcbf3bd6SGeorge Wilson * thread to catch up. If we're past the target size but below the hard 4309dcbf3bd6SGeorge Wilson * limit, we'll only signal the reclaim thread and continue on. 4310fa9e4066Sahrens */ 4311*770499e1SDan Kimmel static void 4312*770499e1SDan Kimmel arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4313fa9e4066Sahrens { 4314*770499e1SDan Kimmel arc_state_t *state = hdr->b_l1hdr.b_state; 4315*770499e1SDan Kimmel arc_buf_contents_t type = arc_buf_type(hdr); 4316fa9e4066Sahrens 431744eda4d7Smaybee arc_adapt(size, state); 4318fa9e4066Sahrens 431944eda4d7Smaybee /* 4320244781f1SPrakash Surya * If arc_size is currently overflowing, and has grown past our 4321244781f1SPrakash Surya * upper limit, we must be adding data faster than the evict 4322244781f1SPrakash Surya * thread can evict. Thus, to ensure we don't compound the 4323244781f1SPrakash Surya * problem by adding more data and forcing arc_size to grow even 4324244781f1SPrakash Surya * further past it's target size, we halt and wait for the 4325244781f1SPrakash Surya * eviction thread to catch up. 4326244781f1SPrakash Surya * 4327244781f1SPrakash Surya * It's also possible that the reclaim thread is unable to evict 4328244781f1SPrakash Surya * enough buffers to get arc_size below the overflow limit (e.g. 4329244781f1SPrakash Surya * due to buffers being un-evictable, or hash lock collisions). 4330244781f1SPrakash Surya * In this case, we want to proceed regardless if we're 4331244781f1SPrakash Surya * overflowing; thus we don't use a while loop here. 433244eda4d7Smaybee */ 4333244781f1SPrakash Surya if (arc_is_overflowing()) { 4334244781f1SPrakash Surya mutex_enter(&arc_reclaim_lock); 4335244781f1SPrakash Surya 4336244781f1SPrakash Surya /* 4337244781f1SPrakash Surya * Now that we've acquired the lock, we may no longer be 4338244781f1SPrakash Surya * over the overflow limit, lets check. 4339244781f1SPrakash Surya * 4340244781f1SPrakash Surya * We're ignoring the case of spurious wake ups. If that 4341244781f1SPrakash Surya * were to happen, it'd let this thread consume an ARC 4342244781f1SPrakash Surya * buffer before it should have (i.e. before we're under 4343244781f1SPrakash Surya * the overflow limit and were signalled by the reclaim 4344244781f1SPrakash Surya * thread). As long as that is a rare occurrence, it 4345244781f1SPrakash Surya * shouldn't cause any harm. 4346244781f1SPrakash Surya */ 4347244781f1SPrakash Surya if (arc_is_overflowing()) { 4348244781f1SPrakash Surya cv_signal(&arc_reclaim_thread_cv); 4349244781f1SPrakash Surya cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 4350ad23a2dbSjohansen } 4351244781f1SPrakash Surya 4352244781f1SPrakash Surya mutex_exit(&arc_reclaim_lock); 435344eda4d7Smaybee } 435444eda4d7Smaybee 4355dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 4356244781f1SPrakash Surya if (type == ARC_BUFC_METADATA) { 4357244781f1SPrakash Surya arc_space_consume(size, ARC_SPACE_META); 4358fa9e4066Sahrens } else { 4359244781f1SPrakash Surya arc_space_consume(size, ARC_SPACE_DATA); 436044eda4d7Smaybee } 4361244781f1SPrakash Surya 436244eda4d7Smaybee /* 436344eda4d7Smaybee * Update the state size. Note that ghost states have a 436444eda4d7Smaybee * "ghost size" and so don't need to be updated. 436544eda4d7Smaybee */ 4366dcbf3bd6SGeorge Wilson if (!GHOST_STATE(state)) { 436744eda4d7Smaybee 4368dcbf3bd6SGeorge Wilson (void) refcount_add_many(&state->arcs_size, size, tag); 4369244781f1SPrakash Surya 4370244781f1SPrakash Surya /* 4371244781f1SPrakash Surya * If this is reached via arc_read, the link is 4372244781f1SPrakash Surya * protected by the hash lock. If reached via 4373244781f1SPrakash Surya * arc_buf_alloc, the header should not be accessed by 4374244781f1SPrakash Surya * any other thread. And, if reached via arc_read_done, 4375244781f1SPrakash Surya * the hash lock will protect it if it's found in the 4376244781f1SPrakash Surya * hash table; otherwise no other thread should be 4377244781f1SPrakash Surya * trying to [add|remove]_reference it. 4378244781f1SPrakash Surya */ 4379244781f1SPrakash Surya if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 438089c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4381dcbf3bd6SGeorge Wilson (void) refcount_add_many(&state->arcs_esize[type], 4382dcbf3bd6SGeorge Wilson size, tag); 4383fa9e4066Sahrens } 4384dcbf3bd6SGeorge Wilson 4385641fbdaeSmaybee /* 4386641fbdaeSmaybee * If we are growing the cache, and we are adding anonymous 438744cb6abcSbmc * data, and we have outgrown arc_p, update arc_p 4388641fbdaeSmaybee */ 438989c86e32SChris Williamson if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 43902fd872a7SPrakash Surya (refcount_count(&arc_anon->arcs_size) + 43912fd872a7SPrakash Surya refcount_count(&arc_mru->arcs_size) > arc_p)) 439244cb6abcSbmc arc_p = MIN(arc_c, arc_p + size); 4393fa9e4066Sahrens } 4394*770499e1SDan Kimmel } 4395*770499e1SDan Kimmel 4396*770499e1SDan Kimmel static void 4397*770499e1SDan Kimmel arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag) 4398*770499e1SDan Kimmel { 4399*770499e1SDan Kimmel arc_free_data_impl(hdr, size, tag); 4400*770499e1SDan Kimmel abd_free(abd); 4401*770499e1SDan Kimmel } 4402*770499e1SDan Kimmel 4403*770499e1SDan Kimmel static void 4404*770499e1SDan Kimmel arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag) 4405*770499e1SDan Kimmel { 4406*770499e1SDan Kimmel arc_buf_contents_t type = arc_buf_type(hdr); 4407*770499e1SDan Kimmel 4408*770499e1SDan Kimmel arc_free_data_impl(hdr, size, tag); 4409*770499e1SDan Kimmel if (type == ARC_BUFC_METADATA) { 4410*770499e1SDan Kimmel zio_buf_free(buf, size); 4411*770499e1SDan Kimmel } else { 4412*770499e1SDan Kimmel ASSERT(type == ARC_BUFC_DATA); 4413*770499e1SDan Kimmel zio_data_buf_free(buf, size); 4414*770499e1SDan Kimmel } 4415dcbf3bd6SGeorge Wilson } 4416dcbf3bd6SGeorge Wilson 4417dcbf3bd6SGeorge Wilson /* 4418dcbf3bd6SGeorge Wilson * Free the arc data buffer. 4419dcbf3bd6SGeorge Wilson */ 4420dcbf3bd6SGeorge Wilson static void 4421*770499e1SDan Kimmel arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4422dcbf3bd6SGeorge Wilson { 4423dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 4424dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 4425dcbf3bd6SGeorge Wilson 4426dcbf3bd6SGeorge Wilson /* protected by hash lock, if in the hash table */ 4427dcbf3bd6SGeorge Wilson if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4428dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4429dcbf3bd6SGeorge Wilson ASSERT(state != arc_anon && state != arc_l2c_only); 4430dcbf3bd6SGeorge Wilson 4431dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 4432dcbf3bd6SGeorge Wilson size, tag); 4433dcbf3bd6SGeorge Wilson } 4434dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_size, size, tag); 4435dcbf3bd6SGeorge Wilson 4436dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 4437dcbf3bd6SGeorge Wilson if (type == ARC_BUFC_METADATA) { 4438dcbf3bd6SGeorge Wilson arc_space_return(size, ARC_SPACE_META); 4439dcbf3bd6SGeorge Wilson } else { 4440dcbf3bd6SGeorge Wilson ASSERT(type == ARC_BUFC_DATA); 4441dcbf3bd6SGeorge Wilson arc_space_return(size, ARC_SPACE_DATA); 4442dcbf3bd6SGeorge Wilson } 4443fa9e4066Sahrens } 4444fa9e4066Sahrens 4445fa9e4066Sahrens /* 4446fa9e4066Sahrens * This routine is called whenever a buffer is accessed. 4447ea8dc4b6Seschrock * NOTE: the hash lock is dropped in this function. 4448fa9e4066Sahrens */ 4449fa9e4066Sahrens static void 44507adb730bSGeorge Wilson arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 4451fa9e4066Sahrens { 4452d3d50737SRafael Vanoni clock_t now; 4453d3d50737SRafael Vanoni 4454fa9e4066Sahrens ASSERT(MUTEX_HELD(hash_lock)); 445589c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 4456fa9e4066Sahrens 445789c86e32SChris Williamson if (hdr->b_l1hdr.b_state == arc_anon) { 4458fa9e4066Sahrens /* 4459fa9e4066Sahrens * This buffer is not in the cache, and does not 4460fa9e4066Sahrens * appear in our "ghost" list. Add the new buffer 4461fa9e4066Sahrens * to the MRU state. 4462fa9e4066Sahrens */ 4463fa9e4066Sahrens 446489c86e32SChris Williamson ASSERT0(hdr->b_l1hdr.b_arc_access); 446589c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 44667adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 44677adb730bSGeorge Wilson arc_change_state(arc_mru, hdr, hash_lock); 4468fa9e4066Sahrens 446989c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mru) { 4470d3d50737SRafael Vanoni now = ddi_get_lbolt(); 4471d3d50737SRafael Vanoni 4472fa9e4066Sahrens /* 447313506d1eSmaybee * If this buffer is here because of a prefetch, then either: 447413506d1eSmaybee * - clear the flag if this is a "referencing" read 447513506d1eSmaybee * (any subsequent access will bump this into the MFU state). 447613506d1eSmaybee * or 447713506d1eSmaybee * - move the buffer to the head of the list if this is 447813506d1eSmaybee * another prefetch (to make it less likely to be evicted). 4479fa9e4066Sahrens */ 448089c86e32SChris Williamson if (HDR_PREFETCH(hdr)) { 448189c86e32SChris Williamson if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4482244781f1SPrakash Surya /* link protected by hash lock */ 4483244781f1SPrakash Surya ASSERT(multilist_link_active( 448489c86e32SChris Williamson &hdr->b_l1hdr.b_arc_node)); 448513506d1eSmaybee } else { 4486dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 448744cb6abcSbmc ARCSTAT_BUMP(arcstat_mru_hits); 448813506d1eSmaybee } 448989c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = now; 4490fa9e4066Sahrens return; 4491fa9e4066Sahrens } 4492fa9e4066Sahrens 4493fa9e4066Sahrens /* 4494fa9e4066Sahrens * This buffer has been "accessed" only once so far, 4495fa9e4066Sahrens * but it is still in the cache. Move it to the MFU 4496fa9e4066Sahrens * state. 4497fa9e4066Sahrens */ 449889c86e32SChris Williamson if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 4499fa9e4066Sahrens /* 4500fa9e4066Sahrens * More than 125ms have passed since we 4501fa9e4066Sahrens * instantiated this buffer. Move it to the 4502fa9e4066Sahrens * most frequently used state. 4503fa9e4066Sahrens */ 450489c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = now; 45057adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 45067adb730bSGeorge Wilson arc_change_state(arc_mfu, hdr, hash_lock); 4507fa9e4066Sahrens } 450844cb6abcSbmc ARCSTAT_BUMP(arcstat_mru_hits); 450989c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 4510fa9e4066Sahrens arc_state_t *new_state; 4511fa9e4066Sahrens /* 4512fa9e4066Sahrens * This buffer has been "accessed" recently, but 4513fa9e4066Sahrens * was evicted from the cache. Move it to the 4514fa9e4066Sahrens * MFU state. 4515fa9e4066Sahrens */ 4516fa9e4066Sahrens 451789c86e32SChris Williamson if (HDR_PREFETCH(hdr)) { 451844cb6abcSbmc new_state = arc_mru; 451989c86e32SChris Williamson if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 4520dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 45217adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4522fa9e4066Sahrens } else { 452344cb6abcSbmc new_state = arc_mfu; 45247adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4525fa9e4066Sahrens } 4526fa9e4066Sahrens 452789c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 45287adb730bSGeorge Wilson arc_change_state(new_state, hdr, hash_lock); 4529fa9e4066Sahrens 453044cb6abcSbmc ARCSTAT_BUMP(arcstat_mru_ghost_hits); 453189c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mfu) { 4532fa9e4066Sahrens /* 4533fa9e4066Sahrens * This buffer has been accessed more than once and is 4534fa9e4066Sahrens * still in the cache. Keep it in the MFU state. 4535fa9e4066Sahrens * 453613506d1eSmaybee * NOTE: an add_reference() that occurred when we did 453713506d1eSmaybee * the arc_read() will have kicked this off the list. 453813506d1eSmaybee * If it was a prefetch, we will explicitly move it to 453913506d1eSmaybee * the head of the list now. 4540fa9e4066Sahrens */ 454189c86e32SChris Williamson if ((HDR_PREFETCH(hdr)) != 0) { 454289c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4543244781f1SPrakash Surya /* link protected by hash_lock */ 4544244781f1SPrakash Surya ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 454513506d1eSmaybee } 454644cb6abcSbmc ARCSTAT_BUMP(arcstat_mfu_hits); 454789c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 454889c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 454944cb6abcSbmc arc_state_t *new_state = arc_mfu; 4550fa9e4066Sahrens /* 4551fa9e4066Sahrens * This buffer has been accessed more than once but has 4552fa9e4066Sahrens * been evicted from the cache. Move it back to the 4553fa9e4066Sahrens * MFU state. 4554fa9e4066Sahrens */ 4555fa9e4066Sahrens 455689c86e32SChris Williamson if (HDR_PREFETCH(hdr)) { 455713506d1eSmaybee /* 455813506d1eSmaybee * This is a prefetch access... 455913506d1eSmaybee * move this block back to the MRU state. 456013506d1eSmaybee */ 456189c86e32SChris Williamson ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 456244cb6abcSbmc new_state = arc_mru; 456313506d1eSmaybee } 456413506d1eSmaybee 456589c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 45667adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 45677adb730bSGeorge Wilson arc_change_state(new_state, hdr, hash_lock); 4568fa9e4066Sahrens 456944cb6abcSbmc ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 457089c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4571fa94a07fSbrendan /* 4572fa94a07fSbrendan * This buffer is on the 2nd Level ARC. 4573fa94a07fSbrendan */ 4574fa94a07fSbrendan 457589c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 45767adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 45777adb730bSGeorge Wilson arc_change_state(arc_mfu, hdr, hash_lock); 4578fa9e4066Sahrens } else { 4579fa9e4066Sahrens ASSERT(!"invalid arc state"); 4580fa9e4066Sahrens } 4581fa9e4066Sahrens } 4582fa9e4066Sahrens 4583fa9e4066Sahrens /* a generic arc_done_func_t which you can use */ 4584fa9e4066Sahrens /* ARGSUSED */ 4585fa9e4066Sahrens void 4586fa9e4066Sahrens arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4587fa9e4066Sahrens { 45883f9d6ad7SLin Ling if (zio == NULL || zio->io_error == 0) 45895602294fSDan Kimmel bcopy(buf->b_data, arg, arc_buf_size(buf)); 4590dcbf3bd6SGeorge Wilson arc_buf_destroy(buf, arg); 4591fa9e4066Sahrens } 4592fa9e4066Sahrens 45930e8c6158Smaybee /* a generic arc_done_func_t */ 4594fa9e4066Sahrens void 4595fa9e4066Sahrens arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4596fa9e4066Sahrens { 4597fa9e4066Sahrens arc_buf_t **bufp = arg; 4598fa9e4066Sahrens if (zio && zio->io_error) { 4599dcbf3bd6SGeorge Wilson arc_buf_destroy(buf, arg); 4600fa9e4066Sahrens *bufp = NULL; 4601fa9e4066Sahrens } else { 4602fa9e4066Sahrens *bufp = buf; 46033f9d6ad7SLin Ling ASSERT(buf->b_data); 4604fa9e4066Sahrens } 4605fa9e4066Sahrens } 4606fa9e4066Sahrens 4607dcbf3bd6SGeorge Wilson static void 4608dcbf3bd6SGeorge Wilson arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) 4609dcbf3bd6SGeorge Wilson { 4610dcbf3bd6SGeorge Wilson if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { 4611dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); 4612dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 4613dcbf3bd6SGeorge Wilson } else { 4614dcbf3bd6SGeorge Wilson if (HDR_COMPRESSION_ENABLED(hdr)) { 4615dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, 4616dcbf3bd6SGeorge Wilson BP_GET_COMPRESS(bp)); 4617dcbf3bd6SGeorge Wilson } 4618dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); 4619dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); 4620dcbf3bd6SGeorge Wilson } 4621dcbf3bd6SGeorge Wilson } 4622dcbf3bd6SGeorge Wilson 4623fa9e4066Sahrens static void 4624fa9e4066Sahrens arc_read_done(zio_t *zio) 4625fa9e4066Sahrens { 4626dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = zio->io_private; 46275d7b4d43SMatthew Ahrens kmutex_t *hash_lock = NULL; 46285602294fSDan Kimmel arc_callback_t *callback_list; 46295602294fSDan Kimmel arc_callback_t *acb; 46305602294fSDan Kimmel boolean_t freeable = B_FALSE; 46315602294fSDan Kimmel boolean_t no_zio_error = (zio->io_error == 0); 4632fa9e4066Sahrens 4633bbf4a8dfSmaybee /* 4634bbf4a8dfSmaybee * The hdr was inserted into hash-table and removed from lists 4635bbf4a8dfSmaybee * prior to starting I/O. We should find this header, since 4636bbf4a8dfSmaybee * it's in the hash table, and it should be legit since it's 4637bbf4a8dfSmaybee * not possible to evict it during the I/O. The only possible 4638bbf4a8dfSmaybee * reason for it not to be found is if we were freed during the 4639bbf4a8dfSmaybee * read. 4640bbf4a8dfSmaybee */ 46415d7b4d43SMatthew Ahrens if (HDR_IN_HASH_TABLE(hdr)) { 46425d7b4d43SMatthew Ahrens ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 46435d7b4d43SMatthew Ahrens ASSERT3U(hdr->b_dva.dva_word[0], ==, 46445d7b4d43SMatthew Ahrens BP_IDENTITY(zio->io_bp)->dva_word[0]); 46455d7b4d43SMatthew Ahrens ASSERT3U(hdr->b_dva.dva_word[1], ==, 46465d7b4d43SMatthew Ahrens BP_IDENTITY(zio->io_bp)->dva_word[1]); 46475d7b4d43SMatthew Ahrens 46485d7b4d43SMatthew Ahrens arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 46495d7b4d43SMatthew Ahrens &hash_lock); 46505d7b4d43SMatthew Ahrens 4651dcbf3bd6SGeorge Wilson ASSERT((found == hdr && 46525d7b4d43SMatthew Ahrens DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 46535d7b4d43SMatthew Ahrens (found == hdr && HDR_L2_READING(hdr))); 4654dcbf3bd6SGeorge Wilson ASSERT3P(hash_lock, !=, NULL); 4655dcbf3bd6SGeorge Wilson } 4656dcbf3bd6SGeorge Wilson 46575602294fSDan Kimmel if (no_zio_error) { 4658dcbf3bd6SGeorge Wilson /* byteswap if necessary */ 4659dcbf3bd6SGeorge Wilson if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 4660dcbf3bd6SGeorge Wilson if (BP_GET_LEVEL(zio->io_bp) > 0) { 4661dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; 4662dcbf3bd6SGeorge Wilson } else { 4663dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = 4664dcbf3bd6SGeorge Wilson DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4665dcbf3bd6SGeorge Wilson } 4666dcbf3bd6SGeorge Wilson } else { 4667dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 4668dcbf3bd6SGeorge Wilson } 46695d7b4d43SMatthew Ahrens } 4670fa94a07fSbrendan 4671dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); 467289c86e32SChris Williamson if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4673dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); 4674fa9e4066Sahrens 467589c86e32SChris Williamson callback_list = hdr->b_l1hdr.b_acb; 4676dcbf3bd6SGeorge Wilson ASSERT3P(callback_list, !=, NULL); 46776b4acc8bSahrens 46785602294fSDan Kimmel if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) { 4679b24ab676SJeff Bonwick /* 4680b24ab676SJeff Bonwick * Only call arc_access on anonymous buffers. This is because 4681b24ab676SJeff Bonwick * if we've issued an I/O for an evicted buffer, we've already 4682b24ab676SJeff Bonwick * called arc_access (to prevent any simultaneous readers from 4683b24ab676SJeff Bonwick * getting confused). 4684b24ab676SJeff Bonwick */ 4685b24ab676SJeff Bonwick arc_access(hdr, hash_lock); 4686b24ab676SJeff Bonwick } 4687b24ab676SJeff Bonwick 46885602294fSDan Kimmel /* 46895602294fSDan Kimmel * If a read request has a callback (i.e. acb_done is not NULL), then we 46905602294fSDan Kimmel * make a buf containing the data according to the parameters which were 46915602294fSDan Kimmel * passed in. The implementation of arc_buf_alloc_impl() ensures that we 46925602294fSDan Kimmel * aren't needlessly decompressing the data multiple times. 46935602294fSDan Kimmel */ 46945602294fSDan Kimmel int callback_cnt = 0; 46955602294fSDan Kimmel for (acb = callback_list; acb != NULL; acb = acb->acb_next) { 46965602294fSDan Kimmel if (!acb->acb_done) 46975602294fSDan Kimmel continue; 46985602294fSDan Kimmel 46995602294fSDan Kimmel /* This is a demand read since prefetches don't use callbacks */ 47005602294fSDan Kimmel callback_cnt++; 47015602294fSDan Kimmel 47025602294fSDan Kimmel int error = arc_buf_alloc_impl(hdr, acb->acb_private, 47035602294fSDan Kimmel acb->acb_compressed, no_zio_error, &acb->acb_buf); 47045602294fSDan Kimmel if (no_zio_error) { 47055602294fSDan Kimmel zio->io_error = error; 4706fa9e4066Sahrens } 4707fa9e4066Sahrens } 470889c86e32SChris Williamson hdr->b_l1hdr.b_acb = NULL; 4709dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 47105602294fSDan Kimmel if (callback_cnt == 0) { 4711dcbf3bd6SGeorge Wilson ASSERT(HDR_PREFETCH(hdr)); 4712dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 4713*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 4714b24ab676SJeff Bonwick } 4715fa9e4066Sahrens 471689c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 471789c86e32SChris Williamson callback_list != NULL); 4718fa9e4066Sahrens 47195602294fSDan Kimmel if (no_zio_error) { 4720dcbf3bd6SGeorge Wilson arc_hdr_verify(hdr, zio->io_bp); 4721dcbf3bd6SGeorge Wilson } else { 4722dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); 472389c86e32SChris Williamson if (hdr->b_l1hdr.b_state != arc_anon) 472444cb6abcSbmc arc_change_state(arc_anon, hdr, hash_lock); 4725ea8dc4b6Seschrock if (HDR_IN_HASH_TABLE(hdr)) 4726ea8dc4b6Seschrock buf_hash_remove(hdr); 472789c86e32SChris Williamson freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4728fa9e4066Sahrens } 4729fa9e4066Sahrens 4730ea8dc4b6Seschrock /* 473113506d1eSmaybee * Broadcast before we drop the hash_lock to avoid the possibility 473213506d1eSmaybee * that the hdr (and hence the cv) might be freed before we get to 473313506d1eSmaybee * the cv_broadcast(). 4734ea8dc4b6Seschrock */ 473589c86e32SChris Williamson cv_broadcast(&hdr->b_l1hdr.b_cv); 4736ea8dc4b6Seschrock 473789c86e32SChris Williamson if (hash_lock != NULL) { 473844eda4d7Smaybee mutex_exit(hash_lock); 4739fa9e4066Sahrens } else { 4740fa9e4066Sahrens /* 4741fa9e4066Sahrens * This block was freed while we waited for the read to 4742fa9e4066Sahrens * complete. It has been removed from the hash table and 4743fa9e4066Sahrens * moved to the anonymous state (so that it won't show up 4744fa9e4066Sahrens * in the cache). 4745fa9e4066Sahrens */ 474689c86e32SChris Williamson ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 474789c86e32SChris Williamson freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4748fa9e4066Sahrens } 4749fa9e4066Sahrens 4750fa9e4066Sahrens /* execute each callback and free its structure */ 4751fa9e4066Sahrens while ((acb = callback_list) != NULL) { 4752fa9e4066Sahrens if (acb->acb_done) 4753fa9e4066Sahrens acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4754fa9e4066Sahrens 4755fa9e4066Sahrens if (acb->acb_zio_dummy != NULL) { 4756fa9e4066Sahrens acb->acb_zio_dummy->io_error = zio->io_error; 4757fa9e4066Sahrens zio_nowait(acb->acb_zio_dummy); 4758fa9e4066Sahrens } 4759fa9e4066Sahrens 4760fa9e4066Sahrens callback_list = acb->acb_next; 4761fa9e4066Sahrens kmem_free(acb, sizeof (arc_callback_t)); 4762fa9e4066Sahrens } 4763fa9e4066Sahrens 4764fa9e4066Sahrens if (freeable) 4765ea8dc4b6Seschrock arc_hdr_destroy(hdr); 4766fa9e4066Sahrens } 4767fa9e4066Sahrens 4768fa9e4066Sahrens /* 4769fc98fea5SBart Coddens * "Read" the block at the specified DVA (in bp) via the 4770fa9e4066Sahrens * cache. If the block is found in the cache, invoke the provided 4771fa9e4066Sahrens * callback immediately and return. Note that the `zio' parameter 4772fa9e4066Sahrens * in the callback will be NULL in this case, since no IO was 4773fa9e4066Sahrens * required. If the block is not in the cache pass the read request 4774fa9e4066Sahrens * on to the spa with a substitute callback function, so that the 4775fa9e4066Sahrens * requested block will be added to the cache. 4776fa9e4066Sahrens * 4777fa9e4066Sahrens * If a read request arrives for a block that has a read in-progress, 4778fa9e4066Sahrens * either wait for the in-progress read to complete (and return the 4779fa9e4066Sahrens * results); or, if this is a read with a "done" func, add a record 4780fa9e4066Sahrens * to the read to invoke the "done" func when the read completes, 4781fa9e4066Sahrens * and return; or just return. 4782fa9e4066Sahrens * 4783fa9e4066Sahrens * arc_read_done() will invoke all the requested "done" functions 4784fa9e4066Sahrens * for readers of this block. 4785fa9e4066Sahrens */ 4786fa9e4066Sahrens int 47871b912ec7SGeorge Wilson arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 47887adb730bSGeorge Wilson void *private, zio_priority_t priority, int zio_flags, 47897adb730bSGeorge Wilson arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4790fa9e4066Sahrens { 47915d7b4d43SMatthew Ahrens arc_buf_hdr_t *hdr = NULL; 47925d7b4d43SMatthew Ahrens kmutex_t *hash_lock = NULL; 4793fa94a07fSbrendan zio_t *rzio; 4794e9103aaeSGarrett D'Amore uint64_t guid = spa_load_guid(spa); 47955602294fSDan Kimmel boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0; 4796fa9e4066Sahrens 47975d7b4d43SMatthew Ahrens ASSERT(!BP_IS_EMBEDDED(bp) || 47985d7b4d43SMatthew Ahrens BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 47995d7b4d43SMatthew Ahrens 4800fa9e4066Sahrens top: 48015d7b4d43SMatthew Ahrens if (!BP_IS_EMBEDDED(bp)) { 48025d7b4d43SMatthew Ahrens /* 48035d7b4d43SMatthew Ahrens * Embedded BP's have no DVA and require no I/O to "read". 48045d7b4d43SMatthew Ahrens * Create an anonymous arc buf to back it. 48055d7b4d43SMatthew Ahrens */ 48065d7b4d43SMatthew Ahrens hdr = buf_hash_find(guid, bp, &hash_lock); 48075d7b4d43SMatthew Ahrens } 48085d7b4d43SMatthew Ahrens 4809*770499e1SDan Kimmel if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) { 4810dcbf3bd6SGeorge Wilson arc_buf_t *buf = NULL; 48117adb730bSGeorge Wilson *arc_flags |= ARC_FLAG_CACHED; 481213506d1eSmaybee 4813fa9e4066Sahrens if (HDR_IO_IN_PROGRESS(hdr)) { 481413506d1eSmaybee 4815cf6106c8SMatthew Ahrens if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 4816cf6106c8SMatthew Ahrens priority == ZIO_PRIORITY_SYNC_READ) { 4817cf6106c8SMatthew Ahrens /* 4818cf6106c8SMatthew Ahrens * This sync read must wait for an 4819cf6106c8SMatthew Ahrens * in-progress async read (e.g. a predictive 4820cf6106c8SMatthew Ahrens * prefetch). Async reads are queued 4821cf6106c8SMatthew Ahrens * separately at the vdev_queue layer, so 4822cf6106c8SMatthew Ahrens * this is a form of priority inversion. 4823cf6106c8SMatthew Ahrens * Ideally, we would "inherit" the demand 4824cf6106c8SMatthew Ahrens * i/o's priority by moving the i/o from 4825cf6106c8SMatthew Ahrens * the async queue to the synchronous queue, 4826cf6106c8SMatthew Ahrens * but there is currently no mechanism to do 4827cf6106c8SMatthew Ahrens * so. Track this so that we can evaluate 4828cf6106c8SMatthew Ahrens * the magnitude of this potential performance 4829cf6106c8SMatthew Ahrens * problem. 4830cf6106c8SMatthew Ahrens * 4831cf6106c8SMatthew Ahrens * Note that if the prefetch i/o is already 4832cf6106c8SMatthew Ahrens * active (has been issued to the device), 4833cf6106c8SMatthew Ahrens * the prefetch improved performance, because 4834cf6106c8SMatthew Ahrens * we issued it sooner than we would have 4835cf6106c8SMatthew Ahrens * without the prefetch. 4836cf6106c8SMatthew Ahrens */ 4837cf6106c8SMatthew Ahrens DTRACE_PROBE1(arc__sync__wait__for__async, 4838cf6106c8SMatthew Ahrens arc_buf_hdr_t *, hdr); 4839cf6106c8SMatthew Ahrens ARCSTAT_BUMP(arcstat_sync_wait_for_async); 4840cf6106c8SMatthew Ahrens } 4841cf6106c8SMatthew Ahrens if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4842dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, 4843dcbf3bd6SGeorge Wilson ARC_FLAG_PREDICTIVE_PREFETCH); 4844cf6106c8SMatthew Ahrens } 4845cf6106c8SMatthew Ahrens 48467adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_WAIT) { 484789c86e32SChris Williamson cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 484813506d1eSmaybee mutex_exit(hash_lock); 484913506d1eSmaybee goto top; 485013506d1eSmaybee } 48517adb730bSGeorge Wilson ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 485213506d1eSmaybee 485313506d1eSmaybee if (done) { 4854cf6106c8SMatthew Ahrens arc_callback_t *acb = NULL; 4855fa9e4066Sahrens 4856fa9e4066Sahrens acb = kmem_zalloc(sizeof (arc_callback_t), 4857fa9e4066Sahrens KM_SLEEP); 4858fa9e4066Sahrens acb->acb_done = done; 4859fa9e4066Sahrens acb->acb_private = private; 48605602294fSDan Kimmel acb->acb_compressed = compressed_read; 4861fa9e4066Sahrens if (pio != NULL) 4862fa9e4066Sahrens acb->acb_zio_dummy = zio_null(pio, 4863a3f829aeSBill Moore spa, NULL, NULL, NULL, zio_flags); 4864fa9e4066Sahrens 4865dcbf3bd6SGeorge Wilson ASSERT3P(acb->acb_done, !=, NULL); 486689c86e32SChris Williamson acb->acb_next = hdr->b_l1hdr.b_acb; 486789c86e32SChris Williamson hdr->b_l1hdr.b_acb = acb; 4868fa9e4066Sahrens mutex_exit(hash_lock); 4869fa9e4066Sahrens return (0); 4870fa9e4066Sahrens } 4871fa9e4066Sahrens mutex_exit(hash_lock); 4872fa9e4066Sahrens return (0); 4873fa9e4066Sahrens } 4874fa9e4066Sahrens 487589c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_state == arc_mru || 487689c86e32SChris Williamson hdr->b_l1hdr.b_state == arc_mfu); 4877fa9e4066Sahrens 4878ea8dc4b6Seschrock if (done) { 4879cf6106c8SMatthew Ahrens if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4880cf6106c8SMatthew Ahrens /* 4881cf6106c8SMatthew Ahrens * This is a demand read which does not have to 4882cf6106c8SMatthew Ahrens * wait for i/o because we did a predictive 4883cf6106c8SMatthew Ahrens * prefetch i/o for it, which has completed. 4884cf6106c8SMatthew Ahrens */ 4885cf6106c8SMatthew Ahrens DTRACE_PROBE1( 4886cf6106c8SMatthew Ahrens arc__demand__hit__predictive__prefetch, 4887cf6106c8SMatthew Ahrens arc_buf_hdr_t *, hdr); 4888cf6106c8SMatthew Ahrens ARCSTAT_BUMP( 4889cf6106c8SMatthew Ahrens arcstat_demand_hit_predictive_prefetch); 4890dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, 4891dcbf3bd6SGeorge Wilson ARC_FLAG_PREDICTIVE_PREFETCH); 4892cf6106c8SMatthew Ahrens } 4893dcbf3bd6SGeorge Wilson ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); 4894dcbf3bd6SGeorge Wilson 48955602294fSDan Kimmel /* Get a buf with the desired data in it. */ 48965602294fSDan Kimmel VERIFY0(arc_buf_alloc_impl(hdr, private, 48975602294fSDan Kimmel compressed_read, B_TRUE, &buf)); 48987adb730bSGeorge Wilson } else if (*arc_flags & ARC_FLAG_PREFETCH && 489989c86e32SChris Williamson refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4900dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 4901fa9e4066Sahrens } 4902fa9e4066Sahrens DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 490344eda4d7Smaybee arc_access(hdr, hash_lock); 49047adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_L2CACHE) 4905dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 490644eda4d7Smaybee mutex_exit(hash_lock); 490744cb6abcSbmc ARCSTAT_BUMP(arcstat_hits); 490889c86e32SChris Williamson ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 490989c86e32SChris Williamson demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 491044cb6abcSbmc data, metadata, hits); 491144cb6abcSbmc 4912fa9e4066Sahrens if (done) 4913fa9e4066Sahrens done(NULL, buf, private); 4914fa9e4066Sahrens } else { 4915dcbf3bd6SGeorge Wilson uint64_t lsize = BP_GET_LSIZE(bp); 4916dcbf3bd6SGeorge Wilson uint64_t psize = BP_GET_PSIZE(bp); 49175d7b4d43SMatthew Ahrens arc_callback_t *acb; 49183a737e0dSbrendan vdev_t *vd = NULL; 4919d5285caeSGeorge Wilson uint64_t addr = 0; 49205a98e54bSBrendan Gregg - Sun Microsystems boolean_t devw = B_FALSE; 4921dcbf3bd6SGeorge Wilson uint64_t size; 4922fa9e4066Sahrens 4923fa9e4066Sahrens if (hdr == NULL) { 4924fa9e4066Sahrens /* this block is not in the cache */ 49255d7b4d43SMatthew Ahrens arc_buf_hdr_t *exists = NULL; 4926ad23a2dbSjohansen arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4927dcbf3bd6SGeorge Wilson hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 4928dcbf3bd6SGeorge Wilson BP_GET_COMPRESS(bp), type); 4929dcbf3bd6SGeorge Wilson 49305d7b4d43SMatthew Ahrens if (!BP_IS_EMBEDDED(bp)) { 49315d7b4d43SMatthew Ahrens hdr->b_dva = *BP_IDENTITY(bp); 49325d7b4d43SMatthew Ahrens hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 49335d7b4d43SMatthew Ahrens exists = buf_hash_insert(hdr, &hash_lock); 49345d7b4d43SMatthew Ahrens } 49355d7b4d43SMatthew Ahrens if (exists != NULL) { 4936fa9e4066Sahrens /* somebody beat us to the hash insert */ 4937fa9e4066Sahrens mutex_exit(hash_lock); 49383f9d6ad7SLin Ling buf_discard_identity(hdr); 4939dcbf3bd6SGeorge Wilson arc_hdr_destroy(hdr); 4940fa9e4066Sahrens goto top; /* restart the IO request */ 4941fa9e4066Sahrens } 4942fa9e4066Sahrens } else { 494389c86e32SChris Williamson /* 494489c86e32SChris Williamson * This block is in the ghost cache. If it was L2-only 494589c86e32SChris Williamson * (and thus didn't have an L1 hdr), we realloc the 494689c86e32SChris Williamson * header to add an L1 hdr. 494789c86e32SChris Williamson */ 494889c86e32SChris Williamson if (!HDR_HAS_L1HDR(hdr)) { 494989c86e32SChris Williamson hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 495089c86e32SChris Williamson hdr_full_cache); 495189c86e32SChris Williamson } 4952*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 495389c86e32SChris Williamson ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4954ea8dc4b6Seschrock ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 495589c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4956244781f1SPrakash Surya ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 49575602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 495813506d1eSmaybee 4959cf6106c8SMatthew Ahrens /* 4960dcbf3bd6SGeorge Wilson * This is a delicate dance that we play here. 4961dcbf3bd6SGeorge Wilson * This hdr is in the ghost list so we access it 4962dcbf3bd6SGeorge Wilson * to move it out of the ghost list before we 4963dcbf3bd6SGeorge Wilson * initiate the read. If it's a prefetch then 4964dcbf3bd6SGeorge Wilson * it won't have a callback so we'll remove the 4965dcbf3bd6SGeorge Wilson * reference that arc_buf_alloc_impl() created. We 4966dcbf3bd6SGeorge Wilson * do this after we've called arc_access() to 4967dcbf3bd6SGeorge Wilson * avoid hitting an assert in remove_reference(). 4968cf6106c8SMatthew Ahrens */ 49697e453561SWilliam Gorrell arc_access(hdr, hash_lock); 4970*770499e1SDan Kimmel arc_hdr_alloc_pabd(hdr); 4971dcbf3bd6SGeorge Wilson } 4972*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 4973dcbf3bd6SGeorge Wilson size = arc_hdr_size(hdr); 4974dcbf3bd6SGeorge Wilson 4975dcbf3bd6SGeorge Wilson /* 4976dcbf3bd6SGeorge Wilson * If compression is enabled on the hdr, then will do 4977dcbf3bd6SGeorge Wilson * RAW I/O and will store the compressed data in the hdr's 4978dcbf3bd6SGeorge Wilson * data block. Otherwise, the hdr's data block will contain 4979dcbf3bd6SGeorge Wilson * the uncompressed data. 4980dcbf3bd6SGeorge Wilson */ 4981dcbf3bd6SGeorge Wilson if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 4982dcbf3bd6SGeorge Wilson zio_flags |= ZIO_FLAG_RAW; 4983fa9e4066Sahrens } 4984fa9e4066Sahrens 4985dcbf3bd6SGeorge Wilson if (*arc_flags & ARC_FLAG_PREFETCH) 4986dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 4987dcbf3bd6SGeorge Wilson if (*arc_flags & ARC_FLAG_L2CACHE) 4988dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 4989dcbf3bd6SGeorge Wilson if (BP_GET_LEVEL(bp) > 0) 4990dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); 4991cf6106c8SMatthew Ahrens if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 4992dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); 499389c86e32SChris Williamson ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 49945614b00aSWilliam Gorrell 4995fa9e4066Sahrens acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4996fa9e4066Sahrens acb->acb_done = done; 4997fa9e4066Sahrens acb->acb_private = private; 49985602294fSDan Kimmel acb->acb_compressed = compressed_read; 4999fa9e4066Sahrens 5000dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 500189c86e32SChris Williamson hdr->b_l1hdr.b_acb = acb; 5002dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5003fa9e4066Sahrens 500489c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr) && 500589c86e32SChris Williamson (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 500689c86e32SChris Williamson devw = hdr->b_l2hdr.b_dev->l2ad_writing; 500789c86e32SChris Williamson addr = hdr->b_l2hdr.b_daddr; 5008e14bb325SJeff Bonwick /* 5009e14bb325SJeff Bonwick * Lock out device removal. 5010e14bb325SJeff Bonwick */ 5011e14bb325SJeff Bonwick if (vdev_is_dead(vd) || 5012e14bb325SJeff Bonwick !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 5013e14bb325SJeff Bonwick vd = NULL; 50143a737e0dSbrendan } 50153a737e0dSbrendan 5016dcbf3bd6SGeorge Wilson if (priority == ZIO_PRIORITY_ASYNC_READ) 5017dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5018dcbf3bd6SGeorge Wilson else 5019dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5020dcbf3bd6SGeorge Wilson 50215d7b4d43SMatthew Ahrens if (hash_lock != NULL) 50225d7b4d43SMatthew Ahrens mutex_exit(hash_lock); 50233a737e0dSbrendan 50243e30c24aSWill Andrews /* 50253e30c24aSWill Andrews * At this point, we have a level 1 cache miss. Try again in 50263e30c24aSWill Andrews * L2ARC if possible. 50273e30c24aSWill Andrews */ 5028dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); 5029dcbf3bd6SGeorge Wilson 50305c28183bSBrendan Gregg - Sun Microsystems DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 5031dcbf3bd6SGeorge Wilson uint64_t, lsize, zbookmark_phys_t *, zb); 503244cb6abcSbmc ARCSTAT_BUMP(arcstat_misses); 503389c86e32SChris Williamson ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 503489c86e32SChris Williamson demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 503544cb6abcSbmc data, metadata, misses); 5036ea8dc4b6Seschrock 50375a98e54bSBrendan Gregg - Sun Microsystems if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 5038fa94a07fSbrendan /* 5039fa94a07fSbrendan * Read from the L2ARC if the following are true: 50403a737e0dSbrendan * 1. The L2ARC vdev was previously cached. 50413a737e0dSbrendan * 2. This buffer still has L2ARC metadata. 50423a737e0dSbrendan * 3. This buffer isn't currently writing to the L2ARC. 50433a737e0dSbrendan * 4. The L2ARC entry wasn't evicted, which may 50443a737e0dSbrendan * also have invalidated the vdev. 50455a98e54bSBrendan Gregg - Sun Microsystems * 5. This isn't prefetch and l2arc_noprefetch is set. 5046fa94a07fSbrendan */ 504789c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr) && 50485a98e54bSBrendan Gregg - Sun Microsystems !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 50495a98e54bSBrendan Gregg - Sun Microsystems !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 5050fa94a07fSbrendan l2arc_read_callback_t *cb; 5051fa94a07fSbrendan 5052c5904d13Seschrock DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 5053c5904d13Seschrock ARCSTAT_BUMP(arcstat_l2_hits); 5054c5904d13Seschrock 5055fa94a07fSbrendan cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 5056fa94a07fSbrendan KM_SLEEP); 5057dcbf3bd6SGeorge Wilson cb->l2rcb_hdr = hdr; 5058fa94a07fSbrendan cb->l2rcb_bp = *bp; 5059fa94a07fSbrendan cb->l2rcb_zb = *zb; 50603baa08fcSek cb->l2rcb_flags = zio_flags; 5061fa94a07fSbrendan 5062d5285caeSGeorge Wilson ASSERT(addr >= VDEV_LABEL_START_SIZE && 5063dcbf3bd6SGeorge Wilson addr + lsize < vd->vdev_psize - 5064d5285caeSGeorge Wilson VDEV_LABEL_END_SIZE); 5065d5285caeSGeorge Wilson 5066fa94a07fSbrendan /* 5067e14bb325SJeff Bonwick * l2arc read. The SCL_L2ARC lock will be 5068e14bb325SJeff Bonwick * released by l2arc_read_done(). 5069aad02571SSaso Kiselkov * Issue a null zio if the underlying buffer 5070aad02571SSaso Kiselkov * was squashed to zero size by compression. 5071fa94a07fSbrendan */ 5072dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), !=, 5073dcbf3bd6SGeorge Wilson ZIO_COMPRESS_EMPTY); 5074dcbf3bd6SGeorge Wilson rzio = zio_read_phys(pio, vd, addr, 5075*770499e1SDan Kimmel size, hdr->b_l1hdr.b_pabd, 5076dcbf3bd6SGeorge Wilson ZIO_CHECKSUM_OFF, 5077dcbf3bd6SGeorge Wilson l2arc_read_done, cb, priority, 5078dcbf3bd6SGeorge Wilson zio_flags | ZIO_FLAG_DONT_CACHE | 5079dcbf3bd6SGeorge Wilson ZIO_FLAG_CANFAIL | 5080dcbf3bd6SGeorge Wilson ZIO_FLAG_DONT_PROPAGATE | 5081dcbf3bd6SGeorge Wilson ZIO_FLAG_DONT_RETRY, B_FALSE); 5082fa94a07fSbrendan DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 5083fa94a07fSbrendan zio_t *, rzio); 5084dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_l2_read_bytes, size); 5085fa94a07fSbrendan 50867adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_NOWAIT) { 50873a737e0dSbrendan zio_nowait(rzio); 50883a737e0dSbrendan return (0); 50893a737e0dSbrendan } 5090fa94a07fSbrendan 50917adb730bSGeorge Wilson ASSERT(*arc_flags & ARC_FLAG_WAIT); 50923a737e0dSbrendan if (zio_wait(rzio) == 0) 50933a737e0dSbrendan return (0); 50943a737e0dSbrendan 50953a737e0dSbrendan /* l2arc read error; goto zio_read() */ 5096fa94a07fSbrendan } else { 5097fa94a07fSbrendan DTRACE_PROBE1(l2arc__miss, 5098fa94a07fSbrendan arc_buf_hdr_t *, hdr); 5099fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_misses); 5100fa94a07fSbrendan if (HDR_L2_WRITING(hdr)) 5101fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_rw_clash); 5102e14bb325SJeff Bonwick spa_config_exit(spa, SCL_L2ARC, vd); 5103fa94a07fSbrendan } 51045a98e54bSBrendan Gregg - Sun Microsystems } else { 510576a25fafSBill Moore if (vd != NULL) 510676a25fafSBill Moore spa_config_exit(spa, SCL_L2ARC, vd); 51075a98e54bSBrendan Gregg - Sun Microsystems if (l2arc_ndev != 0) { 51085a98e54bSBrendan Gregg - Sun Microsystems DTRACE_PROBE1(l2arc__miss, 51095a98e54bSBrendan Gregg - Sun Microsystems arc_buf_hdr_t *, hdr); 51105a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_BUMP(arcstat_l2_misses); 51115a98e54bSBrendan Gregg - Sun Microsystems } 5112fa94a07fSbrendan } 5113c5904d13Seschrock 5114*770499e1SDan Kimmel rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size, 5115dcbf3bd6SGeorge Wilson arc_read_done, hdr, priority, zio_flags, zb); 5116fa9e4066Sahrens 51177adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_WAIT) 5118fa9e4066Sahrens return (zio_wait(rzio)); 5119fa9e4066Sahrens 51207adb730bSGeorge Wilson ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5121fa9e4066Sahrens zio_nowait(rzio); 5122fa9e4066Sahrens } 5123fa9e4066Sahrens return (0); 5124fa9e4066Sahrens } 5125fa9e4066Sahrens 51266e6d5868SMatthew Ahrens /* 51276e6d5868SMatthew Ahrens * Notify the arc that a block was freed, and thus will never be used again. 51286e6d5868SMatthew Ahrens */ 51296e6d5868SMatthew Ahrens void 51306e6d5868SMatthew Ahrens arc_freed(spa_t *spa, const blkptr_t *bp) 51316e6d5868SMatthew Ahrens { 51326e6d5868SMatthew Ahrens arc_buf_hdr_t *hdr; 51336e6d5868SMatthew Ahrens kmutex_t *hash_lock; 51346e6d5868SMatthew Ahrens uint64_t guid = spa_load_guid(spa); 51356e6d5868SMatthew Ahrens 51365d7b4d43SMatthew Ahrens ASSERT(!BP_IS_EMBEDDED(bp)); 51375d7b4d43SMatthew Ahrens 51385d7b4d43SMatthew Ahrens hdr = buf_hash_find(guid, bp, &hash_lock); 51396e6d5868SMatthew Ahrens if (hdr == NULL) 51406e6d5868SMatthew Ahrens return; 51416e6d5868SMatthew Ahrens 5142dcbf3bd6SGeorge Wilson /* 5143dcbf3bd6SGeorge Wilson * We might be trying to free a block that is still doing I/O 5144dcbf3bd6SGeorge Wilson * (i.e. prefetch) or has a reference (i.e. a dedup-ed, 5145dcbf3bd6SGeorge Wilson * dmu_sync-ed block). If this block is being prefetched, then it 5146dcbf3bd6SGeorge Wilson * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr 5147dcbf3bd6SGeorge Wilson * until the I/O completes. A block may also have a reference if it is 5148dcbf3bd6SGeorge Wilson * part of a dedup-ed, dmu_synced write. The dmu_sync() function would 5149dcbf3bd6SGeorge Wilson * have written the new block to its final resting place on disk but 5150dcbf3bd6SGeorge Wilson * without the dedup flag set. This would have left the hdr in the MRU 5151dcbf3bd6SGeorge Wilson * state and discoverable. When the txg finally syncs it detects that 5152dcbf3bd6SGeorge Wilson * the block was overridden in open context and issues an override I/O. 5153dcbf3bd6SGeorge Wilson * Since this is a dedup block, the override I/O will determine if the 5154dcbf3bd6SGeorge Wilson * block is already in the DDT. If so, then it will replace the io_bp 5155dcbf3bd6SGeorge Wilson * with the bp from the DDT and allow the I/O to finish. When the I/O 5156dcbf3bd6SGeorge Wilson * reaches the done callback, dbuf_write_override_done, it will 5157dcbf3bd6SGeorge Wilson * check to see if the io_bp and io_bp_override are identical. 5158dcbf3bd6SGeorge Wilson * If they are not, then it indicates that the bp was replaced with 5159dcbf3bd6SGeorge Wilson * the bp in the DDT and the override bp is freed. This allows 5160dcbf3bd6SGeorge Wilson * us to arrive here with a reference on a block that is being 5161dcbf3bd6SGeorge Wilson * freed. So if we have an I/O in progress, or a reference to 5162dcbf3bd6SGeorge Wilson * this hdr, then we don't destroy the hdr. 5163dcbf3bd6SGeorge Wilson */ 5164dcbf3bd6SGeorge Wilson if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && 5165dcbf3bd6SGeorge Wilson refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { 5166dcbf3bd6SGeorge Wilson arc_change_state(arc_anon, hdr, hash_lock); 5167dcbf3bd6SGeorge Wilson arc_hdr_destroy(hdr); 51686e6d5868SMatthew Ahrens mutex_exit(hash_lock); 5169bbfa8ea8SMatthew Ahrens } else { 5170dcbf3bd6SGeorge Wilson mutex_exit(hash_lock); 5171ea8dc4b6Seschrock } 5172dd6ef538Smaybee 5173ea8dc4b6Seschrock } 5174ea8dc4b6Seschrock 5175fa9e4066Sahrens /* 51763e30c24aSWill Andrews * Release this buffer from the cache, making it an anonymous buffer. This 51773e30c24aSWill Andrews * must be done after a read and prior to modifying the buffer contents. 5178fa9e4066Sahrens * If the buffer has more than one reference, we must make 5179088f3894Sahrens * a new hdr for the buffer. 5180fa9e4066Sahrens */ 5181fa9e4066Sahrens void 5182fa9e4066Sahrens arc_release(arc_buf_t *buf, void *tag) 5183fa9e4066Sahrens { 518489c86e32SChris Williamson arc_buf_hdr_t *hdr = buf->b_hdr; 5185fa9e4066Sahrens 51863f9d6ad7SLin Ling /* 51873f9d6ad7SLin Ling * It would be nice to assert that if it's DMU metadata (level > 51883f9d6ad7SLin Ling * 0 || it's the dnode file), then it must be syncing context. 51893f9d6ad7SLin Ling * But we don't know that information at this level. 51903f9d6ad7SLin Ling */ 51913f9d6ad7SLin Ling 51923f9d6ad7SLin Ling mutex_enter(&buf->b_evict_lock); 51936f83844dSMark Maybee 5194244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 5195244781f1SPrakash Surya 519689c86e32SChris Williamson /* 519789c86e32SChris Williamson * We don't grab the hash lock prior to this check, because if 519889c86e32SChris Williamson * the buffer's header is in the arc_anon state, it won't be 519989c86e32SChris Williamson * linked into the hash table. 520089c86e32SChris Williamson */ 520189c86e32SChris Williamson if (hdr->b_l1hdr.b_state == arc_anon) { 520289c86e32SChris Williamson mutex_exit(&buf->b_evict_lock); 520389c86e32SChris Williamson ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 520489c86e32SChris Williamson ASSERT(!HDR_IN_HASH_TABLE(hdr)); 520589c86e32SChris Williamson ASSERT(!HDR_HAS_L2HDR(hdr)); 5206dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 5207fa9e4066Sahrens 5208dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 520989c86e32SChris Williamson ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 521089c86e32SChris Williamson ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 521189c86e32SChris Williamson 521289c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = 0; 5213dcbf3bd6SGeorge Wilson 5214dcbf3bd6SGeorge Wilson /* 5215dcbf3bd6SGeorge Wilson * If the buf is being overridden then it may already 5216dcbf3bd6SGeorge Wilson * have a hdr that is not empty. 5217dcbf3bd6SGeorge Wilson */ 5218dcbf3bd6SGeorge Wilson buf_discard_identity(hdr); 521989c86e32SChris Williamson arc_buf_thaw(buf); 522089c86e32SChris Williamson 522189c86e32SChris Williamson return; 5222fa9e4066Sahrens } 5223fa9e4066Sahrens 522489c86e32SChris Williamson kmutex_t *hash_lock = HDR_LOCK(hdr); 522589c86e32SChris Williamson mutex_enter(hash_lock); 522689c86e32SChris Williamson 522789c86e32SChris Williamson /* 522889c86e32SChris Williamson * This assignment is only valid as long as the hash_lock is 522989c86e32SChris Williamson * held, we must be careful not to reference state or the 523089c86e32SChris Williamson * b_state field after dropping the lock. 523189c86e32SChris Williamson */ 523289c86e32SChris Williamson arc_state_t *state = hdr->b_l1hdr.b_state; 523389c86e32SChris Williamson ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 523489c86e32SChris Williamson ASSERT3P(state, !=, arc_anon); 523589c86e32SChris Williamson 523689c86e32SChris Williamson /* this buffer is not on any list */ 52375602294fSDan Kimmel ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0); 523889c86e32SChris Williamson 523989c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr)) { 524089c86e32SChris Williamson mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5241244781f1SPrakash Surya 5242244781f1SPrakash Surya /* 5243a52fc310SPrakash Surya * We have to recheck this conditional again now that 5244a52fc310SPrakash Surya * we're holding the l2ad_mtx to prevent a race with 5245a52fc310SPrakash Surya * another thread which might be concurrently calling 5246a52fc310SPrakash Surya * l2arc_evict(). In that case, l2arc_evict() might have 5247a52fc310SPrakash Surya * destroyed the header's L2 portion as we were waiting 5248a52fc310SPrakash Surya * to acquire the l2ad_mtx. 5249244781f1SPrakash Surya */ 5250a52fc310SPrakash Surya if (HDR_HAS_L2HDR(hdr)) 5251a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(hdr); 5252244781f1SPrakash Surya 525389c86e32SChris Williamson mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 52546f83844dSMark Maybee } 52556f83844dSMark Maybee 5256ea8dc4b6Seschrock /* 5257ea8dc4b6Seschrock * Do we have more than one buf? 5258ea8dc4b6Seschrock */ 5259dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_bufcnt > 1) { 5260fa9e4066Sahrens arc_buf_hdr_t *nhdr; 5261ac05c741SMark Maybee uint64_t spa = hdr->b_spa; 5262dcbf3bd6SGeorge Wilson uint64_t psize = HDR_GET_PSIZE(hdr); 5263dcbf3bd6SGeorge Wilson uint64_t lsize = HDR_GET_LSIZE(hdr); 5264dcbf3bd6SGeorge Wilson enum zio_compress compress = HDR_GET_COMPRESS(hdr); 526589c86e32SChris Williamson arc_buf_contents_t type = arc_buf_type(hdr); 5266dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 5267fa9e4066Sahrens 526889c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 5269dcbf3bd6SGeorge Wilson (void) remove_reference(hdr, hash_lock, tag); 5270dcbf3bd6SGeorge Wilson 52715602294fSDan Kimmel if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) { 5272dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5273dcbf3bd6SGeorge Wilson ASSERT(ARC_BUF_LAST(buf)); 5274dcbf3bd6SGeorge Wilson } 5275dcbf3bd6SGeorge Wilson 5276fa9e4066Sahrens /* 52773f9d6ad7SLin Ling * Pull the data off of this hdr and attach it to 5278dcbf3bd6SGeorge Wilson * a new anonymous hdr. Also find the last buffer 5279dcbf3bd6SGeorge Wilson * in the hdr's buffer list. 5280fa9e4066Sahrens */ 52815602294fSDan Kimmel arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 5282dcbf3bd6SGeorge Wilson ASSERT3P(lastbuf, !=, NULL); 5283dcbf3bd6SGeorge Wilson 5284dcbf3bd6SGeorge Wilson /* 5285dcbf3bd6SGeorge Wilson * If the current arc_buf_t and the hdr are sharing their data 52865602294fSDan Kimmel * buffer, then we must stop sharing that block. 5287dcbf3bd6SGeorge Wilson */ 5288dcbf3bd6SGeorge Wilson if (arc_buf_is_shared(buf)) { 5289dcbf3bd6SGeorge Wilson VERIFY(!arc_buf_is_shared(lastbuf)); 5290ea8dc4b6Seschrock 5291dcbf3bd6SGeorge Wilson /* 5292dcbf3bd6SGeorge Wilson * First, sever the block sharing relationship between 52935602294fSDan Kimmel * buf and the arc_buf_hdr_t. 5294dcbf3bd6SGeorge Wilson */ 5295dcbf3bd6SGeorge Wilson arc_unshare_buf(hdr, buf); 52965602294fSDan Kimmel 52975602294fSDan Kimmel /* 5298*770499e1SDan Kimmel * Now we need to recreate the hdr's b_pabd. Since we 52995602294fSDan Kimmel * have lastbuf handy, we try to share with it, but if 5300*770499e1SDan Kimmel * we can't then we allocate a new b_pabd and copy the 53015602294fSDan Kimmel * data from buf into it. 53025602294fSDan Kimmel */ 53035602294fSDan Kimmel if (arc_can_share(hdr, lastbuf)) { 53045602294fSDan Kimmel arc_share_buf(hdr, lastbuf); 53055602294fSDan Kimmel } else { 5306*770499e1SDan Kimmel arc_hdr_alloc_pabd(hdr); 5307*770499e1SDan Kimmel abd_copy_from_buf(hdr->b_l1hdr.b_pabd, 5308*770499e1SDan Kimmel buf->b_data, psize); 53095602294fSDan Kimmel } 5310dcbf3bd6SGeorge Wilson VERIFY3P(lastbuf->b_data, !=, NULL); 5311dcbf3bd6SGeorge Wilson } else if (HDR_SHARED_DATA(hdr)) { 53125602294fSDan Kimmel /* 53135602294fSDan Kimmel * Uncompressed shared buffers are always at the end 53145602294fSDan Kimmel * of the list. Compressed buffers don't have the 53155602294fSDan Kimmel * same requirements. This makes it hard to 53165602294fSDan Kimmel * simply assert that the lastbuf is shared so 53175602294fSDan Kimmel * we rely on the hdr's compression flags to determine 53185602294fSDan Kimmel * if we have a compressed, shared buffer. 53195602294fSDan Kimmel */ 53205602294fSDan Kimmel ASSERT(arc_buf_is_shared(lastbuf) || 53215602294fSDan Kimmel HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 53225602294fSDan Kimmel ASSERT(!ARC_BUF_SHARED(buf)); 5323dcbf3bd6SGeorge Wilson } 5324*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 532589c86e32SChris Williamson ASSERT3P(state, !=, arc_l2c_only); 53262fd872a7SPrakash Surya 5327dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_size, 53285602294fSDan Kimmel arc_buf_size(buf), buf); 53292fd872a7SPrakash Surya 533089c86e32SChris Williamson if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 533189c86e32SChris Williamson ASSERT3P(state, !=, arc_l2c_only); 5332dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 53335602294fSDan Kimmel arc_buf_size(buf), buf); 5334ea8dc4b6Seschrock } 53359253d63dSGeorge Wilson 5336dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt -= 1; 5337c717a561Smaybee arc_cksum_verify(buf); 5338cd1c8b85SMatthew Ahrens arc_buf_unwatch(buf); 5339ea8dc4b6Seschrock 5340fa9e4066Sahrens mutex_exit(hash_lock); 5341fa9e4066Sahrens 5342dcbf3bd6SGeorge Wilson /* 5343*770499e1SDan Kimmel * Allocate a new hdr. The new hdr will contain a b_pabd 5344dcbf3bd6SGeorge Wilson * buffer which will be freed in arc_write(). 5345dcbf3bd6SGeorge Wilson */ 5346dcbf3bd6SGeorge Wilson nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type); 5347dcbf3bd6SGeorge Wilson ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); 5348dcbf3bd6SGeorge Wilson ASSERT0(nhdr->b_l1hdr.b_bufcnt); 5349dcbf3bd6SGeorge Wilson ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt)); 5350dcbf3bd6SGeorge Wilson VERIFY3U(nhdr->b_type, ==, type); 5351dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(nhdr)); 535289c86e32SChris Williamson 535389c86e32SChris Williamson nhdr->b_l1hdr.b_buf = buf; 5354dcbf3bd6SGeorge Wilson nhdr->b_l1hdr.b_bufcnt = 1; 535589c86e32SChris Williamson (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 5356af2c4821Smaybee buf->b_hdr = nhdr; 5357dcbf3bd6SGeorge Wilson 53583f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 5359dcbf3bd6SGeorge Wilson (void) refcount_add_many(&arc_anon->arcs_size, 53605602294fSDan Kimmel arc_buf_size(buf), buf); 5361fa9e4066Sahrens } else { 53623f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 536389c86e32SChris Williamson ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 5364244781f1SPrakash Surya /* protected by hash lock, or hdr is on arc_anon */ 5365244781f1SPrakash Surya ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5366fa9e4066Sahrens ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 536789c86e32SChris Williamson arc_change_state(arc_anon, hdr, hash_lock); 536889c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = 0; 536989c86e32SChris Williamson mutex_exit(hash_lock); 5370fa94a07fSbrendan 53713f9d6ad7SLin Ling buf_discard_identity(hdr); 5372c717a561Smaybee arc_buf_thaw(buf); 5373fa9e4066Sahrens } 5374fa9e4066Sahrens } 5375fa9e4066Sahrens 5376fa9e4066Sahrens int 5377fa9e4066Sahrens arc_released(arc_buf_t *buf) 5378fa9e4066Sahrens { 53796f83844dSMark Maybee int released; 53806f83844dSMark Maybee 53813f9d6ad7SLin Ling mutex_enter(&buf->b_evict_lock); 538289c86e32SChris Williamson released = (buf->b_data != NULL && 538389c86e32SChris Williamson buf->b_hdr->b_l1hdr.b_state == arc_anon); 53843f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 53856f83844dSMark Maybee return (released); 5386ea8dc4b6Seschrock } 5387ea8dc4b6Seschrock 5388ea8dc4b6Seschrock #ifdef ZFS_DEBUG 5389ea8dc4b6Seschrock int 5390ea8dc4b6Seschrock arc_referenced(arc_buf_t *buf) 5391ea8dc4b6Seschrock { 53926f83844dSMark Maybee int referenced; 53936f83844dSMark Maybee 53943f9d6ad7SLin Ling mutex_enter(&buf->b_evict_lock); 539589c86e32SChris Williamson referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 53963f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 53976f83844dSMark Maybee return (referenced); 5398ea8dc4b6Seschrock } 5399ea8dc4b6Seschrock #endif 5400ea8dc4b6Seschrock 5401c717a561Smaybee static void 5402c717a561Smaybee arc_write_ready(zio_t *zio) 5403c717a561Smaybee { 5404c717a561Smaybee arc_write_callback_t *callback = zio->io_private; 5405c717a561Smaybee arc_buf_t *buf = callback->awcb_buf; 54060a4e9518Sgw arc_buf_hdr_t *hdr = buf->b_hdr; 5407dcbf3bd6SGeorge Wilson uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp); 5408c717a561Smaybee 540989c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 541089c86e32SChris Williamson ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 5411dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 5412e14bb325SJeff Bonwick 54130a4e9518Sgw /* 5414dcbf3bd6SGeorge Wilson * If we're reexecuting this zio because the pool suspended, then 5415dcbf3bd6SGeorge Wilson * cleanup any state that was previously set the first time the 54165602294fSDan Kimmel * callback was invoked. 54170a4e9518Sgw */ 5418dcbf3bd6SGeorge Wilson if (zio->io_flags & ZIO_FLAG_REEXECUTED) { 5419dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 5420dcbf3bd6SGeorge Wilson arc_buf_unwatch(buf); 5421*770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 5422dcbf3bd6SGeorge Wilson if (arc_buf_is_shared(buf)) { 5423dcbf3bd6SGeorge Wilson arc_unshare_buf(hdr, buf); 5424dcbf3bd6SGeorge Wilson } else { 5425*770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 5426dcbf3bd6SGeorge Wilson } 54270a4e9518Sgw } 54280a4e9518Sgw } 5429*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5430dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(hdr)); 5431dcbf3bd6SGeorge Wilson ASSERT(!arc_buf_is_shared(buf)); 5432dcbf3bd6SGeorge Wilson 5433dcbf3bd6SGeorge Wilson callback->awcb_ready(zio, buf, callback->awcb_private); 5434dcbf3bd6SGeorge Wilson 5435dcbf3bd6SGeorge Wilson if (HDR_IO_IN_PROGRESS(hdr)) 5436dcbf3bd6SGeorge Wilson ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); 5437dcbf3bd6SGeorge Wilson 5438dcbf3bd6SGeorge Wilson arc_cksum_compute(buf); 5439dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5440dcbf3bd6SGeorge Wilson 5441dcbf3bd6SGeorge Wilson enum zio_compress compress; 5442dcbf3bd6SGeorge Wilson if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5443dcbf3bd6SGeorge Wilson compress = ZIO_COMPRESS_OFF; 5444dcbf3bd6SGeorge Wilson } else { 5445dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp)); 5446dcbf3bd6SGeorge Wilson compress = BP_GET_COMPRESS(zio->io_bp); 5447dcbf3bd6SGeorge Wilson } 5448dcbf3bd6SGeorge Wilson HDR_SET_PSIZE(hdr, psize); 5449dcbf3bd6SGeorge Wilson arc_hdr_set_compress(hdr, compress); 5450dcbf3bd6SGeorge Wilson 5451*770499e1SDan Kimmel 5452dcbf3bd6SGeorge Wilson /* 5453*770499e1SDan Kimmel * Fill the hdr with data. If the hdr is compressed, the data we want 5454*770499e1SDan Kimmel * is available from the zio, otherwise we can take it from the buf. 5455*770499e1SDan Kimmel * 5456*770499e1SDan Kimmel * We might be able to share the buf's data with the hdr here. However, 5457*770499e1SDan Kimmel * doing so would cause the ARC to be full of linear ABDs if we write a 5458*770499e1SDan Kimmel * lot of shareable data. As a compromise, we check whether scattered 5459*770499e1SDan Kimmel * ABDs are allowed, and assume that if they are then the user wants 5460*770499e1SDan Kimmel * the ARC to be primarily filled with them regardless of the data being 5461*770499e1SDan Kimmel * written. Therefore, if they're allowed then we allocate one and copy 5462*770499e1SDan Kimmel * the data into it; otherwise, we share the data directly if we can. 5463dcbf3bd6SGeorge Wilson */ 5464*770499e1SDan Kimmel if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) { 5465*770499e1SDan Kimmel arc_hdr_alloc_pabd(hdr); 5466*770499e1SDan Kimmel 5467*770499e1SDan Kimmel /* 5468*770499e1SDan Kimmel * Ideally, we would always copy the io_abd into b_pabd, but the 5469*770499e1SDan Kimmel * user may have disabled compressed ARC, thus we must check the 5470*770499e1SDan Kimmel * hdr's compression setting rather than the io_bp's. 5471*770499e1SDan Kimmel */ 5472*770499e1SDan Kimmel if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5473*770499e1SDan Kimmel ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=, 5474*770499e1SDan Kimmel ZIO_COMPRESS_OFF); 5475*770499e1SDan Kimmel ASSERT3U(psize, >, 0); 5476*770499e1SDan Kimmel 5477*770499e1SDan Kimmel abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize); 5478*770499e1SDan Kimmel } else { 5479*770499e1SDan Kimmel ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr)); 5480*770499e1SDan Kimmel 5481*770499e1SDan Kimmel abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data, 5482*770499e1SDan Kimmel arc_buf_size(buf)); 5483*770499e1SDan Kimmel } 5484dcbf3bd6SGeorge Wilson } else { 5485*770499e1SDan Kimmel ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd)); 54865602294fSDan Kimmel ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf)); 5487dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5488dcbf3bd6SGeorge Wilson 5489dcbf3bd6SGeorge Wilson arc_share_buf(hdr, buf); 5490dcbf3bd6SGeorge Wilson } 5491*770499e1SDan Kimmel 5492dcbf3bd6SGeorge Wilson arc_hdr_verify(hdr, zio->io_bp); 5493c717a561Smaybee } 5494c717a561Smaybee 54958df0bcf0SPaul Dagnelie static void 54968df0bcf0SPaul Dagnelie arc_write_children_ready(zio_t *zio) 54978df0bcf0SPaul Dagnelie { 54988df0bcf0SPaul Dagnelie arc_write_callback_t *callback = zio->io_private; 54998df0bcf0SPaul Dagnelie arc_buf_t *buf = callback->awcb_buf; 55008df0bcf0SPaul Dagnelie 55018df0bcf0SPaul Dagnelie callback->awcb_children_ready(zio, buf, callback->awcb_private); 55028df0bcf0SPaul Dagnelie } 55038df0bcf0SPaul Dagnelie 550469962b56SMatthew Ahrens /* 550569962b56SMatthew Ahrens * The SPA calls this callback for each physical write that happens on behalf 550669962b56SMatthew Ahrens * of a logical write. See the comment in dbuf_write_physdone() for details. 550769962b56SMatthew Ahrens */ 550869962b56SMatthew Ahrens static void 550969962b56SMatthew Ahrens arc_write_physdone(zio_t *zio) 551069962b56SMatthew Ahrens { 551169962b56SMatthew Ahrens arc_write_callback_t *cb = zio->io_private; 551269962b56SMatthew Ahrens if (cb->awcb_physdone != NULL) 551369962b56SMatthew Ahrens cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 551469962b56SMatthew Ahrens } 551569962b56SMatthew Ahrens 5516fa9e4066Sahrens static void 5517fa9e4066Sahrens arc_write_done(zio_t *zio) 5518fa9e4066Sahrens { 5519c717a561Smaybee arc_write_callback_t *callback = zio->io_private; 5520c717a561Smaybee arc_buf_t *buf = callback->awcb_buf; 5521c717a561Smaybee arc_buf_hdr_t *hdr = buf->b_hdr; 5522fa9e4066Sahrens 5523dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5524b24ab676SJeff Bonwick 5525b24ab676SJeff Bonwick if (zio->io_error == 0) { 5526dcbf3bd6SGeorge Wilson arc_hdr_verify(hdr, zio->io_bp); 5527dcbf3bd6SGeorge Wilson 55285d7b4d43SMatthew Ahrens if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 552943466aaeSMax Grossman buf_discard_identity(hdr); 553043466aaeSMax Grossman } else { 553143466aaeSMax Grossman hdr->b_dva = *BP_IDENTITY(zio->io_bp); 553243466aaeSMax Grossman hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 553343466aaeSMax Grossman } 5534b24ab676SJeff Bonwick } else { 5535dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 5536b24ab676SJeff Bonwick } 5537fa9e4066Sahrens 5538ea8dc4b6Seschrock /* 55395d7b4d43SMatthew Ahrens * If the block to be written was all-zero or compressed enough to be 55405d7b4d43SMatthew Ahrens * embedded in the BP, no write was performed so there will be no 55415d7b4d43SMatthew Ahrens * dva/birth/checksum. The buffer must therefore remain anonymous 55425d7b4d43SMatthew Ahrens * (and uncached). 5543ea8dc4b6Seschrock */ 5544dcbf3bd6SGeorge Wilson if (!HDR_EMPTY(hdr)) { 5545fa9e4066Sahrens arc_buf_hdr_t *exists; 5546fa9e4066Sahrens kmutex_t *hash_lock; 5547fa9e4066Sahrens 55485602294fSDan Kimmel ASSERT3U(zio->io_error, ==, 0); 5549b24ab676SJeff Bonwick 55506b4acc8bSahrens arc_cksum_verify(buf); 55516b4acc8bSahrens 5552fa9e4066Sahrens exists = buf_hash_insert(hdr, &hash_lock); 555389c86e32SChris Williamson if (exists != NULL) { 5554fa9e4066Sahrens /* 5555fa9e4066Sahrens * This can only happen if we overwrite for 5556fa9e4066Sahrens * sync-to-convergence, because we remove 5557fa9e4066Sahrens * buffers from the hash table when we arc_free(). 5558fa9e4066Sahrens */ 5559b24ab676SJeff Bonwick if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 5560b24ab676SJeff Bonwick if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5561b24ab676SJeff Bonwick panic("bad overwrite, hdr=%p exists=%p", 5562b24ab676SJeff Bonwick (void *)hdr, (void *)exists); 556389c86e32SChris Williamson ASSERT(refcount_is_zero( 556489c86e32SChris Williamson &exists->b_l1hdr.b_refcnt)); 5565b24ab676SJeff Bonwick arc_change_state(arc_anon, exists, hash_lock); 5566b24ab676SJeff Bonwick mutex_exit(hash_lock); 5567b24ab676SJeff Bonwick arc_hdr_destroy(exists); 5568b24ab676SJeff Bonwick exists = buf_hash_insert(hdr, &hash_lock); 5569b24ab676SJeff Bonwick ASSERT3P(exists, ==, NULL); 557080901aeaSGeorge Wilson } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 557180901aeaSGeorge Wilson /* nopwrite */ 557280901aeaSGeorge Wilson ASSERT(zio->io_prop.zp_nopwrite); 557380901aeaSGeorge Wilson if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 557480901aeaSGeorge Wilson panic("bad nopwrite, hdr=%p exists=%p", 557580901aeaSGeorge Wilson (void *)hdr, (void *)exists); 5576b24ab676SJeff Bonwick } else { 5577b24ab676SJeff Bonwick /* Dedup */ 5578dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt == 1); 557989c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5580b24ab676SJeff Bonwick ASSERT(BP_GET_DEDUP(zio->io_bp)); 5581b24ab676SJeff Bonwick ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5582ae46e4c7SMatthew Ahrens } 5583fa9e4066Sahrens } 5584dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5585088f3894Sahrens /* if it's not anon, we are doing a scrub */ 558689c86e32SChris Williamson if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5587088f3894Sahrens arc_access(hdr, hash_lock); 558844eda4d7Smaybee mutex_exit(hash_lock); 5589ea8dc4b6Seschrock } else { 5590dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5591fa9e4066Sahrens } 5592ea8dc4b6Seschrock 559389c86e32SChris Williamson ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5594b24ab676SJeff Bonwick callback->awcb_done(zio, buf, callback->awcb_private); 5595fa9e4066Sahrens 5596*770499e1SDan Kimmel abd_put(zio->io_abd); 5597c717a561Smaybee kmem_free(callback, sizeof (arc_write_callback_t)); 5598fa9e4066Sahrens } 5599fa9e4066Sahrens 5600c717a561Smaybee zio_t * 5601dcbf3bd6SGeorge Wilson arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 5602dcbf3bd6SGeorge Wilson boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready, 56038df0bcf0SPaul Dagnelie arc_done_func_t *children_ready, arc_done_func_t *physdone, 560469962b56SMatthew Ahrens arc_done_func_t *done, void *private, zio_priority_t priority, 56057802d7bfSMatthew Ahrens int zio_flags, const zbookmark_phys_t *zb) 5606fa9e4066Sahrens { 5607fa9e4066Sahrens arc_buf_hdr_t *hdr = buf->b_hdr; 5608c717a561Smaybee arc_write_callback_t *callback; 5609e14bb325SJeff Bonwick zio_t *zio; 5610fa9e4066Sahrens 5611dcbf3bd6SGeorge Wilson ASSERT3P(ready, !=, NULL); 5612dcbf3bd6SGeorge Wilson ASSERT3P(done, !=, NULL); 5613fa9e4066Sahrens ASSERT(!HDR_IO_ERROR(hdr)); 561489c86e32SChris Williamson ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5615dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5616dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 56173baa08fcSek if (l2arc) 5618dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 56195602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 56205602294fSDan Kimmel ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_OFF); 56215602294fSDan Kimmel ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf)); 56225602294fSDan Kimmel zio_flags |= ZIO_FLAG_RAW; 56235602294fSDan Kimmel } 5624c717a561Smaybee callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5625c717a561Smaybee callback->awcb_ready = ready; 56268df0bcf0SPaul Dagnelie callback->awcb_children_ready = children_ready; 562769962b56SMatthew Ahrens callback->awcb_physdone = physdone; 5628c717a561Smaybee callback->awcb_done = done; 5629c717a561Smaybee callback->awcb_private = private; 5630c717a561Smaybee callback->awcb_buf = buf; 5631088f3894Sahrens 5632dcbf3bd6SGeorge Wilson /* 5633*770499e1SDan Kimmel * The hdr's b_pabd is now stale, free it now. A new data block 5634dcbf3bd6SGeorge Wilson * will be allocated when the zio pipeline calls arc_write_ready(). 5635dcbf3bd6SGeorge Wilson */ 5636*770499e1SDan Kimmel if (hdr->b_l1hdr.b_pabd != NULL) { 5637dcbf3bd6SGeorge Wilson /* 5638dcbf3bd6SGeorge Wilson * If the buf is currently sharing the data block with 5639dcbf3bd6SGeorge Wilson * the hdr then we need to break that relationship here. 5640dcbf3bd6SGeorge Wilson * The hdr will remain with a NULL data pointer and the 5641dcbf3bd6SGeorge Wilson * buf will take sole ownership of the block. 5642dcbf3bd6SGeorge Wilson */ 5643dcbf3bd6SGeorge Wilson if (arc_buf_is_shared(buf)) { 5644dcbf3bd6SGeorge Wilson arc_unshare_buf(hdr, buf); 5645dcbf3bd6SGeorge Wilson } else { 5646*770499e1SDan Kimmel arc_hdr_free_pabd(hdr); 5647dcbf3bd6SGeorge Wilson } 5648dcbf3bd6SGeorge Wilson VERIFY3P(buf->b_data, !=, NULL); 5649dcbf3bd6SGeorge Wilson arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); 5650dcbf3bd6SGeorge Wilson } 5651dcbf3bd6SGeorge Wilson ASSERT(!arc_buf_is_shared(buf)); 5652*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5653dcbf3bd6SGeorge Wilson 5654*770499e1SDan Kimmel zio = zio_write(pio, spa, txg, bp, 5655*770499e1SDan Kimmel abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)), 56565602294fSDan Kimmel HDR_GET_LSIZE(hdr), arc_buf_size(buf), zp, arc_write_ready, 56578df0bcf0SPaul Dagnelie (children_ready != NULL) ? arc_write_children_ready : NULL, 56588df0bcf0SPaul Dagnelie arc_write_physdone, arc_write_done, callback, 565969962b56SMatthew Ahrens priority, zio_flags, zb); 5660fa9e4066Sahrens 5661c717a561Smaybee return (zio); 5662fa9e4066Sahrens } 5663fa9e4066Sahrens 56641ab7f2deSmaybee static int 566569962b56SMatthew Ahrens arc_memory_throttle(uint64_t reserve, uint64_t txg) 56661ab7f2deSmaybee { 56671ab7f2deSmaybee #ifdef _KERNEL 56681ab7f2deSmaybee uint64_t available_memory = ptob(freemem); 56691ab7f2deSmaybee static uint64_t page_load = 0; 56701ab7f2deSmaybee static uint64_t last_txg = 0; 56711ab7f2deSmaybee 56721ab7f2deSmaybee #if defined(__i386) 56731ab7f2deSmaybee available_memory = 56741ab7f2deSmaybee MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 56751ab7f2deSmaybee #endif 567669962b56SMatthew Ahrens 567769962b56SMatthew Ahrens if (freemem > physmem * arc_lotsfree_percent / 100) 56781ab7f2deSmaybee return (0); 56791ab7f2deSmaybee 56801ab7f2deSmaybee if (txg > last_txg) { 56811ab7f2deSmaybee last_txg = txg; 56821ab7f2deSmaybee page_load = 0; 56831ab7f2deSmaybee } 56841ab7f2deSmaybee /* 56851ab7f2deSmaybee * If we are in pageout, we know that memory is already tight, 56861ab7f2deSmaybee * the arc is already going to be evicting, so we just want to 56871ab7f2deSmaybee * continue to let page writes occur as quickly as possible. 56881ab7f2deSmaybee */ 56891ab7f2deSmaybee if (curproc == proc_pageout) { 56901ab7f2deSmaybee if (page_load > MAX(ptob(minfree), available_memory) / 4) 5691be6fd75aSMatthew Ahrens return (SET_ERROR(ERESTART)); 56921ab7f2deSmaybee /* Note: reserve is inflated, so we deflate */ 56931ab7f2deSmaybee page_load += reserve / 8; 56941ab7f2deSmaybee return (0); 56951ab7f2deSmaybee } else if (page_load > 0 && arc_reclaim_needed()) { 56961ab7f2deSmaybee /* memory is low, delay before restarting */ 56971ab7f2deSmaybee ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5698be6fd75aSMatthew Ahrens return (SET_ERROR(EAGAIN)); 56991ab7f2deSmaybee } 57001ab7f2deSmaybee page_load = 0; 57011ab7f2deSmaybee #endif 57021ab7f2deSmaybee return (0); 57031ab7f2deSmaybee } 57041ab7f2deSmaybee 5705fa9e4066Sahrens void 57061ab7f2deSmaybee arc_tempreserve_clear(uint64_t reserve) 5707fa9e4066Sahrens { 57081ab7f2deSmaybee atomic_add_64(&arc_tempreserve, -reserve); 5709fa9e4066Sahrens ASSERT((int64_t)arc_tempreserve >= 0); 5710fa9e4066Sahrens } 5711fa9e4066Sahrens 5712fa9e4066Sahrens int 57131ab7f2deSmaybee arc_tempreserve_space(uint64_t reserve, uint64_t txg) 5714fa9e4066Sahrens { 57151ab7f2deSmaybee int error; 57162fdbea25SAleksandr Guzovskiy uint64_t anon_size; 57171ab7f2deSmaybee 57181ab7f2deSmaybee if (reserve > arc_c/4 && !arc_no_grow) 57191ab7f2deSmaybee arc_c = MIN(arc_c_max, reserve * 4); 57201ab7f2deSmaybee if (reserve > arc_c) 5721be6fd75aSMatthew Ahrens return (SET_ERROR(ENOMEM)); 5722112fe045Smaybee 57232fdbea25SAleksandr Guzovskiy /* 57242fdbea25SAleksandr Guzovskiy * Don't count loaned bufs as in flight dirty data to prevent long 57252fdbea25SAleksandr Guzovskiy * network delays from blocking transactions that are ready to be 57262fdbea25SAleksandr Guzovskiy * assigned to a txg. 57272fdbea25SAleksandr Guzovskiy */ 57285602294fSDan Kimmel 57295602294fSDan Kimmel /* assert that it has not wrapped around */ 57305602294fSDan Kimmel ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 57315602294fSDan Kimmel 57322fd872a7SPrakash Surya anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 57332fd872a7SPrakash Surya arc_loaned_bytes), 0); 57342fdbea25SAleksandr Guzovskiy 57351ab7f2deSmaybee /* 57361ab7f2deSmaybee * Writes will, almost always, require additional memory allocations 5737f7170741SWill Andrews * in order to compress/encrypt/etc the data. We therefore need to 57381ab7f2deSmaybee * make sure that there is sufficient available memory for this. 57391ab7f2deSmaybee */ 574069962b56SMatthew Ahrens error = arc_memory_throttle(reserve, txg); 574169962b56SMatthew Ahrens if (error != 0) 57421ab7f2deSmaybee return (error); 57431ab7f2deSmaybee 5744fa9e4066Sahrens /* 5745112fe045Smaybee * Throttle writes when the amount of dirty data in the cache 5746112fe045Smaybee * gets too large. We try to keep the cache less than half full 5747112fe045Smaybee * of dirty blocks so that our sync times don't grow too large. 5748112fe045Smaybee * Note: if two requests come in concurrently, we might let them 5749112fe045Smaybee * both succeed, when one of them should fail. Not a huge deal. 5750fa9e4066Sahrens */ 57512fdbea25SAleksandr Guzovskiy 57522fdbea25SAleksandr Guzovskiy if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 57532fdbea25SAleksandr Guzovskiy anon_size > arc_c / 4) { 5754dcbf3bd6SGeorge Wilson uint64_t meta_esize = 5755dcbf3bd6SGeorge Wilson refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5756dcbf3bd6SGeorge Wilson uint64_t data_esize = 5757dcbf3bd6SGeorge Wilson refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 57580e8c6158Smaybee dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 57590e8c6158Smaybee "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5760dcbf3bd6SGeorge Wilson arc_tempreserve >> 10, meta_esize >> 10, 5761dcbf3bd6SGeorge Wilson data_esize >> 10, reserve >> 10, arc_c >> 10); 5762be6fd75aSMatthew Ahrens return (SET_ERROR(ERESTART)); 5763fa9e4066Sahrens } 57641ab7f2deSmaybee atomic_add_64(&arc_tempreserve, reserve); 5765fa9e4066Sahrens return (0); 5766fa9e4066Sahrens } 5767fa9e4066Sahrens 57684076b1bfSPrakash Surya static void 57694076b1bfSPrakash Surya arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 57704076b1bfSPrakash Surya kstat_named_t *evict_data, kstat_named_t *evict_metadata) 57714076b1bfSPrakash Surya { 57722fd872a7SPrakash Surya size->value.ui64 = refcount_count(&state->arcs_size); 5773dcbf3bd6SGeorge Wilson evict_data->value.ui64 = 5774dcbf3bd6SGeorge Wilson refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); 5775dcbf3bd6SGeorge Wilson evict_metadata->value.ui64 = 5776dcbf3bd6SGeorge Wilson refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); 57774076b1bfSPrakash Surya } 57784076b1bfSPrakash Surya 57794076b1bfSPrakash Surya static int 57804076b1bfSPrakash Surya arc_kstat_update(kstat_t *ksp, int rw) 57814076b1bfSPrakash Surya { 57824076b1bfSPrakash Surya arc_stats_t *as = ksp->ks_data; 57834076b1bfSPrakash Surya 57844076b1bfSPrakash Surya if (rw == KSTAT_WRITE) { 57854076b1bfSPrakash Surya return (EACCES); 57864076b1bfSPrakash Surya } else { 57874076b1bfSPrakash Surya arc_kstat_update_state(arc_anon, 57884076b1bfSPrakash Surya &as->arcstat_anon_size, 57894076b1bfSPrakash Surya &as->arcstat_anon_evictable_data, 57904076b1bfSPrakash Surya &as->arcstat_anon_evictable_metadata); 57914076b1bfSPrakash Surya arc_kstat_update_state(arc_mru, 57924076b1bfSPrakash Surya &as->arcstat_mru_size, 57934076b1bfSPrakash Surya &as->arcstat_mru_evictable_data, 57944076b1bfSPrakash Surya &as->arcstat_mru_evictable_metadata); 57954076b1bfSPrakash Surya arc_kstat_update_state(arc_mru_ghost, 57964076b1bfSPrakash Surya &as->arcstat_mru_ghost_size, 57974076b1bfSPrakash Surya &as->arcstat_mru_ghost_evictable_data, 57984076b1bfSPrakash Surya &as->arcstat_mru_ghost_evictable_metadata); 57994076b1bfSPrakash Surya arc_kstat_update_state(arc_mfu, 58004076b1bfSPrakash Surya &as->arcstat_mfu_size, 58014076b1bfSPrakash Surya &as->arcstat_mfu_evictable_data, 58024076b1bfSPrakash Surya &as->arcstat_mfu_evictable_metadata); 58034076b1bfSPrakash Surya arc_kstat_update_state(arc_mfu_ghost, 58044076b1bfSPrakash Surya &as->arcstat_mfu_ghost_size, 58054076b1bfSPrakash Surya &as->arcstat_mfu_ghost_evictable_data, 58064076b1bfSPrakash Surya &as->arcstat_mfu_ghost_evictable_metadata); 58074076b1bfSPrakash Surya } 58084076b1bfSPrakash Surya 58094076b1bfSPrakash Surya return (0); 58104076b1bfSPrakash Surya } 58114076b1bfSPrakash Surya 5812dcbf3bd6SGeorge Wilson /* 5813dcbf3bd6SGeorge Wilson * This function *must* return indices evenly distributed between all 5814dcbf3bd6SGeorge Wilson * sublists of the multilist. This is needed due to how the ARC eviction 5815dcbf3bd6SGeorge Wilson * code is laid out; arc_evict_state() assumes ARC buffers are evenly 5816dcbf3bd6SGeorge Wilson * distributed between all sublists and uses this assumption when 5817dcbf3bd6SGeorge Wilson * deciding which sublist to evict from and how much to evict from it. 5818dcbf3bd6SGeorge Wilson */ 5819dcbf3bd6SGeorge Wilson unsigned int 5820dcbf3bd6SGeorge Wilson arc_state_multilist_index_func(multilist_t *ml, void *obj) 5821dcbf3bd6SGeorge Wilson { 5822dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = obj; 5823dcbf3bd6SGeorge Wilson 5824dcbf3bd6SGeorge Wilson /* 5825dcbf3bd6SGeorge Wilson * We rely on b_dva to generate evenly distributed index 5826dcbf3bd6SGeorge Wilson * numbers using buf_hash below. So, as an added precaution, 5827dcbf3bd6SGeorge Wilson * let's make sure we never add empty buffers to the arc lists. 5828dcbf3bd6SGeorge Wilson */ 5829dcbf3bd6SGeorge Wilson ASSERT(!HDR_EMPTY(hdr)); 5830dcbf3bd6SGeorge Wilson 5831dcbf3bd6SGeorge Wilson /* 5832dcbf3bd6SGeorge Wilson * The assumption here, is the hash value for a given 5833dcbf3bd6SGeorge Wilson * arc_buf_hdr_t will remain constant throughout it's lifetime 5834dcbf3bd6SGeorge Wilson * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 5835dcbf3bd6SGeorge Wilson * Thus, we don't need to store the header's sublist index 5836dcbf3bd6SGeorge Wilson * on insertion, as this index can be recalculated on removal. 5837dcbf3bd6SGeorge Wilson * 5838dcbf3bd6SGeorge Wilson * Also, the low order bits of the hash value are thought to be 5839dcbf3bd6SGeorge Wilson * distributed evenly. Otherwise, in the case that the multilist 5840dcbf3bd6SGeorge Wilson * has a power of two number of sublists, each sublists' usage 5841dcbf3bd6SGeorge Wilson * would not be evenly distributed. 5842dcbf3bd6SGeorge Wilson */ 5843dcbf3bd6SGeorge Wilson return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 5844dcbf3bd6SGeorge Wilson multilist_get_num_sublists(ml)); 5845dcbf3bd6SGeorge Wilson } 5846dcbf3bd6SGeorge Wilson 5847dcbf3bd6SGeorge Wilson static void 5848dcbf3bd6SGeorge Wilson arc_state_init(void) 5849dcbf3bd6SGeorge Wilson { 5850dcbf3bd6SGeorge Wilson arc_anon = &ARC_anon; 5851dcbf3bd6SGeorge Wilson arc_mru = &ARC_mru; 5852dcbf3bd6SGeorge Wilson arc_mru_ghost = &ARC_mru_ghost; 5853dcbf3bd6SGeorge Wilson arc_mfu = &ARC_mfu; 5854dcbf3bd6SGeorge Wilson arc_mfu_ghost = &ARC_mfu_ghost; 5855dcbf3bd6SGeorge Wilson arc_l2c_only = &ARC_l2c_only; 5856dcbf3bd6SGeorge Wilson 585794c2d0ebSMatthew Ahrens arc_mru->arcs_list[ARC_BUFC_METADATA] = 585894c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5859dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 586010fbdecbSMatthew Ahrens arc_state_multilist_index_func); 586194c2d0ebSMatthew Ahrens arc_mru->arcs_list[ARC_BUFC_DATA] = 586294c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5863dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 586410fbdecbSMatthew Ahrens arc_state_multilist_index_func); 586594c2d0ebSMatthew Ahrens arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] = 586694c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5867dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 586810fbdecbSMatthew Ahrens arc_state_multilist_index_func); 586994c2d0ebSMatthew Ahrens arc_mru_ghost->arcs_list[ARC_BUFC_DATA] = 587094c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5871dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 587210fbdecbSMatthew Ahrens arc_state_multilist_index_func); 587394c2d0ebSMatthew Ahrens arc_mfu->arcs_list[ARC_BUFC_METADATA] = 587494c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5875dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 587610fbdecbSMatthew Ahrens arc_state_multilist_index_func); 587794c2d0ebSMatthew Ahrens arc_mfu->arcs_list[ARC_BUFC_DATA] = 587894c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5879dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 588010fbdecbSMatthew Ahrens arc_state_multilist_index_func); 588194c2d0ebSMatthew Ahrens arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] = 588294c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5883dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 588410fbdecbSMatthew Ahrens arc_state_multilist_index_func); 588594c2d0ebSMatthew Ahrens arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] = 588694c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5887dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 588810fbdecbSMatthew Ahrens arc_state_multilist_index_func); 588994c2d0ebSMatthew Ahrens arc_l2c_only->arcs_list[ARC_BUFC_METADATA] = 589094c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5891dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 589210fbdecbSMatthew Ahrens arc_state_multilist_index_func); 589394c2d0ebSMatthew Ahrens arc_l2c_only->arcs_list[ARC_BUFC_DATA] = 589494c2d0ebSMatthew Ahrens multilist_create(sizeof (arc_buf_hdr_t), 5895dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 589610fbdecbSMatthew Ahrens arc_state_multilist_index_func); 5897dcbf3bd6SGeorge Wilson 5898dcbf3bd6SGeorge Wilson refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5899dcbf3bd6SGeorge Wilson refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5900dcbf3bd6SGeorge Wilson refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 5901dcbf3bd6SGeorge Wilson refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 5902dcbf3bd6SGeorge Wilson refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 5903dcbf3bd6SGeorge Wilson refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 5904dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 5905dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 5906dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 5907dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 5908dcbf3bd6SGeorge Wilson refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 5909dcbf3bd6SGeorge Wilson refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 5910dcbf3bd6SGeorge Wilson 5911dcbf3bd6SGeorge Wilson refcount_create(&arc_anon->arcs_size); 5912dcbf3bd6SGeorge Wilson refcount_create(&arc_mru->arcs_size); 5913dcbf3bd6SGeorge Wilson refcount_create(&arc_mru_ghost->arcs_size); 5914dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu->arcs_size); 5915dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu_ghost->arcs_size); 5916dcbf3bd6SGeorge Wilson refcount_create(&arc_l2c_only->arcs_size); 5917dcbf3bd6SGeorge Wilson } 5918dcbf3bd6SGeorge Wilson 5919dcbf3bd6SGeorge Wilson static void 5920dcbf3bd6SGeorge Wilson arc_state_fini(void) 5921dcbf3bd6SGeorge Wilson { 5922dcbf3bd6SGeorge Wilson refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5923dcbf3bd6SGeorge Wilson refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5924dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 5925dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 5926dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 5927dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 5928dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 5929dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 5930dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 5931dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 5932dcbf3bd6SGeorge Wilson refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 5933dcbf3bd6SGeorge Wilson refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 5934dcbf3bd6SGeorge Wilson 5935dcbf3bd6SGeorge Wilson refcount_destroy(&arc_anon->arcs_size); 5936dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru->arcs_size); 5937dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru_ghost->arcs_size); 5938dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu->arcs_size); 5939dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu_ghost->arcs_size); 5940dcbf3bd6SGeorge Wilson refcount_destroy(&arc_l2c_only->arcs_size); 5941dcbf3bd6SGeorge Wilson 594294c2d0ebSMatthew Ahrens multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]); 594394c2d0ebSMatthew Ahrens multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 594494c2d0ebSMatthew Ahrens multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]); 594594c2d0ebSMatthew Ahrens multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 594694c2d0ebSMatthew Ahrens multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]); 594794c2d0ebSMatthew Ahrens multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 594894c2d0ebSMatthew Ahrens multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]); 594994c2d0ebSMatthew Ahrens multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 5950dcbf3bd6SGeorge Wilson } 5951dcbf3bd6SGeorge Wilson 5952dcbf3bd6SGeorge Wilson uint64_t 5953dcbf3bd6SGeorge Wilson arc_max_bytes(void) 5954244781f1SPrakash Surya { 5955dcbf3bd6SGeorge Wilson return (arc_c_max); 5956244781f1SPrakash Surya } 5957244781f1SPrakash Surya 5958fa9e4066Sahrens void 5959fa9e4066Sahrens arc_init(void) 5960fa9e4066Sahrens { 59612ec99e3eSMatthew Ahrens /* 59622ec99e3eSMatthew Ahrens * allmem is "all memory that we could possibly use". 59632ec99e3eSMatthew Ahrens */ 59642ec99e3eSMatthew Ahrens #ifdef _KERNEL 59652ec99e3eSMatthew Ahrens uint64_t allmem = ptob(physmem - swapfs_minfree); 59662ec99e3eSMatthew Ahrens #else 59672ec99e3eSMatthew Ahrens uint64_t allmem = (physmem * PAGESIZE) / 2; 59682ec99e3eSMatthew Ahrens #endif 59692ec99e3eSMatthew Ahrens 5970244781f1SPrakash Surya mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 5971244781f1SPrakash Surya cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 5972244781f1SPrakash Surya cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 5973244781f1SPrakash Surya 597413506d1eSmaybee /* Convert seconds to clock ticks */ 5975b19a79ecSperrin arc_min_prefetch_lifespan = 1 * hz; 597613506d1eSmaybee 5977112fe045Smaybee /* set min cache to 1/32 of all memory, or 64MB, whichever is more */ 59782ec99e3eSMatthew Ahrens arc_c_min = MAX(allmem / 32, 64 << 20); 5979112fe045Smaybee /* set max to 3/4 of all memory, or all but 1GB, whichever is more */ 59802ec99e3eSMatthew Ahrens if (allmem >= 1 << 30) 59812ec99e3eSMatthew Ahrens arc_c_max = allmem - (1 << 30); 5982fa9e4066Sahrens else 598344cb6abcSbmc arc_c_max = arc_c_min; 59842ec99e3eSMatthew Ahrens arc_c_max = MAX(allmem * 3 / 4, arc_c_max); 5985a2eea2e1Sahrens 59868fe00bfbSMatthew Ahrens /* 59878fe00bfbSMatthew Ahrens * In userland, there's only the memory pressure that we artificially 59888fe00bfbSMatthew Ahrens * create (see arc_available_memory()). Don't let arc_c get too 59898fe00bfbSMatthew Ahrens * small, because it can cause transactions to be larger than 59908fe00bfbSMatthew Ahrens * arc_c, causing arc_tempreserve_space() to fail. 59918fe00bfbSMatthew Ahrens */ 59928fe00bfbSMatthew Ahrens #ifndef _KERNEL 59938fe00bfbSMatthew Ahrens arc_c_min = arc_c_max / 2; 59948fe00bfbSMatthew Ahrens #endif 59958fe00bfbSMatthew Ahrens 5996a2eea2e1Sahrens /* 5997a2eea2e1Sahrens * Allow the tunables to override our calculations if they are 5998a2eea2e1Sahrens * reasonable (ie. over 64MB) 5999a2eea2e1Sahrens */ 6000e5961710SMatthew Ahrens if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) { 600144cb6abcSbmc arc_c_max = zfs_arc_max; 6002e5961710SMatthew Ahrens arc_c_min = MIN(arc_c_min, arc_c_max); 6003e5961710SMatthew Ahrens } 60042ec99e3eSMatthew Ahrens if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max) 600544cb6abcSbmc arc_c_min = zfs_arc_min; 6006a2eea2e1Sahrens 600744cb6abcSbmc arc_c = arc_c_max; 600844cb6abcSbmc arc_p = (arc_c >> 1); 6009dcbf3bd6SGeorge Wilson arc_size = 0; 6010fa9e4066Sahrens 60110e8c6158Smaybee /* limit meta-data to 1/4 of the arc capacity */ 60120e8c6158Smaybee arc_meta_limit = arc_c_max / 4; 60131116048bSek 6014af868f46SMatthew Ahrens #ifdef _KERNEL 6015af868f46SMatthew Ahrens /* 6016af868f46SMatthew Ahrens * Metadata is stored in the kernel's heap. Don't let us 6017af868f46SMatthew Ahrens * use more than half the heap for the ARC. 6018af868f46SMatthew Ahrens */ 6019af868f46SMatthew Ahrens arc_meta_limit = MIN(arc_meta_limit, 6020af868f46SMatthew Ahrens vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2); 6021af868f46SMatthew Ahrens #endif 6022af868f46SMatthew Ahrens 60231116048bSek /* Allow the tunable to override if it is reasonable */ 60241116048bSek if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 60251116048bSek arc_meta_limit = zfs_arc_meta_limit; 60261116048bSek 60270e8c6158Smaybee if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 60280e8c6158Smaybee arc_c_min = arc_meta_limit / 2; 60290e8c6158Smaybee 60303a5286a1SMatthew Ahrens if (zfs_arc_meta_min > 0) { 60313a5286a1SMatthew Ahrens arc_meta_min = zfs_arc_meta_min; 60323a5286a1SMatthew Ahrens } else { 60333a5286a1SMatthew Ahrens arc_meta_min = arc_c_min / 2; 60343a5286a1SMatthew Ahrens } 60353a5286a1SMatthew Ahrens 60365a98e54bSBrendan Gregg - Sun Microsystems if (zfs_arc_grow_retry > 0) 60375a98e54bSBrendan Gregg - Sun Microsystems arc_grow_retry = zfs_arc_grow_retry; 60385a98e54bSBrendan Gregg - Sun Microsystems 60395a98e54bSBrendan Gregg - Sun Microsystems if (zfs_arc_shrink_shift > 0) 60405a98e54bSBrendan Gregg - Sun Microsystems arc_shrink_shift = zfs_arc_shrink_shift; 60415a98e54bSBrendan Gregg - Sun Microsystems 60422ec99e3eSMatthew Ahrens /* 60432ec99e3eSMatthew Ahrens * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 60442ec99e3eSMatthew Ahrens */ 60452ec99e3eSMatthew Ahrens if (arc_no_grow_shift >= arc_shrink_shift) 60462ec99e3eSMatthew Ahrens arc_no_grow_shift = arc_shrink_shift - 1; 60472ec99e3eSMatthew Ahrens 60485a98e54bSBrendan Gregg - Sun Microsystems if (zfs_arc_p_min_shift > 0) 60495a98e54bSBrendan Gregg - Sun Microsystems arc_p_min_shift = zfs_arc_p_min_shift; 60505a98e54bSBrendan Gregg - Sun Microsystems 6051fa9e4066Sahrens /* if kmem_flags are set, lets try to use less memory */ 6052fa9e4066Sahrens if (kmem_debugging()) 605344cb6abcSbmc arc_c = arc_c / 2; 605444cb6abcSbmc if (arc_c < arc_c_min) 605544cb6abcSbmc arc_c = arc_c_min; 605644cb6abcSbmc 6057dcbf3bd6SGeorge Wilson arc_state_init(); 6058fa9e4066Sahrens buf_init(); 6059fa9e4066Sahrens 6060dcbf3bd6SGeorge Wilson arc_reclaim_thread_exit = B_FALSE; 6061fa9e4066Sahrens 606244cb6abcSbmc arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 606344cb6abcSbmc sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 606444cb6abcSbmc 606544cb6abcSbmc if (arc_ksp != NULL) { 606644cb6abcSbmc arc_ksp->ks_data = &arc_stats; 60674076b1bfSPrakash Surya arc_ksp->ks_update = arc_kstat_update; 606844cb6abcSbmc kstat_install(arc_ksp); 606944cb6abcSbmc } 607044cb6abcSbmc 6071fa9e4066Sahrens (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 6072fa9e4066Sahrens TS_RUN, minclsyspri); 607349e3519aSmaybee 6074dcbf3bd6SGeorge Wilson arc_dead = B_FALSE; 60753a737e0dSbrendan arc_warm = B_FALSE; 60761ab7f2deSmaybee 607769962b56SMatthew Ahrens /* 607869962b56SMatthew Ahrens * Calculate maximum amount of dirty data per pool. 607969962b56SMatthew Ahrens * 608069962b56SMatthew Ahrens * If it has been set by /etc/system, take that. 608169962b56SMatthew Ahrens * Otherwise, use a percentage of physical memory defined by 608269962b56SMatthew Ahrens * zfs_dirty_data_max_percent (default 10%) with a cap at 608369962b56SMatthew Ahrens * zfs_dirty_data_max_max (default 4GB). 608469962b56SMatthew Ahrens */ 608569962b56SMatthew Ahrens if (zfs_dirty_data_max == 0) { 608669962b56SMatthew Ahrens zfs_dirty_data_max = physmem * PAGESIZE * 608769962b56SMatthew Ahrens zfs_dirty_data_max_percent / 100; 608869962b56SMatthew Ahrens zfs_dirty_data_max = MIN(zfs_dirty_data_max, 608969962b56SMatthew Ahrens zfs_dirty_data_max_max); 609069962b56SMatthew Ahrens } 6091fa9e4066Sahrens } 6092fa9e4066Sahrens 6093fa9e4066Sahrens void 6094fa9e4066Sahrens arc_fini(void) 6095fa9e4066Sahrens { 6096244781f1SPrakash Surya mutex_enter(&arc_reclaim_lock); 6097dcbf3bd6SGeorge Wilson arc_reclaim_thread_exit = B_TRUE; 6098244781f1SPrakash Surya /* 6099244781f1SPrakash Surya * The reclaim thread will set arc_reclaim_thread_exit back to 6100dcbf3bd6SGeorge Wilson * B_FALSE when it is finished exiting; we're waiting for that. 6101244781f1SPrakash Surya */ 6102244781f1SPrakash Surya while (arc_reclaim_thread_exit) { 6103244781f1SPrakash Surya cv_signal(&arc_reclaim_thread_cv); 6104244781f1SPrakash Surya cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 6105244781f1SPrakash Surya } 6106244781f1SPrakash Surya mutex_exit(&arc_reclaim_lock); 6107fa9e4066Sahrens 6108dcbf3bd6SGeorge Wilson /* Use B_TRUE to ensure *all* buffers are evicted */ 6109dcbf3bd6SGeorge Wilson arc_flush(NULL, B_TRUE); 6110fa9e4066Sahrens 6111dcbf3bd6SGeorge Wilson arc_dead = B_TRUE; 6112fa9e4066Sahrens 611344cb6abcSbmc if (arc_ksp != NULL) { 611444cb6abcSbmc kstat_delete(arc_ksp); 611544cb6abcSbmc arc_ksp = NULL; 611644cb6abcSbmc } 611744cb6abcSbmc 6118244781f1SPrakash Surya mutex_destroy(&arc_reclaim_lock); 6119244781f1SPrakash Surya cv_destroy(&arc_reclaim_thread_cv); 6120244781f1SPrakash Surya cv_destroy(&arc_reclaim_waiters_cv); 6121244781f1SPrakash Surya 6122dcbf3bd6SGeorge Wilson arc_state_fini(); 6123fa9e4066Sahrens buf_fini(); 61242fdbea25SAleksandr Guzovskiy 612589c86e32SChris Williamson ASSERT0(arc_loaned_bytes); 6126fa9e4066Sahrens } 6127fa94a07fSbrendan 6128fa94a07fSbrendan /* 6129fa94a07fSbrendan * Level 2 ARC 6130fa94a07fSbrendan * 6131fa94a07fSbrendan * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 6132fa94a07fSbrendan * It uses dedicated storage devices to hold cached data, which are populated 6133fa94a07fSbrendan * using large infrequent writes. The main role of this cache is to boost 6134fa94a07fSbrendan * the performance of random read workloads. The intended L2ARC devices 6135fa94a07fSbrendan * include short-stroked disks, solid state disks, and other media with 6136fa94a07fSbrendan * substantially faster read latency than disk. 6137fa94a07fSbrendan * 6138fa94a07fSbrendan * +-----------------------+ 6139fa94a07fSbrendan * | ARC | 6140fa94a07fSbrendan * +-----------------------+ 6141fa94a07fSbrendan * | ^ ^ 6142fa94a07fSbrendan * | | | 6143fa94a07fSbrendan * l2arc_feed_thread() arc_read() 6144fa94a07fSbrendan * | | | 6145fa94a07fSbrendan * | l2arc read | 6146fa94a07fSbrendan * V | | 6147fa94a07fSbrendan * +---------------+ | 6148fa94a07fSbrendan * | L2ARC | | 6149fa94a07fSbrendan * +---------------+ | 6150fa94a07fSbrendan * | ^ | 6151fa94a07fSbrendan * l2arc_write() | | 6152fa94a07fSbrendan * | | | 6153fa94a07fSbrendan * V | | 6154fa94a07fSbrendan * +-------+ +-------+ 6155fa94a07fSbrendan * | vdev | | vdev | 6156fa94a07fSbrendan * | cache | | cache | 6157fa94a07fSbrendan * +-------+ +-------+ 6158fa94a07fSbrendan * +=========+ .-----. 6159fa94a07fSbrendan * : L2ARC : |-_____-| 6160fa94a07fSbrendan * : devices : | Disks | 6161fa94a07fSbrendan * +=========+ `-_____-' 6162fa94a07fSbrendan * 6163fa94a07fSbrendan * Read requests are satisfied from the following sources, in order: 6164fa94a07fSbrendan * 6165fa94a07fSbrendan * 1) ARC 6166fa94a07fSbrendan * 2) vdev cache of L2ARC devices 6167fa94a07fSbrendan * 3) L2ARC devices 6168fa94a07fSbrendan * 4) vdev cache of disks 6169fa94a07fSbrendan * 5) disks 6170fa94a07fSbrendan * 6171fa94a07fSbrendan * Some L2ARC device types exhibit extremely slow write performance. 6172fa94a07fSbrendan * To accommodate for this there are some significant differences between 6173fa94a07fSbrendan * the L2ARC and traditional cache design: 6174fa94a07fSbrendan * 6175fa94a07fSbrendan * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 6176fa94a07fSbrendan * the ARC behave as usual, freeing buffers and placing headers on ghost 6177fa94a07fSbrendan * lists. The ARC does not send buffers to the L2ARC during eviction as 6178fa94a07fSbrendan * this would add inflated write latencies for all ARC memory pressure. 6179fa94a07fSbrendan * 6180fa94a07fSbrendan * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 6181fa94a07fSbrendan * It does this by periodically scanning buffers from the eviction-end of 6182fa94a07fSbrendan * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 6183aad02571SSaso Kiselkov * not already there. It scans until a headroom of buffers is satisfied, 6184aad02571SSaso Kiselkov * which itself is a buffer for ARC eviction. If a compressible buffer is 6185aad02571SSaso Kiselkov * found during scanning and selected for writing to an L2ARC device, we 6186aad02571SSaso Kiselkov * temporarily boost scanning headroom during the next scan cycle to make 6187aad02571SSaso Kiselkov * sure we adapt to compression effects (which might significantly reduce 6188aad02571SSaso Kiselkov * the data volume we write to L2ARC). The thread that does this is 6189fa94a07fSbrendan * l2arc_feed_thread(), illustrated below; example sizes are included to 6190fa94a07fSbrendan * provide a better sense of ratio than this diagram: 6191fa94a07fSbrendan * 6192fa94a07fSbrendan * head --> tail 6193fa94a07fSbrendan * +---------------------+----------+ 6194fa94a07fSbrendan * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 6195fa94a07fSbrendan * +---------------------+----------+ | o L2ARC eligible 6196fa94a07fSbrendan * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 6197fa94a07fSbrendan * +---------------------+----------+ | 6198fa94a07fSbrendan * 15.9 Gbytes ^ 32 Mbytes | 6199fa94a07fSbrendan * headroom | 6200fa94a07fSbrendan * l2arc_feed_thread() 6201fa94a07fSbrendan * | 6202fa94a07fSbrendan * l2arc write hand <--[oooo]--' 6203fa94a07fSbrendan * | 8 Mbyte 6204fa94a07fSbrendan * | write max 6205fa94a07fSbrendan * V 6206fa94a07fSbrendan * +==============================+ 6207fa94a07fSbrendan * L2ARC dev |####|#|###|###| |####| ... | 6208fa94a07fSbrendan * +==============================+ 6209fa94a07fSbrendan * 32 Gbytes 6210fa94a07fSbrendan * 6211fa94a07fSbrendan * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 6212fa94a07fSbrendan * evicted, then the L2ARC has cached a buffer much sooner than it probably 6213fa94a07fSbrendan * needed to, potentially wasting L2ARC device bandwidth and storage. It is 6214fa94a07fSbrendan * safe to say that this is an uncommon case, since buffers at the end of 6215fa94a07fSbrendan * the ARC lists have moved there due to inactivity. 6216fa94a07fSbrendan * 6217fa94a07fSbrendan * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 6218fa94a07fSbrendan * then the L2ARC simply misses copying some buffers. This serves as a 6219fa94a07fSbrendan * pressure valve to prevent heavy read workloads from both stalling the ARC 6220fa94a07fSbrendan * with waits and clogging the L2ARC with writes. This also helps prevent 6221fa94a07fSbrendan * the potential for the L2ARC to churn if it attempts to cache content too 6222fa94a07fSbrendan * quickly, such as during backups of the entire pool. 6223fa94a07fSbrendan * 62243a737e0dSbrendan * 5. After system boot and before the ARC has filled main memory, there are 62253a737e0dSbrendan * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 62263a737e0dSbrendan * lists can remain mostly static. Instead of searching from tail of these 62273a737e0dSbrendan * lists as pictured, the l2arc_feed_thread() will search from the list heads 62283a737e0dSbrendan * for eligible buffers, greatly increasing its chance of finding them. 62293a737e0dSbrendan * 62303a737e0dSbrendan * The L2ARC device write speed is also boosted during this time so that 62313a737e0dSbrendan * the L2ARC warms up faster. Since there have been no ARC evictions yet, 62323a737e0dSbrendan * there are no L2ARC reads, and no fear of degrading read performance 62333a737e0dSbrendan * through increased writes. 62343a737e0dSbrendan * 62353a737e0dSbrendan * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 6236fa94a07fSbrendan * the vdev queue can aggregate them into larger and fewer writes. Each 6237fa94a07fSbrendan * device is written to in a rotor fashion, sweeping writes through 6238fa94a07fSbrendan * available space then repeating. 6239fa94a07fSbrendan * 62403a737e0dSbrendan * 7. The L2ARC does not store dirty content. It never needs to flush 6241fa94a07fSbrendan * write buffers back to disk based storage. 6242fa94a07fSbrendan * 62433a737e0dSbrendan * 8. If an ARC buffer is written (and dirtied) which also exists in the 6244fa94a07fSbrendan * L2ARC, the now stale L2ARC buffer is immediately dropped. 6245fa94a07fSbrendan * 6246fa94a07fSbrendan * The performance of the L2ARC can be tweaked by a number of tunables, which 6247fa94a07fSbrendan * may be necessary for different workloads: 6248fa94a07fSbrendan * 6249fa94a07fSbrendan * l2arc_write_max max write bytes per interval 62503a737e0dSbrendan * l2arc_write_boost extra write bytes during device warmup 6251fa94a07fSbrendan * l2arc_noprefetch skip caching prefetched buffers 6252fa94a07fSbrendan * l2arc_headroom number of max device writes to precache 6253aad02571SSaso Kiselkov * l2arc_headroom_boost when we find compressed buffers during ARC 6254aad02571SSaso Kiselkov * scanning, we multiply headroom by this 6255aad02571SSaso Kiselkov * percentage factor for the next scan cycle, 6256aad02571SSaso Kiselkov * since more compressed buffers are likely to 6257aad02571SSaso Kiselkov * be present 6258fa94a07fSbrendan * l2arc_feed_secs seconds between L2ARC writing 6259fa94a07fSbrendan * 6260fa94a07fSbrendan * Tunables may be removed or added as future performance improvements are 6261fa94a07fSbrendan * integrated, and also may become zpool properties. 62625a98e54bSBrendan Gregg - Sun Microsystems * 62635a98e54bSBrendan Gregg - Sun Microsystems * There are three key functions that control how the L2ARC warms up: 62645a98e54bSBrendan Gregg - Sun Microsystems * 62655a98e54bSBrendan Gregg - Sun Microsystems * l2arc_write_eligible() check if a buffer is eligible to cache 62665a98e54bSBrendan Gregg - Sun Microsystems * l2arc_write_size() calculate how much to write 62675a98e54bSBrendan Gregg - Sun Microsystems * l2arc_write_interval() calculate sleep delay between writes 62685a98e54bSBrendan Gregg - Sun Microsystems * 62695a98e54bSBrendan Gregg - Sun Microsystems * These three functions determine what to write, how much, and how quickly 62705a98e54bSBrendan Gregg - Sun Microsystems * to send writes. 6271fa94a07fSbrendan */ 6272fa94a07fSbrendan 62735a98e54bSBrendan Gregg - Sun Microsystems static boolean_t 62747adb730bSGeorge Wilson l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 62755a98e54bSBrendan Gregg - Sun Microsystems { 62765a98e54bSBrendan Gregg - Sun Microsystems /* 62775a98e54bSBrendan Gregg - Sun Microsystems * A buffer is *not* eligible for the L2ARC if it: 62785a98e54bSBrendan Gregg - Sun Microsystems * 1. belongs to a different spa. 62795ea40c06SBrendan Gregg - Sun Microsystems * 2. is already cached on the L2ARC. 62805ea40c06SBrendan Gregg - Sun Microsystems * 3. has an I/O in progress (it may be an incomplete read). 62815ea40c06SBrendan Gregg - Sun Microsystems * 4. is flagged not eligible (zfs property). 62825a98e54bSBrendan Gregg - Sun Microsystems */ 628389c86e32SChris Williamson if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) || 62847adb730bSGeorge Wilson HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr)) 62855a98e54bSBrendan Gregg - Sun Microsystems return (B_FALSE); 62865a98e54bSBrendan Gregg - Sun Microsystems 62875a98e54bSBrendan Gregg - Sun Microsystems return (B_TRUE); 62885a98e54bSBrendan Gregg - Sun Microsystems } 62895a98e54bSBrendan Gregg - Sun Microsystems 62905a98e54bSBrendan Gregg - Sun Microsystems static uint64_t 6291aad02571SSaso Kiselkov l2arc_write_size(void) 62925a98e54bSBrendan Gregg - Sun Microsystems { 62935a98e54bSBrendan Gregg - Sun Microsystems uint64_t size; 62945a98e54bSBrendan Gregg - Sun Microsystems 6295aad02571SSaso Kiselkov /* 6296aad02571SSaso Kiselkov * Make sure our globals have meaningful values in case the user 6297aad02571SSaso Kiselkov * altered them. 6298aad02571SSaso Kiselkov */ 6299aad02571SSaso Kiselkov size = l2arc_write_max; 6300aad02571SSaso Kiselkov if (size == 0) { 6301aad02571SSaso Kiselkov cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 6302aad02571SSaso Kiselkov "be greater than zero, resetting it to the default (%d)", 6303aad02571SSaso Kiselkov L2ARC_WRITE_SIZE); 6304aad02571SSaso Kiselkov size = l2arc_write_max = L2ARC_WRITE_SIZE; 6305aad02571SSaso Kiselkov } 63065a98e54bSBrendan Gregg - Sun Microsystems 63075a98e54bSBrendan Gregg - Sun Microsystems if (arc_warm == B_FALSE) 6308aad02571SSaso Kiselkov size += l2arc_write_boost; 63095a98e54bSBrendan Gregg - Sun Microsystems 63105a98e54bSBrendan Gregg - Sun Microsystems return (size); 63115a98e54bSBrendan Gregg - Sun Microsystems 63125a98e54bSBrendan Gregg - Sun Microsystems } 63135a98e54bSBrendan Gregg - Sun Microsystems 63145a98e54bSBrendan Gregg - Sun Microsystems static clock_t 63155a98e54bSBrendan Gregg - Sun Microsystems l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 63165a98e54bSBrendan Gregg - Sun Microsystems { 6317d3d50737SRafael Vanoni clock_t interval, next, now; 63185a98e54bSBrendan Gregg - Sun Microsystems 63195a98e54bSBrendan Gregg - Sun Microsystems /* 63205a98e54bSBrendan Gregg - Sun Microsystems * If the ARC lists are busy, increase our write rate; if the 63215a98e54bSBrendan Gregg - Sun Microsystems * lists are stale, idle back. This is achieved by checking 63225a98e54bSBrendan Gregg - Sun Microsystems * how much we previously wrote - if it was more than half of 63235a98e54bSBrendan Gregg - Sun Microsystems * what we wanted, schedule the next write much sooner. 63245a98e54bSBrendan Gregg - Sun Microsystems */ 63255a98e54bSBrendan Gregg - Sun Microsystems if (l2arc_feed_again && wrote > (wanted / 2)) 63265a98e54bSBrendan Gregg - Sun Microsystems interval = (hz * l2arc_feed_min_ms) / 1000; 63275a98e54bSBrendan Gregg - Sun Microsystems else 63285a98e54bSBrendan Gregg - Sun Microsystems interval = hz * l2arc_feed_secs; 63295a98e54bSBrendan Gregg - Sun Microsystems 6330d3d50737SRafael Vanoni now = ddi_get_lbolt(); 6331d3d50737SRafael Vanoni next = MAX(now, MIN(now + interval, began + interval)); 63325a98e54bSBrendan Gregg - Sun Microsystems 63335a98e54bSBrendan Gregg - Sun Microsystems return (next); 63345a98e54bSBrendan Gregg - Sun Microsystems } 63355a98e54bSBrendan Gregg - Sun Microsystems 6336fa94a07fSbrendan /* 6337fa94a07fSbrendan * Cycle through L2ARC devices. This is how L2ARC load balances. 63383a737e0dSbrendan * If a device is returned, this also returns holding the spa config lock. 6339fa94a07fSbrendan */ 6340fa94a07fSbrendan static l2arc_dev_t * 6341fa94a07fSbrendan l2arc_dev_get_next(void) 6342fa94a07fSbrendan { 63433a737e0dSbrendan l2arc_dev_t *first, *next = NULL; 63443a737e0dSbrendan 63453a737e0dSbrendan /* 63463a737e0dSbrendan * Lock out the removal of spas (spa_namespace_lock), then removal 63473a737e0dSbrendan * of cache devices (l2arc_dev_mtx). Once a device has been selected, 63483a737e0dSbrendan * both locks will be dropped and a spa config lock held instead. 63493a737e0dSbrendan */ 63503a737e0dSbrendan mutex_enter(&spa_namespace_lock); 63513a737e0dSbrendan mutex_enter(&l2arc_dev_mtx); 6352fa94a07fSbrendan 6353c5904d13Seschrock /* if there are no vdevs, there is nothing to do */ 6354c5904d13Seschrock if (l2arc_ndev == 0) 63553a737e0dSbrendan goto out; 6356c5904d13Seschrock 6357c5904d13Seschrock first = NULL; 6358c5904d13Seschrock next = l2arc_dev_last; 6359c5904d13Seschrock do { 6360c5904d13Seschrock /* loop around the list looking for a non-faulted vdev */ 6361c5904d13Seschrock if (next == NULL) { 6362fa94a07fSbrendan next = list_head(l2arc_dev_list); 6363c5904d13Seschrock } else { 6364c5904d13Seschrock next = list_next(l2arc_dev_list, next); 6365c5904d13Seschrock if (next == NULL) 6366c5904d13Seschrock next = list_head(l2arc_dev_list); 6367c5904d13Seschrock } 6368c5904d13Seschrock 6369c5904d13Seschrock /* if we have come back to the start, bail out */ 6370c5904d13Seschrock if (first == NULL) 6371c5904d13Seschrock first = next; 6372c5904d13Seschrock else if (next == first) 6373c5904d13Seschrock break; 6374c5904d13Seschrock 6375c5904d13Seschrock } while (vdev_is_dead(next->l2ad_vdev)); 6376c5904d13Seschrock 6377c5904d13Seschrock /* if we were unable to find any usable vdevs, return NULL */ 6378c5904d13Seschrock if (vdev_is_dead(next->l2ad_vdev)) 63793a737e0dSbrendan next = NULL; 6380fa94a07fSbrendan 6381fa94a07fSbrendan l2arc_dev_last = next; 6382fa94a07fSbrendan 63833a737e0dSbrendan out: 63843a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 63853a737e0dSbrendan 63863a737e0dSbrendan /* 63873a737e0dSbrendan * Grab the config lock to prevent the 'next' device from being 63883a737e0dSbrendan * removed while we are writing to it. 63893a737e0dSbrendan */ 63903a737e0dSbrendan if (next != NULL) 6391e14bb325SJeff Bonwick spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 63923a737e0dSbrendan mutex_exit(&spa_namespace_lock); 63933a737e0dSbrendan 6394fa94a07fSbrendan return (next); 6395fa94a07fSbrendan } 6396fa94a07fSbrendan 63973a737e0dSbrendan /* 63983a737e0dSbrendan * Free buffers that were tagged for destruction. 63993a737e0dSbrendan */ 64003a737e0dSbrendan static void 64013a737e0dSbrendan l2arc_do_free_on_write() 64023a737e0dSbrendan { 64033a737e0dSbrendan list_t *buflist; 64043a737e0dSbrendan l2arc_data_free_t *df, *df_prev; 64053a737e0dSbrendan 64063a737e0dSbrendan mutex_enter(&l2arc_free_on_write_mtx); 64073a737e0dSbrendan buflist = l2arc_free_on_write; 64083a737e0dSbrendan 64093a737e0dSbrendan for (df = list_tail(buflist); df; df = df_prev) { 64103a737e0dSbrendan df_prev = list_prev(buflist, df); 6411*770499e1SDan Kimmel ASSERT3P(df->l2df_abd, !=, NULL); 6412*770499e1SDan Kimmel abd_free(df->l2df_abd); 64133a737e0dSbrendan list_remove(buflist, df); 64143a737e0dSbrendan kmem_free(df, sizeof (l2arc_data_free_t)); 64153a737e0dSbrendan } 64163a737e0dSbrendan 64173a737e0dSbrendan mutex_exit(&l2arc_free_on_write_mtx); 64183a737e0dSbrendan } 64193a737e0dSbrendan 6420fa94a07fSbrendan /* 6421fa94a07fSbrendan * A write to a cache device has completed. Update all headers to allow 6422fa94a07fSbrendan * reads from these buffers to begin. 6423fa94a07fSbrendan */ 6424fa94a07fSbrendan static void 6425fa94a07fSbrendan l2arc_write_done(zio_t *zio) 6426fa94a07fSbrendan { 6427fa94a07fSbrendan l2arc_write_callback_t *cb; 6428fa94a07fSbrendan l2arc_dev_t *dev; 6429fa94a07fSbrendan list_t *buflist; 64307adb730bSGeorge Wilson arc_buf_hdr_t *head, *hdr, *hdr_prev; 6431fa94a07fSbrendan kmutex_t *hash_lock; 64323038a2b4SSaso Kiselkov int64_t bytes_dropped = 0; 6433fa94a07fSbrendan 6434fa94a07fSbrendan cb = zio->io_private; 6435dcbf3bd6SGeorge Wilson ASSERT3P(cb, !=, NULL); 6436fa94a07fSbrendan dev = cb->l2wcb_dev; 6437dcbf3bd6SGeorge Wilson ASSERT3P(dev, !=, NULL); 6438fa94a07fSbrendan head = cb->l2wcb_head; 6439dcbf3bd6SGeorge Wilson ASSERT3P(head, !=, NULL); 644089c86e32SChris Williamson buflist = &dev->l2ad_buflist; 6441dcbf3bd6SGeorge Wilson ASSERT3P(buflist, !=, NULL); 6442fa94a07fSbrendan DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 6443fa94a07fSbrendan l2arc_write_callback_t *, cb); 6444fa94a07fSbrendan 6445fa94a07fSbrendan if (zio->io_error != 0) 6446fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_writes_error); 6447fa94a07fSbrendan 6448fa94a07fSbrendan /* 6449fa94a07fSbrendan * All writes completed, or an error was hit. 6450fa94a07fSbrendan */ 6451244781f1SPrakash Surya top: 6452244781f1SPrakash Surya mutex_enter(&dev->l2ad_mtx); 64537adb730bSGeorge Wilson for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 64547adb730bSGeorge Wilson hdr_prev = list_prev(buflist, hdr); 6455fa94a07fSbrendan 64567adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 6457244781f1SPrakash Surya 6458244781f1SPrakash Surya /* 6459244781f1SPrakash Surya * We cannot use mutex_enter or else we can deadlock 6460244781f1SPrakash Surya * with l2arc_write_buffers (due to swapping the order 6461244781f1SPrakash Surya * the hash lock and l2ad_mtx are taken). 6462244781f1SPrakash Surya */ 6463fa94a07fSbrendan if (!mutex_tryenter(hash_lock)) { 6464fa94a07fSbrendan /* 6465244781f1SPrakash Surya * Missed the hash lock. We must retry so we 6466244781f1SPrakash Surya * don't leave the ARC_FLAG_L2_WRITING bit set. 6467fa94a07fSbrendan */ 6468244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 6469244781f1SPrakash Surya 6470244781f1SPrakash Surya /* 6471244781f1SPrakash Surya * We don't want to rescan the headers we've 6472244781f1SPrakash Surya * already marked as having been written out, so 6473244781f1SPrakash Surya * we reinsert the head node so we can pick up 6474244781f1SPrakash Surya * where we left off. 6475244781f1SPrakash Surya */ 6476244781f1SPrakash Surya list_remove(buflist, head); 6477244781f1SPrakash Surya list_insert_after(buflist, hdr, head); 6478244781f1SPrakash Surya 6479244781f1SPrakash Surya mutex_exit(&dev->l2ad_mtx); 6480244781f1SPrakash Surya 6481244781f1SPrakash Surya /* 6482244781f1SPrakash Surya * We wait for the hash lock to become available 6483244781f1SPrakash Surya * to try and prevent busy waiting, and increase 6484244781f1SPrakash Surya * the chance we'll be able to acquire the lock 6485244781f1SPrakash Surya * the next time around. 6486244781f1SPrakash Surya */ 6487244781f1SPrakash Surya mutex_enter(hash_lock); 6488244781f1SPrakash Surya mutex_exit(hash_lock); 6489244781f1SPrakash Surya goto top; 6490fa94a07fSbrendan } 6491fa94a07fSbrendan 649289c86e32SChris Williamson /* 6493244781f1SPrakash Surya * We could not have been moved into the arc_l2c_only 6494244781f1SPrakash Surya * state while in-flight due to our ARC_FLAG_L2_WRITING 6495244781f1SPrakash Surya * bit being set. Let's just ensure that's being enforced. 6496244781f1SPrakash Surya */ 6497244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 6498244781f1SPrakash Surya 6499fa94a07fSbrendan if (zio->io_error != 0) { 6500fa94a07fSbrendan /* 65013a737e0dSbrendan * Error - drop L2ARC entry. 6502fa94a07fSbrendan */ 65037adb730bSGeorge Wilson list_remove(buflist, hdr); 6504dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 650589c86e32SChris Williamson 6506dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_l2_asize, -arc_hdr_size(hdr)); 6507dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr)); 6508a52fc310SPrakash Surya 6509dcbf3bd6SGeorge Wilson bytes_dropped += arc_hdr_size(hdr); 6510a52fc310SPrakash Surya (void) refcount_remove_many(&dev->l2ad_alloc, 6511dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 6512fa94a07fSbrendan } 6513fa94a07fSbrendan 6514fa94a07fSbrendan /* 6515244781f1SPrakash Surya * Allow ARC to begin reads and ghost list evictions to 6516244781f1SPrakash Surya * this L2ARC entry. 6517fa94a07fSbrendan */ 6518dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); 6519fa94a07fSbrendan 6520fa94a07fSbrendan mutex_exit(hash_lock); 6521fa94a07fSbrendan } 6522fa94a07fSbrendan 6523fa94a07fSbrendan atomic_inc_64(&l2arc_writes_done); 6524fa94a07fSbrendan list_remove(buflist, head); 652589c86e32SChris Williamson ASSERT(!HDR_HAS_L1HDR(head)); 652689c86e32SChris Williamson kmem_cache_free(hdr_l2only_cache, head); 652789c86e32SChris Williamson mutex_exit(&dev->l2ad_mtx); 6528fa94a07fSbrendan 65293038a2b4SSaso Kiselkov vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 65303038a2b4SSaso Kiselkov 65313a737e0dSbrendan l2arc_do_free_on_write(); 6532fa94a07fSbrendan 6533fa94a07fSbrendan kmem_free(cb, sizeof (l2arc_write_callback_t)); 6534fa94a07fSbrendan } 6535fa94a07fSbrendan 6536fa94a07fSbrendan /* 6537fa94a07fSbrendan * A read to a cache device completed. Validate buffer contents before 6538fa94a07fSbrendan * handing over to the regular ARC routines. 6539fa94a07fSbrendan */ 6540fa94a07fSbrendan static void 6541fa94a07fSbrendan l2arc_read_done(zio_t *zio) 6542fa94a07fSbrendan { 6543fa94a07fSbrendan l2arc_read_callback_t *cb; 6544fa94a07fSbrendan arc_buf_hdr_t *hdr; 6545fa94a07fSbrendan kmutex_t *hash_lock; 6546dcbf3bd6SGeorge Wilson boolean_t valid_cksum; 6547fa94a07fSbrendan 6548dcbf3bd6SGeorge Wilson ASSERT3P(zio->io_vd, !=, NULL); 6549e14bb325SJeff Bonwick ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 6550e14bb325SJeff Bonwick 6551e14bb325SJeff Bonwick spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 6552e14bb325SJeff Bonwick 6553fa94a07fSbrendan cb = zio->io_private; 6554dcbf3bd6SGeorge Wilson ASSERT3P(cb, !=, NULL); 6555dcbf3bd6SGeorge Wilson hdr = cb->l2rcb_hdr; 6556dcbf3bd6SGeorge Wilson ASSERT3P(hdr, !=, NULL); 6557fa94a07fSbrendan 6558dcbf3bd6SGeorge Wilson hash_lock = HDR_LOCK(hdr); 6559fa94a07fSbrendan mutex_enter(hash_lock); 65603f9d6ad7SLin Ling ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 6561fa94a07fSbrendan 6562*770499e1SDan Kimmel ASSERT3P(zio->io_abd, !=, NULL); 6563aad02571SSaso Kiselkov 6564fa94a07fSbrendan /* 6565fa94a07fSbrendan * Check this survived the L2ARC journey. 6566fa94a07fSbrendan */ 6567*770499e1SDan Kimmel ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd); 6568dcbf3bd6SGeorge Wilson zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 6569dcbf3bd6SGeorge Wilson zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 6570dcbf3bd6SGeorge Wilson 6571dcbf3bd6SGeorge Wilson valid_cksum = arc_cksum_is_equal(hdr, zio); 6572dcbf3bd6SGeorge Wilson if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 6573fa94a07fSbrendan mutex_exit(hash_lock); 6574dcbf3bd6SGeorge Wilson zio->io_private = hdr; 6575fa94a07fSbrendan arc_read_done(zio); 6576fa94a07fSbrendan } else { 6577fa94a07fSbrendan mutex_exit(hash_lock); 6578fa94a07fSbrendan /* 6579fa94a07fSbrendan * Buffer didn't survive caching. Increment stats and 6580fa94a07fSbrendan * reissue to the original storage device. 6581fa94a07fSbrendan */ 65823a737e0dSbrendan if (zio->io_error != 0) { 6583fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_io_error); 65843a737e0dSbrendan } else { 6585be6fd75aSMatthew Ahrens zio->io_error = SET_ERROR(EIO); 65863a737e0dSbrendan } 6587dcbf3bd6SGeorge Wilson if (!valid_cksum) 6588fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_cksum_bad); 6589fa94a07fSbrendan 6590e14bb325SJeff Bonwick /* 6591e14bb325SJeff Bonwick * If there's no waiter, issue an async i/o to the primary 6592e14bb325SJeff Bonwick * storage now. If there *is* a waiter, the caller must 6593e14bb325SJeff Bonwick * issue the i/o in a context where it's OK to block. 6594e14bb325SJeff Bonwick */ 6595a3f829aeSBill Moore if (zio->io_waiter == NULL) { 6596a3f829aeSBill Moore zio_t *pio = zio_unique_parent(zio); 6597a3f829aeSBill Moore 6598a3f829aeSBill Moore ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6599a3f829aeSBill Moore 6600dcbf3bd6SGeorge Wilson zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, 6601*770499e1SDan Kimmel hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done, 6602dcbf3bd6SGeorge Wilson hdr, zio->io_priority, cb->l2rcb_flags, 6603dcbf3bd6SGeorge Wilson &cb->l2rcb_zb)); 6604a3f829aeSBill Moore } 6605fa94a07fSbrendan } 6606fa94a07fSbrendan 6607fa94a07fSbrendan kmem_free(cb, sizeof (l2arc_read_callback_t)); 6608fa94a07fSbrendan } 6609fa94a07fSbrendan 6610fa94a07fSbrendan /* 6611fa94a07fSbrendan * This is the list priority from which the L2ARC will search for pages to 6612fa94a07fSbrendan * cache. This is used within loops (0..3) to cycle through lists in the 6613fa94a07fSbrendan * desired order. This order can have a significant effect on cache 6614fa94a07fSbrendan * performance. 6615fa94a07fSbrendan * 6616fa94a07fSbrendan * Currently the metadata lists are hit first, MFU then MRU, followed by 6617fa94a07fSbrendan * the data lists. This function returns a locked list, and also returns 6618fa94a07fSbrendan * the lock pointer. 6619fa94a07fSbrendan */ 6620244781f1SPrakash Surya static multilist_sublist_t * 6621244781f1SPrakash Surya l2arc_sublist_lock(int list_num) 6622fa94a07fSbrendan { 6623244781f1SPrakash Surya multilist_t *ml = NULL; 6624244781f1SPrakash Surya unsigned int idx; 6625fa94a07fSbrendan 6626fa94a07fSbrendan ASSERT(list_num >= 0 && list_num <= 3); 6627fa94a07fSbrendan 6628fa94a07fSbrendan switch (list_num) { 6629fa94a07fSbrendan case 0: 663094c2d0ebSMatthew Ahrens ml = arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6631fa94a07fSbrendan break; 6632fa94a07fSbrendan case 1: 663394c2d0ebSMatthew Ahrens ml = arc_mru->arcs_list[ARC_BUFC_METADATA]; 6634fa94a07fSbrendan break; 6635fa94a07fSbrendan case 2: 663694c2d0ebSMatthew Ahrens ml = arc_mfu->arcs_list[ARC_BUFC_DATA]; 6637fa94a07fSbrendan break; 6638fa94a07fSbrendan case 3: 663994c2d0ebSMatthew Ahrens ml = arc_mru->arcs_list[ARC_BUFC_DATA]; 6640fa94a07fSbrendan break; 6641fa94a07fSbrendan } 6642fa94a07fSbrendan 6643244781f1SPrakash Surya /* 6644244781f1SPrakash Surya * Return a randomly-selected sublist. This is acceptable 6645244781f1SPrakash Surya * because the caller feeds only a little bit of data for each 6646244781f1SPrakash Surya * call (8MB). Subsequent calls will result in different 6647244781f1SPrakash Surya * sublists being selected. 6648244781f1SPrakash Surya */ 6649244781f1SPrakash Surya idx = multilist_get_random_index(ml); 6650244781f1SPrakash Surya return (multilist_sublist_lock(ml, idx)); 6651fa94a07fSbrendan } 6652fa94a07fSbrendan 6653fa94a07fSbrendan /* 6654fa94a07fSbrendan * Evict buffers from the device write hand to the distance specified in 6655fa94a07fSbrendan * bytes. This distance may span populated buffers, it may span nothing. 6656fa94a07fSbrendan * This is clearing a region on the L2ARC device ready for writing. 6657fa94a07fSbrendan * If the 'all' boolean is set, every buffer is evicted. 6658fa94a07fSbrendan */ 6659fa94a07fSbrendan static void 6660fa94a07fSbrendan l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6661fa94a07fSbrendan { 6662fa94a07fSbrendan list_t *buflist; 66637adb730bSGeorge Wilson arc_buf_hdr_t *hdr, *hdr_prev; 6664fa94a07fSbrendan kmutex_t *hash_lock; 6665fa94a07fSbrendan uint64_t taddr; 6666fa94a07fSbrendan 666789c86e32SChris Williamson buflist = &dev->l2ad_buflist; 6668fa94a07fSbrendan 6669fa94a07fSbrendan if (!all && dev->l2ad_first) { 6670fa94a07fSbrendan /* 6671fa94a07fSbrendan * This is the first sweep through the device. There is 6672fa94a07fSbrendan * nothing to evict. 6673fa94a07fSbrendan */ 6674fa94a07fSbrendan return; 6675fa94a07fSbrendan } 6676fa94a07fSbrendan 66773a737e0dSbrendan if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6678fa94a07fSbrendan /* 6679fa94a07fSbrendan * When nearing the end of the device, evict to the end 6680fa94a07fSbrendan * before the device write hand jumps to the start. 6681fa94a07fSbrendan */ 6682fa94a07fSbrendan taddr = dev->l2ad_end; 6683fa94a07fSbrendan } else { 6684fa94a07fSbrendan taddr = dev->l2ad_hand + distance; 6685fa94a07fSbrendan } 6686fa94a07fSbrendan DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6687fa94a07fSbrendan uint64_t, taddr, boolean_t, all); 6688fa94a07fSbrendan 6689fa94a07fSbrendan top: 669089c86e32SChris Williamson mutex_enter(&dev->l2ad_mtx); 66917adb730bSGeorge Wilson for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 66927adb730bSGeorge Wilson hdr_prev = list_prev(buflist, hdr); 6693fa94a07fSbrendan 66947adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 6695244781f1SPrakash Surya 6696244781f1SPrakash Surya /* 6697244781f1SPrakash Surya * We cannot use mutex_enter or else we can deadlock 6698244781f1SPrakash Surya * with l2arc_write_buffers (due to swapping the order 6699244781f1SPrakash Surya * the hash lock and l2ad_mtx are taken). 6700244781f1SPrakash Surya */ 6701fa94a07fSbrendan if (!mutex_tryenter(hash_lock)) { 6702fa94a07fSbrendan /* 6703fa94a07fSbrendan * Missed the hash lock. Retry. 6704fa94a07fSbrendan */ 6705fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 670689c86e32SChris Williamson mutex_exit(&dev->l2ad_mtx); 6707fa94a07fSbrendan mutex_enter(hash_lock); 6708fa94a07fSbrendan mutex_exit(hash_lock); 6709fa94a07fSbrendan goto top; 6710fa94a07fSbrendan } 6711fa94a07fSbrendan 67127adb730bSGeorge Wilson if (HDR_L2_WRITE_HEAD(hdr)) { 6713fa94a07fSbrendan /* 6714fa94a07fSbrendan * We hit a write head node. Leave it for 6715fa94a07fSbrendan * l2arc_write_done(). 6716fa94a07fSbrendan */ 67177adb730bSGeorge Wilson list_remove(buflist, hdr); 6718fa94a07fSbrendan mutex_exit(hash_lock); 6719fa94a07fSbrendan continue; 6720fa94a07fSbrendan } 6721fa94a07fSbrendan 672289c86e32SChris Williamson if (!all && HDR_HAS_L2HDR(hdr) && 672389c86e32SChris Williamson (hdr->b_l2hdr.b_daddr > taddr || 672489c86e32SChris Williamson hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6725fa94a07fSbrendan /* 6726fa94a07fSbrendan * We've evicted to the target address, 6727fa94a07fSbrendan * or the end of the device. 6728fa94a07fSbrendan */ 6729fa94a07fSbrendan mutex_exit(hash_lock); 6730fa94a07fSbrendan break; 6731fa94a07fSbrendan } 6732fa94a07fSbrendan 673389c86e32SChris Williamson ASSERT(HDR_HAS_L2HDR(hdr)); 673489c86e32SChris Williamson if (!HDR_HAS_L1HDR(hdr)) { 67357adb730bSGeorge Wilson ASSERT(!HDR_L2_READING(hdr)); 6736fa94a07fSbrendan /* 6737fa94a07fSbrendan * This doesn't exist in the ARC. Destroy. 6738fa94a07fSbrendan * arc_hdr_destroy() will call list_remove() 6739fa94a07fSbrendan * and decrement arcstat_l2_size. 6740fa94a07fSbrendan */ 67417adb730bSGeorge Wilson arc_change_state(arc_anon, hdr, hash_lock); 67427adb730bSGeorge Wilson arc_hdr_destroy(hdr); 6743fa94a07fSbrendan } else { 674489c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 674589c86e32SChris Williamson ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 67463a737e0dSbrendan /* 67473a737e0dSbrendan * Invalidate issued or about to be issued 67483a737e0dSbrendan * reads, since we may be about to write 67493a737e0dSbrendan * over this location. 67503a737e0dSbrendan */ 67517adb730bSGeorge Wilson if (HDR_L2_READING(hdr)) { 67523a737e0dSbrendan ARCSTAT_BUMP(arcstat_l2_evict_reading); 6753dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); 67543a737e0dSbrendan } 67553a737e0dSbrendan 6756244781f1SPrakash Surya /* Ensure this header has finished being written */ 6757244781f1SPrakash Surya ASSERT(!HDR_L2_WRITING(hdr)); 6758a52fc310SPrakash Surya 6759a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(hdr); 6760fa94a07fSbrendan } 6761fa94a07fSbrendan mutex_exit(hash_lock); 6762fa94a07fSbrendan } 676389c86e32SChris Williamson mutex_exit(&dev->l2ad_mtx); 6764fa94a07fSbrendan } 6765fa94a07fSbrendan 6766fa94a07fSbrendan /* 6767fa94a07fSbrendan * Find and write ARC buffers to the L2ARC device. 6768fa94a07fSbrendan * 67697adb730bSGeorge Wilson * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6770fa94a07fSbrendan * for reading until they have completed writing. 6771aad02571SSaso Kiselkov * The headroom_boost is an in-out parameter used to maintain headroom boost 6772aad02571SSaso Kiselkov * state between calls to this function. 6773aad02571SSaso Kiselkov * 6774aad02571SSaso Kiselkov * Returns the number of bytes actually written (which may be smaller than 6775aad02571SSaso Kiselkov * the delta by which the device hand has changed due to alignment). 6776fa94a07fSbrendan */ 67775a98e54bSBrendan Gregg - Sun Microsystems static uint64_t 6778dcbf3bd6SGeorge Wilson l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 6779fa94a07fSbrendan { 67807adb730bSGeorge Wilson arc_buf_hdr_t *hdr, *hdr_prev, *head; 6781dcbf3bd6SGeorge Wilson uint64_t write_asize, write_psize, write_sz, headroom; 6782aad02571SSaso Kiselkov boolean_t full; 6783fa94a07fSbrendan l2arc_write_callback_t *cb; 6784fa94a07fSbrendan zio_t *pio, *wzio; 6785e9103aaeSGarrett D'Amore uint64_t guid = spa_load_guid(spa); 6786fa94a07fSbrendan 6787dcbf3bd6SGeorge Wilson ASSERT3P(dev->l2ad_vdev, !=, NULL); 6788aad02571SSaso Kiselkov 6789fa94a07fSbrendan pio = NULL; 6790dcbf3bd6SGeorge Wilson write_sz = write_asize = write_psize = 0; 6791fa94a07fSbrendan full = B_FALSE; 679289c86e32SChris Williamson head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 6793dcbf3bd6SGeorge Wilson arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); 6794aad02571SSaso Kiselkov 6795fa94a07fSbrendan /* 6796fa94a07fSbrendan * Copy buffers for L2ARC writing. 6797fa94a07fSbrendan */ 6798fa94a07fSbrendan for (int try = 0; try <= 3; try++) { 6799244781f1SPrakash Surya multilist_sublist_t *mls = l2arc_sublist_lock(try); 6800aad02571SSaso Kiselkov uint64_t passed_sz = 0; 6801aad02571SSaso Kiselkov 68023a737e0dSbrendan /* 68033a737e0dSbrendan * L2ARC fast warmup. 68043a737e0dSbrendan * 68053a737e0dSbrendan * Until the ARC is warm and starts to evict, read from the 68063a737e0dSbrendan * head of the ARC lists rather than the tail. 68073a737e0dSbrendan */ 68083a737e0dSbrendan if (arc_warm == B_FALSE) 6809244781f1SPrakash Surya hdr = multilist_sublist_head(mls); 68103a737e0dSbrendan else 6811244781f1SPrakash Surya hdr = multilist_sublist_tail(mls); 68123a737e0dSbrendan 6813aad02571SSaso Kiselkov headroom = target_sz * l2arc_headroom; 6814dcbf3bd6SGeorge Wilson if (zfs_compressed_arc_enabled) 6815aad02571SSaso Kiselkov headroom = (headroom * l2arc_headroom_boost) / 100; 6816aad02571SSaso Kiselkov 68177adb730bSGeorge Wilson for (; hdr; hdr = hdr_prev) { 6818aad02571SSaso Kiselkov kmutex_t *hash_lock; 6819aad02571SSaso Kiselkov 68203a737e0dSbrendan if (arc_warm == B_FALSE) 6821244781f1SPrakash Surya hdr_prev = multilist_sublist_next(mls, hdr); 68223a737e0dSbrendan else 6823244781f1SPrakash Surya hdr_prev = multilist_sublist_prev(mls, hdr); 6824fa94a07fSbrendan 68257adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 6826aad02571SSaso Kiselkov if (!mutex_tryenter(hash_lock)) { 6827fa94a07fSbrendan /* 6828fa94a07fSbrendan * Skip this buffer rather than waiting. 6829fa94a07fSbrendan */ 6830fa94a07fSbrendan continue; 6831fa94a07fSbrendan } 6832fa94a07fSbrendan 6833dcbf3bd6SGeorge Wilson passed_sz += HDR_GET_LSIZE(hdr); 6834fa94a07fSbrendan if (passed_sz > headroom) { 6835fa94a07fSbrendan /* 6836fa94a07fSbrendan * Searched too far. 6837fa94a07fSbrendan */ 6838fa94a07fSbrendan mutex_exit(hash_lock); 6839fa94a07fSbrendan break; 6840fa94a07fSbrendan } 6841fa94a07fSbrendan 68427adb730bSGeorge Wilson if (!l2arc_write_eligible(guid, hdr)) { 6843fa94a07fSbrendan mutex_exit(hash_lock); 6844fa94a07fSbrendan continue; 6845fa94a07fSbrendan } 6846fa94a07fSbrendan 6847dcbf3bd6SGeorge Wilson if ((write_asize + HDR_GET_LSIZE(hdr)) > target_sz) { 6848fa94a07fSbrendan full = B_TRUE; 6849fa94a07fSbrendan mutex_exit(hash_lock); 6850fa94a07fSbrendan break; 6851fa94a07fSbrendan } 6852fa94a07fSbrendan 6853fa94a07fSbrendan if (pio == NULL) { 6854fa94a07fSbrendan /* 6855fa94a07fSbrendan * Insert a dummy header on the buflist so 6856fa94a07fSbrendan * l2arc_write_done() can find where the 6857fa94a07fSbrendan * write buffers begin without searching. 6858fa94a07fSbrendan */ 6859244781f1SPrakash Surya mutex_enter(&dev->l2ad_mtx); 686089c86e32SChris Williamson list_insert_head(&dev->l2ad_buflist, head); 6861244781f1SPrakash Surya mutex_exit(&dev->l2ad_mtx); 6862fa94a07fSbrendan 6863fa94a07fSbrendan cb = kmem_alloc( 6864fa94a07fSbrendan sizeof (l2arc_write_callback_t), KM_SLEEP); 6865fa94a07fSbrendan cb->l2wcb_dev = dev; 6866fa94a07fSbrendan cb->l2wcb_head = head; 6867fa94a07fSbrendan pio = zio_root(spa, l2arc_write_done, cb, 6868fa94a07fSbrendan ZIO_FLAG_CANFAIL); 6869fa94a07fSbrendan } 6870fa94a07fSbrendan 687189c86e32SChris Williamson hdr->b_l2hdr.b_dev = dev; 6872dcbf3bd6SGeorge Wilson hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 6873dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, 6874dcbf3bd6SGeorge Wilson ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR); 6875dcbf3bd6SGeorge Wilson 6876dcbf3bd6SGeorge Wilson mutex_enter(&dev->l2ad_mtx); 6877dcbf3bd6SGeorge Wilson list_insert_head(&dev->l2ad_buflist, hdr); 6878dcbf3bd6SGeorge Wilson mutex_exit(&dev->l2ad_mtx); 6879dcbf3bd6SGeorge Wilson 6880aad02571SSaso Kiselkov /* 6881dcbf3bd6SGeorge Wilson * We rely on the L1 portion of the header below, so 6882dcbf3bd6SGeorge Wilson * it's invalid for this header to have been evicted out 6883dcbf3bd6SGeorge Wilson * of the ghost cache, prior to being written out. The 6884dcbf3bd6SGeorge Wilson * ARC_FLAG_L2_WRITING bit ensures this won't happen. 6885aad02571SSaso Kiselkov */ 6886dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 6887dcbf3bd6SGeorge Wilson 6888dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); 6889*770499e1SDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 6890dcbf3bd6SGeorge Wilson ASSERT3U(arc_hdr_size(hdr), >, 0); 6891dcbf3bd6SGeorge Wilson uint64_t size = arc_hdr_size(hdr); 6892dcbf3bd6SGeorge Wilson 6893dcbf3bd6SGeorge Wilson (void) refcount_add_many(&dev->l2ad_alloc, size, hdr); 6894aad02571SSaso Kiselkov 6895a52fc310SPrakash Surya /* 6896dcbf3bd6SGeorge Wilson * Normally the L2ARC can use the hdr's data, but if 6897dcbf3bd6SGeorge Wilson * we're sharing data between the hdr and one of its 6898dcbf3bd6SGeorge Wilson * bufs, L2ARC needs its own copy of the data so that 6899dcbf3bd6SGeorge Wilson * the ZIO below can't race with the buf consumer. To 6900dcbf3bd6SGeorge Wilson * ensure that this copy will be available for the 6901dcbf3bd6SGeorge Wilson * lifetime of the ZIO and be cleaned up afterwards, we 6902dcbf3bd6SGeorge Wilson * add it to the l2arc_free_on_write queue. 6903a52fc310SPrakash Surya */ 6904*770499e1SDan Kimmel abd_t *to_write; 6905dcbf3bd6SGeorge Wilson if (!HDR_SHARED_DATA(hdr)) { 6906*770499e1SDan Kimmel to_write = hdr->b_l1hdr.b_pabd; 6907dcbf3bd6SGeorge Wilson } else { 6908*770499e1SDan Kimmel to_write = abd_alloc_for_io(size, 6909*770499e1SDan Kimmel HDR_ISTYPE_METADATA(hdr)); 6910*770499e1SDan Kimmel abd_copy(to_write, hdr->b_l1hdr.b_pabd, size); 6911*770499e1SDan Kimmel l2arc_free_abd_on_write(to_write, size, 6912*770499e1SDan Kimmel arc_buf_type(hdr)); 6913dcbf3bd6SGeorge Wilson } 6914dcbf3bd6SGeorge Wilson wzio = zio_write_phys(pio, dev->l2ad_vdev, 6915dcbf3bd6SGeorge Wilson hdr->b_l2hdr.b_daddr, size, to_write, 6916dcbf3bd6SGeorge Wilson ZIO_CHECKSUM_OFF, NULL, hdr, 6917dcbf3bd6SGeorge Wilson ZIO_PRIORITY_ASYNC_WRITE, 6918dcbf3bd6SGeorge Wilson ZIO_FLAG_CANFAIL, B_FALSE); 6919aad02571SSaso Kiselkov 6920dcbf3bd6SGeorge Wilson write_sz += HDR_GET_LSIZE(hdr); 6921dcbf3bd6SGeorge Wilson DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 6922dcbf3bd6SGeorge Wilson zio_t *, wzio); 6923fa94a07fSbrendan 6924dcbf3bd6SGeorge Wilson write_asize += size; 6925fa94a07fSbrendan /* 6926dcbf3bd6SGeorge Wilson * Keep the clock hand suitably device-aligned. 6927fa94a07fSbrendan */ 6928dcbf3bd6SGeorge Wilson uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, 6929dcbf3bd6SGeorge Wilson size); 6930dcbf3bd6SGeorge Wilson write_psize += asize; 6931dcbf3bd6SGeorge Wilson dev->l2ad_hand += asize; 6932fa94a07fSbrendan 6933fa94a07fSbrendan mutex_exit(hash_lock); 6934fa94a07fSbrendan 6935dcbf3bd6SGeorge Wilson (void) zio_nowait(wzio); 6936aad02571SSaso Kiselkov } 6937aad02571SSaso Kiselkov 6938244781f1SPrakash Surya multilist_sublist_unlock(mls); 6939aad02571SSaso Kiselkov 6940aad02571SSaso Kiselkov if (full == B_TRUE) 6941aad02571SSaso Kiselkov break; 6942aad02571SSaso Kiselkov } 6943aad02571SSaso Kiselkov 6944aad02571SSaso Kiselkov /* No buffers selected for writing? */ 6945aad02571SSaso Kiselkov if (pio == NULL) { 6946aad02571SSaso Kiselkov ASSERT0(write_sz); 694789c86e32SChris Williamson ASSERT(!HDR_HAS_L1HDR(head)); 694889c86e32SChris Williamson kmem_cache_free(hdr_l2only_cache, head); 6949aad02571SSaso Kiselkov return (0); 6950aad02571SSaso Kiselkov } 6951aad02571SSaso Kiselkov 6952aad02571SSaso Kiselkov ASSERT3U(write_asize, <=, target_sz); 6953fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_writes_sent); 6954aad02571SSaso Kiselkov ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 6955fa94a07fSbrendan ARCSTAT_INCR(arcstat_l2_size, write_sz); 6956dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_l2_asize, write_asize); 6957dcbf3bd6SGeorge Wilson vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 6958fa94a07fSbrendan 6959fa94a07fSbrendan /* 6960fa94a07fSbrendan * Bump device hand to the device start if it is approaching the end. 6961fa94a07fSbrendan * l2arc_evict() will already have evicted ahead for this case. 6962fa94a07fSbrendan */ 69633a737e0dSbrendan if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 6964fa94a07fSbrendan dev->l2ad_hand = dev->l2ad_start; 6965fa94a07fSbrendan dev->l2ad_first = B_FALSE; 6966fa94a07fSbrendan } 6967fa94a07fSbrendan 69685a98e54bSBrendan Gregg - Sun Microsystems dev->l2ad_writing = B_TRUE; 6969fa94a07fSbrendan (void) zio_wait(pio); 69705a98e54bSBrendan Gregg - Sun Microsystems dev->l2ad_writing = B_FALSE; 69715a98e54bSBrendan Gregg - Sun Microsystems 6972aad02571SSaso Kiselkov return (write_asize); 6973aad02571SSaso Kiselkov } 6974aad02571SSaso Kiselkov 6975fa94a07fSbrendan /* 6976fa94a07fSbrendan * This thread feeds the L2ARC at regular intervals. This is the beating 6977fa94a07fSbrendan * heart of the L2ARC. 6978fa94a07fSbrendan */ 6979fa94a07fSbrendan static void 6980fa94a07fSbrendan l2arc_feed_thread(void) 6981fa94a07fSbrendan { 6982fa94a07fSbrendan callb_cpr_t cpr; 6983fa94a07fSbrendan l2arc_dev_t *dev; 6984fa94a07fSbrendan spa_t *spa; 69855a98e54bSBrendan Gregg - Sun Microsystems uint64_t size, wrote; 6986d3d50737SRafael Vanoni clock_t begin, next = ddi_get_lbolt(); 6987fa94a07fSbrendan 6988fa94a07fSbrendan CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 6989fa94a07fSbrendan 6990fa94a07fSbrendan mutex_enter(&l2arc_feed_thr_lock); 6991fa94a07fSbrendan 6992fa94a07fSbrendan while (l2arc_thread_exit == 0) { 6993fa94a07fSbrendan CALLB_CPR_SAFE_BEGIN(&cpr); 6994fa94a07fSbrendan (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 69955a98e54bSBrendan Gregg - Sun Microsystems next); 6996fa94a07fSbrendan CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 6997d3d50737SRafael Vanoni next = ddi_get_lbolt() + hz; 6998fa94a07fSbrendan 69993a737e0dSbrendan /* 70003a737e0dSbrendan * Quick check for L2ARC devices. 70013a737e0dSbrendan */ 7002c5904d13Seschrock mutex_enter(&l2arc_dev_mtx); 70033a737e0dSbrendan if (l2arc_ndev == 0) { 70043a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 70053a737e0dSbrendan continue; 70063a737e0dSbrendan } 70073a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 7008d3d50737SRafael Vanoni begin = ddi_get_lbolt(); 7009c5904d13Seschrock 7010fa94a07fSbrendan /* 7011c5904d13Seschrock * This selects the next l2arc device to write to, and in 7012c5904d13Seschrock * doing so the next spa to feed from: dev->l2ad_spa. This 70133a737e0dSbrendan * will return NULL if there are now no l2arc devices or if 70143a737e0dSbrendan * they are all faulted. 70153a737e0dSbrendan * 70163a737e0dSbrendan * If a device is returned, its spa's config lock is also 70173a737e0dSbrendan * held to prevent device removal. l2arc_dev_get_next() 70183a737e0dSbrendan * will grab and release l2arc_dev_mtx. 7019fa94a07fSbrendan */ 70203a737e0dSbrendan if ((dev = l2arc_dev_get_next()) == NULL) 7021fa94a07fSbrendan continue; 70223a737e0dSbrendan 70233a737e0dSbrendan spa = dev->l2ad_spa; 7024dcbf3bd6SGeorge Wilson ASSERT3P(spa, !=, NULL); 7025fa94a07fSbrendan 7026f9af39baSGeorge Wilson /* 7027f9af39baSGeorge Wilson * If the pool is read-only then force the feed thread to 7028f9af39baSGeorge Wilson * sleep a little longer. 7029f9af39baSGeorge Wilson */ 7030f9af39baSGeorge Wilson if (!spa_writeable(spa)) { 7031f9af39baSGeorge Wilson next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 7032f9af39baSGeorge Wilson spa_config_exit(spa, SCL_L2ARC, dev); 7033f9af39baSGeorge Wilson continue; 7034f9af39baSGeorge Wilson } 7035f9af39baSGeorge Wilson 7036fa94a07fSbrendan /* 7037fa94a07fSbrendan * Avoid contributing to memory pressure. 7038fa94a07fSbrendan */ 7039fa94a07fSbrendan if (arc_reclaim_needed()) { 7040fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 7041e14bb325SJeff Bonwick spa_config_exit(spa, SCL_L2ARC, dev); 7042fa94a07fSbrendan continue; 7043fa94a07fSbrendan } 7044fa94a07fSbrendan 7045fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_feeds); 7046fa94a07fSbrendan 7047aad02571SSaso Kiselkov size = l2arc_write_size(); 70483a737e0dSbrendan 7049fa94a07fSbrendan /* 7050fa94a07fSbrendan * Evict L2ARC buffers that will be overwritten. 7051fa94a07fSbrendan */ 70523a737e0dSbrendan l2arc_evict(dev, size, B_FALSE); 7053fa94a07fSbrendan 7054fa94a07fSbrendan /* 7055fa94a07fSbrendan * Write ARC buffers. 7056fa94a07fSbrendan */ 7057dcbf3bd6SGeorge Wilson wrote = l2arc_write_buffers(spa, dev, size); 70585a98e54bSBrendan Gregg - Sun Microsystems 70595a98e54bSBrendan Gregg - Sun Microsystems /* 70605a98e54bSBrendan Gregg - Sun Microsystems * Calculate interval between writes. 70615a98e54bSBrendan Gregg - Sun Microsystems */ 70625a98e54bSBrendan Gregg - Sun Microsystems next = l2arc_write_interval(begin, size, wrote); 7063e14bb325SJeff Bonwick spa_config_exit(spa, SCL_L2ARC, dev); 7064fa94a07fSbrendan } 7065fa94a07fSbrendan 7066fa94a07fSbrendan l2arc_thread_exit = 0; 7067fa94a07fSbrendan cv_broadcast(&l2arc_feed_thr_cv); 7068fa94a07fSbrendan CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 7069fa94a07fSbrendan thread_exit(); 7070fa94a07fSbrendan } 7071fa94a07fSbrendan 7072c5904d13Seschrock boolean_t 7073c5904d13Seschrock l2arc_vdev_present(vdev_t *vd) 7074c5904d13Seschrock { 7075c5904d13Seschrock l2arc_dev_t *dev; 7076c5904d13Seschrock 7077c5904d13Seschrock mutex_enter(&l2arc_dev_mtx); 7078c5904d13Seschrock for (dev = list_head(l2arc_dev_list); dev != NULL; 7079c5904d13Seschrock dev = list_next(l2arc_dev_list, dev)) { 7080c5904d13Seschrock if (dev->l2ad_vdev == vd) 7081c5904d13Seschrock break; 7082c5904d13Seschrock } 7083c5904d13Seschrock mutex_exit(&l2arc_dev_mtx); 7084c5904d13Seschrock 7085c5904d13Seschrock return (dev != NULL); 7086c5904d13Seschrock } 7087c5904d13Seschrock 7088fa94a07fSbrendan /* 7089fa94a07fSbrendan * Add a vdev for use by the L2ARC. By this point the spa has already 7090fa94a07fSbrendan * validated the vdev and opened it. 7091fa94a07fSbrendan */ 7092fa94a07fSbrendan void 7093573ca77eSGeorge Wilson l2arc_add_vdev(spa_t *spa, vdev_t *vd) 7094fa94a07fSbrendan { 7095fa94a07fSbrendan l2arc_dev_t *adddev; 7096fa94a07fSbrendan 7097c5904d13Seschrock ASSERT(!l2arc_vdev_present(vd)); 7098c5904d13Seschrock 7099fa94a07fSbrendan /* 7100fa94a07fSbrendan * Create a new l2arc device entry. 7101fa94a07fSbrendan */ 7102fa94a07fSbrendan adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 7103fa94a07fSbrendan adddev->l2ad_spa = spa; 7104fa94a07fSbrendan adddev->l2ad_vdev = vd; 7105573ca77eSGeorge Wilson adddev->l2ad_start = VDEV_LABEL_START_SIZE; 7106573ca77eSGeorge Wilson adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 7107fa94a07fSbrendan adddev->l2ad_hand = adddev->l2ad_start; 7108fa94a07fSbrendan adddev->l2ad_first = B_TRUE; 71095a98e54bSBrendan Gregg - Sun Microsystems adddev->l2ad_writing = B_FALSE; 7110fa94a07fSbrendan 711189c86e32SChris Williamson mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 7112fa94a07fSbrendan /* 7113fa94a07fSbrendan * This is a list of all ARC buffers that are still valid on the 7114fa94a07fSbrendan * device. 7115fa94a07fSbrendan */ 711689c86e32SChris Williamson list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 711789c86e32SChris Williamson offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 7118fa94a07fSbrendan 7119b24ab676SJeff Bonwick vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 7120a52fc310SPrakash Surya refcount_create(&adddev->l2ad_alloc); 7121fa94a07fSbrendan 7122fa94a07fSbrendan /* 7123fa94a07fSbrendan * Add device to global list 7124fa94a07fSbrendan */ 7125fa94a07fSbrendan mutex_enter(&l2arc_dev_mtx); 7126fa94a07fSbrendan list_insert_head(l2arc_dev_list, adddev); 7127fa94a07fSbrendan atomic_inc_64(&l2arc_ndev); 7128fa94a07fSbrendan mutex_exit(&l2arc_dev_mtx); 7129fa94a07fSbrendan } 7130fa94a07fSbrendan 7131fa94a07fSbrendan /* 7132fa94a07fSbrendan * Remove a vdev from the L2ARC. 7133fa94a07fSbrendan */ 7134fa94a07fSbrendan void 7135fa94a07fSbrendan l2arc_remove_vdev(vdev_t *vd) 7136fa94a07fSbrendan { 7137fa94a07fSbrendan l2arc_dev_t *dev, *nextdev, *remdev = NULL; 7138fa94a07fSbrendan 7139fa94a07fSbrendan /* 7140fa94a07fSbrendan * Find the device by vdev 7141fa94a07fSbrendan */ 7142fa94a07fSbrendan mutex_enter(&l2arc_dev_mtx); 7143fa94a07fSbrendan for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 7144fa94a07fSbrendan nextdev = list_next(l2arc_dev_list, dev); 7145fa94a07fSbrendan if (vd == dev->l2ad_vdev) { 7146fa94a07fSbrendan remdev = dev; 7147fa94a07fSbrendan break; 7148fa94a07fSbrendan } 7149fa94a07fSbrendan } 7150dcbf3bd6SGeorge Wilson ASSERT3P(remdev, !=, NULL); 7151fa94a07fSbrendan 7152fa94a07fSbrendan /* 7153fa94a07fSbrendan * Remove device from global list 7154fa94a07fSbrendan */ 7155fa94a07fSbrendan list_remove(l2arc_dev_list, remdev); 7156fa94a07fSbrendan l2arc_dev_last = NULL; /* may have been invalidated */ 71573a737e0dSbrendan atomic_dec_64(&l2arc_ndev); 71583a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 7159fa94a07fSbrendan 7160fa94a07fSbrendan /* 7161fa94a07fSbrendan * Clear all buflists and ARC references. L2ARC device flush. 7162fa94a07fSbrendan */ 7163fa94a07fSbrendan l2arc_evict(remdev, 0, B_TRUE); 716489c86e32SChris Williamson list_destroy(&remdev->l2ad_buflist); 716589c86e32SChris Williamson mutex_destroy(&remdev->l2ad_mtx); 7166a52fc310SPrakash Surya refcount_destroy(&remdev->l2ad_alloc); 7167fa94a07fSbrendan kmem_free(remdev, sizeof (l2arc_dev_t)); 7168fa94a07fSbrendan } 7169fa94a07fSbrendan 7170fa94a07fSbrendan void 7171e14bb325SJeff Bonwick l2arc_init(void) 7172fa94a07fSbrendan { 7173fa94a07fSbrendan l2arc_thread_exit = 0; 7174fa94a07fSbrendan l2arc_ndev = 0; 7175fa94a07fSbrendan l2arc_writes_sent = 0; 7176fa94a07fSbrendan l2arc_writes_done = 0; 7177fa94a07fSbrendan 7178fa94a07fSbrendan mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7179fa94a07fSbrendan cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7180fa94a07fSbrendan mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7181fa94a07fSbrendan mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7182fa94a07fSbrendan 7183fa94a07fSbrendan l2arc_dev_list = &L2ARC_dev_list; 7184fa94a07fSbrendan l2arc_free_on_write = &L2ARC_free_on_write; 7185fa94a07fSbrendan list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7186fa94a07fSbrendan offsetof(l2arc_dev_t, l2ad_node)); 7187fa94a07fSbrendan list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7188fa94a07fSbrendan offsetof(l2arc_data_free_t, l2df_list_node)); 7189fa94a07fSbrendan } 7190fa94a07fSbrendan 7191fa94a07fSbrendan void 7192e14bb325SJeff Bonwick l2arc_fini(void) 7193fa94a07fSbrendan { 71943a737e0dSbrendan /* 71953a737e0dSbrendan * This is called from dmu_fini(), which is called from spa_fini(); 71963a737e0dSbrendan * Because of this, we can assume that all l2arc devices have 71973a737e0dSbrendan * already been removed when the pools themselves were removed. 71983a737e0dSbrendan */ 71993a737e0dSbrendan 72003a737e0dSbrendan l2arc_do_free_on_write(); 72013a737e0dSbrendan 7202fa94a07fSbrendan mutex_destroy(&l2arc_feed_thr_lock); 7203fa94a07fSbrendan cv_destroy(&l2arc_feed_thr_cv); 7204fa94a07fSbrendan mutex_destroy(&l2arc_dev_mtx); 7205fa94a07fSbrendan mutex_destroy(&l2arc_free_on_write_mtx); 7206fa94a07fSbrendan 7207fa94a07fSbrendan list_destroy(l2arc_dev_list); 7208fa94a07fSbrendan list_destroy(l2arc_free_on_write); 7209fa94a07fSbrendan } 7210e14bb325SJeff Bonwick 7211e14bb325SJeff Bonwick void 7212e14bb325SJeff Bonwick l2arc_start(void) 7213e14bb325SJeff Bonwick { 72148ad4d6ddSJeff Bonwick if (!(spa_mode_global & FWRITE)) 7215e14bb325SJeff Bonwick return; 7216e14bb325SJeff Bonwick 7217e14bb325SJeff Bonwick (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7218e14bb325SJeff Bonwick TS_RUN, minclsyspri); 7219e14bb325SJeff Bonwick } 7220e14bb325SJeff Bonwick 7221e14bb325SJeff Bonwick void 7222e14bb325SJeff Bonwick l2arc_stop(void) 7223e14bb325SJeff Bonwick { 72248ad4d6ddSJeff Bonwick if (!(spa_mode_global & FWRITE)) 7225e14bb325SJeff Bonwick return; 7226e14bb325SJeff Bonwick 7227e14bb325SJeff Bonwick mutex_enter(&l2arc_feed_thr_lock); 7228e14bb325SJeff Bonwick cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7229e14bb325SJeff Bonwick l2arc_thread_exit = 1; 7230e14bb325SJeff Bonwick while (l2arc_thread_exit != 0) 7231e14bb325SJeff Bonwick cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7232e14bb325SJeff Bonwick mutex_exit(&l2arc_feed_thr_lock); 7233e14bb325SJeff Bonwick } 7234