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. 248df0bcf0SPaul Dagnelie * Copyright (c) 2011, 2016 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 80*5602294fSDan 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 83*5602294fSDan 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 * 96*5602294fSDan Kimmel * Each ARC state also has a mutex which is used to protect the 97fa9e4066Sahrens * buffer list associated with the state. When attempting to 98*5602294fSDan 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 130*5602294fSDan Kimmel * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and 131dcbf3bd6SGeorge Wilson * also in the arc_buf_hdr_t's private physical data block pointer (b_pdata). 132*5602294fSDan Kimmel * 133*5602294fSDan Kimmel * The L1ARC's data pointer may or may not be uncompressed. The ARC has the 134*5602294fSDan Kimmel * ability to store the physical data (b_pdata) associated with the DVA of the 135*5602294fSDan Kimmel * arc_buf_hdr_t. Since the b_pdata is a copy of the on-disk physical block, 136*5602294fSDan Kimmel * it will match its on-disk compression characteristics. This behavior can be 137*5602294fSDan Kimmel * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the 138*5602294fSDan Kimmel * compressed ARC functionality is disabled, the b_pdata will point to an 139*5602294fSDan Kimmel * uncompressed version of the on-disk data. 140*5602294fSDan Kimmel * 141*5602294fSDan Kimmel * Data in the L1ARC is not accessed by consumers of the ARC directly. Each 142*5602294fSDan Kimmel * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it. 143*5602294fSDan Kimmel * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC 144*5602294fSDan Kimmel * consumer. The ARC will provide references to this data and will keep it 145*5602294fSDan Kimmel * cached until it is no longer in use. The ARC caches only the L1ARC's physical 146*5602294fSDan Kimmel * data block and will evict any arc_buf_t that is no longer referenced. The 147*5602294fSDan Kimmel * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the 148dcbf3bd6SGeorge Wilson * "overhead_size" kstat. 149dcbf3bd6SGeorge Wilson * 150*5602294fSDan Kimmel * Depending on the consumer, an arc_buf_t can be requested in uncompressed or 151*5602294fSDan Kimmel * compressed form. The typical case is that consumers will want uncompressed 152*5602294fSDan Kimmel * data, and when that happens a new data buffer is allocated where the data is 153*5602294fSDan Kimmel * decompressed for them to use. Currently the only consumer who wants 154*5602294fSDan Kimmel * compressed arc_buf_t's is "zfs send", when it streams data exactly as it 155*5602294fSDan Kimmel * exists on disk. When this happens, the arc_buf_t's data buffer is shared 156*5602294fSDan Kimmel * with the arc_buf_hdr_t. 157dcbf3bd6SGeorge Wilson * 158*5602294fSDan Kimmel * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The 159*5602294fSDan Kimmel * first one is owned by a compressed send consumer (and therefore references 160*5602294fSDan Kimmel * the same compressed data buffer as the arc_buf_hdr_t) and the second could be 161*5602294fSDan Kimmel * used by any other consumer (and has its own uncompressed copy of the data 162*5602294fSDan Kimmel * buffer). 163dcbf3bd6SGeorge Wilson * 164*5602294fSDan Kimmel * arc_buf_hdr_t 165*5602294fSDan Kimmel * +-----------+ 166*5602294fSDan Kimmel * | fields | 167*5602294fSDan Kimmel * | common to | 168*5602294fSDan Kimmel * | L1- and | 169*5602294fSDan Kimmel * | L2ARC | 170*5602294fSDan Kimmel * +-----------+ 171*5602294fSDan Kimmel * | l2arc_buf_hdr_t 172*5602294fSDan Kimmel * | | 173*5602294fSDan Kimmel * +-----------+ 174*5602294fSDan Kimmel * | l1arc_buf_hdr_t 175*5602294fSDan Kimmel * | | arc_buf_t 176*5602294fSDan Kimmel * | b_buf +------------>+-----------+ arc_buf_t 177*5602294fSDan Kimmel * | b_pdata +-+ |b_next +---->+-----------+ 178*5602294fSDan Kimmel * +-----------+ | |-----------| |b_next +-->NULL 179*5602294fSDan Kimmel * | |b_comp = T | +-----------+ 180*5602294fSDan Kimmel * | |b_data +-+ |b_comp = F | 181*5602294fSDan Kimmel * | +-----------+ | |b_data +-+ 182*5602294fSDan Kimmel * +->+------+ | +-----------+ | 183*5602294fSDan Kimmel * compressed | | | | 184*5602294fSDan Kimmel * data | |<--------------+ | uncompressed 185*5602294fSDan Kimmel * +------+ compressed, | data 186*5602294fSDan Kimmel * shared +-->+------+ 187*5602294fSDan Kimmel * data | | 188*5602294fSDan Kimmel * | | 189*5602294fSDan Kimmel * +------+ 190dcbf3bd6SGeorge Wilson * 191dcbf3bd6SGeorge Wilson * When a consumer reads a block, the ARC must first look to see if the 192*5602294fSDan Kimmel * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new 193*5602294fSDan Kimmel * arc_buf_t and either copies uncompressed data into a new data buffer from an 194*5602294fSDan Kimmel * existing uncompressed arc_buf_t, decompresses the hdr's b_pdata buffer into a 195*5602294fSDan Kimmel * new data buffer, or shares the hdr's b_pdata buffer, depending on whether the 196*5602294fSDan Kimmel * hdr is compressed and the desired compression characteristics of the 197*5602294fSDan Kimmel * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the 198*5602294fSDan Kimmel * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be 199*5602294fSDan Kimmel * the last buffer in the hdr's b_buf list, however a shared compressed buf can 200*5602294fSDan 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 203*5602294fSDan Kimmel * sharing its data with an arc_buf_t (note that the shared uncompressed buf is 204*5602294fSDan 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 +---->+---------+ 219dcbf3bd6SGeorge Wilson * | b_pdata +-+ |---------| |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*5602294fSDan Kimmel * Writing to the ARC requires that the ARC first discard the hdr's b_pdata 234dcbf3bd6SGeorge Wilson * since the physical block is about to be rewritten. The new data contents 235*5602294fSDan Kimmel * will be contained in the arc_buf_t. As the I/O pipeline performs the write, 236*5602294fSDan Kimmel * it may compress the data before writing it to disk. The ARC will be called 237*5602294fSDan Kimmel * with the transformed data and will bcopy the transformed on-disk block into 238*5602294fSDan Kimmel * a newly allocated b_pdata. Writes are always done into buffers which have 239*5602294fSDan Kimmel * either been loaned (and hence are new and don't have other readers) or 240*5602294fSDan Kimmel * buffers which have been released (and hence have their own hdr, if there 241*5602294fSDan Kimmel * were originally other readers of the buf's original hdr). This ensures that 242*5602294fSDan Kimmel * the ARC only needs to update a single buf and its hdr after a write occurs. 243dcbf3bd6SGeorge Wilson * 244dcbf3bd6SGeorge Wilson * When the L2ARC is in use, it will also take advantage of the b_pdata. The 245dcbf3bd6SGeorge Wilson * L2ARC will always write the contents of b_pdata to the L2ARC. This means 246*5602294fSDan 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 250*5602294fSDan 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> 266244781f1SPrakash Surya #include <sys/multilist.h> 267fa9e4066Sahrens #ifdef _KERNEL 268fa9e4066Sahrens #include <sys/vmsystm.h> 269fa9e4066Sahrens #include <vm/anon.h> 270fa9e4066Sahrens #include <sys/fs/swapnode.h> 271033f9833Sek #include <sys/dnlc.h> 272fa9e4066Sahrens #endif 273fa9e4066Sahrens #include <sys/callb.h> 27444cb6abcSbmc #include <sys/kstat.h> 275b24ab676SJeff Bonwick #include <zfs_fletcher.h> 276fa9e4066Sahrens 277cd1c8b85SMatthew Ahrens #ifndef _KERNEL 278cd1c8b85SMatthew Ahrens /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 279cd1c8b85SMatthew Ahrens boolean_t arc_watch = B_FALSE; 280cd1c8b85SMatthew Ahrens int arc_procfd; 281cd1c8b85SMatthew Ahrens #endif 282cd1c8b85SMatthew Ahrens 283244781f1SPrakash Surya static kmutex_t arc_reclaim_lock; 284244781f1SPrakash Surya static kcondvar_t arc_reclaim_thread_cv; 285244781f1SPrakash Surya static boolean_t arc_reclaim_thread_exit; 286244781f1SPrakash Surya static kcondvar_t arc_reclaim_waiters_cv; 287244781f1SPrakash Surya 2882ec99e3eSMatthew Ahrens uint_t arc_reduce_dnlc_percent = 3; 289fa9e4066Sahrens 29069962b56SMatthew Ahrens /* 291244781f1SPrakash Surya * The number of headers to evict in arc_evict_state_impl() before 292244781f1SPrakash Surya * dropping the sublist lock and evicting from another sublist. A lower 293244781f1SPrakash Surya * value means we're more likely to evict the "correct" header (i.e. the 294244781f1SPrakash Surya * oldest header in the arc state), but comes with higher overhead 295244781f1SPrakash Surya * (i.e. more invocations of arc_evict_state_impl()). 296244781f1SPrakash Surya */ 297244781f1SPrakash Surya int zfs_arc_evict_batch_limit = 10; 298244781f1SPrakash Surya 299244781f1SPrakash Surya /* 300244781f1SPrakash Surya * The number of sublists used for each of the arc state lists. If this 301244781f1SPrakash Surya * is not set to a suitable value by the user, it will be configured to 302244781f1SPrakash Surya * the number of CPUs on the system in arc_init(). 30369962b56SMatthew Ahrens */ 304244781f1SPrakash Surya int zfs_arc_num_sublists_per_state = 0; 30569962b56SMatthew Ahrens 306fa9e4066Sahrens /* number of seconds before growing cache again */ 307fa9e4066Sahrens static int arc_grow_retry = 60; 308fa9e4066Sahrens 309244781f1SPrakash Surya /* shift of arc_c for calculating overflow limit in arc_get_data_buf */ 310244781f1SPrakash Surya int zfs_arc_overflow_shift = 8; 311244781f1SPrakash Surya 3125a98e54bSBrendan Gregg - Sun Microsystems /* shift of arc_c for calculating both min and max arc_p */ 3135a98e54bSBrendan Gregg - Sun Microsystems static int arc_p_min_shift = 4; 3145a98e54bSBrendan Gregg - Sun Microsystems 3155a98e54bSBrendan Gregg - Sun Microsystems /* log2(fraction of arc to reclaim) */ 3162ec99e3eSMatthew Ahrens static int arc_shrink_shift = 7; 3172ec99e3eSMatthew Ahrens 3182ec99e3eSMatthew Ahrens /* 3192ec99e3eSMatthew Ahrens * log2(fraction of ARC which must be free to allow growing). 3202ec99e3eSMatthew Ahrens * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 3212ec99e3eSMatthew Ahrens * when reading a new block into the ARC, we will evict an equal-sized block 3222ec99e3eSMatthew Ahrens * from the ARC. 3232ec99e3eSMatthew Ahrens * 3242ec99e3eSMatthew Ahrens * This must be less than arc_shrink_shift, so that when we shrink the ARC, 3252ec99e3eSMatthew Ahrens * we will still not allow it to grow. 3262ec99e3eSMatthew Ahrens */ 3272ec99e3eSMatthew Ahrens int arc_no_grow_shift = 5; 3282ec99e3eSMatthew Ahrens 3295a98e54bSBrendan Gregg - Sun Microsystems 33013506d1eSmaybee /* 331b19a79ecSperrin * minimum lifespan of a prefetch block in clock ticks 332b19a79ecSperrin * (initialized in arc_init()) 33313506d1eSmaybee */ 334b19a79ecSperrin static int arc_min_prefetch_lifespan; 33513506d1eSmaybee 33669962b56SMatthew Ahrens /* 33769962b56SMatthew Ahrens * If this percent of memory is free, don't throttle. 33869962b56SMatthew Ahrens */ 33969962b56SMatthew Ahrens int arc_lotsfree_percent = 10; 34069962b56SMatthew Ahrens 341fa9e4066Sahrens static int arc_dead; 342fa9e4066Sahrens 3433a737e0dSbrendan /* 3443a737e0dSbrendan * The arc has filled available memory and has now warmed up. 3453a737e0dSbrendan */ 3463a737e0dSbrendan static boolean_t arc_warm; 3473a737e0dSbrendan 3480dd053d7SPrakash Surya /* 3490dd053d7SPrakash Surya * log2 fraction of the zio arena to keep free. 3500dd053d7SPrakash Surya */ 3510dd053d7SPrakash Surya int arc_zio_arena_free_shift = 2; 3520dd053d7SPrakash Surya 353a2eea2e1Sahrens /* 354a2eea2e1Sahrens * These tunables are for performance analysis. 355a2eea2e1Sahrens */ 356a2eea2e1Sahrens uint64_t zfs_arc_max; 357a2eea2e1Sahrens uint64_t zfs_arc_min; 3581116048bSek uint64_t zfs_arc_meta_limit = 0; 3593a5286a1SMatthew Ahrens uint64_t zfs_arc_meta_min = 0; 3605a98e54bSBrendan Gregg - Sun Microsystems int zfs_arc_grow_retry = 0; 3615a98e54bSBrendan Gregg - Sun Microsystems int zfs_arc_shrink_shift = 0; 3625a98e54bSBrendan Gregg - Sun Microsystems int zfs_arc_p_min_shift = 0; 36363e911b6SMatthew Ahrens int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 364a2eea2e1Sahrens 365dcbf3bd6SGeorge Wilson boolean_t zfs_compressed_arc_enabled = B_TRUE; 366dcbf3bd6SGeorge Wilson 367fa9e4066Sahrens /* 368fa94a07fSbrendan * Note that buffers can be in one of 6 states: 369fa9e4066Sahrens * ARC_anon - anonymous (discussed below) 370ea8dc4b6Seschrock * ARC_mru - recently used, currently cached 371ea8dc4b6Seschrock * ARC_mru_ghost - recentely used, no longer in cache 372ea8dc4b6Seschrock * ARC_mfu - frequently used, currently cached 373ea8dc4b6Seschrock * ARC_mfu_ghost - frequently used, no longer in cache 374fa94a07fSbrendan * ARC_l2c_only - exists in L2ARC but not other states 3750e8c6158Smaybee * When there are no active references to the buffer, they are 3760e8c6158Smaybee * are linked onto a list in one of these arc states. These are 3770e8c6158Smaybee * the only buffers that can be evicted or deleted. Within each 3780e8c6158Smaybee * state there are multiple lists, one for meta-data and one for 3790e8c6158Smaybee * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 3800e8c6158Smaybee * etc.) is tracked separately so that it can be managed more 381fa94a07fSbrendan * explicitly: favored over data, limited explicitly. 382fa9e4066Sahrens * 383fa9e4066Sahrens * Anonymous buffers are buffers that are not associated with 384fa9e4066Sahrens * a DVA. These are buffers that hold dirty block copies 385fa9e4066Sahrens * before they are written to stable storage. By definition, 386ea8dc4b6Seschrock * they are "ref'd" and are considered part of arc_mru 387fa9e4066Sahrens * that cannot be freed. Generally, they will aquire a DVA 388ea8dc4b6Seschrock * as they are written and migrate onto the arc_mru list. 389fa94a07fSbrendan * 390fa94a07fSbrendan * The ARC_l2c_only state is for buffers that are in the second 391fa94a07fSbrendan * level ARC but no longer in any of the ARC_m* lists. The second 392fa94a07fSbrendan * level ARC itself may also contain buffers that are in any of 393fa94a07fSbrendan * the ARC_m* states - meaning that a buffer can exist in two 394fa94a07fSbrendan * places. The reason for the ARC_l2c_only state is to keep the 395fa94a07fSbrendan * buffer header in the hash table, so that reads that hit the 396fa94a07fSbrendan * second level ARC benefit from these fast lookups. 397fa9e4066Sahrens */ 398fa9e4066Sahrens 399fa9e4066Sahrens typedef struct arc_state { 400244781f1SPrakash Surya /* 401244781f1SPrakash Surya * list of evictable buffers 402244781f1SPrakash Surya */ 403244781f1SPrakash Surya multilist_t arcs_list[ARC_BUFC_NUMTYPES]; 404244781f1SPrakash Surya /* 405244781f1SPrakash Surya * total amount of evictable data in this state 406244781f1SPrakash Surya */ 407dcbf3bd6SGeorge Wilson refcount_t arcs_esize[ARC_BUFC_NUMTYPES]; 408244781f1SPrakash Surya /* 409244781f1SPrakash Surya * total amount of data in this state; this includes: evictable, 410244781f1SPrakash Surya * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 411244781f1SPrakash Surya */ 4122fd872a7SPrakash Surya refcount_t arcs_size; 413fa9e4066Sahrens } arc_state_t; 414fa9e4066Sahrens 415fa94a07fSbrendan /* The 6 states: */ 416fa9e4066Sahrens static arc_state_t ARC_anon; 417ea8dc4b6Seschrock static arc_state_t ARC_mru; 418ea8dc4b6Seschrock static arc_state_t ARC_mru_ghost; 419ea8dc4b6Seschrock static arc_state_t ARC_mfu; 420ea8dc4b6Seschrock static arc_state_t ARC_mfu_ghost; 421fa94a07fSbrendan static arc_state_t ARC_l2c_only; 422fa9e4066Sahrens 42344cb6abcSbmc typedef struct arc_stats { 42444cb6abcSbmc kstat_named_t arcstat_hits; 42544cb6abcSbmc kstat_named_t arcstat_misses; 42644cb6abcSbmc kstat_named_t arcstat_demand_data_hits; 42744cb6abcSbmc kstat_named_t arcstat_demand_data_misses; 42844cb6abcSbmc kstat_named_t arcstat_demand_metadata_hits; 42944cb6abcSbmc kstat_named_t arcstat_demand_metadata_misses; 43044cb6abcSbmc kstat_named_t arcstat_prefetch_data_hits; 43144cb6abcSbmc kstat_named_t arcstat_prefetch_data_misses; 43244cb6abcSbmc kstat_named_t arcstat_prefetch_metadata_hits; 43344cb6abcSbmc kstat_named_t arcstat_prefetch_metadata_misses; 43444cb6abcSbmc kstat_named_t arcstat_mru_hits; 43544cb6abcSbmc kstat_named_t arcstat_mru_ghost_hits; 43644cb6abcSbmc kstat_named_t arcstat_mfu_hits; 43744cb6abcSbmc kstat_named_t arcstat_mfu_ghost_hits; 43844cb6abcSbmc kstat_named_t arcstat_deleted; 4393e30c24aSWill Andrews /* 4403e30c24aSWill Andrews * Number of buffers that could not be evicted because the hash lock 4413e30c24aSWill Andrews * was held by another thread. The lock may not necessarily be held 4423e30c24aSWill Andrews * by something using the same buffer, since hash locks are shared 4433e30c24aSWill Andrews * by multiple buffers. 4443e30c24aSWill Andrews */ 44544cb6abcSbmc kstat_named_t arcstat_mutex_miss; 4463e30c24aSWill Andrews /* 4473e30c24aSWill Andrews * Number of buffers skipped because they have I/O in progress, are 4483e30c24aSWill Andrews * indrect prefetch buffers that have not lived long enough, or are 4493e30c24aSWill Andrews * not from the spa we're trying to evict from. 4503e30c24aSWill Andrews */ 45144cb6abcSbmc kstat_named_t arcstat_evict_skip; 452244781f1SPrakash Surya /* 453244781f1SPrakash Surya * Number of times arc_evict_state() was unable to evict enough 454244781f1SPrakash Surya * buffers to reach it's target amount. 455244781f1SPrakash Surya */ 456244781f1SPrakash Surya kstat_named_t arcstat_evict_not_enough; 4575ea40c06SBrendan Gregg - Sun Microsystems kstat_named_t arcstat_evict_l2_cached; 4585ea40c06SBrendan Gregg - Sun Microsystems kstat_named_t arcstat_evict_l2_eligible; 4595ea40c06SBrendan Gregg - Sun Microsystems kstat_named_t arcstat_evict_l2_ineligible; 460244781f1SPrakash Surya kstat_named_t arcstat_evict_l2_skip; 46144cb6abcSbmc kstat_named_t arcstat_hash_elements; 46244cb6abcSbmc kstat_named_t arcstat_hash_elements_max; 46344cb6abcSbmc kstat_named_t arcstat_hash_collisions; 46444cb6abcSbmc kstat_named_t arcstat_hash_chains; 46544cb6abcSbmc kstat_named_t arcstat_hash_chain_max; 46644cb6abcSbmc kstat_named_t arcstat_p; 46744cb6abcSbmc kstat_named_t arcstat_c; 46844cb6abcSbmc kstat_named_t arcstat_c_min; 46944cb6abcSbmc kstat_named_t arcstat_c_max; 47044cb6abcSbmc kstat_named_t arcstat_size; 471dcbf3bd6SGeorge Wilson /* 472dcbf3bd6SGeorge Wilson * Number of compressed bytes stored in the arc_buf_hdr_t's b_pdata. 473dcbf3bd6SGeorge Wilson * Note that the compressed bytes may match the uncompressed bytes 474dcbf3bd6SGeorge Wilson * if the block is either not compressed or compressed arc is disabled. 475dcbf3bd6SGeorge Wilson */ 476dcbf3bd6SGeorge Wilson kstat_named_t arcstat_compressed_size; 477dcbf3bd6SGeorge Wilson /* 478dcbf3bd6SGeorge Wilson * Uncompressed size of the data stored in b_pdata. If compressed 479dcbf3bd6SGeorge Wilson * arc is disabled then this value will be identical to the stat 480dcbf3bd6SGeorge Wilson * above. 481dcbf3bd6SGeorge Wilson */ 482dcbf3bd6SGeorge Wilson kstat_named_t arcstat_uncompressed_size; 483dcbf3bd6SGeorge Wilson /* 484dcbf3bd6SGeorge Wilson * Number of bytes stored in all the arc_buf_t's. This is classified 485dcbf3bd6SGeorge Wilson * as "overhead" since this data is typically short-lived and will 486dcbf3bd6SGeorge Wilson * be evicted from the arc when it becomes unreferenced unless the 487dcbf3bd6SGeorge Wilson * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level 488dcbf3bd6SGeorge Wilson * values have been set (see comment in dbuf.c for more information). 489dcbf3bd6SGeorge Wilson */ 490dcbf3bd6SGeorge Wilson kstat_named_t arcstat_overhead_size; 4914076b1bfSPrakash Surya /* 4924076b1bfSPrakash Surya * Number of bytes consumed by internal ARC structures necessary 4934076b1bfSPrakash Surya * for tracking purposes; these structures are not actually 4944076b1bfSPrakash Surya * backed by ARC buffers. This includes arc_buf_hdr_t structures 4954076b1bfSPrakash Surya * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 4964076b1bfSPrakash Surya * caches), and arc_buf_t structures (allocated via arc_buf_t 4974076b1bfSPrakash Surya * cache). 4984076b1bfSPrakash Surya */ 499fa94a07fSbrendan kstat_named_t arcstat_hdr_size; 5004076b1bfSPrakash Surya /* 5014076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers of type equal to 5024076b1bfSPrakash Surya * ARC_BUFC_DATA. This is generally consumed by buffers backing 5034076b1bfSPrakash Surya * on disk user data (e.g. plain file contents). 5044076b1bfSPrakash Surya */ 5055a98e54bSBrendan Gregg - Sun Microsystems kstat_named_t arcstat_data_size; 5064076b1bfSPrakash Surya /* 5074076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers of type equal to 5084076b1bfSPrakash Surya * ARC_BUFC_METADATA. This is generally consumed by buffers 5094076b1bfSPrakash Surya * backing on disk data that is used for internal ZFS 5104076b1bfSPrakash Surya * structures (e.g. ZAP, dnode, indirect blocks, etc). 5114076b1bfSPrakash Surya */ 5124076b1bfSPrakash Surya kstat_named_t arcstat_metadata_size; 5134076b1bfSPrakash Surya /* 5144076b1bfSPrakash Surya * Number of bytes consumed by various buffers and structures 5154076b1bfSPrakash Surya * not actually backed with ARC buffers. This includes bonus 5164076b1bfSPrakash Surya * buffers (allocated directly via zio_buf_* functions), 5174076b1bfSPrakash Surya * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 5184076b1bfSPrakash Surya * cache), and dnode_t structures (allocated via dnode_t cache). 5194076b1bfSPrakash Surya */ 5205a98e54bSBrendan Gregg - Sun Microsystems kstat_named_t arcstat_other_size; 5214076b1bfSPrakash Surya /* 5224076b1bfSPrakash Surya * Total number of bytes consumed by ARC buffers residing in the 5234076b1bfSPrakash Surya * arc_anon state. This includes *all* buffers in the arc_anon 5244076b1bfSPrakash Surya * state; e.g. data, metadata, evictable, and unevictable buffers 5254076b1bfSPrakash Surya * are all included in this value. 5264076b1bfSPrakash Surya */ 5274076b1bfSPrakash Surya kstat_named_t arcstat_anon_size; 5284076b1bfSPrakash Surya /* 5294076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that meet the 5304076b1bfSPrakash Surya * following criteria: backing buffers of type ARC_BUFC_DATA, 5314076b1bfSPrakash Surya * residing in the arc_anon state, and are eligible for eviction 5324076b1bfSPrakash Surya * (e.g. have no outstanding holds on the buffer). 5334076b1bfSPrakash Surya */ 5344076b1bfSPrakash Surya kstat_named_t arcstat_anon_evictable_data; 5354076b1bfSPrakash Surya /* 5364076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that meet the 5374076b1bfSPrakash Surya * following criteria: backing buffers of type ARC_BUFC_METADATA, 5384076b1bfSPrakash Surya * residing in the arc_anon state, and are eligible for eviction 5394076b1bfSPrakash Surya * (e.g. have no outstanding holds on the buffer). 5404076b1bfSPrakash Surya */ 5414076b1bfSPrakash Surya kstat_named_t arcstat_anon_evictable_metadata; 5424076b1bfSPrakash Surya /* 5434076b1bfSPrakash Surya * Total number of bytes consumed by ARC buffers residing in the 5444076b1bfSPrakash Surya * arc_mru state. This includes *all* buffers in the arc_mru 5454076b1bfSPrakash Surya * state; e.g. data, metadata, evictable, and unevictable buffers 5464076b1bfSPrakash Surya * are all included in this value. 5474076b1bfSPrakash Surya */ 5484076b1bfSPrakash Surya kstat_named_t arcstat_mru_size; 5494076b1bfSPrakash Surya /* 5504076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that meet the 5514076b1bfSPrakash Surya * following criteria: backing buffers of type ARC_BUFC_DATA, 5524076b1bfSPrakash Surya * residing in the arc_mru state, and are eligible for eviction 5534076b1bfSPrakash Surya * (e.g. have no outstanding holds on the buffer). 5544076b1bfSPrakash Surya */ 5554076b1bfSPrakash Surya kstat_named_t arcstat_mru_evictable_data; 5564076b1bfSPrakash Surya /* 5574076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that meet the 5584076b1bfSPrakash Surya * following criteria: backing buffers of type ARC_BUFC_METADATA, 5594076b1bfSPrakash Surya * residing in the arc_mru state, and are eligible for eviction 5604076b1bfSPrakash Surya * (e.g. have no outstanding holds on the buffer). 5614076b1bfSPrakash Surya */ 5624076b1bfSPrakash Surya kstat_named_t arcstat_mru_evictable_metadata; 5634076b1bfSPrakash Surya /* 5644076b1bfSPrakash Surya * Total number of bytes that *would have been* consumed by ARC 5654076b1bfSPrakash Surya * buffers in the arc_mru_ghost state. The key thing to note 5664076b1bfSPrakash Surya * here, is the fact that this size doesn't actually indicate 5674076b1bfSPrakash Surya * RAM consumption. The ghost lists only consist of headers and 5684076b1bfSPrakash Surya * don't actually have ARC buffers linked off of these headers. 5694076b1bfSPrakash Surya * Thus, *if* the headers had associated ARC buffers, these 5704076b1bfSPrakash Surya * buffers *would have* consumed this number of bytes. 5714076b1bfSPrakash Surya */ 5724076b1bfSPrakash Surya kstat_named_t arcstat_mru_ghost_size; 5734076b1bfSPrakash Surya /* 5744076b1bfSPrakash Surya * Number of bytes that *would have been* consumed by ARC 5754076b1bfSPrakash Surya * buffers that are eligible for eviction, of type 5764076b1bfSPrakash Surya * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 5774076b1bfSPrakash Surya */ 5784076b1bfSPrakash Surya kstat_named_t arcstat_mru_ghost_evictable_data; 5794076b1bfSPrakash Surya /* 5804076b1bfSPrakash Surya * Number of bytes that *would have been* consumed by ARC 5814076b1bfSPrakash Surya * buffers that are eligible for eviction, of type 5824076b1bfSPrakash Surya * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 5834076b1bfSPrakash Surya */ 5844076b1bfSPrakash Surya kstat_named_t arcstat_mru_ghost_evictable_metadata; 5854076b1bfSPrakash Surya /* 5864076b1bfSPrakash Surya * Total number of bytes consumed by ARC buffers residing in the 5874076b1bfSPrakash Surya * arc_mfu state. This includes *all* buffers in the arc_mfu 5884076b1bfSPrakash Surya * state; e.g. data, metadata, evictable, and unevictable buffers 5894076b1bfSPrakash Surya * are all included in this value. 5904076b1bfSPrakash Surya */ 5914076b1bfSPrakash Surya kstat_named_t arcstat_mfu_size; 5924076b1bfSPrakash Surya /* 5934076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that are eligible for 5944076b1bfSPrakash Surya * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 5954076b1bfSPrakash Surya * state. 5964076b1bfSPrakash Surya */ 5974076b1bfSPrakash Surya kstat_named_t arcstat_mfu_evictable_data; 5984076b1bfSPrakash Surya /* 5994076b1bfSPrakash Surya * Number of bytes consumed by ARC buffers that are eligible for 6004076b1bfSPrakash Surya * eviction, of type ARC_BUFC_METADATA, and reside in the 6014076b1bfSPrakash Surya * arc_mfu state. 6024076b1bfSPrakash Surya */ 6034076b1bfSPrakash Surya kstat_named_t arcstat_mfu_evictable_metadata; 6044076b1bfSPrakash Surya /* 6054076b1bfSPrakash Surya * Total number of bytes that *would have been* consumed by ARC 6064076b1bfSPrakash Surya * buffers in the arc_mfu_ghost state. See the comment above 6074076b1bfSPrakash Surya * arcstat_mru_ghost_size for more details. 6084076b1bfSPrakash Surya */ 6094076b1bfSPrakash Surya kstat_named_t arcstat_mfu_ghost_size; 6104076b1bfSPrakash Surya /* 6114076b1bfSPrakash Surya * Number of bytes that *would have been* consumed by ARC 6124076b1bfSPrakash Surya * buffers that are eligible for eviction, of type 6134076b1bfSPrakash Surya * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 6144076b1bfSPrakash Surya */ 6154076b1bfSPrakash Surya kstat_named_t arcstat_mfu_ghost_evictable_data; 6164076b1bfSPrakash Surya /* 6174076b1bfSPrakash Surya * Number of bytes that *would have been* consumed by ARC 6184076b1bfSPrakash Surya * buffers that are eligible for eviction, of type 6194076b1bfSPrakash Surya * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 6204076b1bfSPrakash Surya */ 6214076b1bfSPrakash Surya kstat_named_t arcstat_mfu_ghost_evictable_metadata; 622fa94a07fSbrendan kstat_named_t arcstat_l2_hits; 623fa94a07fSbrendan kstat_named_t arcstat_l2_misses; 624fa94a07fSbrendan kstat_named_t arcstat_l2_feeds; 625fa94a07fSbrendan kstat_named_t arcstat_l2_rw_clash; 6265a98e54bSBrendan Gregg - Sun Microsystems kstat_named_t arcstat_l2_read_bytes; 6275a98e54bSBrendan Gregg - Sun Microsystems kstat_named_t arcstat_l2_write_bytes; 628fa94a07fSbrendan kstat_named_t arcstat_l2_writes_sent; 629fa94a07fSbrendan kstat_named_t arcstat_l2_writes_done; 630fa94a07fSbrendan kstat_named_t arcstat_l2_writes_error; 631244781f1SPrakash Surya kstat_named_t arcstat_l2_writes_lock_retry; 632fa94a07fSbrendan kstat_named_t arcstat_l2_evict_lock_retry; 633fa94a07fSbrendan kstat_named_t arcstat_l2_evict_reading; 63489c86e32SChris Williamson kstat_named_t arcstat_l2_evict_l1cached; 635fa94a07fSbrendan kstat_named_t arcstat_l2_free_on_write; 636fa94a07fSbrendan kstat_named_t arcstat_l2_abort_lowmem; 637fa94a07fSbrendan kstat_named_t arcstat_l2_cksum_bad; 638fa94a07fSbrendan kstat_named_t arcstat_l2_io_error; 639fa94a07fSbrendan kstat_named_t arcstat_l2_size; 640aad02571SSaso Kiselkov kstat_named_t arcstat_l2_asize; 641fa94a07fSbrendan kstat_named_t arcstat_l2_hdr_size; 6421ab7f2deSmaybee kstat_named_t arcstat_memory_throttle_count; 64320128a08SGeorge Wilson kstat_named_t arcstat_meta_used; 64420128a08SGeorge Wilson kstat_named_t arcstat_meta_limit; 64520128a08SGeorge Wilson kstat_named_t arcstat_meta_max; 6463a5286a1SMatthew Ahrens kstat_named_t arcstat_meta_min; 647cf6106c8SMatthew Ahrens kstat_named_t arcstat_sync_wait_for_async; 648cf6106c8SMatthew Ahrens kstat_named_t arcstat_demand_hit_predictive_prefetch; 64944cb6abcSbmc } arc_stats_t; 65044cb6abcSbmc 65144cb6abcSbmc static arc_stats_t arc_stats = { 65244cb6abcSbmc { "hits", KSTAT_DATA_UINT64 }, 65344cb6abcSbmc { "misses", KSTAT_DATA_UINT64 }, 65444cb6abcSbmc { "demand_data_hits", KSTAT_DATA_UINT64 }, 65544cb6abcSbmc { "demand_data_misses", KSTAT_DATA_UINT64 }, 65644cb6abcSbmc { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 65744cb6abcSbmc { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 65844cb6abcSbmc { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 65944cb6abcSbmc { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 66044cb6abcSbmc { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 66144cb6abcSbmc { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 66244cb6abcSbmc { "mru_hits", KSTAT_DATA_UINT64 }, 66344cb6abcSbmc { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 66444cb6abcSbmc { "mfu_hits", KSTAT_DATA_UINT64 }, 66544cb6abcSbmc { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 66644cb6abcSbmc { "deleted", KSTAT_DATA_UINT64 }, 66744cb6abcSbmc { "mutex_miss", KSTAT_DATA_UINT64 }, 66844cb6abcSbmc { "evict_skip", KSTAT_DATA_UINT64 }, 669244781f1SPrakash Surya { "evict_not_enough", KSTAT_DATA_UINT64 }, 6705ea40c06SBrendan Gregg - Sun Microsystems { "evict_l2_cached", KSTAT_DATA_UINT64 }, 6715ea40c06SBrendan Gregg - Sun Microsystems { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 6725ea40c06SBrendan Gregg - Sun Microsystems { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 673244781f1SPrakash Surya { "evict_l2_skip", KSTAT_DATA_UINT64 }, 67444cb6abcSbmc { "hash_elements", KSTAT_DATA_UINT64 }, 67544cb6abcSbmc { "hash_elements_max", KSTAT_DATA_UINT64 }, 67644cb6abcSbmc { "hash_collisions", KSTAT_DATA_UINT64 }, 67744cb6abcSbmc { "hash_chains", KSTAT_DATA_UINT64 }, 67844cb6abcSbmc { "hash_chain_max", KSTAT_DATA_UINT64 }, 67944cb6abcSbmc { "p", KSTAT_DATA_UINT64 }, 68044cb6abcSbmc { "c", KSTAT_DATA_UINT64 }, 68144cb6abcSbmc { "c_min", KSTAT_DATA_UINT64 }, 68244cb6abcSbmc { "c_max", KSTAT_DATA_UINT64 }, 683fa94a07fSbrendan { "size", KSTAT_DATA_UINT64 }, 684dcbf3bd6SGeorge Wilson { "compressed_size", KSTAT_DATA_UINT64 }, 685dcbf3bd6SGeorge Wilson { "uncompressed_size", KSTAT_DATA_UINT64 }, 686dcbf3bd6SGeorge Wilson { "overhead_size", KSTAT_DATA_UINT64 }, 687fa94a07fSbrendan { "hdr_size", KSTAT_DATA_UINT64 }, 6885a98e54bSBrendan Gregg - Sun Microsystems { "data_size", KSTAT_DATA_UINT64 }, 6894076b1bfSPrakash Surya { "metadata_size", KSTAT_DATA_UINT64 }, 6905a98e54bSBrendan Gregg - Sun Microsystems { "other_size", KSTAT_DATA_UINT64 }, 6914076b1bfSPrakash Surya { "anon_size", KSTAT_DATA_UINT64 }, 6924076b1bfSPrakash Surya { "anon_evictable_data", KSTAT_DATA_UINT64 }, 6934076b1bfSPrakash Surya { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 6944076b1bfSPrakash Surya { "mru_size", KSTAT_DATA_UINT64 }, 6954076b1bfSPrakash Surya { "mru_evictable_data", KSTAT_DATA_UINT64 }, 6964076b1bfSPrakash Surya { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 6974076b1bfSPrakash Surya { "mru_ghost_size", KSTAT_DATA_UINT64 }, 6984076b1bfSPrakash Surya { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 6994076b1bfSPrakash Surya { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 7004076b1bfSPrakash Surya { "mfu_size", KSTAT_DATA_UINT64 }, 7014076b1bfSPrakash Surya { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 7024076b1bfSPrakash Surya { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 7034076b1bfSPrakash Surya { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 7044076b1bfSPrakash Surya { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 7054076b1bfSPrakash Surya { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 706fa94a07fSbrendan { "l2_hits", KSTAT_DATA_UINT64 }, 707fa94a07fSbrendan { "l2_misses", KSTAT_DATA_UINT64 }, 708fa94a07fSbrendan { "l2_feeds", KSTAT_DATA_UINT64 }, 709fa94a07fSbrendan { "l2_rw_clash", KSTAT_DATA_UINT64 }, 7105a98e54bSBrendan Gregg - Sun Microsystems { "l2_read_bytes", KSTAT_DATA_UINT64 }, 7115a98e54bSBrendan Gregg - Sun Microsystems { "l2_write_bytes", KSTAT_DATA_UINT64 }, 712fa94a07fSbrendan { "l2_writes_sent", KSTAT_DATA_UINT64 }, 713fa94a07fSbrendan { "l2_writes_done", KSTAT_DATA_UINT64 }, 714fa94a07fSbrendan { "l2_writes_error", KSTAT_DATA_UINT64 }, 715244781f1SPrakash Surya { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 716fa94a07fSbrendan { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 717fa94a07fSbrendan { "l2_evict_reading", KSTAT_DATA_UINT64 }, 71889c86e32SChris Williamson { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 719fa94a07fSbrendan { "l2_free_on_write", KSTAT_DATA_UINT64 }, 720fa94a07fSbrendan { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 721fa94a07fSbrendan { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 722fa94a07fSbrendan { "l2_io_error", KSTAT_DATA_UINT64 }, 723fa94a07fSbrendan { "l2_size", KSTAT_DATA_UINT64 }, 724aad02571SSaso Kiselkov { "l2_asize", KSTAT_DATA_UINT64 }, 7251ab7f2deSmaybee { "l2_hdr_size", KSTAT_DATA_UINT64 }, 7269253d63dSGeorge Wilson { "memory_throttle_count", KSTAT_DATA_UINT64 }, 72720128a08SGeorge Wilson { "arc_meta_used", KSTAT_DATA_UINT64 }, 72820128a08SGeorge Wilson { "arc_meta_limit", KSTAT_DATA_UINT64 }, 7293a5286a1SMatthew Ahrens { "arc_meta_max", KSTAT_DATA_UINT64 }, 730cf6106c8SMatthew Ahrens { "arc_meta_min", KSTAT_DATA_UINT64 }, 731cf6106c8SMatthew Ahrens { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 732cf6106c8SMatthew Ahrens { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 73344cb6abcSbmc }; 73444cb6abcSbmc 73544cb6abcSbmc #define ARCSTAT(stat) (arc_stats.stat.value.ui64) 73644cb6abcSbmc 73744cb6abcSbmc #define ARCSTAT_INCR(stat, val) \ 738f7170741SWill Andrews atomic_add_64(&arc_stats.stat.value.ui64, (val)) 73944cb6abcSbmc 740b24ab676SJeff Bonwick #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 74144cb6abcSbmc #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 74244cb6abcSbmc 74344cb6abcSbmc #define ARCSTAT_MAX(stat, val) { \ 74444cb6abcSbmc uint64_t m; \ 74544cb6abcSbmc while ((val) > (m = arc_stats.stat.value.ui64) && \ 74644cb6abcSbmc (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 74744cb6abcSbmc continue; \ 74844cb6abcSbmc } 74944cb6abcSbmc 75044cb6abcSbmc #define ARCSTAT_MAXSTAT(stat) \ 75144cb6abcSbmc ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 75244cb6abcSbmc 75344cb6abcSbmc /* 75444cb6abcSbmc * We define a macro to allow ARC hits/misses to be easily broken down by 75544cb6abcSbmc * two separate conditions, giving a total of four different subtypes for 75644cb6abcSbmc * each of hits and misses (so eight statistics total). 75744cb6abcSbmc */ 75844cb6abcSbmc #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 75944cb6abcSbmc if (cond1) { \ 76044cb6abcSbmc if (cond2) { \ 76144cb6abcSbmc ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 76244cb6abcSbmc } else { \ 76344cb6abcSbmc ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 76444cb6abcSbmc } \ 76544cb6abcSbmc } else { \ 76644cb6abcSbmc if (cond2) { \ 76744cb6abcSbmc ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 76844cb6abcSbmc } else { \ 76944cb6abcSbmc ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 77044cb6abcSbmc } \ 77144cb6abcSbmc } 77244cb6abcSbmc 77344cb6abcSbmc kstat_t *arc_ksp; 774b24ab676SJeff Bonwick static arc_state_t *arc_anon; 77544cb6abcSbmc static arc_state_t *arc_mru; 77644cb6abcSbmc static arc_state_t *arc_mru_ghost; 77744cb6abcSbmc static arc_state_t *arc_mfu; 77844cb6abcSbmc static arc_state_t *arc_mfu_ghost; 779fa94a07fSbrendan static arc_state_t *arc_l2c_only; 78044cb6abcSbmc 78144cb6abcSbmc /* 78244cb6abcSbmc * There are several ARC variables that are critical to export as kstats -- 78344cb6abcSbmc * but we don't want to have to grovel around in the kstat whenever we wish to 78444cb6abcSbmc * manipulate them. For these variables, we therefore define them to be in 78544cb6abcSbmc * terms of the statistic variable. This assures that we are not introducing 78644cb6abcSbmc * the possibility of inconsistency by having shadow copies of the variables, 78744cb6abcSbmc * while still allowing the code to be readable. 78844cb6abcSbmc */ 78944cb6abcSbmc #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 79044cb6abcSbmc #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 79144cb6abcSbmc #define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 79244cb6abcSbmc #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 79344cb6abcSbmc #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 79420128a08SGeorge Wilson #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 7953a5286a1SMatthew Ahrens #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 79620128a08SGeorge Wilson #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 79720128a08SGeorge Wilson #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 79844cb6abcSbmc 799dcbf3bd6SGeorge Wilson /* compressed size of entire arc */ 800dcbf3bd6SGeorge Wilson #define arc_compressed_size ARCSTAT(arcstat_compressed_size) 801dcbf3bd6SGeorge Wilson /* uncompressed size of entire arc */ 802dcbf3bd6SGeorge Wilson #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size) 803dcbf3bd6SGeorge Wilson /* number of bytes in the arc from arc_buf_t's */ 804dcbf3bd6SGeorge Wilson #define arc_overhead_size ARCSTAT(arcstat_overhead_size) 805aad02571SSaso Kiselkov 80644cb6abcSbmc static int arc_no_grow; /* Don't try to grow cache size */ 80744cb6abcSbmc static uint64_t arc_tempreserve; 8082fdbea25SAleksandr Guzovskiy static uint64_t arc_loaned_bytes; 809fa9e4066Sahrens 810fa9e4066Sahrens typedef struct arc_callback arc_callback_t; 811fa9e4066Sahrens 812fa9e4066Sahrens struct arc_callback { 813fa9e4066Sahrens void *acb_private; 814c717a561Smaybee arc_done_func_t *acb_done; 815fa9e4066Sahrens arc_buf_t *acb_buf; 816*5602294fSDan Kimmel boolean_t acb_compressed; 817fa9e4066Sahrens zio_t *acb_zio_dummy; 818fa9e4066Sahrens arc_callback_t *acb_next; 819fa9e4066Sahrens }; 820fa9e4066Sahrens 821c717a561Smaybee typedef struct arc_write_callback arc_write_callback_t; 822c717a561Smaybee 823c717a561Smaybee struct arc_write_callback { 824c717a561Smaybee void *awcb_private; 825c717a561Smaybee arc_done_func_t *awcb_ready; 8268df0bcf0SPaul Dagnelie arc_done_func_t *awcb_children_ready; 82769962b56SMatthew Ahrens arc_done_func_t *awcb_physdone; 828c717a561Smaybee arc_done_func_t *awcb_done; 829c717a561Smaybee arc_buf_t *awcb_buf; 830c717a561Smaybee }; 831c717a561Smaybee 83289c86e32SChris Williamson /* 83389c86e32SChris Williamson * ARC buffers are separated into multiple structs as a memory saving measure: 83489c86e32SChris Williamson * - Common fields struct, always defined, and embedded within it: 83589c86e32SChris Williamson * - L2-only fields, always allocated but undefined when not in L2ARC 83689c86e32SChris Williamson * - L1-only fields, only allocated when in L1ARC 83789c86e32SChris Williamson * 83889c86e32SChris Williamson * Buffer in L1 Buffer only in L2 83989c86e32SChris Williamson * +------------------------+ +------------------------+ 84089c86e32SChris Williamson * | arc_buf_hdr_t | | arc_buf_hdr_t | 84189c86e32SChris Williamson * | | | | 84289c86e32SChris Williamson * | | | | 84389c86e32SChris Williamson * | | | | 84489c86e32SChris Williamson * +------------------------+ +------------------------+ 84589c86e32SChris Williamson * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 84689c86e32SChris Williamson * | (undefined if L1-only) | | | 84789c86e32SChris Williamson * +------------------------+ +------------------------+ 84889c86e32SChris Williamson * | l1arc_buf_hdr_t | 84989c86e32SChris Williamson * | | 85089c86e32SChris Williamson * | | 85189c86e32SChris Williamson * | | 85289c86e32SChris Williamson * | | 85389c86e32SChris Williamson * +------------------------+ 85489c86e32SChris Williamson * 85589c86e32SChris Williamson * Because it's possible for the L2ARC to become extremely large, we can wind 85689c86e32SChris Williamson * up eating a lot of memory in L2ARC buffer headers, so the size of a header 85789c86e32SChris Williamson * is minimized by only allocating the fields necessary for an L1-cached buffer 85889c86e32SChris Williamson * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 85989c86e32SChris Williamson * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 86089c86e32SChris Williamson * words in pointers. arc_hdr_realloc() is used to switch a header between 86189c86e32SChris Williamson * these two allocation states. 86289c86e32SChris Williamson */ 86389c86e32SChris Williamson typedef struct l1arc_buf_hdr { 8646b4acc8bSahrens kmutex_t b_freeze_lock; 865dcbf3bd6SGeorge Wilson zio_cksum_t *b_freeze_cksum; 86689c86e32SChris Williamson #ifdef ZFS_DEBUG 86789c86e32SChris Williamson /* 868*5602294fSDan Kimmel * Used for debugging with kmem_flags - by allocating and freeing 86989c86e32SChris Williamson * b_thawed when the buffer is thawed, we get a record of the stack 87089c86e32SChris Williamson * trace that thawed it. 87189c86e32SChris Williamson */ 8723f9d6ad7SLin Ling void *b_thawed; 87389c86e32SChris Williamson #endif 8746b4acc8bSahrens 875fa9e4066Sahrens arc_buf_t *b_buf; 876dcbf3bd6SGeorge Wilson uint32_t b_bufcnt; 87789c86e32SChris Williamson /* for waiting on writes to complete */ 878ad23a2dbSjohansen kcondvar_t b_cv; 879dcbf3bd6SGeorge Wilson uint8_t b_byteswap; 880ad23a2dbSjohansen 881fa9e4066Sahrens /* protected by arc state mutex */ 882fa9e4066Sahrens arc_state_t *b_state; 883244781f1SPrakash Surya multilist_node_t b_arc_node; 884fa9e4066Sahrens 885fa9e4066Sahrens /* updated atomically */ 886fa9e4066Sahrens clock_t b_arc_access; 887fa9e4066Sahrens 888fa9e4066Sahrens /* self protecting */ 889fa9e4066Sahrens refcount_t b_refcnt; 890fa94a07fSbrendan 89189c86e32SChris Williamson arc_callback_t *b_acb; 892dcbf3bd6SGeorge Wilson void *b_pdata; 89389c86e32SChris Williamson } l1arc_buf_hdr_t; 89489c86e32SChris Williamson 89589c86e32SChris Williamson typedef struct l2arc_dev l2arc_dev_t; 89689c86e32SChris Williamson 89789c86e32SChris Williamson typedef struct l2arc_buf_hdr { 89889c86e32SChris Williamson /* protected by arc_buf_hdr mutex */ 89989c86e32SChris Williamson l2arc_dev_t *b_dev; /* L2ARC device */ 90089c86e32SChris Williamson uint64_t b_daddr; /* disk address, offset byte */ 90189c86e32SChris Williamson 902fa94a07fSbrendan list_node_t b_l2node; 90389c86e32SChris Williamson } l2arc_buf_hdr_t; 90489c86e32SChris Williamson 90589c86e32SChris Williamson struct arc_buf_hdr { 90689c86e32SChris Williamson /* protected by hash lock */ 90789c86e32SChris Williamson dva_t b_dva; 90889c86e32SChris Williamson uint64_t b_birth; 90989c86e32SChris Williamson 910dcbf3bd6SGeorge Wilson arc_buf_contents_t b_type; 91189c86e32SChris Williamson arc_buf_hdr_t *b_hash_next; 91289c86e32SChris Williamson arc_flags_t b_flags; 91389c86e32SChris Williamson 914dcbf3bd6SGeorge Wilson /* 915dcbf3bd6SGeorge Wilson * This field stores the size of the data buffer after 916dcbf3bd6SGeorge Wilson * compression, and is set in the arc's zio completion handlers. 917dcbf3bd6SGeorge Wilson * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes). 918dcbf3bd6SGeorge Wilson * 919dcbf3bd6SGeorge Wilson * While the block pointers can store up to 32MB in their psize 920dcbf3bd6SGeorge Wilson * field, we can only store up to 32MB minus 512B. This is due 921dcbf3bd6SGeorge Wilson * to the bp using a bias of 1, whereas we use a bias of 0 (i.e. 922dcbf3bd6SGeorge Wilson * a field of zeros represents 512B in the bp). We can't use a 923dcbf3bd6SGeorge Wilson * bias of 1 since we need to reserve a psize of zero, here, to 924dcbf3bd6SGeorge Wilson * represent holes and embedded blocks. 925dcbf3bd6SGeorge Wilson * 926dcbf3bd6SGeorge Wilson * This isn't a problem in practice, since the maximum size of a 927dcbf3bd6SGeorge Wilson * buffer is limited to 16MB, so we never need to store 32MB in 928dcbf3bd6SGeorge Wilson * this field. Even in the upstream illumos code base, the 929dcbf3bd6SGeorge Wilson * maximum size of a buffer is limited to 16MB. 930dcbf3bd6SGeorge Wilson */ 931dcbf3bd6SGeorge Wilson uint16_t b_psize; 932dcbf3bd6SGeorge Wilson 933dcbf3bd6SGeorge Wilson /* 934dcbf3bd6SGeorge Wilson * This field stores the size of the data buffer before 935dcbf3bd6SGeorge Wilson * compression, and cannot change once set. It is in units 936dcbf3bd6SGeorge Wilson * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes) 937dcbf3bd6SGeorge Wilson */ 938dcbf3bd6SGeorge Wilson uint16_t b_lsize; /* immutable */ 939dcbf3bd6SGeorge Wilson uint64_t b_spa; /* immutable */ 94089c86e32SChris Williamson 94189c86e32SChris Williamson /* L2ARC fields. Undefined when not in L2ARC. */ 94289c86e32SChris Williamson l2arc_buf_hdr_t b_l2hdr; 94389c86e32SChris Williamson /* L1ARC fields. Undefined when in l2arc_only state */ 94489c86e32SChris Williamson l1arc_buf_hdr_t b_l1hdr; 945fa9e4066Sahrens }; 946fa9e4066Sahrens 947ea8dc4b6Seschrock #define GHOST_STATE(state) \ 948fa94a07fSbrendan ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 949fa94a07fSbrendan (state) == arc_l2c_only) 950ea8dc4b6Seschrock 9517adb730bSGeorge Wilson #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 9527adb730bSGeorge Wilson #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 9537adb730bSGeorge Wilson #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 9547adb730bSGeorge Wilson #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 955dcbf3bd6SGeorge Wilson #define HDR_COMPRESSION_ENABLED(hdr) \ 956dcbf3bd6SGeorge Wilson ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC) 95789c86e32SChris Williamson 9587adb730bSGeorge Wilson #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 9597adb730bSGeorge Wilson #define HDR_L2_READING(hdr) \ 960dcbf3bd6SGeorge Wilson (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 961dcbf3bd6SGeorge Wilson ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 9627adb730bSGeorge Wilson #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 9637adb730bSGeorge Wilson #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 9647adb730bSGeorge Wilson #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 965dcbf3bd6SGeorge Wilson #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA) 966fa9e4066Sahrens 96789c86e32SChris Williamson #define HDR_ISTYPE_METADATA(hdr) \ 968dcbf3bd6SGeorge Wilson ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 96989c86e32SChris Williamson #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 97089c86e32SChris Williamson 97189c86e32SChris Williamson #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 97289c86e32SChris Williamson #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 97389c86e32SChris Williamson 974dcbf3bd6SGeorge Wilson /* For storing compression mode in b_flags */ 975dcbf3bd6SGeorge Wilson #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1) 976dcbf3bd6SGeorge Wilson 977dcbf3bd6SGeorge Wilson #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \ 978dcbf3bd6SGeorge Wilson HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS)) 979dcbf3bd6SGeorge Wilson #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \ 980dcbf3bd6SGeorge Wilson HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp)); 981dcbf3bd6SGeorge Wilson 982dcbf3bd6SGeorge Wilson #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL) 983*5602294fSDan Kimmel #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED) 984*5602294fSDan Kimmel #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED) 985dcbf3bd6SGeorge Wilson 986e6c728e1Sbrendan /* 987e6c728e1Sbrendan * Other sizes 988e6c728e1Sbrendan */ 989e6c728e1Sbrendan 99089c86e32SChris Williamson #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 99189c86e32SChris Williamson #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 992e6c728e1Sbrendan 993fa9e4066Sahrens /* 994fa9e4066Sahrens * Hash table routines 995fa9e4066Sahrens */ 996fa9e4066Sahrens 997fa9e4066Sahrens #define HT_LOCK_PAD 64 998fa9e4066Sahrens 999fa9e4066Sahrens struct ht_lock { 1000fa9e4066Sahrens kmutex_t ht_lock; 1001fa9e4066Sahrens #ifdef _KERNEL 1002fa9e4066Sahrens unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 1003fa9e4066Sahrens #endif 1004fa9e4066Sahrens }; 1005fa9e4066Sahrens 1006fa9e4066Sahrens #define BUF_LOCKS 256 1007fa9e4066Sahrens typedef struct buf_hash_table { 1008fa9e4066Sahrens uint64_t ht_mask; 1009fa9e4066Sahrens arc_buf_hdr_t **ht_table; 1010fa9e4066Sahrens struct ht_lock ht_locks[BUF_LOCKS]; 1011fa9e4066Sahrens } buf_hash_table_t; 1012fa9e4066Sahrens 1013fa9e4066Sahrens static buf_hash_table_t buf_hash_table; 1014fa9e4066Sahrens 1015fa9e4066Sahrens #define BUF_HASH_INDEX(spa, dva, birth) \ 1016fa9e4066Sahrens (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 1017fa9e4066Sahrens #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 1018fa9e4066Sahrens #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 10193f9d6ad7SLin Ling #define HDR_LOCK(hdr) \ 10203f9d6ad7SLin Ling (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 1021fa9e4066Sahrens 1022fa9e4066Sahrens uint64_t zfs_crc64_table[256]; 1023fa9e4066Sahrens 1024fa94a07fSbrendan /* 1025fa94a07fSbrendan * Level 2 ARC 1026fa94a07fSbrendan */ 1027fa94a07fSbrendan 1028fa94a07fSbrendan #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 1029aad02571SSaso Kiselkov #define L2ARC_HEADROOM 2 /* num of writes */ 1030aad02571SSaso Kiselkov /* 1031aad02571SSaso Kiselkov * If we discover during ARC scan any buffers to be compressed, we boost 1032aad02571SSaso Kiselkov * our headroom for the next scanning cycle by this percentage multiple. 1033aad02571SSaso Kiselkov */ 1034aad02571SSaso Kiselkov #define L2ARC_HEADROOM_BOOST 200 10355a98e54bSBrendan Gregg - Sun Microsystems #define L2ARC_FEED_SECS 1 /* caching interval secs */ 10365a98e54bSBrendan Gregg - Sun Microsystems #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 1037fa94a07fSbrendan 1038fa94a07fSbrendan #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 1039fa94a07fSbrendan #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 1040fa94a07fSbrendan 1041f7170741SWill Andrews /* L2ARC Performance Tunables */ 1042fa94a07fSbrendan uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 10433a737e0dSbrendan uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1044fa94a07fSbrendan uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1045aad02571SSaso Kiselkov uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1046fa94a07fSbrendan uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 10475a98e54bSBrendan Gregg - Sun Microsystems uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1048fa94a07fSbrendan boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 10495a98e54bSBrendan Gregg - Sun Microsystems boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 10505a98e54bSBrendan Gregg - Sun Microsystems boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1051fa94a07fSbrendan 1052fa94a07fSbrendan /* 1053fa94a07fSbrendan * L2ARC Internals 1054fa94a07fSbrendan */ 105589c86e32SChris Williamson struct l2arc_dev { 1056fa94a07fSbrendan vdev_t *l2ad_vdev; /* vdev */ 1057fa94a07fSbrendan spa_t *l2ad_spa; /* spa */ 1058fa94a07fSbrendan uint64_t l2ad_hand; /* next write location */ 1059fa94a07fSbrendan uint64_t l2ad_start; /* first addr on device */ 1060fa94a07fSbrendan uint64_t l2ad_end; /* last addr on device */ 1061fa94a07fSbrendan boolean_t l2ad_first; /* first sweep through */ 10625a98e54bSBrendan Gregg - Sun Microsystems boolean_t l2ad_writing; /* currently writing */ 106389c86e32SChris Williamson kmutex_t l2ad_mtx; /* lock for buffer list */ 106489c86e32SChris Williamson list_t l2ad_buflist; /* buffer list */ 1065fa94a07fSbrendan list_node_t l2ad_node; /* device list node */ 1066a52fc310SPrakash Surya refcount_t l2ad_alloc; /* allocated bytes */ 106789c86e32SChris Williamson }; 1068fa94a07fSbrendan 1069fa94a07fSbrendan static list_t L2ARC_dev_list; /* device list */ 1070fa94a07fSbrendan static list_t *l2arc_dev_list; /* device list pointer */ 1071fa94a07fSbrendan static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1072fa94a07fSbrendan static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1073fa94a07fSbrendan static list_t L2ARC_free_on_write; /* free after write buf list */ 1074fa94a07fSbrendan static list_t *l2arc_free_on_write; /* free after write list ptr */ 1075fa94a07fSbrendan static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1076fa94a07fSbrendan static uint64_t l2arc_ndev; /* number of devices */ 1077fa94a07fSbrendan 1078fa94a07fSbrendan typedef struct l2arc_read_callback { 1079*5602294fSDan Kimmel arc_buf_hdr_t *l2rcb_hdr; /* read header */ 1080aad02571SSaso Kiselkov blkptr_t l2rcb_bp; /* original blkptr */ 10817802d7bfSMatthew Ahrens zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1082aad02571SSaso Kiselkov int l2rcb_flags; /* original flags */ 1083fa94a07fSbrendan } l2arc_read_callback_t; 1084fa94a07fSbrendan 1085fa94a07fSbrendan typedef struct l2arc_write_callback { 1086fa94a07fSbrendan l2arc_dev_t *l2wcb_dev; /* device info */ 1087fa94a07fSbrendan arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1088fa94a07fSbrendan } l2arc_write_callback_t; 1089fa94a07fSbrendan 1090fa94a07fSbrendan typedef struct l2arc_data_free { 1091fa94a07fSbrendan /* protected by l2arc_free_on_write_mtx */ 1092fa94a07fSbrendan void *l2df_data; 1093fa94a07fSbrendan size_t l2df_size; 1094dcbf3bd6SGeorge Wilson arc_buf_contents_t l2df_type; 1095fa94a07fSbrendan list_node_t l2df_list_node; 1096fa94a07fSbrendan } l2arc_data_free_t; 1097fa94a07fSbrendan 1098fa94a07fSbrendan static kmutex_t l2arc_feed_thr_lock; 1099fa94a07fSbrendan static kcondvar_t l2arc_feed_thr_cv; 1100fa94a07fSbrendan static uint8_t l2arc_thread_exit; 1101fa94a07fSbrendan 1102dcbf3bd6SGeorge Wilson static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); 1103dcbf3bd6SGeorge Wilson static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); 1104dcbf3bd6SGeorge Wilson static void arc_hdr_free_pdata(arc_buf_hdr_t *hdr); 1105dcbf3bd6SGeorge Wilson static void arc_hdr_alloc_pdata(arc_buf_hdr_t *); 11067adb730bSGeorge Wilson static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1107244781f1SPrakash Surya static boolean_t arc_is_overflowing(); 11087adb730bSGeorge Wilson static void arc_buf_watch(arc_buf_t *); 11097adb730bSGeorge Wilson 111089c86e32SChris Williamson static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 111189c86e32SChris Williamson static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1112dcbf3bd6SGeorge Wilson static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1113dcbf3bd6SGeorge Wilson static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 111489c86e32SChris Williamson 11157adb730bSGeorge Wilson static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 11167adb730bSGeorge Wilson static void l2arc_read_done(zio_t *); 1117fa94a07fSbrendan 1118fa9e4066Sahrens static uint64_t 1119ac05c741SMark Maybee buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1120fa9e4066Sahrens { 1121fa9e4066Sahrens uint8_t *vdva = (uint8_t *)dva; 1122fa9e4066Sahrens uint64_t crc = -1ULL; 1123fa9e4066Sahrens int i; 1124fa9e4066Sahrens 1125fa9e4066Sahrens ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1126fa9e4066Sahrens 1127fa9e4066Sahrens for (i = 0; i < sizeof (dva_t); i++) 1128fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1129fa9e4066Sahrens 1130ac05c741SMark Maybee crc ^= (spa>>8) ^ birth; 1131fa9e4066Sahrens 1132fa9e4066Sahrens return (crc); 1133fa9e4066Sahrens } 1134fa9e4066Sahrens 1135dcbf3bd6SGeorge Wilson #define HDR_EMPTY(hdr) \ 1136dcbf3bd6SGeorge Wilson ((hdr)->b_dva.dva_word[0] == 0 && \ 1137dcbf3bd6SGeorge Wilson (hdr)->b_dva.dva_word[1] == 0) 1138fa9e4066Sahrens 1139dcbf3bd6SGeorge Wilson #define HDR_EQUAL(spa, dva, birth, hdr) \ 1140dcbf3bd6SGeorge Wilson ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1141dcbf3bd6SGeorge Wilson ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1142dcbf3bd6SGeorge Wilson ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) 1143fa9e4066Sahrens 11443f9d6ad7SLin Ling static void 11453f9d6ad7SLin Ling buf_discard_identity(arc_buf_hdr_t *hdr) 11463f9d6ad7SLin Ling { 11473f9d6ad7SLin Ling hdr->b_dva.dva_word[0] = 0; 11483f9d6ad7SLin Ling hdr->b_dva.dva_word[1] = 0; 11493f9d6ad7SLin Ling hdr->b_birth = 0; 11503f9d6ad7SLin Ling } 11513f9d6ad7SLin Ling 1152fa9e4066Sahrens static arc_buf_hdr_t * 11535d7b4d43SMatthew Ahrens buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1154fa9e4066Sahrens { 11555d7b4d43SMatthew Ahrens const dva_t *dva = BP_IDENTITY(bp); 11565d7b4d43SMatthew Ahrens uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1157fa9e4066Sahrens uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1158fa9e4066Sahrens kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 11597adb730bSGeorge Wilson arc_buf_hdr_t *hdr; 1160fa9e4066Sahrens 1161fa9e4066Sahrens mutex_enter(hash_lock); 11627adb730bSGeorge Wilson for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 11637adb730bSGeorge Wilson hdr = hdr->b_hash_next) { 1164dcbf3bd6SGeorge Wilson if (HDR_EQUAL(spa, dva, birth, hdr)) { 1165fa9e4066Sahrens *lockp = hash_lock; 11667adb730bSGeorge Wilson return (hdr); 1167fa9e4066Sahrens } 1168fa9e4066Sahrens } 1169fa9e4066Sahrens mutex_exit(hash_lock); 1170fa9e4066Sahrens *lockp = NULL; 1171fa9e4066Sahrens return (NULL); 1172fa9e4066Sahrens } 1173fa9e4066Sahrens 1174fa9e4066Sahrens /* 1175fa9e4066Sahrens * Insert an entry into the hash table. If there is already an element 1176fa9e4066Sahrens * equal to elem in the hash table, then the already existing element 1177fa9e4066Sahrens * will be returned and the new element will not be inserted. 1178fa9e4066Sahrens * Otherwise returns NULL. 117989c86e32SChris Williamson * If lockp == NULL, the caller is assumed to already hold the hash lock. 1180fa9e4066Sahrens */ 1181fa9e4066Sahrens static arc_buf_hdr_t * 11827adb730bSGeorge Wilson buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1183fa9e4066Sahrens { 11847adb730bSGeorge Wilson uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1185fa9e4066Sahrens kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 11867adb730bSGeorge Wilson arc_buf_hdr_t *fhdr; 118744cb6abcSbmc uint32_t i; 1188fa9e4066Sahrens 11897adb730bSGeorge Wilson ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 11907adb730bSGeorge Wilson ASSERT(hdr->b_birth != 0); 11917adb730bSGeorge Wilson ASSERT(!HDR_IN_HASH_TABLE(hdr)); 119289c86e32SChris Williamson 119389c86e32SChris Williamson if (lockp != NULL) { 119489c86e32SChris Williamson *lockp = hash_lock; 119589c86e32SChris Williamson mutex_enter(hash_lock); 119689c86e32SChris Williamson } else { 119789c86e32SChris Williamson ASSERT(MUTEX_HELD(hash_lock)); 119889c86e32SChris Williamson } 119989c86e32SChris Williamson 12007adb730bSGeorge Wilson for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 12017adb730bSGeorge Wilson fhdr = fhdr->b_hash_next, i++) { 1202dcbf3bd6SGeorge Wilson if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 12037adb730bSGeorge Wilson return (fhdr); 1204fa9e4066Sahrens } 1205fa9e4066Sahrens 12067adb730bSGeorge Wilson hdr->b_hash_next = buf_hash_table.ht_table[idx]; 12077adb730bSGeorge Wilson buf_hash_table.ht_table[idx] = hdr; 1208dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1209fa9e4066Sahrens 1210fa9e4066Sahrens /* collect some hash table performance data */ 1211fa9e4066Sahrens if (i > 0) { 121244cb6abcSbmc ARCSTAT_BUMP(arcstat_hash_collisions); 1213fa9e4066Sahrens if (i == 1) 121444cb6abcSbmc ARCSTAT_BUMP(arcstat_hash_chains); 121544cb6abcSbmc 121644cb6abcSbmc ARCSTAT_MAX(arcstat_hash_chain_max, i); 1217fa9e4066Sahrens } 121844cb6abcSbmc 121944cb6abcSbmc ARCSTAT_BUMP(arcstat_hash_elements); 122044cb6abcSbmc ARCSTAT_MAXSTAT(arcstat_hash_elements); 1221fa9e4066Sahrens 1222fa9e4066Sahrens return (NULL); 1223fa9e4066Sahrens } 1224fa9e4066Sahrens 1225fa9e4066Sahrens static void 12267adb730bSGeorge Wilson buf_hash_remove(arc_buf_hdr_t *hdr) 1227fa9e4066Sahrens { 12287adb730bSGeorge Wilson arc_buf_hdr_t *fhdr, **hdrp; 12297adb730bSGeorge Wilson uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1230fa9e4066Sahrens 1231fa9e4066Sahrens ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 12327adb730bSGeorge Wilson ASSERT(HDR_IN_HASH_TABLE(hdr)); 1233fa9e4066Sahrens 12347adb730bSGeorge Wilson hdrp = &buf_hash_table.ht_table[idx]; 12357adb730bSGeorge Wilson while ((fhdr = *hdrp) != hdr) { 1236dcbf3bd6SGeorge Wilson ASSERT3P(fhdr, !=, NULL); 12377adb730bSGeorge Wilson hdrp = &fhdr->b_hash_next; 1238fa9e4066Sahrens } 12397adb730bSGeorge Wilson *hdrp = hdr->b_hash_next; 12407adb730bSGeorge Wilson hdr->b_hash_next = NULL; 1241dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1242fa9e4066Sahrens 1243fa9e4066Sahrens /* collect some hash table performance data */ 124444cb6abcSbmc ARCSTAT_BUMPDOWN(arcstat_hash_elements); 124544cb6abcSbmc 1246fa9e4066Sahrens if (buf_hash_table.ht_table[idx] && 1247fa9e4066Sahrens buf_hash_table.ht_table[idx]->b_hash_next == NULL) 124844cb6abcSbmc ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1249fa9e4066Sahrens } 1250fa9e4066Sahrens 1251fa9e4066Sahrens /* 1252fa9e4066Sahrens * Global data structures and functions for the buf kmem cache. 1253fa9e4066Sahrens */ 125489c86e32SChris Williamson static kmem_cache_t *hdr_full_cache; 125589c86e32SChris Williamson static kmem_cache_t *hdr_l2only_cache; 1256fa9e4066Sahrens static kmem_cache_t *buf_cache; 1257fa9e4066Sahrens 1258fa9e4066Sahrens static void 1259fa9e4066Sahrens buf_fini(void) 1260fa9e4066Sahrens { 1261fa9e4066Sahrens int i; 1262fa9e4066Sahrens 1263fa9e4066Sahrens kmem_free(buf_hash_table.ht_table, 1264fa9e4066Sahrens (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1265fa9e4066Sahrens for (i = 0; i < BUF_LOCKS; i++) 1266fa9e4066Sahrens mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 126789c86e32SChris Williamson kmem_cache_destroy(hdr_full_cache); 126889c86e32SChris Williamson kmem_cache_destroy(hdr_l2only_cache); 1269fa9e4066Sahrens kmem_cache_destroy(buf_cache); 1270fa9e4066Sahrens } 1271fa9e4066Sahrens 1272fa9e4066Sahrens /* 1273fa9e4066Sahrens * Constructor callback - called when the cache is empty 1274fa9e4066Sahrens * and a new buf is requested. 1275fa9e4066Sahrens */ 1276fa9e4066Sahrens /* ARGSUSED */ 1277fa9e4066Sahrens static int 127889c86e32SChris Williamson hdr_full_cons(void *vbuf, void *unused, int kmflag) 1279fa9e4066Sahrens { 12807adb730bSGeorge Wilson arc_buf_hdr_t *hdr = vbuf; 1281fa9e4066Sahrens 128289c86e32SChris Williamson bzero(hdr, HDR_FULL_SIZE); 128389c86e32SChris Williamson cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 128489c86e32SChris Williamson refcount_create(&hdr->b_l1hdr.b_refcnt); 128589c86e32SChris Williamson mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1286244781f1SPrakash Surya multilist_link_init(&hdr->b_l1hdr.b_arc_node); 128789c86e32SChris Williamson arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 128889c86e32SChris Williamson 128989c86e32SChris Williamson return (0); 129089c86e32SChris Williamson } 129189c86e32SChris Williamson 129289c86e32SChris Williamson /* ARGSUSED */ 129389c86e32SChris Williamson static int 129489c86e32SChris Williamson hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 129589c86e32SChris Williamson { 129689c86e32SChris Williamson arc_buf_hdr_t *hdr = vbuf; 129789c86e32SChris Williamson 129889c86e32SChris Williamson bzero(hdr, HDR_L2ONLY_SIZE); 129989c86e32SChris Williamson arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1300fa94a07fSbrendan 1301fa9e4066Sahrens return (0); 1302fa9e4066Sahrens } 1303fa9e4066Sahrens 13046f83844dSMark Maybee /* ARGSUSED */ 13056f83844dSMark Maybee static int 13066f83844dSMark Maybee buf_cons(void *vbuf, void *unused, int kmflag) 13076f83844dSMark Maybee { 13086f83844dSMark Maybee arc_buf_t *buf = vbuf; 13096f83844dSMark Maybee 13106f83844dSMark Maybee bzero(buf, sizeof (arc_buf_t)); 13113f9d6ad7SLin Ling mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 13125a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 13135a98e54bSBrendan Gregg - Sun Microsystems 13146f83844dSMark Maybee return (0); 13156f83844dSMark Maybee } 13166f83844dSMark Maybee 1317fa9e4066Sahrens /* 1318fa9e4066Sahrens * Destructor callback - called when a cached buf is 1319fa9e4066Sahrens * no longer required. 1320fa9e4066Sahrens */ 1321fa9e4066Sahrens /* ARGSUSED */ 1322fa9e4066Sahrens static void 132389c86e32SChris Williamson hdr_full_dest(void *vbuf, void *unused) 132489c86e32SChris Williamson { 132589c86e32SChris Williamson arc_buf_hdr_t *hdr = vbuf; 132689c86e32SChris Williamson 1327dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 132889c86e32SChris Williamson cv_destroy(&hdr->b_l1hdr.b_cv); 132989c86e32SChris Williamson refcount_destroy(&hdr->b_l1hdr.b_refcnt); 133089c86e32SChris Williamson mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1331244781f1SPrakash Surya ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 133289c86e32SChris Williamson arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 133389c86e32SChris Williamson } 133489c86e32SChris Williamson 133589c86e32SChris Williamson /* ARGSUSED */ 133689c86e32SChris Williamson static void 133789c86e32SChris Williamson hdr_l2only_dest(void *vbuf, void *unused) 1338fa9e4066Sahrens { 13397adb730bSGeorge Wilson arc_buf_hdr_t *hdr = vbuf; 1340fa9e4066Sahrens 1341dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 134289c86e32SChris Williamson arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1343fa9e4066Sahrens } 1344fa9e4066Sahrens 13456f83844dSMark Maybee /* ARGSUSED */ 13466f83844dSMark Maybee static void 13476f83844dSMark Maybee buf_dest(void *vbuf, void *unused) 13486f83844dSMark Maybee { 13496f83844dSMark Maybee arc_buf_t *buf = vbuf; 13506f83844dSMark Maybee 13513f9d6ad7SLin Ling mutex_destroy(&buf->b_evict_lock); 13525a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 13536f83844dSMark Maybee } 13546f83844dSMark Maybee 1355fa9e4066Sahrens /* 1356fa9e4066Sahrens * Reclaim callback -- invoked when memory is low. 1357fa9e4066Sahrens */ 1358fa9e4066Sahrens /* ARGSUSED */ 1359fa9e4066Sahrens static void 1360fa9e4066Sahrens hdr_recl(void *unused) 1361fa9e4066Sahrens { 1362fa9e4066Sahrens dprintf("hdr_recl called\n"); 136349e3519aSmaybee /* 136449e3519aSmaybee * umem calls the reclaim func when we destroy the buf cache, 136549e3519aSmaybee * which is after we do arc_fini(). 136649e3519aSmaybee */ 136749e3519aSmaybee if (!arc_dead) 1368244781f1SPrakash Surya cv_signal(&arc_reclaim_thread_cv); 1369fa9e4066Sahrens } 1370fa9e4066Sahrens 1371fa9e4066Sahrens static void 1372fa9e4066Sahrens buf_init(void) 1373fa9e4066Sahrens { 1374fa9e4066Sahrens uint64_t *ct; 1375ea8dc4b6Seschrock uint64_t hsize = 1ULL << 12; 1376fa9e4066Sahrens int i, j; 1377fa9e4066Sahrens 1378fa9e4066Sahrens /* 1379fa9e4066Sahrens * The hash table is big enough to fill all of physical memory 138063e911b6SMatthew Ahrens * with an average block size of zfs_arc_average_blocksize (default 8K). 138163e911b6SMatthew Ahrens * By default, the table will take up 138263e911b6SMatthew Ahrens * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1383fa9e4066Sahrens */ 138463e911b6SMatthew Ahrens while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE) 1385fa9e4066Sahrens hsize <<= 1; 1386ea8dc4b6Seschrock retry: 1387fa9e4066Sahrens buf_hash_table.ht_mask = hsize - 1; 1388ea8dc4b6Seschrock buf_hash_table.ht_table = 1389ea8dc4b6Seschrock kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1390ea8dc4b6Seschrock if (buf_hash_table.ht_table == NULL) { 1391ea8dc4b6Seschrock ASSERT(hsize > (1ULL << 8)); 1392ea8dc4b6Seschrock hsize >>= 1; 1393ea8dc4b6Seschrock goto retry; 1394ea8dc4b6Seschrock } 1395fa9e4066Sahrens 139689c86e32SChris Williamson hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 139789c86e32SChris Williamson 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 139889c86e32SChris Williamson hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 139989c86e32SChris Williamson HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 140089c86e32SChris Williamson NULL, NULL, 0); 1401fa9e4066Sahrens buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 14026f83844dSMark Maybee 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1403fa9e4066Sahrens 1404fa9e4066Sahrens for (i = 0; i < 256; i++) 1405fa9e4066Sahrens for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1406fa9e4066Sahrens *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1407fa9e4066Sahrens 1408fa9e4066Sahrens for (i = 0; i < BUF_LOCKS; i++) { 1409fa9e4066Sahrens mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1410fa9e4066Sahrens NULL, MUTEX_DEFAULT, NULL); 1411fa9e4066Sahrens } 1412fa9e4066Sahrens } 1413fa9e4066Sahrens 1414*5602294fSDan Kimmel /* 1415*5602294fSDan Kimmel * This is the size that the buf occupies in memory. If the buf is compressed, 1416*5602294fSDan Kimmel * it will correspond to the compressed size. You should use this method of 1417*5602294fSDan Kimmel * getting the buf size unless you explicitly need the logical size. 1418*5602294fSDan Kimmel */ 1419*5602294fSDan Kimmel int32_t 1420*5602294fSDan Kimmel arc_buf_size(arc_buf_t *buf) 1421*5602294fSDan Kimmel { 1422*5602294fSDan Kimmel return (ARC_BUF_COMPRESSED(buf) ? 1423*5602294fSDan Kimmel HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr)); 1424*5602294fSDan Kimmel } 1425*5602294fSDan Kimmel 1426*5602294fSDan Kimmel int32_t 1427*5602294fSDan Kimmel arc_buf_lsize(arc_buf_t *buf) 1428*5602294fSDan Kimmel { 1429*5602294fSDan Kimmel return (HDR_GET_LSIZE(buf->b_hdr)); 1430*5602294fSDan Kimmel } 1431*5602294fSDan Kimmel 1432*5602294fSDan Kimmel enum zio_compress 1433*5602294fSDan Kimmel arc_get_compression(arc_buf_t *buf) 1434*5602294fSDan Kimmel { 1435*5602294fSDan Kimmel return (ARC_BUF_COMPRESSED(buf) ? 1436*5602294fSDan Kimmel HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF); 1437*5602294fSDan Kimmel } 1438*5602294fSDan Kimmel 1439dcbf3bd6SGeorge Wilson #define ARC_MINTIME (hz>>4) /* 62 ms */ 1440244781f1SPrakash Surya 1441dcbf3bd6SGeorge Wilson static inline boolean_t 1442dcbf3bd6SGeorge Wilson arc_buf_is_shared(arc_buf_t *buf) 1443dcbf3bd6SGeorge Wilson { 1444dcbf3bd6SGeorge Wilson boolean_t shared = (buf->b_data != NULL && 1445dcbf3bd6SGeorge Wilson buf->b_data == buf->b_hdr->b_l1hdr.b_pdata); 1446dcbf3bd6SGeorge Wilson IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); 1447*5602294fSDan Kimmel IMPLY(shared, ARC_BUF_SHARED(buf)); 1448*5602294fSDan Kimmel IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf)); 1449*5602294fSDan Kimmel 1450*5602294fSDan Kimmel /* 1451*5602294fSDan Kimmel * It would be nice to assert arc_can_share() too, but the "hdr isn't 1452*5602294fSDan Kimmel * already being shared" requirement prevents us from doing that. 1453*5602294fSDan Kimmel */ 1454*5602294fSDan Kimmel 1455dcbf3bd6SGeorge Wilson return (shared); 1456dcbf3bd6SGeorge Wilson } 1457c546f36aSArne Jansen 1458*5602294fSDan Kimmel /* 1459*5602294fSDan Kimmel * Free the checksum associated with this header. If there is no checksum, this 1460*5602294fSDan Kimmel * is a no-op. 1461*5602294fSDan Kimmel */ 1462dcbf3bd6SGeorge Wilson static inline void 1463dcbf3bd6SGeorge Wilson arc_cksum_free(arc_buf_hdr_t *hdr) 1464dcbf3bd6SGeorge Wilson { 1465dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1466dcbf3bd6SGeorge Wilson mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1467dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1468dcbf3bd6SGeorge Wilson kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); 1469dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_freeze_cksum = NULL; 147089c86e32SChris Williamson } 1471dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 147289c86e32SChris Williamson } 147389c86e32SChris Williamson 1474*5602294fSDan Kimmel /* 1475*5602294fSDan Kimmel * Return true iff at least one of the bufs on hdr is not compressed. 1476*5602294fSDan Kimmel */ 1477*5602294fSDan Kimmel static boolean_t 1478*5602294fSDan Kimmel arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr) 1479*5602294fSDan Kimmel { 1480*5602294fSDan Kimmel for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) { 1481*5602294fSDan Kimmel if (!ARC_BUF_COMPRESSED(b)) { 1482*5602294fSDan Kimmel return (B_TRUE); 1483*5602294fSDan Kimmel } 1484*5602294fSDan Kimmel } 1485*5602294fSDan Kimmel return (B_FALSE); 1486*5602294fSDan Kimmel } 1487*5602294fSDan Kimmel 1488*5602294fSDan Kimmel /* 1489*5602294fSDan Kimmel * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data 1490*5602294fSDan Kimmel * matches the checksum that is stored in the hdr. If there is no checksum, 1491*5602294fSDan Kimmel * or if the buf is compressed, this is a no-op. 1492*5602294fSDan Kimmel */ 14936b4acc8bSahrens static void 14946b4acc8bSahrens arc_cksum_verify(arc_buf_t *buf) 14956b4acc8bSahrens { 1496dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 14976b4acc8bSahrens zio_cksum_t zc; 14986b4acc8bSahrens 1499cc60fd72Sahrens if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 15006b4acc8bSahrens return; 15016b4acc8bSahrens 1502*5602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 1503*5602294fSDan Kimmel ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1504*5602294fSDan Kimmel arc_hdr_has_uncompressed_buf(hdr)); 1505*5602294fSDan Kimmel return; 1506*5602294fSDan Kimmel } 1507*5602294fSDan Kimmel 1508dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1509dcbf3bd6SGeorge Wilson 1510dcbf3bd6SGeorge Wilson mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1511dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { 1512dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 15136b4acc8bSahrens return; 15146b4acc8bSahrens } 1515*5602294fSDan Kimmel 1516*5602294fSDan Kimmel fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc); 1517dcbf3bd6SGeorge Wilson if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) 15186b4acc8bSahrens panic("buffer modified while frozen!"); 1519dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 15206b4acc8bSahrens } 15216b4acc8bSahrens 1522dcbf3bd6SGeorge Wilson static boolean_t 1523dcbf3bd6SGeorge Wilson arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) 1524fa94a07fSbrendan { 1525dcbf3bd6SGeorge Wilson enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp); 1526dcbf3bd6SGeorge Wilson boolean_t valid_cksum; 1527fa94a07fSbrendan 1528dcbf3bd6SGeorge Wilson ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); 1529dcbf3bd6SGeorge Wilson VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); 1530dcbf3bd6SGeorge Wilson 1531dcbf3bd6SGeorge Wilson /* 1532dcbf3bd6SGeorge Wilson * We rely on the blkptr's checksum to determine if the block 1533dcbf3bd6SGeorge Wilson * is valid or not. When compressed arc is enabled, the l2arc 1534dcbf3bd6SGeorge Wilson * writes the block to the l2arc just as it appears in the pool. 1535dcbf3bd6SGeorge Wilson * This allows us to use the blkptr's checksum to validate the 1536dcbf3bd6SGeorge Wilson * data that we just read off of the l2arc without having to store 1537dcbf3bd6SGeorge Wilson * a separate checksum in the arc_buf_hdr_t. However, if compressed 1538dcbf3bd6SGeorge Wilson * arc is disabled, then the data written to the l2arc is always 1539dcbf3bd6SGeorge Wilson * uncompressed and won't match the block as it exists in the main 1540dcbf3bd6SGeorge Wilson * pool. When this is the case, we must first compress it if it is 1541dcbf3bd6SGeorge Wilson * compressed on the main pool before we can validate the checksum. 1542dcbf3bd6SGeorge Wilson */ 1543dcbf3bd6SGeorge Wilson if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) { 1544dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1545dcbf3bd6SGeorge Wilson uint64_t lsize = HDR_GET_LSIZE(hdr); 1546dcbf3bd6SGeorge Wilson uint64_t csize; 1547dcbf3bd6SGeorge Wilson 1548dcbf3bd6SGeorge Wilson void *cbuf = zio_buf_alloc(HDR_GET_PSIZE(hdr)); 1549dcbf3bd6SGeorge Wilson csize = zio_compress_data(compress, zio->io_data, cbuf, lsize); 1550dcbf3bd6SGeorge Wilson ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr)); 1551dcbf3bd6SGeorge Wilson if (csize < HDR_GET_PSIZE(hdr)) { 1552dcbf3bd6SGeorge Wilson /* 1553dcbf3bd6SGeorge Wilson * Compressed blocks are always a multiple of the 1554dcbf3bd6SGeorge Wilson * smallest ashift in the pool. Ideally, we would 1555dcbf3bd6SGeorge Wilson * like to round up the csize to the next 1556dcbf3bd6SGeorge Wilson * spa_min_ashift but that value may have changed 1557dcbf3bd6SGeorge Wilson * since the block was last written. Instead, 1558dcbf3bd6SGeorge Wilson * we rely on the fact that the hdr's psize 1559dcbf3bd6SGeorge Wilson * was set to the psize of the block when it was 1560dcbf3bd6SGeorge Wilson * last written. We set the csize to that value 1561dcbf3bd6SGeorge Wilson * and zero out any part that should not contain 1562dcbf3bd6SGeorge Wilson * data. 1563dcbf3bd6SGeorge Wilson */ 1564dcbf3bd6SGeorge Wilson bzero((char *)cbuf + csize, HDR_GET_PSIZE(hdr) - csize); 1565dcbf3bd6SGeorge Wilson csize = HDR_GET_PSIZE(hdr); 1566dcbf3bd6SGeorge Wilson } 1567dcbf3bd6SGeorge Wilson zio_push_transform(zio, cbuf, csize, HDR_GET_PSIZE(hdr), NULL); 1568dcbf3bd6SGeorge Wilson } 1569fa94a07fSbrendan 1570dcbf3bd6SGeorge Wilson /* 1571dcbf3bd6SGeorge Wilson * Block pointers always store the checksum for the logical data. 1572dcbf3bd6SGeorge Wilson * If the block pointer has the gang bit set, then the checksum 1573dcbf3bd6SGeorge Wilson * it represents is for the reconstituted data and not for an 1574dcbf3bd6SGeorge Wilson * individual gang member. The zio pipeline, however, must be able to 1575dcbf3bd6SGeorge Wilson * determine the checksum of each of the gang constituents so it 1576dcbf3bd6SGeorge Wilson * treats the checksum comparison differently than what we need 1577dcbf3bd6SGeorge Wilson * for l2arc blocks. This prevents us from using the 1578dcbf3bd6SGeorge Wilson * zio_checksum_error() interface directly. Instead we must call the 1579dcbf3bd6SGeorge Wilson * zio_checksum_error_impl() so that we can ensure the checksum is 1580dcbf3bd6SGeorge Wilson * generated using the correct checksum algorithm and accounts for the 1581dcbf3bd6SGeorge Wilson * logical I/O size and not just a gang fragment. 1582dcbf3bd6SGeorge Wilson */ 1583dcbf3bd6SGeorge Wilson valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp, 1584dcbf3bd6SGeorge Wilson BP_GET_CHECKSUM(zio->io_bp), zio->io_data, zio->io_size, 1585dcbf3bd6SGeorge Wilson zio->io_offset, NULL) == 0); 1586dcbf3bd6SGeorge Wilson zio_pop_transforms(zio); 1587dcbf3bd6SGeorge Wilson return (valid_cksum); 1588fa94a07fSbrendan } 1589fa94a07fSbrendan 1590*5602294fSDan Kimmel /* 1591*5602294fSDan Kimmel * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a 1592*5602294fSDan Kimmel * checksum and attaches it to the buf's hdr so that we can ensure that the buf 1593*5602294fSDan Kimmel * isn't modified later on. If buf is compressed or there is already a checksum 1594*5602294fSDan Kimmel * on the hdr, this is a no-op (we only checksum uncompressed bufs). 1595*5602294fSDan Kimmel */ 15966b4acc8bSahrens static void 1597dcbf3bd6SGeorge Wilson arc_cksum_compute(arc_buf_t *buf) 15986b4acc8bSahrens { 1599dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 1600dcbf3bd6SGeorge Wilson 1601dcbf3bd6SGeorge Wilson if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 16026b4acc8bSahrens return; 16036b4acc8bSahrens 1604dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1605*5602294fSDan Kimmel 160689c86e32SChris Williamson mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1607dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1608*5602294fSDan Kimmel ASSERT(arc_hdr_has_uncompressed_buf(hdr)); 1609*5602294fSDan Kimmel mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1610*5602294fSDan Kimmel return; 1611*5602294fSDan Kimmel } else if (ARC_BUF_COMPRESSED(buf)) { 1612dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 16136b4acc8bSahrens return; 16146b4acc8bSahrens } 1615*5602294fSDan Kimmel 1616*5602294fSDan Kimmel ASSERT(!ARC_BUF_COMPRESSED(buf)); 1617dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), 1618dcbf3bd6SGeorge Wilson KM_SLEEP); 1619*5602294fSDan Kimmel fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, 1620dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_freeze_cksum); 1621dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1622cd1c8b85SMatthew Ahrens arc_buf_watch(buf); 1623cd1c8b85SMatthew Ahrens } 1624cd1c8b85SMatthew Ahrens 1625cd1c8b85SMatthew Ahrens #ifndef _KERNEL 1626cd1c8b85SMatthew Ahrens typedef struct procctl { 1627cd1c8b85SMatthew Ahrens long cmd; 1628cd1c8b85SMatthew Ahrens prwatch_t prwatch; 1629cd1c8b85SMatthew Ahrens } procctl_t; 1630cd1c8b85SMatthew Ahrens #endif 1631cd1c8b85SMatthew Ahrens 1632cd1c8b85SMatthew Ahrens /* ARGSUSED */ 1633cd1c8b85SMatthew Ahrens static void 1634cd1c8b85SMatthew Ahrens arc_buf_unwatch(arc_buf_t *buf) 1635cd1c8b85SMatthew Ahrens { 1636cd1c8b85SMatthew Ahrens #ifndef _KERNEL 1637cd1c8b85SMatthew Ahrens if (arc_watch) { 1638cd1c8b85SMatthew Ahrens int result; 1639cd1c8b85SMatthew Ahrens procctl_t ctl; 1640cd1c8b85SMatthew Ahrens ctl.cmd = PCWATCH; 1641cd1c8b85SMatthew Ahrens ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1642cd1c8b85SMatthew Ahrens ctl.prwatch.pr_size = 0; 1643cd1c8b85SMatthew Ahrens ctl.prwatch.pr_wflags = 0; 1644cd1c8b85SMatthew Ahrens result = write(arc_procfd, &ctl, sizeof (ctl)); 1645cd1c8b85SMatthew Ahrens ASSERT3U(result, ==, sizeof (ctl)); 1646cd1c8b85SMatthew Ahrens } 1647cd1c8b85SMatthew Ahrens #endif 1648cd1c8b85SMatthew Ahrens } 1649cd1c8b85SMatthew Ahrens 1650cd1c8b85SMatthew Ahrens /* ARGSUSED */ 1651cd1c8b85SMatthew Ahrens static void 1652cd1c8b85SMatthew Ahrens arc_buf_watch(arc_buf_t *buf) 1653cd1c8b85SMatthew Ahrens { 1654cd1c8b85SMatthew Ahrens #ifndef _KERNEL 1655cd1c8b85SMatthew Ahrens if (arc_watch) { 1656cd1c8b85SMatthew Ahrens int result; 1657cd1c8b85SMatthew Ahrens procctl_t ctl; 1658cd1c8b85SMatthew Ahrens ctl.cmd = PCWATCH; 1659cd1c8b85SMatthew Ahrens ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1660*5602294fSDan Kimmel ctl.prwatch.pr_size = arc_buf_size(buf); 1661cd1c8b85SMatthew Ahrens ctl.prwatch.pr_wflags = WA_WRITE; 1662cd1c8b85SMatthew Ahrens result = write(arc_procfd, &ctl, sizeof (ctl)); 1663cd1c8b85SMatthew Ahrens ASSERT3U(result, ==, sizeof (ctl)); 1664cd1c8b85SMatthew Ahrens } 1665cd1c8b85SMatthew Ahrens #endif 16666b4acc8bSahrens } 16676b4acc8bSahrens 166889c86e32SChris Williamson static arc_buf_contents_t 166989c86e32SChris Williamson arc_buf_type(arc_buf_hdr_t *hdr) 167089c86e32SChris Williamson { 1671dcbf3bd6SGeorge Wilson arc_buf_contents_t type; 167289c86e32SChris Williamson if (HDR_ISTYPE_METADATA(hdr)) { 1673dcbf3bd6SGeorge Wilson type = ARC_BUFC_METADATA; 167489c86e32SChris Williamson } else { 1675dcbf3bd6SGeorge Wilson type = ARC_BUFC_DATA; 167689c86e32SChris Williamson } 1677dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 1678dcbf3bd6SGeorge Wilson return (type); 167989c86e32SChris Williamson } 168089c86e32SChris Williamson 1681*5602294fSDan Kimmel boolean_t 1682*5602294fSDan Kimmel arc_is_metadata(arc_buf_t *buf) 1683*5602294fSDan Kimmel { 1684*5602294fSDan Kimmel return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0); 1685*5602294fSDan Kimmel } 1686*5602294fSDan Kimmel 168789c86e32SChris Williamson static uint32_t 168889c86e32SChris Williamson arc_bufc_to_flags(arc_buf_contents_t type) 168989c86e32SChris Williamson { 169089c86e32SChris Williamson switch (type) { 169189c86e32SChris Williamson case ARC_BUFC_DATA: 169289c86e32SChris Williamson /* metadata field is 0 if buffer contains normal data */ 169389c86e32SChris Williamson return (0); 169489c86e32SChris Williamson case ARC_BUFC_METADATA: 169589c86e32SChris Williamson return (ARC_FLAG_BUFC_METADATA); 169689c86e32SChris Williamson default: 169789c86e32SChris Williamson break; 169889c86e32SChris Williamson } 169989c86e32SChris Williamson panic("undefined ARC buffer type!"); 170089c86e32SChris Williamson return ((uint32_t)-1); 170189c86e32SChris Williamson } 170289c86e32SChris Williamson 17036b4acc8bSahrens void 17046b4acc8bSahrens arc_buf_thaw(arc_buf_t *buf) 17056b4acc8bSahrens { 1706dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 1707dcbf3bd6SGeorge Wilson 1708*5602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 1709*5602294fSDan Kimmel ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1710*5602294fSDan Kimmel 1711*5602294fSDan Kimmel arc_cksum_verify(buf); 1712*5602294fSDan Kimmel 1713*5602294fSDan Kimmel /* 1714*5602294fSDan Kimmel * Compressed buffers do not manipulate the b_freeze_cksum or 1715*5602294fSDan Kimmel * allocate b_thawed. 1716*5602294fSDan Kimmel */ 1717*5602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 1718*5602294fSDan Kimmel ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1719*5602294fSDan Kimmel arc_hdr_has_uncompressed_buf(hdr)); 1720*5602294fSDan Kimmel return; 1721fa94a07fSbrendan } 17226b4acc8bSahrens 1723dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1724dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 17253f9d6ad7SLin Ling 1726dcbf3bd6SGeorge Wilson mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 172789c86e32SChris Williamson #ifdef ZFS_DEBUG 17283f9d6ad7SLin Ling if (zfs_flags & ZFS_DEBUG_MODIFY) { 1729dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_thawed != NULL) 1730dcbf3bd6SGeorge Wilson kmem_free(hdr->b_l1hdr.b_thawed, 1); 1731dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 17323f9d6ad7SLin Ling } 173389c86e32SChris Williamson #endif 17343f9d6ad7SLin Ling 1735dcbf3bd6SGeorge Wilson mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1736cd1c8b85SMatthew Ahrens 1737cd1c8b85SMatthew Ahrens arc_buf_unwatch(buf); 17386b4acc8bSahrens } 17396b4acc8bSahrens 17406b4acc8bSahrens void 17416b4acc8bSahrens arc_buf_freeze(arc_buf_t *buf) 17426b4acc8bSahrens { 1743dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 17443f9d6ad7SLin Ling kmutex_t *hash_lock; 17453f9d6ad7SLin Ling 1746cc60fd72Sahrens if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1747cc60fd72Sahrens return; 1748cc60fd72Sahrens 1749*5602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 1750*5602294fSDan Kimmel ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1751*5602294fSDan Kimmel arc_hdr_has_uncompressed_buf(hdr)); 1752*5602294fSDan Kimmel return; 1753*5602294fSDan Kimmel } 1754*5602294fSDan Kimmel 1755dcbf3bd6SGeorge Wilson hash_lock = HDR_LOCK(hdr); 17563f9d6ad7SLin Ling mutex_enter(hash_lock); 17573f9d6ad7SLin Ling 1758dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 1759dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL || 1760dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_state == arc_anon); 1761dcbf3bd6SGeorge Wilson arc_cksum_compute(buf); 17623f9d6ad7SLin Ling mutex_exit(hash_lock); 17636b4acc8bSahrens } 17646b4acc8bSahrens 1765dcbf3bd6SGeorge Wilson /* 1766dcbf3bd6SGeorge Wilson * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, 1767dcbf3bd6SGeorge Wilson * the following functions should be used to ensure that the flags are 1768dcbf3bd6SGeorge Wilson * updated in a thread-safe way. When manipulating the flags either 1769dcbf3bd6SGeorge Wilson * the hash_lock must be held or the hdr must be undiscoverable. This 1770dcbf3bd6SGeorge Wilson * ensures that we're not racing with any other threads when updating 1771dcbf3bd6SGeorge Wilson * the flags. 1772dcbf3bd6SGeorge Wilson */ 1773dcbf3bd6SGeorge Wilson static inline void 1774dcbf3bd6SGeorge Wilson arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1775dcbf3bd6SGeorge Wilson { 1776dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1777dcbf3bd6SGeorge Wilson hdr->b_flags |= flags; 1778dcbf3bd6SGeorge Wilson } 1779dcbf3bd6SGeorge Wilson 1780dcbf3bd6SGeorge Wilson static inline void 1781dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1782dcbf3bd6SGeorge Wilson { 1783dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1784dcbf3bd6SGeorge Wilson hdr->b_flags &= ~flags; 1785dcbf3bd6SGeorge Wilson } 1786dcbf3bd6SGeorge Wilson 1787dcbf3bd6SGeorge Wilson /* 1788dcbf3bd6SGeorge Wilson * Setting the compression bits in the arc_buf_hdr_t's b_flags is 1789dcbf3bd6SGeorge Wilson * done in a special way since we have to clear and set bits 1790dcbf3bd6SGeorge Wilson * at the same time. Consumers that wish to set the compression bits 1791dcbf3bd6SGeorge Wilson * must use this function to ensure that the flags are updated in 1792dcbf3bd6SGeorge Wilson * thread-safe manner. 1793dcbf3bd6SGeorge Wilson */ 1794fa9e4066Sahrens static void 1795dcbf3bd6SGeorge Wilson arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) 1796fa9e4066Sahrens { 1797dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1798fa9e4066Sahrens 1799dcbf3bd6SGeorge Wilson /* 1800dcbf3bd6SGeorge Wilson * Holes and embedded blocks will always have a psize = 0 so 1801dcbf3bd6SGeorge Wilson * we ignore the compression of the blkptr and set the 1802dcbf3bd6SGeorge Wilson * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF. 1803dcbf3bd6SGeorge Wilson * Holes and embedded blocks remain anonymous so we don't 1804dcbf3bd6SGeorge Wilson * want to uncompress them. Mark them as uncompressed. 1805dcbf3bd6SGeorge Wilson */ 1806dcbf3bd6SGeorge Wilson if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { 1807dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1808dcbf3bd6SGeorge Wilson HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 1809dcbf3bd6SGeorge Wilson ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); 1810dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1811dcbf3bd6SGeorge Wilson } else { 1812dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1813dcbf3bd6SGeorge Wilson HDR_SET_COMPRESS(hdr, cmp); 1814dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); 1815dcbf3bd6SGeorge Wilson ASSERT(HDR_COMPRESSION_ENABLED(hdr)); 1816dcbf3bd6SGeorge Wilson } 1817dcbf3bd6SGeorge Wilson } 1818244781f1SPrakash Surya 1819*5602294fSDan Kimmel /* 1820*5602294fSDan Kimmel * Looks for another buf on the same hdr which has the data decompressed, copies 1821*5602294fSDan Kimmel * from it, and returns true. If no such buf exists, returns false. 1822*5602294fSDan Kimmel */ 1823*5602294fSDan Kimmel static boolean_t 1824*5602294fSDan Kimmel arc_buf_try_copy_decompressed_data(arc_buf_t *buf) 1825*5602294fSDan Kimmel { 1826*5602294fSDan Kimmel arc_buf_hdr_t *hdr = buf->b_hdr; 1827*5602294fSDan Kimmel boolean_t copied = B_FALSE; 1828*5602294fSDan Kimmel 1829*5602294fSDan Kimmel ASSERT(HDR_HAS_L1HDR(hdr)); 1830*5602294fSDan Kimmel ASSERT3P(buf->b_data, !=, NULL); 1831*5602294fSDan Kimmel ASSERT(!ARC_BUF_COMPRESSED(buf)); 1832*5602294fSDan Kimmel 1833*5602294fSDan Kimmel for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL; 1834*5602294fSDan Kimmel from = from->b_next) { 1835*5602294fSDan Kimmel /* can't use our own data buffer */ 1836*5602294fSDan Kimmel if (from == buf) { 1837*5602294fSDan Kimmel continue; 1838*5602294fSDan Kimmel } 1839*5602294fSDan Kimmel 1840*5602294fSDan Kimmel if (!ARC_BUF_COMPRESSED(from)) { 1841*5602294fSDan Kimmel bcopy(from->b_data, buf->b_data, arc_buf_size(buf)); 1842*5602294fSDan Kimmel copied = B_TRUE; 1843*5602294fSDan Kimmel break; 1844*5602294fSDan Kimmel } 1845*5602294fSDan Kimmel } 1846*5602294fSDan Kimmel 1847*5602294fSDan Kimmel /* 1848*5602294fSDan Kimmel * There were no decompressed bufs, so there should not be a 1849*5602294fSDan Kimmel * checksum on the hdr either. 1850*5602294fSDan Kimmel */ 1851*5602294fSDan Kimmel EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL); 1852*5602294fSDan Kimmel 1853*5602294fSDan Kimmel return (copied); 1854*5602294fSDan Kimmel } 1855*5602294fSDan Kimmel 1856*5602294fSDan Kimmel /* 1857*5602294fSDan Kimmel * Given a buf that has a data buffer attached to it, this function will 1858*5602294fSDan Kimmel * efficiently fill the buf with data of the specified compression setting from 1859*5602294fSDan Kimmel * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr 1860*5602294fSDan Kimmel * are already sharing a data buf, no copy is performed. 1861*5602294fSDan Kimmel * 1862*5602294fSDan Kimmel * If the buf is marked as compressed but uncompressed data was requested, this 1863*5602294fSDan Kimmel * will allocate a new data buffer for the buf, remove that flag, and fill the 1864*5602294fSDan Kimmel * buf with uncompressed data. You can't request a compressed buf on a hdr with 1865*5602294fSDan Kimmel * uncompressed data, and (since we haven't added support for it yet) if you 1866*5602294fSDan Kimmel * want compressed data your buf must already be marked as compressed and have 1867*5602294fSDan Kimmel * the correct-sized data buffer. 1868*5602294fSDan Kimmel */ 1869dcbf3bd6SGeorge Wilson static int 1870*5602294fSDan Kimmel arc_buf_fill(arc_buf_t *buf, boolean_t compressed) 1871dcbf3bd6SGeorge Wilson { 1872dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 1873*5602294fSDan Kimmel boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 1874dcbf3bd6SGeorge Wilson dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; 187589c86e32SChris Williamson 1876*5602294fSDan Kimmel ASSERT3P(buf->b_data, !=, NULL); 1877*5602294fSDan Kimmel IMPLY(compressed, hdr_compressed); 1878*5602294fSDan Kimmel IMPLY(compressed, ARC_BUF_COMPRESSED(buf)); 1879*5602294fSDan Kimmel 1880*5602294fSDan Kimmel if (hdr_compressed == compressed) { 1881*5602294fSDan Kimmel if (!arc_buf_is_shared(buf)) { 1882*5602294fSDan Kimmel bcopy(hdr->b_l1hdr.b_pdata, buf->b_data, 1883*5602294fSDan Kimmel arc_buf_size(buf)); 1884*5602294fSDan Kimmel } 1885dcbf3bd6SGeorge Wilson } else { 1886*5602294fSDan Kimmel ASSERT(hdr_compressed); 1887*5602294fSDan Kimmel ASSERT(!compressed); 1888dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); 1889*5602294fSDan Kimmel 1890*5602294fSDan Kimmel /* 1891*5602294fSDan Kimmel * If the buf is sharing its data with the hdr, unlink it and 1892*5602294fSDan Kimmel * allocate a new data buffer for the buf. 1893*5602294fSDan Kimmel */ 1894*5602294fSDan Kimmel if (arc_buf_is_shared(buf)) { 1895*5602294fSDan Kimmel ASSERT(ARC_BUF_COMPRESSED(buf)); 1896*5602294fSDan Kimmel 1897*5602294fSDan Kimmel /* We need to give the buf it's own b_data */ 1898*5602294fSDan Kimmel buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 1899*5602294fSDan Kimmel buf->b_data = 1900*5602294fSDan Kimmel arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 1901*5602294fSDan Kimmel arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 1902*5602294fSDan Kimmel 1903*5602294fSDan Kimmel /* Previously overhead was 0; just add new overhead */ 1904*5602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 1905*5602294fSDan Kimmel } else if (ARC_BUF_COMPRESSED(buf)) { 1906*5602294fSDan Kimmel /* We need to reallocate the buf's b_data */ 1907*5602294fSDan Kimmel arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr), 1908*5602294fSDan Kimmel buf); 1909*5602294fSDan Kimmel buf->b_data = 1910*5602294fSDan Kimmel arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 1911*5602294fSDan Kimmel 1912*5602294fSDan Kimmel /* We increased the size of b_data; update overhead */ 1913*5602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, 1914*5602294fSDan Kimmel HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr)); 1915*5602294fSDan Kimmel } 1916*5602294fSDan Kimmel 1917*5602294fSDan Kimmel /* 1918*5602294fSDan Kimmel * Regardless of the buf's previous compression settings, it 1919*5602294fSDan Kimmel * should not be compressed at the end of this function. 1920*5602294fSDan Kimmel */ 1921*5602294fSDan Kimmel buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; 1922*5602294fSDan Kimmel 1923*5602294fSDan Kimmel /* 1924*5602294fSDan Kimmel * Try copying the data from another buf which already has a 1925*5602294fSDan Kimmel * decompressed version. If that's not possible, it's time to 1926*5602294fSDan Kimmel * bite the bullet and decompress the data from the hdr. 1927*5602294fSDan Kimmel */ 1928*5602294fSDan Kimmel if (arc_buf_try_copy_decompressed_data(buf)) { 1929*5602294fSDan Kimmel /* Skip byteswapping and checksumming (already done) */ 1930*5602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL); 1931*5602294fSDan Kimmel return (0); 1932*5602294fSDan Kimmel } else { 1933*5602294fSDan Kimmel int error = zio_decompress_data(HDR_GET_COMPRESS(hdr), 1934*5602294fSDan Kimmel hdr->b_l1hdr.b_pdata, buf->b_data, 1935*5602294fSDan Kimmel HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 1936*5602294fSDan Kimmel 1937*5602294fSDan Kimmel /* 1938*5602294fSDan Kimmel * Absent hardware errors or software bugs, this should 1939*5602294fSDan Kimmel * be impossible, but log it anyway so we can debug it. 1940*5602294fSDan Kimmel */ 1941*5602294fSDan Kimmel if (error != 0) { 1942*5602294fSDan Kimmel zfs_dbgmsg( 1943*5602294fSDan Kimmel "hdr %p, compress %d, psize %d, lsize %d", 1944*5602294fSDan Kimmel hdr, HDR_GET_COMPRESS(hdr), 1945*5602294fSDan Kimmel HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 1946*5602294fSDan Kimmel return (SET_ERROR(EIO)); 1947*5602294fSDan Kimmel } 1948ea8dc4b6Seschrock } 1949fa9e4066Sahrens } 1950*5602294fSDan Kimmel 1951*5602294fSDan Kimmel /* Byteswap the buf's data if necessary */ 1952dcbf3bd6SGeorge Wilson if (bswap != DMU_BSWAP_NUMFUNCS) { 1953dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(hdr)); 1954dcbf3bd6SGeorge Wilson ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); 1955dcbf3bd6SGeorge Wilson dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); 1956dcbf3bd6SGeorge Wilson } 1957*5602294fSDan Kimmel 1958*5602294fSDan Kimmel /* Compute the hdr's checksum if necessary */ 1959dcbf3bd6SGeorge Wilson arc_cksum_compute(buf); 1960*5602294fSDan Kimmel 1961dcbf3bd6SGeorge Wilson return (0); 1962fa9e4066Sahrens } 1963fa9e4066Sahrens 1964*5602294fSDan Kimmel int 1965*5602294fSDan Kimmel arc_decompress(arc_buf_t *buf) 1966*5602294fSDan Kimmel { 1967*5602294fSDan Kimmel return (arc_buf_fill(buf, B_FALSE)); 1968*5602294fSDan Kimmel } 1969*5602294fSDan Kimmel 1970dcbf3bd6SGeorge Wilson /* 1971dcbf3bd6SGeorge Wilson * Return the size of the block, b_pdata, that is stored in the arc_buf_hdr_t. 1972dcbf3bd6SGeorge Wilson */ 1973dcbf3bd6SGeorge Wilson static uint64_t 1974dcbf3bd6SGeorge Wilson arc_hdr_size(arc_buf_hdr_t *hdr) 1975fa9e4066Sahrens { 1976dcbf3bd6SGeorge Wilson uint64_t size; 1977fa9e4066Sahrens 1978dcbf3bd6SGeorge Wilson if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 1979dcbf3bd6SGeorge Wilson HDR_GET_PSIZE(hdr) > 0) { 1980dcbf3bd6SGeorge Wilson size = HDR_GET_PSIZE(hdr); 1981dcbf3bd6SGeorge Wilson } else { 1982dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); 1983dcbf3bd6SGeorge Wilson size = HDR_GET_LSIZE(hdr); 1984dcbf3bd6SGeorge Wilson } 1985dcbf3bd6SGeorge Wilson return (size); 1986dcbf3bd6SGeorge Wilson } 1987fa9e4066Sahrens 1988dcbf3bd6SGeorge Wilson /* 1989dcbf3bd6SGeorge Wilson * Increment the amount of evictable space in the arc_state_t's refcount. 1990dcbf3bd6SGeorge Wilson * We account for the space used by the hdr and the arc buf individually 1991dcbf3bd6SGeorge Wilson * so that we can add and remove them from the refcount individually. 1992dcbf3bd6SGeorge Wilson */ 1993dcbf3bd6SGeorge Wilson static void 1994dcbf3bd6SGeorge Wilson arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) 1995dcbf3bd6SGeorge Wilson { 1996dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 1997244781f1SPrakash Surya 1998dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 19990e8c6158Smaybee 2000dcbf3bd6SGeorge Wilson if (GHOST_STATE(state)) { 2001dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 2002dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2003dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2004*5602294fSDan Kimmel (void) refcount_add_many(&state->arcs_esize[type], 2005*5602294fSDan Kimmel HDR_GET_LSIZE(hdr), hdr); 2006dcbf3bd6SGeorge Wilson return; 2007dcbf3bd6SGeorge Wilson } 2008dcbf3bd6SGeorge Wilson 2009dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(state)); 2010dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_pdata != NULL) { 2011dcbf3bd6SGeorge Wilson (void) refcount_add_many(&state->arcs_esize[type], 2012dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2013dcbf3bd6SGeorge Wilson } 2014dcbf3bd6SGeorge Wilson for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2015dcbf3bd6SGeorge Wilson buf = buf->b_next) { 2016*5602294fSDan Kimmel if (arc_buf_is_shared(buf)) 2017dcbf3bd6SGeorge Wilson continue; 2018*5602294fSDan Kimmel (void) refcount_add_many(&state->arcs_esize[type], 2019*5602294fSDan Kimmel arc_buf_size(buf), buf); 2020fa9e4066Sahrens } 2021fa9e4066Sahrens } 2022fa9e4066Sahrens 2023fa9e4066Sahrens /* 2024dcbf3bd6SGeorge Wilson * Decrement the amount of evictable space in the arc_state_t's refcount. 2025dcbf3bd6SGeorge Wilson * We account for the space used by the hdr and the arc buf individually 2026dcbf3bd6SGeorge Wilson * so that we can add and remove them from the refcount individually. 2027fa9e4066Sahrens */ 2028fa9e4066Sahrens static void 2029*5602294fSDan Kimmel arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) 2030fa9e4066Sahrens { 2031dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 2032dcbf3bd6SGeorge Wilson 2033dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2034dcbf3bd6SGeorge Wilson 2035dcbf3bd6SGeorge Wilson if (GHOST_STATE(state)) { 2036dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 2037dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2038dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2039dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 2040*5602294fSDan Kimmel HDR_GET_LSIZE(hdr), hdr); 2041dcbf3bd6SGeorge Wilson return; 2042dcbf3bd6SGeorge Wilson } 2043dcbf3bd6SGeorge Wilson 2044dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(state)); 2045dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_pdata != NULL) { 2046dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 2047dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2048dcbf3bd6SGeorge Wilson } 2049dcbf3bd6SGeorge Wilson for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2050dcbf3bd6SGeorge Wilson buf = buf->b_next) { 2051*5602294fSDan Kimmel if (arc_buf_is_shared(buf)) 2052dcbf3bd6SGeorge Wilson continue; 2053dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 2054*5602294fSDan Kimmel arc_buf_size(buf), buf); 2055dcbf3bd6SGeorge Wilson } 2056dcbf3bd6SGeorge Wilson } 2057dcbf3bd6SGeorge Wilson 2058dcbf3bd6SGeorge Wilson /* 2059dcbf3bd6SGeorge Wilson * Add a reference to this hdr indicating that someone is actively 2060dcbf3bd6SGeorge Wilson * referencing that memory. When the refcount transitions from 0 to 1, 2061dcbf3bd6SGeorge Wilson * we remove it from the respective arc_state_t list to indicate that 2062dcbf3bd6SGeorge Wilson * it is not evictable. 2063dcbf3bd6SGeorge Wilson */ 2064dcbf3bd6SGeorge Wilson static void 2065dcbf3bd6SGeorge Wilson add_reference(arc_buf_hdr_t *hdr, void *tag) 2066dcbf3bd6SGeorge Wilson { 2067dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2068dcbf3bd6SGeorge Wilson if (!MUTEX_HELD(HDR_LOCK(hdr))) { 2069dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_state == arc_anon); 2070dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2071dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2072dcbf3bd6SGeorge Wilson } 2073dcbf3bd6SGeorge Wilson 2074dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2075dcbf3bd6SGeorge Wilson 2076dcbf3bd6SGeorge Wilson if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 2077dcbf3bd6SGeorge Wilson (state != arc_anon)) { 2078dcbf3bd6SGeorge Wilson /* We don't use the L2-only state list. */ 2079dcbf3bd6SGeorge Wilson if (state != arc_l2c_only) { 2080dcbf3bd6SGeorge Wilson multilist_remove(&state->arcs_list[arc_buf_type(hdr)], 2081dcbf3bd6SGeorge Wilson hdr); 2082*5602294fSDan Kimmel arc_evictable_space_decrement(hdr, state); 2083dcbf3bd6SGeorge Wilson } 2084dcbf3bd6SGeorge Wilson /* remove the prefetch flag if we get a reference */ 2085dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 2086dcbf3bd6SGeorge Wilson } 2087dcbf3bd6SGeorge Wilson } 2088dcbf3bd6SGeorge Wilson 2089dcbf3bd6SGeorge Wilson /* 2090dcbf3bd6SGeorge Wilson * Remove a reference from this hdr. When the reference transitions from 2091dcbf3bd6SGeorge Wilson * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's 2092dcbf3bd6SGeorge Wilson * list making it eligible for eviction. 2093dcbf3bd6SGeorge Wilson */ 2094dcbf3bd6SGeorge Wilson static int 2095dcbf3bd6SGeorge Wilson remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 2096dcbf3bd6SGeorge Wilson { 2097dcbf3bd6SGeorge Wilson int cnt; 2098dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2099dcbf3bd6SGeorge Wilson 2100dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2101dcbf3bd6SGeorge Wilson ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 2102dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(state)); 2103dcbf3bd6SGeorge Wilson 2104dcbf3bd6SGeorge Wilson /* 2105dcbf3bd6SGeorge Wilson * arc_l2c_only counts as a ghost state so we don't need to explicitly 2106dcbf3bd6SGeorge Wilson * check to prevent usage of the arc_l2c_only list. 2107dcbf3bd6SGeorge Wilson */ 2108dcbf3bd6SGeorge Wilson if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 2109dcbf3bd6SGeorge Wilson (state != arc_anon)) { 2110dcbf3bd6SGeorge Wilson multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr); 2111dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 2112dcbf3bd6SGeorge Wilson arc_evictable_space_increment(hdr, state); 2113dcbf3bd6SGeorge Wilson } 2114dcbf3bd6SGeorge Wilson return (cnt); 2115dcbf3bd6SGeorge Wilson } 2116dcbf3bd6SGeorge Wilson 2117dcbf3bd6SGeorge Wilson /* 2118dcbf3bd6SGeorge Wilson * Move the supplied buffer to the indicated state. The hash lock 2119dcbf3bd6SGeorge Wilson * for the buffer must be held by the caller. 2120dcbf3bd6SGeorge Wilson */ 2121dcbf3bd6SGeorge Wilson static void 2122dcbf3bd6SGeorge Wilson arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 2123dcbf3bd6SGeorge Wilson kmutex_t *hash_lock) 2124dcbf3bd6SGeorge Wilson { 2125dcbf3bd6SGeorge Wilson arc_state_t *old_state; 2126dcbf3bd6SGeorge Wilson int64_t refcnt; 2127dcbf3bd6SGeorge Wilson uint32_t bufcnt; 2128dcbf3bd6SGeorge Wilson boolean_t update_old, update_new; 2129dcbf3bd6SGeorge Wilson arc_buf_contents_t buftype = arc_buf_type(hdr); 213089c86e32SChris Williamson 213189c86e32SChris Williamson /* 213289c86e32SChris Williamson * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 213389c86e32SChris Williamson * in arc_read() when bringing a buffer out of the L2ARC. However, the 213489c86e32SChris Williamson * L1 hdr doesn't always exist when we change state to arc_anon before 213589c86e32SChris Williamson * destroying a header, in which case reallocating to add the L1 hdr is 213689c86e32SChris Williamson * pointless. 213789c86e32SChris Williamson */ 213889c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) { 213989c86e32SChris Williamson old_state = hdr->b_l1hdr.b_state; 214089c86e32SChris Williamson refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 2141dcbf3bd6SGeorge Wilson bufcnt = hdr->b_l1hdr.b_bufcnt; 2142dcbf3bd6SGeorge Wilson update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pdata != NULL); 214389c86e32SChris Williamson } else { 214489c86e32SChris Williamson old_state = arc_l2c_only; 214589c86e32SChris Williamson refcnt = 0; 2146dcbf3bd6SGeorge Wilson bufcnt = 0; 2147dcbf3bd6SGeorge Wilson update_old = B_FALSE; 214889c86e32SChris Williamson } 2149dcbf3bd6SGeorge Wilson update_new = update_old; 2150fa9e4066Sahrens 2151fa9e4066Sahrens ASSERT(MUTEX_HELD(hash_lock)); 215269962b56SMatthew Ahrens ASSERT3P(new_state, !=, old_state); 2153dcbf3bd6SGeorge Wilson ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); 2154dcbf3bd6SGeorge Wilson ASSERT(old_state != arc_anon || bufcnt <= 1); 2155fa9e4066Sahrens 2156fa9e4066Sahrens /* 2157fa9e4066Sahrens * If this buffer is evictable, transfer it from the 2158fa9e4066Sahrens * old state list to the new state list. 2159fa9e4066Sahrens */ 2160ea8dc4b6Seschrock if (refcnt == 0) { 216189c86e32SChris Williamson if (old_state != arc_anon && old_state != arc_l2c_only) { 216289c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 2163244781f1SPrakash Surya multilist_remove(&old_state->arcs_list[buftype], hdr); 2164ea8dc4b6Seschrock 2165dcbf3bd6SGeorge Wilson if (GHOST_STATE(old_state)) { 2166dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 2167dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2168dcbf3bd6SGeorge Wilson update_old = B_TRUE; 2169ea8dc4b6Seschrock } 2170*5602294fSDan Kimmel arc_evictable_space_decrement(hdr, old_state); 2171fa9e4066Sahrens } 217289c86e32SChris Williamson if (new_state != arc_anon && new_state != arc_l2c_only) { 2173fa9e4066Sahrens 217489c86e32SChris Williamson /* 217589c86e32SChris Williamson * An L1 header always exists here, since if we're 217689c86e32SChris Williamson * moving to some L1-cached state (i.e. not l2c_only or 217789c86e32SChris Williamson * anonymous), we realloc the header to add an L1hdr 217889c86e32SChris Williamson * beforehand. 217989c86e32SChris Williamson */ 218089c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 2181244781f1SPrakash Surya multilist_insert(&new_state->arcs_list[buftype], hdr); 2182ea8dc4b6Seschrock 2183ea8dc4b6Seschrock if (GHOST_STATE(new_state)) { 2184dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 2185dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2186dcbf3bd6SGeorge Wilson update_new = B_TRUE; 2187ea8dc4b6Seschrock } 2188dcbf3bd6SGeorge Wilson arc_evictable_space_increment(hdr, new_state); 2189fa9e4066Sahrens } 2190fa9e4066Sahrens } 2191fa9e4066Sahrens 2192dcbf3bd6SGeorge Wilson ASSERT(!HDR_EMPTY(hdr)); 21937adb730bSGeorge Wilson if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 21947adb730bSGeorge Wilson buf_hash_remove(hdr); 2195fa9e4066Sahrens 219689c86e32SChris Williamson /* adjust state sizes (ignore arc_l2c_only) */ 21972fd872a7SPrakash Surya 2198dcbf3bd6SGeorge Wilson if (update_new && new_state != arc_l2c_only) { 21992fd872a7SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 22002fd872a7SPrakash Surya if (GHOST_STATE(new_state)) { 2201dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 22022fd872a7SPrakash Surya 22032fd872a7SPrakash Surya /* 2204dcbf3bd6SGeorge Wilson * When moving a header to a ghost state, we first 22052fd872a7SPrakash Surya * remove all arc buffers. Thus, we'll have a 2206dcbf3bd6SGeorge Wilson * bufcnt of zero, and no arc buffer to use for 22072fd872a7SPrakash Surya * the reference. As a result, we use the arc 22082fd872a7SPrakash Surya * header pointer for the reference. 22092fd872a7SPrakash Surya */ 22102fd872a7SPrakash Surya (void) refcount_add_many(&new_state->arcs_size, 2211dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr), hdr); 2212dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 22132fd872a7SPrakash Surya } else { 2214dcbf3bd6SGeorge Wilson uint32_t buffers = 0; 22152fd872a7SPrakash Surya 22162fd872a7SPrakash Surya /* 22172fd872a7SPrakash Surya * Each individual buffer holds a unique reference, 22182fd872a7SPrakash Surya * thus we must remove each of these references one 22192fd872a7SPrakash Surya * at a time. 22202fd872a7SPrakash Surya */ 22212fd872a7SPrakash Surya for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 22222fd872a7SPrakash Surya buf = buf->b_next) { 2223dcbf3bd6SGeorge Wilson ASSERT3U(bufcnt, !=, 0); 2224dcbf3bd6SGeorge Wilson buffers++; 2225dcbf3bd6SGeorge Wilson 2226dcbf3bd6SGeorge Wilson /* 2227dcbf3bd6SGeorge Wilson * When the arc_buf_t is sharing the data 2228dcbf3bd6SGeorge Wilson * block with the hdr, the owner of the 2229dcbf3bd6SGeorge Wilson * reference belongs to the hdr. Only 2230dcbf3bd6SGeorge Wilson * add to the refcount if the arc_buf_t is 2231dcbf3bd6SGeorge Wilson * not shared. 2232dcbf3bd6SGeorge Wilson */ 2233*5602294fSDan Kimmel if (arc_buf_is_shared(buf)) 2234dcbf3bd6SGeorge Wilson continue; 2235dcbf3bd6SGeorge Wilson 22362fd872a7SPrakash Surya (void) refcount_add_many(&new_state->arcs_size, 2237*5602294fSDan Kimmel arc_buf_size(buf), buf); 2238dcbf3bd6SGeorge Wilson } 2239dcbf3bd6SGeorge Wilson ASSERT3U(bufcnt, ==, buffers); 2240dcbf3bd6SGeorge Wilson 2241dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_pdata != NULL) { 2242dcbf3bd6SGeorge Wilson (void) refcount_add_many(&new_state->arcs_size, 2243dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2244dcbf3bd6SGeorge Wilson } else { 2245dcbf3bd6SGeorge Wilson ASSERT(GHOST_STATE(old_state)); 22462fd872a7SPrakash Surya } 22472fd872a7SPrakash Surya } 22482fd872a7SPrakash Surya } 22492fd872a7SPrakash Surya 2250dcbf3bd6SGeorge Wilson if (update_old && old_state != arc_l2c_only) { 22512fd872a7SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 22522fd872a7SPrakash Surya if (GHOST_STATE(old_state)) { 2253dcbf3bd6SGeorge Wilson ASSERT0(bufcnt); 2254*5602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2255dcbf3bd6SGeorge Wilson 22562fd872a7SPrakash Surya /* 22572fd872a7SPrakash Surya * When moving a header off of a ghost state, 2258dcbf3bd6SGeorge Wilson * the header will not contain any arc buffers. 2259dcbf3bd6SGeorge Wilson * We use the arc header pointer for the reference 2260dcbf3bd6SGeorge Wilson * which is exactly what we did when we put the 2261dcbf3bd6SGeorge Wilson * header on the ghost state. 22622fd872a7SPrakash Surya */ 22632fd872a7SPrakash Surya 22642fd872a7SPrakash Surya (void) refcount_remove_many(&old_state->arcs_size, 2265dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr), hdr); 22662fd872a7SPrakash Surya } else { 2267dcbf3bd6SGeorge Wilson uint32_t buffers = 0; 22682fd872a7SPrakash Surya 22692fd872a7SPrakash Surya /* 22702fd872a7SPrakash Surya * Each individual buffer holds a unique reference, 22712fd872a7SPrakash Surya * thus we must remove each of these references one 22722fd872a7SPrakash Surya * at a time. 22732fd872a7SPrakash Surya */ 22742fd872a7SPrakash Surya for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 22752fd872a7SPrakash Surya buf = buf->b_next) { 2276*5602294fSDan Kimmel ASSERT3U(bufcnt, !=, 0); 2277dcbf3bd6SGeorge Wilson buffers++; 2278dcbf3bd6SGeorge Wilson 2279dcbf3bd6SGeorge Wilson /* 2280dcbf3bd6SGeorge Wilson * When the arc_buf_t is sharing the data 2281dcbf3bd6SGeorge Wilson * block with the hdr, the owner of the 2282dcbf3bd6SGeorge Wilson * reference belongs to the hdr. Only 2283dcbf3bd6SGeorge Wilson * add to the refcount if the arc_buf_t is 2284dcbf3bd6SGeorge Wilson * not shared. 2285dcbf3bd6SGeorge Wilson */ 2286*5602294fSDan Kimmel if (arc_buf_is_shared(buf)) 2287dcbf3bd6SGeorge Wilson continue; 2288dcbf3bd6SGeorge Wilson 22892fd872a7SPrakash Surya (void) refcount_remove_many( 2290*5602294fSDan Kimmel &old_state->arcs_size, arc_buf_size(buf), 2291dcbf3bd6SGeorge Wilson buf); 22922fd872a7SPrakash Surya } 2293dcbf3bd6SGeorge Wilson ASSERT3U(bufcnt, ==, buffers); 2294dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2295dcbf3bd6SGeorge Wilson (void) refcount_remove_many( 2296dcbf3bd6SGeorge Wilson &old_state->arcs_size, arc_hdr_size(hdr), hdr); 22972fd872a7SPrakash Surya } 2298fa9e4066Sahrens } 22992fd872a7SPrakash Surya 230089c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) 230189c86e32SChris Williamson hdr->b_l1hdr.b_state = new_state; 2302fa94a07fSbrendan 230389c86e32SChris Williamson /* 230489c86e32SChris Williamson * L2 headers should never be on the L2 state list since they don't 230589c86e32SChris Williamson * have L1 headers allocated. 230689c86e32SChris Williamson */ 2307244781f1SPrakash Surya ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 2308244781f1SPrakash Surya multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 2309fa9e4066Sahrens } 2310fa9e4066Sahrens 23110e8c6158Smaybee void 23125a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(uint64_t space, arc_space_type_t type) 23130e8c6158Smaybee { 23145a98e54bSBrendan Gregg - Sun Microsystems ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 23155a98e54bSBrendan Gregg - Sun Microsystems 23165a98e54bSBrendan Gregg - Sun Microsystems switch (type) { 23175a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_DATA: 23185a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_data_size, space); 23195a98e54bSBrendan Gregg - Sun Microsystems break; 23204076b1bfSPrakash Surya case ARC_SPACE_META: 23214076b1bfSPrakash Surya ARCSTAT_INCR(arcstat_metadata_size, space); 23224076b1bfSPrakash Surya break; 23235a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_OTHER: 23245a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_other_size, space); 23255a98e54bSBrendan Gregg - Sun Microsystems break; 23265a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_HDRS: 23275a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_hdr_size, space); 23285a98e54bSBrendan Gregg - Sun Microsystems break; 23295a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_L2HDRS: 23305a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_l2_hdr_size, space); 23315a98e54bSBrendan Gregg - Sun Microsystems break; 23325a98e54bSBrendan Gregg - Sun Microsystems } 23335a98e54bSBrendan Gregg - Sun Microsystems 23344076b1bfSPrakash Surya if (type != ARC_SPACE_DATA) 23354076b1bfSPrakash Surya ARCSTAT_INCR(arcstat_meta_used, space); 23364076b1bfSPrakash Surya 23370e8c6158Smaybee atomic_add_64(&arc_size, space); 23380e8c6158Smaybee } 23390e8c6158Smaybee 23400e8c6158Smaybee void 23415a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(uint64_t space, arc_space_type_t type) 23420e8c6158Smaybee { 23435a98e54bSBrendan Gregg - Sun Microsystems ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 23445a98e54bSBrendan Gregg - Sun Microsystems 23455a98e54bSBrendan Gregg - Sun Microsystems switch (type) { 23465a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_DATA: 23475a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_data_size, -space); 23485a98e54bSBrendan Gregg - Sun Microsystems break; 23494076b1bfSPrakash Surya case ARC_SPACE_META: 23504076b1bfSPrakash Surya ARCSTAT_INCR(arcstat_metadata_size, -space); 23514076b1bfSPrakash Surya break; 23525a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_OTHER: 23535a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_other_size, -space); 23545a98e54bSBrendan Gregg - Sun Microsystems break; 23555a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_HDRS: 23565a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_hdr_size, -space); 23575a98e54bSBrendan Gregg - Sun Microsystems break; 23585a98e54bSBrendan Gregg - Sun Microsystems case ARC_SPACE_L2HDRS: 23595a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 23605a98e54bSBrendan Gregg - Sun Microsystems break; 23615a98e54bSBrendan Gregg - Sun Microsystems } 23625a98e54bSBrendan Gregg - Sun Microsystems 23634076b1bfSPrakash Surya if (type != ARC_SPACE_DATA) { 23644076b1bfSPrakash Surya ASSERT(arc_meta_used >= space); 23654076b1bfSPrakash Surya if (arc_meta_max < arc_meta_used) 23664076b1bfSPrakash Surya arc_meta_max = arc_meta_used; 23674076b1bfSPrakash Surya ARCSTAT_INCR(arcstat_meta_used, -space); 23684076b1bfSPrakash Surya } 23694076b1bfSPrakash Surya 23700e8c6158Smaybee ASSERT(arc_size >= space); 23710e8c6158Smaybee atomic_add_64(&arc_size, -space); 23720e8c6158Smaybee } 23730e8c6158Smaybee 2374dcbf3bd6SGeorge Wilson /* 2375*5602294fSDan Kimmel * Given a hdr and a buf, returns whether that buf can share its b_data buffer 2376*5602294fSDan Kimmel * with the hdr's b_pdata. 2377dcbf3bd6SGeorge Wilson */ 2378*5602294fSDan Kimmel static boolean_t 2379*5602294fSDan Kimmel arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2380*5602294fSDan Kimmel { 2381*5602294fSDan Kimmel /* 2382*5602294fSDan Kimmel * The criteria for sharing a hdr's data are: 2383*5602294fSDan Kimmel * 1. the hdr's compression matches the buf's compression 2384*5602294fSDan Kimmel * 2. the hdr doesn't need to be byteswapped 2385*5602294fSDan Kimmel * 3. the hdr isn't already being shared 2386*5602294fSDan Kimmel * 4. the buf is either compressed or it is the last buf in the hdr list 2387*5602294fSDan Kimmel * 2388*5602294fSDan Kimmel * Criterion #4 maintains the invariant that shared uncompressed 2389*5602294fSDan Kimmel * bufs must be the final buf in the hdr's b_buf list. Reading this, you 2390*5602294fSDan Kimmel * might ask, "if a compressed buf is allocated first, won't that be the 2391*5602294fSDan Kimmel * last thing in the list?", but in that case it's impossible to create 2392*5602294fSDan Kimmel * a shared uncompressed buf anyway (because the hdr must be compressed 2393*5602294fSDan Kimmel * to have the compressed buf). You might also think that #3 is 2394*5602294fSDan Kimmel * sufficient to make this guarantee, however it's possible 2395*5602294fSDan Kimmel * (specifically in the rare L2ARC write race mentioned in 2396*5602294fSDan Kimmel * arc_buf_alloc_impl()) there will be an existing uncompressed buf that 2397*5602294fSDan Kimmel * is sharable, but wasn't at the time of its allocation. Rather than 2398*5602294fSDan Kimmel * allow a new shared uncompressed buf to be created and then shuffle 2399*5602294fSDan Kimmel * the list around to make it the last element, this simply disallows 2400*5602294fSDan Kimmel * sharing if the new buf isn't the first to be added. 2401*5602294fSDan Kimmel */ 2402*5602294fSDan Kimmel ASSERT3P(buf->b_hdr, ==, hdr); 2403*5602294fSDan Kimmel boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF; 2404*5602294fSDan Kimmel boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0; 2405*5602294fSDan Kimmel return (buf_compressed == hdr_compressed && 2406*5602294fSDan Kimmel hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && 2407*5602294fSDan Kimmel !HDR_SHARED_DATA(hdr) && 2408*5602294fSDan Kimmel (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf))); 2409*5602294fSDan Kimmel } 2410*5602294fSDan Kimmel 2411*5602294fSDan Kimmel /* 2412*5602294fSDan Kimmel * Allocate a buf for this hdr. If you care about the data that's in the hdr, 2413*5602294fSDan Kimmel * or if you want a compressed buffer, pass those flags in. Returns 0 if the 2414*5602294fSDan Kimmel * copy was made successfully, or an error code otherwise. 2415*5602294fSDan Kimmel */ 2416*5602294fSDan Kimmel static int 2417*5602294fSDan Kimmel arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed, 2418*5602294fSDan Kimmel boolean_t fill, arc_buf_t **ret) 2419fa9e4066Sahrens { 2420fa9e4066Sahrens arc_buf_t *buf; 2421fa9e4066Sahrens 2422dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2423dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2424dcbf3bd6SGeorge Wilson VERIFY(hdr->b_type == ARC_BUFC_DATA || 2425dcbf3bd6SGeorge Wilson hdr->b_type == ARC_BUFC_METADATA); 2426*5602294fSDan Kimmel ASSERT3P(ret, !=, NULL); 2427*5602294fSDan Kimmel ASSERT3P(*ret, ==, NULL); 2428dcbf3bd6SGeorge Wilson 2429*5602294fSDan Kimmel buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2430fa9e4066Sahrens buf->b_hdr = hdr; 243144eda4d7Smaybee buf->b_data = NULL; 2432*5602294fSDan Kimmel buf->b_next = hdr->b_l1hdr.b_buf; 2433*5602294fSDan Kimmel buf->b_flags = 0; 243489c86e32SChris Williamson 2435dcbf3bd6SGeorge Wilson add_reference(hdr, tag); 2436dcbf3bd6SGeorge Wilson 2437dcbf3bd6SGeorge Wilson /* 2438dcbf3bd6SGeorge Wilson * We're about to change the hdr's b_flags. We must either 2439dcbf3bd6SGeorge Wilson * hold the hash_lock or be undiscoverable. 2440dcbf3bd6SGeorge Wilson */ 2441dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2442dcbf3bd6SGeorge Wilson 2443dcbf3bd6SGeorge Wilson /* 2444*5602294fSDan Kimmel * Only honor requests for compressed bufs if the hdr is actually 2445*5602294fSDan Kimmel * compressed. 2446*5602294fSDan Kimmel */ 2447*5602294fSDan Kimmel if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 2448*5602294fSDan Kimmel buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; 2449*5602294fSDan Kimmel 2450*5602294fSDan Kimmel /* 2451*5602294fSDan Kimmel * If the hdr's data can be shared then we share the data buffer and 2452*5602294fSDan Kimmel * set the appropriate bit in the hdr's b_flags to indicate the hdr is 2453*5602294fSDan Kimmel * sharing it's b_pdata with the arc_buf_t. Otherwise, we allocate a new 2454*5602294fSDan Kimmel * buffer to store the buf's data. 2455*5602294fSDan Kimmel * 2456*5602294fSDan Kimmel * There is one additional restriction here because we're sharing 2457*5602294fSDan Kimmel * hdr -> buf instead of the usual buf -> hdr: the hdr can't be actively 2458*5602294fSDan Kimmel * involved in an L2ARC write, because if this buf is used by an 2459*5602294fSDan Kimmel * arc_write() then the hdr's data buffer will be released when the 2460*5602294fSDan Kimmel * write completes, even though the L2ARC write might still be using it. 2461dcbf3bd6SGeorge Wilson */ 2462*5602294fSDan Kimmel boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr); 2463*5602294fSDan Kimmel 2464*5602294fSDan Kimmel /* Set up b_data and sharing */ 2465*5602294fSDan Kimmel if (can_share) { 2466dcbf3bd6SGeorge Wilson buf->b_data = hdr->b_l1hdr.b_pdata; 2467*5602294fSDan Kimmel buf->b_flags |= ARC_BUF_FLAG_SHARED; 2468dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2469dcbf3bd6SGeorge Wilson } else { 2470*5602294fSDan Kimmel buf->b_data = 2471*5602294fSDan Kimmel arc_get_data_buf(hdr, arc_buf_size(buf), buf); 2472*5602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2473dcbf3bd6SGeorge Wilson } 2474dcbf3bd6SGeorge Wilson VERIFY3P(buf->b_data, !=, NULL); 247589c86e32SChris Williamson 247689c86e32SChris Williamson hdr->b_l1hdr.b_buf = buf; 2477dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt += 1; 247889c86e32SChris Williamson 2479*5602294fSDan Kimmel /* 2480*5602294fSDan Kimmel * If the user wants the data from the hdr, we need to either copy or 2481*5602294fSDan Kimmel * decompress the data. 2482*5602294fSDan Kimmel */ 2483*5602294fSDan Kimmel if (fill) { 2484*5602294fSDan Kimmel return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0)); 2485*5602294fSDan Kimmel } 2486dcbf3bd6SGeorge Wilson 2487*5602294fSDan Kimmel return (0); 2488*5602294fSDan Kimmel } 2489dcbf3bd6SGeorge Wilson 2490*5602294fSDan Kimmel static char *arc_onloan_tag = "onloan"; 2491fa9e4066Sahrens 2492*5602294fSDan Kimmel static inline void 2493*5602294fSDan Kimmel arc_loaned_bytes_update(int64_t delta) 2494*5602294fSDan Kimmel { 2495*5602294fSDan Kimmel atomic_add_64(&arc_loaned_bytes, delta); 2496dcbf3bd6SGeorge Wilson 2497*5602294fSDan Kimmel /* assert that it did not wrap around */ 2498*5602294fSDan Kimmel ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 2499fa9e4066Sahrens } 2500fa9e4066Sahrens 25012fdbea25SAleksandr Guzovskiy /* 25022fdbea25SAleksandr Guzovskiy * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 25032fdbea25SAleksandr Guzovskiy * flight data by arc_tempreserve_space() until they are "returned". Loaned 25042fdbea25SAleksandr Guzovskiy * buffers must be returned to the arc before they can be used by the DMU or 25052fdbea25SAleksandr Guzovskiy * freed. 25062fdbea25SAleksandr Guzovskiy */ 25072fdbea25SAleksandr Guzovskiy arc_buf_t * 2508*5602294fSDan Kimmel arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size) 25092fdbea25SAleksandr Guzovskiy { 2510*5602294fSDan Kimmel arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag, 2511*5602294fSDan Kimmel is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size); 2512*5602294fSDan Kimmel 2513*5602294fSDan Kimmel arc_loaned_bytes_update(size); 2514*5602294fSDan Kimmel 2515*5602294fSDan Kimmel return (buf); 2516*5602294fSDan Kimmel } 25172fdbea25SAleksandr Guzovskiy 2518*5602294fSDan Kimmel arc_buf_t * 2519*5602294fSDan Kimmel arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize, 2520*5602294fSDan Kimmel enum zio_compress compression_type) 2521*5602294fSDan Kimmel { 2522*5602294fSDan Kimmel arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag, 2523*5602294fSDan Kimmel psize, lsize, compression_type); 2524*5602294fSDan Kimmel 2525*5602294fSDan Kimmel arc_loaned_bytes_update(psize); 25262fdbea25SAleksandr Guzovskiy 25272fdbea25SAleksandr Guzovskiy return (buf); 25282fdbea25SAleksandr Guzovskiy } 25292fdbea25SAleksandr Guzovskiy 2530*5602294fSDan Kimmel 25312fdbea25SAleksandr Guzovskiy /* 25322fdbea25SAleksandr Guzovskiy * Return a loaned arc buffer to the arc. 25332fdbea25SAleksandr Guzovskiy */ 25342fdbea25SAleksandr Guzovskiy void 25352fdbea25SAleksandr Guzovskiy arc_return_buf(arc_buf_t *buf, void *tag) 25362fdbea25SAleksandr Guzovskiy { 25372fdbea25SAleksandr Guzovskiy arc_buf_hdr_t *hdr = buf->b_hdr; 25382fdbea25SAleksandr Guzovskiy 2539dcbf3bd6SGeorge Wilson ASSERT3P(buf->b_data, !=, NULL); 254089c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 254189c86e32SChris Williamson (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 254289c86e32SChris Williamson (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 25432fdbea25SAleksandr Guzovskiy 2544*5602294fSDan Kimmel arc_loaned_bytes_update(-arc_buf_size(buf)); 25452fdbea25SAleksandr Guzovskiy } 25462fdbea25SAleksandr Guzovskiy 2547c242f9a0Schunli zhang - Sun Microsystems - Irvine United States /* Detach an arc_buf from a dbuf (tag) */ 2548c242f9a0Schunli zhang - Sun Microsystems - Irvine United States void 2549c242f9a0Schunli zhang - Sun Microsystems - Irvine United States arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2550c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 255189c86e32SChris Williamson arc_buf_hdr_t *hdr = buf->b_hdr; 2552c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 2553dcbf3bd6SGeorge Wilson ASSERT3P(buf->b_data, !=, NULL); 255489c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 255589c86e32SChris Williamson (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 255689c86e32SChris Williamson (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2557c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 2558*5602294fSDan Kimmel arc_loaned_bytes_update(arc_buf_size(buf)); 2559c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 2560c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 2561dcbf3bd6SGeorge Wilson static void 2562dcbf3bd6SGeorge Wilson l2arc_free_data_on_write(void *data, size_t size, arc_buf_contents_t type) 2563ea8dc4b6Seschrock { 2564dcbf3bd6SGeorge Wilson l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); 2565ea8dc4b6Seschrock 2566dcbf3bd6SGeorge Wilson df->l2df_data = data; 2567dcbf3bd6SGeorge Wilson df->l2df_size = size; 2568dcbf3bd6SGeorge Wilson df->l2df_type = type; 2569dcbf3bd6SGeorge Wilson mutex_enter(&l2arc_free_on_write_mtx); 2570dcbf3bd6SGeorge Wilson list_insert_head(l2arc_free_on_write, df); 2571dcbf3bd6SGeorge Wilson mutex_exit(&l2arc_free_on_write_mtx); 2572dcbf3bd6SGeorge Wilson } 2573b24ab676SJeff Bonwick 2574dcbf3bd6SGeorge Wilson static void 2575dcbf3bd6SGeorge Wilson arc_hdr_free_on_write(arc_buf_hdr_t *hdr) 2576dcbf3bd6SGeorge Wilson { 2577dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2578dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 2579dcbf3bd6SGeorge Wilson uint64_t size = arc_hdr_size(hdr); 2580dcbf3bd6SGeorge Wilson 2581dcbf3bd6SGeorge Wilson /* protected by hash lock, if in the hash table */ 2582dcbf3bd6SGeorge Wilson if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 2583dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2584dcbf3bd6SGeorge Wilson ASSERT(state != arc_anon && state != arc_l2c_only); 2585dcbf3bd6SGeorge Wilson 2586dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 2587dcbf3bd6SGeorge Wilson size, hdr); 2588dcbf3bd6SGeorge Wilson } 2589dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_size, size, hdr); 2590dcbf3bd6SGeorge Wilson 2591dcbf3bd6SGeorge Wilson l2arc_free_data_on_write(hdr->b_l1hdr.b_pdata, size, type); 2592dcbf3bd6SGeorge Wilson } 2593dcbf3bd6SGeorge Wilson 2594dcbf3bd6SGeorge Wilson /* 2595dcbf3bd6SGeorge Wilson * Share the arc_buf_t's data with the hdr. Whenever we are sharing the 2596dcbf3bd6SGeorge Wilson * data buffer, we transfer the refcount ownership to the hdr and update 2597dcbf3bd6SGeorge Wilson * the appropriate kstats. 2598dcbf3bd6SGeorge Wilson */ 2599dcbf3bd6SGeorge Wilson static void 2600dcbf3bd6SGeorge Wilson arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2601dcbf3bd6SGeorge Wilson { 2602dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2603dcbf3bd6SGeorge Wilson 2604*5602294fSDan Kimmel ASSERT(arc_can_share(hdr, buf)); 2605dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2606dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 26079253d63dSGeorge Wilson 26089253d63dSGeorge Wilson /* 2609dcbf3bd6SGeorge Wilson * Start sharing the data buffer. We transfer the 2610dcbf3bd6SGeorge Wilson * refcount ownership to the hdr since it always owns 2611dcbf3bd6SGeorge Wilson * the refcount whenever an arc_buf_t is shared. 26129253d63dSGeorge Wilson */ 2613dcbf3bd6SGeorge Wilson refcount_transfer_ownership(&state->arcs_size, buf, hdr); 2614dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_pdata = buf->b_data; 2615dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2616*5602294fSDan Kimmel buf->b_flags |= ARC_BUF_FLAG_SHARED; 2617dcbf3bd6SGeorge Wilson 2618dcbf3bd6SGeorge Wilson /* 2619dcbf3bd6SGeorge Wilson * Since we've transferred ownership to the hdr we need 2620dcbf3bd6SGeorge Wilson * to increment its compressed and uncompressed kstats and 2621dcbf3bd6SGeorge Wilson * decrement the overhead size. 2622dcbf3bd6SGeorge Wilson */ 2623dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2624dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2625*5602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf)); 2626ea8dc4b6Seschrock } 2627ea8dc4b6Seschrock 2628dcbf3bd6SGeorge Wilson static void 2629dcbf3bd6SGeorge Wilson arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2630ea8dc4b6Seschrock { 2631dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 2632dcbf3bd6SGeorge Wilson 2633dcbf3bd6SGeorge Wilson ASSERT(arc_buf_is_shared(buf)); 2634dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2635dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2636ea8dc4b6Seschrock 26379b23f181Smaybee /* 2638dcbf3bd6SGeorge Wilson * We are no longer sharing this buffer so we need 2639dcbf3bd6SGeorge Wilson * to transfer its ownership to the rightful owner. 26409b23f181Smaybee */ 2641dcbf3bd6SGeorge Wilson refcount_transfer_ownership(&state->arcs_size, hdr, buf); 2642dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2643dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_pdata = NULL; 2644*5602294fSDan Kimmel buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 2645dcbf3bd6SGeorge Wilson 2646dcbf3bd6SGeorge Wilson /* 2647dcbf3bd6SGeorge Wilson * Since the buffer is no longer shared between 2648dcbf3bd6SGeorge Wilson * the arc buf and the hdr, count it as overhead. 2649dcbf3bd6SGeorge Wilson */ 2650dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2651dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2652*5602294fSDan Kimmel ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2653dcbf3bd6SGeorge Wilson } 2654dcbf3bd6SGeorge Wilson 2655dcbf3bd6SGeorge Wilson /* 2656*5602294fSDan Kimmel * Remove an arc_buf_t from the hdr's buf list and return the last 2657*5602294fSDan Kimmel * arc_buf_t on the list. If no buffers remain on the list then return 2658*5602294fSDan Kimmel * NULL. 2659*5602294fSDan Kimmel */ 2660*5602294fSDan Kimmel static arc_buf_t * 2661*5602294fSDan Kimmel arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2662*5602294fSDan Kimmel { 2663*5602294fSDan Kimmel ASSERT(HDR_HAS_L1HDR(hdr)); 2664*5602294fSDan Kimmel ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2665*5602294fSDan Kimmel 2666*5602294fSDan Kimmel arc_buf_t **bufp = &hdr->b_l1hdr.b_buf; 2667*5602294fSDan Kimmel arc_buf_t *lastbuf = NULL; 2668*5602294fSDan Kimmel 2669*5602294fSDan Kimmel /* 2670*5602294fSDan Kimmel * Remove the buf from the hdr list and locate the last 2671*5602294fSDan Kimmel * remaining buffer on the list. 2672*5602294fSDan Kimmel */ 2673*5602294fSDan Kimmel while (*bufp != NULL) { 2674*5602294fSDan Kimmel if (*bufp == buf) 2675*5602294fSDan Kimmel *bufp = buf->b_next; 2676*5602294fSDan Kimmel 2677*5602294fSDan Kimmel /* 2678*5602294fSDan Kimmel * If we've removed a buffer in the middle of 2679*5602294fSDan Kimmel * the list then update the lastbuf and update 2680*5602294fSDan Kimmel * bufp. 2681*5602294fSDan Kimmel */ 2682*5602294fSDan Kimmel if (*bufp != NULL) { 2683*5602294fSDan Kimmel lastbuf = *bufp; 2684*5602294fSDan Kimmel bufp = &(*bufp)->b_next; 2685*5602294fSDan Kimmel } 2686*5602294fSDan Kimmel } 2687*5602294fSDan Kimmel buf->b_next = NULL; 2688*5602294fSDan Kimmel ASSERT3P(lastbuf, !=, buf); 2689*5602294fSDan Kimmel IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL); 2690*5602294fSDan Kimmel IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL); 2691*5602294fSDan Kimmel IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf)); 2692*5602294fSDan Kimmel 2693*5602294fSDan Kimmel return (lastbuf); 2694*5602294fSDan Kimmel } 2695*5602294fSDan Kimmel 2696*5602294fSDan Kimmel /* 2697*5602294fSDan Kimmel * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's 2698*5602294fSDan Kimmel * list and free it. 2699dcbf3bd6SGeorge Wilson */ 2700dcbf3bd6SGeorge Wilson static void 2701*5602294fSDan Kimmel arc_buf_destroy_impl(arc_buf_t *buf) 2702dcbf3bd6SGeorge Wilson { 2703dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = buf->b_hdr; 2704dcbf3bd6SGeorge Wilson 2705dcbf3bd6SGeorge Wilson /* 2706*5602294fSDan Kimmel * Free up the data associated with the buf but only if we're not 2707*5602294fSDan Kimmel * sharing this with the hdr. If we are sharing it with the hdr, the 2708*5602294fSDan Kimmel * hdr is responsible for doing the free. 2709dcbf3bd6SGeorge Wilson */ 2710dcbf3bd6SGeorge Wilson if (buf->b_data != NULL) { 2711dcbf3bd6SGeorge Wilson /* 2712dcbf3bd6SGeorge Wilson * We're about to change the hdr's b_flags. We must either 2713dcbf3bd6SGeorge Wilson * hold the hash_lock or be undiscoverable. 2714dcbf3bd6SGeorge Wilson */ 2715dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2716dcbf3bd6SGeorge Wilson 2717dcbf3bd6SGeorge Wilson arc_cksum_verify(buf); 2718dcbf3bd6SGeorge Wilson arc_buf_unwatch(buf); 2719dcbf3bd6SGeorge Wilson 2720*5602294fSDan Kimmel if (arc_buf_is_shared(buf)) { 2721dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2722dcbf3bd6SGeorge Wilson } else { 2723*5602294fSDan Kimmel uint64_t size = arc_buf_size(buf); 2724dcbf3bd6SGeorge Wilson arc_free_data_buf(hdr, buf->b_data, size, buf); 2725dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_overhead_size, -size); 2726dcbf3bd6SGeorge Wilson } 2727dcbf3bd6SGeorge Wilson buf->b_data = NULL; 2728dcbf3bd6SGeorge Wilson 2729dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 2730dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt -= 1; 2731dcbf3bd6SGeorge Wilson } 2732dcbf3bd6SGeorge Wilson 2733*5602294fSDan Kimmel arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 2734dcbf3bd6SGeorge Wilson 2735*5602294fSDan Kimmel if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) { 2736dcbf3bd6SGeorge Wilson /* 2737*5602294fSDan Kimmel * If the current arc_buf_t is sharing its data buffer with the 2738*5602294fSDan Kimmel * hdr, then reassign the hdr's b_pdata to share it with the new 2739*5602294fSDan Kimmel * buffer at the end of the list. The shared buffer is always 2740*5602294fSDan Kimmel * the last one on the hdr's buffer list. 2741*5602294fSDan Kimmel * 2742*5602294fSDan Kimmel * There is an equivalent case for compressed bufs, but since 2743*5602294fSDan Kimmel * they aren't guaranteed to be the last buf in the list and 2744*5602294fSDan Kimmel * that is an exceedingly rare case, we just allow that space be 2745*5602294fSDan Kimmel * wasted temporarily. 2746dcbf3bd6SGeorge Wilson */ 2747*5602294fSDan Kimmel if (lastbuf != NULL) { 2748*5602294fSDan Kimmel /* Only one buf can be shared at once */ 2749*5602294fSDan Kimmel VERIFY(!arc_buf_is_shared(lastbuf)); 2750*5602294fSDan Kimmel /* hdr is uncompressed so can't have compressed buf */ 2751*5602294fSDan Kimmel VERIFY(!ARC_BUF_COMPRESSED(lastbuf)); 2752dcbf3bd6SGeorge Wilson 2753*5602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2754*5602294fSDan Kimmel arc_hdr_free_pdata(hdr); 2755dcbf3bd6SGeorge Wilson 2756*5602294fSDan Kimmel /* 2757*5602294fSDan Kimmel * We must setup a new shared block between the 2758*5602294fSDan Kimmel * last buffer and the hdr. The data would have 2759*5602294fSDan Kimmel * been allocated by the arc buf so we need to transfer 2760*5602294fSDan Kimmel * ownership to the hdr since it's now being shared. 2761*5602294fSDan Kimmel */ 2762*5602294fSDan Kimmel arc_share_buf(hdr, lastbuf); 2763*5602294fSDan Kimmel } 2764*5602294fSDan Kimmel } else if (HDR_SHARED_DATA(hdr)) { 2765dcbf3bd6SGeorge Wilson /* 2766*5602294fSDan Kimmel * Uncompressed shared buffers are always at the end 2767*5602294fSDan Kimmel * of the list. Compressed buffers don't have the 2768*5602294fSDan Kimmel * same requirements. This makes it hard to 2769*5602294fSDan Kimmel * simply assert that the lastbuf is shared so 2770*5602294fSDan Kimmel * we rely on the hdr's compression flags to determine 2771*5602294fSDan Kimmel * if we have a compressed, shared buffer. 2772dcbf3bd6SGeorge Wilson */ 2773*5602294fSDan Kimmel ASSERT3P(lastbuf, !=, NULL); 2774*5602294fSDan Kimmel ASSERT(arc_buf_is_shared(lastbuf) || 2775*5602294fSDan Kimmel HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 2776dcbf3bd6SGeorge Wilson } 2777dcbf3bd6SGeorge Wilson 2778*5602294fSDan Kimmel /* 2779*5602294fSDan Kimmel * Free the checksum if we're removing the last uncompressed buf from 2780*5602294fSDan Kimmel * this hdr. 2781*5602294fSDan Kimmel */ 2782*5602294fSDan Kimmel if (!arc_hdr_has_uncompressed_buf(hdr)) { 2783dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 2784*5602294fSDan Kimmel } 2785dcbf3bd6SGeorge Wilson 2786dcbf3bd6SGeorge Wilson /* clean up the buf */ 2787dcbf3bd6SGeorge Wilson buf->b_hdr = NULL; 2788dcbf3bd6SGeorge Wilson kmem_cache_free(buf_cache, buf); 2789dcbf3bd6SGeorge Wilson } 2790dcbf3bd6SGeorge Wilson 2791dcbf3bd6SGeorge Wilson static void 2792dcbf3bd6SGeorge Wilson arc_hdr_alloc_pdata(arc_buf_hdr_t *hdr) 2793dcbf3bd6SGeorge Wilson { 2794dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 279589c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 2796dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(hdr)); 2797ea8dc4b6Seschrock 2798dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2799dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_pdata = arc_get_data_buf(hdr, arc_hdr_size(hdr), hdr); 2800dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2801dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 280289c86e32SChris Williamson 2803dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2804dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2805ea8dc4b6Seschrock } 2806ea8dc4b6Seschrock 2807244781f1SPrakash Surya static void 2808dcbf3bd6SGeorge Wilson arc_hdr_free_pdata(arc_buf_hdr_t *hdr) 2809244781f1SPrakash Surya { 2810dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 2811dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2812244781f1SPrakash Surya 2813dcbf3bd6SGeorge Wilson /* 2814dcbf3bd6SGeorge Wilson * If the hdr is currently being written to the l2arc then 2815dcbf3bd6SGeorge Wilson * we defer freeing the data by adding it to the l2arc_free_on_write 2816dcbf3bd6SGeorge Wilson * list. The l2arc will free the data once it's finished 2817dcbf3bd6SGeorge Wilson * writing it to the l2arc device. 2818dcbf3bd6SGeorge Wilson */ 2819dcbf3bd6SGeorge Wilson if (HDR_L2_WRITING(hdr)) { 2820dcbf3bd6SGeorge Wilson arc_hdr_free_on_write(hdr); 2821dcbf3bd6SGeorge Wilson ARCSTAT_BUMP(arcstat_l2_free_on_write); 2822dcbf3bd6SGeorge Wilson } else { 2823dcbf3bd6SGeorge Wilson arc_free_data_buf(hdr, hdr->b_l1hdr.b_pdata, 2824dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 2825dcbf3bd6SGeorge Wilson } 2826dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_pdata = NULL; 2827dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2828dcbf3bd6SGeorge Wilson 2829dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2830dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2831dcbf3bd6SGeorge Wilson } 2832dcbf3bd6SGeorge Wilson 2833dcbf3bd6SGeorge Wilson static arc_buf_hdr_t * 2834dcbf3bd6SGeorge Wilson arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, 2835*5602294fSDan Kimmel enum zio_compress compression_type, arc_buf_contents_t type) 2836dcbf3bd6SGeorge Wilson { 2837dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr; 2838dcbf3bd6SGeorge Wilson 2839dcbf3bd6SGeorge Wilson VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); 2840dcbf3bd6SGeorge Wilson 2841dcbf3bd6SGeorge Wilson hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2842dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 2843dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2844dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL); 2845dcbf3bd6SGeorge Wilson HDR_SET_PSIZE(hdr, psize); 2846dcbf3bd6SGeorge Wilson HDR_SET_LSIZE(hdr, lsize); 2847dcbf3bd6SGeorge Wilson hdr->b_spa = spa; 2848dcbf3bd6SGeorge Wilson hdr->b_type = type; 2849dcbf3bd6SGeorge Wilson hdr->b_flags = 0; 2850dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); 2851*5602294fSDan Kimmel arc_hdr_set_compress(hdr, compression_type); 2852dcbf3bd6SGeorge Wilson 2853dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_state = arc_anon; 2854dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_arc_access = 0; 2855dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt = 0; 2856dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_buf = NULL; 2857dcbf3bd6SGeorge Wilson 2858dcbf3bd6SGeorge Wilson /* 2859dcbf3bd6SGeorge Wilson * Allocate the hdr's buffer. This will contain either 2860dcbf3bd6SGeorge Wilson * the compressed or uncompressed data depending on the block 2861dcbf3bd6SGeorge Wilson * it references and compressed arc enablement. 2862dcbf3bd6SGeorge Wilson */ 2863dcbf3bd6SGeorge Wilson arc_hdr_alloc_pdata(hdr); 2864dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2865dcbf3bd6SGeorge Wilson 2866dcbf3bd6SGeorge Wilson return (hdr); 2867244781f1SPrakash Surya } 2868244781f1SPrakash Surya 2869fa94a07fSbrendan /* 2870dcbf3bd6SGeorge Wilson * Transition between the two allocation states for the arc_buf_hdr struct. 2871dcbf3bd6SGeorge Wilson * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 2872dcbf3bd6SGeorge Wilson * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 2873dcbf3bd6SGeorge Wilson * version is used when a cache buffer is only in the L2ARC in order to reduce 2874dcbf3bd6SGeorge Wilson * memory usage. 2875fa94a07fSbrendan */ 2876dcbf3bd6SGeorge Wilson static arc_buf_hdr_t * 2877dcbf3bd6SGeorge Wilson arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 2878fa94a07fSbrendan { 2879dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L2HDR(hdr)); 2880dcbf3bd6SGeorge Wilson 2881dcbf3bd6SGeorge Wilson arc_buf_hdr_t *nhdr; 2882dcbf3bd6SGeorge Wilson l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2883dcbf3bd6SGeorge Wilson 2884dcbf3bd6SGeorge Wilson ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 2885dcbf3bd6SGeorge Wilson (old == hdr_l2only_cache && new == hdr_full_cache)); 2886dcbf3bd6SGeorge Wilson 2887dcbf3bd6SGeorge Wilson nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 2888dcbf3bd6SGeorge Wilson 2889dcbf3bd6SGeorge Wilson ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 2890dcbf3bd6SGeorge Wilson buf_hash_remove(hdr); 2891dcbf3bd6SGeorge Wilson 2892dcbf3bd6SGeorge Wilson bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 2893dcbf3bd6SGeorge Wilson 2894dcbf3bd6SGeorge Wilson if (new == hdr_full_cache) { 2895dcbf3bd6SGeorge Wilson arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2896dcbf3bd6SGeorge Wilson /* 2897dcbf3bd6SGeorge Wilson * arc_access and arc_change_state need to be aware that a 2898dcbf3bd6SGeorge Wilson * header has just come out of L2ARC, so we set its state to 2899dcbf3bd6SGeorge Wilson * l2c_only even though it's about to change. 2900dcbf3bd6SGeorge Wilson */ 2901dcbf3bd6SGeorge Wilson nhdr->b_l1hdr.b_state = arc_l2c_only; 2902dcbf3bd6SGeorge Wilson 2903dcbf3bd6SGeorge Wilson /* Verify previous threads set to NULL before freeing */ 2904dcbf3bd6SGeorge Wilson ASSERT3P(nhdr->b_l1hdr.b_pdata, ==, NULL); 2905dcbf3bd6SGeorge Wilson } else { 2906dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2907dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 2908dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2909dcbf3bd6SGeorge Wilson 2910dcbf3bd6SGeorge Wilson /* 2911dcbf3bd6SGeorge Wilson * If we've reached here, We must have been called from 2912dcbf3bd6SGeorge Wilson * arc_evict_hdr(), as such we should have already been 2913dcbf3bd6SGeorge Wilson * removed from any ghost list we were previously on 2914dcbf3bd6SGeorge Wilson * (which protects us from racing with arc_evict_state), 2915dcbf3bd6SGeorge Wilson * thus no locking is needed during this check. 2916dcbf3bd6SGeorge Wilson */ 2917dcbf3bd6SGeorge Wilson ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2918cd1c8b85SMatthew Ahrens 2919dcbf3bd6SGeorge Wilson /* 2920dcbf3bd6SGeorge Wilson * A buffer must not be moved into the arc_l2c_only 2921dcbf3bd6SGeorge Wilson * state if it's not finished being written out to the 2922dcbf3bd6SGeorge Wilson * l2arc device. Otherwise, the b_l1hdr.b_pdata field 2923dcbf3bd6SGeorge Wilson * might try to be accessed, even though it was removed. 2924dcbf3bd6SGeorge Wilson */ 2925dcbf3bd6SGeorge Wilson VERIFY(!HDR_L2_WRITING(hdr)); 2926dcbf3bd6SGeorge Wilson VERIFY3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2927fa94a07fSbrendan 2928dcbf3bd6SGeorge Wilson #ifdef ZFS_DEBUG 2929dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_thawed != NULL) { 2930dcbf3bd6SGeorge Wilson kmem_free(hdr->b_l1hdr.b_thawed, 1); 2931dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_thawed = NULL; 2932dcbf3bd6SGeorge Wilson } 2933dcbf3bd6SGeorge Wilson #endif 2934244781f1SPrakash Surya 2935dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2936dcbf3bd6SGeorge Wilson } 2937244781f1SPrakash Surya /* 2938dcbf3bd6SGeorge Wilson * The header has been reallocated so we need to re-insert it into any 2939dcbf3bd6SGeorge Wilson * lists it was on. 2940244781f1SPrakash Surya */ 2941dcbf3bd6SGeorge Wilson (void) buf_hash_insert(nhdr, NULL); 2942244781f1SPrakash Surya 2943dcbf3bd6SGeorge Wilson ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 2944dcbf3bd6SGeorge Wilson 2945dcbf3bd6SGeorge Wilson mutex_enter(&dev->l2ad_mtx); 2946244781f1SPrakash Surya 2947244781f1SPrakash Surya /* 2948dcbf3bd6SGeorge Wilson * We must place the realloc'ed header back into the list at 2949dcbf3bd6SGeorge Wilson * the same spot. Otherwise, if it's placed earlier in the list, 2950dcbf3bd6SGeorge Wilson * l2arc_write_buffers() could find it during the function's 2951dcbf3bd6SGeorge Wilson * write phase, and try to write it out to the l2arc. 2952244781f1SPrakash Surya */ 2953dcbf3bd6SGeorge Wilson list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 2954dcbf3bd6SGeorge Wilson list_remove(&dev->l2ad_buflist, hdr); 2955dcbf3bd6SGeorge Wilson 2956dcbf3bd6SGeorge Wilson mutex_exit(&dev->l2ad_mtx); 2957244781f1SPrakash Surya 2958244781f1SPrakash Surya /* 2959dcbf3bd6SGeorge Wilson * Since we're using the pointer address as the tag when 2960dcbf3bd6SGeorge Wilson * incrementing and decrementing the l2ad_alloc refcount, we 2961dcbf3bd6SGeorge Wilson * must remove the old pointer (that we're about to destroy) and 2962dcbf3bd6SGeorge Wilson * add the new pointer to the refcount. Otherwise we'd remove 2963dcbf3bd6SGeorge Wilson * the wrong pointer address when calling arc_hdr_destroy() later. 2964244781f1SPrakash Surya */ 2965244781f1SPrakash Surya 2966dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); 2967dcbf3bd6SGeorge Wilson (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr); 2968244781f1SPrakash Surya 2969dcbf3bd6SGeorge Wilson buf_discard_identity(hdr); 2970dcbf3bd6SGeorge Wilson kmem_cache_free(old, hdr); 2971244781f1SPrakash Surya 2972dcbf3bd6SGeorge Wilson return (nhdr); 2973244781f1SPrakash Surya } 2974244781f1SPrakash Surya 2975bbfa8ea8SMatthew Ahrens /* 2976dcbf3bd6SGeorge Wilson * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. 2977dcbf3bd6SGeorge Wilson * The buf is returned thawed since we expect the consumer to modify it. 2978bbfa8ea8SMatthew Ahrens */ 2979dcbf3bd6SGeorge Wilson arc_buf_t * 2980*5602294fSDan Kimmel arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size) 2981ea8dc4b6Seschrock { 2982dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, 2983dcbf3bd6SGeorge Wilson ZIO_COMPRESS_OFF, type); 2984dcbf3bd6SGeorge Wilson ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 2985*5602294fSDan Kimmel 2986*5602294fSDan Kimmel arc_buf_t *buf = NULL; 2987*5602294fSDan Kimmel VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf)); 2988dcbf3bd6SGeorge Wilson arc_buf_thaw(buf); 2989*5602294fSDan Kimmel 2990*5602294fSDan Kimmel return (buf); 2991*5602294fSDan Kimmel } 2992*5602294fSDan Kimmel 2993*5602294fSDan Kimmel /* 2994*5602294fSDan Kimmel * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this 2995*5602294fSDan Kimmel * for bufs containing metadata. 2996*5602294fSDan Kimmel */ 2997*5602294fSDan Kimmel arc_buf_t * 2998*5602294fSDan Kimmel arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize, 2999*5602294fSDan Kimmel enum zio_compress compression_type) 3000*5602294fSDan Kimmel { 3001*5602294fSDan Kimmel ASSERT3U(lsize, >, 0); 3002*5602294fSDan Kimmel ASSERT3U(lsize, >=, psize); 3003*5602294fSDan Kimmel ASSERT(compression_type > ZIO_COMPRESS_OFF); 3004*5602294fSDan Kimmel ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS); 3005*5602294fSDan Kimmel 3006*5602294fSDan Kimmel arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 3007*5602294fSDan Kimmel compression_type, ARC_BUFC_DATA); 3008*5602294fSDan Kimmel ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 3009*5602294fSDan Kimmel 3010*5602294fSDan Kimmel arc_buf_t *buf = NULL; 3011*5602294fSDan Kimmel VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf)); 3012*5602294fSDan Kimmel arc_buf_thaw(buf); 3013*5602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3014*5602294fSDan Kimmel 3015dcbf3bd6SGeorge Wilson return (buf); 3016ea8dc4b6Seschrock } 3017ea8dc4b6Seschrock 3018a52fc310SPrakash Surya static void 3019a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 3020a52fc310SPrakash Surya { 3021a52fc310SPrakash Surya l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 3022a52fc310SPrakash Surya l2arc_dev_t *dev = l2hdr->b_dev; 3023dcbf3bd6SGeorge Wilson uint64_t asize = arc_hdr_size(hdr); 3024a52fc310SPrakash Surya 3025a52fc310SPrakash Surya ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 3026a52fc310SPrakash Surya ASSERT(HDR_HAS_L2HDR(hdr)); 3027a52fc310SPrakash Surya 3028a52fc310SPrakash Surya list_remove(&dev->l2ad_buflist, hdr); 3029a52fc310SPrakash Surya 3030dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_l2_asize, -asize); 3031dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr)); 3032a52fc310SPrakash Surya 3033dcbf3bd6SGeorge Wilson vdev_space_update(dev->l2ad_vdev, -asize, 0, 0); 3034a52fc310SPrakash Surya 3035dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&dev->l2ad_alloc, asize, hdr); 3036dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 3037a52fc310SPrakash Surya } 3038a52fc310SPrakash Surya 3039fa9e4066Sahrens static void 3040ea8dc4b6Seschrock arc_hdr_destroy(arc_buf_hdr_t *hdr) 3041fa9e4066Sahrens { 304289c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) { 304389c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_buf == NULL || 3044dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt > 0); 304589c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 304689c86e32SChris Williamson ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 304789c86e32SChris Williamson } 3048ea8dc4b6Seschrock ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 304989c86e32SChris Williamson ASSERT(!HDR_IN_HASH_TABLE(hdr)); 305089c86e32SChris Williamson 3051dcbf3bd6SGeorge Wilson if (!HDR_EMPTY(hdr)) 3052dcbf3bd6SGeorge Wilson buf_discard_identity(hdr); 3053dcbf3bd6SGeorge Wilson 305489c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr)) { 3055a52fc310SPrakash Surya l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3056a52fc310SPrakash Surya boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 3057b24ab676SJeff Bonwick 3058a52fc310SPrakash Surya if (!buflist_held) 3059a52fc310SPrakash Surya mutex_enter(&dev->l2ad_mtx); 306089c86e32SChris Williamson 3061244781f1SPrakash Surya /* 3062a52fc310SPrakash Surya * Even though we checked this conditional above, we 3063a52fc310SPrakash Surya * need to check this again now that we have the 3064a52fc310SPrakash Surya * l2ad_mtx. This is because we could be racing with 3065a52fc310SPrakash Surya * another thread calling l2arc_evict() which might have 3066a52fc310SPrakash Surya * destroyed this header's L2 portion as we were waiting 3067a52fc310SPrakash Surya * to acquire the l2ad_mtx. If that happens, we don't 3068a52fc310SPrakash Surya * want to re-destroy the header's L2 portion. 3069244781f1SPrakash Surya */ 3070a52fc310SPrakash Surya if (HDR_HAS_L2HDR(hdr)) 3071a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(hdr); 3072b24ab676SJeff Bonwick 3073b24ab676SJeff Bonwick if (!buflist_held) 3074a52fc310SPrakash Surya mutex_exit(&dev->l2ad_mtx); 3075fa94a07fSbrendan } 3076fa94a07fSbrendan 3077dcbf3bd6SGeorge Wilson if (HDR_HAS_L1HDR(hdr)) { 3078dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 307989c86e32SChris Williamson 3080dcbf3bd6SGeorge Wilson while (hdr->b_l1hdr.b_buf != NULL) 3081*5602294fSDan Kimmel arc_buf_destroy_impl(hdr->b_l1hdr.b_buf); 308289c86e32SChris Williamson 308389c86e32SChris Williamson #ifdef ZFS_DEBUG 308489c86e32SChris Williamson if (hdr->b_l1hdr.b_thawed != NULL) { 308589c86e32SChris Williamson kmem_free(hdr->b_l1hdr.b_thawed, 1); 308689c86e32SChris Williamson hdr->b_l1hdr.b_thawed = NULL; 308789c86e32SChris Williamson } 308889c86e32SChris Williamson #endif 3089dcbf3bd6SGeorge Wilson 3090dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_pdata != NULL) { 3091dcbf3bd6SGeorge Wilson arc_hdr_free_pdata(hdr); 3092dcbf3bd6SGeorge Wilson } 30933f9d6ad7SLin Ling } 3094ea8dc4b6Seschrock 3095fa9e4066Sahrens ASSERT3P(hdr->b_hash_next, ==, NULL); 309689c86e32SChris Williamson if (HDR_HAS_L1HDR(hdr)) { 3097244781f1SPrakash Surya ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 309889c86e32SChris Williamson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 309989c86e32SChris Williamson kmem_cache_free(hdr_full_cache, hdr); 310089c86e32SChris Williamson } else { 310189c86e32SChris Williamson kmem_cache_free(hdr_l2only_cache, hdr); 310289c86e32SChris Williamson } 3103fa9e4066Sahrens } 3104fa9e4066Sahrens 3105fa9e4066Sahrens void 3106dcbf3bd6SGeorge Wilson arc_buf_destroy(arc_buf_t *buf, void* tag) 3107ea8dc4b6Seschrock { 3108ea8dc4b6Seschrock arc_buf_hdr_t *hdr = buf->b_hdr; 3109ea8dc4b6Seschrock kmutex_t *hash_lock = HDR_LOCK(hdr); 3110fa9e4066Sahrens 311189c86e32SChris Williamson if (hdr->b_l1hdr.b_state == arc_anon) { 3112dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 3113dcbf3bd6SGeorge Wilson ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3114dcbf3bd6SGeorge Wilson VERIFY0(remove_reference(hdr, NULL, tag)); 3115dcbf3bd6SGeorge Wilson arc_hdr_destroy(hdr); 3116dcbf3bd6SGeorge Wilson return; 3117ea8dc4b6Seschrock } 3118ea8dc4b6Seschrock 3119ea8dc4b6Seschrock mutex_enter(hash_lock); 3120dcbf3bd6SGeorge Wilson ASSERT3P(hdr, ==, buf->b_hdr); 3121dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 31223f9d6ad7SLin Ling ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3123dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); 3124dcbf3bd6SGeorge Wilson ASSERT3P(buf->b_data, !=, NULL); 3125ea8dc4b6Seschrock 3126ea8dc4b6Seschrock (void) remove_reference(hdr, hash_lock, tag); 3127*5602294fSDan Kimmel arc_buf_destroy_impl(buf); 3128ea8dc4b6Seschrock mutex_exit(hash_lock); 3129fa9e4066Sahrens } 3130fa9e4066Sahrens 3131fa9e4066Sahrens /* 3132244781f1SPrakash Surya * Evict the arc_buf_hdr that is provided as a parameter. The resultant 3133244781f1SPrakash Surya * state of the header is dependent on it's state prior to entering this 3134244781f1SPrakash Surya * function. The following transitions are possible: 3135874395d5Smaybee * 3136244781f1SPrakash Surya * - arc_mru -> arc_mru_ghost 3137244781f1SPrakash Surya * - arc_mfu -> arc_mfu_ghost 3138244781f1SPrakash Surya * - arc_mru_ghost -> arc_l2c_only 3139244781f1SPrakash Surya * - arc_mru_ghost -> deleted 3140244781f1SPrakash Surya * - arc_mfu_ghost -> arc_l2c_only 3141244781f1SPrakash Surya * - arc_mfu_ghost -> deleted 3142fa9e4066Sahrens */ 3143244781f1SPrakash Surya static int64_t 3144244781f1SPrakash Surya arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3145fa9e4066Sahrens { 3146244781f1SPrakash Surya arc_state_t *evicted_state, *state; 3147244781f1SPrakash Surya int64_t bytes_evicted = 0; 3148fa9e4066Sahrens 3149244781f1SPrakash Surya ASSERT(MUTEX_HELD(hash_lock)); 3150244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 3151fa9e4066Sahrens 3152244781f1SPrakash Surya state = hdr->b_l1hdr.b_state; 3153244781f1SPrakash Surya if (GHOST_STATE(state)) { 3154244781f1SPrakash Surya ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3155dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3156fa9e4066Sahrens 3157244781f1SPrakash Surya /* 3158244781f1SPrakash Surya * l2arc_write_buffers() relies on a header's L1 portion 3159dcbf3bd6SGeorge Wilson * (i.e. its b_pdata field) during its write phase. 3160244781f1SPrakash Surya * Thus, we cannot push a header onto the arc_l2c_only 3161244781f1SPrakash Surya * state (removing it's L1 piece) until the header is 3162244781f1SPrakash Surya * done being written to the l2arc. 3163244781f1SPrakash Surya */ 3164244781f1SPrakash Surya if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 3165244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_l2_skip); 3166244781f1SPrakash Surya return (bytes_evicted); 31673a5286a1SMatthew Ahrens } 3168244781f1SPrakash Surya 3169244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_deleted); 3170dcbf3bd6SGeorge Wilson bytes_evicted += HDR_GET_LSIZE(hdr); 3171244781f1SPrakash Surya 3172244781f1SPrakash Surya DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 3173244781f1SPrakash Surya 3174dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 3175244781f1SPrakash Surya if (HDR_HAS_L2HDR(hdr)) { 3176244781f1SPrakash Surya /* 3177244781f1SPrakash Surya * This buffer is cached on the 2nd Level ARC; 3178244781f1SPrakash Surya * don't destroy the header. 3179244781f1SPrakash Surya */ 3180244781f1SPrakash Surya arc_change_state(arc_l2c_only, hdr, hash_lock); 31813a5286a1SMatthew Ahrens /* 3182244781f1SPrakash Surya * dropping from L1+L2 cached to L2-only, 3183244781f1SPrakash Surya * realloc to remove the L1 header. 31843a5286a1SMatthew Ahrens */ 3185244781f1SPrakash Surya hdr = arc_hdr_realloc(hdr, hdr_full_cache, 3186244781f1SPrakash Surya hdr_l2only_cache); 3187244781f1SPrakash Surya } else { 3188244781f1SPrakash Surya arc_change_state(arc_anon, hdr, hash_lock); 3189244781f1SPrakash Surya arc_hdr_destroy(hdr); 31903a5286a1SMatthew Ahrens } 3191244781f1SPrakash Surya return (bytes_evicted); 31923a5286a1SMatthew Ahrens } 31933a5286a1SMatthew Ahrens 3194244781f1SPrakash Surya ASSERT(state == arc_mru || state == arc_mfu); 3195244781f1SPrakash Surya evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3196244781f1SPrakash Surya 3197244781f1SPrakash Surya /* prefetch buffers have a minimum lifespan */ 3198244781f1SPrakash Surya if (HDR_IO_IN_PROGRESS(hdr) || 3199244781f1SPrakash Surya ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 3200244781f1SPrakash Surya ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 3201244781f1SPrakash Surya arc_min_prefetch_lifespan)) { 3202244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_skip); 3203244781f1SPrakash Surya return (bytes_evicted); 3204244781f1SPrakash Surya } 3205244781f1SPrakash Surya 3206244781f1SPrakash Surya ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3207244781f1SPrakash Surya while (hdr->b_l1hdr.b_buf) { 3208244781f1SPrakash Surya arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3209244781f1SPrakash Surya if (!mutex_tryenter(&buf->b_evict_lock)) { 3210244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_mutex_miss); 3211244781f1SPrakash Surya break; 321213506d1eSmaybee } 3213244781f1SPrakash Surya if (buf->b_data != NULL) 3214dcbf3bd6SGeorge Wilson bytes_evicted += HDR_GET_LSIZE(hdr); 3215dcbf3bd6SGeorge Wilson mutex_exit(&buf->b_evict_lock); 3216*5602294fSDan Kimmel arc_buf_destroy_impl(buf); 3217244781f1SPrakash Surya } 321869962b56SMatthew Ahrens 3219244781f1SPrakash Surya if (HDR_HAS_L2HDR(hdr)) { 3220dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); 3221244781f1SPrakash Surya } else { 3222dcbf3bd6SGeorge Wilson if (l2arc_write_eligible(hdr->b_spa, hdr)) { 3223dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_evict_l2_eligible, 3224dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr)); 3225dcbf3bd6SGeorge Wilson } else { 3226dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_evict_l2_ineligible, 3227dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr)); 3228dcbf3bd6SGeorge Wilson } 3229244781f1SPrakash Surya } 3230244781f1SPrakash Surya 3231dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_bufcnt == 0) { 3232dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 3233dcbf3bd6SGeorge Wilson 3234dcbf3bd6SGeorge Wilson bytes_evicted += arc_hdr_size(hdr); 3235dcbf3bd6SGeorge Wilson 3236dcbf3bd6SGeorge Wilson /* 3237dcbf3bd6SGeorge Wilson * If this hdr is being evicted and has a compressed 3238dcbf3bd6SGeorge Wilson * buffer then we discard it here before we change states. 3239dcbf3bd6SGeorge Wilson * This ensures that the accounting is updated correctly 3240dcbf3bd6SGeorge Wilson * in arc_free_data_buf(). 3241dcbf3bd6SGeorge Wilson */ 3242dcbf3bd6SGeorge Wilson arc_hdr_free_pdata(hdr); 3243dcbf3bd6SGeorge Wilson 3244244781f1SPrakash Surya arc_change_state(evicted_state, hdr, hash_lock); 3245244781f1SPrakash Surya ASSERT(HDR_IN_HASH_TABLE(hdr)); 3246dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 3247244781f1SPrakash Surya DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 3248244781f1SPrakash Surya } 3249244781f1SPrakash Surya 3250244781f1SPrakash Surya return (bytes_evicted); 3251244781f1SPrakash Surya } 3252244781f1SPrakash Surya 3253244781f1SPrakash Surya static uint64_t 3254244781f1SPrakash Surya arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 3255244781f1SPrakash Surya uint64_t spa, int64_t bytes) 3256244781f1SPrakash Surya { 3257244781f1SPrakash Surya multilist_sublist_t *mls; 3258244781f1SPrakash Surya uint64_t bytes_evicted = 0; 3259244781f1SPrakash Surya arc_buf_hdr_t *hdr; 3260244781f1SPrakash Surya kmutex_t *hash_lock; 3261244781f1SPrakash Surya int evict_count = 0; 3262244781f1SPrakash Surya 3263244781f1SPrakash Surya ASSERT3P(marker, !=, NULL); 3264244781f1SPrakash Surya IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3265244781f1SPrakash Surya 3266244781f1SPrakash Surya mls = multilist_sublist_lock(ml, idx); 3267244781f1SPrakash Surya 3268244781f1SPrakash Surya for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 3269244781f1SPrakash Surya hdr = multilist_sublist_prev(mls, marker)) { 3270244781f1SPrakash Surya if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 3271244781f1SPrakash Surya (evict_count >= zfs_arc_evict_batch_limit)) 3272244781f1SPrakash Surya break; 327369962b56SMatthew Ahrens 327469962b56SMatthew Ahrens /* 3275244781f1SPrakash Surya * To keep our iteration location, move the marker 3276244781f1SPrakash Surya * forward. Since we're not holding hdr's hash lock, we 3277244781f1SPrakash Surya * must be very careful and not remove 'hdr' from the 3278244781f1SPrakash Surya * sublist. Otherwise, other consumers might mistake the 3279244781f1SPrakash Surya * 'hdr' as not being on a sublist when they call the 3280244781f1SPrakash Surya * multilist_link_active() function (they all rely on 3281244781f1SPrakash Surya * the hash lock protecting concurrent insertions and 3282244781f1SPrakash Surya * removals). multilist_sublist_move_forward() was 3283244781f1SPrakash Surya * specifically implemented to ensure this is the case 3284244781f1SPrakash Surya * (only 'marker' will be removed and re-inserted). 3285244781f1SPrakash Surya */ 3286244781f1SPrakash Surya multilist_sublist_move_forward(mls, marker); 3287244781f1SPrakash Surya 3288244781f1SPrakash Surya /* 3289244781f1SPrakash Surya * The only case where the b_spa field should ever be 3290244781f1SPrakash Surya * zero, is the marker headers inserted by 3291244781f1SPrakash Surya * arc_evict_state(). It's possible for multiple threads 3292244781f1SPrakash Surya * to be calling arc_evict_state() concurrently (e.g. 3293244781f1SPrakash Surya * dsl_pool_close() and zio_inject_fault()), so we must 3294244781f1SPrakash Surya * skip any markers we see from these other threads. 329569962b56SMatthew Ahrens */ 3296244781f1SPrakash Surya if (hdr->b_spa == 0) 3297244781f1SPrakash Surya continue; 3298244781f1SPrakash Surya 3299244781f1SPrakash Surya /* we're only interested in evicting buffers of a certain spa */ 3300244781f1SPrakash Surya if (spa != 0 && hdr->b_spa != spa) { 3301244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_skip); 330269962b56SMatthew Ahrens continue; 330369962b56SMatthew Ahrens } 330469962b56SMatthew Ahrens 33057adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 33065ea40c06SBrendan Gregg - Sun Microsystems 3307244781f1SPrakash Surya /* 3308244781f1SPrakash Surya * We aren't calling this function from any code path 3309244781f1SPrakash Surya * that would already be holding a hash lock, so we're 3310244781f1SPrakash Surya * asserting on this assumption to be defensive in case 3311244781f1SPrakash Surya * this ever changes. Without this check, it would be 3312244781f1SPrakash Surya * possible to incorrectly increment arcstat_mutex_miss 3313244781f1SPrakash Surya * below (e.g. if the code changed such that we called 3314244781f1SPrakash Surya * this function with a hash lock held). 3315244781f1SPrakash Surya */ 3316244781f1SPrakash Surya ASSERT(!MUTEX_HELD(hash_lock)); 331744cb6abcSbmc 3318244781f1SPrakash Surya if (mutex_tryenter(hash_lock)) { 3319244781f1SPrakash Surya uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 3320244781f1SPrakash Surya mutex_exit(hash_lock); 3321fa9e4066Sahrens 3322244781f1SPrakash Surya bytes_evicted += evicted; 3323fa9e4066Sahrens 3324244781f1SPrakash Surya /* 3325244781f1SPrakash Surya * If evicted is zero, arc_evict_hdr() must have 3326244781f1SPrakash Surya * decided to skip this header, don't increment 3327244781f1SPrakash Surya * evict_count in this case. 3328244781f1SPrakash Surya */ 3329244781f1SPrakash Surya if (evicted != 0) 3330244781f1SPrakash Surya evict_count++; 333144cb6abcSbmc 3332244781f1SPrakash Surya /* 3333244781f1SPrakash Surya * If arc_size isn't overflowing, signal any 3334244781f1SPrakash Surya * threads that might happen to be waiting. 3335244781f1SPrakash Surya * 3336244781f1SPrakash Surya * For each header evicted, we wake up a single 3337244781f1SPrakash Surya * thread. If we used cv_broadcast, we could 3338244781f1SPrakash Surya * wake up "too many" threads causing arc_size 3339244781f1SPrakash Surya * to significantly overflow arc_c; since 3340244781f1SPrakash Surya * arc_get_data_buf() doesn't check for overflow 3341244781f1SPrakash Surya * when it's woken up (it doesn't because it's 3342244781f1SPrakash Surya * possible for the ARC to be overflowing while 3343244781f1SPrakash Surya * full of un-evictable buffers, and the 3344244781f1SPrakash Surya * function should proceed in this case). 3345244781f1SPrakash Surya * 3346244781f1SPrakash Surya * If threads are left sleeping, due to not 3347244781f1SPrakash Surya * using cv_broadcast, they will be woken up 3348244781f1SPrakash Surya * just before arc_reclaim_thread() sleeps. 3349244781f1SPrakash Surya */ 3350244781f1SPrakash Surya mutex_enter(&arc_reclaim_lock); 3351244781f1SPrakash Surya if (!arc_is_overflowing()) 3352244781f1SPrakash Surya cv_signal(&arc_reclaim_waiters_cv); 3353244781f1SPrakash Surya mutex_exit(&arc_reclaim_lock); 3354244781f1SPrakash Surya } else { 3355244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_mutex_miss); 3356244781f1SPrakash Surya } 3357244781f1SPrakash Surya } 3358f4d2e9e6Smaybee 3359244781f1SPrakash Surya multilist_sublist_unlock(mls); 336044cb6abcSbmc 3361244781f1SPrakash Surya return (bytes_evicted); 3362fa9e4066Sahrens } 3363fa9e4066Sahrens 3364fa9e4066Sahrens /* 3365244781f1SPrakash Surya * Evict buffers from the given arc state, until we've removed the 3366244781f1SPrakash Surya * specified number of bytes. Move the removed buffers to the 3367244781f1SPrakash Surya * appropriate evict state. 3368244781f1SPrakash Surya * 3369244781f1SPrakash Surya * This function makes a "best effort". It skips over any buffers 3370244781f1SPrakash Surya * it can't get a hash_lock on, and so, may not catch all candidates. 3371244781f1SPrakash Surya * It may also return without evicting as much space as requested. 3372244781f1SPrakash Surya * 3373244781f1SPrakash Surya * If bytes is specified using the special value ARC_EVICT_ALL, this 3374244781f1SPrakash Surya * will evict all available (i.e. unlocked and evictable) buffers from 3375244781f1SPrakash Surya * the given arc state; which is used by arc_flush(). 3376fa9e4066Sahrens */ 3377244781f1SPrakash Surya static uint64_t 3378244781f1SPrakash Surya arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 3379244781f1SPrakash Surya arc_buf_contents_t type) 3380fa9e4066Sahrens { 3381244781f1SPrakash Surya uint64_t total_evicted = 0; 3382244781f1SPrakash Surya multilist_t *ml = &state->arcs_list[type]; 3383244781f1SPrakash Surya int num_sublists; 3384244781f1SPrakash Surya arc_buf_hdr_t **markers; 3385fa9e4066Sahrens 3386244781f1SPrakash Surya IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3387b802aa8cSSanjeev Bagewadi 3388244781f1SPrakash Surya num_sublists = multilist_get_num_sublists(ml); 3389b802aa8cSSanjeev Bagewadi 3390244781f1SPrakash Surya /* 3391244781f1SPrakash Surya * If we've tried to evict from each sublist, made some 3392244781f1SPrakash Surya * progress, but still have not hit the target number of bytes 3393244781f1SPrakash Surya * to evict, we want to keep trying. The markers allow us to 3394244781f1SPrakash Surya * pick up where we left off for each individual sublist, rather 3395244781f1SPrakash Surya * than starting from the tail each time. 3396244781f1SPrakash Surya */ 3397244781f1SPrakash Surya markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 3398244781f1SPrakash Surya for (int i = 0; i < num_sublists; i++) { 3399244781f1SPrakash Surya markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 340069962b56SMatthew Ahrens 340169962b56SMatthew Ahrens /* 3402244781f1SPrakash Surya * A b_spa of 0 is used to indicate that this header is 3403244781f1SPrakash Surya * a marker. This fact is used in arc_adjust_type() and 3404244781f1SPrakash Surya * arc_evict_state_impl(). 340569962b56SMatthew Ahrens */ 3406244781f1SPrakash Surya markers[i]->b_spa = 0; 3407fa94a07fSbrendan 3408244781f1SPrakash Surya multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3409244781f1SPrakash Surya multilist_sublist_insert_tail(mls, markers[i]); 3410244781f1SPrakash Surya multilist_sublist_unlock(mls); 3411244781f1SPrakash Surya } 3412fa94a07fSbrendan 3413244781f1SPrakash Surya /* 3414244781f1SPrakash Surya * While we haven't hit our target number of bytes to evict, or 3415244781f1SPrakash Surya * we're evicting all available buffers. 3416244781f1SPrakash Surya */ 3417244781f1SPrakash Surya while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 3418244781f1SPrakash Surya /* 3419244781f1SPrakash Surya * Start eviction using a randomly selected sublist, 3420244781f1SPrakash Surya * this is to try and evenly balance eviction across all 3421244781f1SPrakash Surya * sublists. Always starting at the same sublist 3422244781f1SPrakash Surya * (e.g. index 0) would cause evictions to favor certain 3423244781f1SPrakash Surya * sublists over others. 3424244781f1SPrakash Surya */ 3425244781f1SPrakash Surya int sublist_idx = multilist_get_random_index(ml); 3426244781f1SPrakash Surya uint64_t scan_evicted = 0; 3427244781f1SPrakash Surya 3428244781f1SPrakash Surya for (int i = 0; i < num_sublists; i++) { 3429244781f1SPrakash Surya uint64_t bytes_remaining; 3430244781f1SPrakash Surya uint64_t bytes_evicted; 3431244781f1SPrakash Surya 3432244781f1SPrakash Surya if (bytes == ARC_EVICT_ALL) 3433244781f1SPrakash Surya bytes_remaining = ARC_EVICT_ALL; 3434244781f1SPrakash Surya else if (total_evicted < bytes) 3435244781f1SPrakash Surya bytes_remaining = bytes - total_evicted; 3436244781f1SPrakash Surya else 3437fa9e4066Sahrens break; 3438244781f1SPrakash Surya 3439244781f1SPrakash Surya bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 3440244781f1SPrakash Surya markers[sublist_idx], spa, bytes_remaining); 3441244781f1SPrakash Surya 3442244781f1SPrakash Surya scan_evicted += bytes_evicted; 3443244781f1SPrakash Surya total_evicted += bytes_evicted; 3444244781f1SPrakash Surya 3445244781f1SPrakash Surya /* we've reached the end, wrap to the beginning */ 3446244781f1SPrakash Surya if (++sublist_idx >= num_sublists) 3447244781f1SPrakash Surya sublist_idx = 0; 3448244781f1SPrakash Surya } 3449244781f1SPrakash Surya 3450244781f1SPrakash Surya /* 3451244781f1SPrakash Surya * If we didn't evict anything during this scan, we have 3452244781f1SPrakash Surya * no reason to believe we'll evict more during another 3453244781f1SPrakash Surya * scan, so break the loop. 3454244781f1SPrakash Surya */ 3455244781f1SPrakash Surya if (scan_evicted == 0) { 3456244781f1SPrakash Surya /* This isn't possible, let's make that obvious */ 3457244781f1SPrakash Surya ASSERT3S(bytes, !=, 0); 3458244781f1SPrakash Surya 3459b802aa8cSSanjeev Bagewadi /* 3460244781f1SPrakash Surya * When bytes is ARC_EVICT_ALL, the only way to 3461244781f1SPrakash Surya * break the loop is when scan_evicted is zero. 3462244781f1SPrakash Surya * In that case, we actually have evicted enough, 3463244781f1SPrakash Surya * so we don't want to increment the kstat. 3464b802aa8cSSanjeev Bagewadi */ 3465244781f1SPrakash Surya if (bytes != ARC_EVICT_ALL) { 3466244781f1SPrakash Surya ASSERT3S(total_evicted, <, bytes); 3467244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_evict_not_enough); 3468244781f1SPrakash Surya } 3469244781f1SPrakash Surya 3470244781f1SPrakash Surya break; 347169962b56SMatthew Ahrens } 3472244781f1SPrakash Surya } 3473244781f1SPrakash Surya 3474244781f1SPrakash Surya for (int i = 0; i < num_sublists; i++) { 3475244781f1SPrakash Surya multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3476244781f1SPrakash Surya multilist_sublist_remove(mls, markers[i]); 3477244781f1SPrakash Surya multilist_sublist_unlock(mls); 347869962b56SMatthew Ahrens 3479244781f1SPrakash Surya kmem_cache_free(hdr_full_cache, markers[i]); 3480fa9e4066Sahrens } 3481244781f1SPrakash Surya kmem_free(markers, sizeof (*markers) * num_sublists); 3482fa9e4066Sahrens 3483244781f1SPrakash Surya return (total_evicted); 3484244781f1SPrakash Surya } 3485244781f1SPrakash Surya 3486244781f1SPrakash Surya /* 3487244781f1SPrakash Surya * Flush all "evictable" data of the given type from the arc state 3488244781f1SPrakash Surya * specified. This will not evict any "active" buffers (i.e. referenced). 3489244781f1SPrakash Surya * 3490dcbf3bd6SGeorge Wilson * When 'retry' is set to B_FALSE, the function will make a single pass 3491244781f1SPrakash Surya * over the state and evict any buffers that it can. Since it doesn't 3492244781f1SPrakash Surya * continually retry the eviction, it might end up leaving some buffers 3493244781f1SPrakash Surya * in the ARC due to lock misses. 3494244781f1SPrakash Surya * 3495dcbf3bd6SGeorge Wilson * When 'retry' is set to B_TRUE, the function will continually retry the 3496244781f1SPrakash Surya * eviction until *all* evictable buffers have been removed from the 3497244781f1SPrakash Surya * state. As a result, if concurrent insertions into the state are 3498244781f1SPrakash Surya * allowed (e.g. if the ARC isn't shutting down), this function might 3499244781f1SPrakash Surya * wind up in an infinite loop, continually trying to evict buffers. 3500244781f1SPrakash Surya */ 3501244781f1SPrakash Surya static uint64_t 3502244781f1SPrakash Surya arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 3503244781f1SPrakash Surya boolean_t retry) 3504244781f1SPrakash Surya { 3505244781f1SPrakash Surya uint64_t evicted = 0; 3506244781f1SPrakash Surya 3507dcbf3bd6SGeorge Wilson while (refcount_count(&state->arcs_esize[type]) != 0) { 3508244781f1SPrakash Surya evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 3509244781f1SPrakash Surya 3510244781f1SPrakash Surya if (!retry) 3511244781f1SPrakash Surya break; 35120e8c6158Smaybee } 35130e8c6158Smaybee 3514244781f1SPrakash Surya return (evicted); 3515244781f1SPrakash Surya } 3516244781f1SPrakash Surya 3517244781f1SPrakash Surya /* 3518244781f1SPrakash Surya * Evict the specified number of bytes from the state specified, 3519244781f1SPrakash Surya * restricting eviction to the spa and type given. This function 3520244781f1SPrakash Surya * prevents us from trying to evict more from a state's list than 3521244781f1SPrakash Surya * is "evictable", and to skip evicting altogether when passed a 3522244781f1SPrakash Surya * negative value for "bytes". In contrast, arc_evict_state() will 3523244781f1SPrakash Surya * evict everything it can, when passed a negative value for "bytes". 3524244781f1SPrakash Surya */ 3525244781f1SPrakash Surya static uint64_t 3526244781f1SPrakash Surya arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3527244781f1SPrakash Surya arc_buf_contents_t type) 3528244781f1SPrakash Surya { 3529244781f1SPrakash Surya int64_t delta; 3530244781f1SPrakash Surya 3531dcbf3bd6SGeorge Wilson if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) { 3532dcbf3bd6SGeorge Wilson delta = MIN(refcount_count(&state->arcs_esize[type]), bytes); 3533244781f1SPrakash Surya return (arc_evict_state(state, spa, delta, type)); 3534fa9e4066Sahrens } 3535fa9e4066Sahrens 3536244781f1SPrakash Surya return (0); 3537fa9e4066Sahrens } 3538fa9e4066Sahrens 3539244781f1SPrakash Surya /* 3540244781f1SPrakash Surya * Evict metadata buffers from the cache, such that arc_meta_used is 3541244781f1SPrakash Surya * capped by the arc_meta_limit tunable. 3542244781f1SPrakash Surya */ 3543244781f1SPrakash Surya static uint64_t 3544244781f1SPrakash Surya arc_adjust_meta(void) 3545244781f1SPrakash Surya { 3546244781f1SPrakash Surya uint64_t total_evicted = 0; 3547244781f1SPrakash Surya int64_t target; 3548244781f1SPrakash Surya 3549244781f1SPrakash Surya /* 3550244781f1SPrakash Surya * If we're over the meta limit, we want to evict enough 3551244781f1SPrakash Surya * metadata to get back under the meta limit. We don't want to 3552244781f1SPrakash Surya * evict so much that we drop the MRU below arc_p, though. If 3553244781f1SPrakash Surya * we're over the meta limit more than we're over arc_p, we 3554244781f1SPrakash Surya * evict some from the MRU here, and some from the MFU below. 3555244781f1SPrakash Surya */ 3556244781f1SPrakash Surya target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 35572fd872a7SPrakash Surya (int64_t)(refcount_count(&arc_anon->arcs_size) + 35582fd872a7SPrakash Surya refcount_count(&arc_mru->arcs_size) - arc_p)); 3559244781f1SPrakash Surya 3560244781f1SPrakash Surya total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3561244781f1SPrakash Surya 3562244781f1SPrakash Surya /* 3563244781f1SPrakash Surya * Similar to the above, we want to evict enough bytes to get us 3564244781f1SPrakash Surya * below the meta limit, but not so much as to drop us below the 3565*5602294fSDan Kimmel * space allotted to the MFU (which is defined as arc_c - arc_p). 3566244781f1SPrakash Surya */ 3567244781f1SPrakash Surya target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 35682fd872a7SPrakash Surya (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3569244781f1SPrakash Surya 3570244781f1SPrakash Surya total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3571244781f1SPrakash Surya 3572244781f1SPrakash Surya return (total_evicted); 3573244781f1SPrakash Surya } 3574244781f1SPrakash Surya 3575244781f1SPrakash Surya /* 3576244781f1SPrakash Surya * Return the type of the oldest buffer in the given arc state 3577244781f1SPrakash Surya * 3578244781f1SPrakash Surya * This function will select a random sublist of type ARC_BUFC_DATA and 3579244781f1SPrakash Surya * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3580244781f1SPrakash Surya * is compared, and the type which contains the "older" buffer will be 3581244781f1SPrakash Surya * returned. 3582244781f1SPrakash Surya */ 3583244781f1SPrakash Surya static arc_buf_contents_t 3584244781f1SPrakash Surya arc_adjust_type(arc_state_t *state) 3585244781f1SPrakash Surya { 3586244781f1SPrakash Surya multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; 3587244781f1SPrakash Surya multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; 3588244781f1SPrakash Surya int data_idx = multilist_get_random_index(data_ml); 3589244781f1SPrakash Surya int meta_idx = multilist_get_random_index(meta_ml); 3590244781f1SPrakash Surya multilist_sublist_t *data_mls; 3591244781f1SPrakash Surya multilist_sublist_t *meta_mls; 3592244781f1SPrakash Surya arc_buf_contents_t type; 3593244781f1SPrakash Surya arc_buf_hdr_t *data_hdr; 3594244781f1SPrakash Surya arc_buf_hdr_t *meta_hdr; 3595244781f1SPrakash Surya 3596244781f1SPrakash Surya /* 3597244781f1SPrakash Surya * We keep the sublist lock until we're finished, to prevent 3598244781f1SPrakash Surya * the headers from being destroyed via arc_evict_state(). 3599244781f1SPrakash Surya */ 3600244781f1SPrakash Surya data_mls = multilist_sublist_lock(data_ml, data_idx); 3601244781f1SPrakash Surya meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3602244781f1SPrakash Surya 3603244781f1SPrakash Surya /* 3604244781f1SPrakash Surya * These two loops are to ensure we skip any markers that 3605244781f1SPrakash Surya * might be at the tail of the lists due to arc_evict_state(). 3606244781f1SPrakash Surya */ 3607244781f1SPrakash Surya 3608244781f1SPrakash Surya for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3609244781f1SPrakash Surya data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3610244781f1SPrakash Surya if (data_hdr->b_spa != 0) 3611244781f1SPrakash Surya break; 3612244781f1SPrakash Surya } 3613244781f1SPrakash Surya 3614244781f1SPrakash Surya for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3615244781f1SPrakash Surya meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3616244781f1SPrakash Surya if (meta_hdr->b_spa != 0) 3617244781f1SPrakash Surya break; 3618244781f1SPrakash Surya } 3619244781f1SPrakash Surya 3620244781f1SPrakash Surya if (data_hdr == NULL && meta_hdr == NULL) { 3621244781f1SPrakash Surya type = ARC_BUFC_DATA; 3622244781f1SPrakash Surya } else if (data_hdr == NULL) { 3623244781f1SPrakash Surya ASSERT3P(meta_hdr, !=, NULL); 3624244781f1SPrakash Surya type = ARC_BUFC_METADATA; 3625244781f1SPrakash Surya } else if (meta_hdr == NULL) { 3626244781f1SPrakash Surya ASSERT3P(data_hdr, !=, NULL); 3627244781f1SPrakash Surya type = ARC_BUFC_DATA; 3628244781f1SPrakash Surya } else { 3629244781f1SPrakash Surya ASSERT3P(data_hdr, !=, NULL); 3630244781f1SPrakash Surya ASSERT3P(meta_hdr, !=, NULL); 3631244781f1SPrakash Surya 3632244781f1SPrakash Surya /* The headers can't be on the sublist without an L1 header */ 3633244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(data_hdr)); 3634244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3635244781f1SPrakash Surya 3636244781f1SPrakash Surya if (data_hdr->b_l1hdr.b_arc_access < 3637244781f1SPrakash Surya meta_hdr->b_l1hdr.b_arc_access) { 3638244781f1SPrakash Surya type = ARC_BUFC_DATA; 3639244781f1SPrakash Surya } else { 3640244781f1SPrakash Surya type = ARC_BUFC_METADATA; 3641244781f1SPrakash Surya } 3642244781f1SPrakash Surya } 3643244781f1SPrakash Surya 3644244781f1SPrakash Surya multilist_sublist_unlock(meta_mls); 3645244781f1SPrakash Surya multilist_sublist_unlock(data_mls); 3646244781f1SPrakash Surya 3647244781f1SPrakash Surya return (type); 3648244781f1SPrakash Surya } 3649244781f1SPrakash Surya 3650244781f1SPrakash Surya /* 3651244781f1SPrakash Surya * Evict buffers from the cache, such that arc_size is capped by arc_c. 3652244781f1SPrakash Surya */ 3653244781f1SPrakash Surya static uint64_t 3654fa9e4066Sahrens arc_adjust(void) 3655fa9e4066Sahrens { 3656244781f1SPrakash Surya uint64_t total_evicted = 0; 3657244781f1SPrakash Surya uint64_t bytes; 3658244781f1SPrakash Surya int64_t target; 3659fa9e4066Sahrens 36605a98e54bSBrendan Gregg - Sun Microsystems /* 3661244781f1SPrakash Surya * If we're over arc_meta_limit, we want to correct that before 3662244781f1SPrakash Surya * potentially evicting data buffers below. 36635a98e54bSBrendan Gregg - Sun Microsystems */ 3664244781f1SPrakash Surya total_evicted += arc_adjust_meta(); 36655a98e54bSBrendan Gregg - Sun Microsystems 3666244781f1SPrakash Surya /* 3667244781f1SPrakash Surya * Adjust MRU size 3668244781f1SPrakash Surya * 3669244781f1SPrakash Surya * If we're over the target cache size, we want to evict enough 3670244781f1SPrakash Surya * from the list to get back to our target size. We don't want 3671244781f1SPrakash Surya * to evict too much from the MRU, such that it drops below 3672244781f1SPrakash Surya * arc_p. So, if we're over our target cache size more than 3673244781f1SPrakash Surya * the MRU is over arc_p, we'll evict enough to get back to 3674244781f1SPrakash Surya * arc_p here, and then evict more from the MFU below. 3675244781f1SPrakash Surya */ 3676244781f1SPrakash Surya target = MIN((int64_t)(arc_size - arc_c), 36772fd872a7SPrakash Surya (int64_t)(refcount_count(&arc_anon->arcs_size) + 36782fd872a7SPrakash Surya refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 3679fa9e4066Sahrens 3680244781f1SPrakash Surya /* 3681244781f1SPrakash Surya * If we're below arc_meta_min, always prefer to evict data. 3682244781f1SPrakash Surya * Otherwise, try to satisfy the requested number of bytes to 3683244781f1SPrakash Surya * evict from the type which contains older buffers; in an 3684244781f1SPrakash Surya * effort to keep newer buffers in the cache regardless of their 3685244781f1SPrakash Surya * type. If we cannot satisfy the number of bytes from this 3686244781f1SPrakash Surya * type, spill over into the next type. 3687244781f1SPrakash Surya */ 3688244781f1SPrakash Surya if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3689244781f1SPrakash Surya arc_meta_used > arc_meta_min) { 3690244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3691244781f1SPrakash Surya total_evicted += bytes; 36920e8c6158Smaybee 3693244781f1SPrakash Surya /* 3694244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3695244781f1SPrakash Surya * metadata, we try to get the rest from data. 3696244781f1SPrakash Surya */ 3697244781f1SPrakash Surya target -= bytes; 3698244781f1SPrakash Surya 3699244781f1SPrakash Surya total_evicted += 3700244781f1SPrakash Surya arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3701244781f1SPrakash Surya } else { 3702244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3703244781f1SPrakash Surya total_evicted += bytes; 3704244781f1SPrakash Surya 3705244781f1SPrakash Surya /* 3706244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3707244781f1SPrakash Surya * data, we try to get the rest from metadata. 3708244781f1SPrakash Surya */ 3709244781f1SPrakash Surya target -= bytes; 3710244781f1SPrakash Surya 3711244781f1SPrakash Surya total_evicted += 3712244781f1SPrakash Surya arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3713fa9e4066Sahrens } 3714fa9e4066Sahrens 37155a98e54bSBrendan Gregg - Sun Microsystems /* 37165a98e54bSBrendan Gregg - Sun Microsystems * Adjust MFU size 3717244781f1SPrakash Surya * 3718244781f1SPrakash Surya * Now that we've tried to evict enough from the MRU to get its 3719244781f1SPrakash Surya * size back to arc_p, if we're still above the target cache 3720244781f1SPrakash Surya * size, we evict the rest from the MFU. 37215a98e54bSBrendan Gregg - Sun Microsystems */ 3722244781f1SPrakash Surya target = arc_size - arc_c; 3723fa9e4066Sahrens 372431c46cf2SAlek Pinchuk if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3725244781f1SPrakash Surya arc_meta_used > arc_meta_min) { 3726244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3727244781f1SPrakash Surya total_evicted += bytes; 37285a98e54bSBrendan Gregg - Sun Microsystems 3729244781f1SPrakash Surya /* 3730244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3731244781f1SPrakash Surya * metadata, we try to get the rest from data. 3732244781f1SPrakash Surya */ 3733244781f1SPrakash Surya target -= bytes; 3734244781f1SPrakash Surya 3735244781f1SPrakash Surya total_evicted += 3736244781f1SPrakash Surya arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3737244781f1SPrakash Surya } else { 3738244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3739244781f1SPrakash Surya total_evicted += bytes; 3740244781f1SPrakash Surya 3741244781f1SPrakash Surya /* 3742244781f1SPrakash Surya * If we couldn't evict our target number of bytes from 3743244781f1SPrakash Surya * data, we try to get the rest from data. 3744244781f1SPrakash Surya */ 3745244781f1SPrakash Surya target -= bytes; 3746fa9e4066Sahrens 3747244781f1SPrakash Surya total_evicted += 3748244781f1SPrakash Surya arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 37495a98e54bSBrendan Gregg - Sun Microsystems } 3750fa9e4066Sahrens 37515a98e54bSBrendan Gregg - Sun Microsystems /* 37525a98e54bSBrendan Gregg - Sun Microsystems * Adjust ghost lists 3753244781f1SPrakash Surya * 3754244781f1SPrakash Surya * In addition to the above, the ARC also defines target values 3755244781f1SPrakash Surya * for the ghost lists. The sum of the mru list and mru ghost 3756244781f1SPrakash Surya * list should never exceed the target size of the cache, and 3757244781f1SPrakash Surya * the sum of the mru list, mfu list, mru ghost list, and mfu 3758244781f1SPrakash Surya * ghost list should never exceed twice the target size of the 3759244781f1SPrakash Surya * cache. The following logic enforces these limits on the ghost 3760244781f1SPrakash Surya * caches, and evicts from them as needed. 37615a98e54bSBrendan Gregg - Sun Microsystems */ 37622fd872a7SPrakash Surya target = refcount_count(&arc_mru->arcs_size) + 37632fd872a7SPrakash Surya refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3764fa9e4066Sahrens 3765244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3766244781f1SPrakash Surya total_evicted += bytes; 3767fa9e4066Sahrens 3768244781f1SPrakash Surya target -= bytes; 37690e8c6158Smaybee 3770244781f1SPrakash Surya total_evicted += 3771244781f1SPrakash Surya arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 37725a98e54bSBrendan Gregg - Sun Microsystems 3773244781f1SPrakash Surya /* 3774244781f1SPrakash Surya * We assume the sum of the mru list and mfu list is less than 3775244781f1SPrakash Surya * or equal to arc_c (we enforced this above), which means we 3776244781f1SPrakash Surya * can use the simpler of the two equations below: 3777244781f1SPrakash Surya * 3778244781f1SPrakash Surya * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3779244781f1SPrakash Surya * mru ghost + mfu ghost <= arc_c 3780244781f1SPrakash Surya */ 37812fd872a7SPrakash Surya target = refcount_count(&arc_mru_ghost->arcs_size) + 37822fd872a7SPrakash Surya refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3783244781f1SPrakash Surya 3784244781f1SPrakash Surya bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3785244781f1SPrakash Surya total_evicted += bytes; 3786244781f1SPrakash Surya 3787244781f1SPrakash Surya target -= bytes; 3788244781f1SPrakash Surya 3789244781f1SPrakash Surya total_evicted += 3790244781f1SPrakash Surya arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3791244781f1SPrakash Surya 3792244781f1SPrakash Surya return (total_evicted); 3793fa9e4066Sahrens } 3794fa9e4066Sahrens 3795fa9e4066Sahrens void 3796244781f1SPrakash Surya arc_flush(spa_t *spa, boolean_t retry) 3797fa9e4066Sahrens { 3798ac05c741SMark Maybee uint64_t guid = 0; 3799ac05c741SMark Maybee 3800244781f1SPrakash Surya /* 3801dcbf3bd6SGeorge Wilson * If retry is B_TRUE, a spa must not be specified since we have 3802244781f1SPrakash Surya * no good way to determine if all of a spa's buffers have been 3803244781f1SPrakash Surya * evicted from an arc state. 3804244781f1SPrakash Surya */ 3805244781f1SPrakash Surya ASSERT(!retry || spa == 0); 3806244781f1SPrakash Surya 380789c86e32SChris Williamson if (spa != NULL) 3808e9103aaeSGarrett D'Amore guid = spa_load_guid(spa); 3809ac05c741SMark Maybee 3810244781f1SPrakash Surya (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3811244781f1SPrakash Surya (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3812244781f1SPrakash Surya 3813244781f1SPrakash Surya (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3814244781f1SPrakash Surya (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3815874395d5Smaybee 3816244781f1SPrakash Surya (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3817244781f1SPrakash Surya (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3818244781f1SPrakash Surya 3819244781f1SPrakash Surya (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3820244781f1SPrakash Surya (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3821fa9e4066Sahrens } 3822fa9e4066Sahrens 3823fa9e4066Sahrens void 38242ec99e3eSMatthew Ahrens arc_shrink(int64_t to_free) 3825fa9e4066Sahrens { 382644cb6abcSbmc if (arc_c > arc_c_min) { 3827fa9e4066Sahrens 382844cb6abcSbmc if (arc_c > arc_c_min + to_free) 382944cb6abcSbmc atomic_add_64(&arc_c, -to_free); 383049e3519aSmaybee else 383144cb6abcSbmc arc_c = arc_c_min; 383244cb6abcSbmc 383344cb6abcSbmc atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 383444cb6abcSbmc if (arc_c > arc_size) 383544cb6abcSbmc arc_c = MAX(arc_size, arc_c_min); 383644cb6abcSbmc if (arc_p > arc_c) 383744cb6abcSbmc arc_p = (arc_c >> 1); 383844cb6abcSbmc ASSERT(arc_c >= arc_c_min); 383944cb6abcSbmc ASSERT((int64_t)arc_p >= 0); 384049e3519aSmaybee } 3841fa9e4066Sahrens 384244cb6abcSbmc if (arc_size > arc_c) 3843244781f1SPrakash Surya (void) arc_adjust(); 3844fa9e4066Sahrens } 3845fa9e4066Sahrens 38462ec99e3eSMatthew Ahrens typedef enum free_memory_reason_t { 38472ec99e3eSMatthew Ahrens FMR_UNKNOWN, 38482ec99e3eSMatthew Ahrens FMR_NEEDFREE, 38492ec99e3eSMatthew Ahrens FMR_LOTSFREE, 38502ec99e3eSMatthew Ahrens FMR_SWAPFS_MINFREE, 38512ec99e3eSMatthew Ahrens FMR_PAGES_PP_MAXIMUM, 38522ec99e3eSMatthew Ahrens FMR_HEAP_ARENA, 38532ec99e3eSMatthew Ahrens FMR_ZIO_ARENA, 38542ec99e3eSMatthew Ahrens } free_memory_reason_t; 38552ec99e3eSMatthew Ahrens 38562ec99e3eSMatthew Ahrens int64_t last_free_memory; 38572ec99e3eSMatthew Ahrens free_memory_reason_t last_free_reason; 38582ec99e3eSMatthew Ahrens 385994dd93aeSGeorge Wilson /* 38602ec99e3eSMatthew Ahrens * Additional reserve of pages for pp_reserve. 386194dd93aeSGeorge Wilson */ 38622ec99e3eSMatthew Ahrens int64_t arc_pages_pp_reserve = 64; 3863fa9e4066Sahrens 38642ec99e3eSMatthew Ahrens /* 38652ec99e3eSMatthew Ahrens * Additional reserve of pages for swapfs. 38662ec99e3eSMatthew Ahrens */ 38672ec99e3eSMatthew Ahrens int64_t arc_swapfs_reserve = 64; 38683cff2f43Sstans 38692ec99e3eSMatthew Ahrens /* 38702ec99e3eSMatthew Ahrens * Return the amount of memory that can be consumed before reclaim will be 38712ec99e3eSMatthew Ahrens * needed. Positive if there is sufficient free memory, negative indicates 38722ec99e3eSMatthew Ahrens * the amount of memory that needs to be freed up. 38732ec99e3eSMatthew Ahrens */ 38742ec99e3eSMatthew Ahrens static int64_t 38752ec99e3eSMatthew Ahrens arc_available_memory(void) 38762ec99e3eSMatthew Ahrens { 38772ec99e3eSMatthew Ahrens int64_t lowest = INT64_MAX; 38782ec99e3eSMatthew Ahrens int64_t n; 38792ec99e3eSMatthew Ahrens free_memory_reason_t r = FMR_UNKNOWN; 38803cff2f43Sstans 38812ec99e3eSMatthew Ahrens #ifdef _KERNEL 38822ec99e3eSMatthew Ahrens if (needfree > 0) { 38832ec99e3eSMatthew Ahrens n = PAGESIZE * (-needfree); 38842ec99e3eSMatthew Ahrens if (n < lowest) { 38852ec99e3eSMatthew Ahrens lowest = n; 38862ec99e3eSMatthew Ahrens r = FMR_NEEDFREE; 38872ec99e3eSMatthew Ahrens } 38882ec99e3eSMatthew Ahrens } 3889fa9e4066Sahrens 3890fa9e4066Sahrens /* 3891fa9e4066Sahrens * check that we're out of range of the pageout scanner. It starts to 3892fa9e4066Sahrens * schedule paging if freemem is less than lotsfree and needfree. 3893fa9e4066Sahrens * lotsfree is the high-water mark for pageout, and needfree is the 3894fa9e4066Sahrens * number of needed free pages. We add extra pages here to make sure 3895fa9e4066Sahrens * the scanner doesn't start up while we're freeing memory. 3896fa9e4066Sahrens */ 38972ec99e3eSMatthew Ahrens n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 38982ec99e3eSMatthew Ahrens if (n < lowest) { 38992ec99e3eSMatthew Ahrens lowest = n; 39002ec99e3eSMatthew Ahrens r = FMR_LOTSFREE; 39012ec99e3eSMatthew Ahrens } 3902fa9e4066Sahrens 3903fa9e4066Sahrens /* 3904fa9e4066Sahrens * check to make sure that swapfs has enough space so that anon 3905fa94a07fSbrendan * reservations can still succeed. anon_resvmem() checks that the 3906fa9e4066Sahrens * availrmem is greater than swapfs_minfree, and the number of reserved 3907fa9e4066Sahrens * swap pages. We also add a bit of extra here just to prevent 3908fa9e4066Sahrens * circumstances from getting really dire. 3909fa9e4066Sahrens */ 39102ec99e3eSMatthew Ahrens n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 39112ec99e3eSMatthew Ahrens desfree - arc_swapfs_reserve); 39122ec99e3eSMatthew Ahrens if (n < lowest) { 39132ec99e3eSMatthew Ahrens lowest = n; 39142ec99e3eSMatthew Ahrens r = FMR_SWAPFS_MINFREE; 39152ec99e3eSMatthew Ahrens } 39162ec99e3eSMatthew Ahrens 3917fa9e4066Sahrens 3918cf746768SBryan Cantrill /* 3919cf746768SBryan Cantrill * Check that we have enough availrmem that memory locking (e.g., via 3920cf746768SBryan Cantrill * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3921cf746768SBryan Cantrill * stores the number of pages that cannot be locked; when availrmem 3922cf746768SBryan Cantrill * drops below pages_pp_maximum, page locking mechanisms such as 3923cf746768SBryan Cantrill * page_pp_lock() will fail.) 3924cf746768SBryan Cantrill */ 39252ec99e3eSMatthew Ahrens n = PAGESIZE * (availrmem - pages_pp_maximum - 39262ec99e3eSMatthew Ahrens arc_pages_pp_reserve); 39272ec99e3eSMatthew Ahrens if (n < lowest) { 39282ec99e3eSMatthew Ahrens lowest = n; 39292ec99e3eSMatthew Ahrens r = FMR_PAGES_PP_MAXIMUM; 39302ec99e3eSMatthew Ahrens } 3931cf746768SBryan Cantrill 39325dc8af33Smaybee #if defined(__i386) 3933fa9e4066Sahrens /* 3934fa9e4066Sahrens * If we're on an i386 platform, it's possible that we'll exhaust the 3935fa9e4066Sahrens * kernel heap space before we ever run out of available physical 3936fa9e4066Sahrens * memory. Most checks of the size of the heap_area compare against 3937fa9e4066Sahrens * tune.t_minarmem, which is the minimum available real memory that we 3938fa9e4066Sahrens * can have in the system. However, this is generally fixed at 25 pages 3939fa9e4066Sahrens * which is so low that it's useless. In this comparison, we seek to 3940fa9e4066Sahrens * calculate the total heap-size, and reclaim if more than 3/4ths of the 3941fa94a07fSbrendan * heap is allocated. (Or, in the calculation, if less than 1/4th is 3942fa9e4066Sahrens * free) 3943fa9e4066Sahrens */ 39440dd053d7SPrakash Surya n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 39452ec99e3eSMatthew Ahrens (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 39462ec99e3eSMatthew Ahrens if (n < lowest) { 39472ec99e3eSMatthew Ahrens lowest = n; 39482ec99e3eSMatthew Ahrens r = FMR_HEAP_ARENA; 39492ec99e3eSMatthew Ahrens } 3950fa9e4066Sahrens #endif 3951fa9e4066Sahrens 395294dd93aeSGeorge Wilson /* 395394dd93aeSGeorge Wilson * If zio data pages are being allocated out of a separate heap segment, 395494dd93aeSGeorge Wilson * then enforce that the size of available vmem for this arena remains 39550dd053d7SPrakash Surya * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free. 395694dd93aeSGeorge Wilson * 39570dd053d7SPrakash Surya * Note that reducing the arc_zio_arena_free_shift keeps more virtual 39580dd053d7SPrakash Surya * memory (in the zio_arena) free, which can avoid memory 39590dd053d7SPrakash Surya * fragmentation issues. 396094dd93aeSGeorge Wilson */ 39612ec99e3eSMatthew Ahrens if (zio_arena != NULL) { 39620dd053d7SPrakash Surya n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 39630dd053d7SPrakash Surya (vmem_size(zio_arena, VMEM_ALLOC) >> 39640dd053d7SPrakash Surya arc_zio_arena_free_shift); 39652ec99e3eSMatthew Ahrens if (n < lowest) { 39662ec99e3eSMatthew Ahrens lowest = n; 39672ec99e3eSMatthew Ahrens r = FMR_ZIO_ARENA; 39682ec99e3eSMatthew Ahrens } 39692ec99e3eSMatthew Ahrens } 3970fa9e4066Sahrens #else 39712ec99e3eSMatthew Ahrens /* Every 100 calls, free a small amount */ 3972fa9e4066Sahrens if (spa_get_random(100) == 0) 39732ec99e3eSMatthew Ahrens lowest = -1024; 3974fa9e4066Sahrens #endif 39752ec99e3eSMatthew Ahrens 39762ec99e3eSMatthew Ahrens last_free_memory = lowest; 39772ec99e3eSMatthew Ahrens last_free_reason = r; 39782ec99e3eSMatthew Ahrens 39792ec99e3eSMatthew Ahrens return (lowest); 39802ec99e3eSMatthew Ahrens } 39812ec99e3eSMatthew Ahrens 39822ec99e3eSMatthew Ahrens 39832ec99e3eSMatthew Ahrens /* 39842ec99e3eSMatthew Ahrens * Determine if the system is under memory pressure and is asking 3985dcbf3bd6SGeorge Wilson * to reclaim memory. A return value of B_TRUE indicates that the system 39862ec99e3eSMatthew Ahrens * is under memory pressure and that the arc should adjust accordingly. 39872ec99e3eSMatthew Ahrens */ 39882ec99e3eSMatthew Ahrens static boolean_t 39892ec99e3eSMatthew Ahrens arc_reclaim_needed(void) 39902ec99e3eSMatthew Ahrens { 39912ec99e3eSMatthew Ahrens return (arc_available_memory() < 0); 3992fa9e4066Sahrens } 3993fa9e4066Sahrens 3994fa9e4066Sahrens static void 39952ec99e3eSMatthew Ahrens arc_kmem_reap_now(void) 3996fa9e4066Sahrens { 3997fa9e4066Sahrens size_t i; 3998fa9e4066Sahrens kmem_cache_t *prev_cache = NULL; 3999ad23a2dbSjohansen kmem_cache_t *prev_data_cache = NULL; 4000fa9e4066Sahrens extern kmem_cache_t *zio_buf_cache[]; 4001ad23a2dbSjohansen extern kmem_cache_t *zio_data_buf_cache[]; 400283803b51SGeorge Wilson extern kmem_cache_t *range_seg_cache; 4003fa9e4066Sahrens 4004033f9833Sek #ifdef _KERNEL 40050e8c6158Smaybee if (arc_meta_used >= arc_meta_limit) { 40060e8c6158Smaybee /* 40070e8c6158Smaybee * We are exceeding our meta-data cache limit. 40080e8c6158Smaybee * Purge some DNLC entries to release holds on meta-data. 40090e8c6158Smaybee */ 40100e8c6158Smaybee dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 40110e8c6158Smaybee } 40125dc8af33Smaybee #if defined(__i386) 40135dc8af33Smaybee /* 40145dc8af33Smaybee * Reclaim unused memory from all kmem caches. 40155dc8af33Smaybee */ 40165dc8af33Smaybee kmem_reap(); 40175dc8af33Smaybee #endif 4018033f9833Sek #endif 4019033f9833Sek 4020fa9e4066Sahrens for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 4021fa9e4066Sahrens if (zio_buf_cache[i] != prev_cache) { 4022fa9e4066Sahrens prev_cache = zio_buf_cache[i]; 4023fa9e4066Sahrens kmem_cache_reap_now(zio_buf_cache[i]); 4024fa9e4066Sahrens } 4025ad23a2dbSjohansen if (zio_data_buf_cache[i] != prev_data_cache) { 4026ad23a2dbSjohansen prev_data_cache = zio_data_buf_cache[i]; 4027ad23a2dbSjohansen kmem_cache_reap_now(zio_data_buf_cache[i]); 4028ad23a2dbSjohansen } 4029fa9e4066Sahrens } 4030ea8dc4b6Seschrock kmem_cache_reap_now(buf_cache); 403189c86e32SChris Williamson kmem_cache_reap_now(hdr_full_cache); 403289c86e32SChris Williamson kmem_cache_reap_now(hdr_l2only_cache); 403383803b51SGeorge Wilson kmem_cache_reap_now(range_seg_cache); 403494dd93aeSGeorge Wilson 40352ec99e3eSMatthew Ahrens if (zio_arena != NULL) { 40362ec99e3eSMatthew Ahrens /* 40372ec99e3eSMatthew Ahrens * Ask the vmem arena to reclaim unused memory from its 40382ec99e3eSMatthew Ahrens * quantum caches. 40392ec99e3eSMatthew Ahrens */ 404094dd93aeSGeorge Wilson vmem_qcache_reap(zio_arena); 40412ec99e3eSMatthew Ahrens } 4042fa9e4066Sahrens } 4043fa9e4066Sahrens 4044244781f1SPrakash Surya /* 4045244781f1SPrakash Surya * Threads can block in arc_get_data_buf() waiting for this thread to evict 4046244781f1SPrakash Surya * enough data and signal them to proceed. When this happens, the threads in 4047244781f1SPrakash Surya * arc_get_data_buf() are sleeping while holding the hash lock for their 4048244781f1SPrakash Surya * particular arc header. Thus, we must be careful to never sleep on a 4049244781f1SPrakash Surya * hash lock in this thread. This is to prevent the following deadlock: 4050244781f1SPrakash Surya * 4051244781f1SPrakash Surya * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", 4052244781f1SPrakash Surya * waiting for the reclaim thread to signal it. 4053244781f1SPrakash Surya * 4054244781f1SPrakash Surya * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 4055244781f1SPrakash Surya * fails, and goes to sleep forever. 4056244781f1SPrakash Surya * 4057244781f1SPrakash Surya * This possible deadlock is avoided by always acquiring a hash lock 4058244781f1SPrakash Surya * using mutex_tryenter() from arc_reclaim_thread(). 4059244781f1SPrakash Surya */ 4060fa9e4066Sahrens static void 4061fa9e4066Sahrens arc_reclaim_thread(void) 4062fa9e4066Sahrens { 4063a8f6344fSEli Rosenthal hrtime_t growtime = 0; 4064fa9e4066Sahrens callb_cpr_t cpr; 4065fa9e4066Sahrens 4066244781f1SPrakash Surya CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 4067fa9e4066Sahrens 4068244781f1SPrakash Surya mutex_enter(&arc_reclaim_lock); 4069244781f1SPrakash Surya while (!arc_reclaim_thread_exit) { 4070244781f1SPrakash Surya uint64_t evicted = 0; 4071244781f1SPrakash Surya 4072dcbf3bd6SGeorge Wilson /* 4073dcbf3bd6SGeorge Wilson * This is necessary in order for the mdb ::arc dcmd to 4074dcbf3bd6SGeorge Wilson * show up to date information. Since the ::arc command 4075dcbf3bd6SGeorge Wilson * does not call the kstat's update function, without 4076dcbf3bd6SGeorge Wilson * this call, the command may show stale stats for the 4077dcbf3bd6SGeorge Wilson * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 4078dcbf3bd6SGeorge Wilson * with this change, the data might be up to 1 second 4079dcbf3bd6SGeorge Wilson * out of date; but that should suffice. The arc_state_t 4080dcbf3bd6SGeorge Wilson * structures can be queried directly if more accurate 4081dcbf3bd6SGeorge Wilson * information is needed. 4082dcbf3bd6SGeorge Wilson */ 4083dcbf3bd6SGeorge Wilson if (arc_ksp != NULL) 4084dcbf3bd6SGeorge Wilson arc_ksp->ks_update(arc_ksp, KSTAT_READ); 4085dcbf3bd6SGeorge Wilson 4086244781f1SPrakash Surya mutex_exit(&arc_reclaim_lock); 4087244781f1SPrakash Surya 4088405a5a0fSMatthew Ahrens /* 4089405a5a0fSMatthew Ahrens * We call arc_adjust() before (possibly) calling 4090405a5a0fSMatthew Ahrens * arc_kmem_reap_now(), so that we can wake up 4091405a5a0fSMatthew Ahrens * arc_get_data_buf() sooner. 4092405a5a0fSMatthew Ahrens */ 4093405a5a0fSMatthew Ahrens evicted = arc_adjust(); 4094405a5a0fSMatthew Ahrens 4095405a5a0fSMatthew Ahrens int64_t free_memory = arc_available_memory(); 40962ec99e3eSMatthew Ahrens if (free_memory < 0) { 4097fa9e4066Sahrens 40982ec99e3eSMatthew Ahrens arc_no_grow = B_TRUE; 40992ec99e3eSMatthew Ahrens arc_warm = B_TRUE; 4100fa9e4066Sahrens 41012ec99e3eSMatthew Ahrens /* 41022ec99e3eSMatthew Ahrens * Wait at least zfs_grow_retry (default 60) seconds 41032ec99e3eSMatthew Ahrens * before considering growing. 41042ec99e3eSMatthew Ahrens */ 4105a8f6344fSEli Rosenthal growtime = gethrtime() + SEC2NSEC(arc_grow_retry); 4106fa9e4066Sahrens 41072ec99e3eSMatthew Ahrens arc_kmem_reap_now(); 4108fa9e4066Sahrens 41092ec99e3eSMatthew Ahrens /* 41102ec99e3eSMatthew Ahrens * If we are still low on memory, shrink the ARC 41112ec99e3eSMatthew Ahrens * so that we have arc_shrink_min free space. 41122ec99e3eSMatthew Ahrens */ 41132ec99e3eSMatthew Ahrens free_memory = arc_available_memory(); 41142ec99e3eSMatthew Ahrens 41152ec99e3eSMatthew Ahrens int64_t to_free = 41162ec99e3eSMatthew Ahrens (arc_c >> arc_shrink_shift) - free_memory; 41172ec99e3eSMatthew Ahrens if (to_free > 0) { 41182ec99e3eSMatthew Ahrens #ifdef _KERNEL 41192ec99e3eSMatthew Ahrens to_free = MAX(to_free, ptob(needfree)); 41202ec99e3eSMatthew Ahrens #endif 41212ec99e3eSMatthew Ahrens arc_shrink(to_free); 41222ec99e3eSMatthew Ahrens } 41232ec99e3eSMatthew Ahrens } else if (free_memory < arc_c >> arc_no_grow_shift) { 41242ec99e3eSMatthew Ahrens arc_no_grow = B_TRUE; 4125a8f6344fSEli Rosenthal } else if (gethrtime() >= growtime) { 41262ec99e3eSMatthew Ahrens arc_no_grow = B_FALSE; 4127fa9e4066Sahrens } 4128fa9e4066Sahrens 4129244781f1SPrakash Surya mutex_enter(&arc_reclaim_lock); 4130244781f1SPrakash Surya 4131244781f1SPrakash Surya /* 4132244781f1SPrakash Surya * If evicted is zero, we couldn't evict anything via 4133244781f1SPrakash Surya * arc_adjust(). This could be due to hash lock 4134244781f1SPrakash Surya * collisions, but more likely due to the majority of 4135244781f1SPrakash Surya * arc buffers being unevictable. Therefore, even if 4136244781f1SPrakash Surya * arc_size is above arc_c, another pass is unlikely to 4137244781f1SPrakash Surya * be helpful and could potentially cause us to enter an 4138244781f1SPrakash Surya * infinite loop. 4139244781f1SPrakash Surya */ 4140244781f1SPrakash Surya if (arc_size <= arc_c || evicted == 0) { 4141244781f1SPrakash Surya /* 4142244781f1SPrakash Surya * We're either no longer overflowing, or we 4143244781f1SPrakash Surya * can't evict anything more, so we should wake 4144244781f1SPrakash Surya * up any threads before we go to sleep. 4145244781f1SPrakash Surya */ 4146244781f1SPrakash Surya cv_broadcast(&arc_reclaim_waiters_cv); 4147244781f1SPrakash Surya 4148244781f1SPrakash Surya /* 4149244781f1SPrakash Surya * Block until signaled, or after one second (we 4150244781f1SPrakash Surya * might need to perform arc_kmem_reap_now() 4151244781f1SPrakash Surya * even if we aren't being signalled) 4152244781f1SPrakash Surya */ 4153244781f1SPrakash Surya CALLB_CPR_SAFE_BEGIN(&cpr); 4154a8f6344fSEli Rosenthal (void) cv_timedwait_hires(&arc_reclaim_thread_cv, 4155a8f6344fSEli Rosenthal &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 4156244781f1SPrakash Surya CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 4157244781f1SPrakash Surya } 4158244781f1SPrakash Surya } 4159244781f1SPrakash Surya 4160dcbf3bd6SGeorge Wilson arc_reclaim_thread_exit = B_FALSE; 4161244781f1SPrakash Surya cv_broadcast(&arc_reclaim_thread_cv); 4162244781f1SPrakash Surya CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 4163244781f1SPrakash Surya thread_exit(); 4164244781f1SPrakash Surya } 4165244781f1SPrakash Surya 4166ea8dc4b6Seschrock /* 4167ea8dc4b6Seschrock * Adapt arc info given the number of bytes we are trying to add and 4168ea8dc4b6Seschrock * the state that we are comming from. This function is only called 4169ea8dc4b6Seschrock * when we are adding new content to the cache. 4170ea8dc4b6Seschrock */ 4171fa9e4066Sahrens static void 4172ea8dc4b6Seschrock arc_adapt(int bytes, arc_state_t *state) 4173fa9e4066Sahrens { 4174ea8dc4b6Seschrock int mult; 41755a98e54bSBrendan Gregg - Sun Microsystems uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 41762fd872a7SPrakash Surya int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 41772fd872a7SPrakash Surya int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 4178ea8dc4b6Seschrock 4179fa94a07fSbrendan if (state == arc_l2c_only) 4180fa94a07fSbrendan return; 4181fa94a07fSbrendan 4182ea8dc4b6Seschrock ASSERT(bytes > 0); 4183fa9e4066Sahrens /* 4184ea8dc4b6Seschrock * Adapt the target size of the MRU list: 4185ea8dc4b6Seschrock * - if we just hit in the MRU ghost list, then increase 4186ea8dc4b6Seschrock * the target size of the MRU list. 4187ea8dc4b6Seschrock * - if we just hit in the MFU ghost list, then increase 4188ea8dc4b6Seschrock * the target size of the MFU list by decreasing the 4189ea8dc4b6Seschrock * target size of the MRU list. 4190fa9e4066Sahrens */ 419144cb6abcSbmc if (state == arc_mru_ghost) { 41922fd872a7SPrakash Surya mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 41933e4e8481STom Erickson mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 4194ea8dc4b6Seschrock 41955a98e54bSBrendan Gregg - Sun Microsystems arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 419644cb6abcSbmc } else if (state == arc_mfu_ghost) { 41975a98e54bSBrendan Gregg - Sun Microsystems uint64_t delta; 41985a98e54bSBrendan Gregg - Sun Microsystems 41992fd872a7SPrakash Surya mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 42003e4e8481STom Erickson mult = MIN(mult, 10); 4201ea8dc4b6Seschrock 42025a98e54bSBrendan Gregg - Sun Microsystems delta = MIN(bytes * mult, arc_p); 42035a98e54bSBrendan Gregg - Sun Microsystems arc_p = MAX(arc_p_min, arc_p - delta); 4204ea8dc4b6Seschrock } 420544cb6abcSbmc ASSERT((int64_t)arc_p >= 0); 4206fa9e4066Sahrens 4207fa9e4066Sahrens if (arc_reclaim_needed()) { 4208244781f1SPrakash Surya cv_signal(&arc_reclaim_thread_cv); 4209fa9e4066Sahrens return; 4210fa9e4066Sahrens } 4211fa9e4066Sahrens 421244cb6abcSbmc if (arc_no_grow) 4213fa9e4066Sahrens return; 4214fa9e4066Sahrens 421544cb6abcSbmc if (arc_c >= arc_c_max) 4216ea8dc4b6Seschrock return; 4217ea8dc4b6Seschrock 4218fa9e4066Sahrens /* 4219ea8dc4b6Seschrock * If we're within (2 * maxblocksize) bytes of the target 4220ea8dc4b6Seschrock * cache size, increment the target cache size 4221fa9e4066Sahrens */ 422244cb6abcSbmc if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 422344cb6abcSbmc atomic_add_64(&arc_c, (int64_t)bytes); 422444cb6abcSbmc if (arc_c > arc_c_max) 422544cb6abcSbmc arc_c = arc_c_max; 422644cb6abcSbmc else if (state == arc_anon) 422744cb6abcSbmc atomic_add_64(&arc_p, (int64_t)bytes); 422844cb6abcSbmc if (arc_p > arc_c) 422944cb6abcSbmc arc_p = arc_c; 4230fa9e4066Sahrens } 423144cb6abcSbmc ASSERT((int64_t)arc_p >= 0); 4232fa9e4066Sahrens } 4233fa9e4066Sahrens 4234fa9e4066Sahrens /* 4235244781f1SPrakash Surya * Check if arc_size has grown past our upper threshold, determined by 4236244781f1SPrakash Surya * zfs_arc_overflow_shift. 4237fa9e4066Sahrens */ 4238244781f1SPrakash Surya static boolean_t 4239244781f1SPrakash Surya arc_is_overflowing(void) 4240fa9e4066Sahrens { 4241244781f1SPrakash Surya /* Always allow at least one block of overflow */ 4242244781f1SPrakash Surya uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 4243244781f1SPrakash Surya arc_c >> zfs_arc_overflow_shift); 4244fa9e4066Sahrens 4245244781f1SPrakash Surya return (arc_size >= arc_c + overflow); 4246fa9e4066Sahrens } 4247fa9e4066Sahrens 4248fa9e4066Sahrens /* 4249dcbf3bd6SGeorge Wilson * Allocate a block and return it to the caller. If we are hitting the 4250dcbf3bd6SGeorge Wilson * hard limit for the cache size, we must sleep, waiting for the eviction 4251dcbf3bd6SGeorge Wilson * thread to catch up. If we're past the target size but below the hard 4252dcbf3bd6SGeorge Wilson * limit, we'll only signal the reclaim thread and continue on. 4253fa9e4066Sahrens */ 4254dcbf3bd6SGeorge Wilson static void * 4255dcbf3bd6SGeorge Wilson arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4256fa9e4066Sahrens { 4257dcbf3bd6SGeorge Wilson void *datap = NULL; 4258dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 4259dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 4260fa9e4066Sahrens 426144eda4d7Smaybee arc_adapt(size, state); 4262fa9e4066Sahrens 426344eda4d7Smaybee /* 4264244781f1SPrakash Surya * If arc_size is currently overflowing, and has grown past our 4265244781f1SPrakash Surya * upper limit, we must be adding data faster than the evict 4266244781f1SPrakash Surya * thread can evict. Thus, to ensure we don't compound the 4267244781f1SPrakash Surya * problem by adding more data and forcing arc_size to grow even 4268244781f1SPrakash Surya * further past it's target size, we halt and wait for the 4269244781f1SPrakash Surya * eviction thread to catch up. 4270244781f1SPrakash Surya * 4271244781f1SPrakash Surya * It's also possible that the reclaim thread is unable to evict 4272244781f1SPrakash Surya * enough buffers to get arc_size below the overflow limit (e.g. 4273244781f1SPrakash Surya * due to buffers being un-evictable, or hash lock collisions). 4274244781f1SPrakash Surya * In this case, we want to proceed regardless if we're 4275244781f1SPrakash Surya * overflowing; thus we don't use a while loop here. 427644eda4d7Smaybee */ 4277244781f1SPrakash Surya if (arc_is_overflowing()) { 4278244781f1SPrakash Surya mutex_enter(&arc_reclaim_lock); 4279244781f1SPrakash Surya 4280244781f1SPrakash Surya /* 4281244781f1SPrakash Surya * Now that we've acquired the lock, we may no longer be 4282244781f1SPrakash Surya * over the overflow limit, lets check. 4283244781f1SPrakash Surya * 4284244781f1SPrakash Surya * We're ignoring the case of spurious wake ups. If that 4285244781f1SPrakash Surya * were to happen, it'd let this thread consume an ARC 4286244781f1SPrakash Surya * buffer before it should have (i.e. before we're under 4287244781f1SPrakash Surya * the overflow limit and were signalled by the reclaim 4288244781f1SPrakash Surya * thread). As long as that is a rare occurrence, it 4289244781f1SPrakash Surya * shouldn't cause any harm. 4290244781f1SPrakash Surya */ 4291244781f1SPrakash Surya if (arc_is_overflowing()) { 4292244781f1SPrakash Surya cv_signal(&arc_reclaim_thread_cv); 4293244781f1SPrakash Surya cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 4294ad23a2dbSjohansen } 4295244781f1SPrakash Surya 4296244781f1SPrakash Surya mutex_exit(&arc_reclaim_lock); 429744eda4d7Smaybee } 429844eda4d7Smaybee 4299dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 4300244781f1SPrakash Surya if (type == ARC_BUFC_METADATA) { 4301dcbf3bd6SGeorge Wilson datap = zio_buf_alloc(size); 4302244781f1SPrakash Surya arc_space_consume(size, ARC_SPACE_META); 4303fa9e4066Sahrens } else { 4304244781f1SPrakash Surya ASSERT(type == ARC_BUFC_DATA); 4305dcbf3bd6SGeorge Wilson datap = zio_data_buf_alloc(size); 4306244781f1SPrakash Surya arc_space_consume(size, ARC_SPACE_DATA); 430744eda4d7Smaybee } 4308244781f1SPrakash Surya 430944eda4d7Smaybee /* 431044eda4d7Smaybee * Update the state size. Note that ghost states have a 431144eda4d7Smaybee * "ghost size" and so don't need to be updated. 431244eda4d7Smaybee */ 4313dcbf3bd6SGeorge Wilson if (!GHOST_STATE(state)) { 431444eda4d7Smaybee 4315dcbf3bd6SGeorge Wilson (void) refcount_add_many(&state->arcs_size, size, tag); 4316244781f1SPrakash Surya 4317244781f1SPrakash Surya /* 4318244781f1SPrakash Surya * If this is reached via arc_read, the link is 4319244781f1SPrakash Surya * protected by the hash lock. If reached via 4320244781f1SPrakash Surya * arc_buf_alloc, the header should not be accessed by 4321244781f1SPrakash Surya * any other thread. And, if reached via arc_read_done, 4322244781f1SPrakash Surya * the hash lock will protect it if it's found in the 4323244781f1SPrakash Surya * hash table; otherwise no other thread should be 4324244781f1SPrakash Surya * trying to [add|remove]_reference it. 4325244781f1SPrakash Surya */ 4326244781f1SPrakash Surya if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 432789c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4328dcbf3bd6SGeorge Wilson (void) refcount_add_many(&state->arcs_esize[type], 4329dcbf3bd6SGeorge Wilson size, tag); 4330fa9e4066Sahrens } 4331dcbf3bd6SGeorge Wilson 4332641fbdaeSmaybee /* 4333641fbdaeSmaybee * If we are growing the cache, and we are adding anonymous 433444cb6abcSbmc * data, and we have outgrown arc_p, update arc_p 4335641fbdaeSmaybee */ 433689c86e32SChris Williamson if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 43372fd872a7SPrakash Surya (refcount_count(&arc_anon->arcs_size) + 43382fd872a7SPrakash Surya refcount_count(&arc_mru->arcs_size) > arc_p)) 433944cb6abcSbmc arc_p = MIN(arc_c, arc_p + size); 4340fa9e4066Sahrens } 4341dcbf3bd6SGeorge Wilson return (datap); 4342dcbf3bd6SGeorge Wilson } 4343dcbf3bd6SGeorge Wilson 4344dcbf3bd6SGeorge Wilson /* 4345dcbf3bd6SGeorge Wilson * Free the arc data buffer. 4346dcbf3bd6SGeorge Wilson */ 4347dcbf3bd6SGeorge Wilson static void 4348dcbf3bd6SGeorge Wilson arc_free_data_buf(arc_buf_hdr_t *hdr, void *data, uint64_t size, void *tag) 4349dcbf3bd6SGeorge Wilson { 4350dcbf3bd6SGeorge Wilson arc_state_t *state = hdr->b_l1hdr.b_state; 4351dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 4352dcbf3bd6SGeorge Wilson 4353dcbf3bd6SGeorge Wilson /* protected by hash lock, if in the hash table */ 4354dcbf3bd6SGeorge Wilson if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4355dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4356dcbf3bd6SGeorge Wilson ASSERT(state != arc_anon && state != arc_l2c_only); 4357dcbf3bd6SGeorge Wilson 4358dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 4359dcbf3bd6SGeorge Wilson size, tag); 4360dcbf3bd6SGeorge Wilson } 4361dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_size, size, tag); 4362dcbf3bd6SGeorge Wilson 4363dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 4364dcbf3bd6SGeorge Wilson if (type == ARC_BUFC_METADATA) { 4365dcbf3bd6SGeorge Wilson zio_buf_free(data, size); 4366dcbf3bd6SGeorge Wilson arc_space_return(size, ARC_SPACE_META); 4367dcbf3bd6SGeorge Wilson } else { 4368dcbf3bd6SGeorge Wilson ASSERT(type == ARC_BUFC_DATA); 4369dcbf3bd6SGeorge Wilson zio_data_buf_free(data, size); 4370dcbf3bd6SGeorge Wilson arc_space_return(size, ARC_SPACE_DATA); 4371dcbf3bd6SGeorge Wilson } 4372fa9e4066Sahrens } 4373fa9e4066Sahrens 4374fa9e4066Sahrens /* 4375fa9e4066Sahrens * This routine is called whenever a buffer is accessed. 4376ea8dc4b6Seschrock * NOTE: the hash lock is dropped in this function. 4377fa9e4066Sahrens */ 4378fa9e4066Sahrens static void 43797adb730bSGeorge Wilson arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 4380fa9e4066Sahrens { 4381d3d50737SRafael Vanoni clock_t now; 4382d3d50737SRafael Vanoni 4383fa9e4066Sahrens ASSERT(MUTEX_HELD(hash_lock)); 438489c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 4385fa9e4066Sahrens 438689c86e32SChris Williamson if (hdr->b_l1hdr.b_state == arc_anon) { 4387fa9e4066Sahrens /* 4388fa9e4066Sahrens * This buffer is not in the cache, and does not 4389fa9e4066Sahrens * appear in our "ghost" list. Add the new buffer 4390fa9e4066Sahrens * to the MRU state. 4391fa9e4066Sahrens */ 4392fa9e4066Sahrens 439389c86e32SChris Williamson ASSERT0(hdr->b_l1hdr.b_arc_access); 439489c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 43957adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 43967adb730bSGeorge Wilson arc_change_state(arc_mru, hdr, hash_lock); 4397fa9e4066Sahrens 439889c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mru) { 4399d3d50737SRafael Vanoni now = ddi_get_lbolt(); 4400d3d50737SRafael Vanoni 4401fa9e4066Sahrens /* 440213506d1eSmaybee * If this buffer is here because of a prefetch, then either: 440313506d1eSmaybee * - clear the flag if this is a "referencing" read 440413506d1eSmaybee * (any subsequent access will bump this into the MFU state). 440513506d1eSmaybee * or 440613506d1eSmaybee * - move the buffer to the head of the list if this is 440713506d1eSmaybee * another prefetch (to make it less likely to be evicted). 4408fa9e4066Sahrens */ 440989c86e32SChris Williamson if (HDR_PREFETCH(hdr)) { 441089c86e32SChris Williamson if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4411244781f1SPrakash Surya /* link protected by hash lock */ 4412244781f1SPrakash Surya ASSERT(multilist_link_active( 441389c86e32SChris Williamson &hdr->b_l1hdr.b_arc_node)); 441413506d1eSmaybee } else { 4415dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 441644cb6abcSbmc ARCSTAT_BUMP(arcstat_mru_hits); 441713506d1eSmaybee } 441889c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = now; 4419fa9e4066Sahrens return; 4420fa9e4066Sahrens } 4421fa9e4066Sahrens 4422fa9e4066Sahrens /* 4423fa9e4066Sahrens * This buffer has been "accessed" only once so far, 4424fa9e4066Sahrens * but it is still in the cache. Move it to the MFU 4425fa9e4066Sahrens * state. 4426fa9e4066Sahrens */ 442789c86e32SChris Williamson if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 4428fa9e4066Sahrens /* 4429fa9e4066Sahrens * More than 125ms have passed since we 4430fa9e4066Sahrens * instantiated this buffer. Move it to the 4431fa9e4066Sahrens * most frequently used state. 4432fa9e4066Sahrens */ 443389c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = now; 44347adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 44357adb730bSGeorge Wilson arc_change_state(arc_mfu, hdr, hash_lock); 4436fa9e4066Sahrens } 443744cb6abcSbmc ARCSTAT_BUMP(arcstat_mru_hits); 443889c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 4439fa9e4066Sahrens arc_state_t *new_state; 4440fa9e4066Sahrens /* 4441fa9e4066Sahrens * This buffer has been "accessed" recently, but 4442fa9e4066Sahrens * was evicted from the cache. Move it to the 4443fa9e4066Sahrens * MFU state. 4444fa9e4066Sahrens */ 4445fa9e4066Sahrens 444689c86e32SChris Williamson if (HDR_PREFETCH(hdr)) { 444744cb6abcSbmc new_state = arc_mru; 444889c86e32SChris Williamson if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 4449dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 44507adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4451fa9e4066Sahrens } else { 445244cb6abcSbmc new_state = arc_mfu; 44537adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4454fa9e4066Sahrens } 4455fa9e4066Sahrens 445689c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 44577adb730bSGeorge Wilson arc_change_state(new_state, hdr, hash_lock); 4458fa9e4066Sahrens 445944cb6abcSbmc ARCSTAT_BUMP(arcstat_mru_ghost_hits); 446089c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mfu) { 4461fa9e4066Sahrens /* 4462fa9e4066Sahrens * This buffer has been accessed more than once and is 4463fa9e4066Sahrens * still in the cache. Keep it in the MFU state. 4464fa9e4066Sahrens * 446513506d1eSmaybee * NOTE: an add_reference() that occurred when we did 446613506d1eSmaybee * the arc_read() will have kicked this off the list. 446713506d1eSmaybee * If it was a prefetch, we will explicitly move it to 446813506d1eSmaybee * the head of the list now. 4469fa9e4066Sahrens */ 447089c86e32SChris Williamson if ((HDR_PREFETCH(hdr)) != 0) { 447189c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4472244781f1SPrakash Surya /* link protected by hash_lock */ 4473244781f1SPrakash Surya ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 447413506d1eSmaybee } 447544cb6abcSbmc ARCSTAT_BUMP(arcstat_mfu_hits); 447689c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 447789c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 447844cb6abcSbmc arc_state_t *new_state = arc_mfu; 4479fa9e4066Sahrens /* 4480fa9e4066Sahrens * This buffer has been accessed more than once but has 4481fa9e4066Sahrens * been evicted from the cache. Move it back to the 4482fa9e4066Sahrens * MFU state. 4483fa9e4066Sahrens */ 4484fa9e4066Sahrens 448589c86e32SChris Williamson if (HDR_PREFETCH(hdr)) { 448613506d1eSmaybee /* 448713506d1eSmaybee * This is a prefetch access... 448813506d1eSmaybee * move this block back to the MRU state. 448913506d1eSmaybee */ 449089c86e32SChris Williamson ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 449144cb6abcSbmc new_state = arc_mru; 449213506d1eSmaybee } 449313506d1eSmaybee 449489c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 44957adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 44967adb730bSGeorge Wilson arc_change_state(new_state, hdr, hash_lock); 4497fa9e4066Sahrens 449844cb6abcSbmc ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 449989c86e32SChris Williamson } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4500fa94a07fSbrendan /* 4501fa94a07fSbrendan * This buffer is on the 2nd Level ARC. 4502fa94a07fSbrendan */ 4503fa94a07fSbrendan 450489c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 45057adb730bSGeorge Wilson DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 45067adb730bSGeorge Wilson arc_change_state(arc_mfu, hdr, hash_lock); 4507fa9e4066Sahrens } else { 4508fa9e4066Sahrens ASSERT(!"invalid arc state"); 4509fa9e4066Sahrens } 4510fa9e4066Sahrens } 4511fa9e4066Sahrens 4512fa9e4066Sahrens /* a generic arc_done_func_t which you can use */ 4513fa9e4066Sahrens /* ARGSUSED */ 4514fa9e4066Sahrens void 4515fa9e4066Sahrens arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4516fa9e4066Sahrens { 45173f9d6ad7SLin Ling if (zio == NULL || zio->io_error == 0) 4518*5602294fSDan Kimmel bcopy(buf->b_data, arg, arc_buf_size(buf)); 4519dcbf3bd6SGeorge Wilson arc_buf_destroy(buf, arg); 4520fa9e4066Sahrens } 4521fa9e4066Sahrens 45220e8c6158Smaybee /* a generic arc_done_func_t */ 4523fa9e4066Sahrens void 4524fa9e4066Sahrens arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4525fa9e4066Sahrens { 4526fa9e4066Sahrens arc_buf_t **bufp = arg; 4527fa9e4066Sahrens if (zio && zio->io_error) { 4528dcbf3bd6SGeorge Wilson arc_buf_destroy(buf, arg); 4529fa9e4066Sahrens *bufp = NULL; 4530fa9e4066Sahrens } else { 4531fa9e4066Sahrens *bufp = buf; 45323f9d6ad7SLin Ling ASSERT(buf->b_data); 4533fa9e4066Sahrens } 4534fa9e4066Sahrens } 4535fa9e4066Sahrens 4536dcbf3bd6SGeorge Wilson static void 4537dcbf3bd6SGeorge Wilson arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) 4538dcbf3bd6SGeorge Wilson { 4539dcbf3bd6SGeorge Wilson if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { 4540dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); 4541dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 4542dcbf3bd6SGeorge Wilson } else { 4543dcbf3bd6SGeorge Wilson if (HDR_COMPRESSION_ENABLED(hdr)) { 4544dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), ==, 4545dcbf3bd6SGeorge Wilson BP_GET_COMPRESS(bp)); 4546dcbf3bd6SGeorge Wilson } 4547dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); 4548dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); 4549dcbf3bd6SGeorge Wilson } 4550dcbf3bd6SGeorge Wilson } 4551dcbf3bd6SGeorge Wilson 4552fa9e4066Sahrens static void 4553fa9e4066Sahrens arc_read_done(zio_t *zio) 4554fa9e4066Sahrens { 4555dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = zio->io_private; 45565d7b4d43SMatthew Ahrens kmutex_t *hash_lock = NULL; 4557*5602294fSDan Kimmel arc_callback_t *callback_list; 4558*5602294fSDan Kimmel arc_callback_t *acb; 4559*5602294fSDan Kimmel boolean_t freeable = B_FALSE; 4560*5602294fSDan Kimmel boolean_t no_zio_error = (zio->io_error == 0); 4561fa9e4066Sahrens 4562bbf4a8dfSmaybee /* 4563bbf4a8dfSmaybee * The hdr was inserted into hash-table and removed from lists 4564bbf4a8dfSmaybee * prior to starting I/O. We should find this header, since 4565bbf4a8dfSmaybee * it's in the hash table, and it should be legit since it's 4566bbf4a8dfSmaybee * not possible to evict it during the I/O. The only possible 4567bbf4a8dfSmaybee * reason for it not to be found is if we were freed during the 4568bbf4a8dfSmaybee * read. 4569bbf4a8dfSmaybee */ 45705d7b4d43SMatthew Ahrens if (HDR_IN_HASH_TABLE(hdr)) { 45715d7b4d43SMatthew Ahrens ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 45725d7b4d43SMatthew Ahrens ASSERT3U(hdr->b_dva.dva_word[0], ==, 45735d7b4d43SMatthew Ahrens BP_IDENTITY(zio->io_bp)->dva_word[0]); 45745d7b4d43SMatthew Ahrens ASSERT3U(hdr->b_dva.dva_word[1], ==, 45755d7b4d43SMatthew Ahrens BP_IDENTITY(zio->io_bp)->dva_word[1]); 45765d7b4d43SMatthew Ahrens 45775d7b4d43SMatthew Ahrens arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 45785d7b4d43SMatthew Ahrens &hash_lock); 45795d7b4d43SMatthew Ahrens 4580dcbf3bd6SGeorge Wilson ASSERT((found == hdr && 45815d7b4d43SMatthew Ahrens DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 45825d7b4d43SMatthew Ahrens (found == hdr && HDR_L2_READING(hdr))); 4583dcbf3bd6SGeorge Wilson ASSERT3P(hash_lock, !=, NULL); 4584dcbf3bd6SGeorge Wilson } 4585dcbf3bd6SGeorge Wilson 4586*5602294fSDan Kimmel if (no_zio_error) { 4587dcbf3bd6SGeorge Wilson /* byteswap if necessary */ 4588dcbf3bd6SGeorge Wilson if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 4589dcbf3bd6SGeorge Wilson if (BP_GET_LEVEL(zio->io_bp) > 0) { 4590dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; 4591dcbf3bd6SGeorge Wilson } else { 4592dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = 4593dcbf3bd6SGeorge Wilson DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4594dcbf3bd6SGeorge Wilson } 4595dcbf3bd6SGeorge Wilson } else { 4596dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 4597dcbf3bd6SGeorge Wilson } 45985d7b4d43SMatthew Ahrens } 4599fa94a07fSbrendan 4600dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); 460189c86e32SChris Williamson if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4602dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); 4603fa9e4066Sahrens 460489c86e32SChris Williamson callback_list = hdr->b_l1hdr.b_acb; 4605dcbf3bd6SGeorge Wilson ASSERT3P(callback_list, !=, NULL); 46066b4acc8bSahrens 4607*5602294fSDan Kimmel if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) { 4608b24ab676SJeff Bonwick /* 4609b24ab676SJeff Bonwick * Only call arc_access on anonymous buffers. This is because 4610b24ab676SJeff Bonwick * if we've issued an I/O for an evicted buffer, we've already 4611b24ab676SJeff Bonwick * called arc_access (to prevent any simultaneous readers from 4612b24ab676SJeff Bonwick * getting confused). 4613b24ab676SJeff Bonwick */ 4614b24ab676SJeff Bonwick arc_access(hdr, hash_lock); 4615b24ab676SJeff Bonwick } 4616b24ab676SJeff Bonwick 4617*5602294fSDan Kimmel /* 4618*5602294fSDan Kimmel * If a read request has a callback (i.e. acb_done is not NULL), then we 4619*5602294fSDan Kimmel * make a buf containing the data according to the parameters which were 4620*5602294fSDan Kimmel * passed in. The implementation of arc_buf_alloc_impl() ensures that we 4621*5602294fSDan Kimmel * aren't needlessly decompressing the data multiple times. 4622*5602294fSDan Kimmel */ 4623*5602294fSDan Kimmel int callback_cnt = 0; 4624*5602294fSDan Kimmel for (acb = callback_list; acb != NULL; acb = acb->acb_next) { 4625*5602294fSDan Kimmel if (!acb->acb_done) 4626*5602294fSDan Kimmel continue; 4627*5602294fSDan Kimmel 4628*5602294fSDan Kimmel /* This is a demand read since prefetches don't use callbacks */ 4629*5602294fSDan Kimmel callback_cnt++; 4630*5602294fSDan Kimmel 4631*5602294fSDan Kimmel int error = arc_buf_alloc_impl(hdr, acb->acb_private, 4632*5602294fSDan Kimmel acb->acb_compressed, no_zio_error, &acb->acb_buf); 4633*5602294fSDan Kimmel if (no_zio_error) { 4634*5602294fSDan Kimmel zio->io_error = error; 4635fa9e4066Sahrens } 4636fa9e4066Sahrens } 463789c86e32SChris Williamson hdr->b_l1hdr.b_acb = NULL; 4638dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 4639*5602294fSDan Kimmel if (callback_cnt == 0) { 4640dcbf3bd6SGeorge Wilson ASSERT(HDR_PREFETCH(hdr)); 4641dcbf3bd6SGeorge Wilson ASSERT0(hdr->b_l1hdr.b_bufcnt); 4642dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 4643b24ab676SJeff Bonwick } 4644fa9e4066Sahrens 464589c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 464689c86e32SChris Williamson callback_list != NULL); 4647fa9e4066Sahrens 4648*5602294fSDan Kimmel if (no_zio_error) { 4649dcbf3bd6SGeorge Wilson arc_hdr_verify(hdr, zio->io_bp); 4650dcbf3bd6SGeorge Wilson } else { 4651dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); 465289c86e32SChris Williamson if (hdr->b_l1hdr.b_state != arc_anon) 465344cb6abcSbmc arc_change_state(arc_anon, hdr, hash_lock); 4654ea8dc4b6Seschrock if (HDR_IN_HASH_TABLE(hdr)) 4655ea8dc4b6Seschrock buf_hash_remove(hdr); 465689c86e32SChris Williamson freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4657fa9e4066Sahrens } 4658fa9e4066Sahrens 4659ea8dc4b6Seschrock /* 466013506d1eSmaybee * Broadcast before we drop the hash_lock to avoid the possibility 466113506d1eSmaybee * that the hdr (and hence the cv) might be freed before we get to 466213506d1eSmaybee * the cv_broadcast(). 4663ea8dc4b6Seschrock */ 466489c86e32SChris Williamson cv_broadcast(&hdr->b_l1hdr.b_cv); 4665ea8dc4b6Seschrock 466689c86e32SChris Williamson if (hash_lock != NULL) { 466744eda4d7Smaybee mutex_exit(hash_lock); 4668fa9e4066Sahrens } else { 4669fa9e4066Sahrens /* 4670fa9e4066Sahrens * This block was freed while we waited for the read to 4671fa9e4066Sahrens * complete. It has been removed from the hash table and 4672fa9e4066Sahrens * moved to the anonymous state (so that it won't show up 4673fa9e4066Sahrens * in the cache). 4674fa9e4066Sahrens */ 467589c86e32SChris Williamson ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 467689c86e32SChris Williamson freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4677fa9e4066Sahrens } 4678fa9e4066Sahrens 4679fa9e4066Sahrens /* execute each callback and free its structure */ 4680fa9e4066Sahrens while ((acb = callback_list) != NULL) { 4681fa9e4066Sahrens if (acb->acb_done) 4682fa9e4066Sahrens acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4683fa9e4066Sahrens 4684fa9e4066Sahrens if (acb->acb_zio_dummy != NULL) { 4685fa9e4066Sahrens acb->acb_zio_dummy->io_error = zio->io_error; 4686fa9e4066Sahrens zio_nowait(acb->acb_zio_dummy); 4687fa9e4066Sahrens } 4688fa9e4066Sahrens 4689fa9e4066Sahrens callback_list = acb->acb_next; 4690fa9e4066Sahrens kmem_free(acb, sizeof (arc_callback_t)); 4691fa9e4066Sahrens } 4692fa9e4066Sahrens 4693fa9e4066Sahrens if (freeable) 4694ea8dc4b6Seschrock arc_hdr_destroy(hdr); 4695fa9e4066Sahrens } 4696fa9e4066Sahrens 4697fa9e4066Sahrens /* 4698fc98fea5SBart Coddens * "Read" the block at the specified DVA (in bp) via the 4699fa9e4066Sahrens * cache. If the block is found in the cache, invoke the provided 4700fa9e4066Sahrens * callback immediately and return. Note that the `zio' parameter 4701fa9e4066Sahrens * in the callback will be NULL in this case, since no IO was 4702fa9e4066Sahrens * required. If the block is not in the cache pass the read request 4703fa9e4066Sahrens * on to the spa with a substitute callback function, so that the 4704fa9e4066Sahrens * requested block will be added to the cache. 4705fa9e4066Sahrens * 4706fa9e4066Sahrens * If a read request arrives for a block that has a read in-progress, 4707fa9e4066Sahrens * either wait for the in-progress read to complete (and return the 4708fa9e4066Sahrens * results); or, if this is a read with a "done" func, add a record 4709fa9e4066Sahrens * to the read to invoke the "done" func when the read completes, 4710fa9e4066Sahrens * and return; or just return. 4711fa9e4066Sahrens * 4712fa9e4066Sahrens * arc_read_done() will invoke all the requested "done" functions 4713fa9e4066Sahrens * for readers of this block. 4714fa9e4066Sahrens */ 4715fa9e4066Sahrens int 47161b912ec7SGeorge Wilson arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 47177adb730bSGeorge Wilson void *private, zio_priority_t priority, int zio_flags, 47187adb730bSGeorge Wilson arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4719fa9e4066Sahrens { 47205d7b4d43SMatthew Ahrens arc_buf_hdr_t *hdr = NULL; 47215d7b4d43SMatthew Ahrens kmutex_t *hash_lock = NULL; 4722fa94a07fSbrendan zio_t *rzio; 4723e9103aaeSGarrett D'Amore uint64_t guid = spa_load_guid(spa); 4724*5602294fSDan Kimmel boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0; 4725fa9e4066Sahrens 47265d7b4d43SMatthew Ahrens ASSERT(!BP_IS_EMBEDDED(bp) || 47275d7b4d43SMatthew Ahrens BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 47285d7b4d43SMatthew Ahrens 4729fa9e4066Sahrens top: 47305d7b4d43SMatthew Ahrens if (!BP_IS_EMBEDDED(bp)) { 47315d7b4d43SMatthew Ahrens /* 47325d7b4d43SMatthew Ahrens * Embedded BP's have no DVA and require no I/O to "read". 47335d7b4d43SMatthew Ahrens * Create an anonymous arc buf to back it. 47345d7b4d43SMatthew Ahrens */ 47355d7b4d43SMatthew Ahrens hdr = buf_hash_find(guid, bp, &hash_lock); 47365d7b4d43SMatthew Ahrens } 47375d7b4d43SMatthew Ahrens 4738dcbf3bd6SGeorge Wilson if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pdata != NULL) { 4739dcbf3bd6SGeorge Wilson arc_buf_t *buf = NULL; 47407adb730bSGeorge Wilson *arc_flags |= ARC_FLAG_CACHED; 474113506d1eSmaybee 4742fa9e4066Sahrens if (HDR_IO_IN_PROGRESS(hdr)) { 474313506d1eSmaybee 4744cf6106c8SMatthew Ahrens if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 4745cf6106c8SMatthew Ahrens priority == ZIO_PRIORITY_SYNC_READ) { 4746cf6106c8SMatthew Ahrens /* 4747cf6106c8SMatthew Ahrens * This sync read must wait for an 4748cf6106c8SMatthew Ahrens * in-progress async read (e.g. a predictive 4749cf6106c8SMatthew Ahrens * prefetch). Async reads are queued 4750cf6106c8SMatthew Ahrens * separately at the vdev_queue layer, so 4751cf6106c8SMatthew Ahrens * this is a form of priority inversion. 4752cf6106c8SMatthew Ahrens * Ideally, we would "inherit" the demand 4753cf6106c8SMatthew Ahrens * i/o's priority by moving the i/o from 4754cf6106c8SMatthew Ahrens * the async queue to the synchronous queue, 4755cf6106c8SMatthew Ahrens * but there is currently no mechanism to do 4756cf6106c8SMatthew Ahrens * so. Track this so that we can evaluate 4757cf6106c8SMatthew Ahrens * the magnitude of this potential performance 4758cf6106c8SMatthew Ahrens * problem. 4759cf6106c8SMatthew Ahrens * 4760cf6106c8SMatthew Ahrens * Note that if the prefetch i/o is already 4761cf6106c8SMatthew Ahrens * active (has been issued to the device), 4762cf6106c8SMatthew Ahrens * the prefetch improved performance, because 4763cf6106c8SMatthew Ahrens * we issued it sooner than we would have 4764cf6106c8SMatthew Ahrens * without the prefetch. 4765cf6106c8SMatthew Ahrens */ 4766cf6106c8SMatthew Ahrens DTRACE_PROBE1(arc__sync__wait__for__async, 4767cf6106c8SMatthew Ahrens arc_buf_hdr_t *, hdr); 4768cf6106c8SMatthew Ahrens ARCSTAT_BUMP(arcstat_sync_wait_for_async); 4769cf6106c8SMatthew Ahrens } 4770cf6106c8SMatthew Ahrens if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4771dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, 4772dcbf3bd6SGeorge Wilson ARC_FLAG_PREDICTIVE_PREFETCH); 4773cf6106c8SMatthew Ahrens } 4774cf6106c8SMatthew Ahrens 47757adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_WAIT) { 477689c86e32SChris Williamson cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 477713506d1eSmaybee mutex_exit(hash_lock); 477813506d1eSmaybee goto top; 477913506d1eSmaybee } 47807adb730bSGeorge Wilson ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 478113506d1eSmaybee 478213506d1eSmaybee if (done) { 4783cf6106c8SMatthew Ahrens arc_callback_t *acb = NULL; 4784fa9e4066Sahrens 4785fa9e4066Sahrens acb = kmem_zalloc(sizeof (arc_callback_t), 4786fa9e4066Sahrens KM_SLEEP); 4787fa9e4066Sahrens acb->acb_done = done; 4788fa9e4066Sahrens acb->acb_private = private; 4789*5602294fSDan Kimmel acb->acb_compressed = compressed_read; 4790fa9e4066Sahrens if (pio != NULL) 4791fa9e4066Sahrens acb->acb_zio_dummy = zio_null(pio, 4792a3f829aeSBill Moore spa, NULL, NULL, NULL, zio_flags); 4793fa9e4066Sahrens 4794dcbf3bd6SGeorge Wilson ASSERT3P(acb->acb_done, !=, NULL); 479589c86e32SChris Williamson acb->acb_next = hdr->b_l1hdr.b_acb; 479689c86e32SChris Williamson hdr->b_l1hdr.b_acb = acb; 4797fa9e4066Sahrens mutex_exit(hash_lock); 4798fa9e4066Sahrens return (0); 4799fa9e4066Sahrens } 4800fa9e4066Sahrens mutex_exit(hash_lock); 4801fa9e4066Sahrens return (0); 4802fa9e4066Sahrens } 4803fa9e4066Sahrens 480489c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_state == arc_mru || 480589c86e32SChris Williamson hdr->b_l1hdr.b_state == arc_mfu); 4806fa9e4066Sahrens 4807ea8dc4b6Seschrock if (done) { 4808cf6106c8SMatthew Ahrens if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4809cf6106c8SMatthew Ahrens /* 4810cf6106c8SMatthew Ahrens * This is a demand read which does not have to 4811cf6106c8SMatthew Ahrens * wait for i/o because we did a predictive 4812cf6106c8SMatthew Ahrens * prefetch i/o for it, which has completed. 4813cf6106c8SMatthew Ahrens */ 4814cf6106c8SMatthew Ahrens DTRACE_PROBE1( 4815cf6106c8SMatthew Ahrens arc__demand__hit__predictive__prefetch, 4816cf6106c8SMatthew Ahrens arc_buf_hdr_t *, hdr); 4817cf6106c8SMatthew Ahrens ARCSTAT_BUMP( 4818cf6106c8SMatthew Ahrens arcstat_demand_hit_predictive_prefetch); 4819dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, 4820dcbf3bd6SGeorge Wilson ARC_FLAG_PREDICTIVE_PREFETCH); 4821cf6106c8SMatthew Ahrens } 4822dcbf3bd6SGeorge Wilson ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); 4823dcbf3bd6SGeorge Wilson 4824*5602294fSDan Kimmel /* Get a buf with the desired data in it. */ 4825*5602294fSDan Kimmel VERIFY0(arc_buf_alloc_impl(hdr, private, 4826*5602294fSDan Kimmel compressed_read, B_TRUE, &buf)); 48277adb730bSGeorge Wilson } else if (*arc_flags & ARC_FLAG_PREFETCH && 482889c86e32SChris Williamson refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4829dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 4830fa9e4066Sahrens } 4831fa9e4066Sahrens DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 483244eda4d7Smaybee arc_access(hdr, hash_lock); 48337adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_L2CACHE) 4834dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 483544eda4d7Smaybee mutex_exit(hash_lock); 483644cb6abcSbmc ARCSTAT_BUMP(arcstat_hits); 483789c86e32SChris Williamson ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 483889c86e32SChris Williamson demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 483944cb6abcSbmc data, metadata, hits); 484044cb6abcSbmc 4841fa9e4066Sahrens if (done) 4842fa9e4066Sahrens done(NULL, buf, private); 4843fa9e4066Sahrens } else { 4844dcbf3bd6SGeorge Wilson uint64_t lsize = BP_GET_LSIZE(bp); 4845dcbf3bd6SGeorge Wilson uint64_t psize = BP_GET_PSIZE(bp); 48465d7b4d43SMatthew Ahrens arc_callback_t *acb; 48473a737e0dSbrendan vdev_t *vd = NULL; 4848d5285caeSGeorge Wilson uint64_t addr = 0; 48495a98e54bSBrendan Gregg - Sun Microsystems boolean_t devw = B_FALSE; 4850dcbf3bd6SGeorge Wilson uint64_t size; 4851fa9e4066Sahrens 4852fa9e4066Sahrens if (hdr == NULL) { 4853fa9e4066Sahrens /* this block is not in the cache */ 48545d7b4d43SMatthew Ahrens arc_buf_hdr_t *exists = NULL; 4855ad23a2dbSjohansen arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4856dcbf3bd6SGeorge Wilson hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 4857dcbf3bd6SGeorge Wilson BP_GET_COMPRESS(bp), type); 4858dcbf3bd6SGeorge Wilson 48595d7b4d43SMatthew Ahrens if (!BP_IS_EMBEDDED(bp)) { 48605d7b4d43SMatthew Ahrens hdr->b_dva = *BP_IDENTITY(bp); 48615d7b4d43SMatthew Ahrens hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 48625d7b4d43SMatthew Ahrens exists = buf_hash_insert(hdr, &hash_lock); 48635d7b4d43SMatthew Ahrens } 48645d7b4d43SMatthew Ahrens if (exists != NULL) { 4865fa9e4066Sahrens /* somebody beat us to the hash insert */ 4866fa9e4066Sahrens mutex_exit(hash_lock); 48673f9d6ad7SLin Ling buf_discard_identity(hdr); 4868dcbf3bd6SGeorge Wilson arc_hdr_destroy(hdr); 4869fa9e4066Sahrens goto top; /* restart the IO request */ 4870fa9e4066Sahrens } 4871fa9e4066Sahrens } else { 487289c86e32SChris Williamson /* 487389c86e32SChris Williamson * This block is in the ghost cache. If it was L2-only 487489c86e32SChris Williamson * (and thus didn't have an L1 hdr), we realloc the 487589c86e32SChris Williamson * header to add an L1 hdr. 487689c86e32SChris Williamson */ 487789c86e32SChris Williamson if (!HDR_HAS_L1HDR(hdr)) { 487889c86e32SChris Williamson hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 487989c86e32SChris Williamson hdr_full_cache); 488089c86e32SChris Williamson } 4881dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 488289c86e32SChris Williamson ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4883ea8dc4b6Seschrock ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 488489c86e32SChris Williamson ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4885244781f1SPrakash Surya ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 4886*5602294fSDan Kimmel ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 488713506d1eSmaybee 4888cf6106c8SMatthew Ahrens /* 4889dcbf3bd6SGeorge Wilson * This is a delicate dance that we play here. 4890dcbf3bd6SGeorge Wilson * This hdr is in the ghost list so we access it 4891dcbf3bd6SGeorge Wilson * to move it out of the ghost list before we 4892dcbf3bd6SGeorge Wilson * initiate the read. If it's a prefetch then 4893dcbf3bd6SGeorge Wilson * it won't have a callback so we'll remove the 4894dcbf3bd6SGeorge Wilson * reference that arc_buf_alloc_impl() created. We 4895dcbf3bd6SGeorge Wilson * do this after we've called arc_access() to 4896dcbf3bd6SGeorge Wilson * avoid hitting an assert in remove_reference(). 4897cf6106c8SMatthew Ahrens */ 48987e453561SWilliam Gorrell arc_access(hdr, hash_lock); 4899dcbf3bd6SGeorge Wilson arc_hdr_alloc_pdata(hdr); 4900dcbf3bd6SGeorge Wilson } 4901dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 4902dcbf3bd6SGeorge Wilson size = arc_hdr_size(hdr); 4903dcbf3bd6SGeorge Wilson 4904dcbf3bd6SGeorge Wilson /* 4905dcbf3bd6SGeorge Wilson * If compression is enabled on the hdr, then will do 4906dcbf3bd6SGeorge Wilson * RAW I/O and will store the compressed data in the hdr's 4907dcbf3bd6SGeorge Wilson * data block. Otherwise, the hdr's data block will contain 4908dcbf3bd6SGeorge Wilson * the uncompressed data. 4909dcbf3bd6SGeorge Wilson */ 4910dcbf3bd6SGeorge Wilson if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 4911dcbf3bd6SGeorge Wilson zio_flags |= ZIO_FLAG_RAW; 4912fa9e4066Sahrens } 4913fa9e4066Sahrens 4914dcbf3bd6SGeorge Wilson if (*arc_flags & ARC_FLAG_PREFETCH) 4915dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 4916dcbf3bd6SGeorge Wilson if (*arc_flags & ARC_FLAG_L2CACHE) 4917dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 4918dcbf3bd6SGeorge Wilson if (BP_GET_LEVEL(bp) > 0) 4919dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); 4920cf6106c8SMatthew Ahrens if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 4921dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); 492289c86e32SChris Williamson ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 49235614b00aSWilliam Gorrell 4924fa9e4066Sahrens acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4925fa9e4066Sahrens acb->acb_done = done; 4926fa9e4066Sahrens acb->acb_private = private; 4927*5602294fSDan Kimmel acb->acb_compressed = compressed_read; 4928fa9e4066Sahrens 4929dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 493089c86e32SChris Williamson hdr->b_l1hdr.b_acb = acb; 4931dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 4932fa9e4066Sahrens 493389c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr) && 493489c86e32SChris Williamson (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 493589c86e32SChris Williamson devw = hdr->b_l2hdr.b_dev->l2ad_writing; 493689c86e32SChris Williamson addr = hdr->b_l2hdr.b_daddr; 4937e14bb325SJeff Bonwick /* 4938e14bb325SJeff Bonwick * Lock out device removal. 4939e14bb325SJeff Bonwick */ 4940e14bb325SJeff Bonwick if (vdev_is_dead(vd) || 4941e14bb325SJeff Bonwick !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 4942e14bb325SJeff Bonwick vd = NULL; 49433a737e0dSbrendan } 49443a737e0dSbrendan 4945dcbf3bd6SGeorge Wilson if (priority == ZIO_PRIORITY_ASYNC_READ) 4946dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 4947dcbf3bd6SGeorge Wilson else 4948dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 4949dcbf3bd6SGeorge Wilson 49505d7b4d43SMatthew Ahrens if (hash_lock != NULL) 49515d7b4d43SMatthew Ahrens mutex_exit(hash_lock); 49523a737e0dSbrendan 49533e30c24aSWill Andrews /* 49543e30c24aSWill Andrews * At this point, we have a level 1 cache miss. Try again in 49553e30c24aSWill Andrews * L2ARC if possible. 49563e30c24aSWill Andrews */ 4957dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); 4958dcbf3bd6SGeorge Wilson 49595c28183bSBrendan Gregg - Sun Microsystems DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 4960dcbf3bd6SGeorge Wilson uint64_t, lsize, zbookmark_phys_t *, zb); 496144cb6abcSbmc ARCSTAT_BUMP(arcstat_misses); 496289c86e32SChris Williamson ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 496389c86e32SChris Williamson demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 496444cb6abcSbmc data, metadata, misses); 4965ea8dc4b6Seschrock 49665a98e54bSBrendan Gregg - Sun Microsystems if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 4967fa94a07fSbrendan /* 4968fa94a07fSbrendan * Read from the L2ARC if the following are true: 49693a737e0dSbrendan * 1. The L2ARC vdev was previously cached. 49703a737e0dSbrendan * 2. This buffer still has L2ARC metadata. 49713a737e0dSbrendan * 3. This buffer isn't currently writing to the L2ARC. 49723a737e0dSbrendan * 4. The L2ARC entry wasn't evicted, which may 49733a737e0dSbrendan * also have invalidated the vdev. 49745a98e54bSBrendan Gregg - Sun Microsystems * 5. This isn't prefetch and l2arc_noprefetch is set. 4975fa94a07fSbrendan */ 497689c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr) && 49775a98e54bSBrendan Gregg - Sun Microsystems !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 49785a98e54bSBrendan Gregg - Sun Microsystems !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 4979fa94a07fSbrendan l2arc_read_callback_t *cb; 4980fa94a07fSbrendan 4981c5904d13Seschrock DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 4982c5904d13Seschrock ARCSTAT_BUMP(arcstat_l2_hits); 4983c5904d13Seschrock 4984fa94a07fSbrendan cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 4985fa94a07fSbrendan KM_SLEEP); 4986dcbf3bd6SGeorge Wilson cb->l2rcb_hdr = hdr; 4987fa94a07fSbrendan cb->l2rcb_bp = *bp; 4988fa94a07fSbrendan cb->l2rcb_zb = *zb; 49893baa08fcSek cb->l2rcb_flags = zio_flags; 4990fa94a07fSbrendan 4991d5285caeSGeorge Wilson ASSERT(addr >= VDEV_LABEL_START_SIZE && 4992dcbf3bd6SGeorge Wilson addr + lsize < vd->vdev_psize - 4993d5285caeSGeorge Wilson VDEV_LABEL_END_SIZE); 4994d5285caeSGeorge Wilson 4995fa94a07fSbrendan /* 4996e14bb325SJeff Bonwick * l2arc read. The SCL_L2ARC lock will be 4997e14bb325SJeff Bonwick * released by l2arc_read_done(). 4998aad02571SSaso Kiselkov * Issue a null zio if the underlying buffer 4999aad02571SSaso Kiselkov * was squashed to zero size by compression. 5000fa94a07fSbrendan */ 5001dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_COMPRESS(hdr), !=, 5002dcbf3bd6SGeorge Wilson ZIO_COMPRESS_EMPTY); 5003dcbf3bd6SGeorge Wilson rzio = zio_read_phys(pio, vd, addr, 5004dcbf3bd6SGeorge Wilson size, hdr->b_l1hdr.b_pdata, 5005dcbf3bd6SGeorge Wilson ZIO_CHECKSUM_OFF, 5006dcbf3bd6SGeorge Wilson l2arc_read_done, cb, priority, 5007dcbf3bd6SGeorge Wilson zio_flags | ZIO_FLAG_DONT_CACHE | 5008dcbf3bd6SGeorge Wilson ZIO_FLAG_CANFAIL | 5009dcbf3bd6SGeorge Wilson ZIO_FLAG_DONT_PROPAGATE | 5010dcbf3bd6SGeorge Wilson ZIO_FLAG_DONT_RETRY, B_FALSE); 5011fa94a07fSbrendan DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 5012fa94a07fSbrendan zio_t *, rzio); 5013dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_l2_read_bytes, size); 5014fa94a07fSbrendan 50157adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_NOWAIT) { 50163a737e0dSbrendan zio_nowait(rzio); 50173a737e0dSbrendan return (0); 50183a737e0dSbrendan } 5019fa94a07fSbrendan 50207adb730bSGeorge Wilson ASSERT(*arc_flags & ARC_FLAG_WAIT); 50213a737e0dSbrendan if (zio_wait(rzio) == 0) 50223a737e0dSbrendan return (0); 50233a737e0dSbrendan 50243a737e0dSbrendan /* l2arc read error; goto zio_read() */ 5025fa94a07fSbrendan } else { 5026fa94a07fSbrendan DTRACE_PROBE1(l2arc__miss, 5027fa94a07fSbrendan arc_buf_hdr_t *, hdr); 5028fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_misses); 5029fa94a07fSbrendan if (HDR_L2_WRITING(hdr)) 5030fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_rw_clash); 5031e14bb325SJeff Bonwick spa_config_exit(spa, SCL_L2ARC, vd); 5032fa94a07fSbrendan } 50335a98e54bSBrendan Gregg - Sun Microsystems } else { 503476a25fafSBill Moore if (vd != NULL) 503576a25fafSBill Moore spa_config_exit(spa, SCL_L2ARC, vd); 50365a98e54bSBrendan Gregg - Sun Microsystems if (l2arc_ndev != 0) { 50375a98e54bSBrendan Gregg - Sun Microsystems DTRACE_PROBE1(l2arc__miss, 50385a98e54bSBrendan Gregg - Sun Microsystems arc_buf_hdr_t *, hdr); 50395a98e54bSBrendan Gregg - Sun Microsystems ARCSTAT_BUMP(arcstat_l2_misses); 50405a98e54bSBrendan Gregg - Sun Microsystems } 5041fa94a07fSbrendan } 5042c5904d13Seschrock 5043dcbf3bd6SGeorge Wilson rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pdata, size, 5044dcbf3bd6SGeorge Wilson arc_read_done, hdr, priority, zio_flags, zb); 5045fa9e4066Sahrens 50467adb730bSGeorge Wilson if (*arc_flags & ARC_FLAG_WAIT) 5047fa9e4066Sahrens return (zio_wait(rzio)); 5048fa9e4066Sahrens 50497adb730bSGeorge Wilson ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5050fa9e4066Sahrens zio_nowait(rzio); 5051fa9e4066Sahrens } 5052fa9e4066Sahrens return (0); 5053fa9e4066Sahrens } 5054fa9e4066Sahrens 50556e6d5868SMatthew Ahrens /* 50566e6d5868SMatthew Ahrens * Notify the arc that a block was freed, and thus will never be used again. 50576e6d5868SMatthew Ahrens */ 50586e6d5868SMatthew Ahrens void 50596e6d5868SMatthew Ahrens arc_freed(spa_t *spa, const blkptr_t *bp) 50606e6d5868SMatthew Ahrens { 50616e6d5868SMatthew Ahrens arc_buf_hdr_t *hdr; 50626e6d5868SMatthew Ahrens kmutex_t *hash_lock; 50636e6d5868SMatthew Ahrens uint64_t guid = spa_load_guid(spa); 50646e6d5868SMatthew Ahrens 50655d7b4d43SMatthew Ahrens ASSERT(!BP_IS_EMBEDDED(bp)); 50665d7b4d43SMatthew Ahrens 50675d7b4d43SMatthew Ahrens hdr = buf_hash_find(guid, bp, &hash_lock); 50686e6d5868SMatthew Ahrens if (hdr == NULL) 50696e6d5868SMatthew Ahrens return; 50706e6d5868SMatthew Ahrens 5071dcbf3bd6SGeorge Wilson /* 5072dcbf3bd6SGeorge Wilson * We might be trying to free a block that is still doing I/O 5073dcbf3bd6SGeorge Wilson * (i.e. prefetch) or has a reference (i.e. a dedup-ed, 5074dcbf3bd6SGeorge Wilson * dmu_sync-ed block). If this block is being prefetched, then it 5075dcbf3bd6SGeorge Wilson * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr 5076dcbf3bd6SGeorge Wilson * until the I/O completes. A block may also have a reference if it is 5077dcbf3bd6SGeorge Wilson * part of a dedup-ed, dmu_synced write. The dmu_sync() function would 5078dcbf3bd6SGeorge Wilson * have written the new block to its final resting place on disk but 5079dcbf3bd6SGeorge Wilson * without the dedup flag set. This would have left the hdr in the MRU 5080dcbf3bd6SGeorge Wilson * state and discoverable. When the txg finally syncs it detects that 5081dcbf3bd6SGeorge Wilson * the block was overridden in open context and issues an override I/O. 5082dcbf3bd6SGeorge Wilson * Since this is a dedup block, the override I/O will determine if the 5083dcbf3bd6SGeorge Wilson * block is already in the DDT. If so, then it will replace the io_bp 5084dcbf3bd6SGeorge Wilson * with the bp from the DDT and allow the I/O to finish. When the I/O 5085dcbf3bd6SGeorge Wilson * reaches the done callback, dbuf_write_override_done, it will 5086dcbf3bd6SGeorge Wilson * check to see if the io_bp and io_bp_override are identical. 5087dcbf3bd6SGeorge Wilson * If they are not, then it indicates that the bp was replaced with 5088dcbf3bd6SGeorge Wilson * the bp in the DDT and the override bp is freed. This allows 5089dcbf3bd6SGeorge Wilson * us to arrive here with a reference on a block that is being 5090dcbf3bd6SGeorge Wilson * freed. So if we have an I/O in progress, or a reference to 5091dcbf3bd6SGeorge Wilson * this hdr, then we don't destroy the hdr. 5092dcbf3bd6SGeorge Wilson */ 5093dcbf3bd6SGeorge Wilson if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && 5094dcbf3bd6SGeorge Wilson refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { 5095dcbf3bd6SGeorge Wilson arc_change_state(arc_anon, hdr, hash_lock); 5096dcbf3bd6SGeorge Wilson arc_hdr_destroy(hdr); 50976e6d5868SMatthew Ahrens mutex_exit(hash_lock); 5098bbfa8ea8SMatthew Ahrens } else { 5099dcbf3bd6SGeorge Wilson mutex_exit(hash_lock); 5100ea8dc4b6Seschrock } 5101dd6ef538Smaybee 5102ea8dc4b6Seschrock } 5103ea8dc4b6Seschrock 5104fa9e4066Sahrens /* 51053e30c24aSWill Andrews * Release this buffer from the cache, making it an anonymous buffer. This 51063e30c24aSWill Andrews * must be done after a read and prior to modifying the buffer contents. 5107fa9e4066Sahrens * If the buffer has more than one reference, we must make 5108088f3894Sahrens * a new hdr for the buffer. 5109fa9e4066Sahrens */ 5110fa9e4066Sahrens void 5111fa9e4066Sahrens arc_release(arc_buf_t *buf, void *tag) 5112fa9e4066Sahrens { 511389c86e32SChris Williamson arc_buf_hdr_t *hdr = buf->b_hdr; 5114fa9e4066Sahrens 51153f9d6ad7SLin Ling /* 51163f9d6ad7SLin Ling * It would be nice to assert that if it's DMU metadata (level > 51173f9d6ad7SLin Ling * 0 || it's the dnode file), then it must be syncing context. 51183f9d6ad7SLin Ling * But we don't know that information at this level. 51193f9d6ad7SLin Ling */ 51203f9d6ad7SLin Ling 51213f9d6ad7SLin Ling mutex_enter(&buf->b_evict_lock); 51226f83844dSMark Maybee 5123244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 5124244781f1SPrakash Surya 512589c86e32SChris Williamson /* 512689c86e32SChris Williamson * We don't grab the hash lock prior to this check, because if 512789c86e32SChris Williamson * the buffer's header is in the arc_anon state, it won't be 512889c86e32SChris Williamson * linked into the hash table. 512989c86e32SChris Williamson */ 513089c86e32SChris Williamson if (hdr->b_l1hdr.b_state == arc_anon) { 513189c86e32SChris Williamson mutex_exit(&buf->b_evict_lock); 513289c86e32SChris Williamson ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 513389c86e32SChris Williamson ASSERT(!HDR_IN_HASH_TABLE(hdr)); 513489c86e32SChris Williamson ASSERT(!HDR_HAS_L2HDR(hdr)); 5135dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 5136fa9e4066Sahrens 5137dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 513889c86e32SChris Williamson ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 513989c86e32SChris Williamson ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 514089c86e32SChris Williamson 514189c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = 0; 5142dcbf3bd6SGeorge Wilson 5143dcbf3bd6SGeorge Wilson /* 5144dcbf3bd6SGeorge Wilson * If the buf is being overridden then it may already 5145dcbf3bd6SGeorge Wilson * have a hdr that is not empty. 5146dcbf3bd6SGeorge Wilson */ 5147dcbf3bd6SGeorge Wilson buf_discard_identity(hdr); 514889c86e32SChris Williamson arc_buf_thaw(buf); 514989c86e32SChris Williamson 515089c86e32SChris Williamson return; 5151fa9e4066Sahrens } 5152fa9e4066Sahrens 515389c86e32SChris Williamson kmutex_t *hash_lock = HDR_LOCK(hdr); 515489c86e32SChris Williamson mutex_enter(hash_lock); 515589c86e32SChris Williamson 515689c86e32SChris Williamson /* 515789c86e32SChris Williamson * This assignment is only valid as long as the hash_lock is 515889c86e32SChris Williamson * held, we must be careful not to reference state or the 515989c86e32SChris Williamson * b_state field after dropping the lock. 516089c86e32SChris Williamson */ 516189c86e32SChris Williamson arc_state_t *state = hdr->b_l1hdr.b_state; 516289c86e32SChris Williamson ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 516389c86e32SChris Williamson ASSERT3P(state, !=, arc_anon); 516489c86e32SChris Williamson 516589c86e32SChris Williamson /* this buffer is not on any list */ 5166*5602294fSDan Kimmel ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0); 516789c86e32SChris Williamson 516889c86e32SChris Williamson if (HDR_HAS_L2HDR(hdr)) { 516989c86e32SChris Williamson mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5170244781f1SPrakash Surya 5171244781f1SPrakash Surya /* 5172a52fc310SPrakash Surya * We have to recheck this conditional again now that 5173a52fc310SPrakash Surya * we're holding the l2ad_mtx to prevent a race with 5174a52fc310SPrakash Surya * another thread which might be concurrently calling 5175a52fc310SPrakash Surya * l2arc_evict(). In that case, l2arc_evict() might have 5176a52fc310SPrakash Surya * destroyed the header's L2 portion as we were waiting 5177a52fc310SPrakash Surya * to acquire the l2ad_mtx. 5178244781f1SPrakash Surya */ 5179a52fc310SPrakash Surya if (HDR_HAS_L2HDR(hdr)) 5180a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(hdr); 5181244781f1SPrakash Surya 518289c86e32SChris Williamson mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 51836f83844dSMark Maybee } 51846f83844dSMark Maybee 5185ea8dc4b6Seschrock /* 5186ea8dc4b6Seschrock * Do we have more than one buf? 5187ea8dc4b6Seschrock */ 5188dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_bufcnt > 1) { 5189fa9e4066Sahrens arc_buf_hdr_t *nhdr; 5190ac05c741SMark Maybee uint64_t spa = hdr->b_spa; 5191dcbf3bd6SGeorge Wilson uint64_t psize = HDR_GET_PSIZE(hdr); 5192dcbf3bd6SGeorge Wilson uint64_t lsize = HDR_GET_LSIZE(hdr); 5193dcbf3bd6SGeorge Wilson enum zio_compress compress = HDR_GET_COMPRESS(hdr); 519489c86e32SChris Williamson arc_buf_contents_t type = arc_buf_type(hdr); 5195dcbf3bd6SGeorge Wilson VERIFY3U(hdr->b_type, ==, type); 5196fa9e4066Sahrens 519789c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 5198dcbf3bd6SGeorge Wilson (void) remove_reference(hdr, hash_lock, tag); 5199dcbf3bd6SGeorge Wilson 5200*5602294fSDan Kimmel if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) { 5201dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5202dcbf3bd6SGeorge Wilson ASSERT(ARC_BUF_LAST(buf)); 5203dcbf3bd6SGeorge Wilson } 5204dcbf3bd6SGeorge Wilson 5205fa9e4066Sahrens /* 52063f9d6ad7SLin Ling * Pull the data off of this hdr and attach it to 5207dcbf3bd6SGeorge Wilson * a new anonymous hdr. Also find the last buffer 5208dcbf3bd6SGeorge Wilson * in the hdr's buffer list. 5209fa9e4066Sahrens */ 5210*5602294fSDan Kimmel arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 5211dcbf3bd6SGeorge Wilson ASSERT3P(lastbuf, !=, NULL); 5212dcbf3bd6SGeorge Wilson 5213dcbf3bd6SGeorge Wilson /* 5214dcbf3bd6SGeorge Wilson * If the current arc_buf_t and the hdr are sharing their data 5215*5602294fSDan Kimmel * buffer, then we must stop sharing that block. 5216dcbf3bd6SGeorge Wilson */ 5217dcbf3bd6SGeorge Wilson if (arc_buf_is_shared(buf)) { 5218dcbf3bd6SGeorge Wilson VERIFY(!arc_buf_is_shared(lastbuf)); 5219ea8dc4b6Seschrock 5220dcbf3bd6SGeorge Wilson /* 5221dcbf3bd6SGeorge Wilson * First, sever the block sharing relationship between 5222*5602294fSDan Kimmel * buf and the arc_buf_hdr_t. 5223dcbf3bd6SGeorge Wilson */ 5224dcbf3bd6SGeorge Wilson arc_unshare_buf(hdr, buf); 5225*5602294fSDan Kimmel 5226*5602294fSDan Kimmel /* 5227*5602294fSDan Kimmel * Now we need to recreate the hdr's b_pdata. Since we 5228*5602294fSDan Kimmel * have lastbuf handy, we try to share with it, but if 5229*5602294fSDan Kimmel * we can't then we allocate a new b_pdata and copy the 5230*5602294fSDan Kimmel * data from buf into it. 5231*5602294fSDan Kimmel */ 5232*5602294fSDan Kimmel if (arc_can_share(hdr, lastbuf)) { 5233*5602294fSDan Kimmel arc_share_buf(hdr, lastbuf); 5234*5602294fSDan Kimmel } else { 5235*5602294fSDan Kimmel arc_hdr_alloc_pdata(hdr); 5236*5602294fSDan Kimmel bcopy(buf->b_data, hdr->b_l1hdr.b_pdata, psize); 5237*5602294fSDan Kimmel } 5238dcbf3bd6SGeorge Wilson VERIFY3P(lastbuf->b_data, !=, NULL); 5239dcbf3bd6SGeorge Wilson } else if (HDR_SHARED_DATA(hdr)) { 5240*5602294fSDan Kimmel /* 5241*5602294fSDan Kimmel * Uncompressed shared buffers are always at the end 5242*5602294fSDan Kimmel * of the list. Compressed buffers don't have the 5243*5602294fSDan Kimmel * same requirements. This makes it hard to 5244*5602294fSDan Kimmel * simply assert that the lastbuf is shared so 5245*5602294fSDan Kimmel * we rely on the hdr's compression flags to determine 5246*5602294fSDan Kimmel * if we have a compressed, shared buffer. 5247*5602294fSDan Kimmel */ 5248*5602294fSDan Kimmel ASSERT(arc_buf_is_shared(lastbuf) || 5249*5602294fSDan Kimmel HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 5250*5602294fSDan Kimmel ASSERT(!ARC_BUF_SHARED(buf)); 5251dcbf3bd6SGeorge Wilson } 5252dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 525389c86e32SChris Williamson ASSERT3P(state, !=, arc_l2c_only); 52542fd872a7SPrakash Surya 5255dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_size, 5256*5602294fSDan Kimmel arc_buf_size(buf), buf); 52572fd872a7SPrakash Surya 525889c86e32SChris Williamson if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 525989c86e32SChris Williamson ASSERT3P(state, !=, arc_l2c_only); 5260dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&state->arcs_esize[type], 5261*5602294fSDan Kimmel arc_buf_size(buf), buf); 5262ea8dc4b6Seschrock } 52639253d63dSGeorge Wilson 5264dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_bufcnt -= 1; 5265c717a561Smaybee arc_cksum_verify(buf); 5266cd1c8b85SMatthew Ahrens arc_buf_unwatch(buf); 5267ea8dc4b6Seschrock 5268fa9e4066Sahrens mutex_exit(hash_lock); 5269fa9e4066Sahrens 5270dcbf3bd6SGeorge Wilson /* 5271dcbf3bd6SGeorge Wilson * Allocate a new hdr. The new hdr will contain a b_pdata 5272dcbf3bd6SGeorge Wilson * buffer which will be freed in arc_write(). 5273dcbf3bd6SGeorge Wilson */ 5274dcbf3bd6SGeorge Wilson nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type); 5275dcbf3bd6SGeorge Wilson ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); 5276dcbf3bd6SGeorge Wilson ASSERT0(nhdr->b_l1hdr.b_bufcnt); 5277dcbf3bd6SGeorge Wilson ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt)); 5278dcbf3bd6SGeorge Wilson VERIFY3U(nhdr->b_type, ==, type); 5279dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(nhdr)); 528089c86e32SChris Williamson 528189c86e32SChris Williamson nhdr->b_l1hdr.b_buf = buf; 5282dcbf3bd6SGeorge Wilson nhdr->b_l1hdr.b_bufcnt = 1; 528389c86e32SChris Williamson (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 5284af2c4821Smaybee buf->b_hdr = nhdr; 5285dcbf3bd6SGeorge Wilson 52863f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 5287dcbf3bd6SGeorge Wilson (void) refcount_add_many(&arc_anon->arcs_size, 5288*5602294fSDan Kimmel arc_buf_size(buf), buf); 5289fa9e4066Sahrens } else { 52903f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 529189c86e32SChris Williamson ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 5292244781f1SPrakash Surya /* protected by hash lock, or hdr is on arc_anon */ 5293244781f1SPrakash Surya ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5294fa9e4066Sahrens ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 529589c86e32SChris Williamson arc_change_state(arc_anon, hdr, hash_lock); 529689c86e32SChris Williamson hdr->b_l1hdr.b_arc_access = 0; 529789c86e32SChris Williamson mutex_exit(hash_lock); 5298fa94a07fSbrendan 52993f9d6ad7SLin Ling buf_discard_identity(hdr); 5300c717a561Smaybee arc_buf_thaw(buf); 5301fa9e4066Sahrens } 5302fa9e4066Sahrens } 5303fa9e4066Sahrens 5304fa9e4066Sahrens int 5305fa9e4066Sahrens arc_released(arc_buf_t *buf) 5306fa9e4066Sahrens { 53076f83844dSMark Maybee int released; 53086f83844dSMark Maybee 53093f9d6ad7SLin Ling mutex_enter(&buf->b_evict_lock); 531089c86e32SChris Williamson released = (buf->b_data != NULL && 531189c86e32SChris Williamson buf->b_hdr->b_l1hdr.b_state == arc_anon); 53123f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 53136f83844dSMark Maybee return (released); 5314ea8dc4b6Seschrock } 5315ea8dc4b6Seschrock 5316ea8dc4b6Seschrock #ifdef ZFS_DEBUG 5317ea8dc4b6Seschrock int 5318ea8dc4b6Seschrock arc_referenced(arc_buf_t *buf) 5319ea8dc4b6Seschrock { 53206f83844dSMark Maybee int referenced; 53216f83844dSMark Maybee 53223f9d6ad7SLin Ling mutex_enter(&buf->b_evict_lock); 532389c86e32SChris Williamson referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 53243f9d6ad7SLin Ling mutex_exit(&buf->b_evict_lock); 53256f83844dSMark Maybee return (referenced); 5326ea8dc4b6Seschrock } 5327ea8dc4b6Seschrock #endif 5328ea8dc4b6Seschrock 5329c717a561Smaybee static void 5330c717a561Smaybee arc_write_ready(zio_t *zio) 5331c717a561Smaybee { 5332c717a561Smaybee arc_write_callback_t *callback = zio->io_private; 5333c717a561Smaybee arc_buf_t *buf = callback->awcb_buf; 53340a4e9518Sgw arc_buf_hdr_t *hdr = buf->b_hdr; 5335dcbf3bd6SGeorge Wilson uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp); 5336c717a561Smaybee 533789c86e32SChris Williamson ASSERT(HDR_HAS_L1HDR(hdr)); 533889c86e32SChris Williamson ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 5339dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 5340e14bb325SJeff Bonwick 53410a4e9518Sgw /* 5342dcbf3bd6SGeorge Wilson * If we're reexecuting this zio because the pool suspended, then 5343dcbf3bd6SGeorge Wilson * cleanup any state that was previously set the first time the 5344*5602294fSDan Kimmel * callback was invoked. 53450a4e9518Sgw */ 5346dcbf3bd6SGeorge Wilson if (zio->io_flags & ZIO_FLAG_REEXECUTED) { 5347dcbf3bd6SGeorge Wilson arc_cksum_free(hdr); 5348dcbf3bd6SGeorge Wilson arc_buf_unwatch(buf); 5349dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_pdata != NULL) { 5350dcbf3bd6SGeorge Wilson if (arc_buf_is_shared(buf)) { 5351dcbf3bd6SGeorge Wilson arc_unshare_buf(hdr, buf); 5352dcbf3bd6SGeorge Wilson } else { 5353dcbf3bd6SGeorge Wilson arc_hdr_free_pdata(hdr); 5354dcbf3bd6SGeorge Wilson } 53550a4e9518Sgw } 53560a4e9518Sgw } 5357dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5358dcbf3bd6SGeorge Wilson ASSERT(!HDR_SHARED_DATA(hdr)); 5359dcbf3bd6SGeorge Wilson ASSERT(!arc_buf_is_shared(buf)); 5360dcbf3bd6SGeorge Wilson 5361dcbf3bd6SGeorge Wilson callback->awcb_ready(zio, buf, callback->awcb_private); 5362dcbf3bd6SGeorge Wilson 5363dcbf3bd6SGeorge Wilson if (HDR_IO_IN_PROGRESS(hdr)) 5364dcbf3bd6SGeorge Wilson ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); 5365dcbf3bd6SGeorge Wilson 5366dcbf3bd6SGeorge Wilson arc_cksum_compute(buf); 5367dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5368dcbf3bd6SGeorge Wilson 5369dcbf3bd6SGeorge Wilson enum zio_compress compress; 5370dcbf3bd6SGeorge Wilson if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5371dcbf3bd6SGeorge Wilson compress = ZIO_COMPRESS_OFF; 5372dcbf3bd6SGeorge Wilson } else { 5373dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp)); 5374dcbf3bd6SGeorge Wilson compress = BP_GET_COMPRESS(zio->io_bp); 5375dcbf3bd6SGeorge Wilson } 5376dcbf3bd6SGeorge Wilson HDR_SET_PSIZE(hdr, psize); 5377dcbf3bd6SGeorge Wilson arc_hdr_set_compress(hdr, compress); 5378dcbf3bd6SGeorge Wilson 5379dcbf3bd6SGeorge Wilson /* 5380dcbf3bd6SGeorge Wilson * If the hdr is compressed, then copy the compressed 5381dcbf3bd6SGeorge Wilson * zio contents into arc_buf_hdr_t. Otherwise, copy the original 5382dcbf3bd6SGeorge Wilson * data buf into the hdr. Ideally, we would like to always copy the 5383dcbf3bd6SGeorge Wilson * io_data into b_pdata but the user may have disabled compressed 5384dcbf3bd6SGeorge Wilson * arc thus the on-disk block may or may not match what we maintain 5385dcbf3bd6SGeorge Wilson * in the hdr's b_pdata field. 5386dcbf3bd6SGeorge Wilson */ 5387*5602294fSDan Kimmel if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 5388*5602294fSDan Kimmel !ARC_BUF_COMPRESSED(buf)) { 5389*5602294fSDan Kimmel ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=, ZIO_COMPRESS_OFF); 5390dcbf3bd6SGeorge Wilson ASSERT3U(psize, >, 0); 5391dcbf3bd6SGeorge Wilson arc_hdr_alloc_pdata(hdr); 5392dcbf3bd6SGeorge Wilson bcopy(zio->io_data, hdr->b_l1hdr.b_pdata, psize); 5393dcbf3bd6SGeorge Wilson } else { 5394dcbf3bd6SGeorge Wilson ASSERT3P(buf->b_data, ==, zio->io_orig_data); 5395*5602294fSDan Kimmel ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf)); 5396dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5397dcbf3bd6SGeorge Wilson 5398dcbf3bd6SGeorge Wilson /* 5399dcbf3bd6SGeorge Wilson * This hdr is not compressed so we're able to share 5400dcbf3bd6SGeorge Wilson * the arc_buf_t data buffer with the hdr. 5401dcbf3bd6SGeorge Wilson */ 5402dcbf3bd6SGeorge Wilson arc_share_buf(hdr, buf); 5403*5602294fSDan Kimmel ASSERT0(bcmp(zio->io_orig_data, hdr->b_l1hdr.b_pdata, 5404dcbf3bd6SGeorge Wilson HDR_GET_LSIZE(hdr))); 5405dcbf3bd6SGeorge Wilson } 5406dcbf3bd6SGeorge Wilson arc_hdr_verify(hdr, zio->io_bp); 5407c717a561Smaybee } 5408c717a561Smaybee 54098df0bcf0SPaul Dagnelie static void 54108df0bcf0SPaul Dagnelie arc_write_children_ready(zio_t *zio) 54118df0bcf0SPaul Dagnelie { 54128df0bcf0SPaul Dagnelie arc_write_callback_t *callback = zio->io_private; 54138df0bcf0SPaul Dagnelie arc_buf_t *buf = callback->awcb_buf; 54148df0bcf0SPaul Dagnelie 54158df0bcf0SPaul Dagnelie callback->awcb_children_ready(zio, buf, callback->awcb_private); 54168df0bcf0SPaul Dagnelie } 54178df0bcf0SPaul Dagnelie 541869962b56SMatthew Ahrens /* 541969962b56SMatthew Ahrens * The SPA calls this callback for each physical write that happens on behalf 542069962b56SMatthew Ahrens * of a logical write. See the comment in dbuf_write_physdone() for details. 542169962b56SMatthew Ahrens */ 542269962b56SMatthew Ahrens static void 542369962b56SMatthew Ahrens arc_write_physdone(zio_t *zio) 542469962b56SMatthew Ahrens { 542569962b56SMatthew Ahrens arc_write_callback_t *cb = zio->io_private; 542669962b56SMatthew Ahrens if (cb->awcb_physdone != NULL) 542769962b56SMatthew Ahrens cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 542869962b56SMatthew Ahrens } 542969962b56SMatthew Ahrens 5430fa9e4066Sahrens static void 5431fa9e4066Sahrens arc_write_done(zio_t *zio) 5432fa9e4066Sahrens { 5433c717a561Smaybee arc_write_callback_t *callback = zio->io_private; 5434c717a561Smaybee arc_buf_t *buf = callback->awcb_buf; 5435c717a561Smaybee arc_buf_hdr_t *hdr = buf->b_hdr; 5436fa9e4066Sahrens 5437dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5438b24ab676SJeff Bonwick 5439b24ab676SJeff Bonwick if (zio->io_error == 0) { 5440dcbf3bd6SGeorge Wilson arc_hdr_verify(hdr, zio->io_bp); 5441dcbf3bd6SGeorge Wilson 54425d7b4d43SMatthew Ahrens if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 544343466aaeSMax Grossman buf_discard_identity(hdr); 544443466aaeSMax Grossman } else { 544543466aaeSMax Grossman hdr->b_dva = *BP_IDENTITY(zio->io_bp); 544643466aaeSMax Grossman hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 544743466aaeSMax Grossman } 5448b24ab676SJeff Bonwick } else { 5449dcbf3bd6SGeorge Wilson ASSERT(HDR_EMPTY(hdr)); 5450b24ab676SJeff Bonwick } 5451fa9e4066Sahrens 5452ea8dc4b6Seschrock /* 54535d7b4d43SMatthew Ahrens * If the block to be written was all-zero or compressed enough to be 54545d7b4d43SMatthew Ahrens * embedded in the BP, no write was performed so there will be no 54555d7b4d43SMatthew Ahrens * dva/birth/checksum. The buffer must therefore remain anonymous 54565d7b4d43SMatthew Ahrens * (and uncached). 5457ea8dc4b6Seschrock */ 5458dcbf3bd6SGeorge Wilson if (!HDR_EMPTY(hdr)) { 5459fa9e4066Sahrens arc_buf_hdr_t *exists; 5460fa9e4066Sahrens kmutex_t *hash_lock; 5461fa9e4066Sahrens 5462*5602294fSDan Kimmel ASSERT3U(zio->io_error, ==, 0); 5463b24ab676SJeff Bonwick 54646b4acc8bSahrens arc_cksum_verify(buf); 54656b4acc8bSahrens 5466fa9e4066Sahrens exists = buf_hash_insert(hdr, &hash_lock); 546789c86e32SChris Williamson if (exists != NULL) { 5468fa9e4066Sahrens /* 5469fa9e4066Sahrens * This can only happen if we overwrite for 5470fa9e4066Sahrens * sync-to-convergence, because we remove 5471fa9e4066Sahrens * buffers from the hash table when we arc_free(). 5472fa9e4066Sahrens */ 5473b24ab676SJeff Bonwick if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 5474b24ab676SJeff Bonwick if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5475b24ab676SJeff Bonwick panic("bad overwrite, hdr=%p exists=%p", 5476b24ab676SJeff Bonwick (void *)hdr, (void *)exists); 547789c86e32SChris Williamson ASSERT(refcount_is_zero( 547889c86e32SChris Williamson &exists->b_l1hdr.b_refcnt)); 5479b24ab676SJeff Bonwick arc_change_state(arc_anon, exists, hash_lock); 5480b24ab676SJeff Bonwick mutex_exit(hash_lock); 5481b24ab676SJeff Bonwick arc_hdr_destroy(exists); 5482b24ab676SJeff Bonwick exists = buf_hash_insert(hdr, &hash_lock); 5483b24ab676SJeff Bonwick ASSERT3P(exists, ==, NULL); 548480901aeaSGeorge Wilson } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 548580901aeaSGeorge Wilson /* nopwrite */ 548680901aeaSGeorge Wilson ASSERT(zio->io_prop.zp_nopwrite); 548780901aeaSGeorge Wilson if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 548880901aeaSGeorge Wilson panic("bad nopwrite, hdr=%p exists=%p", 548980901aeaSGeorge Wilson (void *)hdr, (void *)exists); 5490b24ab676SJeff Bonwick } else { 5491b24ab676SJeff Bonwick /* Dedup */ 5492dcbf3bd6SGeorge Wilson ASSERT(hdr->b_l1hdr.b_bufcnt == 1); 549389c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5494b24ab676SJeff Bonwick ASSERT(BP_GET_DEDUP(zio->io_bp)); 5495b24ab676SJeff Bonwick ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5496ae46e4c7SMatthew Ahrens } 5497fa9e4066Sahrens } 5498dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5499088f3894Sahrens /* if it's not anon, we are doing a scrub */ 550089c86e32SChris Williamson if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5501088f3894Sahrens arc_access(hdr, hash_lock); 550244eda4d7Smaybee mutex_exit(hash_lock); 5503ea8dc4b6Seschrock } else { 5504dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5505fa9e4066Sahrens } 5506ea8dc4b6Seschrock 550789c86e32SChris Williamson ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5508b24ab676SJeff Bonwick callback->awcb_done(zio, buf, callback->awcb_private); 5509fa9e4066Sahrens 5510c717a561Smaybee kmem_free(callback, sizeof (arc_write_callback_t)); 5511fa9e4066Sahrens } 5512fa9e4066Sahrens 5513c717a561Smaybee zio_t * 5514dcbf3bd6SGeorge Wilson arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 5515dcbf3bd6SGeorge Wilson boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready, 55168df0bcf0SPaul Dagnelie arc_done_func_t *children_ready, arc_done_func_t *physdone, 551769962b56SMatthew Ahrens arc_done_func_t *done, void *private, zio_priority_t priority, 55187802d7bfSMatthew Ahrens int zio_flags, const zbookmark_phys_t *zb) 5519fa9e4066Sahrens { 5520fa9e4066Sahrens arc_buf_hdr_t *hdr = buf->b_hdr; 5521c717a561Smaybee arc_write_callback_t *callback; 5522e14bb325SJeff Bonwick zio_t *zio; 5523fa9e4066Sahrens 5524dcbf3bd6SGeorge Wilson ASSERT3P(ready, !=, NULL); 5525dcbf3bd6SGeorge Wilson ASSERT3P(done, !=, NULL); 5526fa9e4066Sahrens ASSERT(!HDR_IO_ERROR(hdr)); 552789c86e32SChris Williamson ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5528dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5529dcbf3bd6SGeorge Wilson ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 55303baa08fcSek if (l2arc) 5531dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5532*5602294fSDan Kimmel if (ARC_BUF_COMPRESSED(buf)) { 5533*5602294fSDan Kimmel ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_OFF); 5534*5602294fSDan Kimmel ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf)); 5535*5602294fSDan Kimmel zio_flags |= ZIO_FLAG_RAW; 5536*5602294fSDan Kimmel } 5537c717a561Smaybee callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5538c717a561Smaybee callback->awcb_ready = ready; 55398df0bcf0SPaul Dagnelie callback->awcb_children_ready = children_ready; 554069962b56SMatthew Ahrens callback->awcb_physdone = physdone; 5541c717a561Smaybee callback->awcb_done = done; 5542c717a561Smaybee callback->awcb_private = private; 5543c717a561Smaybee callback->awcb_buf = buf; 5544088f3894Sahrens 5545dcbf3bd6SGeorge Wilson /* 5546dcbf3bd6SGeorge Wilson * The hdr's b_pdata is now stale, free it now. A new data block 5547dcbf3bd6SGeorge Wilson * will be allocated when the zio pipeline calls arc_write_ready(). 5548dcbf3bd6SGeorge Wilson */ 5549dcbf3bd6SGeorge Wilson if (hdr->b_l1hdr.b_pdata != NULL) { 5550dcbf3bd6SGeorge Wilson /* 5551dcbf3bd6SGeorge Wilson * If the buf is currently sharing the data block with 5552dcbf3bd6SGeorge Wilson * the hdr then we need to break that relationship here. 5553dcbf3bd6SGeorge Wilson * The hdr will remain with a NULL data pointer and the 5554dcbf3bd6SGeorge Wilson * buf will take sole ownership of the block. 5555dcbf3bd6SGeorge Wilson */ 5556dcbf3bd6SGeorge Wilson if (arc_buf_is_shared(buf)) { 5557dcbf3bd6SGeorge Wilson arc_unshare_buf(hdr, buf); 5558dcbf3bd6SGeorge Wilson } else { 5559dcbf3bd6SGeorge Wilson arc_hdr_free_pdata(hdr); 5560dcbf3bd6SGeorge Wilson } 5561dcbf3bd6SGeorge Wilson VERIFY3P(buf->b_data, !=, NULL); 5562dcbf3bd6SGeorge Wilson arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); 5563dcbf3bd6SGeorge Wilson } 5564dcbf3bd6SGeorge Wilson ASSERT(!arc_buf_is_shared(buf)); 5565dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5566dcbf3bd6SGeorge Wilson 5567*5602294fSDan Kimmel zio = zio_write(pio, spa, txg, bp, buf->b_data, 5568*5602294fSDan Kimmel HDR_GET_LSIZE(hdr), arc_buf_size(buf), zp, arc_write_ready, 55698df0bcf0SPaul Dagnelie (children_ready != NULL) ? arc_write_children_ready : NULL, 55708df0bcf0SPaul Dagnelie arc_write_physdone, arc_write_done, callback, 557169962b56SMatthew Ahrens priority, zio_flags, zb); 5572fa9e4066Sahrens 5573c717a561Smaybee return (zio); 5574fa9e4066Sahrens } 5575fa9e4066Sahrens 55761ab7f2deSmaybee static int 557769962b56SMatthew Ahrens arc_memory_throttle(uint64_t reserve, uint64_t txg) 55781ab7f2deSmaybee { 55791ab7f2deSmaybee #ifdef _KERNEL 55801ab7f2deSmaybee uint64_t available_memory = ptob(freemem); 55811ab7f2deSmaybee static uint64_t page_load = 0; 55821ab7f2deSmaybee static uint64_t last_txg = 0; 55831ab7f2deSmaybee 55841ab7f2deSmaybee #if defined(__i386) 55851ab7f2deSmaybee available_memory = 55861ab7f2deSmaybee MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 55871ab7f2deSmaybee #endif 558869962b56SMatthew Ahrens 558969962b56SMatthew Ahrens if (freemem > physmem * arc_lotsfree_percent / 100) 55901ab7f2deSmaybee return (0); 55911ab7f2deSmaybee 55921ab7f2deSmaybee if (txg > last_txg) { 55931ab7f2deSmaybee last_txg = txg; 55941ab7f2deSmaybee page_load = 0; 55951ab7f2deSmaybee } 55961ab7f2deSmaybee /* 55971ab7f2deSmaybee * If we are in pageout, we know that memory is already tight, 55981ab7f2deSmaybee * the arc is already going to be evicting, so we just want to 55991ab7f2deSmaybee * continue to let page writes occur as quickly as possible. 56001ab7f2deSmaybee */ 56011ab7f2deSmaybee if (curproc == proc_pageout) { 56021ab7f2deSmaybee if (page_load > MAX(ptob(minfree), available_memory) / 4) 5603be6fd75aSMatthew Ahrens return (SET_ERROR(ERESTART)); 56041ab7f2deSmaybee /* Note: reserve is inflated, so we deflate */ 56051ab7f2deSmaybee page_load += reserve / 8; 56061ab7f2deSmaybee return (0); 56071ab7f2deSmaybee } else if (page_load > 0 && arc_reclaim_needed()) { 56081ab7f2deSmaybee /* memory is low, delay before restarting */ 56091ab7f2deSmaybee ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5610be6fd75aSMatthew Ahrens return (SET_ERROR(EAGAIN)); 56111ab7f2deSmaybee } 56121ab7f2deSmaybee page_load = 0; 56131ab7f2deSmaybee #endif 56141ab7f2deSmaybee return (0); 56151ab7f2deSmaybee } 56161ab7f2deSmaybee 5617fa9e4066Sahrens void 56181ab7f2deSmaybee arc_tempreserve_clear(uint64_t reserve) 5619fa9e4066Sahrens { 56201ab7f2deSmaybee atomic_add_64(&arc_tempreserve, -reserve); 5621fa9e4066Sahrens ASSERT((int64_t)arc_tempreserve >= 0); 5622fa9e4066Sahrens } 5623fa9e4066Sahrens 5624fa9e4066Sahrens int 56251ab7f2deSmaybee arc_tempreserve_space(uint64_t reserve, uint64_t txg) 5626fa9e4066Sahrens { 56271ab7f2deSmaybee int error; 56282fdbea25SAleksandr Guzovskiy uint64_t anon_size; 56291ab7f2deSmaybee 56301ab7f2deSmaybee if (reserve > arc_c/4 && !arc_no_grow) 56311ab7f2deSmaybee arc_c = MIN(arc_c_max, reserve * 4); 56321ab7f2deSmaybee if (reserve > arc_c) 5633be6fd75aSMatthew Ahrens return (SET_ERROR(ENOMEM)); 5634112fe045Smaybee 56352fdbea25SAleksandr Guzovskiy /* 56362fdbea25SAleksandr Guzovskiy * Don't count loaned bufs as in flight dirty data to prevent long 56372fdbea25SAleksandr Guzovskiy * network delays from blocking transactions that are ready to be 56382fdbea25SAleksandr Guzovskiy * assigned to a txg. 56392fdbea25SAleksandr Guzovskiy */ 5640*5602294fSDan Kimmel 5641*5602294fSDan Kimmel /* assert that it has not wrapped around */ 5642*5602294fSDan Kimmel ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 5643*5602294fSDan Kimmel 56442fd872a7SPrakash Surya anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 56452fd872a7SPrakash Surya arc_loaned_bytes), 0); 56462fdbea25SAleksandr Guzovskiy 56471ab7f2deSmaybee /* 56481ab7f2deSmaybee * Writes will, almost always, require additional memory allocations 5649f7170741SWill Andrews * in order to compress/encrypt/etc the data. We therefore need to 56501ab7f2deSmaybee * make sure that there is sufficient available memory for this. 56511ab7f2deSmaybee */ 565269962b56SMatthew Ahrens error = arc_memory_throttle(reserve, txg); 565369962b56SMatthew Ahrens if (error != 0) 56541ab7f2deSmaybee return (error); 56551ab7f2deSmaybee 5656fa9e4066Sahrens /* 5657112fe045Smaybee * Throttle writes when the amount of dirty data in the cache 5658112fe045Smaybee * gets too large. We try to keep the cache less than half full 5659112fe045Smaybee * of dirty blocks so that our sync times don't grow too large. 5660112fe045Smaybee * Note: if two requests come in concurrently, we might let them 5661112fe045Smaybee * both succeed, when one of them should fail. Not a huge deal. 5662fa9e4066Sahrens */ 56632fdbea25SAleksandr Guzovskiy 56642fdbea25SAleksandr Guzovskiy if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 56652fdbea25SAleksandr Guzovskiy anon_size > arc_c / 4) { 5666dcbf3bd6SGeorge Wilson uint64_t meta_esize = 5667dcbf3bd6SGeorge Wilson refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5668dcbf3bd6SGeorge Wilson uint64_t data_esize = 5669dcbf3bd6SGeorge Wilson refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 56700e8c6158Smaybee dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 56710e8c6158Smaybee "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5672dcbf3bd6SGeorge Wilson arc_tempreserve >> 10, meta_esize >> 10, 5673dcbf3bd6SGeorge Wilson data_esize >> 10, reserve >> 10, arc_c >> 10); 5674be6fd75aSMatthew Ahrens return (SET_ERROR(ERESTART)); 5675fa9e4066Sahrens } 56761ab7f2deSmaybee atomic_add_64(&arc_tempreserve, reserve); 5677fa9e4066Sahrens return (0); 5678fa9e4066Sahrens } 5679fa9e4066Sahrens 56804076b1bfSPrakash Surya static void 56814076b1bfSPrakash Surya arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 56824076b1bfSPrakash Surya kstat_named_t *evict_data, kstat_named_t *evict_metadata) 56834076b1bfSPrakash Surya { 56842fd872a7SPrakash Surya size->value.ui64 = refcount_count(&state->arcs_size); 5685dcbf3bd6SGeorge Wilson evict_data->value.ui64 = 5686dcbf3bd6SGeorge Wilson refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); 5687dcbf3bd6SGeorge Wilson evict_metadata->value.ui64 = 5688dcbf3bd6SGeorge Wilson refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); 56894076b1bfSPrakash Surya } 56904076b1bfSPrakash Surya 56914076b1bfSPrakash Surya static int 56924076b1bfSPrakash Surya arc_kstat_update(kstat_t *ksp, int rw) 56934076b1bfSPrakash Surya { 56944076b1bfSPrakash Surya arc_stats_t *as = ksp->ks_data; 56954076b1bfSPrakash Surya 56964076b1bfSPrakash Surya if (rw == KSTAT_WRITE) { 56974076b1bfSPrakash Surya return (EACCES); 56984076b1bfSPrakash Surya } else { 56994076b1bfSPrakash Surya arc_kstat_update_state(arc_anon, 57004076b1bfSPrakash Surya &as->arcstat_anon_size, 57014076b1bfSPrakash Surya &as->arcstat_anon_evictable_data, 57024076b1bfSPrakash Surya &as->arcstat_anon_evictable_metadata); 57034076b1bfSPrakash Surya arc_kstat_update_state(arc_mru, 57044076b1bfSPrakash Surya &as->arcstat_mru_size, 57054076b1bfSPrakash Surya &as->arcstat_mru_evictable_data, 57064076b1bfSPrakash Surya &as->arcstat_mru_evictable_metadata); 57074076b1bfSPrakash Surya arc_kstat_update_state(arc_mru_ghost, 57084076b1bfSPrakash Surya &as->arcstat_mru_ghost_size, 57094076b1bfSPrakash Surya &as->arcstat_mru_ghost_evictable_data, 57104076b1bfSPrakash Surya &as->arcstat_mru_ghost_evictable_metadata); 57114076b1bfSPrakash Surya arc_kstat_update_state(arc_mfu, 57124076b1bfSPrakash Surya &as->arcstat_mfu_size, 57134076b1bfSPrakash Surya &as->arcstat_mfu_evictable_data, 57144076b1bfSPrakash Surya &as->arcstat_mfu_evictable_metadata); 57154076b1bfSPrakash Surya arc_kstat_update_state(arc_mfu_ghost, 57164076b1bfSPrakash Surya &as->arcstat_mfu_ghost_size, 57174076b1bfSPrakash Surya &as->arcstat_mfu_ghost_evictable_data, 57184076b1bfSPrakash Surya &as->arcstat_mfu_ghost_evictable_metadata); 57194076b1bfSPrakash Surya } 57204076b1bfSPrakash Surya 57214076b1bfSPrakash Surya return (0); 57224076b1bfSPrakash Surya } 57234076b1bfSPrakash Surya 5724dcbf3bd6SGeorge Wilson /* 5725dcbf3bd6SGeorge Wilson * This function *must* return indices evenly distributed between all 5726dcbf3bd6SGeorge Wilson * sublists of the multilist. This is needed due to how the ARC eviction 5727dcbf3bd6SGeorge Wilson * code is laid out; arc_evict_state() assumes ARC buffers are evenly 5728dcbf3bd6SGeorge Wilson * distributed between all sublists and uses this assumption when 5729dcbf3bd6SGeorge Wilson * deciding which sublist to evict from and how much to evict from it. 5730dcbf3bd6SGeorge Wilson */ 5731dcbf3bd6SGeorge Wilson unsigned int 5732dcbf3bd6SGeorge Wilson arc_state_multilist_index_func(multilist_t *ml, void *obj) 5733dcbf3bd6SGeorge Wilson { 5734dcbf3bd6SGeorge Wilson arc_buf_hdr_t *hdr = obj; 5735dcbf3bd6SGeorge Wilson 5736dcbf3bd6SGeorge Wilson /* 5737dcbf3bd6SGeorge Wilson * We rely on b_dva to generate evenly distributed index 5738dcbf3bd6SGeorge Wilson * numbers using buf_hash below. So, as an added precaution, 5739dcbf3bd6SGeorge Wilson * let's make sure we never add empty buffers to the arc lists. 5740dcbf3bd6SGeorge Wilson */ 5741dcbf3bd6SGeorge Wilson ASSERT(!HDR_EMPTY(hdr)); 5742dcbf3bd6SGeorge Wilson 5743dcbf3bd6SGeorge Wilson /* 5744dcbf3bd6SGeorge Wilson * The assumption here, is the hash value for a given 5745dcbf3bd6SGeorge Wilson * arc_buf_hdr_t will remain constant throughout it's lifetime 5746dcbf3bd6SGeorge Wilson * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 5747dcbf3bd6SGeorge Wilson * Thus, we don't need to store the header's sublist index 5748dcbf3bd6SGeorge Wilson * on insertion, as this index can be recalculated on removal. 5749dcbf3bd6SGeorge Wilson * 5750dcbf3bd6SGeorge Wilson * Also, the low order bits of the hash value are thought to be 5751dcbf3bd6SGeorge Wilson * distributed evenly. Otherwise, in the case that the multilist 5752dcbf3bd6SGeorge Wilson * has a power of two number of sublists, each sublists' usage 5753dcbf3bd6SGeorge Wilson * would not be evenly distributed. 5754dcbf3bd6SGeorge Wilson */ 5755dcbf3bd6SGeorge Wilson return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 5756dcbf3bd6SGeorge Wilson multilist_get_num_sublists(ml)); 5757dcbf3bd6SGeorge Wilson } 5758dcbf3bd6SGeorge Wilson 5759dcbf3bd6SGeorge Wilson static void 5760dcbf3bd6SGeorge Wilson arc_state_init(void) 5761dcbf3bd6SGeorge Wilson { 5762dcbf3bd6SGeorge Wilson arc_anon = &ARC_anon; 5763dcbf3bd6SGeorge Wilson arc_mru = &ARC_mru; 5764dcbf3bd6SGeorge Wilson arc_mru_ghost = &ARC_mru_ghost; 5765dcbf3bd6SGeorge Wilson arc_mfu = &ARC_mfu; 5766dcbf3bd6SGeorge Wilson arc_mfu_ghost = &ARC_mfu_ghost; 5767dcbf3bd6SGeorge Wilson arc_l2c_only = &ARC_l2c_only; 5768dcbf3bd6SGeorge Wilson 5769dcbf3bd6SGeorge Wilson multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 5770dcbf3bd6SGeorge Wilson sizeof (arc_buf_hdr_t), 5771dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5772dcbf3bd6SGeorge Wilson zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5773dcbf3bd6SGeorge Wilson multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 5774dcbf3bd6SGeorge Wilson sizeof (arc_buf_hdr_t), 5775dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5776dcbf3bd6SGeorge Wilson zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5777dcbf3bd6SGeorge Wilson multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 5778dcbf3bd6SGeorge Wilson sizeof (arc_buf_hdr_t), 5779dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5780dcbf3bd6SGeorge Wilson zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5781dcbf3bd6SGeorge Wilson multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 5782dcbf3bd6SGeorge Wilson sizeof (arc_buf_hdr_t), 5783dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5784dcbf3bd6SGeorge Wilson zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5785dcbf3bd6SGeorge Wilson multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 5786dcbf3bd6SGeorge Wilson sizeof (arc_buf_hdr_t), 5787dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5788dcbf3bd6SGeorge Wilson zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5789dcbf3bd6SGeorge Wilson multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 5790dcbf3bd6SGeorge Wilson sizeof (arc_buf_hdr_t), 5791dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5792dcbf3bd6SGeorge Wilson zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5793dcbf3bd6SGeorge Wilson multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 5794dcbf3bd6SGeorge Wilson sizeof (arc_buf_hdr_t), 5795dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5796dcbf3bd6SGeorge Wilson zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5797dcbf3bd6SGeorge Wilson multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 5798dcbf3bd6SGeorge Wilson sizeof (arc_buf_hdr_t), 5799dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5800dcbf3bd6SGeorge Wilson zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5801dcbf3bd6SGeorge Wilson multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], 5802dcbf3bd6SGeorge Wilson sizeof (arc_buf_hdr_t), 5803dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5804dcbf3bd6SGeorge Wilson zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5805dcbf3bd6SGeorge Wilson multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], 5806dcbf3bd6SGeorge Wilson sizeof (arc_buf_hdr_t), 5807dcbf3bd6SGeorge Wilson offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5808dcbf3bd6SGeorge Wilson zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5809dcbf3bd6SGeorge Wilson 5810dcbf3bd6SGeorge Wilson refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5811dcbf3bd6SGeorge Wilson refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5812dcbf3bd6SGeorge Wilson refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 5813dcbf3bd6SGeorge Wilson refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 5814dcbf3bd6SGeorge Wilson refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 5815dcbf3bd6SGeorge Wilson refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 5816dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 5817dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 5818dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 5819dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 5820dcbf3bd6SGeorge Wilson refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 5821dcbf3bd6SGeorge Wilson refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 5822dcbf3bd6SGeorge Wilson 5823dcbf3bd6SGeorge Wilson refcount_create(&arc_anon->arcs_size); 5824dcbf3bd6SGeorge Wilson refcount_create(&arc_mru->arcs_size); 5825dcbf3bd6SGeorge Wilson refcount_create(&arc_mru_ghost->arcs_size); 5826dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu->arcs_size); 5827dcbf3bd6SGeorge Wilson refcount_create(&arc_mfu_ghost->arcs_size); 5828dcbf3bd6SGeorge Wilson refcount_create(&arc_l2c_only->arcs_size); 5829dcbf3bd6SGeorge Wilson } 5830dcbf3bd6SGeorge Wilson 5831dcbf3bd6SGeorge Wilson static void 5832dcbf3bd6SGeorge Wilson arc_state_fini(void) 5833dcbf3bd6SGeorge Wilson { 5834dcbf3bd6SGeorge Wilson refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5835dcbf3bd6SGeorge Wilson refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5836dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 5837dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 5838dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 5839dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 5840dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 5841dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 5842dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 5843dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 5844dcbf3bd6SGeorge Wilson refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 5845dcbf3bd6SGeorge Wilson refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 5846dcbf3bd6SGeorge Wilson 5847dcbf3bd6SGeorge Wilson refcount_destroy(&arc_anon->arcs_size); 5848dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru->arcs_size); 5849dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mru_ghost->arcs_size); 5850dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu->arcs_size); 5851dcbf3bd6SGeorge Wilson refcount_destroy(&arc_mfu_ghost->arcs_size); 5852dcbf3bd6SGeorge Wilson refcount_destroy(&arc_l2c_only->arcs_size); 5853dcbf3bd6SGeorge Wilson 5854dcbf3bd6SGeorge Wilson multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 5855dcbf3bd6SGeorge Wilson multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 5856dcbf3bd6SGeorge Wilson multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 5857dcbf3bd6SGeorge Wilson multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 5858dcbf3bd6SGeorge Wilson multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 5859dcbf3bd6SGeorge Wilson multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 5860dcbf3bd6SGeorge Wilson multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 5861dcbf3bd6SGeorge Wilson multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 5862dcbf3bd6SGeorge Wilson } 5863dcbf3bd6SGeorge Wilson 5864dcbf3bd6SGeorge Wilson uint64_t 5865dcbf3bd6SGeorge Wilson arc_max_bytes(void) 5866244781f1SPrakash Surya { 5867dcbf3bd6SGeorge Wilson return (arc_c_max); 5868244781f1SPrakash Surya } 5869244781f1SPrakash Surya 5870fa9e4066Sahrens void 5871fa9e4066Sahrens arc_init(void) 5872fa9e4066Sahrens { 58732ec99e3eSMatthew Ahrens /* 58742ec99e3eSMatthew Ahrens * allmem is "all memory that we could possibly use". 58752ec99e3eSMatthew Ahrens */ 58762ec99e3eSMatthew Ahrens #ifdef _KERNEL 58772ec99e3eSMatthew Ahrens uint64_t allmem = ptob(physmem - swapfs_minfree); 58782ec99e3eSMatthew Ahrens #else 58792ec99e3eSMatthew Ahrens uint64_t allmem = (physmem * PAGESIZE) / 2; 58802ec99e3eSMatthew Ahrens #endif 58812ec99e3eSMatthew Ahrens 5882244781f1SPrakash Surya mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 5883244781f1SPrakash Surya cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 5884244781f1SPrakash Surya cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 5885244781f1SPrakash Surya 588613506d1eSmaybee /* Convert seconds to clock ticks */ 5887b19a79ecSperrin arc_min_prefetch_lifespan = 1 * hz; 588813506d1eSmaybee 5889112fe045Smaybee /* set min cache to 1/32 of all memory, or 64MB, whichever is more */ 58902ec99e3eSMatthew Ahrens arc_c_min = MAX(allmem / 32, 64 << 20); 5891112fe045Smaybee /* set max to 3/4 of all memory, or all but 1GB, whichever is more */ 58922ec99e3eSMatthew Ahrens if (allmem >= 1 << 30) 58932ec99e3eSMatthew Ahrens arc_c_max = allmem - (1 << 30); 5894fa9e4066Sahrens else 589544cb6abcSbmc arc_c_max = arc_c_min; 58962ec99e3eSMatthew Ahrens arc_c_max = MAX(allmem * 3 / 4, arc_c_max); 5897a2eea2e1Sahrens 58988fe00bfbSMatthew Ahrens /* 58998fe00bfbSMatthew Ahrens * In userland, there's only the memory pressure that we artificially 59008fe00bfbSMatthew Ahrens * create (see arc_available_memory()). Don't let arc_c get too 59018fe00bfbSMatthew Ahrens * small, because it can cause transactions to be larger than 59028fe00bfbSMatthew Ahrens * arc_c, causing arc_tempreserve_space() to fail. 59038fe00bfbSMatthew Ahrens */ 59048fe00bfbSMatthew Ahrens #ifndef _KERNEL 59058fe00bfbSMatthew Ahrens arc_c_min = arc_c_max / 2; 59068fe00bfbSMatthew Ahrens #endif 59078fe00bfbSMatthew Ahrens 5908a2eea2e1Sahrens /* 5909a2eea2e1Sahrens * Allow the tunables to override our calculations if they are 5910a2eea2e1Sahrens * reasonable (ie. over 64MB) 5911a2eea2e1Sahrens */ 5912e5961710SMatthew Ahrens if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) { 591344cb6abcSbmc arc_c_max = zfs_arc_max; 5914e5961710SMatthew Ahrens arc_c_min = MIN(arc_c_min, arc_c_max); 5915e5961710SMatthew Ahrens } 59162ec99e3eSMatthew Ahrens if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max) 591744cb6abcSbmc arc_c_min = zfs_arc_min; 5918a2eea2e1Sahrens 591944cb6abcSbmc arc_c = arc_c_max; 592044cb6abcSbmc arc_p = (arc_c >> 1); 5921dcbf3bd6SGeorge Wilson arc_size = 0; 5922fa9e4066Sahrens 59230e8c6158Smaybee /* limit meta-data to 1/4 of the arc capacity */ 59240e8c6158Smaybee arc_meta_limit = arc_c_max / 4; 59251116048bSek 5926af868f46SMatthew Ahrens #ifdef _KERNEL 5927af868f46SMatthew Ahrens /* 5928af868f46SMatthew Ahrens * Metadata is stored in the kernel's heap. Don't let us 5929af868f46SMatthew Ahrens * use more than half the heap for the ARC. 5930af868f46SMatthew Ahrens */ 5931af868f46SMatthew Ahrens arc_meta_limit = MIN(arc_meta_limit, 5932af868f46SMatthew Ahrens vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2); 5933af868f46SMatthew Ahrens #endif 5934af868f46SMatthew Ahrens 59351116048bSek /* Allow the tunable to override if it is reasonable */ 59361116048bSek if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 59371116048bSek arc_meta_limit = zfs_arc_meta_limit; 59381116048bSek 59390e8c6158Smaybee if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 59400e8c6158Smaybee arc_c_min = arc_meta_limit / 2; 59410e8c6158Smaybee 59423a5286a1SMatthew Ahrens if (zfs_arc_meta_min > 0) { 59433a5286a1SMatthew Ahrens arc_meta_min = zfs_arc_meta_min; 59443a5286a1SMatthew Ahrens } else { 59453a5286a1SMatthew Ahrens arc_meta_min = arc_c_min / 2; 59463a5286a1SMatthew Ahrens } 59473a5286a1SMatthew Ahrens 59485a98e54bSBrendan Gregg - Sun Microsystems if (zfs_arc_grow_retry > 0) 59495a98e54bSBrendan Gregg - Sun Microsystems arc_grow_retry = zfs_arc_grow_retry; 59505a98e54bSBrendan Gregg - Sun Microsystems 59515a98e54bSBrendan Gregg - Sun Microsystems if (zfs_arc_shrink_shift > 0) 59525a98e54bSBrendan Gregg - Sun Microsystems arc_shrink_shift = zfs_arc_shrink_shift; 59535a98e54bSBrendan Gregg - Sun Microsystems 59542ec99e3eSMatthew Ahrens /* 59552ec99e3eSMatthew Ahrens * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 59562ec99e3eSMatthew Ahrens */ 59572ec99e3eSMatthew Ahrens if (arc_no_grow_shift >= arc_shrink_shift) 59582ec99e3eSMatthew Ahrens arc_no_grow_shift = arc_shrink_shift - 1; 59592ec99e3eSMatthew Ahrens 59605a98e54bSBrendan Gregg - Sun Microsystems if (zfs_arc_p_min_shift > 0) 59615a98e54bSBrendan Gregg - Sun Microsystems arc_p_min_shift = zfs_arc_p_min_shift; 59625a98e54bSBrendan Gregg - Sun Microsystems 5963244781f1SPrakash Surya if (zfs_arc_num_sublists_per_state < 1) 5964244781f1SPrakash Surya zfs_arc_num_sublists_per_state = MAX(boot_ncpus, 1); 5965244781f1SPrakash Surya 5966fa9e4066Sahrens /* if kmem_flags are set, lets try to use less memory */ 5967fa9e4066Sahrens if (kmem_debugging()) 596844cb6abcSbmc arc_c = arc_c / 2; 596944cb6abcSbmc if (arc_c < arc_c_min) 597044cb6abcSbmc arc_c = arc_c_min; 597144cb6abcSbmc 5972dcbf3bd6SGeorge Wilson arc_state_init(); 5973fa9e4066Sahrens buf_init(); 5974fa9e4066Sahrens 5975dcbf3bd6SGeorge Wilson arc_reclaim_thread_exit = B_FALSE; 5976fa9e4066Sahrens 597744cb6abcSbmc arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 597844cb6abcSbmc sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 597944cb6abcSbmc 598044cb6abcSbmc if (arc_ksp != NULL) { 598144cb6abcSbmc arc_ksp->ks_data = &arc_stats; 59824076b1bfSPrakash Surya arc_ksp->ks_update = arc_kstat_update; 598344cb6abcSbmc kstat_install(arc_ksp); 598444cb6abcSbmc } 598544cb6abcSbmc 5986fa9e4066Sahrens (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 5987fa9e4066Sahrens TS_RUN, minclsyspri); 598849e3519aSmaybee 5989dcbf3bd6SGeorge Wilson arc_dead = B_FALSE; 59903a737e0dSbrendan arc_warm = B_FALSE; 59911ab7f2deSmaybee 599269962b56SMatthew Ahrens /* 599369962b56SMatthew Ahrens * Calculate maximum amount of dirty data per pool. 599469962b56SMatthew Ahrens * 599569962b56SMatthew Ahrens * If it has been set by /etc/system, take that. 599669962b56SMatthew Ahrens * Otherwise, use a percentage of physical memory defined by 599769962b56SMatthew Ahrens * zfs_dirty_data_max_percent (default 10%) with a cap at 599869962b56SMatthew Ahrens * zfs_dirty_data_max_max (default 4GB). 599969962b56SMatthew Ahrens */ 600069962b56SMatthew Ahrens if (zfs_dirty_data_max == 0) { 600169962b56SMatthew Ahrens zfs_dirty_data_max = physmem * PAGESIZE * 600269962b56SMatthew Ahrens zfs_dirty_data_max_percent / 100; 600369962b56SMatthew Ahrens zfs_dirty_data_max = MIN(zfs_dirty_data_max, 600469962b56SMatthew Ahrens zfs_dirty_data_max_max); 600569962b56SMatthew Ahrens } 6006fa9e4066Sahrens } 6007fa9e4066Sahrens 6008fa9e4066Sahrens void 6009fa9e4066Sahrens arc_fini(void) 6010fa9e4066Sahrens { 6011244781f1SPrakash Surya mutex_enter(&arc_reclaim_lock); 6012dcbf3bd6SGeorge Wilson arc_reclaim_thread_exit = B_TRUE; 6013244781f1SPrakash Surya /* 6014244781f1SPrakash Surya * The reclaim thread will set arc_reclaim_thread_exit back to 6015dcbf3bd6SGeorge Wilson * B_FALSE when it is finished exiting; we're waiting for that. 6016244781f1SPrakash Surya */ 6017244781f1SPrakash Surya while (arc_reclaim_thread_exit) { 6018244781f1SPrakash Surya cv_signal(&arc_reclaim_thread_cv); 6019244781f1SPrakash Surya cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 6020244781f1SPrakash Surya } 6021244781f1SPrakash Surya mutex_exit(&arc_reclaim_lock); 6022fa9e4066Sahrens 6023dcbf3bd6SGeorge Wilson /* Use B_TRUE to ensure *all* buffers are evicted */ 6024dcbf3bd6SGeorge Wilson arc_flush(NULL, B_TRUE); 6025fa9e4066Sahrens 6026dcbf3bd6SGeorge Wilson arc_dead = B_TRUE; 6027fa9e4066Sahrens 602844cb6abcSbmc if (arc_ksp != NULL) { 602944cb6abcSbmc kstat_delete(arc_ksp); 603044cb6abcSbmc arc_ksp = NULL; 603144cb6abcSbmc } 603244cb6abcSbmc 6033244781f1SPrakash Surya mutex_destroy(&arc_reclaim_lock); 6034244781f1SPrakash Surya cv_destroy(&arc_reclaim_thread_cv); 6035244781f1SPrakash Surya cv_destroy(&arc_reclaim_waiters_cv); 6036244781f1SPrakash Surya 6037dcbf3bd6SGeorge Wilson arc_state_fini(); 6038fa9e4066Sahrens buf_fini(); 60392fdbea25SAleksandr Guzovskiy 604089c86e32SChris Williamson ASSERT0(arc_loaned_bytes); 6041fa9e4066Sahrens } 6042fa94a07fSbrendan 6043fa94a07fSbrendan /* 6044fa94a07fSbrendan * Level 2 ARC 6045fa94a07fSbrendan * 6046fa94a07fSbrendan * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 6047fa94a07fSbrendan * It uses dedicated storage devices to hold cached data, which are populated 6048fa94a07fSbrendan * using large infrequent writes. The main role of this cache is to boost 6049fa94a07fSbrendan * the performance of random read workloads. The intended L2ARC devices 6050fa94a07fSbrendan * include short-stroked disks, solid state disks, and other media with 6051fa94a07fSbrendan * substantially faster read latency than disk. 6052fa94a07fSbrendan * 6053fa94a07fSbrendan * +-----------------------+ 6054fa94a07fSbrendan * | ARC | 6055fa94a07fSbrendan * +-----------------------+ 6056fa94a07fSbrendan * | ^ ^ 6057fa94a07fSbrendan * | | | 6058fa94a07fSbrendan * l2arc_feed_thread() arc_read() 6059fa94a07fSbrendan * | | | 6060fa94a07fSbrendan * | l2arc read | 6061fa94a07fSbrendan * V | | 6062fa94a07fSbrendan * +---------------+ | 6063fa94a07fSbrendan * | L2ARC | | 6064fa94a07fSbrendan * +---------------+ | 6065fa94a07fSbrendan * | ^ | 6066fa94a07fSbrendan * l2arc_write() | | 6067fa94a07fSbrendan * | | | 6068fa94a07fSbrendan * V | | 6069fa94a07fSbrendan * +-------+ +-------+ 6070fa94a07fSbrendan * | vdev | | vdev | 6071fa94a07fSbrendan * | cache | | cache | 6072fa94a07fSbrendan * +-------+ +-------+ 6073fa94a07fSbrendan * +=========+ .-----. 6074fa94a07fSbrendan * : L2ARC : |-_____-| 6075fa94a07fSbrendan * : devices : | Disks | 6076fa94a07fSbrendan * +=========+ `-_____-' 6077fa94a07fSbrendan * 6078fa94a07fSbrendan * Read requests are satisfied from the following sources, in order: 6079fa94a07fSbrendan * 6080fa94a07fSbrendan * 1) ARC 6081fa94a07fSbrendan * 2) vdev cache of L2ARC devices 6082fa94a07fSbrendan * 3) L2ARC devices 6083fa94a07fSbrendan * 4) vdev cache of disks 6084fa94a07fSbrendan * 5) disks 6085fa94a07fSbrendan * 6086fa94a07fSbrendan * Some L2ARC device types exhibit extremely slow write performance. 6087fa94a07fSbrendan * To accommodate for this there are some significant differences between 6088fa94a07fSbrendan * the L2ARC and traditional cache design: 6089fa94a07fSbrendan * 6090fa94a07fSbrendan * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 6091fa94a07fSbrendan * the ARC behave as usual, freeing buffers and placing headers on ghost 6092fa94a07fSbrendan * lists. The ARC does not send buffers to the L2ARC during eviction as 6093fa94a07fSbrendan * this would add inflated write latencies for all ARC memory pressure. 6094fa94a07fSbrendan * 6095fa94a07fSbrendan * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 6096fa94a07fSbrendan * It does this by periodically scanning buffers from the eviction-end of 6097fa94a07fSbrendan * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 6098aad02571SSaso Kiselkov * not already there. It scans until a headroom of buffers is satisfied, 6099aad02571SSaso Kiselkov * which itself is a buffer for ARC eviction. If a compressible buffer is 6100aad02571SSaso Kiselkov * found during scanning and selected for writing to an L2ARC device, we 6101aad02571SSaso Kiselkov * temporarily boost scanning headroom during the next scan cycle to make 6102aad02571SSaso Kiselkov * sure we adapt to compression effects (which might significantly reduce 6103aad02571SSaso Kiselkov * the data volume we write to L2ARC). The thread that does this is 6104fa94a07fSbrendan * l2arc_feed_thread(), illustrated below; example sizes are included to 6105fa94a07fSbrendan * provide a better sense of ratio than this diagram: 6106fa94a07fSbrendan * 6107fa94a07fSbrendan * head --> tail 6108fa94a07fSbrendan * +---------------------+----------+ 6109fa94a07fSbrendan * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 6110fa94a07fSbrendan * +---------------------+----------+ | o L2ARC eligible 6111fa94a07fSbrendan * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 6112fa94a07fSbrendan * +---------------------+----------+ | 6113fa94a07fSbrendan * 15.9 Gbytes ^ 32 Mbytes | 6114fa94a07fSbrendan * headroom | 6115fa94a07fSbrendan * l2arc_feed_thread() 6116fa94a07fSbrendan * | 6117fa94a07fSbrendan * l2arc write hand <--[oooo]--' 6118fa94a07fSbrendan * | 8 Mbyte 6119fa94a07fSbrendan * | write max 6120fa94a07fSbrendan * V 6121fa94a07fSbrendan * +==============================+ 6122fa94a07fSbrendan * L2ARC dev |####|#|###|###| |####| ... | 6123fa94a07fSbrendan * +==============================+ 6124fa94a07fSbrendan * 32 Gbytes 6125fa94a07fSbrendan * 6126fa94a07fSbrendan * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 6127fa94a07fSbrendan * evicted, then the L2ARC has cached a buffer much sooner than it probably 6128fa94a07fSbrendan * needed to, potentially wasting L2ARC device bandwidth and storage. It is 6129fa94a07fSbrendan * safe to say that this is an uncommon case, since buffers at the end of 6130fa94a07fSbrendan * the ARC lists have moved there due to inactivity. 6131fa94a07fSbrendan * 6132fa94a07fSbrendan * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 6133fa94a07fSbrendan * then the L2ARC simply misses copying some buffers. This serves as a 6134fa94a07fSbrendan * pressure valve to prevent heavy read workloads from both stalling the ARC 6135fa94a07fSbrendan * with waits and clogging the L2ARC with writes. This also helps prevent 6136fa94a07fSbrendan * the potential for the L2ARC to churn if it attempts to cache content too 6137fa94a07fSbrendan * quickly, such as during backups of the entire pool. 6138fa94a07fSbrendan * 61393a737e0dSbrendan * 5. After system boot and before the ARC has filled main memory, there are 61403a737e0dSbrendan * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 61413a737e0dSbrendan * lists can remain mostly static. Instead of searching from tail of these 61423a737e0dSbrendan * lists as pictured, the l2arc_feed_thread() will search from the list heads 61433a737e0dSbrendan * for eligible buffers, greatly increasing its chance of finding them. 61443a737e0dSbrendan * 61453a737e0dSbrendan * The L2ARC device write speed is also boosted during this time so that 61463a737e0dSbrendan * the L2ARC warms up faster. Since there have been no ARC evictions yet, 61473a737e0dSbrendan * there are no L2ARC reads, and no fear of degrading read performance 61483a737e0dSbrendan * through increased writes. 61493a737e0dSbrendan * 61503a737e0dSbrendan * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 6151fa94a07fSbrendan * the vdev queue can aggregate them into larger and fewer writes. Each 6152fa94a07fSbrendan * device is written to in a rotor fashion, sweeping writes through 6153fa94a07fSbrendan * available space then repeating. 6154fa94a07fSbrendan * 61553a737e0dSbrendan * 7. The L2ARC does not store dirty content. It never needs to flush 6156fa94a07fSbrendan * write buffers back to disk based storage. 6157fa94a07fSbrendan * 61583a737e0dSbrendan * 8. If an ARC buffer is written (and dirtied) which also exists in the 6159fa94a07fSbrendan * L2ARC, the now stale L2ARC buffer is immediately dropped. 6160fa94a07fSbrendan * 6161fa94a07fSbrendan * The performance of the L2ARC can be tweaked by a number of tunables, which 6162fa94a07fSbrendan * may be necessary for different workloads: 6163fa94a07fSbrendan * 6164fa94a07fSbrendan * l2arc_write_max max write bytes per interval 61653a737e0dSbrendan * l2arc_write_boost extra write bytes during device warmup 6166fa94a07fSbrendan * l2arc_noprefetch skip caching prefetched buffers 6167fa94a07fSbrendan * l2arc_headroom number of max device writes to precache 6168aad02571SSaso Kiselkov * l2arc_headroom_boost when we find compressed buffers during ARC 6169aad02571SSaso Kiselkov * scanning, we multiply headroom by this 6170aad02571SSaso Kiselkov * percentage factor for the next scan cycle, 6171aad02571SSaso Kiselkov * since more compressed buffers are likely to 6172aad02571SSaso Kiselkov * be present 6173fa94a07fSbrendan * l2arc_feed_secs seconds between L2ARC writing 6174fa94a07fSbrendan * 6175fa94a07fSbrendan * Tunables may be removed or added as future performance improvements are 6176fa94a07fSbrendan * integrated, and also may become zpool properties. 61775a98e54bSBrendan Gregg - Sun Microsystems * 61785a98e54bSBrendan Gregg - Sun Microsystems * There are three key functions that control how the L2ARC warms up: 61795a98e54bSBrendan Gregg - Sun Microsystems * 61805a98e54bSBrendan Gregg - Sun Microsystems * l2arc_write_eligible() check if a buffer is eligible to cache 61815a98e54bSBrendan Gregg - Sun Microsystems * l2arc_write_size() calculate how much to write 61825a98e54bSBrendan Gregg - Sun Microsystems * l2arc_write_interval() calculate sleep delay between writes 61835a98e54bSBrendan Gregg - Sun Microsystems * 61845a98e54bSBrendan Gregg - Sun Microsystems * These three functions determine what to write, how much, and how quickly 61855a98e54bSBrendan Gregg - Sun Microsystems * to send writes. 6186fa94a07fSbrendan */ 6187fa94a07fSbrendan 61885a98e54bSBrendan Gregg - Sun Microsystems static boolean_t 61897adb730bSGeorge Wilson l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 61905a98e54bSBrendan Gregg - Sun Microsystems { 61915a98e54bSBrendan Gregg - Sun Microsystems /* 61925a98e54bSBrendan Gregg - Sun Microsystems * A buffer is *not* eligible for the L2ARC if it: 61935a98e54bSBrendan Gregg - Sun Microsystems * 1. belongs to a different spa. 61945ea40c06SBrendan Gregg - Sun Microsystems * 2. is already cached on the L2ARC. 61955ea40c06SBrendan Gregg - Sun Microsystems * 3. has an I/O in progress (it may be an incomplete read). 61965ea40c06SBrendan Gregg - Sun Microsystems * 4. is flagged not eligible (zfs property). 61975a98e54bSBrendan Gregg - Sun Microsystems */ 619889c86e32SChris Williamson if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) || 61997adb730bSGeorge Wilson HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr)) 62005a98e54bSBrendan Gregg - Sun Microsystems return (B_FALSE); 62015a98e54bSBrendan Gregg - Sun Microsystems 62025a98e54bSBrendan Gregg - Sun Microsystems return (B_TRUE); 62035a98e54bSBrendan Gregg - Sun Microsystems } 62045a98e54bSBrendan Gregg - Sun Microsystems 62055a98e54bSBrendan Gregg - Sun Microsystems static uint64_t 6206aad02571SSaso Kiselkov l2arc_write_size(void) 62075a98e54bSBrendan Gregg - Sun Microsystems { 62085a98e54bSBrendan Gregg - Sun Microsystems uint64_t size; 62095a98e54bSBrendan Gregg - Sun Microsystems 6210aad02571SSaso Kiselkov /* 6211aad02571SSaso Kiselkov * Make sure our globals have meaningful values in case the user 6212aad02571SSaso Kiselkov * altered them. 6213aad02571SSaso Kiselkov */ 6214aad02571SSaso Kiselkov size = l2arc_write_max; 6215aad02571SSaso Kiselkov if (size == 0) { 6216aad02571SSaso Kiselkov cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 6217aad02571SSaso Kiselkov "be greater than zero, resetting it to the default (%d)", 6218aad02571SSaso Kiselkov L2ARC_WRITE_SIZE); 6219aad02571SSaso Kiselkov size = l2arc_write_max = L2ARC_WRITE_SIZE; 6220aad02571SSaso Kiselkov } 62215a98e54bSBrendan Gregg - Sun Microsystems 62225a98e54bSBrendan Gregg - Sun Microsystems if (arc_warm == B_FALSE) 6223aad02571SSaso Kiselkov size += l2arc_write_boost; 62245a98e54bSBrendan Gregg - Sun Microsystems 62255a98e54bSBrendan Gregg - Sun Microsystems return (size); 62265a98e54bSBrendan Gregg - Sun Microsystems 62275a98e54bSBrendan Gregg - Sun Microsystems } 62285a98e54bSBrendan Gregg - Sun Microsystems 62295a98e54bSBrendan Gregg - Sun Microsystems static clock_t 62305a98e54bSBrendan Gregg - Sun Microsystems l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 62315a98e54bSBrendan Gregg - Sun Microsystems { 6232d3d50737SRafael Vanoni clock_t interval, next, now; 62335a98e54bSBrendan Gregg - Sun Microsystems 62345a98e54bSBrendan Gregg - Sun Microsystems /* 62355a98e54bSBrendan Gregg - Sun Microsystems * If the ARC lists are busy, increase our write rate; if the 62365a98e54bSBrendan Gregg - Sun Microsystems * lists are stale, idle back. This is achieved by checking 62375a98e54bSBrendan Gregg - Sun Microsystems * how much we previously wrote - if it was more than half of 62385a98e54bSBrendan Gregg - Sun Microsystems * what we wanted, schedule the next write much sooner. 62395a98e54bSBrendan Gregg - Sun Microsystems */ 62405a98e54bSBrendan Gregg - Sun Microsystems if (l2arc_feed_again && wrote > (wanted / 2)) 62415a98e54bSBrendan Gregg - Sun Microsystems interval = (hz * l2arc_feed_min_ms) / 1000; 62425a98e54bSBrendan Gregg - Sun Microsystems else 62435a98e54bSBrendan Gregg - Sun Microsystems interval = hz * l2arc_feed_secs; 62445a98e54bSBrendan Gregg - Sun Microsystems 6245d3d50737SRafael Vanoni now = ddi_get_lbolt(); 6246d3d50737SRafael Vanoni next = MAX(now, MIN(now + interval, began + interval)); 62475a98e54bSBrendan Gregg - Sun Microsystems 62485a98e54bSBrendan Gregg - Sun Microsystems return (next); 62495a98e54bSBrendan Gregg - Sun Microsystems } 62505a98e54bSBrendan Gregg - Sun Microsystems 6251fa94a07fSbrendan /* 6252fa94a07fSbrendan * Cycle through L2ARC devices. This is how L2ARC load balances. 62533a737e0dSbrendan * If a device is returned, this also returns holding the spa config lock. 6254fa94a07fSbrendan */ 6255fa94a07fSbrendan static l2arc_dev_t * 6256fa94a07fSbrendan l2arc_dev_get_next(void) 6257fa94a07fSbrendan { 62583a737e0dSbrendan l2arc_dev_t *first, *next = NULL; 62593a737e0dSbrendan 62603a737e0dSbrendan /* 62613a737e0dSbrendan * Lock out the removal of spas (spa_namespace_lock), then removal 62623a737e0dSbrendan * of cache devices (l2arc_dev_mtx). Once a device has been selected, 62633a737e0dSbrendan * both locks will be dropped and a spa config lock held instead. 62643a737e0dSbrendan */ 62653a737e0dSbrendan mutex_enter(&spa_namespace_lock); 62663a737e0dSbrendan mutex_enter(&l2arc_dev_mtx); 6267fa94a07fSbrendan 6268c5904d13Seschrock /* if there are no vdevs, there is nothing to do */ 6269c5904d13Seschrock if (l2arc_ndev == 0) 62703a737e0dSbrendan goto out; 6271c5904d13Seschrock 6272c5904d13Seschrock first = NULL; 6273c5904d13Seschrock next = l2arc_dev_last; 6274c5904d13Seschrock do { 6275c5904d13Seschrock /* loop around the list looking for a non-faulted vdev */ 6276c5904d13Seschrock if (next == NULL) { 6277fa94a07fSbrendan next = list_head(l2arc_dev_list); 6278c5904d13Seschrock } else { 6279c5904d13Seschrock next = list_next(l2arc_dev_list, next); 6280c5904d13Seschrock if (next == NULL) 6281c5904d13Seschrock next = list_head(l2arc_dev_list); 6282c5904d13Seschrock } 6283c5904d13Seschrock 6284c5904d13Seschrock /* if we have come back to the start, bail out */ 6285c5904d13Seschrock if (first == NULL) 6286c5904d13Seschrock first = next; 6287c5904d13Seschrock else if (next == first) 6288c5904d13Seschrock break; 6289c5904d13Seschrock 6290c5904d13Seschrock } while (vdev_is_dead(next->l2ad_vdev)); 6291c5904d13Seschrock 6292c5904d13Seschrock /* if we were unable to find any usable vdevs, return NULL */ 6293c5904d13Seschrock if (vdev_is_dead(next->l2ad_vdev)) 62943a737e0dSbrendan next = NULL; 6295fa94a07fSbrendan 6296fa94a07fSbrendan l2arc_dev_last = next; 6297fa94a07fSbrendan 62983a737e0dSbrendan out: 62993a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 63003a737e0dSbrendan 63013a737e0dSbrendan /* 63023a737e0dSbrendan * Grab the config lock to prevent the 'next' device from being 63033a737e0dSbrendan * removed while we are writing to it. 63043a737e0dSbrendan */ 63053a737e0dSbrendan if (next != NULL) 6306e14bb325SJeff Bonwick spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 63073a737e0dSbrendan mutex_exit(&spa_namespace_lock); 63083a737e0dSbrendan 6309fa94a07fSbrendan return (next); 6310fa94a07fSbrendan } 6311fa94a07fSbrendan 63123a737e0dSbrendan /* 63133a737e0dSbrendan * Free buffers that were tagged for destruction. 63143a737e0dSbrendan */ 63153a737e0dSbrendan static void 63163a737e0dSbrendan l2arc_do_free_on_write() 63173a737e0dSbrendan { 63183a737e0dSbrendan list_t *buflist; 63193a737e0dSbrendan l2arc_data_free_t *df, *df_prev; 63203a737e0dSbrendan 63213a737e0dSbrendan mutex_enter(&l2arc_free_on_write_mtx); 63223a737e0dSbrendan buflist = l2arc_free_on_write; 63233a737e0dSbrendan 63243a737e0dSbrendan for (df = list_tail(buflist); df; df = df_prev) { 63253a737e0dSbrendan df_prev = list_prev(buflist, df); 6326dcbf3bd6SGeorge Wilson ASSERT3P(df->l2df_data, !=, NULL); 6327dcbf3bd6SGeorge Wilson if (df->l2df_type == ARC_BUFC_METADATA) { 6328dcbf3bd6SGeorge Wilson zio_buf_free(df->l2df_data, df->l2df_size); 6329dcbf3bd6SGeorge Wilson } else { 6330dcbf3bd6SGeorge Wilson ASSERT(df->l2df_type == ARC_BUFC_DATA); 6331dcbf3bd6SGeorge Wilson zio_data_buf_free(df->l2df_data, df->l2df_size); 6332dcbf3bd6SGeorge Wilson } 63333a737e0dSbrendan list_remove(buflist, df); 63343a737e0dSbrendan kmem_free(df, sizeof (l2arc_data_free_t)); 63353a737e0dSbrendan } 63363a737e0dSbrendan 63373a737e0dSbrendan mutex_exit(&l2arc_free_on_write_mtx); 63383a737e0dSbrendan } 63393a737e0dSbrendan 6340fa94a07fSbrendan /* 6341fa94a07fSbrendan * A write to a cache device has completed. Update all headers to allow 6342fa94a07fSbrendan * reads from these buffers to begin. 6343fa94a07fSbrendan */ 6344fa94a07fSbrendan static void 6345fa94a07fSbrendan l2arc_write_done(zio_t *zio) 6346fa94a07fSbrendan { 6347fa94a07fSbrendan l2arc_write_callback_t *cb; 6348fa94a07fSbrendan l2arc_dev_t *dev; 6349fa94a07fSbrendan list_t *buflist; 63507adb730bSGeorge Wilson arc_buf_hdr_t *head, *hdr, *hdr_prev; 6351fa94a07fSbrendan kmutex_t *hash_lock; 63523038a2b4SSaso Kiselkov int64_t bytes_dropped = 0; 6353fa94a07fSbrendan 6354fa94a07fSbrendan cb = zio->io_private; 6355dcbf3bd6SGeorge Wilson ASSERT3P(cb, !=, NULL); 6356fa94a07fSbrendan dev = cb->l2wcb_dev; 6357dcbf3bd6SGeorge Wilson ASSERT3P(dev, !=, NULL); 6358fa94a07fSbrendan head = cb->l2wcb_head; 6359dcbf3bd6SGeorge Wilson ASSERT3P(head, !=, NULL); 636089c86e32SChris Williamson buflist = &dev->l2ad_buflist; 6361dcbf3bd6SGeorge Wilson ASSERT3P(buflist, !=, NULL); 6362fa94a07fSbrendan DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 6363fa94a07fSbrendan l2arc_write_callback_t *, cb); 6364fa94a07fSbrendan 6365fa94a07fSbrendan if (zio->io_error != 0) 6366fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_writes_error); 6367fa94a07fSbrendan 6368fa94a07fSbrendan /* 6369fa94a07fSbrendan * All writes completed, or an error was hit. 6370fa94a07fSbrendan */ 6371244781f1SPrakash Surya top: 6372244781f1SPrakash Surya mutex_enter(&dev->l2ad_mtx); 63737adb730bSGeorge Wilson for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 63747adb730bSGeorge Wilson hdr_prev = list_prev(buflist, hdr); 6375fa94a07fSbrendan 63767adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 6377244781f1SPrakash Surya 6378244781f1SPrakash Surya /* 6379244781f1SPrakash Surya * We cannot use mutex_enter or else we can deadlock 6380244781f1SPrakash Surya * with l2arc_write_buffers (due to swapping the order 6381244781f1SPrakash Surya * the hash lock and l2ad_mtx are taken). 6382244781f1SPrakash Surya */ 6383fa94a07fSbrendan if (!mutex_tryenter(hash_lock)) { 6384fa94a07fSbrendan /* 6385244781f1SPrakash Surya * Missed the hash lock. We must retry so we 6386244781f1SPrakash Surya * don't leave the ARC_FLAG_L2_WRITING bit set. 6387fa94a07fSbrendan */ 6388244781f1SPrakash Surya ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 6389244781f1SPrakash Surya 6390244781f1SPrakash Surya /* 6391244781f1SPrakash Surya * We don't want to rescan the headers we've 6392244781f1SPrakash Surya * already marked as having been written out, so 6393244781f1SPrakash Surya * we reinsert the head node so we can pick up 6394244781f1SPrakash Surya * where we left off. 6395244781f1SPrakash Surya */ 6396244781f1SPrakash Surya list_remove(buflist, head); 6397244781f1SPrakash Surya list_insert_after(buflist, hdr, head); 6398244781f1SPrakash Surya 6399244781f1SPrakash Surya mutex_exit(&dev->l2ad_mtx); 6400244781f1SPrakash Surya 6401244781f1SPrakash Surya /* 6402244781f1SPrakash Surya * We wait for the hash lock to become available 6403244781f1SPrakash Surya * to try and prevent busy waiting, and increase 6404244781f1SPrakash Surya * the chance we'll be able to acquire the lock 6405244781f1SPrakash Surya * the next time around. 6406244781f1SPrakash Surya */ 6407244781f1SPrakash Surya mutex_enter(hash_lock); 6408244781f1SPrakash Surya mutex_exit(hash_lock); 6409244781f1SPrakash Surya goto top; 6410fa94a07fSbrendan } 6411fa94a07fSbrendan 641289c86e32SChris Williamson /* 6413244781f1SPrakash Surya * We could not have been moved into the arc_l2c_only 6414244781f1SPrakash Surya * state while in-flight due to our ARC_FLAG_L2_WRITING 6415244781f1SPrakash Surya * bit being set. Let's just ensure that's being enforced. 6416244781f1SPrakash Surya */ 6417244781f1SPrakash Surya ASSERT(HDR_HAS_L1HDR(hdr)); 6418244781f1SPrakash Surya 6419fa94a07fSbrendan if (zio->io_error != 0) { 6420fa94a07fSbrendan /* 64213a737e0dSbrendan * Error - drop L2ARC entry. 6422fa94a07fSbrendan */ 64237adb730bSGeorge Wilson list_remove(buflist, hdr); 6424dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 642589c86e32SChris Williamson 6426dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_l2_asize, -arc_hdr_size(hdr)); 6427dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr)); 6428a52fc310SPrakash Surya 6429dcbf3bd6SGeorge Wilson bytes_dropped += arc_hdr_size(hdr); 6430a52fc310SPrakash Surya (void) refcount_remove_many(&dev->l2ad_alloc, 6431dcbf3bd6SGeorge Wilson arc_hdr_size(hdr), hdr); 6432fa94a07fSbrendan } 6433fa94a07fSbrendan 6434fa94a07fSbrendan /* 6435244781f1SPrakash Surya * Allow ARC to begin reads and ghost list evictions to 6436244781f1SPrakash Surya * this L2ARC entry. 6437fa94a07fSbrendan */ 6438dcbf3bd6SGeorge Wilson arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); 6439fa94a07fSbrendan 6440fa94a07fSbrendan mutex_exit(hash_lock); 6441fa94a07fSbrendan } 6442fa94a07fSbrendan 6443fa94a07fSbrendan atomic_inc_64(&l2arc_writes_done); 6444fa94a07fSbrendan list_remove(buflist, head); 644589c86e32SChris Williamson ASSERT(!HDR_HAS_L1HDR(head)); 644689c86e32SChris Williamson kmem_cache_free(hdr_l2only_cache, head); 644789c86e32SChris Williamson mutex_exit(&dev->l2ad_mtx); 6448fa94a07fSbrendan 64493038a2b4SSaso Kiselkov vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 64503038a2b4SSaso Kiselkov 64513a737e0dSbrendan l2arc_do_free_on_write(); 6452fa94a07fSbrendan 6453fa94a07fSbrendan kmem_free(cb, sizeof (l2arc_write_callback_t)); 6454fa94a07fSbrendan } 6455fa94a07fSbrendan 6456fa94a07fSbrendan /* 6457fa94a07fSbrendan * A read to a cache device completed. Validate buffer contents before 6458fa94a07fSbrendan * handing over to the regular ARC routines. 6459fa94a07fSbrendan */ 6460fa94a07fSbrendan static void 6461fa94a07fSbrendan l2arc_read_done(zio_t *zio) 6462fa94a07fSbrendan { 6463fa94a07fSbrendan l2arc_read_callback_t *cb; 6464fa94a07fSbrendan arc_buf_hdr_t *hdr; 6465fa94a07fSbrendan kmutex_t *hash_lock; 6466dcbf3bd6SGeorge Wilson boolean_t valid_cksum; 6467fa94a07fSbrendan 6468dcbf3bd6SGeorge Wilson ASSERT3P(zio->io_vd, !=, NULL); 6469e14bb325SJeff Bonwick ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 6470e14bb325SJeff Bonwick 6471e14bb325SJeff Bonwick spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 6472e14bb325SJeff Bonwick 6473fa94a07fSbrendan cb = zio->io_private; 6474dcbf3bd6SGeorge Wilson ASSERT3P(cb, !=, NULL); 6475dcbf3bd6SGeorge Wilson hdr = cb->l2rcb_hdr; 6476dcbf3bd6SGeorge Wilson ASSERT3P(hdr, !=, NULL); 6477fa94a07fSbrendan 6478dcbf3bd6SGeorge Wilson hash_lock = HDR_LOCK(hdr); 6479fa94a07fSbrendan mutex_enter(hash_lock); 64803f9d6ad7SLin Ling ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 6481fa94a07fSbrendan 6482dcbf3bd6SGeorge Wilson ASSERT3P(zio->io_data, !=, NULL); 6483aad02571SSaso Kiselkov 6484fa94a07fSbrendan /* 6485fa94a07fSbrendan * Check this survived the L2ARC journey. 6486fa94a07fSbrendan */ 6487dcbf3bd6SGeorge Wilson ASSERT3P(zio->io_data, ==, hdr->b_l1hdr.b_pdata); 6488dcbf3bd6SGeorge Wilson zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 6489dcbf3bd6SGeorge Wilson zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 6490dcbf3bd6SGeorge Wilson 6491dcbf3bd6SGeorge Wilson valid_cksum = arc_cksum_is_equal(hdr, zio); 6492dcbf3bd6SGeorge Wilson if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 6493fa94a07fSbrendan mutex_exit(hash_lock); 6494dcbf3bd6SGeorge Wilson zio->io_private = hdr; 6495fa94a07fSbrendan arc_read_done(zio); 6496fa94a07fSbrendan } else { 6497fa94a07fSbrendan mutex_exit(hash_lock); 6498fa94a07fSbrendan /* 6499fa94a07fSbrendan * Buffer didn't survive caching. Increment stats and 6500fa94a07fSbrendan * reissue to the original storage device. 6501fa94a07fSbrendan */ 65023a737e0dSbrendan if (zio->io_error != 0) { 6503fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_io_error); 65043a737e0dSbrendan } else { 6505be6fd75aSMatthew Ahrens zio->io_error = SET_ERROR(EIO); 65063a737e0dSbrendan } 6507dcbf3bd6SGeorge Wilson if (!valid_cksum) 6508fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_cksum_bad); 6509fa94a07fSbrendan 6510e14bb325SJeff Bonwick /* 6511e14bb325SJeff Bonwick * If there's no waiter, issue an async i/o to the primary 6512e14bb325SJeff Bonwick * storage now. If there *is* a waiter, the caller must 6513e14bb325SJeff Bonwick * issue the i/o in a context where it's OK to block. 6514e14bb325SJeff Bonwick */ 6515a3f829aeSBill Moore if (zio->io_waiter == NULL) { 6516a3f829aeSBill Moore zio_t *pio = zio_unique_parent(zio); 6517a3f829aeSBill Moore 6518a3f829aeSBill Moore ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6519a3f829aeSBill Moore 6520dcbf3bd6SGeorge Wilson zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, 6521dcbf3bd6SGeorge Wilson hdr->b_l1hdr.b_pdata, zio->io_size, arc_read_done, 6522dcbf3bd6SGeorge Wilson hdr, zio->io_priority, cb->l2rcb_flags, 6523dcbf3bd6SGeorge Wilson &cb->l2rcb_zb)); 6524a3f829aeSBill Moore } 6525fa94a07fSbrendan } 6526fa94a07fSbrendan 6527fa94a07fSbrendan kmem_free(cb, sizeof (l2arc_read_callback_t)); 6528fa94a07fSbrendan } 6529fa94a07fSbrendan 6530fa94a07fSbrendan /* 6531fa94a07fSbrendan * This is the list priority from which the L2ARC will search for pages to 6532fa94a07fSbrendan * cache. This is used within loops (0..3) to cycle through lists in the 6533fa94a07fSbrendan * desired order. This order can have a significant effect on cache 6534fa94a07fSbrendan * performance. 6535fa94a07fSbrendan * 6536fa94a07fSbrendan * Currently the metadata lists are hit first, MFU then MRU, followed by 6537fa94a07fSbrendan * the data lists. This function returns a locked list, and also returns 6538fa94a07fSbrendan * the lock pointer. 6539fa94a07fSbrendan */ 6540244781f1SPrakash Surya static multilist_sublist_t * 6541244781f1SPrakash Surya l2arc_sublist_lock(int list_num) 6542fa94a07fSbrendan { 6543244781f1SPrakash Surya multilist_t *ml = NULL; 6544244781f1SPrakash Surya unsigned int idx; 6545fa94a07fSbrendan 6546fa94a07fSbrendan ASSERT(list_num >= 0 && list_num <= 3); 6547fa94a07fSbrendan 6548fa94a07fSbrendan switch (list_num) { 6549fa94a07fSbrendan case 0: 6550244781f1SPrakash Surya ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6551fa94a07fSbrendan break; 6552fa94a07fSbrendan case 1: 6553244781f1SPrakash Surya ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 6554fa94a07fSbrendan break; 6555fa94a07fSbrendan case 2: 6556244781f1SPrakash Surya ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 6557fa94a07fSbrendan break; 6558fa94a07fSbrendan case 3: 6559244781f1SPrakash Surya ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; 6560fa94a07fSbrendan break; 6561fa94a07fSbrendan } 6562fa94a07fSbrendan 6563244781f1SPrakash Surya /* 6564244781f1SPrakash Surya * Return a randomly-selected sublist. This is acceptable 6565244781f1SPrakash Surya * because the caller feeds only a little bit of data for each 6566244781f1SPrakash Surya * call (8MB). Subsequent calls will result in different 6567244781f1SPrakash Surya * sublists being selected. 6568244781f1SPrakash Surya */ 6569244781f1SPrakash Surya idx = multilist_get_random_index(ml); 6570244781f1SPrakash Surya return (multilist_sublist_lock(ml, idx)); 6571fa94a07fSbrendan } 6572fa94a07fSbrendan 6573fa94a07fSbrendan /* 6574fa94a07fSbrendan * Evict buffers from the device write hand to the distance specified in 6575fa94a07fSbrendan * bytes. This distance may span populated buffers, it may span nothing. 6576fa94a07fSbrendan * This is clearing a region on the L2ARC device ready for writing. 6577fa94a07fSbrendan * If the 'all' boolean is set, every buffer is evicted. 6578fa94a07fSbrendan */ 6579fa94a07fSbrendan static void 6580fa94a07fSbrendan l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6581fa94a07fSbrendan { 6582fa94a07fSbrendan list_t *buflist; 65837adb730bSGeorge Wilson arc_buf_hdr_t *hdr, *hdr_prev; 6584fa94a07fSbrendan kmutex_t *hash_lock; 6585fa94a07fSbrendan uint64_t taddr; 6586fa94a07fSbrendan 658789c86e32SChris Williamson buflist = &dev->l2ad_buflist; 6588fa94a07fSbrendan 6589fa94a07fSbrendan if (!all && dev->l2ad_first) { 6590fa94a07fSbrendan /* 6591fa94a07fSbrendan * This is the first sweep through the device. There is 6592fa94a07fSbrendan * nothing to evict. 6593fa94a07fSbrendan */ 6594fa94a07fSbrendan return; 6595fa94a07fSbrendan } 6596fa94a07fSbrendan 65973a737e0dSbrendan if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6598fa94a07fSbrendan /* 6599fa94a07fSbrendan * When nearing the end of the device, evict to the end 6600fa94a07fSbrendan * before the device write hand jumps to the start. 6601fa94a07fSbrendan */ 6602fa94a07fSbrendan taddr = dev->l2ad_end; 6603fa94a07fSbrendan } else { 6604fa94a07fSbrendan taddr = dev->l2ad_hand + distance; 6605fa94a07fSbrendan } 6606fa94a07fSbrendan DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6607fa94a07fSbrendan uint64_t, taddr, boolean_t, all); 6608fa94a07fSbrendan 6609fa94a07fSbrendan top: 661089c86e32SChris Williamson mutex_enter(&dev->l2ad_mtx); 66117adb730bSGeorge Wilson for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 66127adb730bSGeorge Wilson hdr_prev = list_prev(buflist, hdr); 6613fa94a07fSbrendan 66147adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 6615244781f1SPrakash Surya 6616244781f1SPrakash Surya /* 6617244781f1SPrakash Surya * We cannot use mutex_enter or else we can deadlock 6618244781f1SPrakash Surya * with l2arc_write_buffers (due to swapping the order 6619244781f1SPrakash Surya * the hash lock and l2ad_mtx are taken). 6620244781f1SPrakash Surya */ 6621fa94a07fSbrendan if (!mutex_tryenter(hash_lock)) { 6622fa94a07fSbrendan /* 6623fa94a07fSbrendan * Missed the hash lock. Retry. 6624fa94a07fSbrendan */ 6625fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 662689c86e32SChris Williamson mutex_exit(&dev->l2ad_mtx); 6627fa94a07fSbrendan mutex_enter(hash_lock); 6628fa94a07fSbrendan mutex_exit(hash_lock); 6629fa94a07fSbrendan goto top; 6630fa94a07fSbrendan } 6631fa94a07fSbrendan 66327adb730bSGeorge Wilson if (HDR_L2_WRITE_HEAD(hdr)) { 6633fa94a07fSbrendan /* 6634fa94a07fSbrendan * We hit a write head node. Leave it for 6635fa94a07fSbrendan * l2arc_write_done(). 6636fa94a07fSbrendan */ 66377adb730bSGeorge Wilson list_remove(buflist, hdr); 6638fa94a07fSbrendan mutex_exit(hash_lock); 6639fa94a07fSbrendan continue; 6640fa94a07fSbrendan } 6641fa94a07fSbrendan 664289c86e32SChris Williamson if (!all && HDR_HAS_L2HDR(hdr) && 664389c86e32SChris Williamson (hdr->b_l2hdr.b_daddr > taddr || 664489c86e32SChris Williamson hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6645fa94a07fSbrendan /* 6646fa94a07fSbrendan * We've evicted to the target address, 6647fa94a07fSbrendan * or the end of the device. 6648fa94a07fSbrendan */ 6649fa94a07fSbrendan mutex_exit(hash_lock); 6650fa94a07fSbrendan break; 6651fa94a07fSbrendan } 6652fa94a07fSbrendan 665389c86e32SChris Williamson ASSERT(HDR_HAS_L2HDR(hdr)); 665489c86e32SChris Williamson if (!HDR_HAS_L1HDR(hdr)) { 66557adb730bSGeorge Wilson ASSERT(!HDR_L2_READING(hdr)); 6656fa94a07fSbrendan /* 6657fa94a07fSbrendan * This doesn't exist in the ARC. Destroy. 6658fa94a07fSbrendan * arc_hdr_destroy() will call list_remove() 6659fa94a07fSbrendan * and decrement arcstat_l2_size. 6660fa94a07fSbrendan */ 66617adb730bSGeorge Wilson arc_change_state(arc_anon, hdr, hash_lock); 66627adb730bSGeorge Wilson arc_hdr_destroy(hdr); 6663fa94a07fSbrendan } else { 666489c86e32SChris Williamson ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 666589c86e32SChris Williamson ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 66663a737e0dSbrendan /* 66673a737e0dSbrendan * Invalidate issued or about to be issued 66683a737e0dSbrendan * reads, since we may be about to write 66693a737e0dSbrendan * over this location. 66703a737e0dSbrendan */ 66717adb730bSGeorge Wilson if (HDR_L2_READING(hdr)) { 66723a737e0dSbrendan ARCSTAT_BUMP(arcstat_l2_evict_reading); 6673dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); 66743a737e0dSbrendan } 66753a737e0dSbrendan 6676244781f1SPrakash Surya /* Ensure this header has finished being written */ 6677244781f1SPrakash Surya ASSERT(!HDR_L2_WRITING(hdr)); 6678a52fc310SPrakash Surya 6679a52fc310SPrakash Surya arc_hdr_l2hdr_destroy(hdr); 6680fa94a07fSbrendan } 6681fa94a07fSbrendan mutex_exit(hash_lock); 6682fa94a07fSbrendan } 668389c86e32SChris Williamson mutex_exit(&dev->l2ad_mtx); 6684fa94a07fSbrendan } 6685fa94a07fSbrendan 6686fa94a07fSbrendan /* 6687fa94a07fSbrendan * Find and write ARC buffers to the L2ARC device. 6688fa94a07fSbrendan * 66897adb730bSGeorge Wilson * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6690fa94a07fSbrendan * for reading until they have completed writing. 6691aad02571SSaso Kiselkov * The headroom_boost is an in-out parameter used to maintain headroom boost 6692aad02571SSaso Kiselkov * state between calls to this function. 6693aad02571SSaso Kiselkov * 6694aad02571SSaso Kiselkov * Returns the number of bytes actually written (which may be smaller than 6695aad02571SSaso Kiselkov * the delta by which the device hand has changed due to alignment). 6696fa94a07fSbrendan */ 66975a98e54bSBrendan Gregg - Sun Microsystems static uint64_t 6698dcbf3bd6SGeorge Wilson l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 6699fa94a07fSbrendan { 67007adb730bSGeorge Wilson arc_buf_hdr_t *hdr, *hdr_prev, *head; 6701dcbf3bd6SGeorge Wilson uint64_t write_asize, write_psize, write_sz, headroom; 6702aad02571SSaso Kiselkov boolean_t full; 6703fa94a07fSbrendan l2arc_write_callback_t *cb; 6704fa94a07fSbrendan zio_t *pio, *wzio; 6705e9103aaeSGarrett D'Amore uint64_t guid = spa_load_guid(spa); 6706fa94a07fSbrendan 6707dcbf3bd6SGeorge Wilson ASSERT3P(dev->l2ad_vdev, !=, NULL); 6708aad02571SSaso Kiselkov 6709fa94a07fSbrendan pio = NULL; 6710dcbf3bd6SGeorge Wilson write_sz = write_asize = write_psize = 0; 6711fa94a07fSbrendan full = B_FALSE; 671289c86e32SChris Williamson head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 6713dcbf3bd6SGeorge Wilson arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); 6714aad02571SSaso Kiselkov 6715fa94a07fSbrendan /* 6716fa94a07fSbrendan * Copy buffers for L2ARC writing. 6717fa94a07fSbrendan */ 6718fa94a07fSbrendan for (int try = 0; try <= 3; try++) { 6719244781f1SPrakash Surya multilist_sublist_t *mls = l2arc_sublist_lock(try); 6720aad02571SSaso Kiselkov uint64_t passed_sz = 0; 6721aad02571SSaso Kiselkov 67223a737e0dSbrendan /* 67233a737e0dSbrendan * L2ARC fast warmup. 67243a737e0dSbrendan * 67253a737e0dSbrendan * Until the ARC is warm and starts to evict, read from the 67263a737e0dSbrendan * head of the ARC lists rather than the tail. 67273a737e0dSbrendan */ 67283a737e0dSbrendan if (arc_warm == B_FALSE) 6729244781f1SPrakash Surya hdr = multilist_sublist_head(mls); 67303a737e0dSbrendan else 6731244781f1SPrakash Surya hdr = multilist_sublist_tail(mls); 67323a737e0dSbrendan 6733aad02571SSaso Kiselkov headroom = target_sz * l2arc_headroom; 6734dcbf3bd6SGeorge Wilson if (zfs_compressed_arc_enabled) 6735aad02571SSaso Kiselkov headroom = (headroom * l2arc_headroom_boost) / 100; 6736aad02571SSaso Kiselkov 67377adb730bSGeorge Wilson for (; hdr; hdr = hdr_prev) { 6738aad02571SSaso Kiselkov kmutex_t *hash_lock; 6739aad02571SSaso Kiselkov 67403a737e0dSbrendan if (arc_warm == B_FALSE) 6741244781f1SPrakash Surya hdr_prev = multilist_sublist_next(mls, hdr); 67423a737e0dSbrendan else 6743244781f1SPrakash Surya hdr_prev = multilist_sublist_prev(mls, hdr); 6744fa94a07fSbrendan 67457adb730bSGeorge Wilson hash_lock = HDR_LOCK(hdr); 6746aad02571SSaso Kiselkov if (!mutex_tryenter(hash_lock)) { 6747fa94a07fSbrendan /* 6748fa94a07fSbrendan * Skip this buffer rather than waiting. 6749fa94a07fSbrendan */ 6750fa94a07fSbrendan continue; 6751fa94a07fSbrendan } 6752fa94a07fSbrendan 6753dcbf3bd6SGeorge Wilson passed_sz += HDR_GET_LSIZE(hdr); 6754fa94a07fSbrendan if (passed_sz > headroom) { 6755fa94a07fSbrendan /* 6756fa94a07fSbrendan * Searched too far. 6757fa94a07fSbrendan */ 6758fa94a07fSbrendan mutex_exit(hash_lock); 6759fa94a07fSbrendan break; 6760fa94a07fSbrendan } 6761fa94a07fSbrendan 67627adb730bSGeorge Wilson if (!l2arc_write_eligible(guid, hdr)) { 6763fa94a07fSbrendan mutex_exit(hash_lock); 6764fa94a07fSbrendan continue; 6765fa94a07fSbrendan } 6766fa94a07fSbrendan 6767dcbf3bd6SGeorge Wilson if ((write_asize + HDR_GET_LSIZE(hdr)) > target_sz) { 6768fa94a07fSbrendan full = B_TRUE; 6769fa94a07fSbrendan mutex_exit(hash_lock); 6770fa94a07fSbrendan break; 6771fa94a07fSbrendan } 6772fa94a07fSbrendan 6773fa94a07fSbrendan if (pio == NULL) { 6774fa94a07fSbrendan /* 6775fa94a07fSbrendan * Insert a dummy header on the buflist so 6776fa94a07fSbrendan * l2arc_write_done() can find where the 6777fa94a07fSbrendan * write buffers begin without searching. 6778fa94a07fSbrendan */ 6779244781f1SPrakash Surya mutex_enter(&dev->l2ad_mtx); 678089c86e32SChris Williamson list_insert_head(&dev->l2ad_buflist, head); 6781244781f1SPrakash Surya mutex_exit(&dev->l2ad_mtx); 6782fa94a07fSbrendan 6783fa94a07fSbrendan cb = kmem_alloc( 6784fa94a07fSbrendan sizeof (l2arc_write_callback_t), KM_SLEEP); 6785fa94a07fSbrendan cb->l2wcb_dev = dev; 6786fa94a07fSbrendan cb->l2wcb_head = head; 6787fa94a07fSbrendan pio = zio_root(spa, l2arc_write_done, cb, 6788fa94a07fSbrendan ZIO_FLAG_CANFAIL); 6789fa94a07fSbrendan } 6790fa94a07fSbrendan 679189c86e32SChris Williamson hdr->b_l2hdr.b_dev = dev; 6792dcbf3bd6SGeorge Wilson hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 6793dcbf3bd6SGeorge Wilson arc_hdr_set_flags(hdr, 6794dcbf3bd6SGeorge Wilson ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR); 6795dcbf3bd6SGeorge Wilson 6796dcbf3bd6SGeorge Wilson mutex_enter(&dev->l2ad_mtx); 6797dcbf3bd6SGeorge Wilson list_insert_head(&dev->l2ad_buflist, hdr); 6798dcbf3bd6SGeorge Wilson mutex_exit(&dev->l2ad_mtx); 6799dcbf3bd6SGeorge Wilson 6800aad02571SSaso Kiselkov /* 6801dcbf3bd6SGeorge Wilson * We rely on the L1 portion of the header below, so 6802dcbf3bd6SGeorge Wilson * it's invalid for this header to have been evicted out 6803dcbf3bd6SGeorge Wilson * of the ghost cache, prior to being written out. The 6804dcbf3bd6SGeorge Wilson * ARC_FLAG_L2_WRITING bit ensures this won't happen. 6805aad02571SSaso Kiselkov */ 6806dcbf3bd6SGeorge Wilson ASSERT(HDR_HAS_L1HDR(hdr)); 6807dcbf3bd6SGeorge Wilson 6808dcbf3bd6SGeorge Wilson ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); 6809dcbf3bd6SGeorge Wilson ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 6810dcbf3bd6SGeorge Wilson ASSERT3U(arc_hdr_size(hdr), >, 0); 6811dcbf3bd6SGeorge Wilson uint64_t size = arc_hdr_size(hdr); 6812dcbf3bd6SGeorge Wilson 6813dcbf3bd6SGeorge Wilson (void) refcount_add_many(&dev->l2ad_alloc, size, hdr); 6814aad02571SSaso Kiselkov 6815a52fc310SPrakash Surya /* 6816dcbf3bd6SGeorge Wilson * Normally the L2ARC can use the hdr's data, but if 6817dcbf3bd6SGeorge Wilson * we're sharing data between the hdr and one of its 6818dcbf3bd6SGeorge Wilson * bufs, L2ARC needs its own copy of the data so that 6819dcbf3bd6SGeorge Wilson * the ZIO below can't race with the buf consumer. To 6820dcbf3bd6SGeorge Wilson * ensure that this copy will be available for the 6821dcbf3bd6SGeorge Wilson * lifetime of the ZIO and be cleaned up afterwards, we 6822dcbf3bd6SGeorge Wilson * add it to the l2arc_free_on_write queue. 6823a52fc310SPrakash Surya */ 6824dcbf3bd6SGeorge Wilson void *to_write; 6825dcbf3bd6SGeorge Wilson if (!HDR_SHARED_DATA(hdr)) { 6826dcbf3bd6SGeorge Wilson to_write = hdr->b_l1hdr.b_pdata; 6827dcbf3bd6SGeorge Wilson } else { 6828dcbf3bd6SGeorge Wilson arc_buf_contents_t type = arc_buf_type(hdr); 6829dcbf3bd6SGeorge Wilson if (type == ARC_BUFC_METADATA) { 6830dcbf3bd6SGeorge Wilson to_write = zio_buf_alloc(size); 6831dcbf3bd6SGeorge Wilson } else { 6832dcbf3bd6SGeorge Wilson ASSERT3U(type, ==, ARC_BUFC_DATA); 6833dcbf3bd6SGeorge Wilson to_write = zio_data_buf_alloc(size); 6834dcbf3bd6SGeorge Wilson } 6835a52fc310SPrakash Surya 6836dcbf3bd6SGeorge Wilson bcopy(hdr->b_l1hdr.b_pdata, to_write, size); 6837dcbf3bd6SGeorge Wilson l2arc_free_data_on_write(to_write, size, type); 6838dcbf3bd6SGeorge Wilson } 6839dcbf3bd6SGeorge Wilson wzio = zio_write_phys(pio, dev->l2ad_vdev, 6840dcbf3bd6SGeorge Wilson hdr->b_l2hdr.b_daddr, size, to_write, 6841dcbf3bd6SGeorge Wilson ZIO_CHECKSUM_OFF, NULL, hdr, 6842dcbf3bd6SGeorge Wilson ZIO_PRIORITY_ASYNC_WRITE, 6843dcbf3bd6SGeorge Wilson ZIO_FLAG_CANFAIL, B_FALSE); 6844aad02571SSaso Kiselkov 6845dcbf3bd6SGeorge Wilson write_sz += HDR_GET_LSIZE(hdr); 6846dcbf3bd6SGeorge Wilson DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 6847dcbf3bd6SGeorge Wilson zio_t *, wzio); 6848fa94a07fSbrendan 6849dcbf3bd6SGeorge Wilson write_asize += size; 6850fa94a07fSbrendan /* 6851dcbf3bd6SGeorge Wilson * Keep the clock hand suitably device-aligned. 6852fa94a07fSbrendan */ 6853dcbf3bd6SGeorge Wilson uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, 6854dcbf3bd6SGeorge Wilson size); 6855dcbf3bd6SGeorge Wilson write_psize += asize; 6856dcbf3bd6SGeorge Wilson dev->l2ad_hand += asize; 6857fa94a07fSbrendan 6858fa94a07fSbrendan mutex_exit(hash_lock); 6859fa94a07fSbrendan 6860dcbf3bd6SGeorge Wilson (void) zio_nowait(wzio); 6861aad02571SSaso Kiselkov } 6862aad02571SSaso Kiselkov 6863244781f1SPrakash Surya multilist_sublist_unlock(mls); 6864aad02571SSaso Kiselkov 6865aad02571SSaso Kiselkov if (full == B_TRUE) 6866aad02571SSaso Kiselkov break; 6867aad02571SSaso Kiselkov } 6868aad02571SSaso Kiselkov 6869aad02571SSaso Kiselkov /* No buffers selected for writing? */ 6870aad02571SSaso Kiselkov if (pio == NULL) { 6871aad02571SSaso Kiselkov ASSERT0(write_sz); 687289c86e32SChris Williamson ASSERT(!HDR_HAS_L1HDR(head)); 687389c86e32SChris Williamson kmem_cache_free(hdr_l2only_cache, head); 6874aad02571SSaso Kiselkov return (0); 6875aad02571SSaso Kiselkov } 6876aad02571SSaso Kiselkov 6877aad02571SSaso Kiselkov ASSERT3U(write_asize, <=, target_sz); 6878fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_writes_sent); 6879aad02571SSaso Kiselkov ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 6880fa94a07fSbrendan ARCSTAT_INCR(arcstat_l2_size, write_sz); 6881dcbf3bd6SGeorge Wilson ARCSTAT_INCR(arcstat_l2_asize, write_asize); 6882dcbf3bd6SGeorge Wilson vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 6883fa94a07fSbrendan 6884fa94a07fSbrendan /* 6885fa94a07fSbrendan * Bump device hand to the device start if it is approaching the end. 6886fa94a07fSbrendan * l2arc_evict() will already have evicted ahead for this case. 6887fa94a07fSbrendan */ 68883a737e0dSbrendan if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 6889fa94a07fSbrendan dev->l2ad_hand = dev->l2ad_start; 6890fa94a07fSbrendan dev->l2ad_first = B_FALSE; 6891fa94a07fSbrendan } 6892fa94a07fSbrendan 68935a98e54bSBrendan Gregg - Sun Microsystems dev->l2ad_writing = B_TRUE; 6894fa94a07fSbrendan (void) zio_wait(pio); 68955a98e54bSBrendan Gregg - Sun Microsystems dev->l2ad_writing = B_FALSE; 68965a98e54bSBrendan Gregg - Sun Microsystems 6897aad02571SSaso Kiselkov return (write_asize); 6898aad02571SSaso Kiselkov } 6899aad02571SSaso Kiselkov 6900fa94a07fSbrendan /* 6901fa94a07fSbrendan * This thread feeds the L2ARC at regular intervals. This is the beating 6902fa94a07fSbrendan * heart of the L2ARC. 6903fa94a07fSbrendan */ 6904fa94a07fSbrendan static void 6905fa94a07fSbrendan l2arc_feed_thread(void) 6906fa94a07fSbrendan { 6907fa94a07fSbrendan callb_cpr_t cpr; 6908fa94a07fSbrendan l2arc_dev_t *dev; 6909fa94a07fSbrendan spa_t *spa; 69105a98e54bSBrendan Gregg - Sun Microsystems uint64_t size, wrote; 6911d3d50737SRafael Vanoni clock_t begin, next = ddi_get_lbolt(); 6912fa94a07fSbrendan 6913fa94a07fSbrendan CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 6914fa94a07fSbrendan 6915fa94a07fSbrendan mutex_enter(&l2arc_feed_thr_lock); 6916fa94a07fSbrendan 6917fa94a07fSbrendan while (l2arc_thread_exit == 0) { 6918fa94a07fSbrendan CALLB_CPR_SAFE_BEGIN(&cpr); 6919fa94a07fSbrendan (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 69205a98e54bSBrendan Gregg - Sun Microsystems next); 6921fa94a07fSbrendan CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 6922d3d50737SRafael Vanoni next = ddi_get_lbolt() + hz; 6923fa94a07fSbrendan 69243a737e0dSbrendan /* 69253a737e0dSbrendan * Quick check for L2ARC devices. 69263a737e0dSbrendan */ 6927c5904d13Seschrock mutex_enter(&l2arc_dev_mtx); 69283a737e0dSbrendan if (l2arc_ndev == 0) { 69293a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 69303a737e0dSbrendan continue; 69313a737e0dSbrendan } 69323a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 6933d3d50737SRafael Vanoni begin = ddi_get_lbolt(); 6934c5904d13Seschrock 6935fa94a07fSbrendan /* 6936c5904d13Seschrock * This selects the next l2arc device to write to, and in 6937c5904d13Seschrock * doing so the next spa to feed from: dev->l2ad_spa. This 69383a737e0dSbrendan * will return NULL if there are now no l2arc devices or if 69393a737e0dSbrendan * they are all faulted. 69403a737e0dSbrendan * 69413a737e0dSbrendan * If a device is returned, its spa's config lock is also 69423a737e0dSbrendan * held to prevent device removal. l2arc_dev_get_next() 69433a737e0dSbrendan * will grab and release l2arc_dev_mtx. 6944fa94a07fSbrendan */ 69453a737e0dSbrendan if ((dev = l2arc_dev_get_next()) == NULL) 6946fa94a07fSbrendan continue; 69473a737e0dSbrendan 69483a737e0dSbrendan spa = dev->l2ad_spa; 6949dcbf3bd6SGeorge Wilson ASSERT3P(spa, !=, NULL); 6950fa94a07fSbrendan 6951f9af39baSGeorge Wilson /* 6952f9af39baSGeorge Wilson * If the pool is read-only then force the feed thread to 6953f9af39baSGeorge Wilson * sleep a little longer. 6954f9af39baSGeorge Wilson */ 6955f9af39baSGeorge Wilson if (!spa_writeable(spa)) { 6956f9af39baSGeorge Wilson next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 6957f9af39baSGeorge Wilson spa_config_exit(spa, SCL_L2ARC, dev); 6958f9af39baSGeorge Wilson continue; 6959f9af39baSGeorge Wilson } 6960f9af39baSGeorge Wilson 6961fa94a07fSbrendan /* 6962fa94a07fSbrendan * Avoid contributing to memory pressure. 6963fa94a07fSbrendan */ 6964fa94a07fSbrendan if (arc_reclaim_needed()) { 6965fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 6966e14bb325SJeff Bonwick spa_config_exit(spa, SCL_L2ARC, dev); 6967fa94a07fSbrendan continue; 6968fa94a07fSbrendan } 6969fa94a07fSbrendan 6970fa94a07fSbrendan ARCSTAT_BUMP(arcstat_l2_feeds); 6971fa94a07fSbrendan 6972aad02571SSaso Kiselkov size = l2arc_write_size(); 69733a737e0dSbrendan 6974fa94a07fSbrendan /* 6975fa94a07fSbrendan * Evict L2ARC buffers that will be overwritten. 6976fa94a07fSbrendan */ 69773a737e0dSbrendan l2arc_evict(dev, size, B_FALSE); 6978fa94a07fSbrendan 6979fa94a07fSbrendan /* 6980fa94a07fSbrendan * Write ARC buffers. 6981fa94a07fSbrendan */ 6982dcbf3bd6SGeorge Wilson wrote = l2arc_write_buffers(spa, dev, size); 69835a98e54bSBrendan Gregg - Sun Microsystems 69845a98e54bSBrendan Gregg - Sun Microsystems /* 69855a98e54bSBrendan Gregg - Sun Microsystems * Calculate interval between writes. 69865a98e54bSBrendan Gregg - Sun Microsystems */ 69875a98e54bSBrendan Gregg - Sun Microsystems next = l2arc_write_interval(begin, size, wrote); 6988e14bb325SJeff Bonwick spa_config_exit(spa, SCL_L2ARC, dev); 6989fa94a07fSbrendan } 6990fa94a07fSbrendan 6991fa94a07fSbrendan l2arc_thread_exit = 0; 6992fa94a07fSbrendan cv_broadcast(&l2arc_feed_thr_cv); 6993fa94a07fSbrendan CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 6994fa94a07fSbrendan thread_exit(); 6995fa94a07fSbrendan } 6996fa94a07fSbrendan 6997c5904d13Seschrock boolean_t 6998c5904d13Seschrock l2arc_vdev_present(vdev_t *vd) 6999c5904d13Seschrock { 7000c5904d13Seschrock l2arc_dev_t *dev; 7001c5904d13Seschrock 7002c5904d13Seschrock mutex_enter(&l2arc_dev_mtx); 7003c5904d13Seschrock for (dev = list_head(l2arc_dev_list); dev != NULL; 7004c5904d13Seschrock dev = list_next(l2arc_dev_list, dev)) { 7005c5904d13Seschrock if (dev->l2ad_vdev == vd) 7006c5904d13Seschrock break; 7007c5904d13Seschrock } 7008c5904d13Seschrock mutex_exit(&l2arc_dev_mtx); 7009c5904d13Seschrock 7010c5904d13Seschrock return (dev != NULL); 7011c5904d13Seschrock } 7012c5904d13Seschrock 7013fa94a07fSbrendan /* 7014fa94a07fSbrendan * Add a vdev for use by the L2ARC. By this point the spa has already 7015fa94a07fSbrendan * validated the vdev and opened it. 7016fa94a07fSbrendan */ 7017fa94a07fSbrendan void 7018573ca77eSGeorge Wilson l2arc_add_vdev(spa_t *spa, vdev_t *vd) 7019fa94a07fSbrendan { 7020fa94a07fSbrendan l2arc_dev_t *adddev; 7021fa94a07fSbrendan 7022c5904d13Seschrock ASSERT(!l2arc_vdev_present(vd)); 7023c5904d13Seschrock 7024fa94a07fSbrendan /* 7025fa94a07fSbrendan * Create a new l2arc device entry. 7026fa94a07fSbrendan */ 7027fa94a07fSbrendan adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 7028fa94a07fSbrendan adddev->l2ad_spa = spa; 7029fa94a07fSbrendan adddev->l2ad_vdev = vd; 7030573ca77eSGeorge Wilson adddev->l2ad_start = VDEV_LABEL_START_SIZE; 7031573ca77eSGeorge Wilson adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 7032fa94a07fSbrendan adddev->l2ad_hand = adddev->l2ad_start; 7033fa94a07fSbrendan adddev->l2ad_first = B_TRUE; 70345a98e54bSBrendan Gregg - Sun Microsystems adddev->l2ad_writing = B_FALSE; 7035fa94a07fSbrendan 703689c86e32SChris Williamson mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 7037fa94a07fSbrendan /* 7038fa94a07fSbrendan * This is a list of all ARC buffers that are still valid on the 7039fa94a07fSbrendan * device. 7040fa94a07fSbrendan */ 704189c86e32SChris Williamson list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 704289c86e32SChris Williamson offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 7043fa94a07fSbrendan 7044b24ab676SJeff Bonwick vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 7045a52fc310SPrakash Surya refcount_create(&adddev->l2ad_alloc); 7046fa94a07fSbrendan 7047fa94a07fSbrendan /* 7048fa94a07fSbrendan * Add device to global list 7049fa94a07fSbrendan */ 7050fa94a07fSbrendan mutex_enter(&l2arc_dev_mtx); 7051fa94a07fSbrendan list_insert_head(l2arc_dev_list, adddev); 7052fa94a07fSbrendan atomic_inc_64(&l2arc_ndev); 7053fa94a07fSbrendan mutex_exit(&l2arc_dev_mtx); 7054fa94a07fSbrendan } 7055fa94a07fSbrendan 7056fa94a07fSbrendan /* 7057fa94a07fSbrendan * Remove a vdev from the L2ARC. 7058fa94a07fSbrendan */ 7059fa94a07fSbrendan void 7060fa94a07fSbrendan l2arc_remove_vdev(vdev_t *vd) 7061fa94a07fSbrendan { 7062fa94a07fSbrendan l2arc_dev_t *dev, *nextdev, *remdev = NULL; 7063fa94a07fSbrendan 7064fa94a07fSbrendan /* 7065fa94a07fSbrendan * Find the device by vdev 7066fa94a07fSbrendan */ 7067fa94a07fSbrendan mutex_enter(&l2arc_dev_mtx); 7068fa94a07fSbrendan for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 7069fa94a07fSbrendan nextdev = list_next(l2arc_dev_list, dev); 7070fa94a07fSbrendan if (vd == dev->l2ad_vdev) { 7071fa94a07fSbrendan remdev = dev; 7072fa94a07fSbrendan break; 7073fa94a07fSbrendan } 7074fa94a07fSbrendan } 7075dcbf3bd6SGeorge Wilson ASSERT3P(remdev, !=, NULL); 7076fa94a07fSbrendan 7077fa94a07fSbrendan /* 7078fa94a07fSbrendan * Remove device from global list 7079fa94a07fSbrendan */ 7080fa94a07fSbrendan list_remove(l2arc_dev_list, remdev); 7081fa94a07fSbrendan l2arc_dev_last = NULL; /* may have been invalidated */ 70823a737e0dSbrendan atomic_dec_64(&l2arc_ndev); 70833a737e0dSbrendan mutex_exit(&l2arc_dev_mtx); 7084fa94a07fSbrendan 7085fa94a07fSbrendan /* 7086fa94a07fSbrendan * Clear all buflists and ARC references. L2ARC device flush. 7087fa94a07fSbrendan */ 7088fa94a07fSbrendan l2arc_evict(remdev, 0, B_TRUE); 708989c86e32SChris Williamson list_destroy(&remdev->l2ad_buflist); 709089c86e32SChris Williamson mutex_destroy(&remdev->l2ad_mtx); 7091a52fc310SPrakash Surya refcount_destroy(&remdev->l2ad_alloc); 7092fa94a07fSbrendan kmem_free(remdev, sizeof (l2arc_dev_t)); 7093fa94a07fSbrendan } 7094fa94a07fSbrendan 7095fa94a07fSbrendan void 7096e14bb325SJeff Bonwick l2arc_init(void) 7097fa94a07fSbrendan { 7098fa94a07fSbrendan l2arc_thread_exit = 0; 7099fa94a07fSbrendan l2arc_ndev = 0; 7100fa94a07fSbrendan l2arc_writes_sent = 0; 7101fa94a07fSbrendan l2arc_writes_done = 0; 7102fa94a07fSbrendan 7103fa94a07fSbrendan mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7104fa94a07fSbrendan cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7105fa94a07fSbrendan mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7106fa94a07fSbrendan mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7107fa94a07fSbrendan 7108fa94a07fSbrendan l2arc_dev_list = &L2ARC_dev_list; 7109fa94a07fSbrendan l2arc_free_on_write = &L2ARC_free_on_write; 7110fa94a07fSbrendan list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7111fa94a07fSbrendan offsetof(l2arc_dev_t, l2ad_node)); 7112fa94a07fSbrendan list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7113fa94a07fSbrendan offsetof(l2arc_data_free_t, l2df_list_node)); 7114fa94a07fSbrendan } 7115fa94a07fSbrendan 7116fa94a07fSbrendan void 7117e14bb325SJeff Bonwick l2arc_fini(void) 7118fa94a07fSbrendan { 71193a737e0dSbrendan /* 71203a737e0dSbrendan * This is called from dmu_fini(), which is called from spa_fini(); 71213a737e0dSbrendan * Because of this, we can assume that all l2arc devices have 71223a737e0dSbrendan * already been removed when the pools themselves were removed. 71233a737e0dSbrendan */ 71243a737e0dSbrendan 71253a737e0dSbrendan l2arc_do_free_on_write(); 71263a737e0dSbrendan 7127fa94a07fSbrendan mutex_destroy(&l2arc_feed_thr_lock); 7128fa94a07fSbrendan cv_destroy(&l2arc_feed_thr_cv); 7129fa94a07fSbrendan mutex_destroy(&l2arc_dev_mtx); 7130fa94a07fSbrendan mutex_destroy(&l2arc_free_on_write_mtx); 7131fa94a07fSbrendan 7132fa94a07fSbrendan list_destroy(l2arc_dev_list); 7133fa94a07fSbrendan list_destroy(l2arc_free_on_write); 7134fa94a07fSbrendan } 7135e14bb325SJeff Bonwick 7136e14bb325SJeff Bonwick void 7137e14bb325SJeff Bonwick l2arc_start(void) 7138e14bb325SJeff Bonwick { 71398ad4d6ddSJeff Bonwick if (!(spa_mode_global & FWRITE)) 7140e14bb325SJeff Bonwick return; 7141e14bb325SJeff Bonwick 7142e14bb325SJeff Bonwick (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7143e14bb325SJeff Bonwick TS_RUN, minclsyspri); 7144e14bb325SJeff Bonwick } 7145e14bb325SJeff Bonwick 7146e14bb325SJeff Bonwick void 7147e14bb325SJeff Bonwick l2arc_stop(void) 7148e14bb325SJeff Bonwick { 71498ad4d6ddSJeff Bonwick if (!(spa_mode_global & FWRITE)) 7150e14bb325SJeff Bonwick return; 7151e14bb325SJeff Bonwick 7152e14bb325SJeff Bonwick mutex_enter(&l2arc_feed_thr_lock); 7153e14bb325SJeff Bonwick cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7154e14bb325SJeff Bonwick l2arc_thread_exit = 1; 7155e14bb325SJeff Bonwick while (l2arc_thread_exit != 0) 7156e14bb325SJeff Bonwick cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7157e14bb325SJeff Bonwick mutex_exit(&l2arc_feed_thr_lock); 7158e14bb325SJeff Bonwick } 7159