1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2012, Joyent, Inc. All rights reserved.
  24  * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
  25  * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
  26  * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
  27  */
  28 
  29 /*
  30  * DVA-based Adjustable Replacement Cache
  31  *
  32  * While much of the theory of operation used here is
  33  * based on the self-tuning, low overhead replacement cache
  34  * presented by Megiddo and Modha at FAST 2003, there are some
  35  * significant differences:
  36  *
  37  * 1. The Megiddo and Modha model assumes any page is evictable.
  38  * Pages in its cache cannot be "locked" into memory.  This makes
  39  * the eviction algorithm simple: evict the last page in the list.
  40  * This also make the performance characteristics easy to reason
  41  * about.  Our cache is not so simple.  At any given moment, some
  42  * subset of the blocks in the cache are un-evictable because we
  43  * have handed out a reference to them.  Blocks are only evictable
  44  * when there are no external references active.  This makes
  45  * eviction far more problematic:  we choose to evict the evictable
  46  * blocks that are the "lowest" in the list.
  47  *
  48  * There are times when it is not possible to evict the requested
  49  * space.  In these circumstances we are unable to adjust the cache
  50  * size.  To prevent the cache growing unbounded at these times we
  51  * implement a "cache throttle" that slows the flow of new data
  52  * into the cache until we can make space available.
  53  *
  54  * 2. The Megiddo and Modha model assumes a fixed cache size.
  55  * Pages are evicted when the cache is full and there is a cache
  56  * miss.  Our model has a variable sized cache.  It grows with
  57  * high use, but also tries to react to memory pressure from the
  58  * operating system: decreasing its size when system memory is
  59  * tight.
  60  *
  61  * 3. The Megiddo and Modha model assumes a fixed page size. All
  62  * elements of the cache are therefore exactly the same size.  So
  63  * when adjusting the cache size following a cache miss, its simply
  64  * a matter of choosing a single page to evict.  In our model, we
  65  * have variable sized cache blocks (rangeing from 512 bytes to
  66  * 128K bytes).  We therefore choose a set of blocks to evict to make
  67  * space for a cache miss that approximates as closely as possible
  68  * the space used by the new block.
  69  *
  70  * See also:  "ARC: A Self-Tuning, Low Overhead Replacement Cache"
  71  * by N. Megiddo & D. Modha, FAST 2003
  72  */
  73 
  74 /*
  75  * The locking model:
  76  *
  77  * A new reference to a cache buffer can be obtained in two
  78  * ways: 1) via a hash table lookup using the DVA as a key,
  79  * or 2) via one of the ARC lists.  The arc_read() interface
  80  * uses method 1, while the internal arc algorithms for
  81  * adjusting the cache use method 2.  We therefore provide two
  82  * types of locks: 1) the hash table lock array, and 2) the
  83  * arc list locks.
  84  *
  85  * Buffers do not have their own mutexes, rather they rely on the
  86  * hash table mutexes for the bulk of their protection (i.e. most
  87  * fields in the arc_buf_hdr_t are protected by these mutexes).
  88  *
  89  * buf_hash_find() returns the appropriate mutex (held) when it
  90  * locates the requested buffer in the hash table.  It returns
  91  * NULL for the mutex if the buffer was not in the table.
  92  *
  93  * buf_hash_remove() expects the appropriate hash mutex to be
  94  * already held before it is invoked.
  95  *
  96  * Each arc state also has a mutex which is used to protect the
  97  * buffer list associated with the state.  When attempting to
  98  * obtain a hash table lock while holding an arc list lock you
  99  * must use: mutex_tryenter() to avoid deadlock.  Also note that
 100  * the active state mutex must be held before the ghost state mutex.
 101  *
 102  * Arc buffers may have an associated eviction callback function.
 103  * This function will be invoked prior to removing the buffer (e.g.
 104  * in arc_do_user_evicts()).  Note however that the data associated
 105  * with the buffer may be evicted prior to the callback.  The callback
 106  * must be made with *no locks held* (to prevent deadlock).  Additionally,
 107  * the users of callbacks must ensure that their private data is
 108  * protected from simultaneous callbacks from arc_clear_callback()
 109  * and arc_do_user_evicts().
 110  *
 111  * Note that the majority of the performance stats are manipulated
 112  * with atomic operations.
 113  *
 114  * The L2ARC uses the l2ad_mtx on each vdev for the following:
 115  *
 116  *      - L2ARC buflist creation
 117  *      - L2ARC buflist eviction
 118  *      - L2ARC write completion, which walks L2ARC buflists
 119  *      - ARC header destruction, as it removes from L2ARC buflists
 120  *      - ARC header release, as it removes from L2ARC buflists
 121  */
 122 
 123 #include <sys/spa.h>
 124 #include <sys/zio.h>
 125 #include <sys/zio_compress.h>
 126 #include <sys/zfs_context.h>
 127 #include <sys/arc.h>
 128 #include <sys/refcount.h>
 129 #include <sys/vdev.h>
 130 #include <sys/vdev_impl.h>
 131 #include <sys/dsl_pool.h>
 132 #include <sys/multilist.h>
 133 #ifdef _KERNEL
 134 #include <sys/vmsystm.h>
 135 #include <vm/anon.h>
 136 #include <sys/fs/swapnode.h>
 137 #include <sys/dnlc.h>
 138 #endif
 139 #include <sys/callb.h>
 140 #include <sys/kstat.h>
 141 #include <zfs_fletcher.h>
 142 
 143 #ifndef _KERNEL
 144 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
 145 boolean_t arc_watch = B_FALSE;
 146 int arc_procfd;
 147 #endif
 148 
 149 static kmutex_t         arc_reclaim_lock;
 150 static kcondvar_t       arc_reclaim_thread_cv;
 151 static boolean_t        arc_reclaim_thread_exit;
 152 static kcondvar_t       arc_reclaim_waiters_cv;
 153 
 154 static kmutex_t         arc_user_evicts_lock;
 155 static kcondvar_t       arc_user_evicts_cv;
 156 static boolean_t        arc_user_evicts_thread_exit;
 157 
 158 uint_t arc_reduce_dnlc_percent = 3;
 159 
 160 /*
 161  * The number of headers to evict in arc_evict_state_impl() before
 162  * dropping the sublist lock and evicting from another sublist. A lower
 163  * value means we're more likely to evict the "correct" header (i.e. the
 164  * oldest header in the arc state), but comes with higher overhead
 165  * (i.e. more invocations of arc_evict_state_impl()).
 166  */
 167 int zfs_arc_evict_batch_limit = 10;
 168 
 169 /*
 170  * The number of sublists used for each of the arc state lists. If this
 171  * is not set to a suitable value by the user, it will be configured to
 172  * the number of CPUs on the system in arc_init().
 173  */
 174 int zfs_arc_num_sublists_per_state = 0;
 175 
 176 /* number of seconds before growing cache again */
 177 static int              arc_grow_retry = 60;
 178 
 179 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */
 180 int             zfs_arc_overflow_shift = 8;
 181 
 182 /* shift of arc_c for calculating both min and max arc_p */
 183 static int              arc_p_min_shift = 4;
 184 
 185 /* log2(fraction of arc to reclaim) */
 186 static int              arc_shrink_shift = 7;
 187 
 188 /*
 189  * log2(fraction of ARC which must be free to allow growing).
 190  * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
 191  * when reading a new block into the ARC, we will evict an equal-sized block
 192  * from the ARC.
 193  *
 194  * This must be less than arc_shrink_shift, so that when we shrink the ARC,
 195  * we will still not allow it to grow.
 196  */
 197 int                     arc_no_grow_shift = 5;
 198 
 199 
 200 /*
 201  * minimum lifespan of a prefetch block in clock ticks
 202  * (initialized in arc_init())
 203  */
 204 static int              arc_min_prefetch_lifespan;
 205 
 206 /*
 207  * If this percent of memory is free, don't throttle.
 208  */
 209 int arc_lotsfree_percent = 10;
 210 
 211 static int arc_dead;
 212 
 213 /*
 214  * The arc has filled available memory and has now warmed up.
 215  */
 216 static boolean_t arc_warm;
 217 
 218 /*
 219  * These tunables are for performance analysis.
 220  */
 221 uint64_t zfs_arc_max;
 222 uint64_t zfs_arc_min;
 223 uint64_t zfs_arc_meta_limit = 0;
 224 uint64_t zfs_arc_meta_min = 0;
 225 int zfs_arc_grow_retry = 0;
 226 int zfs_arc_shrink_shift = 0;
 227 int zfs_arc_p_min_shift = 0;
 228 int zfs_disable_dup_eviction = 0;
 229 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
 230 
 231 /*
 232  * Note that buffers can be in one of 6 states:
 233  *      ARC_anon        - anonymous (discussed below)
 234  *      ARC_mru         - recently used, currently cached
 235  *      ARC_mru_ghost   - recentely used, no longer in cache
 236  *      ARC_mfu         - frequently used, currently cached
 237  *      ARC_mfu_ghost   - frequently used, no longer in cache
 238  *      ARC_l2c_only    - exists in L2ARC but not other states
 239  * When there are no active references to the buffer, they are
 240  * are linked onto a list in one of these arc states.  These are
 241  * the only buffers that can be evicted or deleted.  Within each
 242  * state there are multiple lists, one for meta-data and one for
 243  * non-meta-data.  Meta-data (indirect blocks, blocks of dnodes,
 244  * etc.) is tracked separately so that it can be managed more
 245  * explicitly: favored over data, limited explicitly.
 246  *
 247  * Anonymous buffers are buffers that are not associated with
 248  * a DVA.  These are buffers that hold dirty block copies
 249  * before they are written to stable storage.  By definition,
 250  * they are "ref'd" and are considered part of arc_mru
 251  * that cannot be freed.  Generally, they will aquire a DVA
 252  * as they are written and migrate onto the arc_mru list.
 253  *
 254  * The ARC_l2c_only state is for buffers that are in the second
 255  * level ARC but no longer in any of the ARC_m* lists.  The second
 256  * level ARC itself may also contain buffers that are in any of
 257  * the ARC_m* states - meaning that a buffer can exist in two
 258  * places.  The reason for the ARC_l2c_only state is to keep the
 259  * buffer header in the hash table, so that reads that hit the
 260  * second level ARC benefit from these fast lookups.
 261  */
 262 
 263 typedef struct arc_state {
 264         /*
 265          * list of evictable buffers
 266          */
 267         multilist_t arcs_list[ARC_BUFC_NUMTYPES];
 268         /*
 269          * total amount of evictable data in this state
 270          */
 271         uint64_t arcs_lsize[ARC_BUFC_NUMTYPES];
 272         /*
 273          * total amount of data in this state; this includes: evictable,
 274          * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
 275          */
 276         refcount_t arcs_size;
 277 } arc_state_t;
 278 
 279 /* The 6 states: */
 280 static arc_state_t ARC_anon;
 281 static arc_state_t ARC_mru;
 282 static arc_state_t ARC_mru_ghost;
 283 static arc_state_t ARC_mfu;
 284 static arc_state_t ARC_mfu_ghost;
 285 static arc_state_t ARC_l2c_only;
 286 
 287 typedef struct arc_stats {
 288         kstat_named_t arcstat_hits;
 289         kstat_named_t arcstat_misses;
 290         kstat_named_t arcstat_demand_data_hits;
 291         kstat_named_t arcstat_demand_data_misses;
 292         kstat_named_t arcstat_demand_metadata_hits;
 293         kstat_named_t arcstat_demand_metadata_misses;
 294         kstat_named_t arcstat_prefetch_data_hits;
 295         kstat_named_t arcstat_prefetch_data_misses;
 296         kstat_named_t arcstat_prefetch_metadata_hits;
 297         kstat_named_t arcstat_prefetch_metadata_misses;
 298         kstat_named_t arcstat_mru_hits;
 299         kstat_named_t arcstat_mru_ghost_hits;
 300         kstat_named_t arcstat_mfu_hits;
 301         kstat_named_t arcstat_mfu_ghost_hits;
 302         kstat_named_t arcstat_deleted;
 303         /*
 304          * Number of buffers that could not be evicted because the hash lock
 305          * was held by another thread.  The lock may not necessarily be held
 306          * by something using the same buffer, since hash locks are shared
 307          * by multiple buffers.
 308          */
 309         kstat_named_t arcstat_mutex_miss;
 310         /*
 311          * Number of buffers skipped because they have I/O in progress, are
 312          * indrect prefetch buffers that have not lived long enough, or are
 313          * not from the spa we're trying to evict from.
 314          */
 315         kstat_named_t arcstat_evict_skip;
 316         /*
 317          * Number of times arc_evict_state() was unable to evict enough
 318          * buffers to reach it's target amount.
 319          */
 320         kstat_named_t arcstat_evict_not_enough;
 321         kstat_named_t arcstat_evict_l2_cached;
 322         kstat_named_t arcstat_evict_l2_eligible;
 323         kstat_named_t arcstat_evict_l2_ineligible;
 324         kstat_named_t arcstat_evict_l2_skip;
 325         kstat_named_t arcstat_hash_elements;
 326         kstat_named_t arcstat_hash_elements_max;
 327         kstat_named_t arcstat_hash_collisions;
 328         kstat_named_t arcstat_hash_chains;
 329         kstat_named_t arcstat_hash_chain_max;
 330         kstat_named_t arcstat_p;
 331         kstat_named_t arcstat_c;
 332         kstat_named_t arcstat_c_min;
 333         kstat_named_t arcstat_c_max;
 334         kstat_named_t arcstat_size;
 335         /*
 336          * Number of bytes consumed by internal ARC structures necessary
 337          * for tracking purposes; these structures are not actually
 338          * backed by ARC buffers. This includes arc_buf_hdr_t structures
 339          * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
 340          * caches), and arc_buf_t structures (allocated via arc_buf_t
 341          * cache).
 342          */
 343         kstat_named_t arcstat_hdr_size;
 344         /*
 345          * Number of bytes consumed by ARC buffers of type equal to
 346          * ARC_BUFC_DATA. This is generally consumed by buffers backing
 347          * on disk user data (e.g. plain file contents).
 348          */
 349         kstat_named_t arcstat_data_size;
 350         /*
 351          * Number of bytes consumed by ARC buffers of type equal to
 352          * ARC_BUFC_METADATA. This is generally consumed by buffers
 353          * backing on disk data that is used for internal ZFS
 354          * structures (e.g. ZAP, dnode, indirect blocks, etc).
 355          */
 356         kstat_named_t arcstat_metadata_size;
 357         /*
 358          * Number of bytes consumed by various buffers and structures
 359          * not actually backed with ARC buffers. This includes bonus
 360          * buffers (allocated directly via zio_buf_* functions),
 361          * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
 362          * cache), and dnode_t structures (allocated via dnode_t cache).
 363          */
 364         kstat_named_t arcstat_other_size;
 365         /*
 366          * Total number of bytes consumed by ARC buffers residing in the
 367          * arc_anon state. This includes *all* buffers in the arc_anon
 368          * state; e.g. data, metadata, evictable, and unevictable buffers
 369          * are all included in this value.
 370          */
 371         kstat_named_t arcstat_anon_size;
 372         /*
 373          * Number of bytes consumed by ARC buffers that meet the
 374          * following criteria: backing buffers of type ARC_BUFC_DATA,
 375          * residing in the arc_anon state, and are eligible for eviction
 376          * (e.g. have no outstanding holds on the buffer).
 377          */
 378         kstat_named_t arcstat_anon_evictable_data;
 379         /*
 380          * Number of bytes consumed by ARC buffers that meet the
 381          * following criteria: backing buffers of type ARC_BUFC_METADATA,
 382          * residing in the arc_anon state, and are eligible for eviction
 383          * (e.g. have no outstanding holds on the buffer).
 384          */
 385         kstat_named_t arcstat_anon_evictable_metadata;
 386         /*
 387          * Total number of bytes consumed by ARC buffers residing in the
 388          * arc_mru state. This includes *all* buffers in the arc_mru
 389          * state; e.g. data, metadata, evictable, and unevictable buffers
 390          * are all included in this value.
 391          */
 392         kstat_named_t arcstat_mru_size;
 393         /*
 394          * Number of bytes consumed by ARC buffers that meet the
 395          * following criteria: backing buffers of type ARC_BUFC_DATA,
 396          * residing in the arc_mru state, and are eligible for eviction
 397          * (e.g. have no outstanding holds on the buffer).
 398          */
 399         kstat_named_t arcstat_mru_evictable_data;
 400         /*
 401          * Number of bytes consumed by ARC buffers that meet the
 402          * following criteria: backing buffers of type ARC_BUFC_METADATA,
 403          * residing in the arc_mru state, and are eligible for eviction
 404          * (e.g. have no outstanding holds on the buffer).
 405          */
 406         kstat_named_t arcstat_mru_evictable_metadata;
 407         /*
 408          * Total number of bytes that *would have been* consumed by ARC
 409          * buffers in the arc_mru_ghost state. The key thing to note
 410          * here, is the fact that this size doesn't actually indicate
 411          * RAM consumption. The ghost lists only consist of headers and
 412          * don't actually have ARC buffers linked off of these headers.
 413          * Thus, *if* the headers had associated ARC buffers, these
 414          * buffers *would have* consumed this number of bytes.
 415          */
 416         kstat_named_t arcstat_mru_ghost_size;
 417         /*
 418          * Number of bytes that *would have been* consumed by ARC
 419          * buffers that are eligible for eviction, of type
 420          * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
 421          */
 422         kstat_named_t arcstat_mru_ghost_evictable_data;
 423         /*
 424          * Number of bytes that *would have been* consumed by ARC
 425          * buffers that are eligible for eviction, of type
 426          * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
 427          */
 428         kstat_named_t arcstat_mru_ghost_evictable_metadata;
 429         /*
 430          * Total number of bytes consumed by ARC buffers residing in the
 431          * arc_mfu state. This includes *all* buffers in the arc_mfu
 432          * state; e.g. data, metadata, evictable, and unevictable buffers
 433          * are all included in this value.
 434          */
 435         kstat_named_t arcstat_mfu_size;
 436         /*
 437          * Number of bytes consumed by ARC buffers that are eligible for
 438          * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
 439          * state.
 440          */
 441         kstat_named_t arcstat_mfu_evictable_data;
 442         /*
 443          * Number of bytes consumed by ARC buffers that are eligible for
 444          * eviction, of type ARC_BUFC_METADATA, and reside in the
 445          * arc_mfu state.
 446          */
 447         kstat_named_t arcstat_mfu_evictable_metadata;
 448         /*
 449          * Total number of bytes that *would have been* consumed by ARC
 450          * buffers in the arc_mfu_ghost state. See the comment above
 451          * arcstat_mru_ghost_size for more details.
 452          */
 453         kstat_named_t arcstat_mfu_ghost_size;
 454         /*
 455          * Number of bytes that *would have been* consumed by ARC
 456          * buffers that are eligible for eviction, of type
 457          * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
 458          */
 459         kstat_named_t arcstat_mfu_ghost_evictable_data;
 460         /*
 461          * Number of bytes that *would have been* consumed by ARC
 462          * buffers that are eligible for eviction, of type
 463          * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
 464          */
 465         kstat_named_t arcstat_mfu_ghost_evictable_metadata;
 466         kstat_named_t arcstat_l2_hits;
 467         kstat_named_t arcstat_l2_misses;
 468         kstat_named_t arcstat_l2_feeds;
 469         kstat_named_t arcstat_l2_rw_clash;
 470         kstat_named_t arcstat_l2_read_bytes;
 471         kstat_named_t arcstat_l2_write_bytes;
 472         kstat_named_t arcstat_l2_writes_sent;
 473         kstat_named_t arcstat_l2_writes_done;
 474         kstat_named_t arcstat_l2_writes_error;
 475         kstat_named_t arcstat_l2_writes_lock_retry;
 476         kstat_named_t arcstat_l2_evict_lock_retry;
 477         kstat_named_t arcstat_l2_evict_reading;
 478         kstat_named_t arcstat_l2_evict_l1cached;
 479         kstat_named_t arcstat_l2_free_on_write;
 480         kstat_named_t arcstat_l2_cdata_free_on_write;
 481         kstat_named_t arcstat_l2_abort_lowmem;
 482         kstat_named_t arcstat_l2_cksum_bad;
 483         kstat_named_t arcstat_l2_io_error;
 484         kstat_named_t arcstat_l2_size;
 485         kstat_named_t arcstat_l2_asize;
 486         kstat_named_t arcstat_l2_hdr_size;
 487         kstat_named_t arcstat_l2_compress_successes;
 488         kstat_named_t arcstat_l2_compress_zeros;
 489         kstat_named_t arcstat_l2_compress_failures;
 490         kstat_named_t arcstat_memory_throttle_count;
 491         kstat_named_t arcstat_duplicate_buffers;
 492         kstat_named_t arcstat_duplicate_buffers_size;
 493         kstat_named_t arcstat_duplicate_reads;
 494         kstat_named_t arcstat_meta_used;
 495         kstat_named_t arcstat_meta_limit;
 496         kstat_named_t arcstat_meta_max;
 497         kstat_named_t arcstat_meta_min;
 498 } arc_stats_t;
 499 
 500 static arc_stats_t arc_stats = {
 501         { "hits",                       KSTAT_DATA_UINT64 },
 502         { "misses",                     KSTAT_DATA_UINT64 },
 503         { "demand_data_hits",           KSTAT_DATA_UINT64 },
 504         { "demand_data_misses",         KSTAT_DATA_UINT64 },
 505         { "demand_metadata_hits",       KSTAT_DATA_UINT64 },
 506         { "demand_metadata_misses",     KSTAT_DATA_UINT64 },
 507         { "prefetch_data_hits",         KSTAT_DATA_UINT64 },
 508         { "prefetch_data_misses",       KSTAT_DATA_UINT64 },
 509         { "prefetch_metadata_hits",     KSTAT_DATA_UINT64 },
 510         { "prefetch_metadata_misses",   KSTAT_DATA_UINT64 },
 511         { "mru_hits",                   KSTAT_DATA_UINT64 },
 512         { "mru_ghost_hits",             KSTAT_DATA_UINT64 },
 513         { "mfu_hits",                   KSTAT_DATA_UINT64 },
 514         { "mfu_ghost_hits",             KSTAT_DATA_UINT64 },
 515         { "deleted",                    KSTAT_DATA_UINT64 },
 516         { "mutex_miss",                 KSTAT_DATA_UINT64 },
 517         { "evict_skip",                 KSTAT_DATA_UINT64 },
 518         { "evict_not_enough",           KSTAT_DATA_UINT64 },
 519         { "evict_l2_cached",            KSTAT_DATA_UINT64 },
 520         { "evict_l2_eligible",          KSTAT_DATA_UINT64 },
 521         { "evict_l2_ineligible",        KSTAT_DATA_UINT64 },
 522         { "evict_l2_skip",              KSTAT_DATA_UINT64 },
 523         { "hash_elements",              KSTAT_DATA_UINT64 },
 524         { "hash_elements_max",          KSTAT_DATA_UINT64 },
 525         { "hash_collisions",            KSTAT_DATA_UINT64 },
 526         { "hash_chains",                KSTAT_DATA_UINT64 },
 527         { "hash_chain_max",             KSTAT_DATA_UINT64 },
 528         { "p",                          KSTAT_DATA_UINT64 },
 529         { "c",                          KSTAT_DATA_UINT64 },
 530         { "c_min",                      KSTAT_DATA_UINT64 },
 531         { "c_max",                      KSTAT_DATA_UINT64 },
 532         { "size",                       KSTAT_DATA_UINT64 },
 533         { "hdr_size",                   KSTAT_DATA_UINT64 },
 534         { "data_size",                  KSTAT_DATA_UINT64 },
 535         { "metadata_size",              KSTAT_DATA_UINT64 },
 536         { "other_size",                 KSTAT_DATA_UINT64 },
 537         { "anon_size",                  KSTAT_DATA_UINT64 },
 538         { "anon_evictable_data",        KSTAT_DATA_UINT64 },
 539         { "anon_evictable_metadata",    KSTAT_DATA_UINT64 },
 540         { "mru_size",                   KSTAT_DATA_UINT64 },
 541         { "mru_evictable_data",         KSTAT_DATA_UINT64 },
 542         { "mru_evictable_metadata",     KSTAT_DATA_UINT64 },
 543         { "mru_ghost_size",             KSTAT_DATA_UINT64 },
 544         { "mru_ghost_evictable_data",   KSTAT_DATA_UINT64 },
 545         { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
 546         { "mfu_size",                   KSTAT_DATA_UINT64 },
 547         { "mfu_evictable_data",         KSTAT_DATA_UINT64 },
 548         { "mfu_evictable_metadata",     KSTAT_DATA_UINT64 },
 549         { "mfu_ghost_size",             KSTAT_DATA_UINT64 },
 550         { "mfu_ghost_evictable_data",   KSTAT_DATA_UINT64 },
 551         { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
 552         { "l2_hits",                    KSTAT_DATA_UINT64 },
 553         { "l2_misses",                  KSTAT_DATA_UINT64 },
 554         { "l2_feeds",                   KSTAT_DATA_UINT64 },
 555         { "l2_rw_clash",                KSTAT_DATA_UINT64 },
 556         { "l2_read_bytes",              KSTAT_DATA_UINT64 },
 557         { "l2_write_bytes",             KSTAT_DATA_UINT64 },
 558         { "l2_writes_sent",             KSTAT_DATA_UINT64 },
 559         { "l2_writes_done",             KSTAT_DATA_UINT64 },
 560         { "l2_writes_error",            KSTAT_DATA_UINT64 },
 561         { "l2_writes_lock_retry",       KSTAT_DATA_UINT64 },
 562         { "l2_evict_lock_retry",        KSTAT_DATA_UINT64 },
 563         { "l2_evict_reading",           KSTAT_DATA_UINT64 },
 564         { "l2_evict_l1cached",          KSTAT_DATA_UINT64 },
 565         { "l2_free_on_write",           KSTAT_DATA_UINT64 },
 566         { "l2_cdata_free_on_write",     KSTAT_DATA_UINT64 },
 567         { "l2_abort_lowmem",            KSTAT_DATA_UINT64 },
 568         { "l2_cksum_bad",               KSTAT_DATA_UINT64 },
 569         { "l2_io_error",                KSTAT_DATA_UINT64 },
 570         { "l2_size",                    KSTAT_DATA_UINT64 },
 571         { "l2_asize",                   KSTAT_DATA_UINT64 },
 572         { "l2_hdr_size",                KSTAT_DATA_UINT64 },
 573         { "l2_compress_successes",      KSTAT_DATA_UINT64 },
 574         { "l2_compress_zeros",          KSTAT_DATA_UINT64 },
 575         { "l2_compress_failures",       KSTAT_DATA_UINT64 },
 576         { "memory_throttle_count",      KSTAT_DATA_UINT64 },
 577         { "duplicate_buffers",          KSTAT_DATA_UINT64 },
 578         { "duplicate_buffers_size",     KSTAT_DATA_UINT64 },
 579         { "duplicate_reads",            KSTAT_DATA_UINT64 },
 580         { "arc_meta_used",              KSTAT_DATA_UINT64 },
 581         { "arc_meta_limit",             KSTAT_DATA_UINT64 },
 582         { "arc_meta_max",               KSTAT_DATA_UINT64 },
 583         { "arc_meta_min",               KSTAT_DATA_UINT64 }
 584 };
 585 
 586 #define ARCSTAT(stat)   (arc_stats.stat.value.ui64)
 587 
 588 #define ARCSTAT_INCR(stat, val) \
 589         atomic_add_64(&arc_stats.stat.value.ui64, (val))
 590 
 591 #define ARCSTAT_BUMP(stat)      ARCSTAT_INCR(stat, 1)
 592 #define ARCSTAT_BUMPDOWN(stat)  ARCSTAT_INCR(stat, -1)
 593 
 594 #define ARCSTAT_MAX(stat, val) {                                        \
 595         uint64_t m;                                                     \
 596         while ((val) > (m = arc_stats.stat.value.ui64) &&            \
 597             (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val))))     \
 598                 continue;                                               \
 599 }
 600 
 601 #define ARCSTAT_MAXSTAT(stat) \
 602         ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
 603 
 604 /*
 605  * We define a macro to allow ARC hits/misses to be easily broken down by
 606  * two separate conditions, giving a total of four different subtypes for
 607  * each of hits and misses (so eight statistics total).
 608  */
 609 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
 610         if (cond1) {                                                    \
 611                 if (cond2) {                                            \
 612                         ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
 613                 } else {                                                \
 614                         ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
 615                 }                                                       \
 616         } else {                                                        \
 617                 if (cond2) {                                            \
 618                         ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
 619                 } else {                                                \
 620                         ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
 621                 }                                                       \
 622         }
 623 
 624 kstat_t                 *arc_ksp;
 625 static arc_state_t      *arc_anon;
 626 static arc_state_t      *arc_mru;
 627 static arc_state_t      *arc_mru_ghost;
 628 static arc_state_t      *arc_mfu;
 629 static arc_state_t      *arc_mfu_ghost;
 630 static arc_state_t      *arc_l2c_only;
 631 
 632 /*
 633  * There are several ARC variables that are critical to export as kstats --
 634  * but we don't want to have to grovel around in the kstat whenever we wish to
 635  * manipulate them.  For these variables, we therefore define them to be in
 636  * terms of the statistic variable.  This assures that we are not introducing
 637  * the possibility of inconsistency by having shadow copies of the variables,
 638  * while still allowing the code to be readable.
 639  */
 640 #define arc_size        ARCSTAT(arcstat_size)   /* actual total arc size */
 641 #define arc_p           ARCSTAT(arcstat_p)      /* target size of MRU */
 642 #define arc_c           ARCSTAT(arcstat_c)      /* target size of cache */
 643 #define arc_c_min       ARCSTAT(arcstat_c_min)  /* min target cache size */
 644 #define arc_c_max       ARCSTAT(arcstat_c_max)  /* max target cache size */
 645 #define arc_meta_limit  ARCSTAT(arcstat_meta_limit) /* max size for metadata */
 646 #define arc_meta_min    ARCSTAT(arcstat_meta_min) /* min size for metadata */
 647 #define arc_meta_used   ARCSTAT(arcstat_meta_used) /* size of metadata */
 648 #define arc_meta_max    ARCSTAT(arcstat_meta_max) /* max size of metadata */
 649 
 650 #define L2ARC_IS_VALID_COMPRESS(_c_) \
 651         ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
 652 
 653 static int              arc_no_grow;    /* Don't try to grow cache size */
 654 static uint64_t         arc_tempreserve;
 655 static uint64_t         arc_loaned_bytes;
 656 
 657 typedef struct arc_callback arc_callback_t;
 658 
 659 struct arc_callback {
 660         void                    *acb_private;
 661         arc_done_func_t         *acb_done;
 662         arc_buf_t               *acb_buf;
 663         zio_t                   *acb_zio_dummy;
 664         arc_callback_t          *acb_next;
 665 };
 666 
 667 typedef struct arc_write_callback arc_write_callback_t;
 668 
 669 struct arc_write_callback {
 670         void            *awcb_private;
 671         arc_done_func_t *awcb_ready;
 672         arc_done_func_t *awcb_physdone;
 673         arc_done_func_t *awcb_done;
 674         arc_buf_t       *awcb_buf;
 675 };
 676 
 677 /*
 678  * ARC buffers are separated into multiple structs as a memory saving measure:
 679  *   - Common fields struct, always defined, and embedded within it:
 680  *       - L2-only fields, always allocated but undefined when not in L2ARC
 681  *       - L1-only fields, only allocated when in L1ARC
 682  *
 683  *           Buffer in L1                     Buffer only in L2
 684  *    +------------------------+          +------------------------+
 685  *    | arc_buf_hdr_t          |          | arc_buf_hdr_t          |
 686  *    |                        |          |                        |
 687  *    |                        |          |                        |
 688  *    |                        |          |                        |
 689  *    +------------------------+          +------------------------+
 690  *    | l2arc_buf_hdr_t        |          | l2arc_buf_hdr_t        |
 691  *    | (undefined if L1-only) |          |                        |
 692  *    +------------------------+          +------------------------+
 693  *    | l1arc_buf_hdr_t        |
 694  *    |                        |
 695  *    |                        |
 696  *    |                        |
 697  *    |                        |
 698  *    +------------------------+
 699  *
 700  * Because it's possible for the L2ARC to become extremely large, we can wind
 701  * up eating a lot of memory in L2ARC buffer headers, so the size of a header
 702  * is minimized by only allocating the fields necessary for an L1-cached buffer
 703  * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
 704  * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
 705  * words in pointers. arc_hdr_realloc() is used to switch a header between
 706  * these two allocation states.
 707  */
 708 typedef struct l1arc_buf_hdr {
 709         kmutex_t                b_freeze_lock;
 710 #ifdef ZFS_DEBUG
 711         /*
 712          * used for debugging wtih kmem_flags - by allocating and freeing
 713          * b_thawed when the buffer is thawed, we get a record of the stack
 714          * trace that thawed it.
 715          */
 716         void                    *b_thawed;
 717 #endif
 718 
 719         arc_buf_t               *b_buf;
 720         uint32_t                b_datacnt;
 721         /* for waiting on writes to complete */
 722         kcondvar_t              b_cv;
 723 
 724         /* protected by arc state mutex */
 725         arc_state_t             *b_state;
 726         multilist_node_t        b_arc_node;
 727 
 728         /* updated atomically */
 729         clock_t                 b_arc_access;
 730 
 731         /* self protecting */
 732         refcount_t              b_refcnt;
 733 
 734         arc_callback_t          *b_acb;
 735         /* temporary buffer holder for in-flight compressed data */
 736         void                    *b_tmp_cdata;
 737 } l1arc_buf_hdr_t;
 738 
 739 typedef struct l2arc_dev l2arc_dev_t;
 740 
 741 typedef struct l2arc_buf_hdr {
 742         /* protected by arc_buf_hdr mutex */
 743         l2arc_dev_t             *b_dev;         /* L2ARC device */
 744         uint64_t                b_daddr;        /* disk address, offset byte */
 745         /* real alloc'd buffer size depending on b_compress applied */
 746         int32_t                 b_asize;
 747 
 748         list_node_t             b_l2node;
 749 } l2arc_buf_hdr_t;
 750 
 751 struct arc_buf_hdr {
 752         /* protected by hash lock */
 753         dva_t                   b_dva;
 754         uint64_t                b_birth;
 755         /*
 756          * Even though this checksum is only set/verified when a buffer is in
 757          * the L1 cache, it needs to be in the set of common fields because it
 758          * must be preserved from the time before a buffer is written out to
 759          * L2ARC until after it is read back in.
 760          */
 761         zio_cksum_t             *b_freeze_cksum;
 762 
 763         arc_buf_hdr_t           *b_hash_next;
 764         arc_flags_t             b_flags;
 765 
 766         /* immutable */
 767         int32_t                 b_size;
 768         uint64_t                b_spa;
 769 
 770         /* L2ARC fields. Undefined when not in L2ARC. */
 771         l2arc_buf_hdr_t         b_l2hdr;
 772         /* L1ARC fields. Undefined when in l2arc_only state */
 773         l1arc_buf_hdr_t         b_l1hdr;
 774 };
 775 
 776 static arc_buf_t *arc_eviction_list;
 777 static arc_buf_hdr_t arc_eviction_hdr;
 778 
 779 #define GHOST_STATE(state)      \
 780         ((state) == arc_mru_ghost || (state) == arc_mfu_ghost ||        \
 781         (state) == arc_l2c_only)
 782 
 783 #define HDR_IN_HASH_TABLE(hdr)  ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
 784 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
 785 #define HDR_IO_ERROR(hdr)       ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
 786 #define HDR_PREFETCH(hdr)       ((hdr)->b_flags & ARC_FLAG_PREFETCH)
 787 #define HDR_FREED_IN_READ(hdr)  ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ)
 788 #define HDR_BUF_AVAILABLE(hdr)  ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE)
 789 
 790 #define HDR_L2CACHE(hdr)        ((hdr)->b_flags & ARC_FLAG_L2CACHE)
 791 #define HDR_L2COMPRESS(hdr)     ((hdr)->b_flags & ARC_FLAG_L2COMPRESS)
 792 #define HDR_L2_READING(hdr)     \
 793             (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) &&       \
 794             ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
 795 #define HDR_L2_WRITING(hdr)     ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
 796 #define HDR_L2_EVICTED(hdr)     ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
 797 #define HDR_L2_WRITE_HEAD(hdr)  ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
 798 
 799 #define HDR_ISTYPE_METADATA(hdr)        \
 800             ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
 801 #define HDR_ISTYPE_DATA(hdr)    (!HDR_ISTYPE_METADATA(hdr))
 802 
 803 #define HDR_HAS_L1HDR(hdr)      ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
 804 #define HDR_HAS_L2HDR(hdr)      ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
 805 
 806 /* For storing compression mode in b_flags */
 807 #define HDR_COMPRESS_OFFSET     24
 808 #define HDR_COMPRESS_NBITS      7
 809 
 810 #define HDR_GET_COMPRESS(hdr)   ((enum zio_compress)BF32_GET(hdr->b_flags, \
 811             HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS))
 812 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET(hdr->b_flags, \
 813             HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS, (cmp))
 814 
 815 /*
 816  * Other sizes
 817  */
 818 
 819 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
 820 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
 821 
 822 /*
 823  * Hash table routines
 824  */
 825 
 826 #define HT_LOCK_PAD     64
 827 
 828 struct ht_lock {
 829         kmutex_t        ht_lock;
 830 #ifdef _KERNEL
 831         unsigned char   pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
 832 #endif
 833 };
 834 
 835 #define BUF_LOCKS 256
 836 typedef struct buf_hash_table {
 837         uint64_t ht_mask;
 838         arc_buf_hdr_t **ht_table;
 839         struct ht_lock ht_locks[BUF_LOCKS];
 840 } buf_hash_table_t;
 841 
 842 static buf_hash_table_t buf_hash_table;
 843 
 844 #define BUF_HASH_INDEX(spa, dva, birth) \
 845         (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
 846 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
 847 #define BUF_HASH_LOCK(idx)      (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
 848 #define HDR_LOCK(hdr) \
 849         (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
 850 
 851 uint64_t zfs_crc64_table[256];
 852 
 853 /*
 854  * Level 2 ARC
 855  */
 856 
 857 #define L2ARC_WRITE_SIZE        (8 * 1024 * 1024)       /* initial write max */
 858 #define L2ARC_HEADROOM          2                       /* num of writes */
 859 /*
 860  * If we discover during ARC scan any buffers to be compressed, we boost
 861  * our headroom for the next scanning cycle by this percentage multiple.
 862  */
 863 #define L2ARC_HEADROOM_BOOST    200
 864 #define L2ARC_FEED_SECS         1               /* caching interval secs */
 865 #define L2ARC_FEED_MIN_MS       200             /* min caching interval ms */
 866 
 867 /*
 868  * Used to distinguish headers that are being process by
 869  * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk
 870  * address. This can happen when the header is added to the l2arc's list
 871  * of buffers to write in the first stage of l2arc_write_buffers(), but
 872  * has not yet been written out which happens in the second stage of
 873  * l2arc_write_buffers().
 874  */
 875 #define L2ARC_ADDR_UNSET        ((uint64_t)(-1))
 876 
 877 #define l2arc_writes_sent       ARCSTAT(arcstat_l2_writes_sent)
 878 #define l2arc_writes_done       ARCSTAT(arcstat_l2_writes_done)
 879 
 880 /* L2ARC Performance Tunables */
 881 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE;    /* default max write size */
 882 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE;  /* extra write during warmup */
 883 uint64_t l2arc_headroom = L2ARC_HEADROOM;       /* number of dev writes */
 884 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
 885 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS;     /* interval seconds */
 886 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
 887 boolean_t l2arc_noprefetch = B_TRUE;            /* don't cache prefetch bufs */
 888 boolean_t l2arc_feed_again = B_TRUE;            /* turbo warmup */
 889 boolean_t l2arc_norw = B_TRUE;                  /* no reads during writes */
 890 
 891 /*
 892  * L2ARC Internals
 893  */
 894 struct l2arc_dev {
 895         vdev_t                  *l2ad_vdev;     /* vdev */
 896         spa_t                   *l2ad_spa;      /* spa */
 897         uint64_t                l2ad_hand;      /* next write location */
 898         uint64_t                l2ad_start;     /* first addr on device */
 899         uint64_t                l2ad_end;       /* last addr on device */
 900         boolean_t               l2ad_first;     /* first sweep through */
 901         boolean_t               l2ad_writing;   /* currently writing */
 902         kmutex_t                l2ad_mtx;       /* lock for buffer list */
 903         list_t                  l2ad_buflist;   /* buffer list */
 904         list_node_t             l2ad_node;      /* device list node */
 905         refcount_t              l2ad_alloc;     /* allocated bytes */
 906 };
 907 
 908 static list_t L2ARC_dev_list;                   /* device list */
 909 static list_t *l2arc_dev_list;                  /* device list pointer */
 910 static kmutex_t l2arc_dev_mtx;                  /* device list mutex */
 911 static l2arc_dev_t *l2arc_dev_last;             /* last device used */
 912 static list_t L2ARC_free_on_write;              /* free after write buf list */
 913 static list_t *l2arc_free_on_write;             /* free after write list ptr */
 914 static kmutex_t l2arc_free_on_write_mtx;        /* mutex for list */
 915 static uint64_t l2arc_ndev;                     /* number of devices */
 916 
 917 typedef struct l2arc_read_callback {
 918         arc_buf_t               *l2rcb_buf;             /* read buffer */
 919         spa_t                   *l2rcb_spa;             /* spa */
 920         blkptr_t                l2rcb_bp;               /* original blkptr */
 921         zbookmark_phys_t        l2rcb_zb;               /* original bookmark */
 922         int                     l2rcb_flags;            /* original flags */
 923         enum zio_compress       l2rcb_compress;         /* applied compress */
 924 } l2arc_read_callback_t;
 925 
 926 typedef struct l2arc_write_callback {
 927         l2arc_dev_t     *l2wcb_dev;             /* device info */
 928         arc_buf_hdr_t   *l2wcb_head;            /* head of write buflist */
 929 } l2arc_write_callback_t;
 930 
 931 typedef struct l2arc_data_free {
 932         /* protected by l2arc_free_on_write_mtx */
 933         void            *l2df_data;
 934         size_t          l2df_size;
 935         void            (*l2df_func)(void *, size_t);
 936         list_node_t     l2df_list_node;
 937 } l2arc_data_free_t;
 938 
 939 static kmutex_t l2arc_feed_thr_lock;
 940 static kcondvar_t l2arc_feed_thr_cv;
 941 static uint8_t l2arc_thread_exit;
 942 
 943 static void arc_get_data_buf(arc_buf_t *);
 944 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
 945 static boolean_t arc_is_overflowing();
 946 static void arc_buf_watch(arc_buf_t *);
 947 
 948 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
 949 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
 950 
 951 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
 952 static void l2arc_read_done(zio_t *);
 953 
 954 static boolean_t l2arc_compress_buf(arc_buf_hdr_t *);
 955 static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress);
 956 static void l2arc_release_cdata_buf(arc_buf_hdr_t *);
 957 
 958 static uint64_t
 959 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
 960 {
 961         uint8_t *vdva = (uint8_t *)dva;
 962         uint64_t crc = -1ULL;
 963         int i;
 964 
 965         ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
 966 
 967         for (i = 0; i < sizeof (dva_t); i++)
 968                 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
 969 
 970         crc ^= (spa>>8) ^ birth;
 971 
 972         return (crc);
 973 }
 974 
 975 #define BUF_EMPTY(buf)                                          \
 976         ((buf)->b_dva.dva_word[0] == 0 &&                    \
 977         (buf)->b_dva.dva_word[1] == 0)
 978 
 979 #define BUF_EQUAL(spa, dva, birth, buf)                         \
 980         ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) &&       \
 981         ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) &&       \
 982         ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
 983 
 984 static void
 985 buf_discard_identity(arc_buf_hdr_t *hdr)
 986 {
 987         hdr->b_dva.dva_word[0] = 0;
 988         hdr->b_dva.dva_word[1] = 0;
 989         hdr->b_birth = 0;
 990 }
 991 
 992 static arc_buf_hdr_t *
 993 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
 994 {
 995         const dva_t *dva = BP_IDENTITY(bp);
 996         uint64_t birth = BP_PHYSICAL_BIRTH(bp);
 997         uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
 998         kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
 999         arc_buf_hdr_t *hdr;
1000 
1001         mutex_enter(hash_lock);
1002         for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1003             hdr = hdr->b_hash_next) {
1004                 if (BUF_EQUAL(spa, dva, birth, hdr)) {
1005                         *lockp = hash_lock;
1006                         return (hdr);
1007                 }
1008         }
1009         mutex_exit(hash_lock);
1010         *lockp = NULL;
1011         return (NULL);
1012 }
1013 
1014 /*
1015  * Insert an entry into the hash table.  If there is already an element
1016  * equal to elem in the hash table, then the already existing element
1017  * will be returned and the new element will not be inserted.
1018  * Otherwise returns NULL.
1019  * If lockp == NULL, the caller is assumed to already hold the hash lock.
1020  */
1021 static arc_buf_hdr_t *
1022 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1023 {
1024         uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1025         kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1026         arc_buf_hdr_t *fhdr;
1027         uint32_t i;
1028 
1029         ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1030         ASSERT(hdr->b_birth != 0);
1031         ASSERT(!HDR_IN_HASH_TABLE(hdr));
1032 
1033         if (lockp != NULL) {
1034                 *lockp = hash_lock;
1035                 mutex_enter(hash_lock);
1036         } else {
1037                 ASSERT(MUTEX_HELD(hash_lock));
1038         }
1039 
1040         for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1041             fhdr = fhdr->b_hash_next, i++) {
1042                 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1043                         return (fhdr);
1044         }
1045 
1046         hdr->b_hash_next = buf_hash_table.ht_table[idx];
1047         buf_hash_table.ht_table[idx] = hdr;
1048         hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
1049 
1050         /* collect some hash table performance data */
1051         if (i > 0) {
1052                 ARCSTAT_BUMP(arcstat_hash_collisions);
1053                 if (i == 1)
1054                         ARCSTAT_BUMP(arcstat_hash_chains);
1055 
1056                 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1057         }
1058 
1059         ARCSTAT_BUMP(arcstat_hash_elements);
1060         ARCSTAT_MAXSTAT(arcstat_hash_elements);
1061 
1062         return (NULL);
1063 }
1064 
1065 static void
1066 buf_hash_remove(arc_buf_hdr_t *hdr)
1067 {
1068         arc_buf_hdr_t *fhdr, **hdrp;
1069         uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1070 
1071         ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1072         ASSERT(HDR_IN_HASH_TABLE(hdr));
1073 
1074         hdrp = &buf_hash_table.ht_table[idx];
1075         while ((fhdr = *hdrp) != hdr) {
1076                 ASSERT(fhdr != NULL);
1077                 hdrp = &fhdr->b_hash_next;
1078         }
1079         *hdrp = hdr->b_hash_next;
1080         hdr->b_hash_next = NULL;
1081         hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE;
1082 
1083         /* collect some hash table performance data */
1084         ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1085 
1086         if (buf_hash_table.ht_table[idx] &&
1087             buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1088                 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1089 }
1090 
1091 /*
1092  * Global data structures and functions for the buf kmem cache.
1093  */
1094 static kmem_cache_t *hdr_full_cache;
1095 static kmem_cache_t *hdr_l2only_cache;
1096 static kmem_cache_t *buf_cache;
1097 
1098 static void
1099 buf_fini(void)
1100 {
1101         int i;
1102 
1103         kmem_free(buf_hash_table.ht_table,
1104             (buf_hash_table.ht_mask + 1) * sizeof (void *));
1105         for (i = 0; i < BUF_LOCKS; i++)
1106                 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1107         kmem_cache_destroy(hdr_full_cache);
1108         kmem_cache_destroy(hdr_l2only_cache);
1109         kmem_cache_destroy(buf_cache);
1110 }
1111 
1112 /*
1113  * Constructor callback - called when the cache is empty
1114  * and a new buf is requested.
1115  */
1116 /* ARGSUSED */
1117 static int
1118 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1119 {
1120         arc_buf_hdr_t *hdr = vbuf;
1121 
1122         bzero(hdr, HDR_FULL_SIZE);
1123         cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1124         refcount_create(&hdr->b_l1hdr.b_refcnt);
1125         mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1126         multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1127         arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1128 
1129         return (0);
1130 }
1131 
1132 /* ARGSUSED */
1133 static int
1134 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1135 {
1136         arc_buf_hdr_t *hdr = vbuf;
1137 
1138         bzero(hdr, HDR_L2ONLY_SIZE);
1139         arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1140 
1141         return (0);
1142 }
1143 
1144 /* ARGSUSED */
1145 static int
1146 buf_cons(void *vbuf, void *unused, int kmflag)
1147 {
1148         arc_buf_t *buf = vbuf;
1149 
1150         bzero(buf, sizeof (arc_buf_t));
1151         mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1152         arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1153 
1154         return (0);
1155 }
1156 
1157 /*
1158  * Destructor callback - called when a cached buf is
1159  * no longer required.
1160  */
1161 /* ARGSUSED */
1162 static void
1163 hdr_full_dest(void *vbuf, void *unused)
1164 {
1165         arc_buf_hdr_t *hdr = vbuf;
1166 
1167         ASSERT(BUF_EMPTY(hdr));
1168         cv_destroy(&hdr->b_l1hdr.b_cv);
1169         refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1170         mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1171         ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1172         arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1173 }
1174 
1175 /* ARGSUSED */
1176 static void
1177 hdr_l2only_dest(void *vbuf, void *unused)
1178 {
1179         arc_buf_hdr_t *hdr = vbuf;
1180 
1181         ASSERT(BUF_EMPTY(hdr));
1182         arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1183 }
1184 
1185 /* ARGSUSED */
1186 static void
1187 buf_dest(void *vbuf, void *unused)
1188 {
1189         arc_buf_t *buf = vbuf;
1190 
1191         mutex_destroy(&buf->b_evict_lock);
1192         arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1193 }
1194 
1195 /*
1196  * Reclaim callback -- invoked when memory is low.
1197  */
1198 /* ARGSUSED */
1199 static void
1200 hdr_recl(void *unused)
1201 {
1202         dprintf("hdr_recl called\n");
1203         /*
1204          * umem calls the reclaim func when we destroy the buf cache,
1205          * which is after we do arc_fini().
1206          */
1207         if (!arc_dead)
1208                 cv_signal(&arc_reclaim_thread_cv);
1209 }
1210 
1211 static void
1212 buf_init(void)
1213 {
1214         uint64_t *ct;
1215         uint64_t hsize = 1ULL << 12;
1216         int i, j;
1217 
1218         /*
1219          * The hash table is big enough to fill all of physical memory
1220          * with an average block size of zfs_arc_average_blocksize (default 8K).
1221          * By default, the table will take up
1222          * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1223          */
1224         while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
1225                 hsize <<= 1;
1226 retry:
1227         buf_hash_table.ht_mask = hsize - 1;
1228         buf_hash_table.ht_table =
1229             kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1230         if (buf_hash_table.ht_table == NULL) {
1231                 ASSERT(hsize > (1ULL << 8));
1232                 hsize >>= 1;
1233                 goto retry;
1234         }
1235 
1236         hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1237             0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1238         hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1239             HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1240             NULL, NULL, 0);
1241         buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1242             0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1243 
1244         for (i = 0; i < 256; i++)
1245                 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1246                         *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1247 
1248         for (i = 0; i < BUF_LOCKS; i++) {
1249                 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1250                     NULL, MUTEX_DEFAULT, NULL);
1251         }
1252 }
1253 
1254 /*
1255  * Transition between the two allocation states for the arc_buf_hdr struct.
1256  * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
1257  * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
1258  * version is used when a cache buffer is only in the L2ARC in order to reduce
1259  * memory usage.
1260  */
1261 static arc_buf_hdr_t *
1262 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
1263 {
1264         ASSERT(HDR_HAS_L2HDR(hdr));
1265 
1266         arc_buf_hdr_t *nhdr;
1267         l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1268 
1269         ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
1270             (old == hdr_l2only_cache && new == hdr_full_cache));
1271 
1272         nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
1273 
1274         ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
1275         buf_hash_remove(hdr);
1276 
1277         bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
1278 
1279         if (new == hdr_full_cache) {
1280                 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1281                 /*
1282                  * arc_access and arc_change_state need to be aware that a
1283                  * header has just come out of L2ARC, so we set its state to
1284                  * l2c_only even though it's about to change.
1285                  */
1286                 nhdr->b_l1hdr.b_state = arc_l2c_only;
1287 
1288                 /* Verify previous threads set to NULL before freeing */
1289                 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1290         } else {
1291                 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1292                 ASSERT0(hdr->b_l1hdr.b_datacnt);
1293 
1294                 /*
1295                  * If we've reached here, We must have been called from
1296                  * arc_evict_hdr(), as such we should have already been
1297                  * removed from any ghost list we were previously on
1298                  * (which protects us from racing with arc_evict_state),
1299                  * thus no locking is needed during this check.
1300                  */
1301                 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1302 
1303                 /*
1304                  * A buffer must not be moved into the arc_l2c_only
1305                  * state if it's not finished being written out to the
1306                  * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field
1307                  * might try to be accessed, even though it was removed.
1308                  */
1309                 VERIFY(!HDR_L2_WRITING(hdr));
1310                 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1311 
1312 #ifdef ZFS_DEBUG
1313                 if (hdr->b_l1hdr.b_thawed != NULL)
1314                         kmem_free(hdr->b_l1hdr.b_thawed, 1);
1315                         hdr->b_l1hdr.b_thawed = NULL;
1316 #endif
1317 
1318                 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR;
1319         }
1320         /*
1321          * The header has been reallocated so we need to re-insert it into any
1322          * lists it was on.
1323          */
1324         (void) buf_hash_insert(nhdr, NULL);
1325 
1326         ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
1327 
1328         mutex_enter(&dev->l2ad_mtx);
1329 
1330         /*
1331          * We must place the realloc'ed header back into the list at
1332          * the same spot. Otherwise, if it's placed earlier in the list,
1333          * l2arc_write_buffers() could find it during the function's
1334          * write phase, and try to write it out to the l2arc.
1335          */
1336         list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
1337         list_remove(&dev->l2ad_buflist, hdr);
1338 
1339         mutex_exit(&dev->l2ad_mtx);
1340 
1341         /*
1342          * Since we're using the pointer address as the tag when
1343          * incrementing and decrementing the l2ad_alloc refcount, we
1344          * must remove the old pointer (that we're about to destroy) and
1345          * add the new pointer to the refcount. Otherwise we'd remove
1346          * the wrong pointer address when calling arc_hdr_destroy() later.
1347          */
1348 
1349         (void) refcount_remove_many(&dev->l2ad_alloc,
1350             hdr->b_l2hdr.b_asize, hdr);
1351 
1352         (void) refcount_add_many(&dev->l2ad_alloc,
1353             nhdr->b_l2hdr.b_asize, nhdr);
1354 
1355         buf_discard_identity(hdr);
1356         hdr->b_freeze_cksum = NULL;
1357         kmem_cache_free(old, hdr);
1358 
1359         return (nhdr);
1360 }
1361 
1362 
1363 #define ARC_MINTIME     (hz>>4) /* 62 ms */
1364 
1365 static void
1366 arc_cksum_verify(arc_buf_t *buf)
1367 {
1368         zio_cksum_t zc;
1369 
1370         if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1371                 return;
1372 
1373         mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1374         if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) {
1375                 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1376                 return;
1377         }
1378         fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1379         if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1380                 panic("buffer modified while frozen!");
1381         mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1382 }
1383 
1384 static int
1385 arc_cksum_equal(arc_buf_t *buf)
1386 {
1387         zio_cksum_t zc;
1388         int equal;
1389 
1390         mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1391         fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1392         equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1393         mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1394 
1395         return (equal);
1396 }
1397 
1398 static void
1399 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1400 {
1401         if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1402                 return;
1403 
1404         mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1405         if (buf->b_hdr->b_freeze_cksum != NULL) {
1406                 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1407                 return;
1408         }
1409         buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1410         fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1411             buf->b_hdr->b_freeze_cksum);
1412         mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1413         arc_buf_watch(buf);
1414 }
1415 
1416 #ifndef _KERNEL
1417 typedef struct procctl {
1418         long cmd;
1419         prwatch_t prwatch;
1420 } procctl_t;
1421 #endif
1422 
1423 /* ARGSUSED */
1424 static void
1425 arc_buf_unwatch(arc_buf_t *buf)
1426 {
1427 #ifndef _KERNEL
1428         if (arc_watch) {
1429                 int result;
1430                 procctl_t ctl;
1431                 ctl.cmd = PCWATCH;
1432                 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1433                 ctl.prwatch.pr_size = 0;
1434                 ctl.prwatch.pr_wflags = 0;
1435                 result = write(arc_procfd, &ctl, sizeof (ctl));
1436                 ASSERT3U(result, ==, sizeof (ctl));
1437         }
1438 #endif
1439 }
1440 
1441 /* ARGSUSED */
1442 static void
1443 arc_buf_watch(arc_buf_t *buf)
1444 {
1445 #ifndef _KERNEL
1446         if (arc_watch) {
1447                 int result;
1448                 procctl_t ctl;
1449                 ctl.cmd = PCWATCH;
1450                 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1451                 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1452                 ctl.prwatch.pr_wflags = WA_WRITE;
1453                 result = write(arc_procfd, &ctl, sizeof (ctl));
1454                 ASSERT3U(result, ==, sizeof (ctl));
1455         }
1456 #endif
1457 }
1458 
1459 static arc_buf_contents_t
1460 arc_buf_type(arc_buf_hdr_t *hdr)
1461 {
1462         if (HDR_ISTYPE_METADATA(hdr)) {
1463                 return (ARC_BUFC_METADATA);
1464         } else {
1465                 return (ARC_BUFC_DATA);
1466         }
1467 }
1468 
1469 static uint32_t
1470 arc_bufc_to_flags(arc_buf_contents_t type)
1471 {
1472         switch (type) {
1473         case ARC_BUFC_DATA:
1474                 /* metadata field is 0 if buffer contains normal data */
1475                 return (0);
1476         case ARC_BUFC_METADATA:
1477                 return (ARC_FLAG_BUFC_METADATA);
1478         default:
1479                 break;
1480         }
1481         panic("undefined ARC buffer type!");
1482         return ((uint32_t)-1);
1483 }
1484 
1485 void
1486 arc_buf_thaw(arc_buf_t *buf)
1487 {
1488         if (zfs_flags & ZFS_DEBUG_MODIFY) {
1489                 if (buf->b_hdr->b_l1hdr.b_state != arc_anon)
1490                         panic("modifying non-anon buffer!");
1491                 if (HDR_IO_IN_PROGRESS(buf->b_hdr))
1492                         panic("modifying buffer while i/o in progress!");
1493                 arc_cksum_verify(buf);
1494         }
1495 
1496         mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1497         if (buf->b_hdr->b_freeze_cksum != NULL) {
1498                 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1499                 buf->b_hdr->b_freeze_cksum = NULL;
1500         }
1501 
1502 #ifdef ZFS_DEBUG
1503         if (zfs_flags & ZFS_DEBUG_MODIFY) {
1504                 if (buf->b_hdr->b_l1hdr.b_thawed != NULL)
1505                         kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1);
1506                 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1507         }
1508 #endif
1509 
1510         mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1511 
1512         arc_buf_unwatch(buf);
1513 }
1514 
1515 void
1516 arc_buf_freeze(arc_buf_t *buf)
1517 {
1518         kmutex_t *hash_lock;
1519 
1520         if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1521                 return;
1522 
1523         hash_lock = HDR_LOCK(buf->b_hdr);
1524         mutex_enter(hash_lock);
1525 
1526         ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1527             buf->b_hdr->b_l1hdr.b_state == arc_anon);
1528         arc_cksum_compute(buf, B_FALSE);
1529         mutex_exit(hash_lock);
1530 
1531 }
1532 
1533 static void
1534 add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1535 {
1536         ASSERT(HDR_HAS_L1HDR(hdr));
1537         ASSERT(MUTEX_HELD(hash_lock));
1538         arc_state_t *state = hdr->b_l1hdr.b_state;
1539 
1540         if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
1541             (state != arc_anon)) {
1542                 /* We don't use the L2-only state list. */
1543                 if (state != arc_l2c_only) {
1544                         arc_buf_contents_t type = arc_buf_type(hdr);
1545                         uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt;
1546                         multilist_t *list = &state->arcs_list[type];
1547                         uint64_t *size = &state->arcs_lsize[type];
1548 
1549                         multilist_remove(list, hdr);
1550 
1551                         if (GHOST_STATE(state)) {
1552                                 ASSERT0(hdr->b_l1hdr.b_datacnt);
1553                                 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
1554                                 delta = hdr->b_size;
1555                         }
1556                         ASSERT(delta > 0);
1557                         ASSERT3U(*size, >=, delta);
1558                         atomic_add_64(size, -delta);
1559                 }
1560                 /* remove the prefetch flag if we get a reference */
1561                 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
1562         }
1563 }
1564 
1565 static int
1566 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1567 {
1568         int cnt;
1569         arc_state_t *state = hdr->b_l1hdr.b_state;
1570 
1571         ASSERT(HDR_HAS_L1HDR(hdr));
1572         ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1573         ASSERT(!GHOST_STATE(state));
1574 
1575         /*
1576          * arc_l2c_only counts as a ghost state so we don't need to explicitly
1577          * check to prevent usage of the arc_l2c_only list.
1578          */
1579         if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
1580             (state != arc_anon)) {
1581                 arc_buf_contents_t type = arc_buf_type(hdr);
1582                 multilist_t *list = &state->arcs_list[type];
1583                 uint64_t *size = &state->arcs_lsize[type];
1584 
1585                 multilist_insert(list, hdr);
1586 
1587                 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
1588                 atomic_add_64(size, hdr->b_size *
1589                     hdr->b_l1hdr.b_datacnt);
1590         }
1591         return (cnt);
1592 }
1593 
1594 /*
1595  * Move the supplied buffer to the indicated state. The hash lock
1596  * for the buffer must be held by the caller.
1597  */
1598 static void
1599 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
1600     kmutex_t *hash_lock)
1601 {
1602         arc_state_t *old_state;
1603         int64_t refcnt;
1604         uint32_t datacnt;
1605         uint64_t from_delta, to_delta;
1606         arc_buf_contents_t buftype = arc_buf_type(hdr);
1607 
1608         /*
1609          * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
1610          * in arc_read() when bringing a buffer out of the L2ARC.  However, the
1611          * L1 hdr doesn't always exist when we change state to arc_anon before
1612          * destroying a header, in which case reallocating to add the L1 hdr is
1613          * pointless.
1614          */
1615         if (HDR_HAS_L1HDR(hdr)) {
1616                 old_state = hdr->b_l1hdr.b_state;
1617                 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
1618                 datacnt = hdr->b_l1hdr.b_datacnt;
1619         } else {
1620                 old_state = arc_l2c_only;
1621                 refcnt = 0;
1622                 datacnt = 0;
1623         }
1624 
1625         ASSERT(MUTEX_HELD(hash_lock));
1626         ASSERT3P(new_state, !=, old_state);
1627         ASSERT(refcnt == 0 || datacnt > 0);
1628         ASSERT(!GHOST_STATE(new_state) || datacnt == 0);
1629         ASSERT(old_state != arc_anon || datacnt <= 1);
1630 
1631         from_delta = to_delta = datacnt * hdr->b_size;
1632 
1633         /*
1634          * If this buffer is evictable, transfer it from the
1635          * old state list to the new state list.
1636          */
1637         if (refcnt == 0) {
1638                 if (old_state != arc_anon && old_state != arc_l2c_only) {
1639                         uint64_t *size = &old_state->arcs_lsize[buftype];
1640 
1641                         ASSERT(HDR_HAS_L1HDR(hdr));
1642                         multilist_remove(&old_state->arcs_list[buftype], hdr);
1643 
1644                         /*
1645                          * If prefetching out of the ghost cache,
1646                          * we will have a non-zero datacnt.
1647                          */
1648                         if (GHOST_STATE(old_state) && datacnt == 0) {
1649                                 /* ghost elements have a ghost size */
1650                                 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1651                                 from_delta = hdr->b_size;
1652                         }
1653                         ASSERT3U(*size, >=, from_delta);
1654                         atomic_add_64(size, -from_delta);
1655                 }
1656                 if (new_state != arc_anon && new_state != arc_l2c_only) {
1657                         uint64_t *size = &new_state->arcs_lsize[buftype];
1658 
1659                         /*
1660                          * An L1 header always exists here, since if we're
1661                          * moving to some L1-cached state (i.e. not l2c_only or
1662                          * anonymous), we realloc the header to add an L1hdr
1663                          * beforehand.
1664                          */
1665                         ASSERT(HDR_HAS_L1HDR(hdr));
1666                         multilist_insert(&new_state->arcs_list[buftype], hdr);
1667 
1668                         /* ghost elements have a ghost size */
1669                         if (GHOST_STATE(new_state)) {
1670                                 ASSERT0(datacnt);
1671                                 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1672                                 to_delta = hdr->b_size;
1673                         }
1674                         atomic_add_64(size, to_delta);
1675                 }
1676         }
1677 
1678         ASSERT(!BUF_EMPTY(hdr));
1679         if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
1680                 buf_hash_remove(hdr);
1681 
1682         /* adjust state sizes (ignore arc_l2c_only) */
1683 
1684         if (to_delta && new_state != arc_l2c_only) {
1685                 ASSERT(HDR_HAS_L1HDR(hdr));
1686                 if (GHOST_STATE(new_state)) {
1687                         ASSERT0(datacnt);
1688 
1689                         /*
1690                          * We moving a header to a ghost state, we first
1691                          * remove all arc buffers. Thus, we'll have a
1692                          * datacnt of zero, and no arc buffer to use for
1693                          * the reference. As a result, we use the arc
1694                          * header pointer for the reference.
1695                          */
1696                         (void) refcount_add_many(&new_state->arcs_size,
1697                             hdr->b_size, hdr);
1698                 } else {
1699                         ASSERT3U(datacnt, !=, 0);
1700 
1701                         /*
1702                          * Each individual buffer holds a unique reference,
1703                          * thus we must remove each of these references one
1704                          * at a time.
1705                          */
1706                         for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1707                             buf = buf->b_next) {
1708                                 (void) refcount_add_many(&new_state->arcs_size,
1709                                     hdr->b_size, buf);
1710                         }
1711                 }
1712         }
1713 
1714         if (from_delta && old_state != arc_l2c_only) {
1715                 ASSERT(HDR_HAS_L1HDR(hdr));
1716                 if (GHOST_STATE(old_state)) {
1717                         /*
1718                          * When moving a header off of a ghost state,
1719                          * there's the possibility for datacnt to be
1720                          * non-zero. This is because we first add the
1721                          * arc buffer to the header prior to changing
1722                          * the header's state. Since we used the header
1723                          * for the reference when putting the header on
1724                          * the ghost state, we must balance that and use
1725                          * the header when removing off the ghost state
1726                          * (even though datacnt is non zero).
1727                          */
1728 
1729                         IMPLY(datacnt == 0, new_state == arc_anon ||
1730                             new_state == arc_l2c_only);
1731 
1732                         (void) refcount_remove_many(&old_state->arcs_size,
1733                             hdr->b_size, hdr);
1734                 } else {
1735                         ASSERT3P(datacnt, !=, 0);
1736 
1737                         /*
1738                          * Each individual buffer holds a unique reference,
1739                          * thus we must remove each of these references one
1740                          * at a time.
1741                          */
1742                         for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1743                             buf = buf->b_next) {
1744                                 (void) refcount_remove_many(
1745                                     &old_state->arcs_size, hdr->b_size, buf);
1746                         }
1747                 }
1748         }
1749 
1750         if (HDR_HAS_L1HDR(hdr))
1751                 hdr->b_l1hdr.b_state = new_state;
1752 
1753         /*
1754          * L2 headers should never be on the L2 state list since they don't
1755          * have L1 headers allocated.
1756          */
1757         ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
1758             multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
1759 }
1760 
1761 void
1762 arc_space_consume(uint64_t space, arc_space_type_t type)
1763 {
1764         ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1765 
1766         switch (type) {
1767         case ARC_SPACE_DATA:
1768                 ARCSTAT_INCR(arcstat_data_size, space);
1769                 break;
1770         case ARC_SPACE_META:
1771                 ARCSTAT_INCR(arcstat_metadata_size, space);
1772                 break;
1773         case ARC_SPACE_OTHER:
1774                 ARCSTAT_INCR(arcstat_other_size, space);
1775                 break;
1776         case ARC_SPACE_HDRS:
1777                 ARCSTAT_INCR(arcstat_hdr_size, space);
1778                 break;
1779         case ARC_SPACE_L2HDRS:
1780                 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1781                 break;
1782         }
1783 
1784         if (type != ARC_SPACE_DATA)
1785                 ARCSTAT_INCR(arcstat_meta_used, space);
1786 
1787         atomic_add_64(&arc_size, space);
1788 }
1789 
1790 void
1791 arc_space_return(uint64_t space, arc_space_type_t type)
1792 {
1793         ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1794 
1795         switch (type) {
1796         case ARC_SPACE_DATA:
1797                 ARCSTAT_INCR(arcstat_data_size, -space);
1798                 break;
1799         case ARC_SPACE_META:
1800                 ARCSTAT_INCR(arcstat_metadata_size, -space);
1801                 break;
1802         case ARC_SPACE_OTHER:
1803                 ARCSTAT_INCR(arcstat_other_size, -space);
1804                 break;
1805         case ARC_SPACE_HDRS:
1806                 ARCSTAT_INCR(arcstat_hdr_size, -space);
1807                 break;
1808         case ARC_SPACE_L2HDRS:
1809                 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1810                 break;
1811         }
1812 
1813         if (type != ARC_SPACE_DATA) {
1814                 ASSERT(arc_meta_used >= space);
1815                 if (arc_meta_max < arc_meta_used)
1816                         arc_meta_max = arc_meta_used;
1817                 ARCSTAT_INCR(arcstat_meta_used, -space);
1818         }
1819 
1820         ASSERT(arc_size >= space);
1821         atomic_add_64(&arc_size, -space);
1822 }
1823 
1824 arc_buf_t *
1825 arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type)
1826 {
1827         arc_buf_hdr_t *hdr;
1828         arc_buf_t *buf;
1829 
1830         ASSERT3U(size, >, 0);
1831         hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
1832         ASSERT(BUF_EMPTY(hdr));
1833         ASSERT3P(hdr->b_freeze_cksum, ==, NULL);
1834         hdr->b_size = size;
1835         hdr->b_spa = spa_load_guid(spa);
1836 
1837         buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1838         buf->b_hdr = hdr;
1839         buf->b_data = NULL;
1840         buf->b_efunc = NULL;
1841         buf->b_private = NULL;
1842         buf->b_next = NULL;
1843 
1844         hdr->b_flags = arc_bufc_to_flags(type);
1845         hdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1846 
1847         hdr->b_l1hdr.b_buf = buf;
1848         hdr->b_l1hdr.b_state = arc_anon;
1849         hdr->b_l1hdr.b_arc_access = 0;
1850         hdr->b_l1hdr.b_datacnt = 1;
1851         hdr->b_l1hdr.b_tmp_cdata = NULL;
1852 
1853         arc_get_data_buf(buf);
1854         ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
1855         (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
1856 
1857         return (buf);
1858 }
1859 
1860 static char *arc_onloan_tag = "onloan";
1861 
1862 /*
1863  * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1864  * flight data by arc_tempreserve_space() until they are "returned". Loaned
1865  * buffers must be returned to the arc before they can be used by the DMU or
1866  * freed.
1867  */
1868 arc_buf_t *
1869 arc_loan_buf(spa_t *spa, int size)
1870 {
1871         arc_buf_t *buf;
1872 
1873         buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1874 
1875         atomic_add_64(&arc_loaned_bytes, size);
1876         return (buf);
1877 }
1878 
1879 /*
1880  * Return a loaned arc buffer to the arc.
1881  */
1882 void
1883 arc_return_buf(arc_buf_t *buf, void *tag)
1884 {
1885         arc_buf_hdr_t *hdr = buf->b_hdr;
1886 
1887         ASSERT(buf->b_data != NULL);
1888         ASSERT(HDR_HAS_L1HDR(hdr));
1889         (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
1890         (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
1891 
1892         atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1893 }
1894 
1895 /* Detach an arc_buf from a dbuf (tag) */
1896 void
1897 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1898 {
1899         arc_buf_hdr_t *hdr = buf->b_hdr;
1900 
1901         ASSERT(buf->b_data != NULL);
1902         ASSERT(HDR_HAS_L1HDR(hdr));
1903         (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
1904         (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
1905         buf->b_efunc = NULL;
1906         buf->b_private = NULL;
1907 
1908         atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1909 }
1910 
1911 static arc_buf_t *
1912 arc_buf_clone(arc_buf_t *from)
1913 {
1914         arc_buf_t *buf;
1915         arc_buf_hdr_t *hdr = from->b_hdr;
1916         uint64_t size = hdr->b_size;
1917 
1918         ASSERT(HDR_HAS_L1HDR(hdr));
1919         ASSERT(hdr->b_l1hdr.b_state != arc_anon);
1920 
1921         buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1922         buf->b_hdr = hdr;
1923         buf->b_data = NULL;
1924         buf->b_efunc = NULL;
1925         buf->b_private = NULL;
1926         buf->b_next = hdr->b_l1hdr.b_buf;
1927         hdr->b_l1hdr.b_buf = buf;
1928         arc_get_data_buf(buf);
1929         bcopy(from->b_data, buf->b_data, size);
1930 
1931         /*
1932          * This buffer already exists in the arc so create a duplicate
1933          * copy for the caller.  If the buffer is associated with user data
1934          * then track the size and number of duplicates.  These stats will be
1935          * updated as duplicate buffers are created and destroyed.
1936          */
1937         if (HDR_ISTYPE_DATA(hdr)) {
1938                 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1939                 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1940         }
1941         hdr->b_l1hdr.b_datacnt += 1;
1942         return (buf);
1943 }
1944 
1945 void
1946 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1947 {
1948         arc_buf_hdr_t *hdr;
1949         kmutex_t *hash_lock;
1950 
1951         /*
1952          * Check to see if this buffer is evicted.  Callers
1953          * must verify b_data != NULL to know if the add_ref
1954          * was successful.
1955          */
1956         mutex_enter(&buf->b_evict_lock);
1957         if (buf->b_data == NULL) {
1958                 mutex_exit(&buf->b_evict_lock);
1959                 return;
1960         }
1961         hash_lock = HDR_LOCK(buf->b_hdr);
1962         mutex_enter(hash_lock);
1963         hdr = buf->b_hdr;
1964         ASSERT(HDR_HAS_L1HDR(hdr));
1965         ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1966         mutex_exit(&buf->b_evict_lock);
1967 
1968         ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
1969             hdr->b_l1hdr.b_state == arc_mfu);
1970 
1971         add_reference(hdr, hash_lock, tag);
1972         DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1973         arc_access(hdr, hash_lock);
1974         mutex_exit(hash_lock);
1975         ARCSTAT_BUMP(arcstat_hits);
1976         ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
1977             demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
1978             data, metadata, hits);
1979 }
1980 
1981 static void
1982 arc_buf_free_on_write(void *data, size_t size,
1983     void (*free_func)(void *, size_t))
1984 {
1985         l2arc_data_free_t *df;
1986 
1987         df = kmem_alloc(sizeof (*df), KM_SLEEP);
1988         df->l2df_data = data;
1989         df->l2df_size = size;
1990         df->l2df_func = free_func;
1991         mutex_enter(&l2arc_free_on_write_mtx);
1992         list_insert_head(l2arc_free_on_write, df);
1993         mutex_exit(&l2arc_free_on_write_mtx);
1994 }
1995 
1996 /*
1997  * Free the arc data buffer.  If it is an l2arc write in progress,
1998  * the buffer is placed on l2arc_free_on_write to be freed later.
1999  */
2000 static void
2001 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
2002 {
2003         arc_buf_hdr_t *hdr = buf->b_hdr;
2004 
2005         if (HDR_L2_WRITING(hdr)) {
2006                 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func);
2007                 ARCSTAT_BUMP(arcstat_l2_free_on_write);
2008         } else {
2009                 free_func(buf->b_data, hdr->b_size);
2010         }
2011 }
2012 
2013 static void
2014 arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr)
2015 {
2016         ASSERT(HDR_HAS_L2HDR(hdr));
2017         ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx));
2018 
2019         /*
2020          * The b_tmp_cdata field is linked off of the b_l1hdr, so if
2021          * that doesn't exist, the header is in the arc_l2c_only state,
2022          * and there isn't anything to free (it's already been freed).
2023          */
2024         if (!HDR_HAS_L1HDR(hdr))
2025                 return;
2026 
2027         /*
2028          * The header isn't being written to the l2arc device, thus it
2029          * shouldn't have a b_tmp_cdata to free.
2030          */
2031         if (!HDR_L2_WRITING(hdr)) {
2032                 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2033                 return;
2034         }
2035 
2036         /*
2037          * The header does not have compression enabled. This can be due
2038          * to the buffer not being compressible, or because we're
2039          * freeing the buffer before the second phase of
2040          * l2arc_write_buffer() has started (which does the compression
2041          * step). In either case, b_tmp_cdata does not point to a
2042          * separately compressed buffer, so there's nothing to free (it
2043          * points to the same buffer as the arc_buf_t's b_data field).
2044          */
2045         if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) {
2046                 hdr->b_l1hdr.b_tmp_cdata = NULL;
2047                 return;
2048         }
2049 
2050         /*
2051          * There's nothing to free since the buffer was all zero's and
2052          * compressed to a zero length buffer.
2053          */
2054         if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_EMPTY) {
2055                 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2056                 return;
2057         }
2058 
2059         ASSERT(L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr)));
2060 
2061         arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata,
2062             hdr->b_size, zio_data_buf_free);
2063 
2064         ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write);
2065         hdr->b_l1hdr.b_tmp_cdata = NULL;
2066 }
2067 
2068 /*
2069  * Free up buf->b_data and if 'remove' is set, then pull the
2070  * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
2071  */
2072 static void
2073 arc_buf_destroy(arc_buf_t *buf, boolean_t remove)
2074 {
2075         arc_buf_t **bufp;
2076 
2077         /* free up data associated with the buf */
2078         if (buf->b_data != NULL) {
2079                 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
2080                 uint64_t size = buf->b_hdr->b_size;
2081                 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
2082 
2083                 arc_cksum_verify(buf);
2084                 arc_buf_unwatch(buf);
2085 
2086                 if (type == ARC_BUFC_METADATA) {
2087                         arc_buf_data_free(buf, zio_buf_free);
2088                         arc_space_return(size, ARC_SPACE_META);
2089                 } else {
2090                         ASSERT(type == ARC_BUFC_DATA);
2091                         arc_buf_data_free(buf, zio_data_buf_free);
2092                         arc_space_return(size, ARC_SPACE_DATA);
2093                 }
2094 
2095                 /* protected by hash lock, if in the hash table */
2096                 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) {
2097                         uint64_t *cnt = &state->arcs_lsize[type];
2098 
2099                         ASSERT(refcount_is_zero(
2100                             &buf->b_hdr->b_l1hdr.b_refcnt));
2101                         ASSERT(state != arc_anon && state != arc_l2c_only);
2102 
2103                         ASSERT3U(*cnt, >=, size);
2104                         atomic_add_64(cnt, -size);
2105                 }
2106 
2107                 (void) refcount_remove_many(&state->arcs_size, size, buf);
2108                 buf->b_data = NULL;
2109 
2110                 /*
2111                  * If we're destroying a duplicate buffer make sure
2112                  * that the appropriate statistics are updated.
2113                  */
2114                 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 &&
2115                     HDR_ISTYPE_DATA(buf->b_hdr)) {
2116                         ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
2117                         ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
2118                 }
2119                 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0);
2120                 buf->b_hdr->b_l1hdr.b_datacnt -= 1;
2121         }
2122 
2123         /* only remove the buf if requested */
2124         if (!remove)
2125                 return;
2126 
2127         /* remove the buf from the hdr list */
2128         for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf;
2129             bufp = &(*bufp)->b_next)
2130                 continue;
2131         *bufp = buf->b_next;
2132         buf->b_next = NULL;
2133 
2134         ASSERT(buf->b_efunc == NULL);
2135 
2136         /* clean up the buf */
2137         buf->b_hdr = NULL;
2138         kmem_cache_free(buf_cache, buf);
2139 }
2140 
2141 static void
2142 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
2143 {
2144         l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
2145         l2arc_dev_t *dev = l2hdr->b_dev;
2146 
2147         ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
2148         ASSERT(HDR_HAS_L2HDR(hdr));
2149 
2150         list_remove(&dev->l2ad_buflist, hdr);
2151 
2152         /*
2153          * We don't want to leak the b_tmp_cdata buffer that was
2154          * allocated in l2arc_write_buffers()
2155          */
2156         arc_buf_l2_cdata_free(hdr);
2157 
2158         /*
2159          * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then
2160          * this header is being processed by l2arc_write_buffers() (i.e.
2161          * it's in the first stage of l2arc_write_buffers()).
2162          * Re-affirming that truth here, just to serve as a reminder. If
2163          * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or
2164          * may not have its HDR_L2_WRITING flag set. (the write may have
2165          * completed, in which case HDR_L2_WRITING will be false and the
2166          * b_daddr field will point to the address of the buffer on disk).
2167          */
2168         IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr));
2169 
2170         /*
2171          * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with
2172          * l2arc_write_buffers(). Since we've just removed this header
2173          * from the l2arc buffer list, this header will never reach the
2174          * second stage of l2arc_write_buffers(), which increments the
2175          * accounting stats for this header. Thus, we must be careful
2176          * not to decrement them for this header either.
2177          */
2178         if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) {
2179                 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
2180                 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
2181 
2182                 vdev_space_update(dev->l2ad_vdev,
2183                     -l2hdr->b_asize, 0, 0);
2184 
2185                 (void) refcount_remove_many(&dev->l2ad_alloc,
2186                     l2hdr->b_asize, hdr);
2187         }
2188 
2189         hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
2190 }
2191 
2192 static void
2193 arc_hdr_destroy(arc_buf_hdr_t *hdr)
2194 {
2195         if (HDR_HAS_L1HDR(hdr)) {
2196                 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
2197                     hdr->b_l1hdr.b_datacnt > 0);
2198                 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2199                 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2200         }
2201         ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2202         ASSERT(!HDR_IN_HASH_TABLE(hdr));
2203 
2204         if (HDR_HAS_L2HDR(hdr)) {
2205                 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2206                 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
2207 
2208                 if (!buflist_held)
2209                         mutex_enter(&dev->l2ad_mtx);
2210 
2211                 /*
2212                  * Even though we checked this conditional above, we
2213                  * need to check this again now that we have the
2214                  * l2ad_mtx. This is because we could be racing with
2215                  * another thread calling l2arc_evict() which might have
2216                  * destroyed this header's L2 portion as we were waiting
2217                  * to acquire the l2ad_mtx. If that happens, we don't
2218                  * want to re-destroy the header's L2 portion.
2219                  */
2220                 if (HDR_HAS_L2HDR(hdr))
2221                         arc_hdr_l2hdr_destroy(hdr);
2222 
2223                 if (!buflist_held)
2224                         mutex_exit(&dev->l2ad_mtx);
2225         }
2226 
2227         if (!BUF_EMPTY(hdr))
2228                 buf_discard_identity(hdr);
2229 
2230         if (hdr->b_freeze_cksum != NULL) {
2231                 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
2232                 hdr->b_freeze_cksum = NULL;
2233         }
2234 
2235         if (HDR_HAS_L1HDR(hdr)) {
2236                 while (hdr->b_l1hdr.b_buf) {
2237                         arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2238 
2239                         if (buf->b_efunc != NULL) {
2240                                 mutex_enter(&arc_user_evicts_lock);
2241                                 mutex_enter(&buf->b_evict_lock);
2242                                 ASSERT(buf->b_hdr != NULL);
2243                                 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE);
2244                                 hdr->b_l1hdr.b_buf = buf->b_next;
2245                                 buf->b_hdr = &arc_eviction_hdr;
2246                                 buf->b_next = arc_eviction_list;
2247                                 arc_eviction_list = buf;
2248                                 mutex_exit(&buf->b_evict_lock);
2249                                 cv_signal(&arc_user_evicts_cv);
2250                                 mutex_exit(&arc_user_evicts_lock);
2251                         } else {
2252                                 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE);
2253                         }
2254                 }
2255 #ifdef ZFS_DEBUG
2256                 if (hdr->b_l1hdr.b_thawed != NULL) {
2257                         kmem_free(hdr->b_l1hdr.b_thawed, 1);
2258                         hdr->b_l1hdr.b_thawed = NULL;
2259                 }
2260 #endif
2261         }
2262 
2263         ASSERT3P(hdr->b_hash_next, ==, NULL);
2264         if (HDR_HAS_L1HDR(hdr)) {
2265                 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
2266                 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
2267                 kmem_cache_free(hdr_full_cache, hdr);
2268         } else {
2269                 kmem_cache_free(hdr_l2only_cache, hdr);
2270         }
2271 }
2272 
2273 void
2274 arc_buf_free(arc_buf_t *buf, void *tag)
2275 {
2276         arc_buf_hdr_t *hdr = buf->b_hdr;
2277         int hashed = hdr->b_l1hdr.b_state != arc_anon;
2278 
2279         ASSERT(buf->b_efunc == NULL);
2280         ASSERT(buf->b_data != NULL);
2281 
2282         if (hashed) {
2283                 kmutex_t *hash_lock = HDR_LOCK(hdr);
2284 
2285                 mutex_enter(hash_lock);
2286                 hdr = buf->b_hdr;
2287                 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2288 
2289                 (void) remove_reference(hdr, hash_lock, tag);
2290                 if (hdr->b_l1hdr.b_datacnt > 1) {
2291                         arc_buf_destroy(buf, TRUE);
2292                 } else {
2293                         ASSERT(buf == hdr->b_l1hdr.b_buf);
2294                         ASSERT(buf->b_efunc == NULL);
2295                         hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2296                 }
2297                 mutex_exit(hash_lock);
2298         } else if (HDR_IO_IN_PROGRESS(hdr)) {
2299                 int destroy_hdr;
2300                 /*
2301                  * We are in the middle of an async write.  Don't destroy
2302                  * this buffer unless the write completes before we finish
2303                  * decrementing the reference count.
2304                  */
2305                 mutex_enter(&arc_user_evicts_lock);
2306                 (void) remove_reference(hdr, NULL, tag);
2307                 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2308                 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
2309                 mutex_exit(&arc_user_evicts_lock);
2310                 if (destroy_hdr)
2311                         arc_hdr_destroy(hdr);
2312         } else {
2313                 if (remove_reference(hdr, NULL, tag) > 0)
2314                         arc_buf_destroy(buf, TRUE);
2315                 else
2316                         arc_hdr_destroy(hdr);
2317         }
2318 }
2319 
2320 boolean_t
2321 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
2322 {
2323         arc_buf_hdr_t *hdr = buf->b_hdr;
2324         kmutex_t *hash_lock = HDR_LOCK(hdr);
2325         boolean_t no_callback = (buf->b_efunc == NULL);
2326 
2327         if (hdr->b_l1hdr.b_state == arc_anon) {
2328                 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
2329                 arc_buf_free(buf, tag);
2330                 return (no_callback);
2331         }
2332 
2333         mutex_enter(hash_lock);
2334         hdr = buf->b_hdr;
2335         ASSERT(hdr->b_l1hdr.b_datacnt > 0);
2336         ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2337         ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2338         ASSERT(buf->b_data != NULL);
2339 
2340         (void) remove_reference(hdr, hash_lock, tag);
2341         if (hdr->b_l1hdr.b_datacnt > 1) {
2342                 if (no_callback)
2343                         arc_buf_destroy(buf, TRUE);
2344         } else if (no_callback) {
2345                 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL);
2346                 ASSERT(buf->b_efunc == NULL);
2347                 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2348         }
2349         ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 ||
2350             refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2351         mutex_exit(hash_lock);
2352         return (no_callback);
2353 }
2354 
2355 int32_t
2356 arc_buf_size(arc_buf_t *buf)
2357 {
2358         return (buf->b_hdr->b_size);
2359 }
2360 
2361 /*
2362  * Called from the DMU to determine if the current buffer should be
2363  * evicted. In order to ensure proper locking, the eviction must be initiated
2364  * from the DMU. Return true if the buffer is associated with user data and
2365  * duplicate buffers still exist.
2366  */
2367 boolean_t
2368 arc_buf_eviction_needed(arc_buf_t *buf)
2369 {
2370         arc_buf_hdr_t *hdr;
2371         boolean_t evict_needed = B_FALSE;
2372 
2373         if (zfs_disable_dup_eviction)
2374                 return (B_FALSE);
2375 
2376         mutex_enter(&buf->b_evict_lock);
2377         hdr = buf->b_hdr;
2378         if (hdr == NULL) {
2379                 /*
2380                  * We are in arc_do_user_evicts(); let that function
2381                  * perform the eviction.
2382                  */
2383                 ASSERT(buf->b_data == NULL);
2384                 mutex_exit(&buf->b_evict_lock);
2385                 return (B_FALSE);
2386         } else if (buf->b_data == NULL) {
2387                 /*
2388                  * We have already been added to the arc eviction list;
2389                  * recommend eviction.
2390                  */
2391                 ASSERT3P(hdr, ==, &arc_eviction_hdr);
2392                 mutex_exit(&buf->b_evict_lock);
2393                 return (B_TRUE);
2394         }
2395 
2396         if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr))
2397                 evict_needed = B_TRUE;
2398 
2399         mutex_exit(&buf->b_evict_lock);
2400         return (evict_needed);
2401 }
2402 
2403 /*
2404  * Evict the arc_buf_hdr that is provided as a parameter. The resultant
2405  * state of the header is dependent on it's state prior to entering this
2406  * function. The following transitions are possible:
2407  *
2408  *    - arc_mru -> arc_mru_ghost
2409  *    - arc_mfu -> arc_mfu_ghost
2410  *    - arc_mru_ghost -> arc_l2c_only
2411  *    - arc_mru_ghost -> deleted
2412  *    - arc_mfu_ghost -> arc_l2c_only
2413  *    - arc_mfu_ghost -> deleted
2414  */
2415 static int64_t
2416 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
2417 {
2418         arc_state_t *evicted_state, *state;
2419         int64_t bytes_evicted = 0;
2420 
2421         ASSERT(MUTEX_HELD(hash_lock));
2422         ASSERT(HDR_HAS_L1HDR(hdr));
2423 
2424         state = hdr->b_l1hdr.b_state;
2425         if (GHOST_STATE(state)) {
2426                 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2427                 ASSERT(hdr->b_l1hdr.b_buf == NULL);
2428 
2429                 /*
2430                  * l2arc_write_buffers() relies on a header's L1 portion
2431                  * (i.e. it's b_tmp_cdata field) during it's write phase.
2432                  * Thus, we cannot push a header onto the arc_l2c_only
2433                  * state (removing it's L1 piece) until the header is
2434                  * done being written to the l2arc.
2435                  */
2436                 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
2437                         ARCSTAT_BUMP(arcstat_evict_l2_skip);
2438                         return (bytes_evicted);
2439                 }
2440 
2441                 ARCSTAT_BUMP(arcstat_deleted);
2442                 bytes_evicted += hdr->b_size;
2443 
2444                 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
2445 
2446                 if (HDR_HAS_L2HDR(hdr)) {
2447                         /*
2448                          * This buffer is cached on the 2nd Level ARC;
2449                          * don't destroy the header.
2450                          */
2451                         arc_change_state(arc_l2c_only, hdr, hash_lock);
2452                         /*
2453                          * dropping from L1+L2 cached to L2-only,
2454                          * realloc to remove the L1 header.
2455                          */
2456                         hdr = arc_hdr_realloc(hdr, hdr_full_cache,
2457                             hdr_l2only_cache);
2458                 } else {
2459                         arc_change_state(arc_anon, hdr, hash_lock);
2460                         arc_hdr_destroy(hdr);
2461                 }
2462                 return (bytes_evicted);
2463         }
2464 
2465         ASSERT(state == arc_mru || state == arc_mfu);
2466         evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
2467 
2468         /* prefetch buffers have a minimum lifespan */
2469         if (HDR_IO_IN_PROGRESS(hdr) ||
2470             ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
2471             ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
2472             arc_min_prefetch_lifespan)) {
2473                 ARCSTAT_BUMP(arcstat_evict_skip);
2474                 return (bytes_evicted);
2475         }
2476 
2477         ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
2478         ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0);
2479         while (hdr->b_l1hdr.b_buf) {
2480                 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2481                 if (!mutex_tryenter(&buf->b_evict_lock)) {
2482                         ARCSTAT_BUMP(arcstat_mutex_miss);
2483                         break;
2484                 }
2485                 if (buf->b_data != NULL)
2486                         bytes_evicted += hdr->b_size;
2487                 if (buf->b_efunc != NULL) {
2488                         mutex_enter(&arc_user_evicts_lock);
2489                         arc_buf_destroy(buf, FALSE);
2490                         hdr->b_l1hdr.b_buf = buf->b_next;
2491                         buf->b_hdr = &arc_eviction_hdr;
2492                         buf->b_next = arc_eviction_list;
2493                         arc_eviction_list = buf;
2494                         cv_signal(&arc_user_evicts_cv);
2495                         mutex_exit(&arc_user_evicts_lock);
2496                         mutex_exit(&buf->b_evict_lock);
2497                 } else {
2498                         mutex_exit(&buf->b_evict_lock);
2499                         arc_buf_destroy(buf, TRUE);
2500                 }
2501         }
2502 
2503         if (HDR_HAS_L2HDR(hdr)) {
2504                 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size);
2505         } else {
2506                 if (l2arc_write_eligible(hdr->b_spa, hdr))
2507                         ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size);
2508                 else
2509                         ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size);
2510         }
2511 
2512         if (hdr->b_l1hdr.b_datacnt == 0) {
2513                 arc_change_state(evicted_state, hdr, hash_lock);
2514                 ASSERT(HDR_IN_HASH_TABLE(hdr));
2515                 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
2516                 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
2517                 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
2518         }
2519 
2520         return (bytes_evicted);
2521 }
2522 
2523 static uint64_t
2524 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
2525     uint64_t spa, int64_t bytes)
2526 {
2527         multilist_sublist_t *mls;
2528         uint64_t bytes_evicted = 0;
2529         arc_buf_hdr_t *hdr;
2530         kmutex_t *hash_lock;
2531         int evict_count = 0;
2532 
2533         ASSERT3P(marker, !=, NULL);
2534         IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2535 
2536         mls = multilist_sublist_lock(ml, idx);
2537 
2538         for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
2539             hdr = multilist_sublist_prev(mls, marker)) {
2540                 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
2541                     (evict_count >= zfs_arc_evict_batch_limit))
2542                         break;
2543 
2544                 /*
2545                  * To keep our iteration location, move the marker
2546                  * forward. Since we're not holding hdr's hash lock, we
2547                  * must be very careful and not remove 'hdr' from the
2548                  * sublist. Otherwise, other consumers might mistake the
2549                  * 'hdr' as not being on a sublist when they call the
2550                  * multilist_link_active() function (they all rely on
2551                  * the hash lock protecting concurrent insertions and
2552                  * removals). multilist_sublist_move_forward() was
2553                  * specifically implemented to ensure this is the case
2554                  * (only 'marker' will be removed and re-inserted).
2555                  */
2556                 multilist_sublist_move_forward(mls, marker);
2557 
2558                 /*
2559                  * The only case where the b_spa field should ever be
2560                  * zero, is the marker headers inserted by
2561                  * arc_evict_state(). It's possible for multiple threads
2562                  * to be calling arc_evict_state() concurrently (e.g.
2563                  * dsl_pool_close() and zio_inject_fault()), so we must
2564                  * skip any markers we see from these other threads.
2565                  */
2566                 if (hdr->b_spa == 0)
2567                         continue;
2568 
2569                 /* we're only interested in evicting buffers of a certain spa */
2570                 if (spa != 0 && hdr->b_spa != spa) {
2571                         ARCSTAT_BUMP(arcstat_evict_skip);
2572                         continue;
2573                 }
2574 
2575                 hash_lock = HDR_LOCK(hdr);
2576 
2577                 /*
2578                  * We aren't calling this function from any code path
2579                  * that would already be holding a hash lock, so we're
2580                  * asserting on this assumption to be defensive in case
2581                  * this ever changes. Without this check, it would be
2582                  * possible to incorrectly increment arcstat_mutex_miss
2583                  * below (e.g. if the code changed such that we called
2584                  * this function with a hash lock held).
2585                  */
2586                 ASSERT(!MUTEX_HELD(hash_lock));
2587 
2588                 if (mutex_tryenter(hash_lock)) {
2589                         uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
2590                         mutex_exit(hash_lock);
2591 
2592                         bytes_evicted += evicted;
2593 
2594                         /*
2595                          * If evicted is zero, arc_evict_hdr() must have
2596                          * decided to skip this header, don't increment
2597                          * evict_count in this case.
2598                          */
2599                         if (evicted != 0)
2600                                 evict_count++;
2601 
2602                         /*
2603                          * If arc_size isn't overflowing, signal any
2604                          * threads that might happen to be waiting.
2605                          *
2606                          * For each header evicted, we wake up a single
2607                          * thread. If we used cv_broadcast, we could
2608                          * wake up "too many" threads causing arc_size
2609                          * to significantly overflow arc_c; since
2610                          * arc_get_data_buf() doesn't check for overflow
2611                          * when it's woken up (it doesn't because it's
2612                          * possible for the ARC to be overflowing while
2613                          * full of un-evictable buffers, and the
2614                          * function should proceed in this case).
2615                          *
2616                          * If threads are left sleeping, due to not
2617                          * using cv_broadcast, they will be woken up
2618                          * just before arc_reclaim_thread() sleeps.
2619                          */
2620                         mutex_enter(&arc_reclaim_lock);
2621                         if (!arc_is_overflowing())
2622                                 cv_signal(&arc_reclaim_waiters_cv);
2623                         mutex_exit(&arc_reclaim_lock);
2624                 } else {
2625                         ARCSTAT_BUMP(arcstat_mutex_miss);
2626                 }
2627         }
2628 
2629         multilist_sublist_unlock(mls);
2630 
2631         return (bytes_evicted);
2632 }
2633 
2634 /*
2635  * Evict buffers from the given arc state, until we've removed the
2636  * specified number of bytes. Move the removed buffers to the
2637  * appropriate evict state.
2638  *
2639  * This function makes a "best effort". It skips over any buffers
2640  * it can't get a hash_lock on, and so, may not catch all candidates.
2641  * It may also return without evicting as much space as requested.
2642  *
2643  * If bytes is specified using the special value ARC_EVICT_ALL, this
2644  * will evict all available (i.e. unlocked and evictable) buffers from
2645  * the given arc state; which is used by arc_flush().
2646  */
2647 static uint64_t
2648 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
2649     arc_buf_contents_t type)
2650 {
2651         uint64_t total_evicted = 0;
2652         multilist_t *ml = &state->arcs_list[type];
2653         int num_sublists;
2654         arc_buf_hdr_t **markers;
2655 
2656         IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2657 
2658         num_sublists = multilist_get_num_sublists(ml);
2659 
2660         /*
2661          * If we've tried to evict from each sublist, made some
2662          * progress, but still have not hit the target number of bytes
2663          * to evict, we want to keep trying. The markers allow us to
2664          * pick up where we left off for each individual sublist, rather
2665          * than starting from the tail each time.
2666          */
2667         markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
2668         for (int i = 0; i < num_sublists; i++) {
2669                 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
2670 
2671                 /*
2672                  * A b_spa of 0 is used to indicate that this header is
2673                  * a marker. This fact is used in arc_adjust_type() and
2674                  * arc_evict_state_impl().
2675                  */
2676                 markers[i]->b_spa = 0;
2677 
2678                 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2679                 multilist_sublist_insert_tail(mls, markers[i]);
2680                 multilist_sublist_unlock(mls);
2681         }
2682 
2683         /*
2684          * While we haven't hit our target number of bytes to evict, or
2685          * we're evicting all available buffers.
2686          */
2687         while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
2688                 /*
2689                  * Start eviction using a randomly selected sublist,
2690                  * this is to try and evenly balance eviction across all
2691                  * sublists. Always starting at the same sublist
2692                  * (e.g. index 0) would cause evictions to favor certain
2693                  * sublists over others.
2694                  */
2695                 int sublist_idx = multilist_get_random_index(ml);
2696                 uint64_t scan_evicted = 0;
2697 
2698                 for (int i = 0; i < num_sublists; i++) {
2699                         uint64_t bytes_remaining;
2700                         uint64_t bytes_evicted;
2701 
2702                         if (bytes == ARC_EVICT_ALL)
2703                                 bytes_remaining = ARC_EVICT_ALL;
2704                         else if (total_evicted < bytes)
2705                                 bytes_remaining = bytes - total_evicted;
2706                         else
2707                                 break;
2708 
2709                         bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
2710                             markers[sublist_idx], spa, bytes_remaining);
2711 
2712                         scan_evicted += bytes_evicted;
2713                         total_evicted += bytes_evicted;
2714 
2715                         /* we've reached the end, wrap to the beginning */
2716                         if (++sublist_idx >= num_sublists)
2717                                 sublist_idx = 0;
2718                 }
2719 
2720                 /*
2721                  * If we didn't evict anything during this scan, we have
2722                  * no reason to believe we'll evict more during another
2723                  * scan, so break the loop.
2724                  */
2725                 if (scan_evicted == 0) {
2726                         /* This isn't possible, let's make that obvious */
2727                         ASSERT3S(bytes, !=, 0);
2728 
2729                         /*
2730                          * When bytes is ARC_EVICT_ALL, the only way to
2731                          * break the loop is when scan_evicted is zero.
2732                          * In that case, we actually have evicted enough,
2733                          * so we don't want to increment the kstat.
2734                          */
2735                         if (bytes != ARC_EVICT_ALL) {
2736                                 ASSERT3S(total_evicted, <, bytes);
2737                                 ARCSTAT_BUMP(arcstat_evict_not_enough);
2738                         }
2739 
2740                         break;
2741                 }
2742         }
2743 
2744         for (int i = 0; i < num_sublists; i++) {
2745                 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2746                 multilist_sublist_remove(mls, markers[i]);
2747                 multilist_sublist_unlock(mls);
2748 
2749                 kmem_cache_free(hdr_full_cache, markers[i]);
2750         }
2751         kmem_free(markers, sizeof (*markers) * num_sublists);
2752 
2753         return (total_evicted);
2754 }
2755 
2756 /*
2757  * Flush all "evictable" data of the given type from the arc state
2758  * specified. This will not evict any "active" buffers (i.e. referenced).
2759  *
2760  * When 'retry' is set to FALSE, the function will make a single pass
2761  * over the state and evict any buffers that it can. Since it doesn't
2762  * continually retry the eviction, it might end up leaving some buffers
2763  * in the ARC due to lock misses.
2764  *
2765  * When 'retry' is set to TRUE, the function will continually retry the
2766  * eviction until *all* evictable buffers have been removed from the
2767  * state. As a result, if concurrent insertions into the state are
2768  * allowed (e.g. if the ARC isn't shutting down), this function might
2769  * wind up in an infinite loop, continually trying to evict buffers.
2770  */
2771 static uint64_t
2772 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
2773     boolean_t retry)
2774 {
2775         uint64_t evicted = 0;
2776 
2777         while (state->arcs_lsize[type] != 0) {
2778                 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
2779 
2780                 if (!retry)
2781                         break;
2782         }
2783 
2784         return (evicted);
2785 }
2786 
2787 /*
2788  * Evict the specified number of bytes from the state specified,
2789  * restricting eviction to the spa and type given. This function
2790  * prevents us from trying to evict more from a state's list than
2791  * is "evictable", and to skip evicting altogether when passed a
2792  * negative value for "bytes". In contrast, arc_evict_state() will
2793  * evict everything it can, when passed a negative value for "bytes".
2794  */
2795 static uint64_t
2796 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
2797     arc_buf_contents_t type)
2798 {
2799         int64_t delta;
2800 
2801         if (bytes > 0 && state->arcs_lsize[type] > 0) {
2802                 delta = MIN(state->arcs_lsize[type], bytes);
2803                 return (arc_evict_state(state, spa, delta, type));
2804         }
2805 
2806         return (0);
2807 }
2808 
2809 /*
2810  * Evict metadata buffers from the cache, such that arc_meta_used is
2811  * capped by the arc_meta_limit tunable.
2812  */
2813 static uint64_t
2814 arc_adjust_meta(void)
2815 {
2816         uint64_t total_evicted = 0;
2817         int64_t target;
2818 
2819         /*
2820          * If we're over the meta limit, we want to evict enough
2821          * metadata to get back under the meta limit. We don't want to
2822          * evict so much that we drop the MRU below arc_p, though. If
2823          * we're over the meta limit more than we're over arc_p, we
2824          * evict some from the MRU here, and some from the MFU below.
2825          */
2826         target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
2827             (int64_t)(refcount_count(&arc_anon->arcs_size) +
2828             refcount_count(&arc_mru->arcs_size) - arc_p));
2829 
2830         total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
2831 
2832         /*
2833          * Similar to the above, we want to evict enough bytes to get us
2834          * below the meta limit, but not so much as to drop us below the
2835          * space alloted to the MFU (which is defined as arc_c - arc_p).
2836          */
2837         target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
2838             (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
2839 
2840         total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
2841 
2842         return (total_evicted);
2843 }
2844 
2845 /*
2846  * Return the type of the oldest buffer in the given arc state
2847  *
2848  * This function will select a random sublist of type ARC_BUFC_DATA and
2849  * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
2850  * is compared, and the type which contains the "older" buffer will be
2851  * returned.
2852  */
2853 static arc_buf_contents_t
2854 arc_adjust_type(arc_state_t *state)
2855 {
2856         multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
2857         multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
2858         int data_idx = multilist_get_random_index(data_ml);
2859         int meta_idx = multilist_get_random_index(meta_ml);
2860         multilist_sublist_t *data_mls;
2861         multilist_sublist_t *meta_mls;
2862         arc_buf_contents_t type;
2863         arc_buf_hdr_t *data_hdr;
2864         arc_buf_hdr_t *meta_hdr;
2865 
2866         /*
2867          * We keep the sublist lock until we're finished, to prevent
2868          * the headers from being destroyed via arc_evict_state().
2869          */
2870         data_mls = multilist_sublist_lock(data_ml, data_idx);
2871         meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
2872 
2873         /*
2874          * These two loops are to ensure we skip any markers that
2875          * might be at the tail of the lists due to arc_evict_state().
2876          */
2877 
2878         for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
2879             data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
2880                 if (data_hdr->b_spa != 0)
2881                         break;
2882         }
2883 
2884         for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
2885             meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
2886                 if (meta_hdr->b_spa != 0)
2887                         break;
2888         }
2889 
2890         if (data_hdr == NULL && meta_hdr == NULL) {
2891                 type = ARC_BUFC_DATA;
2892         } else if (data_hdr == NULL) {
2893                 ASSERT3P(meta_hdr, !=, NULL);
2894                 type = ARC_BUFC_METADATA;
2895         } else if (meta_hdr == NULL) {
2896                 ASSERT3P(data_hdr, !=, NULL);
2897                 type = ARC_BUFC_DATA;
2898         } else {
2899                 ASSERT3P(data_hdr, !=, NULL);
2900                 ASSERT3P(meta_hdr, !=, NULL);
2901 
2902                 /* The headers can't be on the sublist without an L1 header */
2903                 ASSERT(HDR_HAS_L1HDR(data_hdr));
2904                 ASSERT(HDR_HAS_L1HDR(meta_hdr));
2905 
2906                 if (data_hdr->b_l1hdr.b_arc_access <
2907                     meta_hdr->b_l1hdr.b_arc_access) {
2908                         type = ARC_BUFC_DATA;
2909                 } else {
2910                         type = ARC_BUFC_METADATA;
2911                 }
2912         }
2913 
2914         multilist_sublist_unlock(meta_mls);
2915         multilist_sublist_unlock(data_mls);
2916 
2917         return (type);
2918 }
2919 
2920 /*
2921  * Evict buffers from the cache, such that arc_size is capped by arc_c.
2922  */
2923 static uint64_t
2924 arc_adjust(void)
2925 {
2926         uint64_t total_evicted = 0;
2927         uint64_t bytes;
2928         int64_t target;
2929 
2930         /*
2931          * If we're over arc_meta_limit, we want to correct that before
2932          * potentially evicting data buffers below.
2933          */
2934         total_evicted += arc_adjust_meta();
2935 
2936         /*
2937          * Adjust MRU size
2938          *
2939          * If we're over the target cache size, we want to evict enough
2940          * from the list to get back to our target size. We don't want
2941          * to evict too much from the MRU, such that it drops below
2942          * arc_p. So, if we're over our target cache size more than
2943          * the MRU is over arc_p, we'll evict enough to get back to
2944          * arc_p here, and then evict more from the MFU below.
2945          */
2946         target = MIN((int64_t)(arc_size - arc_c),
2947             (int64_t)(refcount_count(&arc_anon->arcs_size) +
2948             refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
2949 
2950         /*
2951          * If we're below arc_meta_min, always prefer to evict data.
2952          * Otherwise, try to satisfy the requested number of bytes to
2953          * evict from the type which contains older buffers; in an
2954          * effort to keep newer buffers in the cache regardless of their
2955          * type. If we cannot satisfy the number of bytes from this
2956          * type, spill over into the next type.
2957          */
2958         if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
2959             arc_meta_used > arc_meta_min) {
2960                 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
2961                 total_evicted += bytes;
2962 
2963                 /*
2964                  * If we couldn't evict our target number of bytes from
2965                  * metadata, we try to get the rest from data.
2966                  */
2967                 target -= bytes;
2968 
2969                 total_evicted +=
2970                     arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
2971         } else {
2972                 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
2973                 total_evicted += bytes;
2974 
2975                 /*
2976                  * If we couldn't evict our target number of bytes from
2977                  * data, we try to get the rest from metadata.
2978                  */
2979                 target -= bytes;
2980 
2981                 total_evicted +=
2982                     arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
2983         }
2984 
2985         /*
2986          * Adjust MFU size
2987          *
2988          * Now that we've tried to evict enough from the MRU to get its
2989          * size back to arc_p, if we're still above the target cache
2990          * size, we evict the rest from the MFU.
2991          */
2992         target = arc_size - arc_c;
2993 
2994         if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
2995             arc_meta_used > arc_meta_min) {
2996                 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
2997                 total_evicted += bytes;
2998 
2999                 /*
3000                  * If we couldn't evict our target number of bytes from
3001                  * metadata, we try to get the rest from data.
3002                  */
3003                 target -= bytes;
3004 
3005                 total_evicted +=
3006                     arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3007         } else {
3008                 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3009                 total_evicted += bytes;
3010 
3011                 /*
3012                  * If we couldn't evict our target number of bytes from
3013                  * data, we try to get the rest from data.
3014                  */
3015                 target -= bytes;
3016 
3017                 total_evicted +=
3018                     arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3019         }
3020 
3021         /*
3022          * Adjust ghost lists
3023          *
3024          * In addition to the above, the ARC also defines target values
3025          * for the ghost lists. The sum of the mru list and mru ghost
3026          * list should never exceed the target size of the cache, and
3027          * the sum of the mru list, mfu list, mru ghost list, and mfu
3028          * ghost list should never exceed twice the target size of the
3029          * cache. The following logic enforces these limits on the ghost
3030          * caches, and evicts from them as needed.
3031          */
3032         target = refcount_count(&arc_mru->arcs_size) +
3033             refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3034 
3035         bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3036         total_evicted += bytes;
3037 
3038         target -= bytes;
3039 
3040         total_evicted +=
3041             arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3042 
3043         /*
3044          * We assume the sum of the mru list and mfu list is less than
3045          * or equal to arc_c (we enforced this above), which means we
3046          * can use the simpler of the two equations below:
3047          *
3048          *      mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3049          *                  mru ghost + mfu ghost <= arc_c
3050          */
3051         target = refcount_count(&arc_mru_ghost->arcs_size) +
3052             refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3053 
3054         bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3055         total_evicted += bytes;
3056 
3057         target -= bytes;
3058 
3059         total_evicted +=
3060             arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3061 
3062         return (total_evicted);
3063 }
3064 
3065 static void
3066 arc_do_user_evicts(void)
3067 {
3068         mutex_enter(&arc_user_evicts_lock);
3069         while (arc_eviction_list != NULL) {
3070                 arc_buf_t *buf = arc_eviction_list;
3071                 arc_eviction_list = buf->b_next;
3072                 mutex_enter(&buf->b_evict_lock);
3073                 buf->b_hdr = NULL;
3074                 mutex_exit(&buf->b_evict_lock);
3075                 mutex_exit(&arc_user_evicts_lock);
3076 
3077                 if (buf->b_efunc != NULL)
3078                         VERIFY0(buf->b_efunc(buf->b_private));
3079 
3080                 buf->b_efunc = NULL;
3081                 buf->b_private = NULL;
3082                 kmem_cache_free(buf_cache, buf);
3083                 mutex_enter(&arc_user_evicts_lock);
3084         }
3085         mutex_exit(&arc_user_evicts_lock);
3086 }
3087 
3088 void
3089 arc_flush(spa_t *spa, boolean_t retry)
3090 {
3091         uint64_t guid = 0;
3092 
3093         /*
3094          * If retry is TRUE, a spa must not be specified since we have
3095          * no good way to determine if all of a spa's buffers have been
3096          * evicted from an arc state.
3097          */
3098         ASSERT(!retry || spa == 0);
3099 
3100         if (spa != NULL)
3101                 guid = spa_load_guid(spa);
3102 
3103         (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3104         (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3105 
3106         (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3107         (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3108 
3109         (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3110         (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3111 
3112         (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3113         (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3114 
3115         arc_do_user_evicts();
3116         ASSERT(spa || arc_eviction_list == NULL);
3117 }
3118 
3119 void
3120 arc_shrink(int64_t to_free)
3121 {
3122         if (arc_c > arc_c_min) {
3123 
3124                 if (arc_c > arc_c_min + to_free)
3125                         atomic_add_64(&arc_c, -to_free);
3126                 else
3127                         arc_c = arc_c_min;
3128 
3129                 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3130                 if (arc_c > arc_size)
3131                         arc_c = MAX(arc_size, arc_c_min);
3132                 if (arc_p > arc_c)
3133                         arc_p = (arc_c >> 1);
3134                 ASSERT(arc_c >= arc_c_min);
3135                 ASSERT((int64_t)arc_p >= 0);
3136         }
3137 
3138         if (arc_size > arc_c)
3139                 (void) arc_adjust();
3140 }
3141 
3142 typedef enum free_memory_reason_t {
3143         FMR_UNKNOWN,
3144         FMR_NEEDFREE,
3145         FMR_LOTSFREE,
3146         FMR_SWAPFS_MINFREE,
3147         FMR_PAGES_PP_MAXIMUM,
3148         FMR_HEAP_ARENA,
3149         FMR_ZIO_ARENA,
3150 } free_memory_reason_t;
3151 
3152 int64_t last_free_memory;
3153 free_memory_reason_t last_free_reason;
3154 
3155 /*
3156  * Additional reserve of pages for pp_reserve.
3157  */
3158 int64_t arc_pages_pp_reserve = 64;
3159 
3160 /*
3161  * Additional reserve of pages for swapfs.
3162  */
3163 int64_t arc_swapfs_reserve = 64;
3164 
3165 /*
3166  * Return the amount of memory that can be consumed before reclaim will be
3167  * needed.  Positive if there is sufficient free memory, negative indicates
3168  * the amount of memory that needs to be freed up.
3169  */
3170 static int64_t
3171 arc_available_memory(void)
3172 {
3173         int64_t lowest = INT64_MAX;
3174         int64_t n;
3175         free_memory_reason_t r = FMR_UNKNOWN;
3176 
3177 #ifdef _KERNEL
3178         if (needfree > 0) {
3179                 n = PAGESIZE * (-needfree);
3180                 if (n < lowest) {
3181                         lowest = n;
3182                         r = FMR_NEEDFREE;
3183                 }
3184         }
3185 
3186         /*
3187          * check that we're out of range of the pageout scanner.  It starts to
3188          * schedule paging if freemem is less than lotsfree and needfree.
3189          * lotsfree is the high-water mark for pageout, and needfree is the
3190          * number of needed free pages.  We add extra pages here to make sure
3191          * the scanner doesn't start up while we're freeing memory.
3192          */
3193         n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
3194         if (n < lowest) {
3195                 lowest = n;
3196                 r = FMR_LOTSFREE;
3197         }
3198 
3199         /*
3200          * check to make sure that swapfs has enough space so that anon
3201          * reservations can still succeed. anon_resvmem() checks that the
3202          * availrmem is greater than swapfs_minfree, and the number of reserved
3203          * swap pages.  We also add a bit of extra here just to prevent
3204          * circumstances from getting really dire.
3205          */
3206         n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
3207             desfree - arc_swapfs_reserve);
3208         if (n < lowest) {
3209                 lowest = n;
3210                 r = FMR_SWAPFS_MINFREE;
3211         }
3212 
3213 
3214         /*
3215          * Check that we have enough availrmem that memory locking (e.g., via
3216          * mlock(3C) or memcntl(2)) can still succeed.  (pages_pp_maximum
3217          * stores the number of pages that cannot be locked; when availrmem
3218          * drops below pages_pp_maximum, page locking mechanisms such as
3219          * page_pp_lock() will fail.)
3220          */
3221         n = PAGESIZE * (availrmem - pages_pp_maximum -
3222             arc_pages_pp_reserve);
3223         if (n < lowest) {
3224                 lowest = n;
3225                 r = FMR_PAGES_PP_MAXIMUM;
3226         }
3227 
3228 #if defined(__i386)
3229         /*
3230          * If we're on an i386 platform, it's possible that we'll exhaust the
3231          * kernel heap space before we ever run out of available physical
3232          * memory.  Most checks of the size of the heap_area compare against
3233          * tune.t_minarmem, which is the minimum available real memory that we
3234          * can have in the system.  However, this is generally fixed at 25 pages
3235          * which is so low that it's useless.  In this comparison, we seek to
3236          * calculate the total heap-size, and reclaim if more than 3/4ths of the
3237          * heap is allocated.  (Or, in the calculation, if less than 1/4th is
3238          * free)
3239          */
3240         n = vmem_size(heap_arena, VMEM_FREE) -
3241             (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
3242         if (n < lowest) {
3243                 lowest = n;
3244                 r = FMR_HEAP_ARENA;
3245         }
3246 #endif
3247 
3248         /*
3249          * If zio data pages are being allocated out of a separate heap segment,
3250          * then enforce that the size of available vmem for this arena remains
3251          * above about 1/16th free.
3252          *
3253          * Note: The 1/16th arena free requirement was put in place
3254          * to aggressively evict memory from the arc in order to avoid
3255          * memory fragmentation issues.
3256          */
3257         if (zio_arena != NULL) {
3258                 n = vmem_size(zio_arena, VMEM_FREE) -
3259                     (vmem_size(zio_arena, VMEM_ALLOC) >> 4);
3260                 if (n < lowest) {
3261                         lowest = n;
3262                         r = FMR_ZIO_ARENA;
3263                 }
3264         }
3265 #else
3266         /* Every 100 calls, free a small amount */
3267         if (spa_get_random(100) == 0)
3268                 lowest = -1024;
3269 #endif
3270 
3271         last_free_memory = lowest;
3272         last_free_reason = r;
3273 
3274         return (lowest);
3275 }
3276 
3277 
3278 /*
3279  * Determine if the system is under memory pressure and is asking
3280  * to reclaim memory. A return value of TRUE indicates that the system
3281  * is under memory pressure and that the arc should adjust accordingly.
3282  */
3283 static boolean_t
3284 arc_reclaim_needed(void)
3285 {
3286         return (arc_available_memory() < 0);
3287 }
3288 
3289 static void
3290 arc_kmem_reap_now(void)
3291 {
3292         size_t                  i;
3293         kmem_cache_t            *prev_cache = NULL;
3294         kmem_cache_t            *prev_data_cache = NULL;
3295         extern kmem_cache_t     *zio_buf_cache[];
3296         extern kmem_cache_t     *zio_data_buf_cache[];
3297         extern kmem_cache_t     *range_seg_cache;
3298 
3299 #ifdef _KERNEL
3300         if (arc_meta_used >= arc_meta_limit) {
3301                 /*
3302                  * We are exceeding our meta-data cache limit.
3303                  * Purge some DNLC entries to release holds on meta-data.
3304                  */
3305                 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
3306         }
3307 #if defined(__i386)
3308         /*
3309          * Reclaim unused memory from all kmem caches.
3310          */
3311         kmem_reap();
3312 #endif
3313 #endif
3314 
3315         for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
3316                 if (zio_buf_cache[i] != prev_cache) {
3317                         prev_cache = zio_buf_cache[i];
3318                         kmem_cache_reap_now(zio_buf_cache[i]);
3319                 }
3320                 if (zio_data_buf_cache[i] != prev_data_cache) {
3321                         prev_data_cache = zio_data_buf_cache[i];
3322                         kmem_cache_reap_now(zio_data_buf_cache[i]);
3323                 }
3324         }
3325         kmem_cache_reap_now(buf_cache);
3326         kmem_cache_reap_now(hdr_full_cache);
3327         kmem_cache_reap_now(hdr_l2only_cache);
3328         kmem_cache_reap_now(range_seg_cache);
3329 
3330         if (zio_arena != NULL) {
3331                 /*
3332                  * Ask the vmem arena to reclaim unused memory from its
3333                  * quantum caches.
3334                  */
3335                 vmem_qcache_reap(zio_arena);
3336         }
3337 }
3338 
3339 /*
3340  * Threads can block in arc_get_data_buf() waiting for this thread to evict
3341  * enough data and signal them to proceed. When this happens, the threads in
3342  * arc_get_data_buf() are sleeping while holding the hash lock for their
3343  * particular arc header. Thus, we must be careful to never sleep on a
3344  * hash lock in this thread. This is to prevent the following deadlock:
3345  *
3346  *  - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
3347  *    waiting for the reclaim thread to signal it.
3348  *
3349  *  - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
3350  *    fails, and goes to sleep forever.
3351  *
3352  * This possible deadlock is avoided by always acquiring a hash lock
3353  * using mutex_tryenter() from arc_reclaim_thread().
3354  */
3355 static void
3356 arc_reclaim_thread(void)
3357 {
3358         clock_t                 growtime = 0;
3359         callb_cpr_t             cpr;
3360 
3361         CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
3362 
3363         mutex_enter(&arc_reclaim_lock);
3364         while (!arc_reclaim_thread_exit) {
3365                 int64_t free_memory = arc_available_memory();
3366                 uint64_t evicted = 0;
3367 
3368                 mutex_exit(&arc_reclaim_lock);
3369 
3370                 if (free_memory < 0) {
3371 
3372                         arc_no_grow = B_TRUE;
3373                         arc_warm = B_TRUE;
3374 
3375                         /*
3376                          * Wait at least zfs_grow_retry (default 60) seconds
3377                          * before considering growing.
3378                          */
3379                         growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
3380 
3381                         arc_kmem_reap_now();
3382 
3383                         /*
3384                          * If we are still low on memory, shrink the ARC
3385                          * so that we have arc_shrink_min free space.
3386                          */
3387                         free_memory = arc_available_memory();
3388 
3389                         int64_t to_free =
3390                             (arc_c >> arc_shrink_shift) - free_memory;
3391                         if (to_free > 0) {
3392 #ifdef _KERNEL
3393                                 to_free = MAX(to_free, ptob(needfree));
3394 #endif
3395                                 arc_shrink(to_free);
3396                         }
3397                 } else if (free_memory < arc_c >> arc_no_grow_shift) {
3398                         arc_no_grow = B_TRUE;
3399                 } else if (ddi_get_lbolt() >= growtime) {
3400                         arc_no_grow = B_FALSE;
3401                 }
3402 
3403                 evicted = arc_adjust();
3404 
3405                 mutex_enter(&arc_reclaim_lock);
3406 
3407                 /*
3408                  * If evicted is zero, we couldn't evict anything via
3409                  * arc_adjust(). This could be due to hash lock
3410                  * collisions, but more likely due to the majority of
3411                  * arc buffers being unevictable. Therefore, even if
3412                  * arc_size is above arc_c, another pass is unlikely to
3413                  * be helpful and could potentially cause us to enter an
3414                  * infinite loop.
3415                  */
3416                 if (arc_size <= arc_c || evicted == 0) {
3417                         /*
3418                          * We're either no longer overflowing, or we
3419                          * can't evict anything more, so we should wake
3420                          * up any threads before we go to sleep.
3421                          */
3422                         cv_broadcast(&arc_reclaim_waiters_cv);
3423 
3424                         /*
3425                          * Block until signaled, or after one second (we
3426                          * might need to perform arc_kmem_reap_now()
3427                          * even if we aren't being signalled)
3428                          */
3429                         CALLB_CPR_SAFE_BEGIN(&cpr);
3430                         (void) cv_timedwait(&arc_reclaim_thread_cv,
3431                             &arc_reclaim_lock, ddi_get_lbolt() + hz);
3432                         CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
3433                 }
3434         }
3435 
3436         arc_reclaim_thread_exit = FALSE;
3437         cv_broadcast(&arc_reclaim_thread_cv);
3438         CALLB_CPR_EXIT(&cpr);               /* drops arc_reclaim_lock */
3439         thread_exit();
3440 }
3441 
3442 static void
3443 arc_user_evicts_thread(void)
3444 {
3445         callb_cpr_t cpr;
3446 
3447         CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG);
3448 
3449         mutex_enter(&arc_user_evicts_lock);
3450         while (!arc_user_evicts_thread_exit) {
3451                 mutex_exit(&arc_user_evicts_lock);
3452 
3453                 arc_do_user_evicts();
3454 
3455                 /*
3456                  * This is necessary in order for the mdb ::arc dcmd to
3457                  * show up to date information. Since the ::arc command
3458                  * does not call the kstat's update function, without
3459                  * this call, the command may show stale stats for the
3460                  * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
3461                  * with this change, the data might be up to 1 second
3462                  * out of date; but that should suffice. The arc_state_t
3463                  * structures can be queried directly if more accurate
3464                  * information is needed.
3465                  */
3466                 if (arc_ksp != NULL)
3467                         arc_ksp->ks_update(arc_ksp, KSTAT_READ);
3468 
3469                 mutex_enter(&arc_user_evicts_lock);
3470 
3471                 /*
3472                  * Block until signaled, or after one second (we need to
3473                  * call the arc's kstat update function regularly).
3474                  */
3475                 CALLB_CPR_SAFE_BEGIN(&cpr);
3476                 (void) cv_timedwait(&arc_user_evicts_cv,
3477                     &arc_user_evicts_lock, ddi_get_lbolt() + hz);
3478                 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock);
3479         }
3480 
3481         arc_user_evicts_thread_exit = FALSE;
3482         cv_broadcast(&arc_user_evicts_cv);
3483         CALLB_CPR_EXIT(&cpr);               /* drops arc_user_evicts_lock */
3484         thread_exit();
3485 }
3486 
3487 /*
3488  * Adapt arc info given the number of bytes we are trying to add and
3489  * the state that we are comming from.  This function is only called
3490  * when we are adding new content to the cache.
3491  */
3492 static void
3493 arc_adapt(int bytes, arc_state_t *state)
3494 {
3495         int mult;
3496         uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
3497         int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
3498         int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
3499 
3500         if (state == arc_l2c_only)
3501                 return;
3502 
3503         ASSERT(bytes > 0);
3504         /*
3505          * Adapt the target size of the MRU list:
3506          *      - if we just hit in the MRU ghost list, then increase
3507          *        the target size of the MRU list.
3508          *      - if we just hit in the MFU ghost list, then increase
3509          *        the target size of the MFU list by decreasing the
3510          *        target size of the MRU list.
3511          */
3512         if (state == arc_mru_ghost) {
3513                 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
3514                 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
3515 
3516                 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
3517         } else if (state == arc_mfu_ghost) {
3518                 uint64_t delta;
3519 
3520                 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
3521                 mult = MIN(mult, 10);
3522 
3523                 delta = MIN(bytes * mult, arc_p);
3524                 arc_p = MAX(arc_p_min, arc_p - delta);
3525         }
3526         ASSERT((int64_t)arc_p >= 0);
3527 
3528         if (arc_reclaim_needed()) {
3529                 cv_signal(&arc_reclaim_thread_cv);
3530                 return;
3531         }
3532 
3533         if (arc_no_grow)
3534                 return;
3535 
3536         if (arc_c >= arc_c_max)
3537                 return;
3538 
3539         /*
3540          * If we're within (2 * maxblocksize) bytes of the target
3541          * cache size, increment the target cache size
3542          */
3543         if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
3544                 atomic_add_64(&arc_c, (int64_t)bytes);
3545                 if (arc_c > arc_c_max)
3546                         arc_c = arc_c_max;
3547                 else if (state == arc_anon)
3548                         atomic_add_64(&arc_p, (int64_t)bytes);
3549                 if (arc_p > arc_c)
3550                         arc_p = arc_c;
3551         }
3552         ASSERT((int64_t)arc_p >= 0);
3553 }
3554 
3555 /*
3556  * Check if arc_size has grown past our upper threshold, determined by
3557  * zfs_arc_overflow_shift.
3558  */
3559 static boolean_t
3560 arc_is_overflowing(void)
3561 {
3562         /* Always allow at least one block of overflow */
3563         uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
3564             arc_c >> zfs_arc_overflow_shift);
3565 
3566         return (arc_size >= arc_c + overflow);
3567 }
3568 
3569 /*
3570  * The buffer, supplied as the first argument, needs a data block. If we
3571  * are hitting the hard limit for the cache size, we must sleep, waiting
3572  * for the eviction thread to catch up. If we're past the target size
3573  * but below the hard limit, we'll only signal the reclaim thread and
3574  * continue on.
3575  */
3576 static void
3577 arc_get_data_buf(arc_buf_t *buf)
3578 {
3579         arc_state_t             *state = buf->b_hdr->b_l1hdr.b_state;
3580         uint64_t                size = buf->b_hdr->b_size;
3581         arc_buf_contents_t      type = arc_buf_type(buf->b_hdr);
3582 
3583         arc_adapt(size, state);
3584 
3585         /*
3586          * If arc_size is currently overflowing, and has grown past our
3587          * upper limit, we must be adding data faster than the evict
3588          * thread can evict. Thus, to ensure we don't compound the
3589          * problem by adding more data and forcing arc_size to grow even
3590          * further past it's target size, we halt and wait for the
3591          * eviction thread to catch up.
3592          *
3593          * It's also possible that the reclaim thread is unable to evict
3594          * enough buffers to get arc_size below the overflow limit (e.g.
3595          * due to buffers being un-evictable, or hash lock collisions).
3596          * In this case, we want to proceed regardless if we're
3597          * overflowing; thus we don't use a while loop here.
3598          */
3599         if (arc_is_overflowing()) {
3600                 mutex_enter(&arc_reclaim_lock);
3601 
3602                 /*
3603                  * Now that we've acquired the lock, we may no longer be
3604                  * over the overflow limit, lets check.
3605                  *
3606                  * We're ignoring the case of spurious wake ups. If that
3607                  * were to happen, it'd let this thread consume an ARC
3608                  * buffer before it should have (i.e. before we're under
3609                  * the overflow limit and were signalled by the reclaim
3610                  * thread). As long as that is a rare occurrence, it
3611                  * shouldn't cause any harm.
3612                  */
3613                 if (arc_is_overflowing()) {
3614                         cv_signal(&arc_reclaim_thread_cv);
3615                         cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
3616                 }
3617 
3618                 mutex_exit(&arc_reclaim_lock);
3619         }
3620 
3621         if (type == ARC_BUFC_METADATA) {
3622                 buf->b_data = zio_buf_alloc(size);
3623                 arc_space_consume(size, ARC_SPACE_META);
3624         } else {
3625                 ASSERT(type == ARC_BUFC_DATA);
3626                 buf->b_data = zio_data_buf_alloc(size);
3627                 arc_space_consume(size, ARC_SPACE_DATA);
3628         }
3629 
3630         /*
3631          * Update the state size.  Note that ghost states have a
3632          * "ghost size" and so don't need to be updated.
3633          */
3634         if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) {
3635                 arc_buf_hdr_t *hdr = buf->b_hdr;
3636                 arc_state_t *state = hdr->b_l1hdr.b_state;
3637 
3638                 (void) refcount_add_many(&state->arcs_size, size, buf);
3639 
3640                 /*
3641                  * If this is reached via arc_read, the link is
3642                  * protected by the hash lock. If reached via
3643                  * arc_buf_alloc, the header should not be accessed by
3644                  * any other thread. And, if reached via arc_read_done,
3645                  * the hash lock will protect it if it's found in the
3646                  * hash table; otherwise no other thread should be
3647                  * trying to [add|remove]_reference it.
3648                  */
3649                 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3650                         ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3651                         atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type],
3652                             size);
3653                 }
3654                 /*
3655                  * If we are growing the cache, and we are adding anonymous
3656                  * data, and we have outgrown arc_p, update arc_p
3657                  */
3658                 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
3659                     (refcount_count(&arc_anon->arcs_size) +
3660                     refcount_count(&arc_mru->arcs_size) > arc_p))
3661                         arc_p = MIN(arc_c, arc_p + size);
3662         }
3663 }
3664 
3665 /*
3666  * This routine is called whenever a buffer is accessed.
3667  * NOTE: the hash lock is dropped in this function.
3668  */
3669 static void
3670 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3671 {
3672         clock_t now;
3673 
3674         ASSERT(MUTEX_HELD(hash_lock));
3675         ASSERT(HDR_HAS_L1HDR(hdr));
3676 
3677         if (hdr->b_l1hdr.b_state == arc_anon) {
3678                 /*
3679                  * This buffer is not in the cache, and does not
3680                  * appear in our "ghost" list.  Add the new buffer
3681                  * to the MRU state.
3682                  */
3683 
3684                 ASSERT0(hdr->b_l1hdr.b_arc_access);
3685                 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3686                 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3687                 arc_change_state(arc_mru, hdr, hash_lock);
3688 
3689         } else if (hdr->b_l1hdr.b_state == arc_mru) {
3690                 now = ddi_get_lbolt();
3691 
3692                 /*
3693                  * If this buffer is here because of a prefetch, then either:
3694                  * - clear the flag if this is a "referencing" read
3695                  *   (any subsequent access will bump this into the MFU state).
3696                  * or
3697                  * - move the buffer to the head of the list if this is
3698                  *   another prefetch (to make it less likely to be evicted).
3699                  */
3700                 if (HDR_PREFETCH(hdr)) {
3701                         if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
3702                                 /* link protected by hash lock */
3703                                 ASSERT(multilist_link_active(
3704                                     &hdr->b_l1hdr.b_arc_node));
3705                         } else {
3706                                 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3707                                 ARCSTAT_BUMP(arcstat_mru_hits);
3708                         }
3709                         hdr->b_l1hdr.b_arc_access = now;
3710                         return;
3711                 }
3712 
3713                 /*
3714                  * This buffer has been "accessed" only once so far,
3715                  * but it is still in the cache. Move it to the MFU
3716                  * state.
3717                  */
3718                 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
3719                         /*
3720                          * More than 125ms have passed since we
3721                          * instantiated this buffer.  Move it to the
3722                          * most frequently used state.
3723                          */
3724                         hdr->b_l1hdr.b_arc_access = now;
3725                         DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3726                         arc_change_state(arc_mfu, hdr, hash_lock);
3727                 }
3728                 ARCSTAT_BUMP(arcstat_mru_hits);
3729         } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
3730                 arc_state_t     *new_state;
3731                 /*
3732                  * This buffer has been "accessed" recently, but
3733                  * was evicted from the cache.  Move it to the
3734                  * MFU state.
3735                  */
3736 
3737                 if (HDR_PREFETCH(hdr)) {
3738                         new_state = arc_mru;
3739                         if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
3740                                 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3741                         DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3742                 } else {
3743                         new_state = arc_mfu;
3744                         DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3745                 }
3746 
3747                 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3748                 arc_change_state(new_state, hdr, hash_lock);
3749 
3750                 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
3751         } else if (hdr->b_l1hdr.b_state == arc_mfu) {
3752                 /*
3753                  * This buffer has been accessed more than once and is
3754                  * still in the cache.  Keep it in the MFU state.
3755                  *
3756                  * NOTE: an add_reference() that occurred when we did
3757                  * the arc_read() will have kicked this off the list.
3758                  * If it was a prefetch, we will explicitly move it to
3759                  * the head of the list now.
3760                  */
3761                 if ((HDR_PREFETCH(hdr)) != 0) {
3762                         ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3763                         /* link protected by hash_lock */
3764                         ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3765                 }
3766                 ARCSTAT_BUMP(arcstat_mfu_hits);
3767                 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3768         } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
3769                 arc_state_t     *new_state = arc_mfu;
3770                 /*
3771                  * This buffer has been accessed more than once but has
3772                  * been evicted from the cache.  Move it back to the
3773                  * MFU state.
3774                  */
3775 
3776                 if (HDR_PREFETCH(hdr)) {
3777                         /*
3778                          * This is a prefetch access...
3779                          * move this block back to the MRU state.
3780                          */
3781                         ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3782                         new_state = arc_mru;
3783                 }
3784 
3785                 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3786                 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3787                 arc_change_state(new_state, hdr, hash_lock);
3788 
3789                 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
3790         } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
3791                 /*
3792                  * This buffer is on the 2nd Level ARC.
3793                  */
3794 
3795                 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3796                 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3797                 arc_change_state(arc_mfu, hdr, hash_lock);
3798         } else {
3799                 ASSERT(!"invalid arc state");
3800         }
3801 }
3802 
3803 /* a generic arc_done_func_t which you can use */
3804 /* ARGSUSED */
3805 void
3806 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
3807 {
3808         if (zio == NULL || zio->io_error == 0)
3809                 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
3810         VERIFY(arc_buf_remove_ref(buf, arg));
3811 }
3812 
3813 /* a generic arc_done_func_t */
3814 void
3815 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
3816 {
3817         arc_buf_t **bufp = arg;
3818         if (zio && zio->io_error) {
3819                 VERIFY(arc_buf_remove_ref(buf, arg));
3820                 *bufp = NULL;
3821         } else {
3822                 *bufp = buf;
3823                 ASSERT(buf->b_data);
3824         }
3825 }
3826 
3827 static void
3828 arc_read_done(zio_t *zio)
3829 {
3830         arc_buf_hdr_t   *hdr;
3831         arc_buf_t       *buf;
3832         arc_buf_t       *abuf;  /* buffer we're assigning to callback */
3833         kmutex_t        *hash_lock = NULL;
3834         arc_callback_t  *callback_list, *acb;
3835         int             freeable = FALSE;
3836 
3837         buf = zio->io_private;
3838         hdr = buf->b_hdr;
3839 
3840         /*
3841          * The hdr was inserted into hash-table and removed from lists
3842          * prior to starting I/O.  We should find this header, since
3843          * it's in the hash table, and it should be legit since it's
3844          * not possible to evict it during the I/O.  The only possible
3845          * reason for it not to be found is if we were freed during the
3846          * read.
3847          */
3848         if (HDR_IN_HASH_TABLE(hdr)) {
3849                 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
3850                 ASSERT3U(hdr->b_dva.dva_word[0], ==,
3851                     BP_IDENTITY(zio->io_bp)->dva_word[0]);
3852                 ASSERT3U(hdr->b_dva.dva_word[1], ==,
3853                     BP_IDENTITY(zio->io_bp)->dva_word[1]);
3854 
3855                 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
3856                     &hash_lock);
3857 
3858                 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) &&
3859                     hash_lock == NULL) ||
3860                     (found == hdr &&
3861                     DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
3862                     (found == hdr && HDR_L2_READING(hdr)));
3863         }
3864 
3865         hdr->b_flags &= ~ARC_FLAG_L2_EVICTED;
3866         if (l2arc_noprefetch && HDR_PREFETCH(hdr))
3867                 hdr->b_flags &= ~ARC_FLAG_L2CACHE;
3868 
3869         /* byteswap if necessary */
3870         callback_list = hdr->b_l1hdr.b_acb;
3871         ASSERT(callback_list != NULL);
3872         if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
3873                 dmu_object_byteswap_t bswap =
3874                     DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
3875                 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
3876                     byteswap_uint64_array :
3877                     dmu_ot_byteswap[bswap].ob_func;
3878                 func(buf->b_data, hdr->b_size);
3879         }
3880 
3881         arc_cksum_compute(buf, B_FALSE);
3882         arc_buf_watch(buf);
3883 
3884         if (hash_lock && zio->io_error == 0 &&
3885             hdr->b_l1hdr.b_state == arc_anon) {
3886                 /*
3887                  * Only call arc_access on anonymous buffers.  This is because
3888                  * if we've issued an I/O for an evicted buffer, we've already
3889                  * called arc_access (to prevent any simultaneous readers from
3890                  * getting confused).
3891                  */
3892                 arc_access(hdr, hash_lock);
3893         }
3894 
3895         /* create copies of the data buffer for the callers */
3896         abuf = buf;
3897         for (acb = callback_list; acb; acb = acb->acb_next) {
3898                 if (acb->acb_done) {
3899                         if (abuf == NULL) {
3900                                 ARCSTAT_BUMP(arcstat_duplicate_reads);
3901                                 abuf = arc_buf_clone(buf);
3902                         }
3903                         acb->acb_buf = abuf;
3904                         abuf = NULL;
3905                 }
3906         }
3907         hdr->b_l1hdr.b_acb = NULL;
3908         hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
3909         ASSERT(!HDR_BUF_AVAILABLE(hdr));
3910         if (abuf == buf) {
3911                 ASSERT(buf->b_efunc == NULL);
3912                 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
3913                 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
3914         }
3915 
3916         ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
3917             callback_list != NULL);
3918 
3919         if (zio->io_error != 0) {
3920                 hdr->b_flags |= ARC_FLAG_IO_ERROR;
3921                 if (hdr->b_l1hdr.b_state != arc_anon)
3922                         arc_change_state(arc_anon, hdr, hash_lock);
3923                 if (HDR_IN_HASH_TABLE(hdr))
3924                         buf_hash_remove(hdr);
3925                 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
3926         }
3927 
3928         /*
3929          * Broadcast before we drop the hash_lock to avoid the possibility
3930          * that the hdr (and hence the cv) might be freed before we get to
3931          * the cv_broadcast().
3932          */
3933         cv_broadcast(&hdr->b_l1hdr.b_cv);
3934 
3935         if (hash_lock != NULL) {
3936                 mutex_exit(hash_lock);
3937         } else {
3938                 /*
3939                  * This block was freed while we waited for the read to
3940                  * complete.  It has been removed from the hash table and
3941                  * moved to the anonymous state (so that it won't show up
3942                  * in the cache).
3943                  */
3944                 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3945                 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
3946         }
3947 
3948         /* execute each callback and free its structure */
3949         while ((acb = callback_list) != NULL) {
3950                 if (acb->acb_done)
3951                         acb->acb_done(zio, acb->acb_buf, acb->acb_private);
3952 
3953                 if (acb->acb_zio_dummy != NULL) {
3954                         acb->acb_zio_dummy->io_error = zio->io_error;
3955                         zio_nowait(acb->acb_zio_dummy);
3956                 }
3957 
3958                 callback_list = acb->acb_next;
3959                 kmem_free(acb, sizeof (arc_callback_t));
3960         }
3961 
3962         if (freeable)
3963                 arc_hdr_destroy(hdr);
3964 }
3965 
3966 /*
3967  * "Read" the block at the specified DVA (in bp) via the
3968  * cache.  If the block is found in the cache, invoke the provided
3969  * callback immediately and return.  Note that the `zio' parameter
3970  * in the callback will be NULL in this case, since no IO was
3971  * required.  If the block is not in the cache pass the read request
3972  * on to the spa with a substitute callback function, so that the
3973  * requested block will be added to the cache.
3974  *
3975  * If a read request arrives for a block that has a read in-progress,
3976  * either wait for the in-progress read to complete (and return the
3977  * results); or, if this is a read with a "done" func, add a record
3978  * to the read to invoke the "done" func when the read completes,
3979  * and return; or just return.
3980  *
3981  * arc_read_done() will invoke all the requested "done" functions
3982  * for readers of this block.
3983  */
3984 int
3985 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
3986     void *private, zio_priority_t priority, int zio_flags,
3987     arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
3988 {
3989         arc_buf_hdr_t *hdr = NULL;
3990         arc_buf_t *buf = NULL;
3991         kmutex_t *hash_lock = NULL;
3992         zio_t *rzio;
3993         uint64_t guid = spa_load_guid(spa);
3994 
3995         ASSERT(!BP_IS_EMBEDDED(bp) ||
3996             BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
3997 
3998 top:
3999         if (!BP_IS_EMBEDDED(bp)) {
4000                 /*
4001                  * Embedded BP's have no DVA and require no I/O to "read".
4002                  * Create an anonymous arc buf to back it.
4003                  */
4004                 hdr = buf_hash_find(guid, bp, &hash_lock);
4005         }
4006 
4007         if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) {
4008 
4009                 *arc_flags |= ARC_FLAG_CACHED;
4010 
4011                 if (HDR_IO_IN_PROGRESS(hdr)) {
4012 
4013                         if (*arc_flags & ARC_FLAG_WAIT) {
4014                                 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
4015                                 mutex_exit(hash_lock);
4016                                 goto top;
4017                         }
4018                         ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4019 
4020                         if (done) {
4021                                 arc_callback_t  *acb = NULL;
4022 
4023                                 acb = kmem_zalloc(sizeof (arc_callback_t),
4024                                     KM_SLEEP);
4025                                 acb->acb_done = done;
4026                                 acb->acb_private = private;
4027                                 if (pio != NULL)
4028                                         acb->acb_zio_dummy = zio_null(pio,
4029                                             spa, NULL, NULL, NULL, zio_flags);
4030 
4031                                 ASSERT(acb->acb_done != NULL);
4032                                 acb->acb_next = hdr->b_l1hdr.b_acb;
4033                                 hdr->b_l1hdr.b_acb = acb;
4034                                 add_reference(hdr, hash_lock, private);
4035                                 mutex_exit(hash_lock);
4036                                 return (0);
4037                         }
4038                         mutex_exit(hash_lock);
4039                         return (0);
4040                 }
4041 
4042                 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4043                     hdr->b_l1hdr.b_state == arc_mfu);
4044 
4045                 if (done) {
4046                         add_reference(hdr, hash_lock, private);
4047                         /*
4048                          * If this block is already in use, create a new
4049                          * copy of the data so that we will be guaranteed
4050                          * that arc_release() will always succeed.
4051                          */
4052                         buf = hdr->b_l1hdr.b_buf;
4053                         ASSERT(buf);
4054                         ASSERT(buf->b_data);
4055                         if (HDR_BUF_AVAILABLE(hdr)) {
4056                                 ASSERT(buf->b_efunc == NULL);
4057                                 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4058                         } else {
4059                                 buf = arc_buf_clone(buf);
4060                         }
4061 
4062                 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
4063                     refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4064                         hdr->b_flags |= ARC_FLAG_PREFETCH;
4065                 }
4066                 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4067                 arc_access(hdr, hash_lock);
4068                 if (*arc_flags & ARC_FLAG_L2CACHE)
4069                         hdr->b_flags |= ARC_FLAG_L2CACHE;
4070                 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4071                         hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4072                 mutex_exit(hash_lock);
4073                 ARCSTAT_BUMP(arcstat_hits);
4074                 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4075                     demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4076                     data, metadata, hits);
4077 
4078                 if (done)
4079                         done(NULL, buf, private);
4080         } else {
4081                 uint64_t size = BP_GET_LSIZE(bp);
4082                 arc_callback_t *acb;
4083                 vdev_t *vd = NULL;
4084                 uint64_t addr = 0;
4085                 boolean_t devw = B_FALSE;
4086                 enum zio_compress b_compress = ZIO_COMPRESS_OFF;
4087                 int32_t b_asize = 0;
4088 
4089                 if (hdr == NULL) {
4090                         /* this block is not in the cache */
4091                         arc_buf_hdr_t *exists = NULL;
4092                         arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
4093                         buf = arc_buf_alloc(spa, size, private, type);
4094                         hdr = buf->b_hdr;
4095                         if (!BP_IS_EMBEDDED(bp)) {
4096                                 hdr->b_dva = *BP_IDENTITY(bp);
4097                                 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
4098                                 exists = buf_hash_insert(hdr, &hash_lock);
4099                         }
4100                         if (exists != NULL) {
4101                                 /* somebody beat us to the hash insert */
4102                                 mutex_exit(hash_lock);
4103                                 buf_discard_identity(hdr);
4104                                 (void) arc_buf_remove_ref(buf, private);
4105                                 goto top; /* restart the IO request */
4106                         }
4107 
4108                         /* if this is a prefetch, we don't have a reference */
4109                         if (*arc_flags & ARC_FLAG_PREFETCH) {
4110                                 (void) remove_reference(hdr, hash_lock,
4111                                     private);
4112                                 hdr->b_flags |= ARC_FLAG_PREFETCH;
4113                         }
4114                         if (*arc_flags & ARC_FLAG_L2CACHE)
4115                                 hdr->b_flags |= ARC_FLAG_L2CACHE;
4116                         if (*arc_flags & ARC_FLAG_L2COMPRESS)
4117                                 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4118                         if (BP_GET_LEVEL(bp) > 0)
4119                                 hdr->b_flags |= ARC_FLAG_INDIRECT;
4120                 } else {
4121                         /*
4122                          * This block is in the ghost cache. If it was L2-only
4123                          * (and thus didn't have an L1 hdr), we realloc the
4124                          * header to add an L1 hdr.
4125                          */
4126                         if (!HDR_HAS_L1HDR(hdr)) {
4127                                 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
4128                                     hdr_full_cache);
4129                         }
4130 
4131                         ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
4132                         ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4133                         ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4134                         ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
4135 
4136                         /* if this is a prefetch, we don't have a reference */
4137                         if (*arc_flags & ARC_FLAG_PREFETCH)
4138                                 hdr->b_flags |= ARC_FLAG_PREFETCH;
4139                         else
4140                                 add_reference(hdr, hash_lock, private);
4141                         if (*arc_flags & ARC_FLAG_L2CACHE)
4142                                 hdr->b_flags |= ARC_FLAG_L2CACHE;
4143                         if (*arc_flags & ARC_FLAG_L2COMPRESS)
4144                                 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4145                         buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
4146                         buf->b_hdr = hdr;
4147                         buf->b_data = NULL;
4148                         buf->b_efunc = NULL;
4149                         buf->b_private = NULL;
4150                         buf->b_next = NULL;
4151                         hdr->b_l1hdr.b_buf = buf;
4152                         ASSERT0(hdr->b_l1hdr.b_datacnt);
4153                         hdr->b_l1hdr.b_datacnt = 1;
4154                         arc_get_data_buf(buf);
4155                         arc_access(hdr, hash_lock);
4156                 }
4157 
4158                 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
4159 
4160                 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
4161                 acb->acb_done = done;
4162                 acb->acb_private = private;
4163 
4164                 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4165                 hdr->b_l1hdr.b_acb = acb;
4166                 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4167 
4168                 if (HDR_HAS_L2HDR(hdr) &&
4169                     (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
4170                         devw = hdr->b_l2hdr.b_dev->l2ad_writing;
4171                         addr = hdr->b_l2hdr.b_daddr;
4172                         b_compress = HDR_GET_COMPRESS(hdr);
4173                         b_asize = hdr->b_l2hdr.b_asize;
4174                         /*
4175                          * Lock out device removal.
4176                          */
4177                         if (vdev_is_dead(vd) ||
4178                             !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
4179                                 vd = NULL;
4180                 }
4181 
4182                 if (hash_lock != NULL)
4183                         mutex_exit(hash_lock);
4184 
4185                 /*
4186                  * At this point, we have a level 1 cache miss.  Try again in
4187                  * L2ARC if possible.
4188                  */
4189                 ASSERT3U(hdr->b_size, ==, size);
4190                 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
4191                     uint64_t, size, zbookmark_phys_t *, zb);
4192                 ARCSTAT_BUMP(arcstat_misses);
4193                 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4194                     demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4195                     data, metadata, misses);
4196 
4197                 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
4198                         /*
4199                          * Read from the L2ARC if the following are true:
4200                          * 1. The L2ARC vdev was previously cached.
4201                          * 2. This buffer still has L2ARC metadata.
4202                          * 3. This buffer isn't currently writing to the L2ARC.
4203                          * 4. The L2ARC entry wasn't evicted, which may
4204                          *    also have invalidated the vdev.
4205                          * 5. This isn't prefetch and l2arc_noprefetch is set.
4206                          */
4207                         if (HDR_HAS_L2HDR(hdr) &&
4208                             !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
4209                             !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
4210                                 l2arc_read_callback_t *cb;
4211 
4212                                 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
4213                                 ARCSTAT_BUMP(arcstat_l2_hits);
4214 
4215                                 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
4216                                     KM_SLEEP);
4217                                 cb->l2rcb_buf = buf;
4218                                 cb->l2rcb_spa = spa;
4219                                 cb->l2rcb_bp = *bp;
4220                                 cb->l2rcb_zb = *zb;
4221                                 cb->l2rcb_flags = zio_flags;
4222                                 cb->l2rcb_compress = b_compress;
4223 
4224                                 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
4225                                     addr + size < vd->vdev_psize -
4226                                     VDEV_LABEL_END_SIZE);
4227 
4228                                 /*
4229                                  * l2arc read.  The SCL_L2ARC lock will be
4230                                  * released by l2arc_read_done().
4231                                  * Issue a null zio if the underlying buffer
4232                                  * was squashed to zero size by compression.
4233                                  */
4234                                 if (b_compress == ZIO_COMPRESS_EMPTY) {
4235                                         rzio = zio_null(pio, spa, vd,
4236                                             l2arc_read_done, cb,
4237                                             zio_flags | ZIO_FLAG_DONT_CACHE |
4238                                             ZIO_FLAG_CANFAIL |
4239                                             ZIO_FLAG_DONT_PROPAGATE |
4240                                             ZIO_FLAG_DONT_RETRY);
4241                                 } else {
4242                                         rzio = zio_read_phys(pio, vd, addr,
4243                                             b_asize, buf->b_data,
4244                                             ZIO_CHECKSUM_OFF,
4245                                             l2arc_read_done, cb, priority,
4246                                             zio_flags | ZIO_FLAG_DONT_CACHE |
4247                                             ZIO_FLAG_CANFAIL |
4248                                             ZIO_FLAG_DONT_PROPAGATE |
4249                                             ZIO_FLAG_DONT_RETRY, B_FALSE);
4250                                 }
4251                                 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
4252                                     zio_t *, rzio);
4253                                 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize);
4254 
4255                                 if (*arc_flags & ARC_FLAG_NOWAIT) {
4256                                         zio_nowait(rzio);
4257                                         return (0);
4258                                 }
4259 
4260                                 ASSERT(*arc_flags & ARC_FLAG_WAIT);
4261                                 if (zio_wait(rzio) == 0)
4262                                         return (0);
4263 
4264                                 /* l2arc read error; goto zio_read() */
4265                         } else {
4266                                 DTRACE_PROBE1(l2arc__miss,
4267                                     arc_buf_hdr_t *, hdr);
4268                                 ARCSTAT_BUMP(arcstat_l2_misses);
4269                                 if (HDR_L2_WRITING(hdr))
4270                                         ARCSTAT_BUMP(arcstat_l2_rw_clash);
4271                                 spa_config_exit(spa, SCL_L2ARC, vd);
4272                         }
4273                 } else {
4274                         if (vd != NULL)
4275                                 spa_config_exit(spa, SCL_L2ARC, vd);
4276                         if (l2arc_ndev != 0) {
4277                                 DTRACE_PROBE1(l2arc__miss,
4278                                     arc_buf_hdr_t *, hdr);
4279                                 ARCSTAT_BUMP(arcstat_l2_misses);
4280                         }
4281                 }
4282 
4283                 rzio = zio_read(pio, spa, bp, buf->b_data, size,
4284                     arc_read_done, buf, priority, zio_flags, zb);
4285 
4286                 if (*arc_flags & ARC_FLAG_WAIT)
4287                         return (zio_wait(rzio));
4288 
4289                 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4290                 zio_nowait(rzio);
4291         }
4292         return (0);
4293 }
4294 
4295 void
4296 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
4297 {
4298         ASSERT(buf->b_hdr != NULL);
4299         ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon);
4300         ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) ||
4301             func == NULL);
4302         ASSERT(buf->b_efunc == NULL);
4303         ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
4304 
4305         buf->b_efunc = func;
4306         buf->b_private = private;
4307 }
4308 
4309 /*
4310  * Notify the arc that a block was freed, and thus will never be used again.
4311  */
4312 void
4313 arc_freed(spa_t *spa, const blkptr_t *bp)
4314 {
4315         arc_buf_hdr_t *hdr;
4316         kmutex_t *hash_lock;
4317         uint64_t guid = spa_load_guid(spa);
4318 
4319         ASSERT(!BP_IS_EMBEDDED(bp));
4320 
4321         hdr = buf_hash_find(guid, bp, &hash_lock);
4322         if (hdr == NULL)
4323                 return;
4324         if (HDR_BUF_AVAILABLE(hdr)) {
4325                 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
4326                 add_reference(hdr, hash_lock, FTAG);
4327                 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4328                 mutex_exit(hash_lock);
4329 
4330                 arc_release(buf, FTAG);
4331                 (void) arc_buf_remove_ref(buf, FTAG);
4332         } else {
4333                 mutex_exit(hash_lock);
4334         }
4335 
4336 }
4337 
4338 /*
4339  * Clear the user eviction callback set by arc_set_callback(), first calling
4340  * it if it exists.  Because the presence of a callback keeps an arc_buf cached
4341  * clearing the callback may result in the arc_buf being destroyed.  However,
4342  * it will not result in the *last* arc_buf being destroyed, hence the data
4343  * will remain cached in the ARC. We make a copy of the arc buffer here so
4344  * that we can process the callback without holding any locks.
4345  *
4346  * It's possible that the callback is already in the process of being cleared
4347  * by another thread.  In this case we can not clear the callback.
4348  *
4349  * Returns B_TRUE if the callback was successfully called and cleared.
4350  */
4351 boolean_t
4352 arc_clear_callback(arc_buf_t *buf)
4353 {
4354         arc_buf_hdr_t *hdr;
4355         kmutex_t *hash_lock;
4356         arc_evict_func_t *efunc = buf->b_efunc;
4357         void *private = buf->b_private;
4358 
4359         mutex_enter(&buf->b_evict_lock);
4360         hdr = buf->b_hdr;
4361         if (hdr == NULL) {
4362                 /*
4363                  * We are in arc_do_user_evicts().
4364                  */
4365                 ASSERT(buf->b_data == NULL);
4366                 mutex_exit(&buf->b_evict_lock);
4367                 return (B_FALSE);
4368         } else if (buf->b_data == NULL) {
4369                 /*
4370                  * We are on the eviction list; process this buffer now
4371                  * but let arc_do_user_evicts() do the reaping.
4372                  */
4373                 buf->b_efunc = NULL;
4374                 mutex_exit(&buf->b_evict_lock);
4375                 VERIFY0(efunc(private));
4376                 return (B_TRUE);
4377         }
4378         hash_lock = HDR_LOCK(hdr);
4379         mutex_enter(hash_lock);
4380         hdr = buf->b_hdr;
4381         ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4382 
4383         ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <,
4384             hdr->b_l1hdr.b_datacnt);
4385         ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4386             hdr->b_l1hdr.b_state == arc_mfu);
4387 
4388         buf->b_efunc = NULL;
4389         buf->b_private = NULL;
4390 
4391         if (hdr->b_l1hdr.b_datacnt > 1) {
4392                 mutex_exit(&buf->b_evict_lock);
4393                 arc_buf_destroy(buf, TRUE);
4394         } else {
4395                 ASSERT(buf == hdr->b_l1hdr.b_buf);
4396                 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
4397                 mutex_exit(&buf->b_evict_lock);
4398         }
4399 
4400         mutex_exit(hash_lock);
4401         VERIFY0(efunc(private));
4402         return (B_TRUE);
4403 }
4404 
4405 /*
4406  * Release this buffer from the cache, making it an anonymous buffer.  This
4407  * must be done after a read and prior to modifying the buffer contents.
4408  * If the buffer has more than one reference, we must make
4409  * a new hdr for the buffer.
4410  */
4411 void
4412 arc_release(arc_buf_t *buf, void *tag)
4413 {
4414         arc_buf_hdr_t *hdr = buf->b_hdr;
4415 
4416         /*
4417          * It would be nice to assert that if it's DMU metadata (level >
4418          * 0 || it's the dnode file), then it must be syncing context.
4419          * But we don't know that information at this level.
4420          */
4421 
4422         mutex_enter(&buf->b_evict_lock);
4423 
4424         ASSERT(HDR_HAS_L1HDR(hdr));
4425 
4426         /*
4427          * We don't grab the hash lock prior to this check, because if
4428          * the buffer's header is in the arc_anon state, it won't be
4429          * linked into the hash table.
4430          */
4431         if (hdr->b_l1hdr.b_state == arc_anon) {
4432                 mutex_exit(&buf->b_evict_lock);
4433                 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4434                 ASSERT(!HDR_IN_HASH_TABLE(hdr));
4435                 ASSERT(!HDR_HAS_L2HDR(hdr));
4436                 ASSERT(BUF_EMPTY(hdr));
4437 
4438                 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1);
4439                 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
4440                 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
4441 
4442                 ASSERT3P(buf->b_efunc, ==, NULL);
4443                 ASSERT3P(buf->b_private, ==, NULL);
4444 
4445                 hdr->b_l1hdr.b_arc_access = 0;
4446                 arc_buf_thaw(buf);
4447 
4448                 return;
4449         }
4450 
4451         kmutex_t *hash_lock = HDR_LOCK(hdr);
4452         mutex_enter(hash_lock);
4453 
4454         /*
4455          * This assignment is only valid as long as the hash_lock is
4456          * held, we must be careful not to reference state or the
4457          * b_state field after dropping the lock.
4458          */
4459         arc_state_t *state = hdr->b_l1hdr.b_state;
4460         ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4461         ASSERT3P(state, !=, arc_anon);
4462 
4463         /* this buffer is not on any list */
4464         ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0);
4465 
4466         if (HDR_HAS_L2HDR(hdr)) {
4467                 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4468 
4469                 /*
4470                  * We have to recheck this conditional again now that
4471                  * we're holding the l2ad_mtx to prevent a race with
4472                  * another thread which might be concurrently calling
4473                  * l2arc_evict(). In that case, l2arc_evict() might have
4474                  * destroyed the header's L2 portion as we were waiting
4475                  * to acquire the l2ad_mtx.
4476                  */
4477                 if (HDR_HAS_L2HDR(hdr))
4478                         arc_hdr_l2hdr_destroy(hdr);
4479 
4480                 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4481         }
4482 
4483         /*
4484          * Do we have more than one buf?
4485          */
4486         if (hdr->b_l1hdr.b_datacnt > 1) {
4487                 arc_buf_hdr_t *nhdr;
4488                 arc_buf_t **bufp;
4489                 uint64_t blksz = hdr->b_size;
4490                 uint64_t spa = hdr->b_spa;
4491                 arc_buf_contents_t type = arc_buf_type(hdr);
4492                 uint32_t flags = hdr->b_flags;
4493 
4494                 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
4495                 /*
4496                  * Pull the data off of this hdr and attach it to
4497                  * a new anonymous hdr.
4498                  */
4499                 (void) remove_reference(hdr, hash_lock, tag);
4500                 bufp = &hdr->b_l1hdr.b_buf;
4501                 while (*bufp != buf)
4502                         bufp = &(*bufp)->b_next;
4503                 *bufp = buf->b_next;
4504                 buf->b_next = NULL;
4505 
4506                 ASSERT3P(state, !=, arc_l2c_only);
4507 
4508                 (void) refcount_remove_many(
4509                     &state->arcs_size, hdr->b_size, buf);
4510 
4511                 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
4512                         ASSERT3P(state, !=, arc_l2c_only);
4513                         uint64_t *size = &state->arcs_lsize[type];
4514                         ASSERT3U(*size, >=, hdr->b_size);
4515                         atomic_add_64(size, -hdr->b_size);
4516                 }
4517 
4518                 /*
4519                  * We're releasing a duplicate user data buffer, update
4520                  * our statistics accordingly.
4521                  */
4522                 if (HDR_ISTYPE_DATA(hdr)) {
4523                         ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
4524                         ARCSTAT_INCR(arcstat_duplicate_buffers_size,
4525                             -hdr->b_size);
4526                 }
4527                 hdr->b_l1hdr.b_datacnt -= 1;
4528                 arc_cksum_verify(buf);
4529                 arc_buf_unwatch(buf);
4530 
4531                 mutex_exit(hash_lock);
4532 
4533                 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
4534                 nhdr->b_size = blksz;
4535                 nhdr->b_spa = spa;
4536 
4537                 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING;
4538                 nhdr->b_flags |= arc_bufc_to_flags(type);
4539                 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
4540 
4541                 nhdr->b_l1hdr.b_buf = buf;
4542                 nhdr->b_l1hdr.b_datacnt = 1;
4543                 nhdr->b_l1hdr.b_state = arc_anon;
4544                 nhdr->b_l1hdr.b_arc_access = 0;
4545                 nhdr->b_l1hdr.b_tmp_cdata = NULL;
4546                 nhdr->b_freeze_cksum = NULL;
4547 
4548                 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
4549                 buf->b_hdr = nhdr;
4550                 mutex_exit(&buf->b_evict_lock);
4551                 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf);
4552         } else {
4553                 mutex_exit(&buf->b_evict_lock);
4554                 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
4555                 /* protected by hash lock, or hdr is on arc_anon */
4556                 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4557                 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4558                 arc_change_state(arc_anon, hdr, hash_lock);
4559                 hdr->b_l1hdr.b_arc_access = 0;
4560                 mutex_exit(hash_lock);
4561 
4562                 buf_discard_identity(hdr);
4563                 arc_buf_thaw(buf);
4564         }
4565         buf->b_efunc = NULL;
4566         buf->b_private = NULL;
4567 }
4568 
4569 int
4570 arc_released(arc_buf_t *buf)
4571 {
4572         int released;
4573 
4574         mutex_enter(&buf->b_evict_lock);
4575         released = (buf->b_data != NULL &&
4576             buf->b_hdr->b_l1hdr.b_state == arc_anon);
4577         mutex_exit(&buf->b_evict_lock);
4578         return (released);
4579 }
4580 
4581 #ifdef ZFS_DEBUG
4582 int
4583 arc_referenced(arc_buf_t *buf)
4584 {
4585         int referenced;
4586 
4587         mutex_enter(&buf->b_evict_lock);
4588         referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
4589         mutex_exit(&buf->b_evict_lock);
4590         return (referenced);
4591 }
4592 #endif
4593 
4594 static void
4595 arc_write_ready(zio_t *zio)
4596 {
4597         arc_write_callback_t *callback = zio->io_private;
4598         arc_buf_t *buf = callback->awcb_buf;
4599         arc_buf_hdr_t *hdr = buf->b_hdr;
4600 
4601         ASSERT(HDR_HAS_L1HDR(hdr));
4602         ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
4603         ASSERT(hdr->b_l1hdr.b_datacnt > 0);
4604         callback->awcb_ready(zio, buf, callback->awcb_private);
4605 
4606         /*
4607          * If the IO is already in progress, then this is a re-write
4608          * attempt, so we need to thaw and re-compute the cksum.
4609          * It is the responsibility of the callback to handle the
4610          * accounting for any re-write attempt.
4611          */
4612         if (HDR_IO_IN_PROGRESS(hdr)) {
4613                 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
4614                 if (hdr->b_freeze_cksum != NULL) {
4615                         kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
4616                         hdr->b_freeze_cksum = NULL;
4617                 }
4618                 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
4619         }
4620         arc_cksum_compute(buf, B_FALSE);
4621         hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4622 }
4623 
4624 /*
4625  * The SPA calls this callback for each physical write that happens on behalf
4626  * of a logical write.  See the comment in dbuf_write_physdone() for details.
4627  */
4628 static void
4629 arc_write_physdone(zio_t *zio)
4630 {
4631         arc_write_callback_t *cb = zio->io_private;
4632         if (cb->awcb_physdone != NULL)
4633                 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
4634 }
4635 
4636 static void
4637 arc_write_done(zio_t *zio)
4638 {
4639         arc_write_callback_t *callback = zio->io_private;
4640         arc_buf_t *buf = callback->awcb_buf;
4641         arc_buf_hdr_t *hdr = buf->b_hdr;
4642 
4643         ASSERT(hdr->b_l1hdr.b_acb == NULL);
4644 
4645         if (zio->io_error == 0) {
4646                 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
4647                         buf_discard_identity(hdr);
4648                 } else {
4649                         hdr->b_dva = *BP_IDENTITY(zio->io_bp);
4650                         hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
4651                 }
4652         } else {
4653                 ASSERT(BUF_EMPTY(hdr));
4654         }
4655 
4656         /*
4657          * If the block to be written was all-zero or compressed enough to be
4658          * embedded in the BP, no write was performed so there will be no
4659          * dva/birth/checksum.  The buffer must therefore remain anonymous
4660          * (and uncached).
4661          */
4662         if (!BUF_EMPTY(hdr)) {
4663                 arc_buf_hdr_t *exists;
4664                 kmutex_t *hash_lock;
4665 
4666                 ASSERT(zio->io_error == 0);
4667 
4668                 arc_cksum_verify(buf);
4669 
4670                 exists = buf_hash_insert(hdr, &hash_lock);
4671                 if (exists != NULL) {
4672                         /*
4673                          * This can only happen if we overwrite for
4674                          * sync-to-convergence, because we remove
4675                          * buffers from the hash table when we arc_free().
4676                          */
4677                         if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
4678                                 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
4679                                         panic("bad overwrite, hdr=%p exists=%p",
4680                                             (void *)hdr, (void *)exists);
4681                                 ASSERT(refcount_is_zero(
4682                                     &exists->b_l1hdr.b_refcnt));
4683                                 arc_change_state(arc_anon, exists, hash_lock);
4684                                 mutex_exit(hash_lock);
4685                                 arc_hdr_destroy(exists);
4686                                 exists = buf_hash_insert(hdr, &hash_lock);
4687                                 ASSERT3P(exists, ==, NULL);
4688                         } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
4689                                 /* nopwrite */
4690                                 ASSERT(zio->io_prop.zp_nopwrite);
4691                                 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
4692                                         panic("bad nopwrite, hdr=%p exists=%p",
4693                                             (void *)hdr, (void *)exists);
4694                         } else {
4695                                 /* Dedup */
4696                                 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
4697                                 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
4698                                 ASSERT(BP_GET_DEDUP(zio->io_bp));
4699                                 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
4700                         }
4701                 }
4702                 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4703                 /* if it's not anon, we are doing a scrub */
4704                 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
4705                         arc_access(hdr, hash_lock);
4706                 mutex_exit(hash_lock);
4707         } else {
4708                 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4709         }
4710 
4711         ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4712         callback->awcb_done(zio, buf, callback->awcb_private);
4713 
4714         kmem_free(callback, sizeof (arc_write_callback_t));
4715 }
4716 
4717 zio_t *
4718 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
4719     blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
4720     const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
4721     arc_done_func_t *done, void *private, zio_priority_t priority,
4722     int zio_flags, const zbookmark_phys_t *zb)
4723 {
4724         arc_buf_hdr_t *hdr = buf->b_hdr;
4725         arc_write_callback_t *callback;
4726         zio_t *zio;
4727 
4728         ASSERT(ready != NULL);
4729         ASSERT(done != NULL);
4730         ASSERT(!HDR_IO_ERROR(hdr));
4731         ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4732         ASSERT(hdr->b_l1hdr.b_acb == NULL);
4733         ASSERT(hdr->b_l1hdr.b_datacnt > 0);
4734         if (l2arc)
4735                 hdr->b_flags |= ARC_FLAG_L2CACHE;
4736         if (l2arc_compress)
4737                 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4738         callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
4739         callback->awcb_ready = ready;
4740         callback->awcb_physdone = physdone;
4741         callback->awcb_done = done;
4742         callback->awcb_private = private;
4743         callback->awcb_buf = buf;
4744 
4745         zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
4746             arc_write_ready, arc_write_physdone, arc_write_done, callback,
4747             priority, zio_flags, zb);
4748 
4749         return (zio);
4750 }
4751 
4752 static int
4753 arc_memory_throttle(uint64_t reserve, uint64_t txg)
4754 {
4755 #ifdef _KERNEL
4756         uint64_t available_memory = ptob(freemem);
4757         static uint64_t page_load = 0;
4758         static uint64_t last_txg = 0;
4759 
4760 #if defined(__i386)
4761         available_memory =
4762             MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
4763 #endif
4764 
4765         if (freemem > physmem * arc_lotsfree_percent / 100)
4766                 return (0);
4767 
4768         if (txg > last_txg) {
4769                 last_txg = txg;
4770                 page_load = 0;
4771         }
4772         /*
4773          * If we are in pageout, we know that memory is already tight,
4774          * the arc is already going to be evicting, so we just want to
4775          * continue to let page writes occur as quickly as possible.
4776          */
4777         if (curproc == proc_pageout) {
4778                 if (page_load > MAX(ptob(minfree), available_memory) / 4)
4779                         return (SET_ERROR(ERESTART));
4780                 /* Note: reserve is inflated, so we deflate */
4781                 page_load += reserve / 8;
4782                 return (0);
4783         } else if (page_load > 0 && arc_reclaim_needed()) {
4784                 /* memory is low, delay before restarting */
4785                 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
4786                 return (SET_ERROR(EAGAIN));
4787         }
4788         page_load = 0;
4789 #endif
4790         return (0);
4791 }
4792 
4793 void
4794 arc_tempreserve_clear(uint64_t reserve)
4795 {
4796         atomic_add_64(&arc_tempreserve, -reserve);
4797         ASSERT((int64_t)arc_tempreserve >= 0);
4798 }
4799 
4800 int
4801 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
4802 {
4803         int error;
4804         uint64_t anon_size;
4805 
4806         if (reserve > arc_c/4 && !arc_no_grow)
4807                 arc_c = MIN(arc_c_max, reserve * 4);
4808         if (reserve > arc_c)
4809                 return (SET_ERROR(ENOMEM));
4810 
4811         /*
4812          * Don't count loaned bufs as in flight dirty data to prevent long
4813          * network delays from blocking transactions that are ready to be
4814          * assigned to a txg.
4815          */
4816         anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
4817             arc_loaned_bytes), 0);
4818 
4819         /*
4820          * Writes will, almost always, require additional memory allocations
4821          * in order to compress/encrypt/etc the data.  We therefore need to
4822          * make sure that there is sufficient available memory for this.
4823          */
4824         error = arc_memory_throttle(reserve, txg);
4825         if (error != 0)
4826                 return (error);
4827 
4828         /*
4829          * Throttle writes when the amount of dirty data in the cache
4830          * gets too large.  We try to keep the cache less than half full
4831          * of dirty blocks so that our sync times don't grow too large.
4832          * Note: if two requests come in concurrently, we might let them
4833          * both succeed, when one of them should fail.  Not a huge deal.
4834          */
4835 
4836         if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
4837             anon_size > arc_c / 4) {
4838                 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
4839                     "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
4840                     arc_tempreserve>>10,
4841                     arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
4842                     arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
4843                     reserve>>10, arc_c>>10);
4844                 return (SET_ERROR(ERESTART));
4845         }
4846         atomic_add_64(&arc_tempreserve, reserve);
4847         return (0);
4848 }
4849 
4850 static void
4851 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
4852     kstat_named_t *evict_data, kstat_named_t *evict_metadata)
4853 {
4854         size->value.ui64 = refcount_count(&state->arcs_size);
4855         evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
4856         evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
4857 }
4858 
4859 static int
4860 arc_kstat_update(kstat_t *ksp, int rw)
4861 {
4862         arc_stats_t *as = ksp->ks_data;
4863 
4864         if (rw == KSTAT_WRITE) {
4865                 return (EACCES);
4866         } else {
4867                 arc_kstat_update_state(arc_anon,
4868                     &as->arcstat_anon_size,
4869                     &as->arcstat_anon_evictable_data,
4870                     &as->arcstat_anon_evictable_metadata);
4871                 arc_kstat_update_state(arc_mru,
4872                     &as->arcstat_mru_size,
4873                     &as->arcstat_mru_evictable_data,
4874                     &as->arcstat_mru_evictable_metadata);
4875                 arc_kstat_update_state(arc_mru_ghost,
4876                     &as->arcstat_mru_ghost_size,
4877                     &as->arcstat_mru_ghost_evictable_data,
4878                     &as->arcstat_mru_ghost_evictable_metadata);
4879                 arc_kstat_update_state(arc_mfu,
4880                     &as->arcstat_mfu_size,
4881                     &as->arcstat_mfu_evictable_data,
4882                     &as->arcstat_mfu_evictable_metadata);
4883                 arc_kstat_update_state(arc_mfu_ghost,
4884                     &as->arcstat_mfu_ghost_size,
4885                     &as->arcstat_mfu_ghost_evictable_data,
4886                     &as->arcstat_mfu_ghost_evictable_metadata);
4887         }
4888 
4889         return (0);
4890 }
4891 
4892 /*
4893  * This function *must* return indices evenly distributed between all
4894  * sublists of the multilist. This is needed due to how the ARC eviction
4895  * code is laid out; arc_evict_state() assumes ARC buffers are evenly
4896  * distributed between all sublists and uses this assumption when
4897  * deciding which sublist to evict from and how much to evict from it.
4898  */
4899 unsigned int
4900 arc_state_multilist_index_func(multilist_t *ml, void *obj)
4901 {
4902         arc_buf_hdr_t *hdr = obj;
4903 
4904         /*
4905          * We rely on b_dva to generate evenly distributed index
4906          * numbers using buf_hash below. So, as an added precaution,
4907          * let's make sure we never add empty buffers to the arc lists.
4908          */
4909         ASSERT(!BUF_EMPTY(hdr));
4910 
4911         /*
4912          * The assumption here, is the hash value for a given
4913          * arc_buf_hdr_t will remain constant throughout it's lifetime
4914          * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
4915          * Thus, we don't need to store the header's sublist index
4916          * on insertion, as this index can be recalculated on removal.
4917          *
4918          * Also, the low order bits of the hash value are thought to be
4919          * distributed evenly. Otherwise, in the case that the multilist
4920          * has a power of two number of sublists, each sublists' usage
4921          * would not be evenly distributed.
4922          */
4923         return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
4924             multilist_get_num_sublists(ml));
4925 }
4926 
4927 void
4928 arc_init(void)
4929 {
4930         /*
4931          * allmem is "all memory that we could possibly use".
4932          */
4933 #ifdef _KERNEL
4934         uint64_t allmem = ptob(physmem - swapfs_minfree);
4935 #else
4936         uint64_t allmem = (physmem * PAGESIZE) / 2;
4937 #endif
4938 
4939         mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
4940         cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
4941         cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
4942 
4943         mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
4944         cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL);
4945 
4946         /* Convert seconds to clock ticks */
4947         arc_min_prefetch_lifespan = 1 * hz;
4948 
4949         /* Start out with 1/8 of all memory */
4950         arc_c = allmem / 8;
4951 
4952 #ifdef _KERNEL
4953         /*
4954          * On architectures where the physical memory can be larger
4955          * than the addressable space (intel in 32-bit mode), we may
4956          * need to limit the cache to 1/8 of VM size.
4957          */
4958         arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
4959 #endif
4960 
4961         /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
4962         arc_c_min = MAX(allmem / 32, 64 << 20);
4963         /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
4964         if (allmem >= 1 << 30)
4965                 arc_c_max = allmem - (1 << 30);
4966         else
4967                 arc_c_max = arc_c_min;
4968         arc_c_max = MAX(allmem * 3 / 4, arc_c_max);
4969 
4970         /*
4971          * Allow the tunables to override our calculations if they are
4972          * reasonable (ie. over 64MB)
4973          */
4974         if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem)
4975                 arc_c_max = zfs_arc_max;
4976         if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max)
4977                 arc_c_min = zfs_arc_min;
4978 
4979         arc_c = arc_c_max;
4980         arc_p = (arc_c >> 1);
4981 
4982         /* limit meta-data to 1/4 of the arc capacity */
4983         arc_meta_limit = arc_c_max / 4;
4984 
4985         /* Allow the tunable to override if it is reasonable */
4986         if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
4987                 arc_meta_limit = zfs_arc_meta_limit;
4988 
4989         if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
4990                 arc_c_min = arc_meta_limit / 2;
4991 
4992         if (zfs_arc_meta_min > 0) {
4993                 arc_meta_min = zfs_arc_meta_min;
4994         } else {
4995                 arc_meta_min = arc_c_min / 2;
4996         }
4997 
4998         if (zfs_arc_grow_retry > 0)
4999                 arc_grow_retry = zfs_arc_grow_retry;
5000 
5001         if (zfs_arc_shrink_shift > 0)
5002                 arc_shrink_shift = zfs_arc_shrink_shift;
5003 
5004         /*
5005          * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
5006          */
5007         if (arc_no_grow_shift >= arc_shrink_shift)
5008                 arc_no_grow_shift = arc_shrink_shift - 1;
5009 
5010         if (zfs_arc_p_min_shift > 0)
5011                 arc_p_min_shift = zfs_arc_p_min_shift;
5012 
5013         if (zfs_arc_num_sublists_per_state < 1)
5014                 zfs_arc_num_sublists_per_state = MAX(boot_ncpus, 1);
5015 
5016         /* if kmem_flags are set, lets try to use less memory */
5017         if (kmem_debugging())
5018                 arc_c = arc_c / 2;
5019         if (arc_c < arc_c_min)
5020                 arc_c = arc_c_min;
5021 
5022         arc_anon = &ARC_anon;
5023         arc_mru = &ARC_mru;
5024         arc_mru_ghost = &ARC_mru_ghost;
5025         arc_mfu = &ARC_mfu;
5026         arc_mfu_ghost = &ARC_mfu_ghost;
5027         arc_l2c_only = &ARC_l2c_only;
5028         arc_size = 0;
5029 
5030         multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
5031             sizeof (arc_buf_hdr_t),
5032             offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5033             zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5034         multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
5035             sizeof (arc_buf_hdr_t),
5036             offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5037             zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5038         multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
5039             sizeof (arc_buf_hdr_t),
5040             offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5041             zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5042         multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
5043             sizeof (arc_buf_hdr_t),
5044             offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5045             zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5046         multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
5047             sizeof (arc_buf_hdr_t),
5048             offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5049             zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5050         multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
5051             sizeof (arc_buf_hdr_t),
5052             offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5053             zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5054         multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
5055             sizeof (arc_buf_hdr_t),
5056             offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5057             zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5058         multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
5059             sizeof (arc_buf_hdr_t),
5060             offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5061             zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5062         multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
5063             sizeof (arc_buf_hdr_t),
5064             offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5065             zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5066         multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
5067             sizeof (arc_buf_hdr_t),
5068             offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5069             zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5070 
5071         refcount_create(&arc_anon->arcs_size);
5072         refcount_create(&arc_mru->arcs_size);
5073         refcount_create(&arc_mru_ghost->arcs_size);
5074         refcount_create(&arc_mfu->arcs_size);
5075         refcount_create(&arc_mfu_ghost->arcs_size);
5076         refcount_create(&arc_l2c_only->arcs_size);
5077 
5078         buf_init();
5079 
5080         arc_reclaim_thread_exit = FALSE;
5081         arc_user_evicts_thread_exit = FALSE;
5082         arc_eviction_list = NULL;
5083         bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
5084 
5085         arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
5086             sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
5087 
5088         if (arc_ksp != NULL) {
5089                 arc_ksp->ks_data = &arc_stats;
5090                 arc_ksp->ks_update = arc_kstat_update;
5091                 kstat_install(arc_ksp);
5092         }
5093 
5094         (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
5095             TS_RUN, minclsyspri);
5096 
5097         (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0,
5098             TS_RUN, minclsyspri);
5099 
5100         arc_dead = FALSE;
5101         arc_warm = B_FALSE;
5102 
5103         /*
5104          * Calculate maximum amount of dirty data per pool.
5105          *
5106          * If it has been set by /etc/system, take that.
5107          * Otherwise, use a percentage of physical memory defined by
5108          * zfs_dirty_data_max_percent (default 10%) with a cap at
5109          * zfs_dirty_data_max_max (default 4GB).
5110          */
5111         if (zfs_dirty_data_max == 0) {
5112                 zfs_dirty_data_max = physmem * PAGESIZE *
5113                     zfs_dirty_data_max_percent / 100;
5114                 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
5115                     zfs_dirty_data_max_max);
5116         }
5117 }
5118 
5119 void
5120 arc_fini(void)
5121 {
5122         mutex_enter(&arc_reclaim_lock);
5123         arc_reclaim_thread_exit = TRUE;
5124         /*
5125          * The reclaim thread will set arc_reclaim_thread_exit back to
5126          * FALSE when it is finished exiting; we're waiting for that.
5127          */
5128         while (arc_reclaim_thread_exit) {
5129                 cv_signal(&arc_reclaim_thread_cv);
5130                 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
5131         }
5132         mutex_exit(&arc_reclaim_lock);
5133 
5134         mutex_enter(&arc_user_evicts_lock);
5135         arc_user_evicts_thread_exit = TRUE;
5136         /*
5137          * The user evicts thread will set arc_user_evicts_thread_exit
5138          * to FALSE when it is finished exiting; we're waiting for that.
5139          */
5140         while (arc_user_evicts_thread_exit) {
5141                 cv_signal(&arc_user_evicts_cv);
5142                 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock);
5143         }
5144         mutex_exit(&arc_user_evicts_lock);
5145 
5146         /* Use TRUE to ensure *all* buffers are evicted */
5147         arc_flush(NULL, TRUE);
5148 
5149         arc_dead = TRUE;
5150 
5151         if (arc_ksp != NULL) {
5152                 kstat_delete(arc_ksp);
5153                 arc_ksp = NULL;
5154         }
5155 
5156         mutex_destroy(&arc_reclaim_lock);
5157         cv_destroy(&arc_reclaim_thread_cv);
5158         cv_destroy(&arc_reclaim_waiters_cv);
5159 
5160         mutex_destroy(&arc_user_evicts_lock);
5161         cv_destroy(&arc_user_evicts_cv);
5162 
5163         refcount_destroy(&arc_anon->arcs_size);
5164         refcount_destroy(&arc_mru->arcs_size);
5165         refcount_destroy(&arc_mru_ghost->arcs_size);
5166         refcount_destroy(&arc_mfu->arcs_size);
5167         refcount_destroy(&arc_mfu_ghost->arcs_size);
5168         refcount_destroy(&arc_l2c_only->arcs_size);
5169 
5170         multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
5171         multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
5172         multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
5173         multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
5174         multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
5175         multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
5176         multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
5177         multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
5178 
5179         buf_fini();
5180 
5181         ASSERT0(arc_loaned_bytes);
5182 }
5183 
5184 /*
5185  * Level 2 ARC
5186  *
5187  * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
5188  * It uses dedicated storage devices to hold cached data, which are populated
5189  * using large infrequent writes.  The main role of this cache is to boost
5190  * the performance of random read workloads.  The intended L2ARC devices
5191  * include short-stroked disks, solid state disks, and other media with
5192  * substantially faster read latency than disk.
5193  *
5194  *                 +-----------------------+
5195  *                 |         ARC           |
5196  *                 +-----------------------+
5197  *                    |         ^     ^
5198  *                    |         |     |
5199  *      l2arc_feed_thread()    arc_read()
5200  *                    |         |     |
5201  *                    |  l2arc read   |
5202  *                    V         |     |
5203  *               +---------------+    |
5204  *               |     L2ARC     |    |
5205  *               +---------------+    |
5206  *                   |    ^           |
5207  *          l2arc_write() |           |
5208  *                   |    |           |
5209  *                   V    |           |
5210  *                 +-------+      +-------+
5211  *                 | vdev  |      | vdev  |
5212  *                 | cache |      | cache |
5213  *                 +-------+      +-------+
5214  *                 +=========+     .-----.
5215  *                 :  L2ARC  :    |-_____-|
5216  *                 : devices :    | Disks |
5217  *                 +=========+    `-_____-'
5218  *
5219  * Read requests are satisfied from the following sources, in order:
5220  *
5221  *      1) ARC
5222  *      2) vdev cache of L2ARC devices
5223  *      3) L2ARC devices
5224  *      4) vdev cache of disks
5225  *      5) disks
5226  *
5227  * Some L2ARC device types exhibit extremely slow write performance.
5228  * To accommodate for this there are some significant differences between
5229  * the L2ARC and traditional cache design:
5230  *
5231  * 1. There is no eviction path from the ARC to the L2ARC.  Evictions from
5232  * the ARC behave as usual, freeing buffers and placing headers on ghost
5233  * lists.  The ARC does not send buffers to the L2ARC during eviction as
5234  * this would add inflated write latencies for all ARC memory pressure.
5235  *
5236  * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
5237  * It does this by periodically scanning buffers from the eviction-end of
5238  * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
5239  * not already there. It scans until a headroom of buffers is satisfied,
5240  * which itself is a buffer for ARC eviction. If a compressible buffer is
5241  * found during scanning and selected for writing to an L2ARC device, we
5242  * temporarily boost scanning headroom during the next scan cycle to make
5243  * sure we adapt to compression effects (which might significantly reduce
5244  * the data volume we write to L2ARC). The thread that does this is
5245  * l2arc_feed_thread(), illustrated below; example sizes are included to
5246  * provide a better sense of ratio than this diagram:
5247  *
5248  *             head -->                        tail
5249  *              +---------------------+----------+
5250  *      ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->.   # already on L2ARC
5251  *              +---------------------+----------+   |   o L2ARC eligible
5252  *      ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->|   : ARC buffer
5253  *              +---------------------+----------+   |
5254  *                   15.9 Gbytes      ^ 32 Mbytes    |
5255  *                                 headroom          |
5256  *                                            l2arc_feed_thread()
5257  *                                                   |
5258  *                       l2arc write hand <--[oooo]--'
5259  *                               |           8 Mbyte
5260  *                               |          write max
5261  *                               V
5262  *                +==============================+
5263  *      L2ARC dev |####|#|###|###|    |####| ... |
5264  *                +==============================+
5265  *                           32 Gbytes
5266  *
5267  * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
5268  * evicted, then the L2ARC has cached a buffer much sooner than it probably
5269  * needed to, potentially wasting L2ARC device bandwidth and storage.  It is
5270  * safe to say that this is an uncommon case, since buffers at the end of
5271  * the ARC lists have moved there due to inactivity.
5272  *
5273  * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
5274  * then the L2ARC simply misses copying some buffers.  This serves as a
5275  * pressure valve to prevent heavy read workloads from both stalling the ARC
5276  * with waits and clogging the L2ARC with writes.  This also helps prevent
5277  * the potential for the L2ARC to churn if it attempts to cache content too
5278  * quickly, such as during backups of the entire pool.
5279  *
5280  * 5. After system boot and before the ARC has filled main memory, there are
5281  * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
5282  * lists can remain mostly static.  Instead of searching from tail of these
5283  * lists as pictured, the l2arc_feed_thread() will search from the list heads
5284  * for eligible buffers, greatly increasing its chance of finding them.
5285  *
5286  * The L2ARC device write speed is also boosted during this time so that
5287  * the L2ARC warms up faster.  Since there have been no ARC evictions yet,
5288  * there are no L2ARC reads, and no fear of degrading read performance
5289  * through increased writes.
5290  *
5291  * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
5292  * the vdev queue can aggregate them into larger and fewer writes.  Each
5293  * device is written to in a rotor fashion, sweeping writes through
5294  * available space then repeating.
5295  *
5296  * 7. The L2ARC does not store dirty content.  It never needs to flush
5297  * write buffers back to disk based storage.
5298  *
5299  * 8. If an ARC buffer is written (and dirtied) which also exists in the
5300  * L2ARC, the now stale L2ARC buffer is immediately dropped.
5301  *
5302  * The performance of the L2ARC can be tweaked by a number of tunables, which
5303  * may be necessary for different workloads:
5304  *
5305  *      l2arc_write_max         max write bytes per interval
5306  *      l2arc_write_boost       extra write bytes during device warmup
5307  *      l2arc_noprefetch        skip caching prefetched buffers
5308  *      l2arc_headroom          number of max device writes to precache
5309  *      l2arc_headroom_boost    when we find compressed buffers during ARC
5310  *                              scanning, we multiply headroom by this
5311  *                              percentage factor for the next scan cycle,
5312  *                              since more compressed buffers are likely to
5313  *                              be present
5314  *      l2arc_feed_secs         seconds between L2ARC writing
5315  *
5316  * Tunables may be removed or added as future performance improvements are
5317  * integrated, and also may become zpool properties.
5318  *
5319  * There are three key functions that control how the L2ARC warms up:
5320  *
5321  *      l2arc_write_eligible()  check if a buffer is eligible to cache
5322  *      l2arc_write_size()      calculate how much to write
5323  *      l2arc_write_interval()  calculate sleep delay between writes
5324  *
5325  * These three functions determine what to write, how much, and how quickly
5326  * to send writes.
5327  */
5328 
5329 static boolean_t
5330 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
5331 {
5332         /*
5333          * A buffer is *not* eligible for the L2ARC if it:
5334          * 1. belongs to a different spa.
5335          * 2. is already cached on the L2ARC.
5336          * 3. has an I/O in progress (it may be an incomplete read).
5337          * 4. is flagged not eligible (zfs property).
5338          */
5339         if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
5340             HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
5341                 return (B_FALSE);
5342 
5343         return (B_TRUE);
5344 }
5345 
5346 static uint64_t
5347 l2arc_write_size(void)
5348 {
5349         uint64_t size;
5350 
5351         /*
5352          * Make sure our globals have meaningful values in case the user
5353          * altered them.
5354          */
5355         size = l2arc_write_max;
5356         if (size == 0) {
5357                 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
5358                     "be greater than zero, resetting it to the default (%d)",
5359                     L2ARC_WRITE_SIZE);
5360                 size = l2arc_write_max = L2ARC_WRITE_SIZE;
5361         }
5362 
5363         if (arc_warm == B_FALSE)
5364                 size += l2arc_write_boost;
5365 
5366         return (size);
5367 
5368 }
5369 
5370 static clock_t
5371 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
5372 {
5373         clock_t interval, next, now;
5374 
5375         /*
5376          * If the ARC lists are busy, increase our write rate; if the
5377          * lists are stale, idle back.  This is achieved by checking
5378          * how much we previously wrote - if it was more than half of
5379          * what we wanted, schedule the next write much sooner.
5380          */
5381         if (l2arc_feed_again && wrote > (wanted / 2))
5382                 interval = (hz * l2arc_feed_min_ms) / 1000;
5383         else
5384                 interval = hz * l2arc_feed_secs;
5385 
5386         now = ddi_get_lbolt();
5387         next = MAX(now, MIN(now + interval, began + interval));
5388 
5389         return (next);
5390 }
5391 
5392 /*
5393  * Cycle through L2ARC devices.  This is how L2ARC load balances.
5394  * If a device is returned, this also returns holding the spa config lock.
5395  */
5396 static l2arc_dev_t *
5397 l2arc_dev_get_next(void)
5398 {
5399         l2arc_dev_t *first, *next = NULL;
5400 
5401         /*
5402          * Lock out the removal of spas (spa_namespace_lock), then removal
5403          * of cache devices (l2arc_dev_mtx).  Once a device has been selected,
5404          * both locks will be dropped and a spa config lock held instead.
5405          */
5406         mutex_enter(&spa_namespace_lock);
5407         mutex_enter(&l2arc_dev_mtx);
5408 
5409         /* if there are no vdevs, there is nothing to do */
5410         if (l2arc_ndev == 0)
5411                 goto out;
5412 
5413         first = NULL;
5414         next = l2arc_dev_last;
5415         do {
5416                 /* loop around the list looking for a non-faulted vdev */
5417                 if (next == NULL) {
5418                         next = list_head(l2arc_dev_list);
5419                 } else {
5420                         next = list_next(l2arc_dev_list, next);
5421                         if (next == NULL)
5422                                 next = list_head(l2arc_dev_list);
5423                 }
5424 
5425                 /* if we have come back to the start, bail out */
5426                 if (first == NULL)
5427                         first = next;
5428                 else if (next == first)
5429                         break;
5430 
5431         } while (vdev_is_dead(next->l2ad_vdev));
5432 
5433         /* if we were unable to find any usable vdevs, return NULL */
5434         if (vdev_is_dead(next->l2ad_vdev))
5435                 next = NULL;
5436 
5437         l2arc_dev_last = next;
5438 
5439 out:
5440         mutex_exit(&l2arc_dev_mtx);
5441 
5442         /*
5443          * Grab the config lock to prevent the 'next' device from being
5444          * removed while we are writing to it.
5445          */
5446         if (next != NULL)
5447                 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
5448         mutex_exit(&spa_namespace_lock);
5449 
5450         return (next);
5451 }
5452 
5453 /*
5454  * Free buffers that were tagged for destruction.
5455  */
5456 static void
5457 l2arc_do_free_on_write()
5458 {
5459         list_t *buflist;
5460         l2arc_data_free_t *df, *df_prev;
5461 
5462         mutex_enter(&l2arc_free_on_write_mtx);
5463         buflist = l2arc_free_on_write;
5464 
5465         for (df = list_tail(buflist); df; df = df_prev) {
5466                 df_prev = list_prev(buflist, df);
5467                 ASSERT(df->l2df_data != NULL);
5468                 ASSERT(df->l2df_func != NULL);
5469                 df->l2df_func(df->l2df_data, df->l2df_size);
5470                 list_remove(buflist, df);
5471                 kmem_free(df, sizeof (l2arc_data_free_t));
5472         }
5473 
5474         mutex_exit(&l2arc_free_on_write_mtx);
5475 }
5476 
5477 /*
5478  * A write to a cache device has completed.  Update all headers to allow
5479  * reads from these buffers to begin.
5480  */
5481 static void
5482 l2arc_write_done(zio_t *zio)
5483 {
5484         l2arc_write_callback_t *cb;
5485         l2arc_dev_t *dev;
5486         list_t *buflist;
5487         arc_buf_hdr_t *head, *hdr, *hdr_prev;
5488         kmutex_t *hash_lock;
5489         int64_t bytes_dropped = 0;
5490 
5491         cb = zio->io_private;
5492         ASSERT(cb != NULL);
5493         dev = cb->l2wcb_dev;
5494         ASSERT(dev != NULL);
5495         head = cb->l2wcb_head;
5496         ASSERT(head != NULL);
5497         buflist = &dev->l2ad_buflist;
5498         ASSERT(buflist != NULL);
5499         DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
5500             l2arc_write_callback_t *, cb);
5501 
5502         if (zio->io_error != 0)
5503                 ARCSTAT_BUMP(arcstat_l2_writes_error);
5504 
5505         /*
5506          * All writes completed, or an error was hit.
5507          */
5508 top:
5509         mutex_enter(&dev->l2ad_mtx);
5510         for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
5511                 hdr_prev = list_prev(buflist, hdr);
5512 
5513                 hash_lock = HDR_LOCK(hdr);
5514 
5515                 /*
5516                  * We cannot use mutex_enter or else we can deadlock
5517                  * with l2arc_write_buffers (due to swapping the order
5518                  * the hash lock and l2ad_mtx are taken).
5519                  */
5520                 if (!mutex_tryenter(hash_lock)) {
5521                         /*
5522                          * Missed the hash lock. We must retry so we
5523                          * don't leave the ARC_FLAG_L2_WRITING bit set.
5524                          */
5525                         ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
5526 
5527                         /*
5528                          * We don't want to rescan the headers we've
5529                          * already marked as having been written out, so
5530                          * we reinsert the head node so we can pick up
5531                          * where we left off.
5532                          */
5533                         list_remove(buflist, head);
5534                         list_insert_after(buflist, hdr, head);
5535 
5536                         mutex_exit(&dev->l2ad_mtx);
5537 
5538                         /*
5539                          * We wait for the hash lock to become available
5540                          * to try and prevent busy waiting, and increase
5541                          * the chance we'll be able to acquire the lock
5542                          * the next time around.
5543                          */
5544                         mutex_enter(hash_lock);
5545                         mutex_exit(hash_lock);
5546                         goto top;
5547                 }
5548 
5549                 /*
5550                  * We could not have been moved into the arc_l2c_only
5551                  * state while in-flight due to our ARC_FLAG_L2_WRITING
5552                  * bit being set. Let's just ensure that's being enforced.
5553                  */
5554                 ASSERT(HDR_HAS_L1HDR(hdr));
5555 
5556                 /*
5557                  * We may have allocated a buffer for L2ARC compression,
5558                  * we must release it to avoid leaking this data.
5559                  */
5560                 l2arc_release_cdata_buf(hdr);
5561 
5562                 if (zio->io_error != 0) {
5563                         /*
5564                          * Error - drop L2ARC entry.
5565                          */
5566                         list_remove(buflist, hdr);
5567                         hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
5568 
5569                         ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize);
5570                         ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
5571 
5572                         bytes_dropped += hdr->b_l2hdr.b_asize;
5573                         (void) refcount_remove_many(&dev->l2ad_alloc,
5574                             hdr->b_l2hdr.b_asize, hdr);
5575                 }
5576 
5577                 /*
5578                  * Allow ARC to begin reads and ghost list evictions to
5579                  * this L2ARC entry.
5580                  */
5581                 hdr->b_flags &= ~ARC_FLAG_L2_WRITING;
5582 
5583                 mutex_exit(hash_lock);
5584         }
5585 
5586         atomic_inc_64(&l2arc_writes_done);
5587         list_remove(buflist, head);
5588         ASSERT(!HDR_HAS_L1HDR(head));
5589         kmem_cache_free(hdr_l2only_cache, head);
5590         mutex_exit(&dev->l2ad_mtx);
5591 
5592         vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
5593 
5594         l2arc_do_free_on_write();
5595 
5596         kmem_free(cb, sizeof (l2arc_write_callback_t));
5597 }
5598 
5599 /*
5600  * A read to a cache device completed.  Validate buffer contents before
5601  * handing over to the regular ARC routines.
5602  */
5603 static void
5604 l2arc_read_done(zio_t *zio)
5605 {
5606         l2arc_read_callback_t *cb;
5607         arc_buf_hdr_t *hdr;
5608         arc_buf_t *buf;
5609         kmutex_t *hash_lock;
5610         int equal;
5611 
5612         ASSERT(zio->io_vd != NULL);
5613         ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
5614 
5615         spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
5616 
5617         cb = zio->io_private;
5618         ASSERT(cb != NULL);
5619         buf = cb->l2rcb_buf;
5620         ASSERT(buf != NULL);
5621 
5622         hash_lock = HDR_LOCK(buf->b_hdr);
5623         mutex_enter(hash_lock);
5624         hdr = buf->b_hdr;
5625         ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5626 
5627         /*
5628          * If the buffer was compressed, decompress it first.
5629          */
5630         if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
5631                 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
5632         ASSERT(zio->io_data != NULL);
5633 
5634         /*
5635          * Check this survived the L2ARC journey.
5636          */
5637         equal = arc_cksum_equal(buf);
5638         if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
5639                 mutex_exit(hash_lock);
5640                 zio->io_private = buf;
5641                 zio->io_bp_copy = cb->l2rcb_bp;   /* XXX fix in L2ARC 2.0 */
5642                 zio->io_bp = &zio->io_bp_copy;        /* XXX fix in L2ARC 2.0 */
5643                 arc_read_done(zio);
5644         } else {
5645                 mutex_exit(hash_lock);
5646                 /*
5647                  * Buffer didn't survive caching.  Increment stats and
5648                  * reissue to the original storage device.
5649                  */
5650                 if (zio->io_error != 0) {
5651                         ARCSTAT_BUMP(arcstat_l2_io_error);
5652                 } else {
5653                         zio->io_error = SET_ERROR(EIO);
5654                 }
5655                 if (!equal)
5656                         ARCSTAT_BUMP(arcstat_l2_cksum_bad);
5657 
5658                 /*
5659                  * If there's no waiter, issue an async i/o to the primary
5660                  * storage now.  If there *is* a waiter, the caller must
5661                  * issue the i/o in a context where it's OK to block.
5662                  */
5663                 if (zio->io_waiter == NULL) {
5664                         zio_t *pio = zio_unique_parent(zio);
5665 
5666                         ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
5667 
5668                         zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
5669                             buf->b_data, zio->io_size, arc_read_done, buf,
5670                             zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
5671                 }
5672         }
5673 
5674         kmem_free(cb, sizeof (l2arc_read_callback_t));
5675 }
5676 
5677 /*
5678  * This is the list priority from which the L2ARC will search for pages to
5679  * cache.  This is used within loops (0..3) to cycle through lists in the
5680  * desired order.  This order can have a significant effect on cache
5681  * performance.
5682  *
5683  * Currently the metadata lists are hit first, MFU then MRU, followed by
5684  * the data lists.  This function returns a locked list, and also returns
5685  * the lock pointer.
5686  */
5687 static multilist_sublist_t *
5688 l2arc_sublist_lock(int list_num)
5689 {
5690         multilist_t *ml = NULL;
5691         unsigned int idx;
5692 
5693         ASSERT(list_num >= 0 && list_num <= 3);
5694 
5695         switch (list_num) {
5696         case 0:
5697                 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
5698                 break;
5699         case 1:
5700                 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
5701                 break;
5702         case 2:
5703                 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
5704                 break;
5705         case 3:
5706                 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
5707                 break;
5708         }
5709 
5710         /*
5711          * Return a randomly-selected sublist. This is acceptable
5712          * because the caller feeds only a little bit of data for each
5713          * call (8MB). Subsequent calls will result in different
5714          * sublists being selected.
5715          */
5716         idx = multilist_get_random_index(ml);
5717         return (multilist_sublist_lock(ml, idx));
5718 }
5719 
5720 /*
5721  * Evict buffers from the device write hand to the distance specified in
5722  * bytes.  This distance may span populated buffers, it may span nothing.
5723  * This is clearing a region on the L2ARC device ready for writing.
5724  * If the 'all' boolean is set, every buffer is evicted.
5725  */
5726 static void
5727 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
5728 {
5729         list_t *buflist;
5730         arc_buf_hdr_t *hdr, *hdr_prev;
5731         kmutex_t *hash_lock;
5732         uint64_t taddr;
5733 
5734         buflist = &dev->l2ad_buflist;
5735 
5736         if (!all && dev->l2ad_first) {
5737                 /*
5738                  * This is the first sweep through the device.  There is
5739                  * nothing to evict.
5740                  */
5741                 return;
5742         }
5743 
5744         if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
5745                 /*
5746                  * When nearing the end of the device, evict to the end
5747                  * before the device write hand jumps to the start.
5748                  */
5749                 taddr = dev->l2ad_end;
5750         } else {
5751                 taddr = dev->l2ad_hand + distance;
5752         }
5753         DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
5754             uint64_t, taddr, boolean_t, all);
5755 
5756 top:
5757         mutex_enter(&dev->l2ad_mtx);
5758         for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
5759                 hdr_prev = list_prev(buflist, hdr);
5760 
5761                 hash_lock = HDR_LOCK(hdr);
5762 
5763                 /*
5764                  * We cannot use mutex_enter or else we can deadlock
5765                  * with l2arc_write_buffers (due to swapping the order
5766                  * the hash lock and l2ad_mtx are taken).
5767                  */
5768                 if (!mutex_tryenter(hash_lock)) {
5769                         /*
5770                          * Missed the hash lock.  Retry.
5771                          */
5772                         ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
5773                         mutex_exit(&dev->l2ad_mtx);
5774                         mutex_enter(hash_lock);
5775                         mutex_exit(hash_lock);
5776                         goto top;
5777                 }
5778 
5779                 if (HDR_L2_WRITE_HEAD(hdr)) {
5780                         /*
5781                          * We hit a write head node.  Leave it for
5782                          * l2arc_write_done().
5783                          */
5784                         list_remove(buflist, hdr);
5785                         mutex_exit(hash_lock);
5786                         continue;
5787                 }
5788 
5789                 if (!all && HDR_HAS_L2HDR(hdr) &&
5790                     (hdr->b_l2hdr.b_daddr > taddr ||
5791                     hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
5792                         /*
5793                          * We've evicted to the target address,
5794                          * or the end of the device.
5795                          */
5796                         mutex_exit(hash_lock);
5797                         break;
5798                 }
5799 
5800                 ASSERT(HDR_HAS_L2HDR(hdr));
5801                 if (!HDR_HAS_L1HDR(hdr)) {
5802                         ASSERT(!HDR_L2_READING(hdr));
5803                         /*
5804                          * This doesn't exist in the ARC.  Destroy.
5805                          * arc_hdr_destroy() will call list_remove()
5806                          * and decrement arcstat_l2_size.
5807                          */
5808                         arc_change_state(arc_anon, hdr, hash_lock);
5809                         arc_hdr_destroy(hdr);
5810                 } else {
5811                         ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
5812                         ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
5813                         /*
5814                          * Invalidate issued or about to be issued
5815                          * reads, since we may be about to write
5816                          * over this location.
5817                          */
5818                         if (HDR_L2_READING(hdr)) {
5819                                 ARCSTAT_BUMP(arcstat_l2_evict_reading);
5820                                 hdr->b_flags |= ARC_FLAG_L2_EVICTED;
5821                         }
5822 
5823                         /* Ensure this header has finished being written */
5824                         ASSERT(!HDR_L2_WRITING(hdr));
5825                         ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
5826 
5827                         arc_hdr_l2hdr_destroy(hdr);
5828                 }
5829                 mutex_exit(hash_lock);
5830         }
5831         mutex_exit(&dev->l2ad_mtx);
5832 }
5833 
5834 /*
5835  * Find and write ARC buffers to the L2ARC device.
5836  *
5837  * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
5838  * for reading until they have completed writing.
5839  * The headroom_boost is an in-out parameter used to maintain headroom boost
5840  * state between calls to this function.
5841  *
5842  * Returns the number of bytes actually written (which may be smaller than
5843  * the delta by which the device hand has changed due to alignment).
5844  */
5845 static uint64_t
5846 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
5847     boolean_t *headroom_boost)
5848 {
5849         arc_buf_hdr_t *hdr, *hdr_prev, *head;
5850         uint64_t write_asize, write_psize, write_sz, headroom,
5851             buf_compress_minsz;
5852         void *buf_data;
5853         boolean_t full;
5854         l2arc_write_callback_t *cb;
5855         zio_t *pio, *wzio;
5856         uint64_t guid = spa_load_guid(spa);
5857         const boolean_t do_headroom_boost = *headroom_boost;
5858 
5859         ASSERT(dev->l2ad_vdev != NULL);
5860 
5861         /* Lower the flag now, we might want to raise it again later. */
5862         *headroom_boost = B_FALSE;
5863 
5864         pio = NULL;
5865         write_sz = write_asize = write_psize = 0;
5866         full = B_FALSE;
5867         head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
5868         head->b_flags |= ARC_FLAG_L2_WRITE_HEAD;
5869         head->b_flags |= ARC_FLAG_HAS_L2HDR;
5870 
5871         /*
5872          * We will want to try to compress buffers that are at least 2x the
5873          * device sector size.
5874          */
5875         buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
5876 
5877         /*
5878          * Copy buffers for L2ARC writing.
5879          */
5880         for (int try = 0; try <= 3; try++) {
5881                 multilist_sublist_t *mls = l2arc_sublist_lock(try);
5882                 uint64_t passed_sz = 0;
5883 
5884                 /*
5885                  * L2ARC fast warmup.
5886                  *
5887                  * Until the ARC is warm and starts to evict, read from the
5888                  * head of the ARC lists rather than the tail.
5889                  */
5890                 if (arc_warm == B_FALSE)
5891                         hdr = multilist_sublist_head(mls);
5892                 else
5893                         hdr = multilist_sublist_tail(mls);
5894 
5895                 headroom = target_sz * l2arc_headroom;
5896                 if (do_headroom_boost)
5897                         headroom = (headroom * l2arc_headroom_boost) / 100;
5898 
5899                 for (; hdr; hdr = hdr_prev) {
5900                         kmutex_t *hash_lock;
5901                         uint64_t buf_sz;
5902 
5903                         if (arc_warm == B_FALSE)
5904                                 hdr_prev = multilist_sublist_next(mls, hdr);
5905                         else
5906                                 hdr_prev = multilist_sublist_prev(mls, hdr);
5907 
5908                         hash_lock = HDR_LOCK(hdr);
5909                         if (!mutex_tryenter(hash_lock)) {
5910                                 /*
5911                                  * Skip this buffer rather than waiting.
5912                                  */
5913                                 continue;
5914                         }
5915 
5916                         passed_sz += hdr->b_size;
5917                         if (passed_sz > headroom) {
5918                                 /*
5919                                  * Searched too far.
5920                                  */
5921                                 mutex_exit(hash_lock);
5922                                 break;
5923                         }
5924 
5925                         if (!l2arc_write_eligible(guid, hdr)) {
5926                                 mutex_exit(hash_lock);
5927                                 continue;
5928                         }
5929 
5930                         if ((write_sz + hdr->b_size) > target_sz) {
5931                                 full = B_TRUE;
5932                                 mutex_exit(hash_lock);
5933                                 break;
5934                         }
5935 
5936                         if (pio == NULL) {
5937                                 /*
5938                                  * Insert a dummy header on the buflist so
5939                                  * l2arc_write_done() can find where the
5940                                  * write buffers begin without searching.
5941                                  */
5942                                 mutex_enter(&dev->l2ad_mtx);
5943                                 list_insert_head(&dev->l2ad_buflist, head);
5944                                 mutex_exit(&dev->l2ad_mtx);
5945 
5946                                 cb = kmem_alloc(
5947                                     sizeof (l2arc_write_callback_t), KM_SLEEP);
5948                                 cb->l2wcb_dev = dev;
5949                                 cb->l2wcb_head = head;
5950                                 pio = zio_root(spa, l2arc_write_done, cb,
5951                                     ZIO_FLAG_CANFAIL);
5952                         }
5953 
5954                         /*
5955                          * Create and add a new L2ARC header.
5956                          */
5957                         hdr->b_l2hdr.b_dev = dev;
5958                         hdr->b_flags |= ARC_FLAG_L2_WRITING;
5959                         /*
5960                          * Temporarily stash the data buffer in b_tmp_cdata.
5961                          * The subsequent write step will pick it up from
5962                          * there. This is because can't access b_l1hdr.b_buf
5963                          * without holding the hash_lock, which we in turn
5964                          * can't access without holding the ARC list locks
5965                          * (which we want to avoid during compression/writing).
5966                          */
5967                         HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
5968                         hdr->b_l2hdr.b_asize = hdr->b_size;
5969                         hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data;
5970 
5971                         /*
5972                          * Explicitly set the b_daddr field to a known
5973                          * value which means "invalid address". This
5974                          * enables us to differentiate which stage of
5975                          * l2arc_write_buffers() the particular header
5976                          * is in (e.g. this loop, or the one below).
5977                          * ARC_FLAG_L2_WRITING is not enough to make
5978                          * this distinction, and we need to know in
5979                          * order to do proper l2arc vdev accounting in
5980                          * arc_release() and arc_hdr_destroy().
5981                          *
5982                          * Note, we can't use a new flag to distinguish
5983                          * the two stages because we don't hold the
5984                          * header's hash_lock below, in the second stage
5985                          * of this function. Thus, we can't simply
5986                          * change the b_flags field to denote that the
5987                          * IO has been sent. We can change the b_daddr
5988                          * field of the L2 portion, though, since we'll
5989                          * be holding the l2ad_mtx; which is why we're
5990                          * using it to denote the header's state change.
5991                          */
5992                         hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET;
5993 
5994                         buf_sz = hdr->b_size;
5995                         hdr->b_flags |= ARC_FLAG_HAS_L2HDR;
5996 
5997                         mutex_enter(&dev->l2ad_mtx);
5998                         list_insert_head(&dev->l2ad_buflist, hdr);
5999                         mutex_exit(&dev->l2ad_mtx);
6000 
6001                         /*
6002                          * Compute and store the buffer cksum before
6003                          * writing.  On debug the cksum is verified first.
6004                          */
6005                         arc_cksum_verify(hdr->b_l1hdr.b_buf);
6006                         arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE);
6007 
6008                         mutex_exit(hash_lock);
6009 
6010                         write_sz += buf_sz;
6011                 }
6012 
6013                 multilist_sublist_unlock(mls);
6014 
6015                 if (full == B_TRUE)
6016                         break;
6017         }
6018 
6019         /* No buffers selected for writing? */
6020         if (pio == NULL) {
6021                 ASSERT0(write_sz);
6022                 ASSERT(!HDR_HAS_L1HDR(head));
6023                 kmem_cache_free(hdr_l2only_cache, head);
6024                 return (0);
6025         }
6026 
6027         mutex_enter(&dev->l2ad_mtx);
6028 
6029         /*
6030          * Now start writing the buffers. We're starting at the write head
6031          * and work backwards, retracing the course of the buffer selector
6032          * loop above.
6033          */
6034         for (hdr = list_prev(&dev->l2ad_buflist, head); hdr;
6035             hdr = list_prev(&dev->l2ad_buflist, hdr)) {
6036                 uint64_t buf_sz;
6037 
6038                 /*
6039                  * We rely on the L1 portion of the header below, so
6040                  * it's invalid for this header to have been evicted out
6041                  * of the ghost cache, prior to being written out. The
6042                  * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6043                  */
6044                 ASSERT(HDR_HAS_L1HDR(hdr));
6045 
6046                 /*
6047                  * We shouldn't need to lock the buffer here, since we flagged
6048                  * it as ARC_FLAG_L2_WRITING in the previous step, but we must
6049                  * take care to only access its L2 cache parameters. In
6050                  * particular, hdr->l1hdr.b_buf may be invalid by now due to
6051                  * ARC eviction.
6052                  */
6053                 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
6054 
6055                 if ((HDR_L2COMPRESS(hdr)) &&
6056                     hdr->b_l2hdr.b_asize >= buf_compress_minsz) {
6057                         if (l2arc_compress_buf(hdr)) {
6058                                 /*
6059                                  * If compression succeeded, enable headroom
6060                                  * boost on the next scan cycle.
6061                                  */
6062                                 *headroom_boost = B_TRUE;
6063                         }
6064                 }
6065 
6066                 /*
6067                  * Pick up the buffer data we had previously stashed away
6068                  * (and now potentially also compressed).
6069                  */
6070                 buf_data = hdr->b_l1hdr.b_tmp_cdata;
6071                 buf_sz = hdr->b_l2hdr.b_asize;
6072 
6073                 /*
6074                  * We need to do this regardless if buf_sz is zero or
6075                  * not, otherwise, when this l2hdr is evicted we'll
6076                  * remove a reference that was never added.
6077                  */
6078                 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr);
6079 
6080                 /* Compression may have squashed the buffer to zero length. */
6081                 if (buf_sz != 0) {
6082                         uint64_t buf_p_sz;
6083 
6084                         wzio = zio_write_phys(pio, dev->l2ad_vdev,
6085                             dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
6086                             NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
6087                             ZIO_FLAG_CANFAIL, B_FALSE);
6088 
6089                         DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
6090                             zio_t *, wzio);
6091                         (void) zio_nowait(wzio);
6092 
6093                         write_asize += buf_sz;
6094 
6095                         /*
6096                          * Keep the clock hand suitably device-aligned.
6097                          */
6098                         buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
6099                         write_psize += buf_p_sz;
6100                         dev->l2ad_hand += buf_p_sz;
6101                 }
6102         }
6103 
6104         mutex_exit(&dev->l2ad_mtx);
6105 
6106         ASSERT3U(write_asize, <=, target_sz);
6107         ARCSTAT_BUMP(arcstat_l2_writes_sent);
6108         ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
6109         ARCSTAT_INCR(arcstat_l2_size, write_sz);
6110         ARCSTAT_INCR(arcstat_l2_asize, write_asize);
6111         vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0);
6112 
6113         /*
6114          * Bump device hand to the device start if it is approaching the end.
6115          * l2arc_evict() will already have evicted ahead for this case.
6116          */
6117         if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
6118                 dev->l2ad_hand = dev->l2ad_start;
6119                 dev->l2ad_first = B_FALSE;
6120         }
6121 
6122         dev->l2ad_writing = B_TRUE;
6123         (void) zio_wait(pio);
6124         dev->l2ad_writing = B_FALSE;
6125 
6126         return (write_asize);
6127 }
6128 
6129 /*
6130  * Compresses an L2ARC buffer.
6131  * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its
6132  * size in l2hdr->b_asize. This routine tries to compress the data and
6133  * depending on the compression result there are three possible outcomes:
6134  * *) The buffer was incompressible. The original l2hdr contents were left
6135  *    untouched and are ready for writing to an L2 device.
6136  * *) The buffer was all-zeros, so there is no need to write it to an L2
6137  *    device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
6138  *    set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
6139  * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
6140  *    data buffer which holds the compressed data to be written, and b_asize
6141  *    tells us how much data there is. b_compress is set to the appropriate
6142  *    compression algorithm. Once writing is done, invoke
6143  *    l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
6144  *
6145  * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
6146  * buffer was incompressible).
6147  */
6148 static boolean_t
6149 l2arc_compress_buf(arc_buf_hdr_t *hdr)
6150 {
6151         void *cdata;
6152         size_t csize, len, rounded;
6153         ASSERT(HDR_HAS_L2HDR(hdr));
6154         l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
6155 
6156         ASSERT(HDR_HAS_L1HDR(hdr));
6157         ASSERT(HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF);
6158         ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6159 
6160         len = l2hdr->b_asize;
6161         cdata = zio_data_buf_alloc(len);
6162         ASSERT3P(cdata, !=, NULL);
6163         csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata,
6164             cdata, l2hdr->b_asize);
6165 
6166         rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE);
6167         if (rounded > csize) {
6168                 bzero((char *)cdata + csize, rounded - csize);
6169                 csize = rounded;
6170         }
6171 
6172         if (csize == 0) {
6173                 /* zero block, indicate that there's nothing to write */
6174                 zio_data_buf_free(cdata, len);
6175                 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_EMPTY);
6176                 l2hdr->b_asize = 0;
6177                 hdr->b_l1hdr.b_tmp_cdata = NULL;
6178                 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
6179                 return (B_TRUE);
6180         } else if (csize > 0 && csize < len) {
6181                 /*
6182                  * Compression succeeded, we'll keep the cdata around for
6183                  * writing and release it afterwards.
6184                  */
6185                 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_LZ4);
6186                 l2hdr->b_asize = csize;
6187                 hdr->b_l1hdr.b_tmp_cdata = cdata;
6188                 ARCSTAT_BUMP(arcstat_l2_compress_successes);
6189                 return (B_TRUE);
6190         } else {
6191                 /*
6192                  * Compression failed, release the compressed buffer.
6193                  * l2hdr will be left unmodified.
6194                  */
6195                 zio_data_buf_free(cdata, len);
6196                 ARCSTAT_BUMP(arcstat_l2_compress_failures);
6197                 return (B_FALSE);
6198         }
6199 }
6200 
6201 /*
6202  * Decompresses a zio read back from an l2arc device. On success, the
6203  * underlying zio's io_data buffer is overwritten by the uncompressed
6204  * version. On decompression error (corrupt compressed stream), the
6205  * zio->io_error value is set to signal an I/O error.
6206  *
6207  * Please note that the compressed data stream is not checksummed, so
6208  * if the underlying device is experiencing data corruption, we may feed
6209  * corrupt data to the decompressor, so the decompressor needs to be
6210  * able to handle this situation (LZ4 does).
6211  */
6212 static void
6213 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
6214 {
6215         ASSERT(L2ARC_IS_VALID_COMPRESS(c));
6216 
6217         if (zio->io_error != 0) {
6218                 /*
6219                  * An io error has occured, just restore the original io
6220                  * size in preparation for a main pool read.
6221                  */
6222                 zio->io_orig_size = zio->io_size = hdr->b_size;
6223                 return;
6224         }
6225 
6226         if (c == ZIO_COMPRESS_EMPTY) {
6227                 /*
6228                  * An empty buffer results in a null zio, which means we
6229                  * need to fill its io_data after we're done restoring the
6230                  * buffer's contents.
6231                  */
6232                 ASSERT(hdr->b_l1hdr.b_buf != NULL);
6233                 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size);
6234                 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data;
6235         } else {
6236                 ASSERT(zio->io_data != NULL);
6237                 /*
6238                  * We copy the compressed data from the start of the arc buffer
6239                  * (the zio_read will have pulled in only what we need, the
6240                  * rest is garbage which we will overwrite at decompression)
6241                  * and then decompress back to the ARC data buffer. This way we
6242                  * can minimize copying by simply decompressing back over the
6243                  * original compressed data (rather than decompressing to an
6244                  * aux buffer and then copying back the uncompressed buffer,
6245                  * which is likely to be much larger).
6246                  */
6247                 uint64_t csize;
6248                 void *cdata;
6249 
6250                 csize = zio->io_size;
6251                 cdata = zio_data_buf_alloc(csize);
6252                 bcopy(zio->io_data, cdata, csize);
6253                 if (zio_decompress_data(c, cdata, zio->io_data, csize,
6254                     hdr->b_size) != 0)
6255                         zio->io_error = EIO;
6256                 zio_data_buf_free(cdata, csize);
6257         }
6258 
6259         /* Restore the expected uncompressed IO size. */
6260         zio->io_orig_size = zio->io_size = hdr->b_size;
6261 }
6262 
6263 /*
6264  * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
6265  * This buffer serves as a temporary holder of compressed data while
6266  * the buffer entry is being written to an l2arc device. Once that is
6267  * done, we can dispose of it.
6268  */
6269 static void
6270 l2arc_release_cdata_buf(arc_buf_hdr_t *hdr)
6271 {
6272         enum zio_compress comp = HDR_GET_COMPRESS(hdr);
6273 
6274         ASSERT(HDR_HAS_L1HDR(hdr));
6275         ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp));
6276 
6277         if (comp == ZIO_COMPRESS_OFF) {
6278                 /*
6279                  * In this case, b_tmp_cdata points to the same buffer
6280                  * as the arc_buf_t's b_data field. We don't want to
6281                  * free it, since the arc_buf_t will handle that.
6282                  */
6283                 hdr->b_l1hdr.b_tmp_cdata = NULL;
6284         } else if (comp == ZIO_COMPRESS_EMPTY) {
6285                 /*
6286                  * In this case, b_tmp_cdata was compressed to an empty
6287                  * buffer, thus there's nothing to free and b_tmp_cdata
6288                  * should have been set to NULL in l2arc_write_buffers().
6289                  */
6290                 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
6291         } else {
6292                 /*
6293                  * If the data was compressed, then we've allocated a
6294                  * temporary buffer for it, so now we need to release it.
6295                  */
6296                 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6297                 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata,
6298                     hdr->b_size);
6299                 hdr->b_l1hdr.b_tmp_cdata = NULL;
6300         }
6301 
6302 }
6303 
6304 /*
6305  * This thread feeds the L2ARC at regular intervals.  This is the beating
6306  * heart of the L2ARC.
6307  */
6308 static void
6309 l2arc_feed_thread(void)
6310 {
6311         callb_cpr_t cpr;
6312         l2arc_dev_t *dev;
6313         spa_t *spa;
6314         uint64_t size, wrote;
6315         clock_t begin, next = ddi_get_lbolt();
6316         boolean_t headroom_boost = B_FALSE;
6317 
6318         CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
6319 
6320         mutex_enter(&l2arc_feed_thr_lock);
6321 
6322         while (l2arc_thread_exit == 0) {
6323                 CALLB_CPR_SAFE_BEGIN(&cpr);
6324                 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
6325                     next);
6326                 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
6327                 next = ddi_get_lbolt() + hz;
6328 
6329                 /*
6330                  * Quick check for L2ARC devices.
6331                  */
6332                 mutex_enter(&l2arc_dev_mtx);
6333                 if (l2arc_ndev == 0) {
6334                         mutex_exit(&l2arc_dev_mtx);
6335                         continue;
6336                 }
6337                 mutex_exit(&l2arc_dev_mtx);
6338                 begin = ddi_get_lbolt();
6339 
6340                 /*
6341                  * This selects the next l2arc device to write to, and in
6342                  * doing so the next spa to feed from: dev->l2ad_spa.   This
6343                  * will return NULL if there are now no l2arc devices or if
6344                  * they are all faulted.
6345                  *
6346                  * If a device is returned, its spa's config lock is also
6347                  * held to prevent device removal.  l2arc_dev_get_next()
6348                  * will grab and release l2arc_dev_mtx.
6349                  */
6350                 if ((dev = l2arc_dev_get_next()) == NULL)
6351                         continue;
6352 
6353                 spa = dev->l2ad_spa;
6354                 ASSERT(spa != NULL);
6355 
6356                 /*
6357                  * If the pool is read-only then force the feed thread to
6358                  * sleep a little longer.
6359                  */
6360                 if (!spa_writeable(spa)) {
6361                         next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
6362                         spa_config_exit(spa, SCL_L2ARC, dev);
6363                         continue;
6364                 }
6365 
6366                 /*
6367                  * Avoid contributing to memory pressure.
6368                  */
6369                 if (arc_reclaim_needed()) {
6370                         ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
6371                         spa_config_exit(spa, SCL_L2ARC, dev);
6372                         continue;
6373                 }
6374 
6375                 ARCSTAT_BUMP(arcstat_l2_feeds);
6376 
6377                 size = l2arc_write_size();
6378 
6379                 /*
6380                  * Evict L2ARC buffers that will be overwritten.
6381                  */
6382                 l2arc_evict(dev, size, B_FALSE);
6383 
6384                 /*
6385                  * Write ARC buffers.
6386                  */
6387                 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
6388 
6389                 /*
6390                  * Calculate interval between writes.
6391                  */
6392                 next = l2arc_write_interval(begin, size, wrote);
6393                 spa_config_exit(spa, SCL_L2ARC, dev);
6394         }
6395 
6396         l2arc_thread_exit = 0;
6397         cv_broadcast(&l2arc_feed_thr_cv);
6398         CALLB_CPR_EXIT(&cpr);               /* drops l2arc_feed_thr_lock */
6399         thread_exit();
6400 }
6401 
6402 boolean_t
6403 l2arc_vdev_present(vdev_t *vd)
6404 {
6405         l2arc_dev_t *dev;
6406 
6407         mutex_enter(&l2arc_dev_mtx);
6408         for (dev = list_head(l2arc_dev_list); dev != NULL;
6409             dev = list_next(l2arc_dev_list, dev)) {
6410                 if (dev->l2ad_vdev == vd)
6411                         break;
6412         }
6413         mutex_exit(&l2arc_dev_mtx);
6414 
6415         return (dev != NULL);
6416 }
6417 
6418 /*
6419  * Add a vdev for use by the L2ARC.  By this point the spa has already
6420  * validated the vdev and opened it.
6421  */
6422 void
6423 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
6424 {
6425         l2arc_dev_t *adddev;
6426 
6427         ASSERT(!l2arc_vdev_present(vd));
6428 
6429         /*
6430          * Create a new l2arc device entry.
6431          */
6432         adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
6433         adddev->l2ad_spa = spa;
6434         adddev->l2ad_vdev = vd;
6435         adddev->l2ad_start = VDEV_LABEL_START_SIZE;
6436         adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
6437         adddev->l2ad_hand = adddev->l2ad_start;
6438         adddev->l2ad_first = B_TRUE;
6439         adddev->l2ad_writing = B_FALSE;
6440 
6441         mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
6442         /*
6443          * This is a list of all ARC buffers that are still valid on the
6444          * device.
6445          */
6446         list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
6447             offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
6448 
6449         vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
6450         refcount_create(&adddev->l2ad_alloc);
6451 
6452         /*
6453          * Add device to global list
6454          */
6455         mutex_enter(&l2arc_dev_mtx);
6456         list_insert_head(l2arc_dev_list, adddev);
6457         atomic_inc_64(&l2arc_ndev);
6458         mutex_exit(&l2arc_dev_mtx);
6459 }
6460 
6461 /*
6462  * Remove a vdev from the L2ARC.
6463  */
6464 void
6465 l2arc_remove_vdev(vdev_t *vd)
6466 {
6467         l2arc_dev_t *dev, *nextdev, *remdev = NULL;
6468 
6469         /*
6470          * Find the device by vdev
6471          */
6472         mutex_enter(&l2arc_dev_mtx);
6473         for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
6474                 nextdev = list_next(l2arc_dev_list, dev);
6475                 if (vd == dev->l2ad_vdev) {
6476                         remdev = dev;
6477                         break;
6478                 }
6479         }
6480         ASSERT(remdev != NULL);
6481 
6482         /*
6483          * Remove device from global list
6484          */
6485         list_remove(l2arc_dev_list, remdev);
6486         l2arc_dev_last = NULL;          /* may have been invalidated */
6487         atomic_dec_64(&l2arc_ndev);
6488         mutex_exit(&l2arc_dev_mtx);
6489 
6490         /*
6491          * Clear all buflists and ARC references.  L2ARC device flush.
6492          */
6493         l2arc_evict(remdev, 0, B_TRUE);
6494         list_destroy(&remdev->l2ad_buflist);
6495         mutex_destroy(&remdev->l2ad_mtx);
6496         refcount_destroy(&remdev->l2ad_alloc);
6497         kmem_free(remdev, sizeof (l2arc_dev_t));
6498 }
6499 
6500 void
6501 l2arc_init(void)
6502 {
6503         l2arc_thread_exit = 0;
6504         l2arc_ndev = 0;
6505         l2arc_writes_sent = 0;
6506         l2arc_writes_done = 0;
6507 
6508         mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
6509         cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
6510         mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
6511         mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
6512 
6513         l2arc_dev_list = &L2ARC_dev_list;
6514         l2arc_free_on_write = &L2ARC_free_on_write;
6515         list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
6516             offsetof(l2arc_dev_t, l2ad_node));
6517         list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
6518             offsetof(l2arc_data_free_t, l2df_list_node));
6519 }
6520 
6521 void
6522 l2arc_fini(void)
6523 {
6524         /*
6525          * This is called from dmu_fini(), which is called from spa_fini();
6526          * Because of this, we can assume that all l2arc devices have
6527          * already been removed when the pools themselves were removed.
6528          */
6529 
6530         l2arc_do_free_on_write();
6531 
6532         mutex_destroy(&l2arc_feed_thr_lock);
6533         cv_destroy(&l2arc_feed_thr_cv);
6534         mutex_destroy(&l2arc_dev_mtx);
6535         mutex_destroy(&l2arc_free_on_write_mtx);
6536 
6537         list_destroy(l2arc_dev_list);
6538         list_destroy(l2arc_free_on_write);
6539 }
6540 
6541 void
6542 l2arc_start(void)
6543 {
6544         if (!(spa_mode_global & FWRITE))
6545                 return;
6546 
6547         (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
6548             TS_RUN, minclsyspri);
6549 }
6550 
6551 void
6552 l2arc_stop(void)
6553 {
6554         if (!(spa_mode_global & FWRITE))
6555                 return;
6556 
6557         mutex_enter(&l2arc_feed_thr_lock);
6558         cv_signal(&l2arc_feed_thr_cv);      /* kick thread out of startup */
6559         l2arc_thread_exit = 1;
6560         while (l2arc_thread_exit != 0)
6561                 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
6562         mutex_exit(&l2arc_feed_thr_lock);
6563 }