xref: /illumos-gate/usr/src/uts/common/fs/zfs/arc.c

/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License (the "License").
 * You may not use this file except in compliance with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */
/*
 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
 * Copyright (c) 2019, Joyent, Inc.
 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
 * Copyright 2017 Nexenta Systems, Inc.  All rights reserved.
 */

/*
 * DVA-based Adjustable Replacement Cache
 *
 * While much of the theory of operation used here is
 * based on the self-tuning, low overhead replacement cache
 * presented by Megiddo and Modha at FAST 2003, there are some
 * significant differences:
 *
 * 1. The Megiddo and Modha model assumes any page is evictable.
 * Pages in its cache cannot be "locked" into memory.  This makes
 * the eviction algorithm simple: evict the last page in the list.
 * This also make the performance characteristics easy to reason
 * about.  Our cache is not so simple.  At any given moment, some
 * subset of the blocks in the cache are un-evictable because we
 * have handed out a reference to them.  Blocks are only evictable
 * when there are no external references active.  This makes
 * eviction far more problematic:  we choose to evict the evictable
 * blocks that are the "lowest" in the list.
 *
 * There are times when it is not possible to evict the requested
 * space.  In these circumstances we are unable to adjust the cache
 * size.  To prevent the cache growing unbounded at these times we
 * implement a "cache throttle" that slows the flow of new data
 * into the cache until we can make space available.
 *
 * 2. The Megiddo and Modha model assumes a fixed cache size.
 * Pages are evicted when the cache is full and there is a cache
 * miss.  Our model has a variable sized cache.  It grows with
 * high use, but also tries to react to memory pressure from the
 * operating system: decreasing its size when system memory is
 * tight.
 *
 * 3. The Megiddo and Modha model assumes a fixed page size. All
 * elements of the cache are therefore exactly the same size.  So
 * when adjusting the cache size following a cache miss, its simply
 * a matter of choosing a single page to evict.  In our model, we
 * have variable sized cache blocks (rangeing from 512 bytes to
 * 128K bytes).  We therefore choose a set of blocks to evict to make
 * space for a cache miss that approximates as closely as possible
 * the space used by the new block.
 *
 * See also:  "ARC: A Self-Tuning, Low Overhead Replacement Cache"
 * by N. Megiddo & D. Modha, FAST 2003
 */

/*
 * The locking model:
 *
 * A new reference to a cache buffer can be obtained in two
 * ways: 1) via a hash table lookup using the DVA as a key,
 * or 2) via one of the ARC lists.  The arc_read() interface
 * uses method 1, while the internal ARC algorithms for
 * adjusting the cache use method 2.  We therefore provide two
 * types of locks: 1) the hash table lock array, and 2) the
 * ARC list locks.
 *
 * Buffers do not have their own mutexes, rather they rely on the
 * hash table mutexes for the bulk of their protection (i.e. most
 * fields in the arc_buf_hdr_t are protected by these mutexes).
 *
 * buf_hash_find() returns the appropriate mutex (held) when it
 * locates the requested buffer in the hash table.  It returns
 * NULL for the mutex if the buffer was not in the table.
 *
 * buf_hash_remove() expects the appropriate hash mutex to be
 * already held before it is invoked.
 *
 * Each ARC state also has a mutex which is used to protect the
 * buffer list associated with the state.  When attempting to
 * obtain a hash table lock while holding an ARC list lock you
 * must use: mutex_tryenter() to avoid deadlock.  Also note that
 * the active state mutex must be held before the ghost state mutex.
 *
 * Note that the majority of the performance stats are manipulated
 * with atomic operations.
 *
 * The L2ARC uses the l2ad_mtx on each vdev for the following:
 *
 *	- L2ARC buflist creation
 *	- L2ARC buflist eviction
 *	- L2ARC write completion, which walks L2ARC buflists
 *	- ARC header destruction, as it removes from L2ARC buflists
 *	- ARC header release, as it removes from L2ARC buflists
 */

/*
 * ARC operation:
 *
 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
 * This structure can point either to a block that is still in the cache or to
 * one that is only accessible in an L2 ARC device, or it can provide
 * information about a block that was recently evicted. If a block is
 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
 * information to retrieve it from the L2ARC device. This information is
 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
 * that is in this state cannot access the data directly.
 *
 * Blocks that are actively being referenced or have not been evicted
 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
 * the arc_buf_hdr_t that will point to the data block in memory. A block can
 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
 *
 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
 * ability to store the physical data (b_pabd) associated with the DVA of the
 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
 * it will match its on-disk compression characteristics. This behavior can be
 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
 * compressed ARC functionality is disabled, the b_pabd will point to an
 * uncompressed version of the on-disk data.
 *
 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
 * consumer. The ARC will provide references to this data and will keep it
 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
 * data block and will evict any arc_buf_t that is no longer referenced. The
 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
 * "overhead_size" kstat.
 *
 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
 * compressed form. The typical case is that consumers will want uncompressed
 * data, and when that happens a new data buffer is allocated where the data is
 * decompressed for them to use. Currently the only consumer who wants
 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
 * with the arc_buf_hdr_t.
 *
 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
 * first one is owned by a compressed send consumer (and therefore references
 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
 * used by any other consumer (and has its own uncompressed copy of the data
 * buffer).
 *
 *   arc_buf_hdr_t
 *   +-----------+
 *   | fields    |
 *   | common to |
 *   | L1- and   |
 *   | L2ARC     |
 *   +-----------+
 *   | l2arc_buf_hdr_t
 *   |           |
 *   +-----------+
 *   | l1arc_buf_hdr_t
 *   |           |              arc_buf_t
 *   | b_buf     +------------>+-----------+      arc_buf_t
 *   | b_pabd    +-+           |b_next     +---->+-----------+
 *   +-----------+ |           |-----------|     |b_next     +-->NULL
 *                 |           |b_comp = T |     +-----------+
 *                 |           |b_data     +-+   |b_comp = F |
 *                 |           +-----------+ |   |b_data     +-+
 *                 +->+------+               |   +-----------+ |
 *        compressed  |      |               |                 |
 *           data     |      |<--------------+                 | uncompressed
 *                    +------+          compressed,            |     data
 *                                        shared               +-->+------+
 *                                         data                    |      |
 *                                                                 |      |
 *                                                                 +------+
 *
 * When a consumer reads a block, the ARC must first look to see if the
 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
 * arc_buf_t and either copies uncompressed data into a new data buffer from an
 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
 * hdr is compressed and the desired compression characteristics of the
 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
 * be anywhere in the hdr's list.
 *
 * The diagram below shows an example of an uncompressed ARC hdr that is
 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
 * the last element in the buf list):
 *
 *                arc_buf_hdr_t
 *                +-----------+
 *                |           |
 *                |           |
 *                |           |
 *                +-----------+
 * l2arc_buf_hdr_t|           |
 *                |           |
 *                +-----------+
 * l1arc_buf_hdr_t|           |
 *                |           |                 arc_buf_t    (shared)
 *                |    b_buf  +------------>+---------+      arc_buf_t
 *                |           |             |b_next   +---->+---------+
 *                |  b_pabd   +-+           |---------|     |b_next   +-->NULL
 *                +-----------+ |           |         |     +---------+
 *                              |           |b_data   +-+   |         |
 *                              |           +---------+ |   |b_data   +-+
 *                              +->+------+             |   +---------+ |
 *                                 |      |             |               |
 *                   uncompressed  |      |             |               |
 *                        data     +------+             |               |
 *                                    ^                 +->+------+     |
 *                                    |       uncompressed |      |     |
 *                                    |           data     |      |     |
 *                                    |                    +------+     |
 *                                    +---------------------------------+
 *
 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
 * since the physical block is about to be rewritten. The new data contents
 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
 * it may compress the data before writing it to disk. The ARC will be called
 * with the transformed data and will bcopy the transformed on-disk block into
 * a newly allocated b_pabd. Writes are always done into buffers which have
 * either been loaned (and hence are new and don't have other readers) or
 * buffers which have been released (and hence have their own hdr, if there
 * were originally other readers of the buf's original hdr). This ensures that
 * the ARC only needs to update a single buf and its hdr after a write occurs.
 *
 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
 * that when compressed ARC is enabled that the L2ARC blocks are identical
 * to the on-disk block in the main data pool. This provides a significant
 * advantage since the ARC can leverage the bp's checksum when reading from the
 * L2ARC to determine if the contents are valid. However, if the compressed
 * ARC is disabled, then the L2ARC's block must be transformed to look
 * like the physical block in the main data pool before comparing the
 * checksum and determining its validity.
 *
 * The L1ARC has a slightly different system for storing encrypted data.
 * Raw (encrypted + possibly compressed) data has a few subtle differences from
 * data that is just compressed. The biggest difference is that it is not
 * possible to decrypt encrypted data (or visa versa) if the keys aren't loaded.
 * The other difference is that encryption cannot be treated as a suggestion.
 * If a caller would prefer compressed data, but they actually wind up with
 * uncompressed data the worst thing that could happen is there might be a
 * performance hit. If the caller requests encrypted data, however, we must be
 * sure they actually get it or else secret information could be leaked. Raw
 * data is stored in hdr->b_crypt_hdr.b_rabd. An encrypted header, therefore,
 * may have both an encrypted version and a decrypted version of its data at
 * once. When a caller needs a raw arc_buf_t, it is allocated and the data is
 * copied out of this header. To avoid complications with b_pabd, raw buffers
 * cannot be shared.
 */

#include <sys/spa.h>
#include <sys/zio.h>
#include <sys/spa_impl.h>
#include <sys/zio_compress.h>
#include <sys/zio_checksum.h>
#include <sys/zfs_context.h>
#include <sys/arc.h>
#include <sys/refcount.h>
#include <sys/vdev.h>
#include <sys/vdev_impl.h>
#include <sys/dsl_pool.h>
#include <sys/zio_checksum.h>
#include <sys/multilist.h>
#include <sys/abd.h>
#include <sys/zil.h>
#include <sys/fm/fs/zfs.h>
#ifdef _KERNEL
#include <sys/vmsystm.h>
#include <vm/anon.h>
#include <sys/fs/swapnode.h>
#include <sys/dnlc.h>
#endif
#include <sys/callb.h>
#include <sys/kstat.h>
#include <sys/zthr.h>
#include <zfs_fletcher.h>
#include <sys/aggsum.h>
#include <sys/cityhash.h>
#include <sys/param.h>

#ifndef _KERNEL
/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
boolean_t arc_watch = B_FALSE;
int arc_procfd;
#endif

/*
 * This thread's job is to keep enough free memory in the system, by
 * calling arc_kmem_reap_now() plus arc_shrink(), which improves
 * arc_available_memory().
 */
static zthr_t		*arc_reap_zthr;

/*
 * This thread's job is to keep arc_size under arc_c, by calling
 * arc_adjust(), which improves arc_is_overflowing().
 */
static zthr_t		*arc_adjust_zthr;

static kmutex_t		arc_adjust_lock;
static kcondvar_t	arc_adjust_waiters_cv;
static boolean_t	arc_adjust_needed = B_FALSE;

uint_t arc_reduce_dnlc_percent = 3;

/*
 * The number of headers to evict in arc_evict_state_impl() before
 * dropping the sublist lock and evicting from another sublist. A lower
 * value means we're more likely to evict the "correct" header (i.e. the
 * oldest header in the arc state), but comes with higher overhead
 * (i.e. more invocations of arc_evict_state_impl()).
 */
int zfs_arc_evict_batch_limit = 10;

/* number of seconds before growing cache again */
int arc_grow_retry = 60;

/*
 * Minimum time between calls to arc_kmem_reap_soon().  Note that this will
 * be converted to ticks, so with the default hz=100, a setting of 15 ms
 * will actually wait 2 ticks, or 20ms.
 */
int arc_kmem_cache_reap_retry_ms = 1000;

/* shift of arc_c for calculating overflow limit in arc_get_data_impl */
int zfs_arc_overflow_shift = 8;

/* shift of arc_c for calculating both min and max arc_p */
int arc_p_min_shift = 4;

/* log2(fraction of arc to reclaim) */
int arc_shrink_shift = 7;

/*
 * log2(fraction of ARC which must be free to allow growing).
 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
 * when reading a new block into the ARC, we will evict an equal-sized block
 * from the ARC.
 *
 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
 * we will still not allow it to grow.
 */
int			arc_no_grow_shift = 5;


/*
 * minimum lifespan of a prefetch block in clock ticks
 * (initialized in arc_init())
 */
static int		zfs_arc_min_prefetch_ms = 1;
static int		zfs_arc_min_prescient_prefetch_ms = 6;

/*
 * If this percent of memory is free, don't throttle.
 */
int arc_lotsfree_percent = 10;

static boolean_t arc_initialized;

/*
 * The arc has filled available memory and has now warmed up.
 */
static boolean_t arc_warm;

/*
 * log2 fraction of the zio arena to keep free.
 */
int arc_zio_arena_free_shift = 2;

/*
 * These tunables are for performance analysis.
 */
uint64_t zfs_arc_max;
uint64_t zfs_arc_min;
uint64_t zfs_arc_meta_limit = 0;
uint64_t zfs_arc_meta_min = 0;
int zfs_arc_grow_retry = 0;
int zfs_arc_shrink_shift = 0;
int zfs_arc_p_min_shift = 0;
int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */

/*
 * ARC dirty data constraints for arc_tempreserve_space() throttle
 */
uint_t zfs_arc_dirty_limit_percent = 50;	/* total dirty data limit */
uint_t zfs_arc_anon_limit_percent = 25;		/* anon block dirty limit */
uint_t zfs_arc_pool_dirty_percent = 20;		/* each pool's anon allowance */

boolean_t zfs_compressed_arc_enabled = B_TRUE;

/*
 * Note that buffers can be in one of 6 states:
 *	ARC_anon	- anonymous (discussed below)
 *	ARC_mru		- recently used, currently cached
 *	ARC_mru_ghost	- recentely used, no longer in cache
 *	ARC_mfu		- frequently used, currently cached
 *	ARC_mfu_ghost	- frequently used, no longer in cache
 *	ARC_l2c_only	- exists in L2ARC but not other states
 * When there are no active references to the buffer, they are
 * are linked onto a list in one of these arc states.  These are
 * the only buffers that can be evicted or deleted.  Within each
 * state there are multiple lists, one for meta-data and one for
 * non-meta-data.  Meta-data (indirect blocks, blocks of dnodes,
 * etc.) is tracked separately so that it can be managed more
 * explicitly: favored over data, limited explicitly.
 *
 * Anonymous buffers are buffers that are not associated with
 * a DVA.  These are buffers that hold dirty block copies
 * before they are written to stable storage.  By definition,
 * they are "ref'd" and are considered part of arc_mru
 * that cannot be freed.  Generally, they will aquire a DVA
 * as they are written and migrate onto the arc_mru list.
 *
 * The ARC_l2c_only state is for buffers that are in the second
 * level ARC but no longer in any of the ARC_m* lists.  The second
 * level ARC itself may also contain buffers that are in any of
 * the ARC_m* states - meaning that a buffer can exist in two
 * places.  The reason for the ARC_l2c_only state is to keep the
 * buffer header in the hash table, so that reads that hit the
 * second level ARC benefit from these fast lookups.
 */

typedef struct arc_state {
	/*
	 * list of evictable buffers
	 */
	multilist_t *arcs_list[ARC_BUFC_NUMTYPES];
	/*
	 * total amount of evictable data in this state
	 */
	zfs_refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
	/*
	 * total amount of data in this state; this includes: evictable,
	 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
	 */
	zfs_refcount_t arcs_size;
} arc_state_t;

/* The 6 states: */
static arc_state_t ARC_anon;
static arc_state_t ARC_mru;
static arc_state_t ARC_mru_ghost;
static arc_state_t ARC_mfu;
static arc_state_t ARC_mfu_ghost;
static arc_state_t ARC_l2c_only;

typedef struct arc_stats {
	kstat_named_t arcstat_hits;
	kstat_named_t arcstat_misses;
	kstat_named_t arcstat_demand_data_hits;
	kstat_named_t arcstat_demand_data_misses;
	kstat_named_t arcstat_demand_metadata_hits;
	kstat_named_t arcstat_demand_metadata_misses;
	kstat_named_t arcstat_prefetch_data_hits;
	kstat_named_t arcstat_prefetch_data_misses;
	kstat_named_t arcstat_prefetch_metadata_hits;
	kstat_named_t arcstat_prefetch_metadata_misses;
	kstat_named_t arcstat_mru_hits;
	kstat_named_t arcstat_mru_ghost_hits;
	kstat_named_t arcstat_mfu_hits;
	kstat_named_t arcstat_mfu_ghost_hits;
	kstat_named_t arcstat_deleted;
	/*
	 * Number of buffers that could not be evicted because the hash lock
	 * was held by another thread.  The lock may not necessarily be held
	 * by something using the same buffer, since hash locks are shared
	 * by multiple buffers.
	 */
	kstat_named_t arcstat_mutex_miss;
	/*
	 * Number of buffers skipped when updating the access state due to the
	 * header having already been released after acquiring the hash lock.
	 */
	kstat_named_t arcstat_access_skip;
	/*
	 * Number of buffers skipped because they have I/O in progress, are
	 * indirect prefetch buffers that have not lived long enough, or are
	 * not from the spa we're trying to evict from.
	 */
	kstat_named_t arcstat_evict_skip;
	/*
	 * Number of times arc_evict_state() was unable to evict enough
	 * buffers to reach its target amount.
	 */
	kstat_named_t arcstat_evict_not_enough;
	kstat_named_t arcstat_evict_l2_cached;
	kstat_named_t arcstat_evict_l2_eligible;
	kstat_named_t arcstat_evict_l2_ineligible;
	kstat_named_t arcstat_evict_l2_skip;
	kstat_named_t arcstat_hash_elements;
	kstat_named_t arcstat_hash_elements_max;
	kstat_named_t arcstat_hash_collisions;
	kstat_named_t arcstat_hash_chains;
	kstat_named_t arcstat_hash_chain_max;
	kstat_named_t arcstat_p;
	kstat_named_t arcstat_c;
	kstat_named_t arcstat_c_min;
	kstat_named_t arcstat_c_max;
	/* Not updated directly; only synced in arc_kstat_update. */
	kstat_named_t arcstat_size;
	/*
	 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
	 * Note that the compressed bytes may match the uncompressed bytes
	 * if the block is either not compressed or compressed arc is disabled.
	 */
	kstat_named_t arcstat_compressed_size;
	/*
	 * Uncompressed size of the data stored in b_pabd. If compressed
	 * arc is disabled then this value will be identical to the stat
	 * above.
	 */
	kstat_named_t arcstat_uncompressed_size;
	/*
	 * Number of bytes stored in all the arc_buf_t's. This is classified
	 * as "overhead" since this data is typically short-lived and will
	 * be evicted from the arc when it becomes unreferenced unless the
	 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
	 * values have been set (see comment in dbuf.c for more information).
	 */
	kstat_named_t arcstat_overhead_size;
	/*
	 * Number of bytes consumed by internal ARC structures necessary
	 * for tracking purposes; these structures are not actually
	 * backed by ARC buffers. This includes arc_buf_hdr_t structures
	 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
	 * caches), and arc_buf_t structures (allocated via arc_buf_t
	 * cache).
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_hdr_size;
	/*
	 * Number of bytes consumed by ARC buffers of type equal to
	 * ARC_BUFC_DATA. This is generally consumed by buffers backing
	 * on disk user data (e.g. plain file contents).
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_data_size;
	/*
	 * Number of bytes consumed by ARC buffers of type equal to
	 * ARC_BUFC_METADATA. This is generally consumed by buffers
	 * backing on disk data that is used for internal ZFS
	 * structures (e.g. ZAP, dnode, indirect blocks, etc).
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_metadata_size;
	/*
	 * Number of bytes consumed by various buffers and structures
	 * not actually backed with ARC buffers. This includes bonus
	 * buffers (allocated directly via zio_buf_* functions),
	 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
	 * cache), and dnode_t structures (allocated via dnode_t cache).
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_other_size;
	/*
	 * Total number of bytes consumed by ARC buffers residing in the
	 * arc_anon state. This includes *all* buffers in the arc_anon
	 * state; e.g. data, metadata, evictable, and unevictable buffers
	 * are all included in this value.
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_anon_size;
	/*
	 * Number of bytes consumed by ARC buffers that meet the
	 * following criteria: backing buffers of type ARC_BUFC_DATA,
	 * residing in the arc_anon state, and are eligible for eviction
	 * (e.g. have no outstanding holds on the buffer).
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_anon_evictable_data;
	/*
	 * Number of bytes consumed by ARC buffers that meet the
	 * following criteria: backing buffers of type ARC_BUFC_METADATA,
	 * residing in the arc_anon state, and are eligible for eviction
	 * (e.g. have no outstanding holds on the buffer).
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_anon_evictable_metadata;
	/*
	 * Total number of bytes consumed by ARC buffers residing in the
	 * arc_mru state. This includes *all* buffers in the arc_mru
	 * state; e.g. data, metadata, evictable, and unevictable buffers
	 * are all included in this value.
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_mru_size;
	/*
	 * Number of bytes consumed by ARC buffers that meet the
	 * following criteria: backing buffers of type ARC_BUFC_DATA,
	 * residing in the arc_mru state, and are eligible for eviction
	 * (e.g. have no outstanding holds on the buffer).
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_mru_evictable_data;
	/*
	 * Number of bytes consumed by ARC buffers that meet the
	 * following criteria: backing buffers of type ARC_BUFC_METADATA,
	 * residing in the arc_mru state, and are eligible for eviction
	 * (e.g. have no outstanding holds on the buffer).
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_mru_evictable_metadata;
	/*
	 * Total number of bytes that *would have been* consumed by ARC
	 * buffers in the arc_mru_ghost state. The key thing to note
	 * here, is the fact that this size doesn't actually indicate
	 * RAM consumption. The ghost lists only consist of headers and
	 * don't actually have ARC buffers linked off of these headers.
	 * Thus, *if* the headers had associated ARC buffers, these
	 * buffers *would have* consumed this number of bytes.
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_mru_ghost_size;
	/*
	 * Number of bytes that *would have been* consumed by ARC
	 * buffers that are eligible for eviction, of type
	 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_mru_ghost_evictable_data;
	/*
	 * Number of bytes that *would have been* consumed by ARC
	 * buffers that are eligible for eviction, of type
	 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_mru_ghost_evictable_metadata;
	/*
	 * Total number of bytes consumed by ARC buffers residing in the
	 * arc_mfu state. This includes *all* buffers in the arc_mfu
	 * state; e.g. data, metadata, evictable, and unevictable buffers
	 * are all included in this value.
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_mfu_size;
	/*
	 * Number of bytes consumed by ARC buffers that are eligible for
	 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
	 * state.
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_mfu_evictable_data;
	/*
	 * Number of bytes consumed by ARC buffers that are eligible for
	 * eviction, of type ARC_BUFC_METADATA, and reside in the
	 * arc_mfu state.
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_mfu_evictable_metadata;
	/*
	 * Total number of bytes that *would have been* consumed by ARC
	 * buffers in the arc_mfu_ghost state. See the comment above
	 * arcstat_mru_ghost_size for more details.
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_mfu_ghost_size;
	/*
	 * Number of bytes that *would have been* consumed by ARC
	 * buffers that are eligible for eviction, of type
	 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_mfu_ghost_evictable_data;
	/*
	 * Number of bytes that *would have been* consumed by ARC
	 * buffers that are eligible for eviction, of type
	 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
	 * Not updated directly; only synced in arc_kstat_update.
	 */
	kstat_named_t arcstat_mfu_ghost_evictable_metadata;
	kstat_named_t arcstat_l2_hits;
	kstat_named_t arcstat_l2_misses;
	kstat_named_t arcstat_l2_feeds;
	kstat_named_t arcstat_l2_rw_clash;
	kstat_named_t arcstat_l2_read_bytes;
	kstat_named_t arcstat_l2_write_bytes;
	kstat_named_t arcstat_l2_writes_sent;
	kstat_named_t arcstat_l2_writes_done;
	kstat_named_t arcstat_l2_writes_error;
	kstat_named_t arcstat_l2_writes_lock_retry;
	kstat_named_t arcstat_l2_evict_lock_retry;
	kstat_named_t arcstat_l2_evict_reading;
	kstat_named_t arcstat_l2_evict_l1cached;
	kstat_named_t arcstat_l2_free_on_write;
	kstat_named_t arcstat_l2_abort_lowmem;
	kstat_named_t arcstat_l2_cksum_bad;
	kstat_named_t arcstat_l2_io_error;
	kstat_named_t arcstat_l2_lsize;
	kstat_named_t arcstat_l2_psize;
	/* Not updated directly; only synced in arc_kstat_update. */
	kstat_named_t arcstat_l2_hdr_size;
	kstat_named_t arcstat_memory_throttle_count;
	/* Not updated directly; only synced in arc_kstat_update. */
	kstat_named_t arcstat_meta_used;
	kstat_named_t arcstat_meta_limit;
	kstat_named_t arcstat_meta_max;
	kstat_named_t arcstat_meta_min;
	kstat_named_t arcstat_async_upgrade_sync;
	kstat_named_t arcstat_demand_hit_predictive_prefetch;
	kstat_named_t arcstat_demand_hit_prescient_prefetch;
} arc_stats_t;

static arc_stats_t arc_stats = {
	{ "hits",			KSTAT_DATA_UINT64 },
	{ "misses",			KSTAT_DATA_UINT64 },
	{ "demand_data_hits",		KSTAT_DATA_UINT64 },
	{ "demand_data_misses",		KSTAT_DATA_UINT64 },
	{ "demand_metadata_hits",	KSTAT_DATA_UINT64 },
	{ "demand_metadata_misses",	KSTAT_DATA_UINT64 },
	{ "prefetch_data_hits",		KSTAT_DATA_UINT64 },
	{ "prefetch_data_misses",	KSTAT_DATA_UINT64 },
	{ "prefetch_metadata_hits",	KSTAT_DATA_UINT64 },
	{ "prefetch_metadata_misses",	KSTAT_DATA_UINT64 },
	{ "mru_hits",			KSTAT_DATA_UINT64 },
	{ "mru_ghost_hits",		KSTAT_DATA_UINT64 },
	{ "mfu_hits",			KSTAT_DATA_UINT64 },
	{ "mfu_ghost_hits",		KSTAT_DATA_UINT64 },
	{ "deleted",			KSTAT_DATA_UINT64 },
	{ "mutex_miss",			KSTAT_DATA_UINT64 },
	{ "access_skip",		KSTAT_DATA_UINT64 },
	{ "evict_skip",			KSTAT_DATA_UINT64 },
	{ "evict_not_enough",		KSTAT_DATA_UINT64 },
	{ "evict_l2_cached",		KSTAT_DATA_UINT64 },
	{ "evict_l2_eligible",		KSTAT_DATA_UINT64 },
	{ "evict_l2_ineligible",	KSTAT_DATA_UINT64 },
	{ "evict_l2_skip",		KSTAT_DATA_UINT64 },
	{ "hash_elements",		KSTAT_DATA_UINT64 },
	{ "hash_elements_max",		KSTAT_DATA_UINT64 },
	{ "hash_collisions",		KSTAT_DATA_UINT64 },
	{ "hash_chains",		KSTAT_DATA_UINT64 },
	{ "hash_chain_max",		KSTAT_DATA_UINT64 },
	{ "p",				KSTAT_DATA_UINT64 },
	{ "c",				KSTAT_DATA_UINT64 },
	{ "c_min",			KSTAT_DATA_UINT64 },
	{ "c_max",			KSTAT_DATA_UINT64 },
	{ "size",			KSTAT_DATA_UINT64 },
	{ "compressed_size",		KSTAT_DATA_UINT64 },
	{ "uncompressed_size",		KSTAT_DATA_UINT64 },
	{ "overhead_size",		KSTAT_DATA_UINT64 },
	{ "hdr_size",			KSTAT_DATA_UINT64 },
	{ "data_size",			KSTAT_DATA_UINT64 },
	{ "metadata_size",		KSTAT_DATA_UINT64 },
	{ "other_size",			KSTAT_DATA_UINT64 },
	{ "anon_size",			KSTAT_DATA_UINT64 },
	{ "anon_evictable_data",	KSTAT_DATA_UINT64 },
	{ "anon_evictable_metadata",	KSTAT_DATA_UINT64 },
	{ "mru_size",			KSTAT_DATA_UINT64 },
	{ "mru_evictable_data",		KSTAT_DATA_UINT64 },
	{ "mru_evictable_metadata",	KSTAT_DATA_UINT64 },
	{ "mru_ghost_size",		KSTAT_DATA_UINT64 },
	{ "mru_ghost_evictable_data",	KSTAT_DATA_UINT64 },
	{ "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
	{ "mfu_size",			KSTAT_DATA_UINT64 },
	{ "mfu_evictable_data",		KSTAT_DATA_UINT64 },
	{ "mfu_evictable_metadata",	KSTAT_DATA_UINT64 },
	{ "mfu_ghost_size",		KSTAT_DATA_UINT64 },
	{ "mfu_ghost_evictable_data",	KSTAT_DATA_UINT64 },
	{ "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
	{ "l2_hits",			KSTAT_DATA_UINT64 },
	{ "l2_misses",			KSTAT_DATA_UINT64 },
	{ "l2_feeds",			KSTAT_DATA_UINT64 },
	{ "l2_rw_clash",		KSTAT_DATA_UINT64 },
	{ "l2_read_bytes",		KSTAT_DATA_UINT64 },
	{ "l2_write_bytes",		KSTAT_DATA_UINT64 },
	{ "l2_writes_sent",		KSTAT_DATA_UINT64 },
	{ "l2_writes_done",		KSTAT_DATA_UINT64 },
	{ "l2_writes_error",		KSTAT_DATA_UINT64 },
	{ "l2_writes_lock_retry",	KSTAT_DATA_UINT64 },
	{ "l2_evict_lock_retry",	KSTAT_DATA_UINT64 },
	{ "l2_evict_reading",		KSTAT_DATA_UINT64 },
	{ "l2_evict_l1cached",		KSTAT_DATA_UINT64 },
	{ "l2_free_on_write",		KSTAT_DATA_UINT64 },
	{ "l2_abort_lowmem",		KSTAT_DATA_UINT64 },
	{ "l2_cksum_bad",		KSTAT_DATA_UINT64 },
	{ "l2_io_error",		KSTAT_DATA_UINT64 },
	{ "l2_size",			KSTAT_DATA_UINT64 },
	{ "l2_asize",			KSTAT_DATA_UINT64 },
	{ "l2_hdr_size",		KSTAT_DATA_UINT64 },
	{ "memory_throttle_count",	KSTAT_DATA_UINT64 },
	{ "arc_meta_used",		KSTAT_DATA_UINT64 },
	{ "arc_meta_limit",		KSTAT_DATA_UINT64 },
	{ "arc_meta_max",		KSTAT_DATA_UINT64 },
	{ "arc_meta_min",		KSTAT_DATA_UINT64 },
	{ "async_upgrade_sync",		KSTAT_DATA_UINT64 },
	{ "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
	{ "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
};

#define	ARCSTAT(stat)	(arc_stats.stat.value.ui64)

#define	ARCSTAT_INCR(stat, val) \
	atomic_add_64(&arc_stats.stat.value.ui64, (val))

#define	ARCSTAT_BUMP(stat)	ARCSTAT_INCR(stat, 1)
#define	ARCSTAT_BUMPDOWN(stat)	ARCSTAT_INCR(stat, -1)

#define	ARCSTAT_MAX(stat, val) {					\
	uint64_t m;							\
	while ((val) > (m = arc_stats.stat.value.ui64) &&		\
	    (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val))))	\
		continue;						\
}

#define	ARCSTAT_MAXSTAT(stat) \
	ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)

/*
 * We define a macro to allow ARC hits/misses to be easily broken down by
 * two separate conditions, giving a total of four different subtypes for
 * each of hits and misses (so eight statistics total).
 */
#define	ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
	if (cond1) {							\
		if (cond2) {						\
			ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
		} else {						\
			ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
		}							\
	} else {							\
		if (cond2) {						\
			ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
		} else {						\
			ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
		}							\
	}

kstat_t			*arc_ksp;
static arc_state_t	*arc_anon;
static arc_state_t	*arc_mru;
static arc_state_t	*arc_mru_ghost;
static arc_state_t	*arc_mfu;
static arc_state_t	*arc_mfu_ghost;
static arc_state_t	*arc_l2c_only;

/*
 * There are several ARC variables that are critical to export as kstats --
 * but we don't want to have to grovel around in the kstat whenever we wish to
 * manipulate them.  For these variables, we therefore define them to be in
 * terms of the statistic variable.  This assures that we are not introducing
 * the possibility of inconsistency by having shadow copies of the variables,
 * while still allowing the code to be readable.
 */
#define	arc_p		ARCSTAT(arcstat_p)	/* target size of MRU */
#define	arc_c		ARCSTAT(arcstat_c)	/* target size of cache */
#define	arc_c_min	ARCSTAT(arcstat_c_min)	/* min target cache size */
#define	arc_c_max	ARCSTAT(arcstat_c_max)	/* max target cache size */
#define	arc_meta_limit	ARCSTAT(arcstat_meta_limit) /* max size for metadata */
#define	arc_meta_min	ARCSTAT(arcstat_meta_min) /* min size for metadata */
#define	arc_meta_max	ARCSTAT(arcstat_meta_max) /* max size of metadata */

/* compressed size of entire arc */
#define	arc_compressed_size	ARCSTAT(arcstat_compressed_size)
/* uncompressed size of entire arc */
#define	arc_uncompressed_size	ARCSTAT(arcstat_uncompressed_size)
/* number of bytes in the arc from arc_buf_t's */
#define	arc_overhead_size	ARCSTAT(arcstat_overhead_size)

/*
 * There are also some ARC variables that we want to export, but that are
 * updated so often that having the canonical representation be the statistic
 * variable causes a performance bottleneck. We want to use aggsum_t's for these
 * instead, but still be able to export the kstat in the same way as before.
 * The solution is to always use the aggsum version, except in the kstat update
 * callback.
 */
aggsum_t arc_size;
aggsum_t arc_meta_used;
aggsum_t astat_data_size;
aggsum_t astat_metadata_size;
aggsum_t astat_hdr_size;
aggsum_t astat_other_size;
aggsum_t astat_l2_hdr_size;

static int		arc_no_grow;	/* Don't try to grow cache size */
static hrtime_t		arc_growtime;
static uint64_t		arc_tempreserve;
static uint64_t		arc_loaned_bytes;

typedef struct arc_callback arc_callback_t;

struct arc_callback {
	void			*acb_private;
	arc_read_done_func_t	*acb_done;
	arc_buf_t		*acb_buf;
	boolean_t		acb_encrypted;
	boolean_t		acb_compressed;
	boolean_t		acb_noauth;
	zbookmark_phys_t	acb_zb;
	zio_t			*acb_zio_dummy;
	zio_t			*acb_zio_head;
	arc_callback_t		*acb_next;
};

typedef struct arc_write_callback arc_write_callback_t;

struct arc_write_callback {
	void			*awcb_private;
	arc_write_done_func_t	*awcb_ready;
	arc_write_done_func_t	*awcb_children_ready;
	arc_write_done_func_t	*awcb_physdone;
	arc_write_done_func_t	*awcb_done;
	arc_buf_t		*awcb_buf;
};

/*
 * ARC buffers are separated into multiple structs as a memory saving measure:
 *   - Common fields struct, always defined, and embedded within it:
 *       - L2-only fields, always allocated but undefined when not in L2ARC
 *       - L1-only fields, only allocated when in L1ARC
 *
 *           Buffer in L1                     Buffer only in L2
 *    +------------------------+          +------------------------+
 *    | arc_buf_hdr_t          |          | arc_buf_hdr_t          |
 *    |                        |          |                        |
 *    |                        |          |                        |
 *    |                        |          |                        |
 *    +------------------------+          +------------------------+
 *    | l2arc_buf_hdr_t        |          | l2arc_buf_hdr_t        |
 *    | (undefined if L1-only) |          |                        |
 *    +------------------------+          +------------------------+
 *    | l1arc_buf_hdr_t        |
 *    |                        |
 *    |                        |
 *    |                        |
 *    |                        |
 *    +------------------------+
 *
 * Because it's possible for the L2ARC to become extremely large, we can wind
 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
 * is minimized by only allocating the fields necessary for an L1-cached buffer
 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
 * words in pointers. arc_hdr_realloc() is used to switch a header between
 * these two allocation states.
 */
typedef struct l1arc_buf_hdr {
	kmutex_t		b_freeze_lock;
	zio_cksum_t		*b_freeze_cksum;
#ifdef ZFS_DEBUG
	/*
	 * Used for debugging with kmem_flags - by allocating and freeing
	 * b_thawed when the buffer is thawed, we get a record of the stack
	 * trace that thawed it.
	 */
	void			*b_thawed;
#endif

	arc_buf_t		*b_buf;
	uint32_t		b_bufcnt;
	/* for waiting on writes to complete */
	kcondvar_t		b_cv;
	uint8_t			b_byteswap;

	/* protected by arc state mutex */
	arc_state_t		*b_state;
	multilist_node_t	b_arc_node;

	/* updated atomically */
	clock_t			b_arc_access;

	/* self protecting */
	zfs_refcount_t		b_refcnt;

	arc_callback_t		*b_acb;
	abd_t			*b_pabd;
} l1arc_buf_hdr_t;

/*
 * Encrypted blocks will need to be stored encrypted on the L2ARC
 * disk as they appear in the main pool. In order for this to work we
 * need to pass around the encryption parameters so they can be used
 * to write data to the L2ARC. This struct is only defined in the
 * arc_buf_hdr_t if the L1 header is defined and has the ARC_FLAG_ENCRYPTED
 * flag set.
 */
typedef struct arc_buf_hdr_crypt {
	abd_t		*b_rabd;		/* raw encrypted data */
	dmu_object_type_t	b_ot;		/* object type */
	uint32_t		b_ebufcnt;	/* number or encryped buffers */

	/* dsobj for looking up encryption key for l2arc encryption */
	uint64_t		b_dsobj;	/* for looking up key */

	/* encryption parameters */
	uint8_t		b_salt[ZIO_DATA_SALT_LEN];
	uint8_t		b_iv[ZIO_DATA_IV_LEN];

	/*
	 * Technically this could be removed since we will always be able to
	 * get the mac from the bp when we need it. However, it is inconvenient
	 * for callers of arc code to have to pass a bp in all the time. This
	 * also allows us to assert that L2ARC data is properly encrypted to
	 * match the data in the main storage pool.
	 */
	uint8_t		b_mac[ZIO_DATA_MAC_LEN];
} arc_buf_hdr_crypt_t;

typedef struct l2arc_dev l2arc_dev_t;

typedef struct l2arc_buf_hdr {
	/* protected by arc_buf_hdr mutex */
	l2arc_dev_t		*b_dev;		/* L2ARC device */
	uint64_t		b_daddr;	/* disk address, offset byte */

	list_node_t		b_l2node;
} l2arc_buf_hdr_t;

struct arc_buf_hdr {
	/* protected by hash lock */
	dva_t			b_dva;
	uint64_t		b_birth;

	arc_buf_contents_t	b_type;
	arc_buf_hdr_t		*b_hash_next;
	arc_flags_t		b_flags;

	/*
	 * This field stores the size of the data buffer after
	 * compression, and is set in the arc's zio completion handlers.
	 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
	 *
	 * While the block pointers can store up to 32MB in their psize
	 * field, we can only store up to 32MB minus 512B. This is due
	 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
	 * a field of zeros represents 512B in the bp). We can't use a
	 * bias of 1 since we need to reserve a psize of zero, here, to
	 * represent holes and embedded blocks.
	 *
	 * This isn't a problem in practice, since the maximum size of a
	 * buffer is limited to 16MB, so we never need to store 32MB in
	 * this field. Even in the upstream illumos code base, the
	 * maximum size of a buffer is limited to 16MB.
	 */
	uint16_t		b_psize;

	/*
	 * This field stores the size of the data buffer before
	 * compression, and cannot change once set. It is in units
	 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
	 */
	uint16_t		b_lsize;	/* immutable */
	uint64_t		b_spa;		/* immutable */

	/* L2ARC fields. Undefined when not in L2ARC. */
	l2arc_buf_hdr_t		b_l2hdr;
	/* L1ARC fields. Undefined when in l2arc_only state */
	l1arc_buf_hdr_t		b_l1hdr;
	/*
	 * Encryption parameters. Defined only when ARC_FLAG_ENCRYPTED
	 * is set and the L1 header exists.
	 */
	arc_buf_hdr_crypt_t b_crypt_hdr;
};

#define	GHOST_STATE(state)	\
	((state) == arc_mru_ghost || (state) == arc_mfu_ghost ||	\
	(state) == arc_l2c_only)

#define	HDR_IN_HASH_TABLE(hdr)	((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
#define	HDR_IO_IN_PROGRESS(hdr)	((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
#define	HDR_IO_ERROR(hdr)	((hdr)->b_flags & ARC_FLAG_IO_ERROR)
#define	HDR_PREFETCH(hdr)	((hdr)->b_flags & ARC_FLAG_PREFETCH)
#define	HDR_PRESCIENT_PREFETCH(hdr)	\
	((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
#define	HDR_COMPRESSION_ENABLED(hdr)	\
	((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)

#define	HDR_L2CACHE(hdr)	((hdr)->b_flags & ARC_FLAG_L2CACHE)
#define	HDR_L2_READING(hdr)	\
	(((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) &&	\
	((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
#define	HDR_L2_WRITING(hdr)	((hdr)->b_flags & ARC_FLAG_L2_WRITING)
#define	HDR_L2_EVICTED(hdr)	((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
#define	HDR_L2_WRITE_HEAD(hdr)	((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
#define	HDR_PROTECTED(hdr)	((hdr)->b_flags & ARC_FLAG_PROTECTED)
#define	HDR_NOAUTH(hdr)		((hdr)->b_flags & ARC_FLAG_NOAUTH)
#define	HDR_SHARED_DATA(hdr)	((hdr)->b_flags & ARC_FLAG_SHARED_DATA)

#define	HDR_ISTYPE_METADATA(hdr)	\
	((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
#define	HDR_ISTYPE_DATA(hdr)	(!HDR_ISTYPE_METADATA(hdr))

#define	HDR_HAS_L1HDR(hdr)	((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
#define	HDR_HAS_L2HDR(hdr)	((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
#define	HDR_HAS_RABD(hdr)	\
	(HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) &&	\
	(hdr)->b_crypt_hdr.b_rabd != NULL)
#define	HDR_ENCRYPTED(hdr)	\
	(HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
#define	HDR_AUTHENTICATED(hdr)	\
	(HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))

/* For storing compression mode in b_flags */
#define	HDR_COMPRESS_OFFSET	(highbit64(ARC_FLAG_COMPRESS_0) - 1)

#define	HDR_GET_COMPRESS(hdr)	((enum zio_compress)BF32_GET((hdr)->b_flags, \
	HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
#define	HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
	HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));

#define	ARC_BUF_LAST(buf)	((buf)->b_next == NULL)
#define	ARC_BUF_SHARED(buf)	((buf)->b_flags & ARC_BUF_FLAG_SHARED)
#define	ARC_BUF_COMPRESSED(buf)	((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
#define	ARC_BUF_ENCRYPTED(buf)	((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED)

/*
 * Other sizes
 */

#define	HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
#define	HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr))
#define	HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))

/*
 * Hash table routines
 */

#define	HT_LOCK_PAD	64

struct ht_lock {
	kmutex_t	ht_lock;
#ifdef _KERNEL
	unsigned char	pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
#endif
};

#define	BUF_LOCKS 256
typedef struct buf_hash_table {
	uint64_t ht_mask;
	arc_buf_hdr_t **ht_table;
	struct ht_lock ht_locks[BUF_LOCKS];
} buf_hash_table_t;

static buf_hash_table_t buf_hash_table;

#define	BUF_HASH_INDEX(spa, dva, birth) \
	(buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
#define	BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
#define	BUF_HASH_LOCK(idx)	(&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
#define	HDR_LOCK(hdr) \
	(BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))

uint64_t zfs_crc64_table[256];

/*
 * Level 2 ARC
 */

#define	L2ARC_WRITE_SIZE	(8 * 1024 * 1024)	/* initial write max */
#define	L2ARC_HEADROOM		2			/* num of writes */
/*
 * If we discover during ARC scan any buffers to be compressed, we boost
 * our headroom for the next scanning cycle by this percentage multiple.
 */
#define	L2ARC_HEADROOM_BOOST	200
#define	L2ARC_FEED_SECS		1		/* caching interval secs */
#define	L2ARC_FEED_MIN_MS	200		/* min caching interval ms */

#define	l2arc_writes_sent	ARCSTAT(arcstat_l2_writes_sent)
#define	l2arc_writes_done	ARCSTAT(arcstat_l2_writes_done)

/* L2ARC Performance Tunables */
uint64_t l2arc_write_max = L2ARC_WRITE_SIZE;	/* default max write size */
uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE;	/* extra write during warmup */
uint64_t l2arc_headroom = L2ARC_HEADROOM;	/* number of dev writes */
uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
uint64_t l2arc_feed_secs = L2ARC_FEED_SECS;	/* interval seconds */
uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS;	/* min interval milliseconds */
boolean_t l2arc_noprefetch = B_TRUE;		/* don't cache prefetch bufs */
boolean_t l2arc_feed_again = B_TRUE;		/* turbo warmup */
boolean_t l2arc_norw = B_TRUE;			/* no reads during writes */

/*
 * L2ARC Internals
 */
struct l2arc_dev {
	vdev_t			*l2ad_vdev;	/* vdev */
	spa_t			*l2ad_spa;	/* spa */
	uint64_t		l2ad_hand;	/* next write location */
	uint64_t		l2ad_start;	/* first addr on device */
	uint64_t		l2ad_end;	/* last addr on device */
	boolean_t		l2ad_first;	/* first sweep through */
	boolean_t		l2ad_writing;	/* currently writing */
	kmutex_t		l2ad_mtx;	/* lock for buffer list */
	list_t			l2ad_buflist;	/* buffer list */
	list_node_t		l2ad_node;	/* device list node */
	zfs_refcount_t		l2ad_alloc;	/* allocated bytes */
};

static list_t L2ARC_dev_list;			/* device list */
static list_t *l2arc_dev_list;			/* device list pointer */
static kmutex_t l2arc_dev_mtx;			/* device list mutex */
static l2arc_dev_t *l2arc_dev_last;		/* last device used */
static list_t L2ARC_free_on_write;		/* free after write buf list */
static list_t *l2arc_free_on_write;		/* free after write list ptr */
static kmutex_t l2arc_free_on_write_mtx;	/* mutex for list */
static uint64_t l2arc_ndev;			/* number of devices */

typedef struct l2arc_read_callback {
	arc_buf_hdr_t		*l2rcb_hdr;		/* read header */
	blkptr_t		l2rcb_bp;		/* original blkptr */
	zbookmark_phys_t	l2rcb_zb;		/* original bookmark */
	int			l2rcb_flags;		/* original flags */
	abd_t			*l2rcb_abd;		/* temporary buffer */
} l2arc_read_callback_t;

typedef struct l2arc_write_callback {
	l2arc_dev_t	*l2wcb_dev;		/* device info */
	arc_buf_hdr_t	*l2wcb_head;		/* head of write buflist */
} l2arc_write_callback_t;

typedef struct l2arc_data_free {
	/* protected by l2arc_free_on_write_mtx */
	abd_t		*l2df_abd;
	size_t		l2df_size;
	arc_buf_contents_t l2df_type;
	list_node_t	l2df_list_node;
} l2arc_data_free_t;

static kmutex_t l2arc_feed_thr_lock;
static kcondvar_t l2arc_feed_thr_cv;
static uint8_t l2arc_thread_exit;

static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
typedef enum arc_fill_flags {
	ARC_FILL_LOCKED		= 1 << 0, /* hdr lock is held */
	ARC_FILL_COMPRESSED	= 1 << 1, /* fill with compressed data */
	ARC_FILL_ENCRYPTED	= 1 << 2, /* fill with encrypted data */
	ARC_FILL_NOAUTH		= 1 << 3, /* don't attempt to authenticate */
	ARC_FILL_IN_PLACE	= 1 << 4  /* fill in place (special case) */
} arc_fill_flags_t;

static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
static void arc_hdr_free_pabd(arc_buf_hdr_t *, boolean_t);
static void arc_hdr_alloc_pabd(arc_buf_hdr_t *, boolean_t);
static void arc_access(arc_buf_hdr_t *, kmutex_t *);
static boolean_t arc_is_overflowing();
static void arc_buf_watch(arc_buf_t *);

static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);

static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
static void l2arc_read_done(zio_t *);

/*
 * The arc_all_memory function is a ZoL enhancement that lives in their OSL
 * code. In user-space code, which is used primarily for testing, we return
 * half of all memory.
 */
uint64_t
arc_all_memory(void)
{
#ifdef _KERNEL
	return (ptob(physmem));
#else
	return ((sysconf(_SC_PAGESIZE) * sysconf(_SC_PHYS_PAGES)) / 2);
#endif
}

/*
 * We use Cityhash for this. It's fast, and has good hash properties without
 * requiring any large static buffers.
 */
static uint64_t
buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
{
	return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
}

#define	HDR_EMPTY(hdr)						\
	((hdr)->b_dva.dva_word[0] == 0 &&			\
	(hdr)->b_dva.dva_word[1] == 0)

#define	HDR_EQUAL(spa, dva, birth, hdr)				\
	((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) &&	\
	((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) &&	\
	((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)

static void
buf_discard_identity(arc_buf_hdr_t *hdr)
{
	hdr->b_dva.dva_word[0] = 0;
	hdr->b_dva.dva_word[1] = 0;
	hdr->b_birth = 0;
}

static arc_buf_hdr_t *
buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
{
	const dva_t *dva = BP_IDENTITY(bp);
	uint64_t birth = BP_PHYSICAL_BIRTH(bp);
	uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
	arc_buf_hdr_t *hdr;

	mutex_enter(hash_lock);
	for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
	    hdr = hdr->b_hash_next) {
		if (HDR_EQUAL(spa, dva, birth, hdr)) {
			*lockp = hash_lock;
			return (hdr);
		}
	}
	mutex_exit(hash_lock);
	*lockp = NULL;
	return (NULL);
}

/*
 * Insert an entry into the hash table.  If there is already an element
 * equal to elem in the hash table, then the already existing element
 * will be returned and the new element will not be inserted.
 * Otherwise returns NULL.
 * If lockp == NULL, the caller is assumed to already hold the hash lock.
 */
static arc_buf_hdr_t *
buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
{
	uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
	arc_buf_hdr_t *fhdr;
	uint32_t i;

	ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
	ASSERT(hdr->b_birth != 0);
	ASSERT(!HDR_IN_HASH_TABLE(hdr));

	if (lockp != NULL) {
		*lockp = hash_lock;
		mutex_enter(hash_lock);
	} else {
		ASSERT(MUTEX_HELD(hash_lock));
	}

	for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
	    fhdr = fhdr->b_hash_next, i++) {
		if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
			return (fhdr);
	}

	hdr->b_hash_next = buf_hash_table.ht_table[idx];
	buf_hash_table.ht_table[idx] = hdr;
	arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);

	/* collect some hash table performance data */
	if (i > 0) {
		ARCSTAT_BUMP(arcstat_hash_collisions);
		if (i == 1)
			ARCSTAT_BUMP(arcstat_hash_chains);

		ARCSTAT_MAX(arcstat_hash_chain_max, i);
	}

	ARCSTAT_BUMP(arcstat_hash_elements);
	ARCSTAT_MAXSTAT(arcstat_hash_elements);

	return (NULL);
}

static void
buf_hash_remove(arc_buf_hdr_t *hdr)
{
	arc_buf_hdr_t *fhdr, **hdrp;
	uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);

	ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
	ASSERT(HDR_IN_HASH_TABLE(hdr));

	hdrp = &buf_hash_table.ht_table[idx];
	while ((fhdr = *hdrp) != hdr) {
		ASSERT3P(fhdr, !=, NULL);
		hdrp = &fhdr->b_hash_next;
	}
	*hdrp = hdr->b_hash_next;
	hdr->b_hash_next = NULL;
	arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);

	/* collect some hash table performance data */
	ARCSTAT_BUMPDOWN(arcstat_hash_elements);

	if (buf_hash_table.ht_table[idx] &&
	    buf_hash_table.ht_table[idx]->b_hash_next == NULL)
		ARCSTAT_BUMPDOWN(arcstat_hash_chains);
}

/*
 * Global data structures and functions for the buf kmem cache.
 */

static kmem_cache_t *hdr_full_cache;
static kmem_cache_t *hdr_full_crypt_cache;
static kmem_cache_t *hdr_l2only_cache;
static kmem_cache_t *buf_cache;

static void
buf_fini(void)
{
	int i;

	kmem_free(buf_hash_table.ht_table,
	    (buf_hash_table.ht_mask + 1) * sizeof (void *));
	for (i = 0; i < BUF_LOCKS; i++)
		mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
	kmem_cache_destroy(hdr_full_cache);
	kmem_cache_destroy(hdr_full_crypt_cache);
	kmem_cache_destroy(hdr_l2only_cache);
	kmem_cache_destroy(buf_cache);
}

/*
 * Constructor callback - called when the cache is empty
 * and a new buf is requested.
 */
/* ARGSUSED */
static int
hdr_full_cons(void *vbuf, void *unused, int kmflag)
{
	arc_buf_hdr_t *hdr = vbuf;

	bzero(hdr, HDR_FULL_SIZE);
	hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
	cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
	zfs_refcount_create(&hdr->b_l1hdr.b_refcnt);
	mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
	multilist_link_init(&hdr->b_l1hdr.b_arc_node);
	arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);

	return (0);
}

/* ARGSUSED */
static int
hdr_full_crypt_cons(void *vbuf, void *unused, int kmflag)
{
	arc_buf_hdr_t *hdr = vbuf;

	(void) hdr_full_cons(vbuf, unused, kmflag);
	bzero(&hdr->b_crypt_hdr, sizeof (hdr->b_crypt_hdr));
	arc_space_consume(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);

	return (0);
}

/* ARGSUSED */
static int
hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
{
	arc_buf_hdr_t *hdr = vbuf;

	bzero(hdr, HDR_L2ONLY_SIZE);
	arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);

	return (0);
}

/* ARGSUSED */
static int
buf_cons(void *vbuf, void *unused, int kmflag)
{
	arc_buf_t *buf = vbuf;

	bzero(buf, sizeof (arc_buf_t));
	mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
	arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);

	return (0);
}

/*
 * Destructor callback - called when a cached buf is
 * no longer required.
 */
/* ARGSUSED */
static void
hdr_full_dest(void *vbuf, void *unused)
{
	arc_buf_hdr_t *hdr = vbuf;

	ASSERT(HDR_EMPTY(hdr));
	cv_destroy(&hdr->b_l1hdr.b_cv);
	zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt);
	mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
	ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
	arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
}

/* ARGSUSED */
static void
hdr_full_crypt_dest(void *vbuf, void *unused)
{
	arc_buf_hdr_t *hdr = vbuf;

	hdr_full_dest(hdr, unused);
	arc_space_return(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
}

/* ARGSUSED */
static void
hdr_l2only_dest(void *vbuf, void *unused)
{
	arc_buf_hdr_t *hdr = vbuf;

	ASSERT(HDR_EMPTY(hdr));
	arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
}

/* ARGSUSED */
static void
buf_dest(void *vbuf, void *unused)
{
	arc_buf_t *buf = vbuf;

	mutex_destroy(&buf->b_evict_lock);
	arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
}

/*
 * Reclaim callback -- invoked when memory is low.
 */
/* ARGSUSED */
static void
hdr_recl(void *unused)
{
	dprintf("hdr_recl called\n");
	/*
	 * umem calls the reclaim func when we destroy the buf cache,
	 * which is after we do arc_fini().
	 */
	if (arc_initialized)
		zthr_wakeup(arc_reap_zthr);
}

static void
buf_init(void)
{
	uint64_t *ct;
	uint64_t hsize = 1ULL << 12;
	int i, j;

	/*
	 * The hash table is big enough to fill all of physical memory
	 * with an average block size of zfs_arc_average_blocksize (default 8K).
	 * By default, the table will take up
	 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
	 */
	while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
		hsize <<= 1;
retry:
	buf_hash_table.ht_mask = hsize - 1;
	buf_hash_table.ht_table =
	    kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
	if (buf_hash_table.ht_table == NULL) {
		ASSERT(hsize > (1ULL << 8));
		hsize >>= 1;
		goto retry;
	}

	hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
	    0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
	hdr_full_crypt_cache = kmem_cache_create("arc_buf_hdr_t_full_crypt",
	    HDR_FULL_CRYPT_SIZE, 0, hdr_full_crypt_cons, hdr_full_crypt_dest,
	    hdr_recl, NULL, NULL, 0);
	hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
	    HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
	    NULL, NULL, 0);
	buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
	    0, buf_cons, buf_dest, NULL, NULL, NULL, 0);

	for (i = 0; i < 256; i++)
		for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
			*ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);

	for (i = 0; i < BUF_LOCKS; i++) {
		mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
		    NULL, MUTEX_DEFAULT, NULL);
	}
}

/*
 * This is the size that the buf occupies in memory. If the buf is compressed,
 * it will correspond to the compressed size. You should use this method of
 * getting the buf size unless you explicitly need the logical size.
 */
int32_t
arc_buf_size(arc_buf_t *buf)
{
	return (ARC_BUF_COMPRESSED(buf) ?
	    HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
}

int32_t
arc_buf_lsize(arc_buf_t *buf)
{
	return (HDR_GET_LSIZE(buf->b_hdr));
}

/*
 * This function will return B_TRUE if the buffer is encrypted in memory.
 * This buffer can be decrypted by calling arc_untransform().
 */
boolean_t
arc_is_encrypted(arc_buf_t *buf)
{
	return (ARC_BUF_ENCRYPTED(buf) != 0);
}

/*
 * Returns B_TRUE if the buffer represents data that has not had its MAC
 * verified yet.
 */
boolean_t
arc_is_unauthenticated(arc_buf_t *buf)
{
	return (HDR_NOAUTH(buf->b_hdr) != 0);
}

void
arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt,
    uint8_t *iv, uint8_t *mac)
{
	arc_buf_hdr_t *hdr = buf->b_hdr;

	ASSERT(HDR_PROTECTED(hdr));

	bcopy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
	bcopy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
	bcopy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
	*byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
	    /* CONSTCOND */
	    ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
}

/*
 * Indicates how this buffer is compressed in memory. If it is not compressed
 * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
 * arc_untransform() as long as it is also unencrypted.
 */
enum zio_compress
arc_get_compression(arc_buf_t *buf)
{
	return (ARC_BUF_COMPRESSED(buf) ?
	    HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
}

#define	ARC_MINTIME	(hz>>4) /* 62 ms */

/*
 * Return the compression algorithm used to store this data in the ARC. If ARC
 * compression is enabled or this is an encrypted block, this will be the same
 * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
 */
static inline enum zio_compress
arc_hdr_get_compress(arc_buf_hdr_t *hdr)
{
	return (HDR_COMPRESSION_ENABLED(hdr) ?
	    HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF);
}

static inline boolean_t
arc_buf_is_shared(arc_buf_t *buf)
{
	boolean_t shared = (buf->b_data != NULL &&
	    buf->b_hdr->b_l1hdr.b_pabd != NULL &&
	    abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
	    buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
	IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
	IMPLY(shared, ARC_BUF_SHARED(buf));
	IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));

	/*
	 * It would be nice to assert arc_can_share() too, but the "hdr isn't
	 * already being shared" requirement prevents us from doing that.
	 */

	return (shared);
}

/*
 * Free the checksum associated with this header. If there is no checksum, this
 * is a no-op.
 */
static inline void
arc_cksum_free(arc_buf_hdr_t *hdr)
{
	ASSERT(HDR_HAS_L1HDR(hdr));

	mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
	if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
		kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
		hdr->b_l1hdr.b_freeze_cksum = NULL;
	}
	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
}

/*
 * Return true iff at least one of the bufs on hdr is not compressed.
 * Encrypted buffers count as compressed.
 */
static boolean_t
arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
{
	ASSERT(hdr->b_l1hdr.b_state == arc_anon ||
	    MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));

	for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
		if (!ARC_BUF_COMPRESSED(b)) {
			return (B_TRUE);
		}
	}
	return (B_FALSE);
}

/*
 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
 * matches the checksum that is stored in the hdr. If there is no checksum,
 * or if the buf is compressed, this is a no-op.
 */
static void
arc_cksum_verify(arc_buf_t *buf)
{
	arc_buf_hdr_t *hdr = buf->b_hdr;
	zio_cksum_t zc;

	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
		return;

	if (ARC_BUF_COMPRESSED(buf))
		return;

	ASSERT(HDR_HAS_L1HDR(hdr));

	mutex_enter(&hdr->b_l1hdr.b_freeze_lock);

	if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
		mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
		return;
	}

	fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
	if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
		panic("buffer modified while frozen!");
	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
}

/*
 * This function makes the assumption that data stored in the L2ARC
 * will be transformed exactly as it is in the main pool. Because of
 * this we can verify the checksum against the reading process's bp.
 */
static boolean_t
arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
{
	enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
	boolean_t valid_cksum;

	ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
	VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));

	/*
	 * We rely on the blkptr's checksum to determine if the block
	 * is valid or not. When compressed arc is enabled, the l2arc
	 * writes the block to the l2arc just as it appears in the pool.
	 * This allows us to use the blkptr's checksum to validate the
	 * data that we just read off of the l2arc without having to store
	 * a separate checksum in the arc_buf_hdr_t. However, if compressed
	 * arc is disabled, then the data written to the l2arc is always
	 * uncompressed and won't match the block as it exists in the main
	 * pool. When this is the case, we must first compress it if it is
	 * compressed on the main pool before we can validate the checksum.
	 */
	if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
		ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
		uint64_t lsize = HDR_GET_LSIZE(hdr);
		uint64_t csize;

		abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
		csize = zio_compress_data(compress, zio->io_abd,
		    abd_to_buf(cdata), lsize);

		ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
		if (csize < HDR_GET_PSIZE(hdr)) {
			/*
			 * Compressed blocks are always a multiple of the
			 * smallest ashift in the pool. Ideally, we would
			 * like to round up the csize to the next
			 * spa_min_ashift but that value may have changed
			 * since the block was last written. Instead,
			 * we rely on the fact that the hdr's psize
			 * was set to the psize of the block when it was
			 * last written. We set the csize to that value
			 * and zero out any part that should not contain
			 * data.
			 */
			abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
			csize = HDR_GET_PSIZE(hdr);
		}
		zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
	}

	/*
	 * Block pointers always store the checksum for the logical data.
	 * If the block pointer has the gang bit set, then the checksum
	 * it represents is for the reconstituted data and not for an
	 * individual gang member. The zio pipeline, however, must be able to
	 * determine the checksum of each of the gang constituents so it
	 * treats the checksum comparison differently than what we need
	 * for l2arc blocks. This prevents us from using the
	 * zio_checksum_error() interface directly. Instead we must call the
	 * zio_checksum_error_impl() so that we can ensure the checksum is
	 * generated using the correct checksum algorithm and accounts for the
	 * logical I/O size and not just a gang fragment.
	 */
	valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
	    BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
	    zio->io_offset, NULL) == 0);
	zio_pop_transforms(zio);
	return (valid_cksum);
}

/*
 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
 * isn't modified later on. If buf is compressed or there is already a checksum
 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
 */
static void
arc_cksum_compute(arc_buf_t *buf)
{
	arc_buf_hdr_t *hdr = buf->b_hdr;

	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
		return;

	ASSERT(HDR_HAS_L1HDR(hdr));

	mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
	if (hdr->b_l1hdr.b_freeze_cksum != NULL || ARC_BUF_COMPRESSED(buf)) {
		mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
		return;
	}

	ASSERT(!ARC_BUF_ENCRYPTED(buf));
	ASSERT(!ARC_BUF_COMPRESSED(buf));
	hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
	    KM_SLEEP);
	fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
	    hdr->b_l1hdr.b_freeze_cksum);
	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
	arc_buf_watch(buf);
}

#ifndef _KERNEL
typedef struct procctl {
	long cmd;
	prwatch_t prwatch;
} procctl_t;
#endif

/* ARGSUSED */
static void
arc_buf_unwatch(arc_buf_t *buf)
{
#ifndef _KERNEL
	if (arc_watch) {
		int result;
		procctl_t ctl;
		ctl.cmd = PCWATCH;
		ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
		ctl.prwatch.pr_size = 0;
		ctl.prwatch.pr_wflags = 0;
		result = write(arc_procfd, &ctl, sizeof (ctl));
		ASSERT3U(result, ==, sizeof (ctl));
	}
#endif
}

/* ARGSUSED */
static void
arc_buf_watch(arc_buf_t *buf)
{
#ifndef _KERNEL
	if (arc_watch) {
		int result;
		procctl_t ctl;
		ctl.cmd = PCWATCH;
		ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
		ctl.prwatch.pr_size = arc_buf_size(buf);
		ctl.prwatch.pr_wflags = WA_WRITE;
		result = write(arc_procfd, &ctl, sizeof (ctl));
		ASSERT3U(result, ==, sizeof (ctl));
	}
#endif
}

static arc_buf_contents_t
arc_buf_type(arc_buf_hdr_t *hdr)
{
	arc_buf_contents_t type;
	if (HDR_ISTYPE_METADATA(hdr)) {
		type = ARC_BUFC_METADATA;
	} else {
		type = ARC_BUFC_DATA;
	}
	VERIFY3U(hdr->b_type, ==, type);
	return (type);
}

boolean_t
arc_is_metadata(arc_buf_t *buf)
{
	return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
}

static uint32_t
arc_bufc_to_flags(arc_buf_contents_t type)
{
	switch (type) {
	case ARC_BUFC_DATA:
		/* metadata field is 0 if buffer contains normal data */
		return (0);
	case ARC_BUFC_METADATA:
		return (ARC_FLAG_BUFC_METADATA);
	default:
		break;
	}
	panic("undefined ARC buffer type!");
	return ((uint32_t)-1);
}

void
arc_buf_thaw(arc_buf_t *buf)
{
	arc_buf_hdr_t *hdr = buf->b_hdr;

	ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
	ASSERT(!HDR_IO_IN_PROGRESS(hdr));

	arc_cksum_verify(buf);

	/*
	 * Compressed buffers do not manipulate the b_freeze_cksum.
	 */
	if (ARC_BUF_COMPRESSED(buf))
		return;

	ASSERT(HDR_HAS_L1HDR(hdr));
	arc_cksum_free(hdr);

	mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
#ifdef ZFS_DEBUG
	if (zfs_flags & ZFS_DEBUG_MODIFY) {
		if (hdr->b_l1hdr.b_thawed != NULL)
			kmem_free(hdr->b_l1hdr.b_thawed, 1);
		hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
	}
#endif

	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);

	arc_buf_unwatch(buf);
}

void
arc_buf_freeze(arc_buf_t *buf)
{
	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
		return;

	if (ARC_BUF_COMPRESSED(buf))
		return;

	ASSERT(HDR_HAS_L1HDR(buf->b_hdr));
	arc_cksum_compute(buf);
}

/*
 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
 * the following functions should be used to ensure that the flags are
 * updated in a thread-safe way. When manipulating the flags either
 * the hash_lock must be held or the hdr must be undiscoverable. This
 * ensures that we're not racing with any other threads when updating
 * the flags.
 */
static inline void
arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
{
	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
	hdr->b_flags |= flags;
}

static inline void
arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
{
	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
	hdr->b_flags &= ~flags;
}

/*
 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
 * done in a special way since we have to clear and set bits
 * at the same time. Consumers that wish to set the compression bits
 * must use this function to ensure that the flags are updated in
 * thread-safe manner.
 */
static void
arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
{
	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));

	/*
	 * Holes and embedded blocks will always have a psize = 0 so
	 * we ignore the compression of the blkptr and set the
	 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
	 * Holes and embedded blocks remain anonymous so we don't
	 * want to uncompress them. Mark them as uncompressed.
	 */
	if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
		arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
		ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
	} else {
		arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
		ASSERT(HDR_COMPRESSION_ENABLED(hdr));
	}

	HDR_SET_COMPRESS(hdr, cmp);
	ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
}

/*
 * Looks for another buf on the same hdr which has the data decompressed, copies
 * from it, and returns true. If no such buf exists, returns false.
 */
static boolean_t
arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
{
	arc_buf_hdr_t *hdr = buf->b_hdr;
	boolean_t copied = B_FALSE;

	ASSERT(HDR_HAS_L1HDR(hdr));
	ASSERT3P(buf->b_data, !=, NULL);
	ASSERT(!ARC_BUF_COMPRESSED(buf));

	for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
	    from = from->b_next) {
		/* can't use our own data buffer */
		if (from == buf) {
			continue;
		}

		if (!ARC_BUF_COMPRESSED(from)) {
			bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
			copied = B_TRUE;
			break;
		}
	}

	/*
	 * Note: With encryption support, the following assertion is no longer
	 * necessarily valid. If we receive two back to back raw snapshots
	 * (send -w), the second receive can use a hdr with a cksum already
	 * calculated. This happens via:
	 *    dmu_recv_stream() -> receive_read_record() -> arc_loan_raw_buf()
	 * The rsend/send_mixed_raw test case exercises this code path.
	 *
	 * There were no decompressed bufs, so there should not be a
	 * checksum on the hdr either.
	 * EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
	 */

	return (copied);
}

/*
 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
 */
static uint64_t
arc_hdr_size(arc_buf_hdr_t *hdr)
{
	uint64_t size;

	if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
	    HDR_GET_PSIZE(hdr) > 0) {
		size = HDR_GET_PSIZE(hdr);
	} else {
		ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
		size = HDR_GET_LSIZE(hdr);
	}
	return (size);
}

static int
arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj)
{
	int ret;
	uint64_t csize;
	uint64_t lsize = HDR_GET_LSIZE(hdr);
	uint64_t psize = HDR_GET_PSIZE(hdr);
	void *tmpbuf = NULL;
	abd_t *abd = hdr->b_l1hdr.b_pabd;

	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
	ASSERT(HDR_AUTHENTICATED(hdr));
	ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);

	/*
	 * The MAC is calculated on the compressed data that is stored on disk.
	 * However, if compressed arc is disabled we will only have the
	 * decompressed data available to us now. Compress it into a temporary
	 * abd so we can verify the MAC. The performance overhead of this will
	 * be relatively low, since most objects in an encrypted objset will
	 * be encrypted (instead of authenticated) anyway.
	 */
	if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
	    !HDR_COMPRESSION_ENABLED(hdr)) {
		tmpbuf = zio_buf_alloc(lsize);
		abd = abd_get_from_buf(tmpbuf, lsize);
		abd_take_ownership_of_buf(abd, B_TRUE);

		csize = zio_compress_data(HDR_GET_COMPRESS(hdr),
		    hdr->b_l1hdr.b_pabd, tmpbuf, lsize);
		ASSERT3U(csize, <=, psize);
		abd_zero_off(abd, csize, psize - csize);
	}

	/*
	 * Authentication is best effort. We authenticate whenever the key is
	 * available. If we succeed we clear ARC_FLAG_NOAUTH.
	 */
	if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) {
		ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
		ASSERT3U(lsize, ==, psize);
		ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd,
		    psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
	} else {
		ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize,
		    hdr->b_crypt_hdr.b_mac);
	}

	if (ret == 0)
		arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH);
	else if (ret != ENOENT)
		goto error;

	if (tmpbuf != NULL)
		abd_free(abd);

	return (0);

error:
	if (tmpbuf != NULL)
		abd_free(abd);

	return (ret);
}

/*
 * This function will take a header that only has raw encrypted data in
 * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
 * b_l1hdr.b_pabd. If designated in the header flags, this function will
 * also decompress the data.
 */
static int
arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb)
{
	int ret;
	abd_t *cabd = NULL;
	void *tmp = NULL;
	boolean_t no_crypt = B_FALSE;
	boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);

	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
	ASSERT(HDR_ENCRYPTED(hdr));

	arc_hdr_alloc_pabd(hdr, B_FALSE);

	ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot,
	    B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv,
	    hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd,
	    hdr->b_crypt_hdr.b_rabd, &no_crypt);
	if (ret != 0)
		goto error;

	if (no_crypt) {
		abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd,
		    HDR_GET_PSIZE(hdr));
	}

	/*
	 * If this header has disabled arc compression but the b_pabd is
	 * compressed after decrypting it, we need to decompress the newly
	 * decrypted data.
	 */
	if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
	    !HDR_COMPRESSION_ENABLED(hdr)) {
		/*
		 * We want to make sure that we are correctly honoring the
		 * zfs_abd_scatter_enabled setting, so we allocate an abd here
		 * and then loan a buffer from it, rather than allocating a
		 * linear buffer and wrapping it in an abd later.
		 */
		cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
		tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));

		ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
		    hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
		    HDR_GET_LSIZE(hdr));
		if (ret != 0) {
			abd_return_buf(cabd, tmp, arc_hdr_size(hdr));
			goto error;
		}

		abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
		arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
		    arc_hdr_size(hdr), hdr);
		hdr->b_l1hdr.b_pabd = cabd;
	}

	return (0);

error:
	arc_hdr_free_pabd(hdr, B_FALSE);
	if (cabd != NULL)
		arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr);

	return (ret);
}

/*
 * This function is called during arc_buf_fill() to prepare the header's
 * abd plaintext pointer for use. This involves authenticated protected
 * data and decrypting encrypted data into the plaintext abd.
 */
static int
arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa,
    const zbookmark_phys_t *zb, boolean_t noauth)
{
	int ret;

	ASSERT(HDR_PROTECTED(hdr));

	if (hash_lock != NULL)
		mutex_enter(hash_lock);

	if (HDR_NOAUTH(hdr) && !noauth) {
		/*
		 * The caller requested authenticated data but our data has
		 * not been authenticated yet. Verify the MAC now if we can.
		 */
		ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset);
		if (ret != 0)
			goto error;
	} else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) {
		/*
		 * If we only have the encrypted version of the data, but the
		 * unencrypted version was requested we take this opportunity
		 * to store the decrypted version in the header for future use.
		 */
		ret = arc_hdr_decrypt(hdr, spa, zb);
		if (ret != 0)
			goto error;
	}

	ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);

	if (hash_lock != NULL)
		mutex_exit(hash_lock);

	return (0);

error:
	if (hash_lock != NULL)
		mutex_exit(hash_lock);

	return (ret);
}

/*
 * This function is used by the dbuf code to decrypt bonus buffers in place.
 * The dbuf code itself doesn't have any locking for decrypting a shared dnode
 * block, so we use the hash lock here to protect against concurrent calls to
 * arc_buf_fill().
 */
/* ARGSUSED */
static void
arc_buf_untransform_in_place(arc_buf_t *buf, kmutex_t *hash_lock)
{
	arc_buf_hdr_t *hdr = buf->b_hdr;

	ASSERT(HDR_ENCRYPTED(hdr));
	ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
	ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);

	zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data,
	    arc_buf_size(buf));
	buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
	buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
	hdr->b_crypt_hdr.b_ebufcnt -= 1;
}

/*
 * Given a buf that has a data buffer attached to it, this function will
 * efficiently fill the buf with data of the specified compression setting from
 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
 * are already sharing a data buf, no copy is performed.
 *
 * If the buf is marked as compressed but uncompressed data was requested, this
 * will allocate a new data buffer for the buf, remove that flag, and fill the
 * buf with uncompressed data. You can't request a compressed buf on a hdr with
 * uncompressed data, and (since we haven't added support for it yet) if you
 * want compressed data your buf must already be marked as compressed and have
 * the correct-sized data buffer.
 */
static int
arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
    arc_fill_flags_t flags)
{
	int error = 0;
	arc_buf_hdr_t *hdr = buf->b_hdr;
	boolean_t hdr_compressed =
	    (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
	boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0;
	boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0;
	dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
	kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr);

	ASSERT3P(buf->b_data, !=, NULL);
	IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf));
	IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
	IMPLY(encrypted, HDR_ENCRYPTED(hdr));
	IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf));
	IMPLY(encrypted, ARC_BUF_COMPRESSED(buf));
	IMPLY(encrypted, !ARC_BUF_SHARED(buf));

	/*
	 * If the caller wanted encrypted data we just need to copy it from
	 * b_rabd and potentially byteswap it. We won't be able to do any
	 * further transforms on it.
	 */
	if (encrypted) {
		ASSERT(HDR_HAS_RABD(hdr));
		abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd,
		    HDR_GET_PSIZE(hdr));
		goto byteswap;
	}

	/*
	 * Adjust encrypted and authenticated headers to accomodate
	 * the request if needed. Dnode blocks (ARC_FILL_IN_PLACE) are
	 * allowed to fail decryption due to keys not being loaded
	 * without being marked as an IO error.
	 */
	if (HDR_PROTECTED(hdr)) {
		error = arc_fill_hdr_crypt(hdr, hash_lock, spa,
		    zb, !!(flags & ARC_FILL_NOAUTH));
		if (error == EACCES && (flags & ARC_FILL_IN_PLACE) != 0) {
			return (error);
		} else if (error != 0) {
			if (hash_lock != NULL)
				mutex_enter(hash_lock);
			arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
			if (hash_lock != NULL)
				mutex_exit(hash_lock);
			return (error);
		}
	}

	/*
	 * There is a special case here for dnode blocks which are
	 * decrypting their bonus buffers. These blocks may request to
	 * be decrypted in-place. This is necessary because there may
	 * be many dnodes pointing into this buffer and there is
	 * currently no method to synchronize replacing the backing
	 * b_data buffer and updating all of the pointers. Here we use
	 * the hash lock to ensure there are no races. If the need
	 * arises for other types to be decrypted in-place, they must
	 * add handling here as well.
	 */
	if ((flags & ARC_FILL_IN_PLACE) != 0) {
		ASSERT(!hdr_compressed);
		ASSERT(!compressed);
		ASSERT(!encrypted);

		if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) {
			ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);

			if (hash_lock != NULL)
				mutex_enter(hash_lock);
			arc_buf_untransform_in_place(buf, hash_lock);
			if (hash_lock != NULL)
				mutex_exit(hash_lock);

			/* Compute the hdr's checksum if necessary */
			arc_cksum_compute(buf);
		}

		return (0);
	}

	if (hdr_compressed == compressed) {
		if (!arc_buf_is_shared(buf)) {
			abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
			    arc_buf_size(buf));
		}
	} else {
		ASSERT(hdr_compressed);
		ASSERT(!compressed);
		ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));

		/*
		 * If the buf is sharing its data with the hdr, unlink it and
		 * allocate a new data buffer for the buf.
		 */
		if (arc_buf_is_shared(buf)) {
			ASSERT(ARC_BUF_COMPRESSED(buf));

			/* We need to give the buf its own b_data */
			buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
			buf->b_data =
			    arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
			arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);

			/* Previously overhead was 0; just add new overhead */
			ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
		} else if (ARC_BUF_COMPRESSED(buf)) {
			/* We need to reallocate the buf's b_data */
			arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
			    buf);
			buf->b_data =
			    arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);

			/* We increased the size of b_data; update overhead */
			ARCSTAT_INCR(arcstat_overhead_size,
			    HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
		}

		/*
		 * Regardless of the buf's previous compression settings, it
		 * should not be compressed at the end of this function.
		 */
		buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;

		/*
		 * Try copying the data from another buf which already has a
		 * decompressed version. If that's not possible, it's time to
		 * bite the bullet and decompress the data from the hdr.
		 */
		if (arc_buf_try_copy_decompressed_data(buf)) {
			/* Skip byteswapping and checksumming (already done) */
			ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
			return (0);
		} else {
			error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
			    hdr->b_l1hdr.b_pabd, buf->b_data,
			    HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));

			/*
			 * Absent hardware errors or software bugs, this should
			 * be impossible, but log it anyway so we can debug it.
			 */
			if (error != 0) {
				zfs_dbgmsg(
				    "hdr %p, compress %d, psize %d, lsize %d",
				    hdr, arc_hdr_get_compress(hdr),
				    HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
				if (hash_lock != NULL)
					mutex_enter(hash_lock);
				arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
				if (hash_lock != NULL)
					mutex_exit(hash_lock);
				return (SET_ERROR(EIO));
			}
		}
	}

byteswap:
	/* Byteswap the buf's data if necessary */
	if (bswap != DMU_BSWAP_NUMFUNCS) {
		ASSERT(!HDR_SHARED_DATA(hdr));
		ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
		dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
	}

	/* Compute the hdr's checksum if necessary */
	arc_cksum_compute(buf);

	return (0);
}

/*
 * If this function is being called to decrypt an encrypted buffer or verify an
 * authenticated one, the key must be loaded and a mapping must be made
 * available in the keystore via spa_keystore_create_mapping() or one of its
 * callers.
 */
int
arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
    boolean_t in_place)
{
	int ret;
	arc_fill_flags_t flags = 0;

	if (in_place)
		flags |= ARC_FILL_IN_PLACE;

	ret = arc_buf_fill(buf, spa, zb, flags);
	if (ret == ECKSUM) {
		/*
		 * Convert authentication and decryption errors to EIO
		 * (and generate an ereport) before leaving the ARC.
		 */
		ret = SET_ERROR(EIO);
		spa_log_error(spa, zb);
		zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
		    spa, NULL, zb, NULL, 0, 0);
	}

	return (ret);
}

/*
 * Increment the amount of evictable space in the arc_state_t's refcount.
 * We account for the space used by the hdr and the arc buf individually
 * so that we can add and remove them from the refcount individually.
 */
static void
arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
{
	arc_buf_contents_t type = arc_buf_type(hdr);

	ASSERT(HDR_HAS_L1HDR(hdr));

	if (GHOST_STATE(state)) {
		ASSERT0(hdr->b_l1hdr.b_bufcnt);
		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
		ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
		ASSERT(!HDR_HAS_RABD(hdr));
		(void) zfs_refcount_add_many(&state->arcs_esize[type],
		    HDR_GET_LSIZE(hdr), hdr);
		return;
	}

	ASSERT(!GHOST_STATE(state));
	if (hdr->b_l1hdr.b_pabd != NULL) {
		(void) zfs_refcount_add_many(&state->arcs_esize[type],
		    arc_hdr_size(hdr), hdr);
	}
	if (HDR_HAS_RABD(hdr)) {
		(void) zfs_refcount_add_many(&state->arcs_esize[type],
		    HDR_GET_PSIZE(hdr), hdr);
	}
	for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
	    buf = buf->b_next) {
		if (arc_buf_is_shared(buf))
			continue;
		(void) zfs_refcount_add_many(&state->arcs_esize[type],
		    arc_buf_size(buf), buf);
	}
}

/*
 * Decrement the amount of evictable space in the arc_state_t's refcount.
 * We account for the space used by the hdr and the arc buf individually
 * so that we can add and remove them from the refcount individually.
 */
static void
arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
{
	arc_buf_contents_t type = arc_buf_type(hdr);

	ASSERT(HDR_HAS_L1HDR(hdr));

	if (GHOST_STATE(state)) {
		ASSERT0(hdr->b_l1hdr.b_bufcnt);
		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
		ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
		ASSERT(!HDR_HAS_RABD(hdr));
		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
		    HDR_GET_LSIZE(hdr), hdr);
		return;
	}

	ASSERT(!GHOST_STATE(state));
	if (hdr->b_l1hdr.b_pabd != NULL) {
		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
		    arc_hdr_size(hdr), hdr);
	}
	if (HDR_HAS_RABD(hdr)) {
		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
		    HDR_GET_PSIZE(hdr), hdr);
	}
	for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
	    buf = buf->b_next) {
		if (arc_buf_is_shared(buf))
			continue;
		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
		    arc_buf_size(buf), buf);
	}
}

/*
 * Add a reference to this hdr indicating that someone is actively
 * referencing that memory. When the refcount transitions from 0 to 1,
 * we remove it from the respective arc_state_t list to indicate that
 * it is not evictable.
 */
static void
add_reference(arc_buf_hdr_t *hdr, void *tag)
{
	ASSERT(HDR_HAS_L1HDR(hdr));
	if (!MUTEX_HELD(HDR_LOCK(hdr))) {
		ASSERT(hdr->b_l1hdr.b_state == arc_anon);
		ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
	}

	arc_state_t *state = hdr->b_l1hdr.b_state;

	if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
	    (state != arc_anon)) {
		/* We don't use the L2-only state list. */
		if (state != arc_l2c_only) {
			multilist_remove(state->arcs_list[arc_buf_type(hdr)],
			    hdr);
			arc_evictable_space_decrement(hdr, state);
		}
		/* remove the prefetch flag if we get a reference */
		arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
	}
}

/*
 * Remove a reference from this hdr. When the reference transitions from
 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
 * list making it eligible for eviction.
 */
static int
remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
{
	int cnt;
	arc_state_t *state = hdr->b_l1hdr.b_state;

	ASSERT(HDR_HAS_L1HDR(hdr));
	ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
	ASSERT(!GHOST_STATE(state));

	/*
	 * arc_l2c_only counts as a ghost state so we don't need to explicitly
	 * check to prevent usage of the arc_l2c_only list.
	 */
	if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
	    (state != arc_anon)) {
		multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
		ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
		arc_evictable_space_increment(hdr, state);
	}
	return (cnt);
}

/*
 * Move the supplied buffer to the indicated state. The hash lock
 * for the buffer must be held by the caller.
 */
static void
arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
    kmutex_t *hash_lock)
{
	arc_state_t *old_state;
	int64_t refcnt;
	uint32_t bufcnt;
	boolean_t update_old, update_new;
	arc_buf_contents_t buftype = arc_buf_type(hdr);

	/*
	 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
	 * in arc_read() when bringing a buffer out of the L2ARC.  However, the
	 * L1 hdr doesn't always exist when we change state to arc_anon before
	 * destroying a header, in which case reallocating to add the L1 hdr is
	 * pointless.
	 */
	if (HDR_HAS_L1HDR(hdr)) {
		old_state = hdr->b_l1hdr.b_state;
		refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt);
		bufcnt = hdr->b_l1hdr.b_bufcnt;

		update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL ||
		    HDR_HAS_RABD(hdr));
	} else {
		old_state = arc_l2c_only;
		refcnt = 0;
		bufcnt = 0;
		update_old = B_FALSE;
	}
	update_new = update_old;

	ASSERT(MUTEX_HELD(hash_lock));
	ASSERT3P(new_state, !=, old_state);
	ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
	ASSERT(old_state != arc_anon || bufcnt <= 1);

	/*
	 * If this buffer is evictable, transfer it from the
	 * old state list to the new state list.
	 */
	if (refcnt == 0) {
		if (old_state != arc_anon && old_state != arc_l2c_only) {
			ASSERT(HDR_HAS_L1HDR(hdr));
			multilist_remove(old_state->arcs_list[buftype], hdr);

			if (GHOST_STATE(old_state)) {
				ASSERT0(bufcnt);
				ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
				update_old = B_TRUE;
			}
			arc_evictable_space_decrement(hdr, old_state);
		}
		if (new_state != arc_anon && new_state != arc_l2c_only) {

			/*
			 * An L1 header always exists here, since if we're
			 * moving to some L1-cached state (i.e. not l2c_only or
			 * anonymous), we realloc the header to add an L1hdr
			 * beforehand.
			 */
			ASSERT(HDR_HAS_L1HDR(hdr));
			multilist_insert(new_state->arcs_list[buftype], hdr);

			if (GHOST_STATE(new_state)) {
				ASSERT0(bufcnt);
				ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
				update_new = B_TRUE;
			}
			arc_evictable_space_increment(hdr, new_state);
		}
	}

	ASSERT(!HDR_EMPTY(hdr));
	if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
		buf_hash_remove(hdr);

	/* adjust state sizes (ignore arc_l2c_only) */

	if (update_new && new_state != arc_l2c_only) {
		ASSERT(HDR_HAS_L1HDR(hdr));
		if (GHOST_STATE(new_state)) {
			ASSERT0(bufcnt);

			/*
			 * When moving a header to a ghost state, we first
			 * remove all arc buffers. Thus, we'll have a
			 * bufcnt of zero, and no arc buffer to use for
			 * the reference. As a result, we use the arc
			 * header pointer for the reference.
			 */
			(void) zfs_refcount_add_many(&new_state->arcs_size,
			    HDR_GET_LSIZE(hdr), hdr);
			ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
			ASSERT(!HDR_HAS_RABD(hdr));
		} else {
			uint32_t buffers = 0;

			/*
			 * Each individual buffer holds a unique reference,
			 * thus we must remove each of these references one
			 * at a time.
			 */
			for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
			    buf = buf->b_next) {
				ASSERT3U(bufcnt, !=, 0);
				buffers++;

				/*
				 * When the arc_buf_t is sharing the data
				 * block with the hdr, the owner of the
				 * reference belongs to the hdr. Only
				 * add to the refcount if the arc_buf_t is
				 * not shared.
				 */
				if (arc_buf_is_shared(buf))
					continue;

				(void) zfs_refcount_add_many(
				    &new_state->arcs_size,
				    arc_buf_size(buf), buf);
			}
			ASSERT3U(bufcnt, ==, buffers);

			if (hdr->b_l1hdr.b_pabd != NULL) {
				(void) zfs_refcount_add_many(
				    &new_state->arcs_size,
				    arc_hdr_size(hdr), hdr);
			}

			if (HDR_HAS_RABD(hdr)) {
				(void) zfs_refcount_add_many(
				    &new_state->arcs_size,
				    HDR_GET_PSIZE(hdr), hdr);
			}
		}
	}

	if (update_old && old_state != arc_l2c_only) {
		ASSERT(HDR_HAS_L1HDR(hdr));
		if (GHOST_STATE(old_state)) {
			ASSERT0(bufcnt);
			ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
			ASSERT(!HDR_HAS_RABD(hdr));

			/*
			 * When moving a header off of a ghost state,
			 * the header will not contain any arc buffers.
			 * We use the arc header pointer for the reference
			 * which is exactly what we did when we put the
			 * header on the ghost state.
			 */

			(void) zfs_refcount_remove_many(&old_state->arcs_size,
			    HDR_GET_LSIZE(hdr), hdr);
		} else {
			uint32_t buffers = 0;

			/*
			 * Each individual buffer holds a unique reference,
			 * thus we must remove each of these references one
			 * at a time.
			 */
			for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
			    buf = buf->b_next) {
				ASSERT3U(bufcnt, !=, 0);
				buffers++;

				/*
				 * When the arc_buf_t is sharing the data
				 * block with the hdr, the owner of the
				 * reference belongs to the hdr. Only
				 * add to the refcount if the arc_buf_t is
				 * not shared.
				 */
				if (arc_buf_is_shared(buf))
					continue;

				(void) zfs_refcount_remove_many(
				    &old_state->arcs_size, arc_buf_size(buf),
				    buf);
			}
			ASSERT3U(bufcnt, ==, buffers);
			ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
			    HDR_HAS_RABD(hdr));

			if (hdr->b_l1hdr.b_pabd != NULL) {
				(void) zfs_refcount_remove_many(
				    &old_state->arcs_size, arc_hdr_size(hdr),
				    hdr);
			}

			if (HDR_HAS_RABD(hdr)) {
				(void) zfs_refcount_remove_many(
				    &old_state->arcs_size, HDR_GET_PSIZE(hdr),
				    hdr);
			}
		}
	}

	if (HDR_HAS_L1HDR(hdr))
		hdr->b_l1hdr.b_state = new_state;

	/*
	 * L2 headers should never be on the L2 state list since they don't
	 * have L1 headers allocated.
	 */
	ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
	    multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
}

void
arc_space_consume(uint64_t space, arc_space_type_t type)
{
	ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);

	switch (type) {
	case ARC_SPACE_DATA:
		aggsum_add(&astat_data_size, space);
		break;
	case ARC_SPACE_META:
		aggsum_add(&astat_metadata_size, space);
		break;
	case ARC_SPACE_OTHER:
		aggsum_add(&astat_other_size, space);
		break;
	case ARC_SPACE_HDRS:
		aggsum_add(&astat_hdr_size, space);
		break;
	case ARC_SPACE_L2HDRS:
		aggsum_add(&astat_l2_hdr_size, space);
		break;
	}

	if (type != ARC_SPACE_DATA)
		aggsum_add(&arc_meta_used, space);

	aggsum_add(&arc_size, space);
}

void
arc_space_return(uint64_t space, arc_space_type_t type)
{
	ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);

	switch (type) {
	case ARC_SPACE_DATA:
		aggsum_add(&astat_data_size, -space);
		break;
	case ARC_SPACE_META:
		aggsum_add(&astat_metadata_size, -space);
		break;
	case ARC_SPACE_OTHER:
		aggsum_add(&astat_other_size, -space);
		break;
	case ARC_SPACE_HDRS:
		aggsum_add(&astat_hdr_size, -space);
		break;
	case ARC_SPACE_L2HDRS:
		aggsum_add(&astat_l2_hdr_size, -space);
		break;
	}

	if (type != ARC_SPACE_DATA) {
		ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
		/*
		 * We use the upper bound here rather than the precise value
		 * because the arc_meta_max value doesn't need to be
		 * precise. It's only consumed by humans via arcstats.
		 */
		if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
			arc_meta_max = aggsum_upper_bound(&arc_meta_used);
		aggsum_add(&arc_meta_used, -space);
	}

	ASSERT(aggsum_compare(&arc_size, space) >= 0);
	aggsum_add(&arc_size, -space);
}

/*
 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
 * with the hdr's b_pabd.
 */
static boolean_t
arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
{
	/*
	 * The criteria for sharing a hdr's data are:
	 * 1. the buffer is not encrypted
	 * 2. the hdr's compression matches the buf's compression
	 * 3. the hdr doesn't need to be byteswapped
	 * 4. the hdr isn't already being shared
	 * 5. the buf is either compressed or it is the last buf in the hdr list
	 *
	 * Criterion #5 maintains the invariant that shared uncompressed
	 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
	 * might ask, "if a compressed buf is allocated first, won't that be the
	 * last thing in the list?", but in that case it's impossible to create
	 * a shared uncompressed buf anyway (because the hdr must be compressed
	 * to have the compressed buf). You might also think that #3 is
	 * sufficient to make this guarantee, however it's possible
	 * (specifically in the rare L2ARC write race mentioned in
	 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
	 * is sharable, but wasn't at the time of its allocation. Rather than
	 * allow a new shared uncompressed buf to be created and then shuffle
	 * the list around to make it the last element, this simply disallows
	 * sharing if the new buf isn't the first to be added.
	 */
	ASSERT3P(buf->b_hdr, ==, hdr);
	boolean_t hdr_compressed = arc_hdr_get_compress(hdr) !=
	    ZIO_COMPRESS_OFF;
	boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
	return (!ARC_BUF_ENCRYPTED(buf) &&
	    buf_compressed == hdr_compressed &&
	    hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
	    !HDR_SHARED_DATA(hdr) &&
	    (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
}

/*
 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
 * copy was made successfully, or an error code otherwise.
 */
static int
arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb,
    void *tag, boolean_t encrypted, boolean_t compressed, boolean_t noauth,
    boolean_t fill, arc_buf_t **ret)
{
	arc_buf_t *buf;
	arc_fill_flags_t flags = ARC_FILL_LOCKED;

	ASSERT(HDR_HAS_L1HDR(hdr));
	ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
	VERIFY(hdr->b_type == ARC_BUFC_DATA ||
	    hdr->b_type == ARC_BUFC_METADATA);
	ASSERT3P(ret, !=, NULL);
	ASSERT3P(*ret, ==, NULL);
	IMPLY(encrypted, compressed);

	buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
	buf->b_hdr = hdr;
	buf->b_data = NULL;
	buf->b_next = hdr->b_l1hdr.b_buf;
	buf->b_flags = 0;

	add_reference(hdr, tag);

	/*
	 * We're about to change the hdr's b_flags. We must either
	 * hold the hash_lock or be undiscoverable.
	 */
	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));

	/*
	 * Only honor requests for compressed bufs if the hdr is actually
	 * compressed. This must be overriden if the buffer is encrypted since
	 * encrypted buffers cannot be decompressed.
	 */
	if (encrypted) {
		buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
		buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED;
		flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED;
	} else if (compressed &&
	    arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
		buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
		flags |= ARC_FILL_COMPRESSED;
	}

	if (noauth) {
		ASSERT0(encrypted);
		flags |= ARC_FILL_NOAUTH;
	}

	/*
	 * If the hdr's data can be shared then we share the data buffer and
	 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
	 * allocate a new buffer to store the buf's data.
	 *
	 * There are two additional restrictions here because we're sharing
	 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
	 * actively involved in an L2ARC write, because if this buf is used by
	 * an arc_write() then the hdr's data buffer will be released when the
	 * write completes, even though the L2ARC write might still be using it.
	 * Second, the hdr's ABD must be linear so that the buf's user doesn't
	 * need to be ABD-aware.
	 */
	boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
	    hdr->b_l1hdr.b_pabd != NULL && abd_is_linear(hdr->b_l1hdr.b_pabd);

	/* Set up b_data and sharing */
	if (can_share) {
		buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
		buf->b_flags |= ARC_BUF_FLAG_SHARED;
		arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
	} else {
		buf->b_data =
		    arc_get_data_buf(hdr, arc_buf_size(buf), buf);
		ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
	}
	VERIFY3P(buf->b_data, !=, NULL);

	hdr->b_l1hdr.b_buf = buf;
	hdr->b_l1hdr.b_bufcnt += 1;
	if (encrypted)
		hdr->b_crypt_hdr.b_ebufcnt += 1;

	/*
	 * If the user wants the data from the hdr, we need to either copy or
	 * decompress the data.
	 */
	if (fill) {
		ASSERT3P(zb, !=, NULL);
		return (arc_buf_fill(buf, spa, zb, flags));
	}

	return (0);
}

static char *arc_onloan_tag = "onloan";

static inline void
arc_loaned_bytes_update(int64_t delta)
{
	atomic_add_64(&arc_loaned_bytes, delta);

	/* assert that it did not wrap around */
	ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
}

/*
 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
 * flight data by arc_tempreserve_space() until they are "returned". Loaned
 * buffers must be returned to the arc before they can be used by the DMU or
 * freed.
 */
arc_buf_t *
arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
{
	arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
	    is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);

	arc_loaned_bytes_update(arc_buf_size(buf));

	return (buf);
}

arc_buf_t *
arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
    enum zio_compress compression_type)
{
	arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
	    psize, lsize, compression_type);

	arc_loaned_bytes_update(arc_buf_size(buf));

	return (buf);
}

arc_buf_t *
arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder,
    const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
    dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
    enum zio_compress compression_type)
{
	arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj,
	    byteorder, salt, iv, mac, ot, psize, lsize, compression_type);

	atomic_add_64(&arc_loaned_bytes, psize);
	return (buf);
}


/*
 * Return a loaned arc buffer to the arc.
 */
void
arc_return_buf(arc_buf_t *buf, void *tag)
{
	arc_buf_hdr_t *hdr = buf->b_hdr;

	ASSERT3P(buf->b_data, !=, NULL);
	ASSERT(HDR_HAS_L1HDR(hdr));
	(void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
	(void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);

	arc_loaned_bytes_update(-arc_buf_size(buf));
}

/* Detach an arc_buf from a dbuf (tag) */
void
arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
{
	arc_buf_hdr_t *hdr = buf->b_hdr;

	ASSERT3P(buf->b_data, !=, NULL);
	ASSERT(HDR_HAS_L1HDR(hdr));
	(void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
	(void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);

	arc_loaned_bytes_update(arc_buf_size(buf));
}

static void
l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
{
	l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);

	df->l2df_abd = abd;
	df->l2df_size = size;
	df->l2df_type = type;
	mutex_enter(&l2arc_free_on_write_mtx);
	list_insert_head(l2arc_free_on_write, df);
	mutex_exit(&l2arc_free_on_write_mtx);
}

static void
arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata)
{
	arc_state_t *state = hdr->b_l1hdr.b_state;
	arc_buf_contents_t type = arc_buf_type(hdr);
	uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);

	/* protected by hash lock, if in the hash table */
	if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
		ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
		ASSERT(state != arc_anon && state != arc_l2c_only);

		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
		    size, hdr);
	}
	(void) zfs_refcount_remove_many(&state->arcs_size, size, hdr);
	if (type == ARC_BUFC_METADATA) {
		arc_space_return(size, ARC_SPACE_META);
	} else {
		ASSERT(type == ARC_BUFC_DATA);
		arc_space_return(size, ARC_SPACE_DATA);
	}

	if (free_rdata) {
		l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type);
	} else {
		l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
	}
}

/*
 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
 * data buffer, we transfer the refcount ownership to the hdr and update
 * the appropriate kstats.
 */
static void
arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
{
	/* LINTED */
	arc_state_t *state = hdr->b_l1hdr.b_state;

	ASSERT(arc_can_share(hdr, buf));
	ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
	ASSERT(!ARC_BUF_ENCRYPTED(buf));
	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));

	/*
	 * Start sharing the data buffer. We transfer the
	 * refcount ownership to the hdr since it always owns
	 * the refcount whenever an arc_buf_t is shared.
	 */
	zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
	    arc_hdr_size(hdr), buf, hdr);
	hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
	abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
	    HDR_ISTYPE_METADATA(hdr));
	arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
	buf->b_flags |= ARC_BUF_FLAG_SHARED;

	/*
	 * Since we've transferred ownership to the hdr we need
	 * to increment its compressed and uncompressed kstats and
	 * decrement the overhead size.
	 */
	ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
	ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
	ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
}

static void
arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
{
	/* LINTED */
	arc_state_t *state = hdr->b_l1hdr.b_state;

	ASSERT(arc_buf_is_shared(buf));
	ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));

	/*
	 * We are no longer sharing this buffer so we need
	 * to transfer its ownership to the rightful owner.
	 */
	zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
	    arc_hdr_size(hdr), hdr, buf);
	arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
	abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
	abd_put(hdr->b_l1hdr.b_pabd);
	hdr->b_l1hdr.b_pabd = NULL;
	buf->b_flags &= ~ARC_BUF_FLAG_SHARED;

	/*
	 * Since the buffer is no longer shared between
	 * the arc buf and the hdr, count it as overhead.
	 */
	ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
	ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
	ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
}

/*
 * Remove an arc_buf_t from the hdr's buf list and return the last
 * arc_buf_t on the list. If no buffers remain on the list then return
 * NULL.
 */
static arc_buf_t *
arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
{
	arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
	arc_buf_t *lastbuf = NULL;

	ASSERT(HDR_HAS_L1HDR(hdr));
	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));

	/*
	 * Remove the buf from the hdr list and locate the last
	 * remaining buffer on the list.
	 */
	while (*bufp != NULL) {
		if (*bufp == buf)
			*bufp = buf->b_next;

		/*
		 * If we've removed a buffer in the middle of
		 * the list then update the lastbuf and update
		 * bufp.
		 */
		if (*bufp != NULL) {
			lastbuf = *bufp;
			bufp = &(*bufp)->b_next;
		}
	}
	buf->b_next = NULL;
	ASSERT3P(lastbuf, !=, buf);
	IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
	IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
	IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));

	return (lastbuf);
}

/*
 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
 * list and free it.
 */
static void
arc_buf_destroy_impl(arc_buf_t *buf)
{
	arc_buf_hdr_t *hdr = buf->b_hdr;

	/*
	 * Free up the data associated with the buf but only if we're not
	 * sharing this with the hdr. If we are sharing it with the hdr, the
	 * hdr is responsible for doing the free.
	 */
	if (buf->b_data != NULL) {
		/*
		 * We're about to change the hdr's b_flags. We must either
		 * hold the hash_lock or be undiscoverable.
		 */
		ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));

		arc_cksum_verify(buf);
		arc_buf_unwatch(buf);

		if (arc_buf_is_shared(buf)) {
			arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
		} else {
			uint64_t size = arc_buf_size(buf);
			arc_free_data_buf(hdr, buf->b_data, size, buf);
			ARCSTAT_INCR(arcstat_overhead_size, -size);
		}
		buf->b_data = NULL;

		ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
		hdr->b_l1hdr.b_bufcnt -= 1;

		if (ARC_BUF_ENCRYPTED(buf)) {
			hdr->b_crypt_hdr.b_ebufcnt -= 1;

			/*
			 * If we have no more encrypted buffers and we've
			 * already gotten a copy of the decrypted data we can
			 * free b_rabd to save some space.
			 */
			if (hdr->b_crypt_hdr.b_ebufcnt == 0 &&
			    HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd != NULL &&
			    !HDR_IO_IN_PROGRESS(hdr)) {
				arc_hdr_free_pabd(hdr, B_TRUE);
			}
		}
	}

	arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);

	if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
		/*
		 * If the current arc_buf_t is sharing its data buffer with the
		 * hdr, then reassign the hdr's b_pabd to share it with the new
		 * buffer at the end of the list. The shared buffer is always
		 * the last one on the hdr's buffer list.
		 *
		 * There is an equivalent case for compressed bufs, but since
		 * they aren't guaranteed to be the last buf in the list and
		 * that is an exceedingly rare case, we just allow that space be
		 * wasted temporarily. We must also be careful not to share
		 * encrypted buffers, since they cannot be shared.
		 */
		if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) {
			/* Only one buf can be shared at once */
			VERIFY(!arc_buf_is_shared(lastbuf));
			/* hdr is uncompressed so can't have compressed buf */
			VERIFY(!ARC_BUF_COMPRESSED(lastbuf));

			ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
			arc_hdr_free_pabd(hdr, B_FALSE);

			/*
			 * We must setup a new shared block between the
			 * last buffer and the hdr. The data would have
			 * been allocated by the arc buf so we need to transfer
			 * ownership to the hdr since it's now being shared.
			 */
			arc_share_buf(hdr, lastbuf);
		}
	} else if (HDR_SHARED_DATA(hdr)) {
		/*
		 * Uncompressed shared buffers are always at the end
		 * of the list. Compressed buffers don't have the
		 * same requirements. This makes it hard to
		 * simply assert that the lastbuf is shared so
		 * we rely on the hdr's compression flags to determine
		 * if we have a compressed, shared buffer.
		 */
		ASSERT3P(lastbuf, !=, NULL);
		ASSERT(arc_buf_is_shared(lastbuf) ||
		    arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
	}

	/*
	 * Free the checksum if we're removing the last uncompressed buf from
	 * this hdr.
	 */
	if (!arc_hdr_has_uncompressed_buf(hdr)) {
		arc_cksum_free(hdr);
	}

	/* clean up the buf */
	buf->b_hdr = NULL;
	kmem_cache_free(buf_cache, buf);
}

static void
arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr, boolean_t alloc_rdata)
{
	uint64_t size;

	ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
	ASSERT(HDR_HAS_L1HDR(hdr));
	ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata);
	IMPLY(alloc_rdata, HDR_PROTECTED(hdr));

	if (alloc_rdata) {
		size = HDR_GET_PSIZE(hdr);
		ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL);
		hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr);
		ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL);
	} else {
		size = arc_hdr_size(hdr);
		ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
		hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr);
		ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
	}

	ARCSTAT_INCR(arcstat_compressed_size, size);
	ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
}

static void
arc_hdr_free_pabd(arc_buf_hdr_t *hdr, boolean_t free_rdata)
{
	uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);

	ASSERT(HDR_HAS_L1HDR(hdr));
	ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
	IMPLY(free_rdata, HDR_HAS_RABD(hdr));


	/*
	 * If the hdr is currently being written to the l2arc then
	 * we defer freeing the data by adding it to the l2arc_free_on_write
	 * list. The l2arc will free the data once it's finished
	 * writing it to the l2arc device.
	 */
	if (HDR_L2_WRITING(hdr)) {
		arc_hdr_free_on_write(hdr, free_rdata);
		ARCSTAT_BUMP(arcstat_l2_free_on_write);
	} else if (free_rdata) {
		arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr);
	} else {
		arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
		    size, hdr);
	}

	if (free_rdata) {
		hdr->b_crypt_hdr.b_rabd = NULL;
	} else {
		hdr->b_l1hdr.b_pabd = NULL;
	}

	if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr))
		hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;

	ARCSTAT_INCR(arcstat_compressed_size, -size);
	ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
}

static arc_buf_hdr_t *
arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
    boolean_t protected, enum zio_compress compression_type,
    arc_buf_contents_t type, boolean_t alloc_rdata)
{
	arc_buf_hdr_t *hdr;

	VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
	if (protected) {
		hdr = kmem_cache_alloc(hdr_full_crypt_cache, KM_PUSHPAGE);
	} else {
		hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
	}
	ASSERT(HDR_EMPTY(hdr));
	ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
	ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
	HDR_SET_PSIZE(hdr, psize);
	HDR_SET_LSIZE(hdr, lsize);
	hdr->b_spa = spa;
	hdr->b_type = type;
	hdr->b_flags = 0;
	arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
	arc_hdr_set_compress(hdr, compression_type);
	if (protected)
		arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);

	hdr->b_l1hdr.b_state = arc_anon;
	hdr->b_l1hdr.b_arc_access = 0;
	hdr->b_l1hdr.b_bufcnt = 0;
	hdr->b_l1hdr.b_buf = NULL;

	/*
	 * Allocate the hdr's buffer. This will contain either
	 * the compressed or uncompressed data depending on the block
	 * it references and compressed arc enablement.
	 */
	arc_hdr_alloc_pabd(hdr, alloc_rdata);
	ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));

	return (hdr);
}

/*
 * Transition between the two allocation states for the arc_buf_hdr struct.
 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
 * version is used when a cache buffer is only in the L2ARC in order to reduce
 * memory usage.
 */
static arc_buf_hdr_t *
arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
{
	ASSERT(HDR_HAS_L2HDR(hdr));

	arc_buf_hdr_t *nhdr;
	l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;

	ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
	    (old == hdr_l2only_cache && new == hdr_full_cache));

	/*
	 * if the caller wanted a new full header and the header is to be
	 * encrypted we will actually allocate the header from the full crypt
	 * cache instead. The same applies to freeing from the old cache.
	 */
	if (HDR_PROTECTED(hdr) && new == hdr_full_cache)
		new = hdr_full_crypt_cache;
	if (HDR_PROTECTED(hdr) && old == hdr_full_cache)
		old = hdr_full_crypt_cache;

	nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);

	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
	buf_hash_remove(hdr);

	bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);

	if (new == hdr_full_cache || new == hdr_full_crypt_cache) {
		arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
		/*
		 * arc_access and arc_change_state need to be aware that a
		 * header has just come out of L2ARC, so we set its state to
		 * l2c_only even though it's about to change.
		 */
		nhdr->b_l1hdr.b_state = arc_l2c_only;

		/* Verify previous threads set to NULL before freeing */
		ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
		ASSERT(!HDR_HAS_RABD(hdr));
	} else {
		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
		ASSERT0(hdr->b_l1hdr.b_bufcnt);
		ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);

		/*
		 * If we've reached here, We must have been called from
		 * arc_evict_hdr(), as such we should have already been
		 * removed from any ghost list we were previously on
		 * (which protects us from racing with arc_evict_state),
		 * thus no locking is needed during this check.
		 */
		ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));

		/*
		 * A buffer must not be moved into the arc_l2c_only
		 * state if it's not finished being written out to the
		 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
		 * might try to be accessed, even though it was removed.
		 */
		VERIFY(!HDR_L2_WRITING(hdr));
		VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
		ASSERT(!HDR_HAS_RABD(hdr));

#ifdef ZFS_DEBUG
		if (hdr->b_l1hdr.b_thawed != NULL) {
			kmem_free(hdr->b_l1hdr.b_thawed, 1);
			hdr->b_l1hdr.b_thawed = NULL;
		}
#endif

		arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
	}
	/*
	 * The header has been reallocated so we need to re-insert it into any
	 * lists it was on.
	 */
	(void) buf_hash_insert(nhdr, NULL);

	ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));

	mutex_enter(&dev->l2ad_mtx);

	/*
	 * We must place the realloc'ed header back into the list at
	 * the same spot. Otherwise, if it's placed earlier in the list,
	 * l2arc_write_buffers() could find it during the function's
	 * write phase, and try to write it out to the l2arc.
	 */
	list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
	list_remove(&dev->l2ad_buflist, hdr);

	mutex_exit(&dev->l2ad_mtx);

	/*
	 * Since we're using the pointer address as the tag when
	 * incrementing and decrementing the l2ad_alloc refcount, we
	 * must remove the old pointer (that we're about to destroy) and
	 * add the new pointer to the refcount. Otherwise we'd remove
	 * the wrong pointer address when calling arc_hdr_destroy() later.
	 */

	(void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
	    hdr);
	(void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr),
	    nhdr);

	buf_discard_identity(hdr);
	kmem_cache_free(old, hdr);

	return (nhdr);
}

/*
 * This function allows an L1 header to be reallocated as a crypt
 * header and vice versa. If we are going to a crypt header, the
 * new fields will be zeroed out.
 */
static arc_buf_hdr_t *
arc_hdr_realloc_crypt(arc_buf_hdr_t *hdr, boolean_t need_crypt)
{
	arc_buf_hdr_t *nhdr;
	arc_buf_t *buf;
	kmem_cache_t *ncache, *ocache;

	ASSERT(HDR_HAS_L1HDR(hdr));
	ASSERT3U(!!HDR_PROTECTED(hdr), !=, need_crypt);
	ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
	ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
	ASSERT(!list_link_active(&hdr->b_l2hdr.b_l2node));
	ASSERT3P(hdr->b_hash_next, ==, NULL);

	if (need_crypt) {
		ncache = hdr_full_crypt_cache;
		ocache = hdr_full_cache;
	} else {
		ncache = hdr_full_cache;
		ocache = hdr_full_crypt_cache;
	}

	nhdr = kmem_cache_alloc(ncache, KM_PUSHPAGE);

	/*
	 * Copy all members that aren't locks or condvars to the new header.
	 * No lists are pointing to us (as we asserted above), so we don't
	 * need to worry about the list nodes.
	 */
	nhdr->b_dva = hdr->b_dva;
	nhdr->b_birth = hdr->b_birth;
	nhdr->b_type = hdr->b_type;
	nhdr->b_flags = hdr->b_flags;
	nhdr->b_psize = hdr->b_psize;
	nhdr->b_lsize = hdr->b_lsize;
	nhdr->b_spa = hdr->b_spa;
	nhdr->b_l2hdr.b_dev = hdr->b_l2hdr.b_dev;
	nhdr->b_l2hdr.b_daddr = hdr->b_l2hdr.b_daddr;
	nhdr->b_l1hdr.b_freeze_cksum = hdr->b_l1hdr.b_freeze_cksum;
	nhdr->b_l1hdr.b_bufcnt = hdr->b_l1hdr.b_bufcnt;
	nhdr->b_l1hdr.b_byteswap = hdr->b_l1hdr.b_byteswap;
	nhdr->b_l1hdr.b_state = hdr->b_l1hdr.b_state;
	nhdr->b_l1hdr.b_arc_access = hdr->b_l1hdr.b_arc_access;
	nhdr->b_l1hdr.b_acb = hdr->b_l1hdr.b_acb;
	nhdr->b_l1hdr.b_pabd = hdr->b_l1hdr.b_pabd;
#ifdef ZFS_DEBUG
	if (hdr->b_l1hdr.b_thawed != NULL) {
		nhdr->b_l1hdr.b_thawed = hdr->b_l1hdr.b_thawed;
		hdr->b_l1hdr.b_thawed = NULL;
	}
#endif

	/*
	 * This refcount_add() exists only to ensure that the individual
	 * arc buffers always point to a header that is referenced, avoiding
	 * a small race condition that could trigger ASSERTs.
	 */
	(void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, FTAG);
	nhdr->b_l1hdr.b_buf = hdr->b_l1hdr.b_buf;
	for (buf = nhdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) {
		mutex_enter(&buf->b_evict_lock);
		buf->b_hdr = nhdr;
		mutex_exit(&buf->b_evict_lock);
	}
	zfs_refcount_transfer(&nhdr->b_l1hdr.b_refcnt, &hdr->b_l1hdr.b_refcnt);
	(void) zfs_refcount_remove(&nhdr->b_l1hdr.b_refcnt, FTAG);
	ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));

	if (need_crypt) {
		arc_hdr_set_flags(nhdr, ARC_FLAG_PROTECTED);
	} else {
		arc_hdr_clear_flags(nhdr, ARC_FLAG_PROTECTED);
	}

	/* unset all members of the original hdr */
	bzero(&hdr->b_dva, sizeof (dva_t));
	hdr->b_birth = 0;
	hdr->b_type = ARC_BUFC_INVALID;
	hdr->b_flags = 0;
	hdr->b_psize = 0;
	hdr->b_lsize = 0;
	hdr->b_spa = 0;
	hdr->b_l2hdr.b_dev = NULL;
	hdr->b_l2hdr.b_daddr = 0;
	hdr->b_l1hdr.b_freeze_cksum = NULL;
	hdr->b_l1hdr.b_buf = NULL;
	hdr->b_l1hdr.b_bufcnt = 0;
	hdr->b_l1hdr.b_byteswap = 0;
	hdr->b_l1hdr.b_state = NULL;
	hdr->b_l1hdr.b_arc_access = 0;
	hdr->b_l1hdr.b_acb = NULL;
	hdr->b_l1hdr.b_pabd = NULL;

	if (ocache == hdr_full_crypt_cache) {
		ASSERT(!HDR_HAS_RABD(hdr));
		hdr->b_crypt_hdr.b_ot = DMU_OT_NONE;
		hdr->b_crypt_hdr.b_ebufcnt = 0;
		hdr->b_crypt_hdr.b_dsobj = 0;
		bzero(hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
		bzero(hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
		bzero(hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
	}

	buf_discard_identity(hdr);
	kmem_cache_free(ocache, hdr);

	return (nhdr);
}

/*
 * This function is used by the send / receive code to convert a newly
 * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
 * is also used to allow the root objset block to be uupdated without altering
 * its embedded MACs. Both block types will always be uncompressed so we do not
 * have to worry about compression type or psize.
 */
void
arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder,
    dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv,
    const uint8_t *mac)
{
	arc_buf_hdr_t *hdr = buf->b_hdr;

	ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET);
	ASSERT(HDR_HAS_L1HDR(hdr));
	ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);

	buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED);
	if (!HDR_PROTECTED(hdr))
		hdr = arc_hdr_realloc_crypt(hdr, B_TRUE);
	hdr->b_crypt_hdr.b_dsobj = dsobj;
	hdr->b_crypt_hdr.b_ot = ot;
	hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
	    DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
	if (!arc_hdr_has_uncompressed_buf(hdr))
		arc_cksum_free(hdr);

	if (salt != NULL)
		bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
	if (iv != NULL)
		bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
	if (mac != NULL)
		bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
}

/*
 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
 * The buf is returned thawed since we expect the consumer to modify it.
 */
arc_buf_t *
arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
{
	arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
	    B_FALSE, ZIO_COMPRESS_OFF, type, B_FALSE);
	ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));

	arc_buf_t *buf = NULL;
	VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE,
	    B_FALSE, B_FALSE, &buf));
	arc_buf_thaw(buf);

	return (buf);
}

/*
 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
 * for bufs containing metadata.
 */
arc_buf_t *
arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
    enum zio_compress compression_type)
{
	ASSERT3U(lsize, >, 0);
	ASSERT3U(lsize, >=, psize);
	ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF);
	ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);

	arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
	    B_FALSE, compression_type, ARC_BUFC_DATA, B_FALSE);
	ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));

	arc_buf_t *buf = NULL;
	VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE,
	    B_TRUE, B_FALSE, B_FALSE, &buf));
	arc_buf_thaw(buf);
	ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);

	if (!arc_buf_is_shared(buf)) {
		/*
		 * To ensure that the hdr has the correct data in it if we call
		 * arc_untransform() on this buf before it's been written to
		 * disk, it's easiest if we just set up sharing between the
		 * buf and the hdr.
		 */
		ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
		arc_hdr_free_pabd(hdr, B_FALSE);
		arc_share_buf(hdr, buf);
	}

	return (buf);
}

arc_buf_t *
arc_alloc_raw_buf(spa_t *spa, void *tag, uint64_t dsobj, boolean_t byteorder,
    const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
    dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
    enum zio_compress compression_type)
{
	arc_buf_hdr_t *hdr;
	arc_buf_t *buf;
	arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ?
	    ARC_BUFC_METADATA : ARC_BUFC_DATA;

	ASSERT3U(lsize, >, 0);
	ASSERT3U(lsize, >=, psize);
	ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF);
	ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);

	hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE,
	    compression_type, type, B_TRUE);
	ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));

	hdr->b_crypt_hdr.b_dsobj = dsobj;
	hdr->b_crypt_hdr.b_ot = ot;
	hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
	    DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
	bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
	bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
	bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);

	/*
	 * This buffer will be considered encrypted even if the ot is not an
	 * encrypted type. It will become authenticated instead in
	 * arc_write_ready().
	 */
	buf = NULL;
	VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE,
	    B_FALSE, B_FALSE, &buf));
	arc_buf_thaw(buf);
	ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);

	return (buf);
}

static void
arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
{
	l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
	l2arc_dev_t *dev = l2hdr->b_dev;
	uint64_t psize = HDR_GET_PSIZE(hdr);
	uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);

	ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
	ASSERT(HDR_HAS_L2HDR(hdr));

	list_remove(&dev->l2ad_buflist, hdr);

	ARCSTAT_INCR(arcstat_l2_psize, -psize);
	ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));

	vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);

	(void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
	    hdr);
	arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
}

static void
arc_hdr_destroy(arc_buf_hdr_t *hdr)
{
	if (HDR_HAS_L1HDR(hdr)) {
		ASSERT(hdr->b_l1hdr.b_buf == NULL ||
		    hdr->b_l1hdr.b_bufcnt > 0);
		ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
		ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
	}
	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
	ASSERT(!HDR_IN_HASH_TABLE(hdr));

	if (!HDR_EMPTY(hdr))
		buf_discard_identity(hdr);

	if (HDR_HAS_L2HDR(hdr)) {
		l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
		boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);

		if (!buflist_held)
			mutex_enter(&dev->l2ad_mtx);

		/*
		 * Even though we checked this conditional above, we
		 * need to check this again now that we have the
		 * l2ad_mtx. This is because we could be racing with
		 * another thread calling l2arc_evict() which might have
		 * destroyed this header's L2 portion as we were waiting
		 * to acquire the l2ad_mtx. If that happens, we don't
		 * want to re-destroy the header's L2 portion.
		 */
		if (HDR_HAS_L2HDR(hdr))
			arc_hdr_l2hdr_destroy(hdr);

		if (!buflist_held)
			mutex_exit(&dev->l2ad_mtx);
	}

	if (HDR_HAS_L1HDR(hdr)) {
		arc_cksum_free(hdr);

		while (hdr->b_l1hdr.b_buf != NULL)
			arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);

#ifdef ZFS_DEBUG
		if (hdr->b_l1hdr.b_thawed != NULL) {
			kmem_free(hdr->b_l1hdr.b_thawed, 1);
			hdr->b_l1hdr.b_thawed = NULL;
		}
#endif

		if (hdr->b_l1hdr.b_pabd != NULL) {
			arc_hdr_free_pabd(hdr, B_FALSE);
		}

		if (HDR_HAS_RABD(hdr))
			arc_hdr_free_pabd(hdr, B_TRUE);
	}

	ASSERT3P(hdr->b_hash_next, ==, NULL);
	if (HDR_HAS_L1HDR(hdr)) {
		ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
		ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);

		if (!HDR_PROTECTED(hdr)) {
			kmem_cache_free(hdr_full_cache, hdr);
		} else {
			kmem_cache_free(hdr_full_crypt_cache, hdr);
		}
	} else {
		kmem_cache_free(hdr_l2only_cache, hdr);
	}
}

void
arc_buf_destroy(arc_buf_t *buf, void* tag)
{
	arc_buf_hdr_t *hdr = buf->b_hdr;
	kmutex_t *hash_lock = HDR_LOCK(hdr);

	if (hdr->b_l1hdr.b_state == arc_anon) {
		ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
		VERIFY0(remove_reference(hdr, NULL, tag));
		arc_hdr_destroy(hdr);
		return;
	}

	mutex_enter(hash_lock);
	ASSERT3P(hdr, ==, buf->b_hdr);
	ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
	ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
	ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
	ASSERT3P(buf->b_data, !=, NULL);

	(void) remove_reference(hdr, hash_lock, tag);
	arc_buf_destroy_impl(buf);
	mutex_exit(hash_lock);
}

/*
 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
 * state of the header is dependent on its state prior to entering this
 * function. The following transitions are possible:
 *
 *    - arc_mru -> arc_mru_ghost
 *    - arc_mfu -> arc_mfu_ghost
 *    - arc_mru_ghost -> arc_l2c_only
 *    - arc_mru_ghost -> deleted
 *    - arc_mfu_ghost -> arc_l2c_only
 *    - arc_mfu_ghost -> deleted
 */
static int64_t
arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
{
	arc_state_t *evicted_state, *state;
	int64_t bytes_evicted = 0;
	int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
	    zfs_arc_min_prescient_prefetch_ms : zfs_arc_min_prefetch_ms;

	ASSERT(MUTEX_HELD(hash_lock));
	ASSERT(HDR_HAS_L1HDR(hdr));

	state = hdr->b_l1hdr.b_state;
	if (GHOST_STATE(state)) {
		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);

		/*
		 * l2arc_write_buffers() relies on a header's L1 portion
		 * (i.e. its b_pabd field) during its write phase.
		 * Thus, we cannot push a header onto the arc_l2c_only
		 * state (removing its L1 piece) until the header is
		 * done being written to the l2arc.
		 */
		if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
			ARCSTAT_BUMP(arcstat_evict_l2_skip);
			return (bytes_evicted);
		}

		ARCSTAT_BUMP(arcstat_deleted);
		bytes_evicted += HDR_GET_LSIZE(hdr);

		DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);

		if (HDR_HAS_L2HDR(hdr)) {
			ASSERT(hdr->b_l1hdr.b_pabd == NULL);
			ASSERT(!HDR_HAS_RABD(hdr));
			/*
			 * This buffer is cached on the 2nd Level ARC;
			 * don't destroy the header.
			 */
			arc_change_state(arc_l2c_only, hdr, hash_lock);
			/*
			 * dropping from L1+L2 cached to L2-only,
			 * realloc to remove the L1 header.
			 */
			hdr = arc_hdr_realloc(hdr, hdr_full_cache,
			    hdr_l2only_cache);
		} else {
			arc_change_state(arc_anon, hdr, hash_lock);
			arc_hdr_destroy(hdr);
		}
		return (bytes_evicted);
	}

	ASSERT(state == arc_mru || state == arc_mfu);
	evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;

	/* prefetch buffers have a minimum lifespan */
	if (HDR_IO_IN_PROGRESS(hdr) ||
	    ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
	    ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < min_lifetime * hz)) {
		ARCSTAT_BUMP(arcstat_evict_skip);
		return (bytes_evicted);
	}

	ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
	while (hdr->b_l1hdr.b_buf) {
		arc_buf_t *buf = hdr->b_l1hdr.b_buf;
		if (!mutex_tryenter(&buf->b_evict_lock)) {
			ARCSTAT_BUMP(arcstat_mutex_miss);
			break;
		}
		if (buf->b_data != NULL)
			bytes_evicted += HDR_GET_LSIZE(hdr);
		mutex_exit(&buf->b_evict_lock);
		arc_buf_destroy_impl(buf);
	}

	if (HDR_HAS_L2HDR(hdr)) {
		ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
	} else {
		if (l2arc_write_eligible(hdr->b_spa, hdr)) {
			ARCSTAT_INCR(arcstat_evict_l2_eligible,
			    HDR_GET_LSIZE(hdr));
		} else {
			ARCSTAT_INCR(arcstat_evict_l2_ineligible,
			    HDR_GET_LSIZE(hdr));
		}
	}

	if (hdr->b_l1hdr.b_bufcnt == 0) {
		arc_cksum_free(hdr);

		bytes_evicted += arc_hdr_size(hdr);

		/*
		 * If this hdr is being evicted and has a compressed
		 * buffer then we discard it here before we change states.
		 * This ensures that the accounting is updated correctly
		 * in arc_free_data_impl().
		 */
		if (hdr->b_l1hdr.b_pabd != NULL)
			arc_hdr_free_pabd(hdr, B_FALSE);

		if (HDR_HAS_RABD(hdr))
			arc_hdr_free_pabd(hdr, B_TRUE);

		arc_change_state(evicted_state, hdr, hash_lock);
		ASSERT(HDR_IN_HASH_TABLE(hdr));
		arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
		DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
	}

	return (bytes_evicted);
}

static uint64_t
arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
    uint64_t spa, int64_t bytes)
{
	multilist_sublist_t *mls;
	uint64_t bytes_evicted = 0;
	arc_buf_hdr_t *hdr;
	kmutex_t *hash_lock;
	int evict_count = 0;

	ASSERT3P(marker, !=, NULL);
	IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);

	mls = multilist_sublist_lock(ml, idx);

	for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
	    hdr = multilist_sublist_prev(mls, marker)) {
		if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
		    (evict_count >= zfs_arc_evict_batch_limit))
			break;

		/*
		 * To keep our iteration location, move the marker
		 * forward. Since we're not holding hdr's hash lock, we
		 * must be very careful and not remove 'hdr' from the
		 * sublist. Otherwise, other consumers might mistake the
		 * 'hdr' as not being on a sublist when they call the
		 * multilist_link_active() function (they all rely on
		 * the hash lock protecting concurrent insertions and
		 * removals). multilist_sublist_move_forward() was
		 * specifically implemented to ensure this is the case
		 * (only 'marker' will be removed and re-inserted).
		 */
		multilist_sublist_move_forward(mls, marker);

		/*
		 * The only case where the b_spa field should ever be
		 * zero, is the marker headers inserted by
		 * arc_evict_state(). It's possible for multiple threads
		 * to be calling arc_evict_state() concurrently (e.g.
		 * dsl_pool_close() and zio_inject_fault()), so we must
		 * skip any markers we see from these other threads.
		 */
		if (hdr->b_spa == 0)
			continue;

		/* we're only interested in evicting buffers of a certain spa */
		if (spa != 0 && hdr->b_spa != spa) {
			ARCSTAT_BUMP(arcstat_evict_skip);
			continue;
		}

		hash_lock = HDR_LOCK(hdr);

		/*
		 * We aren't calling this function from any code path
		 * that would already be holding a hash lock, so we're
		 * asserting on this assumption to be defensive in case
		 * this ever changes. Without this check, it would be
		 * possible to incorrectly increment arcstat_mutex_miss
		 * below (e.g. if the code changed such that we called
		 * this function with a hash lock held).
		 */
		ASSERT(!MUTEX_HELD(hash_lock));

		if (mutex_tryenter(hash_lock)) {
			uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
			mutex_exit(hash_lock);

			bytes_evicted += evicted;

			/*
			 * If evicted is zero, arc_evict_hdr() must have
			 * decided to skip this header, don't increment
			 * evict_count in this case.
			 */
			if (evicted != 0)
				evict_count++;

			/*
			 * If arc_size isn't overflowing, signal any
			 * threads that might happen to be waiting.
			 *
			 * For each header evicted, we wake up a single
			 * thread. If we used cv_broadcast, we could
			 * wake up "too many" threads causing arc_size
			 * to significantly overflow arc_c; since
			 * arc_get_data_impl() doesn't check for overflow
			 * when it's woken up (it doesn't because it's
			 * possible for the ARC to be overflowing while
			 * full of un-evictable buffers, and the
			 * function should proceed in this case).
			 *
			 * If threads are left sleeping, due to not
			 * using cv_broadcast here, they will be woken
			 * up via cv_broadcast in arc_adjust_cb() just
			 * before arc_adjust_zthr sleeps.
			 */
			mutex_enter(&arc_adjust_lock);
			if (!arc_is_overflowing())
				cv_signal(&arc_adjust_waiters_cv);
			mutex_exit(&arc_adjust_lock);
		} else {
			ARCSTAT_BUMP(arcstat_mutex_miss);
		}
	}

	multilist_sublist_unlock(mls);

	return (bytes_evicted);
}

/*
 * Evict buffers from the given arc state, until we've removed the
 * specified number of bytes. Move the removed buffers to the
 * appropriate evict state.
 *
 * This function makes a "best effort". It skips over any buffers
 * it can't get a hash_lock on, and so, may not catch all candidates.
 * It may also return without evicting as much space as requested.
 *
 * If bytes is specified using the special value ARC_EVICT_ALL, this
 * will evict all available (i.e. unlocked and evictable) buffers from
 * the given arc state; which is used by arc_flush().
 */
static uint64_t
arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
    arc_buf_contents_t type)
{
	uint64_t total_evicted = 0;
	multilist_t *ml = state->arcs_list[type];
	int num_sublists;
	arc_buf_hdr_t **markers;

	IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);

	num_sublists = multilist_get_num_sublists(ml);

	/*
	 * If we've tried to evict from each sublist, made some
	 * progress, but still have not hit the target number of bytes
	 * to evict, we want to keep trying. The markers allow us to
	 * pick up where we left off for each individual sublist, rather
	 * than starting from the tail each time.
	 */
	markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
	for (int i = 0; i < num_sublists; i++) {
		markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);

		/*
		 * A b_spa of 0 is used to indicate that this header is
		 * a marker. This fact is used in arc_adjust_type() and
		 * arc_evict_state_impl().
		 */
		markers[i]->b_spa = 0;

		multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
		multilist_sublist_insert_tail(mls, markers[i]);
		multilist_sublist_unlock(mls);
	}

	/*
	 * While we haven't hit our target number of bytes to evict, or
	 * we're evicting all available buffers.
	 */
	while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
		/*
		 * Start eviction using a randomly selected sublist,
		 * this is to try and evenly balance eviction across all
		 * sublists. Always starting at the same sublist
		 * (e.g. index 0) would cause evictions to favor certain
		 * sublists over others.
		 */
		int sublist_idx = multilist_get_random_index(ml);
		uint64_t scan_evicted = 0;

		for (int i = 0; i < num_sublists; i++) {
			uint64_t bytes_remaining;
			uint64_t bytes_evicted;

			if (bytes == ARC_EVICT_ALL)
				bytes_remaining = ARC_EVICT_ALL;
			else if (total_evicted < bytes)
				bytes_remaining = bytes - total_evicted;
			else
				break;

			bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
			    markers[sublist_idx], spa, bytes_remaining);

			scan_evicted += bytes_evicted;
			total_evicted += bytes_evicted;

			/* we've reached the end, wrap to the beginning */
			if (++sublist_idx >= num_sublists)
				sublist_idx = 0;
		}

		/*
		 * If we didn't evict anything during this scan, we have
		 * no reason to believe we'll evict more during another
		 * scan, so break the loop.
		 */
		if (scan_evicted == 0) {
			/* This isn't possible, let's make that obvious */
			ASSERT3S(bytes, !=, 0);

			/*
			 * When bytes is ARC_EVICT_ALL, the only way to
			 * break the loop is when scan_evicted is zero.
			 * In that case, we actually have evicted enough,
			 * so we don't want to increment the kstat.
			 */
			if (bytes != ARC_EVICT_ALL) {
				ASSERT3S(total_evicted, <, bytes);
				ARCSTAT_BUMP(arcstat_evict_not_enough);
			}

			break;
		}
	}

	for (int i = 0; i < num_sublists; i++) {
		multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
		multilist_sublist_remove(mls, markers[i]);
		multilist_sublist_unlock(mls);

		kmem_cache_free(hdr_full_cache, markers[i]);
	}
	kmem_free(markers, sizeof (*markers) * num_sublists);

	return (total_evicted);
}

/*
 * Flush all "evictable" data of the given type from the arc state
 * specified. This will not evict any "active" buffers (i.e. referenced).
 *
 * When 'retry' is set to B_FALSE, the function will make a single pass
 * over the state and evict any buffers that it can. Since it doesn't
 * continually retry the eviction, it might end up leaving some buffers
 * in the ARC due to lock misses.
 *
 * When 'retry' is set to B_TRUE, the function will continually retry the
 * eviction until *all* evictable buffers have been removed from the
 * state. As a result, if concurrent insertions into the state are
 * allowed (e.g. if the ARC isn't shutting down), this function might
 * wind up in an infinite loop, continually trying to evict buffers.
 */
static uint64_t
arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
    boolean_t retry)
{
	uint64_t evicted = 0;

	while (zfs_refcount_count(&state->arcs_esize[type]) != 0) {
		evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);

		if (!retry)
			break;
	}

	return (evicted);
}

/*
 * Evict the specified number of bytes from the state specified,
 * restricting eviction to the spa and type given. This function
 * prevents us from trying to evict more from a state's list than
 * is "evictable", and to skip evicting altogether when passed a
 * negative value for "bytes". In contrast, arc_evict_state() will
 * evict everything it can, when passed a negative value for "bytes".
 */
static uint64_t
arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
    arc_buf_contents_t type)
{
	int64_t delta;

	if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) {
		delta = MIN(zfs_refcount_count(&state->arcs_esize[type]),
		    bytes);
		return (arc_evict_state(state, spa, delta, type));
	}

	return (0);
}

/*
 * Evict metadata buffers from the cache, such that arc_meta_used is
 * capped by the arc_meta_limit tunable.
 */
static uint64_t
arc_adjust_meta(uint64_t meta_used)
{
	uint64_t total_evicted = 0;
	int64_t target;

	/*
	 * If we're over the meta limit, we want to evict enough
	 * metadata to get back under the meta limit. We don't want to
	 * evict so much that we drop the MRU below arc_p, though. If
	 * we're over the meta limit more than we're over arc_p, we
	 * evict some from the MRU here, and some from the MFU below.
	 */
	target = MIN((int64_t)(meta_used - arc_meta_limit),
	    (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
	    zfs_refcount_count(&arc_mru->arcs_size) - arc_p));

	total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);

	/*
	 * Similar to the above, we want to evict enough bytes to get us
	 * below the meta limit, but not so much as to drop us below the
	 * space allotted to the MFU (which is defined as arc_c - arc_p).
	 */
	target = MIN((int64_t)(meta_used - arc_meta_limit),
	    (int64_t)(zfs_refcount_count(&arc_mfu->arcs_size) -
	    (arc_c - arc_p)));

	total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);

	return (total_evicted);
}

/*
 * Return the type of the oldest buffer in the given arc state
 *
 * This function will select a random sublist of type ARC_BUFC_DATA and
 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
 * is compared, and the type which contains the "older" buffer will be
 * returned.
 */
static arc_buf_contents_t
arc_adjust_type(arc_state_t *state)
{
	multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
	multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
	int data_idx = multilist_get_random_index(data_ml);
	int meta_idx = multilist_get_random_index(meta_ml);
	multilist_sublist_t *data_mls;
	multilist_sublist_t *meta_mls;
	arc_buf_contents_t type;
	arc_buf_hdr_t *data_hdr;
	arc_buf_hdr_t *meta_hdr;

	/*
	 * We keep the sublist lock until we're finished, to prevent
	 * the headers from being destroyed via arc_evict_state().
	 */
	data_mls = multilist_sublist_lock(data_ml, data_idx);
	meta_mls = multilist_sublist_lock(meta_ml, meta_idx);

	/*
	 * These two loops are to ensure we skip any markers that
	 * might be at the tail of the lists due to arc_evict_state().
	 */

	for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
	    data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
		if (data_hdr->b_spa != 0)
			break;
	}

	for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
	    meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
		if (meta_hdr->b_spa != 0)
			break;
	}

	if (data_hdr == NULL && meta_hdr == NULL) {
		type = ARC_BUFC_DATA;
	} else if (data_hdr == NULL) {
		ASSERT3P(meta_hdr, !=, NULL);
		type = ARC_BUFC_METADATA;
	} else if (meta_hdr == NULL) {
		ASSERT3P(data_hdr, !=, NULL);
		type = ARC_BUFC_DATA;
	} else {
		ASSERT3P(data_hdr, !=, NULL);
		ASSERT3P(meta_hdr, !=, NULL);

		/* The headers can't be on the sublist without an L1 header */
		ASSERT(HDR_HAS_L1HDR(data_hdr));
		ASSERT(HDR_HAS_L1HDR(meta_hdr));

		if (data_hdr->b_l1hdr.b_arc_access <
		    meta_hdr->b_l1hdr.b_arc_access) {
			type = ARC_BUFC_DATA;
		} else {
			type = ARC_BUFC_METADATA;
		}
	}

	multilist_sublist_unlock(meta_mls);
	multilist_sublist_unlock(data_mls);

	return (type);
}

/*
 * Evict buffers from the cache, such that arc_size is capped by arc_c.
 */
static uint64_t
arc_adjust(void)
{
	uint64_t total_evicted = 0;
	uint64_t bytes;
	int64_t target;
	uint64_t asize = aggsum_value(&arc_size);
	uint64_t ameta = aggsum_value(&arc_meta_used);

	/*
	 * If we're over arc_meta_limit, we want to correct that before
	 * potentially evicting data buffers below.
	 */
	total_evicted += arc_adjust_meta(ameta);

	/*
	 * Adjust MRU size
	 *
	 * If we're over the target cache size, we want to evict enough
	 * from the list to get back to our target size. We don't want
	 * to evict too much from the MRU, such that it drops below
	 * arc_p. So, if we're over our target cache size more than
	 * the MRU is over arc_p, we'll evict enough to get back to
	 * arc_p here, and then evict more from the MFU below.
	 */
	target = MIN((int64_t)(asize - arc_c),
	    (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
	    zfs_refcount_count(&arc_mru->arcs_size) + ameta - arc_p));

	/*
	 * If we're below arc_meta_min, always prefer to evict data.
	 * Otherwise, try to satisfy the requested number of bytes to
	 * evict from the type which contains older buffers; in an
	 * effort to keep newer buffers in the cache regardless of their
	 * type. If we cannot satisfy the number of bytes from this
	 * type, spill over into the next type.
	 */
	if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
	    ameta > arc_meta_min) {
		bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
		total_evicted += bytes;

		/*
		 * If we couldn't evict our target number of bytes from
		 * metadata, we try to get the rest from data.
		 */
		target -= bytes;

		total_evicted +=
		    arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
	} else {
		bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
		total_evicted += bytes;

		/*
		 * If we couldn't evict our target number of bytes from
		 * data, we try to get the rest from metadata.
		 */
		target -= bytes;

		total_evicted +=
		    arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
	}

	/*
	 * Adjust MFU size
	 *
	 * Now that we've tried to evict enough from the MRU to get its
	 * size back to arc_p, if we're still above the target cache
	 * size, we evict the rest from the MFU.
	 */
	target = asize - arc_c;

	if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
	    ameta > arc_meta_min) {
		bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
		total_evicted += bytes;

		/*
		 * If we couldn't evict our target number of bytes from
		 * metadata, we try to get the rest from data.
		 */
		target -= bytes;

		total_evicted +=
		    arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
	} else {
		bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
		total_evicted += bytes;

		/*
		 * If we couldn't evict our target number of bytes from
		 * data, we try to get the rest from data.
		 */
		target -= bytes;

		total_evicted +=
		    arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
	}

	/*
	 * Adjust ghost lists
	 *
	 * In addition to the above, the ARC also defines target values
	 * for the ghost lists. The sum of the mru list and mru ghost
	 * list should never exceed the target size of the cache, and
	 * the sum of the mru list, mfu list, mru ghost list, and mfu
	 * ghost list should never exceed twice the target size of the
	 * cache. The following logic enforces these limits on the ghost
	 * caches, and evicts from them as needed.
	 */
	target = zfs_refcount_count(&arc_mru->arcs_size) +
	    zfs_refcount_count(&arc_mru_ghost->arcs_size) - arc_c;

	bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
	total_evicted += bytes;

	target -= bytes;

	total_evicted +=
	    arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);

	/*
	 * We assume the sum of the mru list and mfu list is less than
	 * or equal to arc_c (we enforced this above), which means we
	 * can use the simpler of the two equations below:
	 *
	 *	mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
	 *		    mru ghost + mfu ghost <= arc_c
	 */
	target = zfs_refcount_count(&arc_mru_ghost->arcs_size) +
	    zfs_refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;

	bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
	total_evicted += bytes;

	target -= bytes;

	total_evicted +=
	    arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);

	return (total_evicted);
}

void
arc_flush(spa_t *spa, boolean_t retry)
{
	uint64_t guid = 0;

	/*
	 * If retry is B_TRUE, a spa must not be specified since we have
	 * no good way to determine if all of a spa's buffers have been
	 * evicted from an arc state.
	 */
	ASSERT(!retry || spa == 0);

	if (spa != NULL)
		guid = spa_load_guid(spa);

	(void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
	(void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);

	(void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
	(void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);

	(void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
	(void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);

	(void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
	(void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
}

static void
arc_reduce_target_size(int64_t to_free)
{
	uint64_t asize = aggsum_value(&arc_size);
	if (arc_c > arc_c_min) {

		if (arc_c > arc_c_min + to_free)
			atomic_add_64(&arc_c, -to_free);
		else
			arc_c = arc_c_min;

		atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
		if (asize < arc_c)
			arc_c = MAX(asize, arc_c_min);
		if (arc_p > arc_c)
			arc_p = (arc_c >> 1);
		ASSERT(arc_c >= arc_c_min);
		ASSERT((int64_t)arc_p >= 0);
	}

	if (asize > arc_c) {
		/* See comment in arc_adjust_cb_check() on why lock+flag */
		mutex_enter(&arc_adjust_lock);
		arc_adjust_needed = B_TRUE;
		mutex_exit(&arc_adjust_lock);
		zthr_wakeup(arc_adjust_zthr);
	}
}

typedef enum free_memory_reason_t {
	FMR_UNKNOWN,
	FMR_NEEDFREE,
	FMR_LOTSFREE,
	FMR_SWAPFS_MINFREE,
	FMR_PAGES_PP_MAXIMUM,
	FMR_HEAP_ARENA,
	FMR_ZIO_ARENA,
} free_memory_reason_t;

int64_t last_free_memory;
free_memory_reason_t last_free_reason;

/*
 * Additional reserve of pages for pp_reserve.
 */
int64_t arc_pages_pp_reserve = 64;

/*
 * Additional reserve of pages for swapfs.
 */
int64_t arc_swapfs_reserve = 64;

/*
 * Return the amount of memory that can be consumed before reclaim will be
 * needed.  Positive if there is sufficient free memory, negative indicates
 * the amount of memory that needs to be freed up.
 */
static int64_t
arc_available_memory(void)
{
	int64_t lowest = INT64_MAX;
	int64_t n;
	free_memory_reason_t r = FMR_UNKNOWN;

#ifdef _KERNEL
	if (needfree > 0) {
		n = PAGESIZE * (-needfree);
		if (n < lowest) {
			lowest = n;
			r = FMR_NEEDFREE;
		}
	}

	/*
	 * check that we're out of range of the pageout scanner.  It starts to
	 * schedule paging if freemem is less than lotsfree and needfree.
	 * lotsfree is the high-water mark for pageout, and needfree is the
	 * number of needed free pages.  We add extra pages here to make sure
	 * the scanner doesn't start up while we're freeing memory.
	 */
	n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
	if (n < lowest) {
		lowest = n;
		r = FMR_LOTSFREE;
	}

	/*
	 * check to make sure that swapfs has enough space so that anon
	 * reservations can still succeed. anon_resvmem() checks that the
	 * availrmem is greater than swapfs_minfree, and the number of reserved
	 * swap pages.  We also add a bit of extra here just to prevent
	 * circumstances from getting really dire.
	 */
	n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
	    desfree - arc_swapfs_reserve);
	if (n < lowest) {
		lowest = n;
		r = FMR_SWAPFS_MINFREE;
	}


	/*
	 * Check that we have enough availrmem that memory locking (e.g., via
	 * mlock(3C) or memcntl(2)) can still succeed.  (pages_pp_maximum
	 * stores the number of pages that cannot be locked; when availrmem
	 * drops below pages_pp_maximum, page locking mechanisms such as
	 * page_pp_lock() will fail.)
	 */
	n = PAGESIZE * (availrmem - pages_pp_maximum -
	    arc_pages_pp_reserve);
	if (n < lowest) {
		lowest = n;
		r = FMR_PAGES_PP_MAXIMUM;
	}

#if defined(__i386)
	/*
	 * If we're on an i386 platform, it's possible that we'll exhaust the
	 * kernel heap space before we ever run out of available physical
	 * memory.  Most checks of the size of the heap_area compare against
	 * tune.t_minarmem, which is the minimum available real memory that we
	 * can have in the system.  However, this is generally fixed at 25 pages
	 * which is so low that it's useless.  In this comparison, we seek to
	 * calculate the total heap-size, and reclaim if more than 3/4ths of the
	 * heap is allocated.  (Or, in the calculation, if less than 1/4th is
	 * free)
	 */
	n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
	    (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
	if (n < lowest) {
		lowest = n;
		r = FMR_HEAP_ARENA;
	}
#endif

	/*
	 * If zio data pages are being allocated out of a separate heap segment,
	 * then enforce that the size of available vmem for this arena remains
	 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
	 *
	 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
	 * memory (in the zio_arena) free, which can avoid memory
	 * fragmentation issues.
	 */
	if (zio_arena != NULL) {
		n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
		    (vmem_size(zio_arena, VMEM_ALLOC) >>
		    arc_zio_arena_free_shift);
		if (n < lowest) {
			lowest = n;
			r = FMR_ZIO_ARENA;
		}
	}
#else
	/* Every 100 calls, free a small amount */
	if (spa_get_random(100) == 0)
		lowest = -1024;
#endif

	last_free_memory = lowest;
	last_free_reason = r;

	return (lowest);
}


/*
 * Determine if the system is under memory pressure and is asking
 * to reclaim memory. A return value of B_TRUE indicates that the system
 * is under memory pressure and that the arc should adjust accordingly.
 */
static boolean_t
arc_reclaim_needed(void)
{
	return (arc_available_memory() < 0);
}

static void
arc_kmem_reap_soon(void)
{
	size_t			i;
	kmem_cache_t		*prev_cache = NULL;
	kmem_cache_t		*prev_data_cache = NULL;
	extern kmem_cache_t	*zio_buf_cache[];
	extern kmem_cache_t	*zio_data_buf_cache[];
	extern kmem_cache_t	*zfs_btree_leaf_cache;
	extern kmem_cache_t	*abd_chunk_cache;

#ifdef _KERNEL
	if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) {
		/*
		 * We are exceeding our meta-data cache limit.
		 * Purge some DNLC entries to release holds on meta-data.
		 */
		dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
	}
#if defined(__i386)
	/*
	 * Reclaim unused memory from all kmem caches.
	 */
	kmem_reap();
#endif
#endif

	for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
		if (zio_buf_cache[i] != prev_cache) {
			prev_cache = zio_buf_cache[i];
			kmem_cache_reap_soon(zio_buf_cache[i]);
		}
		if (zio_data_buf_cache[i] != prev_data_cache) {
			prev_data_cache = zio_data_buf_cache[i];
			kmem_cache_reap_soon(zio_data_buf_cache[i]);
		}
	}
	kmem_cache_reap_soon(abd_chunk_cache);
	kmem_cache_reap_soon(buf_cache);
	kmem_cache_reap_soon(hdr_full_cache);
	kmem_cache_reap_soon(hdr_l2only_cache);
	kmem_cache_reap_soon(zfs_btree_leaf_cache);

	if (zio_arena != NULL) {
		/*
		 * Ask the vmem arena to reclaim unused memory from its
		 * quantum caches.
		 */
		vmem_qcache_reap(zio_arena);
	}
}

/* ARGSUSED */
static boolean_t
arc_adjust_cb_check(void *arg, zthr_t *zthr)
{
	/*
	 * This is necessary in order for the mdb ::arc dcmd to
	 * show up to date information. Since the ::arc command
	 * does not call the kstat's update function, without
	 * this call, the command may show stale stats for the
	 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
	 * with this change, the data might be up to 1 second
	 * out of date(the arc_adjust_zthr has a maximum sleep
	 * time of 1 second); but that should suffice.  The
	 * arc_state_t structures can be queried directly if more
	 * accurate information is needed.
	 */
	if (arc_ksp != NULL)
		arc_ksp->ks_update(arc_ksp, KSTAT_READ);

	/*
	 * We have to rely on arc_get_data_impl() to tell us when to adjust,
	 * rather than checking if we are overflowing here, so that we are
	 * sure to not leave arc_get_data_impl() waiting on
	 * arc_adjust_waiters_cv.  If we have become "not overflowing" since
	 * arc_get_data_impl() checked, we need to wake it up.  We could
	 * broadcast the CV here, but arc_get_data_impl() may have not yet
	 * gone to sleep.  We would need to use a mutex to ensure that this
	 * function doesn't broadcast until arc_get_data_impl() has gone to
	 * sleep (e.g. the arc_adjust_lock).  However, the lock ordering of
	 * such a lock would necessarily be incorrect with respect to the
	 * zthr_lock, which is held before this function is called, and is
	 * held by arc_get_data_impl() when it calls zthr_wakeup().
	 */
	return (arc_adjust_needed);
}

/*
 * Keep arc_size under arc_c by running arc_adjust which evicts data
 * from the ARC.
 */
/* ARGSUSED */
static void
arc_adjust_cb(void *arg, zthr_t *zthr)
{
	uint64_t evicted = 0;

	/* Evict from cache */
	evicted = arc_adjust();

	/*
	 * If evicted is zero, we couldn't evict anything
	 * via arc_adjust(). This could be due to hash lock
	 * collisions, but more likely due to the majority of
	 * arc buffers being unevictable. Therefore, even if
	 * arc_size is above arc_c, another pass is unlikely to
	 * be helpful and could potentially cause us to enter an
	 * infinite loop.  Additionally, zthr_iscancelled() is
	 * checked here so that if the arc is shutting down, the
	 * broadcast will wake any remaining arc adjust waiters.
	 */
	mutex_enter(&arc_adjust_lock);
	arc_adjust_needed = !zthr_iscancelled(arc_adjust_zthr) &&
	    evicted > 0 && aggsum_compare(&arc_size, arc_c) > 0;
	if (!arc_adjust_needed) {
		/*
		 * We're either no longer overflowing, or we
		 * can't evict anything more, so we should wake
		 * up any waiters.
		 */
		cv_broadcast(&arc_adjust_waiters_cv);
	}
	mutex_exit(&arc_adjust_lock);
}

/* ARGSUSED */
static boolean_t
arc_reap_cb_check(void *arg, zthr_t *zthr)
{
	int64_t free_memory = arc_available_memory();

	/*
	 * If a kmem reap is already active, don't schedule more.  We must
	 * check for this because kmem_cache_reap_soon() won't actually
	 * block on the cache being reaped (this is to prevent callers from
	 * becoming implicitly blocked by a system-wide kmem reap -- which,
	 * on a system with many, many full magazines, can take minutes).
	 */
	if (!kmem_cache_reap_active() &&
	    free_memory < 0) {
		arc_no_grow = B_TRUE;
		arc_warm = B_TRUE;
		/*
		 * Wait at least zfs_grow_retry (default 60) seconds
		 * before considering growing.
		 */
		arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
		return (B_TRUE);
	} else if (free_memory < arc_c >> arc_no_grow_shift) {
		arc_no_grow = B_TRUE;
	} else if (gethrtime() >= arc_growtime) {
		arc_no_grow = B_FALSE;
	}

	return (B_FALSE);
}

/*
 * Keep enough free memory in the system by reaping the ARC's kmem
 * caches.  To cause more slabs to be reapable, we may reduce the
 * target size of the cache (arc_c), causing the arc_adjust_cb()
 * to free more buffers.
 */
/* ARGSUSED */
static void
arc_reap_cb(void *arg, zthr_t *zthr)
{
	int64_t free_memory;

	/*
	 * Kick off asynchronous kmem_reap()'s of all our caches.
	 */
	arc_kmem_reap_soon();

	/*
	 * Wait at least arc_kmem_cache_reap_retry_ms between
	 * arc_kmem_reap_soon() calls. Without this check it is possible to
	 * end up in a situation where we spend lots of time reaping
	 * caches, while we're near arc_c_min.  Waiting here also gives the
	 * subsequent free memory check a chance of finding that the
	 * asynchronous reap has already freed enough memory, and we don't
	 * need to call arc_reduce_target_size().
	 */
	delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000);

	/*
	 * Reduce the target size as needed to maintain the amount of free
	 * memory in the system at a fraction of the arc_size (1/128th by
	 * default).  If oversubscribed (free_memory < 0) then reduce the
	 * target arc_size by the deficit amount plus the fractional
	 * amount.  If free memory is positive but less then the fractional
	 * amount, reduce by what is needed to hit the fractional amount.
	 */
	free_memory = arc_available_memory();

	int64_t to_free =
	    (arc_c >> arc_shrink_shift) - free_memory;
	if (to_free > 0) {
#ifdef _KERNEL
		to_free = MAX(to_free, ptob(needfree));
#endif
		arc_reduce_target_size(to_free);
	}
}

/*
 * Adapt arc info given the number of bytes we are trying to add and
 * the state that we are coming from.  This function is only called
 * when we are adding new content to the cache.
 */
static void
arc_adapt(int bytes, arc_state_t *state)
{
	int mult;
	uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
	int64_t mrug_size = zfs_refcount_count(&arc_mru_ghost->arcs_size);
	int64_t mfug_size = zfs_refcount_count(&arc_mfu_ghost->arcs_size);

	if (state == arc_l2c_only)
		return;

	ASSERT(bytes > 0);
	/*
	 * Adapt the target size of the MRU list:
	 *	- if we just hit in the MRU ghost list, then increase
	 *	  the target size of the MRU list.
	 *	- if we just hit in the MFU ghost list, then increase
	 *	  the target size of the MFU list by decreasing the
	 *	  target size of the MRU list.
	 */
	if (state == arc_mru_ghost) {
		mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
		mult = MIN(mult, 10); /* avoid wild arc_p adjustment */

		arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
	} else if (state == arc_mfu_ghost) {
		uint64_t delta;

		mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
		mult = MIN(mult, 10);

		delta = MIN(bytes * mult, arc_p);
		arc_p = MAX(arc_p_min, arc_p - delta);
	}
	ASSERT((int64_t)arc_p >= 0);

	/*
	 * Wake reap thread if we do not have any available memory
	 */
	if (arc_reclaim_needed()) {
		zthr_wakeup(arc_reap_zthr);
		return;
	}


	if (arc_no_grow)
		return;

	if (arc_c >= arc_c_max)
		return;

	/*
	 * If we're within (2 * maxblocksize) bytes of the target
	 * cache size, increment the target cache size
	 */
	if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >
	    0) {
		atomic_add_64(&arc_c, (int64_t)bytes);
		if (arc_c > arc_c_max)
			arc_c = arc_c_max;
		else if (state == arc_anon)
			atomic_add_64(&arc_p, (int64_t)bytes);
		if (arc_p > arc_c)
			arc_p = arc_c;
	}
	ASSERT((int64_t)arc_p >= 0);
}

/*
 * Check if arc_size has grown past our upper threshold, determined by
 * zfs_arc_overflow_shift.
 */
static boolean_t
arc_is_overflowing(void)
{
	/* Always allow at least one block of overflow */
	uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
	    arc_c >> zfs_arc_overflow_shift);

	/*
	 * We just compare the lower bound here for performance reasons. Our
	 * primary goals are to make sure that the arc never grows without
	 * bound, and that it can reach its maximum size. This check
	 * accomplishes both goals. The maximum amount we could run over by is
	 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
	 * in the ARC. In practice, that's in the tens of MB, which is low
	 * enough to be safe.
	 */
	return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
}

static abd_t *
arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
{
	arc_buf_contents_t type = arc_buf_type(hdr);

	arc_get_data_impl(hdr, size, tag);
	if (type == ARC_BUFC_METADATA) {
		return (abd_alloc(size, B_TRUE));
	} else {
		ASSERT(type == ARC_BUFC_DATA);
		return (abd_alloc(size, B_FALSE));
	}
}

static void *
arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
{
	arc_buf_contents_t type = arc_buf_type(hdr);

	arc_get_data_impl(hdr, size, tag);
	if (type == ARC_BUFC_METADATA) {
		return (zio_buf_alloc(size));
	} else {
		ASSERT(type == ARC_BUFC_DATA);
		return (zio_data_buf_alloc(size));
	}
}

/*
 * Allocate a block and return it to the caller. If we are hitting the
 * hard limit for the cache size, we must sleep, waiting for the eviction
 * thread to catch up. If we're past the target size but below the hard
 * limit, we'll only signal the reclaim thread and continue on.
 */
static void
arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
{
	arc_state_t *state = hdr->b_l1hdr.b_state;
	arc_buf_contents_t type = arc_buf_type(hdr);

	arc_adapt(size, state);

	/*
	 * If arc_size is currently overflowing, and has grown past our
	 * upper limit, we must be adding data faster than the evict
	 * thread can evict. Thus, to ensure we don't compound the
	 * problem by adding more data and forcing arc_size to grow even
	 * further past its target size, we halt and wait for the
	 * eviction thread to catch up.
	 *
	 * It's also possible that the reclaim thread is unable to evict
	 * enough buffers to get arc_size below the overflow limit (e.g.
	 * due to buffers being un-evictable, or hash lock collisions).
	 * In this case, we want to proceed regardless if we're
	 * overflowing; thus we don't use a while loop here.
	 */
	if (arc_is_overflowing()) {
		mutex_enter(&arc_adjust_lock);

		/*
		 * Now that we've acquired the lock, we may no longer be
		 * over the overflow limit, lets check.
		 *
		 * We're ignoring the case of spurious wake ups. If that
		 * were to happen, it'd let this thread consume an ARC
		 * buffer before it should have (i.e. before we're under
		 * the overflow limit and were signalled by the reclaim
		 * thread). As long as that is a rare occurrence, it
		 * shouldn't cause any harm.
		 */
		if (arc_is_overflowing()) {
			arc_adjust_needed = B_TRUE;
			zthr_wakeup(arc_adjust_zthr);
			(void) cv_wait(&arc_adjust_waiters_cv,
			    &arc_adjust_lock);
		}
		mutex_exit(&arc_adjust_lock);
	}

	VERIFY3U(hdr->b_type, ==, type);
	if (type == ARC_BUFC_METADATA) {
		arc_space_consume(size, ARC_SPACE_META);
	} else {
		arc_space_consume(size, ARC_SPACE_DATA);
	}

	/*
	 * Update the state size.  Note that ghost states have a
	 * "ghost size" and so don't need to be updated.
	 */
	if (!GHOST_STATE(state)) {

		(void) zfs_refcount_add_many(&state->arcs_size, size, tag);

		/*
		 * If this is reached via arc_read, the link is
		 * protected by the hash lock. If reached via
		 * arc_buf_alloc, the header should not be accessed by
		 * any other thread. And, if reached via arc_read_done,
		 * the hash lock will protect it if it's found in the
		 * hash table; otherwise no other thread should be
		 * trying to [add|remove]_reference it.
		 */
		if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
			ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
			(void) zfs_refcount_add_many(&state->arcs_esize[type],
			    size, tag);
		}

		/*
		 * If we are growing the cache, and we are adding anonymous
		 * data, and we have outgrown arc_p, update arc_p
		 */
		if (aggsum_compare(&arc_size, arc_c) < 0 &&
		    hdr->b_l1hdr.b_state == arc_anon &&
		    (zfs_refcount_count(&arc_anon->arcs_size) +
		    zfs_refcount_count(&arc_mru->arcs_size) > arc_p))
			arc_p = MIN(arc_c, arc_p + size);
	}
}

static void
arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
{
	arc_free_data_impl(hdr, size, tag);
	abd_free(abd);
}

static void
arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
{
	arc_buf_contents_t type = arc_buf_type(hdr);

	arc_free_data_impl(hdr, size, tag);
	if (type == ARC_BUFC_METADATA) {
		zio_buf_free(buf, size);
	} else {
		ASSERT(type == ARC_BUFC_DATA);
		zio_data_buf_free(buf, size);
	}
}

/*
 * Free the arc data buffer.
 */
static void
arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
{
	arc_state_t *state = hdr->b_l1hdr.b_state;
	arc_buf_contents_t type = arc_buf_type(hdr);

	/* protected by hash lock, if in the hash table */
	if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
		ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
		ASSERT(state != arc_anon && state != arc_l2c_only);

		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
		    size, tag);
	}
	(void) zfs_refcount_remove_many(&state->arcs_size, size, tag);

	VERIFY3U(hdr->b_type, ==, type);
	if (type == ARC_BUFC_METADATA) {
		arc_space_return(size, ARC_SPACE_META);
	} else {
		ASSERT(type == ARC_BUFC_DATA);
		arc_space_return(size, ARC_SPACE_DATA);
	}
}

/*
 * This routine is called whenever a buffer is accessed.
 * NOTE: the hash lock is dropped in this function.
 */
static void
arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
{
	clock_t now;

	ASSERT(MUTEX_HELD(hash_lock));
	ASSERT(HDR_HAS_L1HDR(hdr));

	if (hdr->b_l1hdr.b_state == arc_anon) {
		/*
		 * This buffer is not in the cache, and does not
		 * appear in our "ghost" list.  Add the new buffer
		 * to the MRU state.
		 */

		ASSERT0(hdr->b_l1hdr.b_arc_access);
		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
		DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
		arc_change_state(arc_mru, hdr, hash_lock);

	} else if (hdr->b_l1hdr.b_state == arc_mru) {
		now = ddi_get_lbolt();

		/*
		 * If this buffer is here because of a prefetch, then either:
		 * - clear the flag if this is a "referencing" read
		 *   (any subsequent access will bump this into the MFU state).
		 * or
		 * - move the buffer to the head of the list if this is
		 *   another prefetch (to make it less likely to be evicted).
		 */
		if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
			if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
				/* link protected by hash lock */
				ASSERT(multilist_link_active(
				    &hdr->b_l1hdr.b_arc_node));
			} else {
				arc_hdr_clear_flags(hdr,
				    ARC_FLAG_PREFETCH |
				    ARC_FLAG_PRESCIENT_PREFETCH);
				ARCSTAT_BUMP(arcstat_mru_hits);
			}
			hdr->b_l1hdr.b_arc_access = now;
			return;
		}

		/*
		 * This buffer has been "accessed" only once so far,
		 * but it is still in the cache. Move it to the MFU
		 * state.
		 */
		if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
			/*
			 * More than 125ms have passed since we
			 * instantiated this buffer.  Move it to the
			 * most frequently used state.
			 */
			hdr->b_l1hdr.b_arc_access = now;
			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
			arc_change_state(arc_mfu, hdr, hash_lock);
		}
		ARCSTAT_BUMP(arcstat_mru_hits);
	} else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
		arc_state_t	*new_state;
		/*
		 * This buffer has been "accessed" recently, but
		 * was evicted from the cache.  Move it to the
		 * MFU state.
		 */

		if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
			new_state = arc_mru;
			if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
				arc_hdr_clear_flags(hdr,
				    ARC_FLAG_PREFETCH |
				    ARC_FLAG_PRESCIENT_PREFETCH);
			}
			DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
		} else {
			new_state = arc_mfu;
			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
		}

		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
		arc_change_state(new_state, hdr, hash_lock);

		ARCSTAT_BUMP(arcstat_mru_ghost_hits);
	} else if (hdr->b_l1hdr.b_state == arc_mfu) {
		/*
		 * This buffer has been accessed more than once and is
		 * still in the cache.  Keep it in the MFU state.
		 *
		 * NOTE: an add_reference() that occurred when we did
		 * the arc_read() will have kicked this off the list.
		 * If it was a prefetch, we will explicitly move it to
		 * the head of the list now.
		 */
		ARCSTAT_BUMP(arcstat_mfu_hits);
		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
	} else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
		arc_state_t	*new_state = arc_mfu;
		/*
		 * This buffer has been accessed more than once but has
		 * been evicted from the cache.  Move it back to the
		 * MFU state.
		 */

		if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
			/*
			 * This is a prefetch access...
			 * move this block back to the MRU state.
			 */
			new_state = arc_mru;
		}

		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
		arc_change_state(new_state, hdr, hash_lock);

		ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
	} else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
		/*
		 * This buffer is on the 2nd Level ARC.
		 */

		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
		arc_change_state(arc_mfu, hdr, hash_lock);
	} else {
		ASSERT(!"invalid arc state");
	}
}

/*
 * This routine is called by dbuf_hold() to update the arc_access() state
 * which otherwise would be skipped for entries in the dbuf cache.
 */
void
arc_buf_access(arc_buf_t *buf)
{
	mutex_enter(&buf->b_evict_lock);
	arc_buf_hdr_t *hdr = buf->b_hdr;

	/*
	 * Avoid taking the hash_lock when possible as an optimization.
	 * The header must be checked again under the hash_lock in order
	 * to handle the case where it is concurrently being released.
	 */
	if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
		mutex_exit(&buf->b_evict_lock);
		return;
	}

	kmutex_t *hash_lock = HDR_LOCK(hdr);
	mutex_enter(hash_lock);

	if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
		mutex_exit(hash_lock);
		mutex_exit(&buf->b_evict_lock);
		ARCSTAT_BUMP(arcstat_access_skip);
		return;
	}

	mutex_exit(&buf->b_evict_lock);

	ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
	    hdr->b_l1hdr.b_state == arc_mfu);

	DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
	arc_access(hdr, hash_lock);
	mutex_exit(hash_lock);

	ARCSTAT_BUMP(arcstat_hits);
	ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
	    demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
}

/* a generic arc_read_done_func_t which you can use */
/* ARGSUSED */
void
arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
    arc_buf_t *buf, void *arg)
{
	if (buf == NULL)
		return;

	bcopy(buf->b_data, arg, arc_buf_size(buf));
	arc_buf_destroy(buf, arg);
}

/* a generic arc_read_done_func_t */
void
arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
    arc_buf_t *buf, void *arg)
{
	arc_buf_t **bufp = arg;

	if (buf == NULL) {
		ASSERT(zio == NULL || zio->io_error != 0);
		*bufp = NULL;
	} else {
		ASSERT(zio == NULL || zio->io_error == 0);
		*bufp = buf;
		ASSERT(buf->b_data != NULL);
	}
}

static void
arc_hdr_verify(arc_buf_hdr_t *hdr, const blkptr_t *bp)
{
	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
		ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
		ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF);
	} else {
		if (HDR_COMPRESSION_ENABLED(hdr)) {
			ASSERT3U(arc_hdr_get_compress(hdr), ==,
			    BP_GET_COMPRESS(bp));
		}
		ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
		ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
		ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp));
	}
}

/*
 * XXX this should be changed to return an error, and callers
 * re-read from disk on failure (on nondebug bits).
 */
static void
arc_hdr_verify_checksum(spa_t *spa, arc_buf_hdr_t *hdr, const blkptr_t *bp)
{
	arc_hdr_verify(hdr, bp);
	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
		return;
	int err = 0;
	abd_t *abd = NULL;
	if (BP_IS_ENCRYPTED(bp)) {
		if (HDR_HAS_RABD(hdr)) {
			abd = hdr->b_crypt_hdr.b_rabd;
		}
	} else if (HDR_COMPRESSION_ENABLED(hdr)) {
		abd = hdr->b_l1hdr.b_pabd;
	}
	if (abd != NULL) {
		/*
		 * The offset is only used for labels, which are not
		 * cached in the ARC, so it doesn't matter what we
		 * pass for the offset parameter.
		 */
		int psize = HDR_GET_PSIZE(hdr);
		err = zio_checksum_error_impl(spa, bp,
		    BP_GET_CHECKSUM(bp), abd, psize, 0, NULL);
		if (err != 0) {
			/*
			 * Use abd_copy_to_buf() rather than
			 * abd_borrow_buf_copy() so that we are sure to
			 * include the buf in crash dumps.
			 */
			void *buf = kmem_alloc(psize, KM_SLEEP);
			abd_copy_to_buf(buf, abd, psize);
			panic("checksum of cached data doesn't match BP "
			    "err=%u hdr=%p bp=%p abd=%p buf=%p",
			    err, (void *)hdr, (void *)bp, (void *)abd, buf);
		}
	}
}

static void
arc_read_done(zio_t *zio)
{
	blkptr_t	*bp = zio->io_bp;
	arc_buf_hdr_t	*hdr = zio->io_private;
	kmutex_t	*hash_lock = NULL;
	arc_callback_t	*callback_list;
	arc_callback_t	*acb;
	boolean_t	freeable = B_FALSE;

	/*
	 * The hdr was inserted into hash-table and removed from lists
	 * prior to starting I/O.  We should find this header, since
	 * it's in the hash table, and it should be legit since it's
	 * not possible to evict it during the I/O.  The only possible
	 * reason for it not to be found is if we were freed during the
	 * read.
	 */
	if (HDR_IN_HASH_TABLE(hdr)) {
		ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
		ASSERT3U(hdr->b_dva.dva_word[0], ==,
		    BP_IDENTITY(zio->io_bp)->dva_word[0]);
		ASSERT3U(hdr->b_dva.dva_word[1], ==,
		    BP_IDENTITY(zio->io_bp)->dva_word[1]);

		arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
		    &hash_lock);

		ASSERT((found == hdr &&
		    DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
		    (found == hdr && HDR_L2_READING(hdr)));
		ASSERT3P(hash_lock, !=, NULL);
	}

	if (BP_IS_PROTECTED(bp)) {
		hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
		hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
		zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
		    hdr->b_crypt_hdr.b_iv);

		if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) {
			void *tmpbuf;

			tmpbuf = abd_borrow_buf_copy(zio->io_abd,
			    sizeof (zil_chain_t));
			zio_crypt_decode_mac_zil(tmpbuf,
			    hdr->b_crypt_hdr.b_mac);
			abd_return_buf(zio->io_abd, tmpbuf,
			    sizeof (zil_chain_t));
		} else {
			zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
		}
	}

	if (zio->io_error == 0) {
		/* byteswap if necessary */
		if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
			if (BP_GET_LEVEL(zio->io_bp) > 0) {
				hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
			} else {
				hdr->b_l1hdr.b_byteswap =
				    DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
			}
		} else {
			hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
		}
	}

	arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
	if (l2arc_noprefetch && HDR_PREFETCH(hdr))
		arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);

	callback_list = hdr->b_l1hdr.b_acb;
	ASSERT3P(callback_list, !=, NULL);

	if (hash_lock && zio->io_error == 0 &&
	    hdr->b_l1hdr.b_state == arc_anon) {
		/*
		 * Only call arc_access on anonymous buffers.  This is because
		 * if we've issued an I/O for an evicted buffer, we've already
		 * called arc_access (to prevent any simultaneous readers from
		 * getting confused).
		 */
		arc_access(hdr, hash_lock);
	}

	/*
	 * If a read request has a callback (i.e. acb_done is not NULL), then we
	 * make a buf containing the data according to the parameters which were
	 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
	 * aren't needlessly decompressing the data multiple times.
	 */
	int callback_cnt = 0;
	for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
		if (!acb->acb_done)
			continue;

		callback_cnt++;

		if (zio->io_error != 0)
			continue;

		int error = arc_buf_alloc_impl(hdr, zio->io_spa,
		    &acb->acb_zb, acb->acb_private, acb->acb_encrypted,
		    acb->acb_compressed, acb->acb_noauth, B_TRUE,
		    &acb->acb_buf);

		/*
		 * Assert non-speculative zios didn't fail because an
		 * encryption key wasn't loaded
		 */
		ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) ||
		    error != EACCES);

		/*
		 * If we failed to decrypt, report an error now (as the zio
		 * layer would have done if it had done the transforms).
		 */
		if (error == ECKSUM) {
			ASSERT(BP_IS_PROTECTED(bp));
			error = SET_ERROR(EIO);
			if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
				spa_log_error(zio->io_spa, &acb->acb_zb);
				zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
				    zio->io_spa, NULL, &acb->acb_zb, zio, 0, 0);
			}
		}

		if (error != 0) {
			/*
			 * Decompression failed.  Set io_error
			 * so that when we call acb_done (below),
			 * we will indicate that the read failed.
			 * Note that in the unusual case where one
			 * callback is compressed and another
			 * uncompressed, we will mark all of them
			 * as failed, even though the uncompressed
			 * one can't actually fail.  In this case,
			 * the hdr will not be anonymous, because
			 * if there are multiple callbacks, it's
			 * because multiple threads found the same
			 * arc buf in the hash table.
			 */
			zio->io_error = error;
		}
	}

	/*
	 * If there are multiple callbacks, we must have the hash lock,
	 * because the only way for multiple threads to find this hdr is
	 * in the hash table.  This ensures that if there are multiple
	 * callbacks, the hdr is not anonymous.  If it were anonymous,
	 * we couldn't use arc_buf_destroy() in the error case below.
	 */
	ASSERT(callback_cnt < 2 || hash_lock != NULL);

	hdr->b_l1hdr.b_acb = NULL;
	arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
	if (callback_cnt == 0)
		ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));

	ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
	    callback_list != NULL);

	if (zio->io_error == 0) {
		arc_hdr_verify(hdr, zio->io_bp);
	} else {
		arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
		if (hdr->b_l1hdr.b_state != arc_anon)
			arc_change_state(arc_anon, hdr, hash_lock);
		if (HDR_IN_HASH_TABLE(hdr))
			buf_hash_remove(hdr);
		freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
	}

	/*
	 * Broadcast before we drop the hash_lock to avoid the possibility
	 * that the hdr (and hence the cv) might be freed before we get to
	 * the cv_broadcast().
	 */
	cv_broadcast(&hdr->b_l1hdr.b_cv);

	if (hash_lock != NULL) {
		mutex_exit(hash_lock);
	} else {
		/*
		 * This block was freed while we waited for the read to
		 * complete.  It has been removed from the hash table and
		 * moved to the anonymous state (so that it won't show up
		 * in the cache).
		 */
		ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
		freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
	}

	/* execute each callback and free its structure */
	while ((acb = callback_list) != NULL) {

		if (acb->acb_done != NULL) {
			if (zio->io_error != 0 && acb->acb_buf != NULL) {
				/*
				 * If arc_buf_alloc_impl() fails during
				 * decompression, the buf will still be
				 * allocated, and needs to be freed here.
				 */
				arc_buf_destroy(acb->acb_buf, acb->acb_private);
				acb->acb_buf = NULL;
			}
			acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
			    acb->acb_buf, acb->acb_private);
		}

		if (acb->acb_zio_dummy != NULL) {
			acb->acb_zio_dummy->io_error = zio->io_error;
			zio_nowait(acb->acb_zio_dummy);
		}

		callback_list = acb->acb_next;
		kmem_free(acb, sizeof (arc_callback_t));
	}

	if (freeable)
		arc_hdr_destroy(hdr);
}

/*
 * "Read" the block at the specified DVA (in bp) via the
 * cache.  If the block is found in the cache, invoke the provided
 * callback immediately and return.  Note that the `zio' parameter
 * in the callback will be NULL in this case, since no IO was
 * required.  If the block is not in the cache pass the read request
 * on to the spa with a substitute callback function, so that the
 * requested block will be added to the cache.
 *
 * If a read request arrives for a block that has a read in-progress,
 * either wait for the in-progress read to complete (and return the
 * results); or, if this is a read with a "done" func, add a record
 * to the read to invoke the "done" func when the read completes,
 * and return; or just return.
 *
 * arc_read_done() will invoke all the requested "done" functions
 * for readers of this block.
 */
int
arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_read_done_func_t *done,
    void *private, zio_priority_t priority, int zio_flags,
    arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
{
	arc_buf_hdr_t *hdr = NULL;
	kmutex_t *hash_lock = NULL;
	zio_t *rzio;
	uint64_t guid = spa_load_guid(spa);
	boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0;
	boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) &&
	    (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
	boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) &&
	    (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
	int rc = 0;

	ASSERT(!BP_IS_EMBEDDED(bp) ||
	    BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);

top:
	if (!BP_IS_EMBEDDED(bp)) {
		/*
		 * Embedded BP's have no DVA and require no I/O to "read".
		 * Create an anonymous arc buf to back it.
		 */
		hdr = buf_hash_find(guid, bp, &hash_lock);
	}

	/*
	 * Determine if we have an L1 cache hit or a cache miss. For simplicity
	 * we maintain encrypted data seperately from compressed / uncompressed
	 * data. If the user is requesting raw encrypted data and we don't have
	 * that in the header we will read from disk to guarantee that we can
	 * get it even if the encryption keys aren't loaded.
	 */
	if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) ||
	    (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) {
		arc_buf_t *buf = NULL;
		*arc_flags |= ARC_FLAG_CACHED;

		if (HDR_IO_IN_PROGRESS(hdr)) {
			zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;

			ASSERT3P(head_zio, !=, NULL);
			if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
			    priority == ZIO_PRIORITY_SYNC_READ) {
				/*
				 * This is a sync read that needs to wait for
				 * an in-flight async read. Request that the
				 * zio have its priority upgraded.
				 */
				zio_change_priority(head_zio, priority);
				DTRACE_PROBE1(arc__async__upgrade__sync,
				    arc_buf_hdr_t *, hdr);
				ARCSTAT_BUMP(arcstat_async_upgrade_sync);
			}
			if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
				arc_hdr_clear_flags(hdr,
				    ARC_FLAG_PREDICTIVE_PREFETCH);
			}

			if (*arc_flags & ARC_FLAG_WAIT) {
				cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
				mutex_exit(hash_lock);
				goto top;
			}
			ASSERT(*arc_flags & ARC_FLAG_NOWAIT);

			if (done) {
				arc_callback_t *acb = NULL;

				acb = kmem_zalloc(sizeof (arc_callback_t),
				    KM_SLEEP);
				acb->acb_done = done;
				acb->acb_private = private;
				acb->acb_compressed = compressed_read;
				acb->acb_encrypted = encrypted_read;
				acb->acb_noauth = noauth_read;
				acb->acb_zb = *zb;
				if (pio != NULL)
					acb->acb_zio_dummy = zio_null(pio,
					    spa, NULL, NULL, NULL, zio_flags);

				ASSERT3P(acb->acb_done, !=, NULL);
				acb->acb_zio_head = head_zio;
				acb->acb_next = hdr->b_l1hdr.b_acb;
				hdr->b_l1hdr.b_acb = acb;
				mutex_exit(hash_lock);
				return (0);
			}
			mutex_exit(hash_lock);
			return (0);
		}

		ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
		    hdr->b_l1hdr.b_state == arc_mfu);

		if (done) {
			if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
				/*
				 * This is a demand read which does not have to
				 * wait for i/o because we did a predictive
				 * prefetch i/o for it, which has completed.
				 */
				DTRACE_PROBE1(
				    arc__demand__hit__predictive__prefetch,
				    arc_buf_hdr_t *, hdr);
				ARCSTAT_BUMP(
				    arcstat_demand_hit_predictive_prefetch);
				arc_hdr_clear_flags(hdr,
				    ARC_FLAG_PREDICTIVE_PREFETCH);
			}

			if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
				ARCSTAT_BUMP(
				    arcstat_demand_hit_prescient_prefetch);
				arc_hdr_clear_flags(hdr,
				    ARC_FLAG_PRESCIENT_PREFETCH);
			}

			ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));

			arc_hdr_verify_checksum(spa, hdr, bp);

			/* Get a buf with the desired data in it. */
			rc = arc_buf_alloc_impl(hdr, spa, zb, private,
			    encrypted_read, compressed_read, noauth_read,
			    B_TRUE, &buf);
			if (rc == ECKSUM) {
				/*
				 * Convert authentication and decryption errors
				 * to EIO (and generate an ereport if needed)
				 * before leaving the ARC.
				 */
				rc = SET_ERROR(EIO);
				if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) {
					spa_log_error(spa, zb);
					zfs_ereport_post(
					    FM_EREPORT_ZFS_AUTHENTICATION,
					    spa, NULL, zb, NULL, 0, 0);
				}
			}
			if (rc != 0) {
				(void) remove_reference(hdr, hash_lock,
				    private);
				arc_buf_destroy_impl(buf);
				buf = NULL;
			}
			/* assert any errors weren't due to unloaded keys */
			ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
			    rc != EACCES);
		} else if (*arc_flags & ARC_FLAG_PREFETCH &&
		    zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
			arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
		}
		DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
		arc_access(hdr, hash_lock);
		if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
			arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
		if (*arc_flags & ARC_FLAG_L2CACHE)
			arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
		mutex_exit(hash_lock);
		ARCSTAT_BUMP(arcstat_hits);
		ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
		    demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
		    data, metadata, hits);

		if (done)
			done(NULL, zb, bp, buf, private);
	} else {
		uint64_t lsize = BP_GET_LSIZE(bp);
		uint64_t psize = BP_GET_PSIZE(bp);
		arc_callback_t *acb;
		vdev_t *vd = NULL;
		uint64_t addr = 0;
		boolean_t devw = B_FALSE;
		uint64_t size;
		abd_t *hdr_abd;

		if (hdr == NULL) {
			/* this block is not in the cache */
			arc_buf_hdr_t *exists = NULL;
			arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
			hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
			    BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), type,
			    encrypted_read);

			if (!BP_IS_EMBEDDED(bp)) {
				hdr->b_dva = *BP_IDENTITY(bp);
				hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
				exists = buf_hash_insert(hdr, &hash_lock);
			}
			if (exists != NULL) {
				/* somebody beat us to the hash insert */
				mutex_exit(hash_lock);
				buf_discard_identity(hdr);
				arc_hdr_destroy(hdr);
				goto top; /* restart the IO request */
			}
		} else {
			/*
			 * This block is in the ghost cache or encrypted data
			 * was requested and we didn't have it. If it was
			 * L2-only (and thus didn't have an L1 hdr),
			 * we realloc the header to add an L1 hdr.
			 */
			if (!HDR_HAS_L1HDR(hdr)) {
				hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
				    hdr_full_cache);
			}

			if (GHOST_STATE(hdr->b_l1hdr.b_state)) {
				ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
				ASSERT(!HDR_HAS_RABD(hdr));
				ASSERT(!HDR_IO_IN_PROGRESS(hdr));
				ASSERT0(zfs_refcount_count(
				    &hdr->b_l1hdr.b_refcnt));
				ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
				ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
			} else if (HDR_IO_IN_PROGRESS(hdr)) {
				/*
				 * If this header already had an IO in progress
				 * and we are performing another IO to fetch
				 * encrypted data we must wait until the first
				 * IO completes so as not to confuse
				 * arc_read_done(). This should be very rare
				 * and so the performance impact shouldn't
				 * matter.
				 */
				cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
				mutex_exit(hash_lock);
				goto top;
			}

			/*
			 * This is a delicate dance that we play here.
			 * This hdr might be in the ghost list so we access
			 * it to move it out of the ghost list before we
			 * initiate the read. If it's a prefetch then
			 * it won't have a callback so we'll remove the
			 * reference that arc_buf_alloc_impl() created. We
			 * do this after we've called arc_access() to
			 * avoid hitting an assert in remove_reference().
			 */
			arc_access(hdr, hash_lock);
			arc_hdr_alloc_pabd(hdr, encrypted_read);
		}

		if (encrypted_read) {
			ASSERT(HDR_HAS_RABD(hdr));
			size = HDR_GET_PSIZE(hdr);
			hdr_abd = hdr->b_crypt_hdr.b_rabd;
			zio_flags |= ZIO_FLAG_RAW;
		} else {
			ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
			size = arc_hdr_size(hdr);
			hdr_abd = hdr->b_l1hdr.b_pabd;

			if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
				zio_flags |= ZIO_FLAG_RAW_COMPRESS;
			}

			/*
			 * For authenticated bp's, we do not ask the ZIO layer
			 * to authenticate them since this will cause the entire
			 * IO to fail if the key isn't loaded. Instead, we
			 * defer authentication until arc_buf_fill(), which will
			 * verify the data when the key is available.
			 */
			if (BP_IS_AUTHENTICATED(bp))
				zio_flags |= ZIO_FLAG_RAW_ENCRYPT;
		}

		if (*arc_flags & ARC_FLAG_PREFETCH &&
		    zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))
			arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
		if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
			arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);

		if (*arc_flags & ARC_FLAG_L2CACHE)
			arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
		if (BP_IS_AUTHENTICATED(bp))
			arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
		if (BP_GET_LEVEL(bp) > 0)
			arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
		if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
			arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
		ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));

		acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
		acb->acb_done = done;
		acb->acb_private = private;
		acb->acb_compressed = compressed_read;
		acb->acb_encrypted = encrypted_read;
		acb->acb_noauth = noauth_read;
		acb->acb_zb = *zb;

		ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
		hdr->b_l1hdr.b_acb = acb;
		arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);

		if (HDR_HAS_L2HDR(hdr) &&
		    (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
			devw = hdr->b_l2hdr.b_dev->l2ad_writing;
			addr = hdr->b_l2hdr.b_daddr;
			/*
			 * Lock out L2ARC device removal.
			 */
			if (vdev_is_dead(vd) ||
			    !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
				vd = NULL;
		}

		/*
		 * We count both async reads and scrub IOs as asynchronous so
		 * that both can be upgraded in the event of a cache hit while
		 * the read IO is still in-flight.
		 */
		if (priority == ZIO_PRIORITY_ASYNC_READ ||
		    priority == ZIO_PRIORITY_SCRUB)
			arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
		else
			arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);

		/*
		 * At this point, we have a level 1 cache miss.  Try again in
		 * L2ARC if possible.
		 */
		ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);

		DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
		    uint64_t, lsize, zbookmark_phys_t *, zb);
		ARCSTAT_BUMP(arcstat_misses);
		ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
		    demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
		    data, metadata, misses);

		if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
			/*
			 * Read from the L2ARC if the following are true:
			 * 1. The L2ARC vdev was previously cached.
			 * 2. This buffer still has L2ARC metadata.
			 * 3. This buffer isn't currently writing to the L2ARC.
			 * 4. The L2ARC entry wasn't evicted, which may
			 *    also have invalidated the vdev.
			 * 5. This isn't prefetch and l2arc_noprefetch is set.
			 */
			if (HDR_HAS_L2HDR(hdr) &&
			    !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
			    !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
				l2arc_read_callback_t *cb;
				abd_t *abd;
				uint64_t asize;

				DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
				ARCSTAT_BUMP(arcstat_l2_hits);

				cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
				    KM_SLEEP);
				cb->l2rcb_hdr = hdr;
				cb->l2rcb_bp = *bp;
				cb->l2rcb_zb = *zb;
				cb->l2rcb_flags = zio_flags;

				asize = vdev_psize_to_asize(vd, size);
				if (asize != size) {
					abd = abd_alloc_for_io(asize,
					    HDR_ISTYPE_METADATA(hdr));
					cb->l2rcb_abd = abd;
				} else {
					abd = hdr_abd;
				}

				ASSERT(addr >= VDEV_LABEL_START_SIZE &&
				    addr + asize <= vd->vdev_psize -
				    VDEV_LABEL_END_SIZE);

				/*
				 * l2arc read.  The SCL_L2ARC lock will be
				 * released by l2arc_read_done().
				 * Issue a null zio if the underlying buffer
				 * was squashed to zero size by compression.
				 */
				ASSERT3U(arc_hdr_get_compress(hdr), !=,
				    ZIO_COMPRESS_EMPTY);
				rzio = zio_read_phys(pio, vd, addr,
				    asize, abd,
				    ZIO_CHECKSUM_OFF,
				    l2arc_read_done, cb, priority,
				    zio_flags | ZIO_FLAG_DONT_CACHE |
				    ZIO_FLAG_CANFAIL |
				    ZIO_FLAG_DONT_PROPAGATE |
				    ZIO_FLAG_DONT_RETRY, B_FALSE);
				acb->acb_zio_head = rzio;

				if (hash_lock != NULL)
					mutex_exit(hash_lock);

				DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
				    zio_t *, rzio);
				ARCSTAT_INCR(arcstat_l2_read_bytes,
				    HDR_GET_PSIZE(hdr));

				if (*arc_flags & ARC_FLAG_NOWAIT) {
					zio_nowait(rzio);
					return (0);
				}

				ASSERT(*arc_flags & ARC_FLAG_WAIT);
				if (zio_wait(rzio) == 0)
					return (0);

				/* l2arc read error; goto zio_read() */
				if (hash_lock != NULL)
					mutex_enter(hash_lock);
			} else {
				DTRACE_PROBE1(l2arc__miss,
				    arc_buf_hdr_t *, hdr);
				ARCSTAT_BUMP(arcstat_l2_misses);
				if (HDR_L2_WRITING(hdr))
					ARCSTAT_BUMP(arcstat_l2_rw_clash);
				spa_config_exit(spa, SCL_L2ARC, vd);
			}
		} else {
			if (vd != NULL)
				spa_config_exit(spa, SCL_L2ARC, vd);
			if (l2arc_ndev != 0) {
				DTRACE_PROBE1(l2arc__miss,
				    arc_buf_hdr_t *, hdr);
				ARCSTAT_BUMP(arcstat_l2_misses);
			}
		}

		rzio = zio_read(pio, spa, bp, hdr_abd, size,
		    arc_read_done, hdr, priority, zio_flags, zb);
		acb->acb_zio_head = rzio;

		if (hash_lock != NULL)
			mutex_exit(hash_lock);

		if (*arc_flags & ARC_FLAG_WAIT)
			return (zio_wait(rzio));

		ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
		zio_nowait(rzio);
	}
	return (rc);
}

/*
 * Notify the arc that a block was freed, and thus will never be used again.
 */
void
arc_freed(spa_t *spa, const blkptr_t *bp)
{
	arc_buf_hdr_t *hdr;
	kmutex_t *hash_lock;
	uint64_t guid = spa_load_guid(spa);

	ASSERT(!BP_IS_EMBEDDED(bp));

	hdr = buf_hash_find(guid, bp, &hash_lock);
	if (hdr == NULL)
		return;

	/*
	 * We might be trying to free a block that is still doing I/O
	 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
	 * dmu_sync-ed block). If this block is being prefetched, then it
	 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
	 * until the I/O completes. A block may also have a reference if it is
	 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
	 * have written the new block to its final resting place on disk but
	 * without the dedup flag set. This would have left the hdr in the MRU
	 * state and discoverable. When the txg finally syncs it detects that
	 * the block was overridden in open context and issues an override I/O.
	 * Since this is a dedup block, the override I/O will determine if the
	 * block is already in the DDT. If so, then it will replace the io_bp
	 * with the bp from the DDT and allow the I/O to finish. When the I/O
	 * reaches the done callback, dbuf_write_override_done, it will
	 * check to see if the io_bp and io_bp_override are identical.
	 * If they are not, then it indicates that the bp was replaced with
	 * the bp in the DDT and the override bp is freed. This allows
	 * us to arrive here with a reference on a block that is being
	 * freed. So if we have an I/O in progress, or a reference to
	 * this hdr, then we don't destroy the hdr.
	 */
	if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
	    zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
		arc_change_state(arc_anon, hdr, hash_lock);
		arc_hdr_destroy(hdr);
		mutex_exit(hash_lock);
	} else {
		mutex_exit(hash_lock);
	}

}

/*
 * Release this buffer from the cache, making it an anonymous buffer.  This
 * must be done after a read and prior to modifying the buffer contents.
 * If the buffer has more than one reference, we must make
 * a new hdr for the buffer.
 */
void
arc_release(arc_buf_t *buf, void *tag)
{
	arc_buf_hdr_t *hdr = buf->b_hdr;

	/*
	 * It would be nice to assert that if its DMU metadata (level >
	 * 0 || it's the dnode file), then it must be syncing context.
	 * But we don't know that information at this level.
	 */

	mutex_enter(&buf->b_evict_lock);

	ASSERT(HDR_HAS_L1HDR(hdr));

	/*
	 * We don't grab the hash lock prior to this check, because if
	 * the buffer's header is in the arc_anon state, it won't be
	 * linked into the hash table.
	 */
	if (hdr->b_l1hdr.b_state == arc_anon) {
		mutex_exit(&buf->b_evict_lock);
		/*
		 * If we are called from dmu_convert_mdn_block_to_raw(),
		 * a write might be in progress.  This is OK because
		 * the caller won't change the content of this buffer,
		 * only the flags (via arc_convert_to_raw()).
		 */
		/* ASSERT(!HDR_IO_IN_PROGRESS(hdr)); */
		ASSERT(!HDR_IN_HASH_TABLE(hdr));
		ASSERT(!HDR_HAS_L2HDR(hdr));
		ASSERT(HDR_EMPTY(hdr));

		ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
		ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
		ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));

		hdr->b_l1hdr.b_arc_access = 0;

		/*
		 * If the buf is being overridden then it may already
		 * have a hdr that is not empty.
		 */
		buf_discard_identity(hdr);
		arc_buf_thaw(buf);

		return;
	}

	kmutex_t *