xref: /illumos-gate/usr/src/uts/common/fs/zfs/arc.c (revision f43aa5fa)

/*
 * 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.
 * Copyright (c) 2011, 2019, Delphix. All rights reserved.
 * Copyright (c) 2020, George Amanakis. All rights reserved.
 * Copyright (c) 2020, The FreeBSD Foundation [1]
 *
 * [1] Portions of this software were developed by Allan Jude
 *     under sponsorship from the FreeBSD Foundation.
 */

/*
 * 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/arc_impl.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;

/* 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;

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_eligible_mfu",	KSTAT_DATA_UINT64 },
	{ "evict_l2_eligible_mru",	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_prefetch_asize",		KSTAT_DATA_UINT64 },
	{ "l2_mru_asize",		KSTAT_DATA_UINT64 },
	{ "l2_mfu_asize",		KSTAT_DATA_UINT64 },
	{ "l2_bufc_data_asize",		KSTAT_DATA_UINT64 },
	{ "l2_bufc_metadata_asize",	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 },
	{ "l2_log_blk_writes",		KSTAT_DATA_UINT64 },
	{ "l2_log_blk_avg_asize",	KSTAT_DATA_UINT64 },
	{ "l2_log_blk_asize",		KSTAT_DATA_UINT64 },
	{ "l2_log_blk_count",		KSTAT_DATA_UINT64 },
	{ "l2_data_to_meta_ratio",	KSTAT_DATA_UINT64 },
	{ "l2_rebuild_success",		KSTAT_DATA_UINT64 },
	{ "l2_rebuild_unsupported",	KSTAT_DATA_UINT64 },
	{ "l2_rebuild_io_errors",	KSTAT_DATA_UINT64 },
	{ "l2_rebuild_dh_errors",	KSTAT_DATA_UINT64 },
	{ "l2_rebuild_cksum_lb_errors",	KSTAT_DATA_UINT64 },
	{ "l2_rebuild_lowmem",		KSTAT_DATA_UINT64 },
	{ "l2_rebuild_size",		KSTAT_DATA_UINT64 },
	{ "l2_rebuild_asize",		KSTAT_DATA_UINT64 },
	{ "l2_rebuild_bufs",		KSTAT_DATA_UINT64 },
	{ "l2_rebuild_bufs_precached",	KSTAT_DATA_UINT64 },
	{ "l2_rebuild_log_blks",	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_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);\
		}							\
	}

/*
 * This macro allows us to use kstats as floating averages. Each time we
 * update this kstat, we first factor it and the update value by
 * ARCSTAT_AVG_FACTOR to shrink the new value's contribution to the overall
 * average. This macro assumes that integer loads and stores are atomic, but
 * is not safe for multiple writers updating the kstat in parallel (only the
 * last writer's update will remain).
 */
#define	ARCSTAT_F_AVG_FACTOR	3
#define	ARCSTAT_F_AVG(stat, value) \
	do { \
		uint64_t x = ARCSTAT(stat); \
		x = x - x / ARCSTAT_F_AVG_FACTOR + \
		    (value) / ARCSTAT_F_AVG_FACTOR; \
		ARCSTAT(stat) = x; \
		_NOTE(CONSTCOND) \
	} while (0)

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 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;

#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 */

/*
 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
 * and each of the state has two types: data and metadata.
 */
#define	L2ARC_FEED_TYPES	4


#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 */
int l2arc_meta_percent = 33;			/* limit on headers size */

/*
 * L2ARC Internals
 */
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_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 kmutex_t l2arc_rebuild_thr_lock;
static kcondvar_t l2arc_rebuild_thr_cv;

enum arc_hdr_alloc_flags {
	ARC_HDR_ALLOC_RDATA = 0x1,
	ARC_HDR_DO_ADAPT = 0x2,
};


static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *, boolean_t);
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 *, boolean_t);
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 *, int);
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 l2arc_dev_t *l2arc_vdev_get(vdev_t *vd);

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 *);
static void l2arc_do_free_on_write(void);
static void l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
    boolean_t state_only);

#define	l2arc_hdr_arcstats_increment(hdr) \
	l2arc_hdr_arcstats_update((hdr), B_TRUE, B_FALSE)
#define	l2arc_hdr_arcstats_decrement(hdr) \
	l2arc_hdr_arcstats_update((hdr), B_FALSE, B_FALSE)
#define	l2arc_hdr_arcstats_increment_state(hdr) \
	l2arc_hdr_arcstats_update((hdr), B_TRUE, B_TRUE)
#define	l2arc_hdr_arcstats_decrement_state(hdr) \
	l2arc_hdr_arcstats_update((hdr), B_FALSE, B_TRUE)

/*
 * 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_EMPTY_OR_LOCKED(hdr)				\
	(HDR_EMPTY(hdr) || MUTEX_HELD(HDR_LOCK(hdr)))

#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);
}

/*
 * l2arc_mfuonly : A ZFS module parameter that controls whether only MFU
 *		metadata and data are cached from ARC into L2ARC.
 */
int l2arc_mfuonly = 0;

/*
 * 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 || HDR_EMPTY_OR_LOCKED(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(HDR_EMPTY_OR_LOCKED(hdr));
	hdr->b_flags |= flags;
}

static inline void
arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
{
	ASSERT(HDR_EMPTY_OR_LOCKED(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(HDR_EMPTY_OR_LOCKED(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(HDR_EMPTY_OR_LOCKED(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(HDR_EMPTY_OR_LOCKED(hdr));
	ASSERT(HDR_ENCRYPTED(hdr));

	arc_hdr_alloc_pabd(hdr, ARC_HDR_DO_ADAPT);

	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, B_TRUE);
		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(HDR_EMPTY_OR_LOCKED(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);
		(void) 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 (!HDR_EMPTY(hdr) && !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 */
		if (HDR_HAS_L2HDR(hdr))
			l2arc_hdr_arcstats_decrement_state(hdr);
		arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
		if (HDR_HAS_L2HDR(hdr))
			l2arc_hdr_arcstats_increment_state(hdr);
	}
}

/*
 * 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;

		if (HDR_HAS_L2HDR(hdr) && new_state != arc_l2c_only) {
			l2arc_hdr_arcstats_decrement_state(hdr);
			hdr->b_l2hdr.b_arcs_state = new_state->arcs_state;
			l2arc_hdr_arcstats_increment_state(hdr);
		}
	}

	/*
	 * 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(HDR_EMPTY_OR_LOCKED(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);
}

/*
 * Performance tuning of L2ARC persistence:
 *
 * l2arc_rebuild_enabled : A ZFS module parameter that controls whether adding
 *		an L2ARC device (either at pool import or later) will attempt
 *		to rebuild L2ARC buffer contents.
 * l2arc_rebuild_blocks_min_l2size : A ZFS module parameter that controls
 *		whether log blocks are written to the L2ARC device. If the L2ARC
 *		device is less than 1GB, the amount of data l2arc_evict()
 *		evicts is significant compared to the amount of restored L2ARC
 *		data. In this case do not write log blocks in L2ARC in order
 *		not to waste space.
 */
int l2arc_rebuild_enabled = B_TRUE;
unsigned long l2arc_rebuild_blocks_min_l2size = 1024 * 1024 * 1024;

/* L2ARC persistence rebuild control routines. */
void l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen);
static void l2arc_dev_rebuild_start(l2arc_dev_t *dev);
static int l2arc_rebuild(l2arc_dev_t *dev);

/* L2ARC persistence read I/O routines. */
static int l2arc_dev_hdr_read(l2arc_dev_t *dev);
static int l2arc_log_blk_read(l2arc_dev_t *dev,
    const l2arc_log_blkptr_t *this_lp, const l2arc_log_blkptr_t *next_lp,
    l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
    zio_t *this_io, zio_t **next_io);
static zio_t *l2arc_log_blk_fetch(vdev_t *vd,
    const l2arc_log_blkptr_t *lp, l2arc_log_blk_phys_t *lb);
static void l2arc_log_blk_fetch_abort(zio_t *zio);

/* L2ARC persistence block restoration routines. */
static void l2arc_log_blk_restore(l2arc_dev_t *dev,
    const l2arc_log_blk_phys_t *lb, uint64_t lb_asize);
static void l2arc_hdr_restore(const l2arc_log_ent_phys_t *le,
    l2arc_dev_t *dev);

/* L2ARC persistence write I/O routines. */
static void l2arc_dev_hdr_update(l2arc_dev_t *dev);
static void l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio,
    l2arc_write_callback_t *cb);

/* L2ARC persistence auxilliary routines. */
boolean_t l2arc_log_blkptr_valid(l2arc_dev_t *dev,
    const l2arc_log_blkptr_t *lbp);
static boolean_t l2arc_log_blk_insert(l2arc_dev_t *dev,
    const arc_buf_hdr_t *ab);
boolean_t l2arc_range_check_overlap(uint64_t bottom,
    uint64_t top, uint64_t check);
static void l2arc_blk_fetch_done(zio_t *zio);
static inline uint64_t
    l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev);

/*
 * 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(HDR_EMPTY_OR_LOCKED(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(HDR_EMPTY_OR_LOCKED(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(HDR_EMPTY_OR_LOCKED(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(HDR_EMPTY_OR_LOCKED(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, int alloc_flags)
{
	uint64_t size;
	boolean_t alloc_rdata = ((alloc_flags & ARC_HDR_ALLOC_RDATA) != 0);
	boolean_t do_adapt = ((alloc_flags & ARC_HDR_DO_ADAPT) != 0);

	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,
		    do_adapt);
		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,
		    do_adapt);
		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;
	int flags = ARC_HDR_DO_ADAPT;

	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);
	}
	flags |= alloc_rdata ? ARC_HDR_ALLOC_RDATA : 0;
	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, flags);
	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,