1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2018, Joyent, Inc.
24 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2017 Nexenta Systems, Inc.  All rights reserved.
27 */
28
29/*
30 * DVA-based Adjustable Replacement Cache
31 *
32 * While much of the theory of operation used here is
33 * based on the self-tuning, low overhead replacement cache
34 * presented by Megiddo and Modha at FAST 2003, there are some
35 * significant differences:
36 *
37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 * Pages in its cache cannot be "locked" into memory.  This makes
39 * the eviction algorithm simple: evict the last page in the list.
40 * This also make the performance characteristics easy to reason
41 * about.  Our cache is not so simple.  At any given moment, some
42 * subset of the blocks in the cache are un-evictable because we
43 * have handed out a reference to them.  Blocks are only evictable
44 * when there are no external references active.  This makes
45 * eviction far more problematic:  we choose to evict the evictable
46 * blocks that are the "lowest" in the list.
47 *
48 * There are times when it is not possible to evict the requested
49 * space.  In these circumstances we are unable to adjust the cache
50 * size.  To prevent the cache growing unbounded at these times we
51 * implement a "cache throttle" that slows the flow of new data
52 * into the cache until we can make space available.
53 *
54 * 2. The Megiddo and Modha model assumes a fixed cache size.
55 * Pages are evicted when the cache is full and there is a cache
56 * miss.  Our model has a variable sized cache.  It grows with
57 * high use, but also tries to react to memory pressure from the
58 * operating system: decreasing its size when system memory is
59 * tight.
60 *
61 * 3. The Megiddo and Modha model assumes a fixed page size. All
62 * elements of the cache are therefore exactly the same size.  So
63 * when adjusting the cache size following a cache miss, its simply
64 * a matter of choosing a single page to evict.  In our model, we
65 * have variable sized cache blocks (rangeing from 512 bytes to
66 * 128K bytes).  We therefore choose a set of blocks to evict to make
67 * space for a cache miss that approximates as closely as possible
68 * the space used by the new block.
69 *
70 * See also:  "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71 * by N. Megiddo & D. Modha, FAST 2003
72 */
73
74/*
75 * The locking model:
76 *
77 * A new reference to a cache buffer can be obtained in two
78 * ways: 1) via a hash table lookup using the DVA as a key,
79 * or 2) via one of the ARC lists.  The arc_read() interface
80 * uses method 1, while the internal ARC algorithms for
81 * adjusting the cache use method 2.  We therefore provide two
82 * types of locks: 1) the hash table lock array, and 2) the
83 * ARC list locks.
84 *
85 * Buffers do not have their own mutexes, rather they rely on the
86 * hash table mutexes for the bulk of their protection (i.e. most
87 * fields in the arc_buf_hdr_t are protected by these mutexes).
88 *
89 * buf_hash_find() returns the appropriate mutex (held) when it
90 * locates the requested buffer in the hash table.  It returns
91 * NULL for the mutex if the buffer was not in the table.
92 *
93 * buf_hash_remove() expects the appropriate hash mutex to be
94 * already held before it is invoked.
95 *
96 * Each ARC state also has a mutex which is used to protect the
97 * buffer list associated with the state.  When attempting to
98 * obtain a hash table lock while holding an ARC list lock you
99 * must use: mutex_tryenter() to avoid deadlock.  Also note that
100 * the active state mutex must be held before the ghost state mutex.
101 *
102 * Note that the majority of the performance stats are manipulated
103 * with atomic operations.
104 *
105 * The L2ARC uses the l2ad_mtx on each vdev for the following:
106 *
107 *	- L2ARC buflist creation
108 *	- L2ARC buflist eviction
109 *	- L2ARC write completion, which walks L2ARC buflists
110 *	- ARC header destruction, as it removes from L2ARC buflists
111 *	- ARC header release, as it removes from L2ARC buflists
112 */
113
114/*
115 * ARC operation:
116 *
117 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
118 * This structure can point either to a block that is still in the cache or to
119 * one that is only accessible in an L2 ARC device, or it can provide
120 * information about a block that was recently evicted. If a block is
121 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
122 * information to retrieve it from the L2ARC device. This information is
123 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
124 * that is in this state cannot access the data directly.
125 *
126 * Blocks that are actively being referenced or have not been evicted
127 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
128 * the arc_buf_hdr_t that will point to the data block in memory. A block can
129 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
130 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
131 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
132 *
133 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
134 * ability to store the physical data (b_pabd) associated with the DVA of the
135 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
136 * it will match its on-disk compression characteristics. This behavior can be
137 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
138 * compressed ARC functionality is disabled, the b_pabd will point to an
139 * uncompressed version of the on-disk data.
140 *
141 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
142 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
143 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
144 * consumer. The ARC will provide references to this data and will keep it
145 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
146 * data block and will evict any arc_buf_t that is no longer referenced. The
147 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
148 * "overhead_size" kstat.
149 *
150 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
151 * compressed form. The typical case is that consumers will want uncompressed
152 * data, and when that happens a new data buffer is allocated where the data is
153 * decompressed for them to use. Currently the only consumer who wants
154 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
155 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
156 * with the arc_buf_hdr_t.
157 *
158 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
159 * first one is owned by a compressed send consumer (and therefore references
160 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
161 * used by any other consumer (and has its own uncompressed copy of the data
162 * buffer).
163 *
164 *   arc_buf_hdr_t
165 *   +-----------+
166 *   | fields    |
167 *   | common to |
168 *   | L1- and   |
169 *   | L2ARC     |
170 *   +-----------+
171 *   | l2arc_buf_hdr_t
172 *   |           |
173 *   +-----------+
174 *   | l1arc_buf_hdr_t
175 *   |           |              arc_buf_t
176 *   | b_buf     +------------>+-----------+      arc_buf_t
177 *   | b_pabd    +-+           |b_next     +---->+-----------+
178 *   +-----------+ |           |-----------|     |b_next     +-->NULL
179 *                 |           |b_comp = T |     +-----------+
180 *                 |           |b_data     +-+   |b_comp = F |
181 *                 |           +-----------+ |   |b_data     +-+
182 *                 +->+------+               |   +-----------+ |
183 *        compressed  |      |               |                 |
184 *           data     |      |<--------------+                 | uncompressed
185 *                    +------+          compressed,            |     data
186 *                                        shared               +-->+------+
187 *                                         data                    |      |
188 *                                                                 |      |
189 *                                                                 +------+
190 *
191 * When a consumer reads a block, the ARC must first look to see if the
192 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
193 * arc_buf_t and either copies uncompressed data into a new data buffer from an
194 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
195 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
196 * hdr is compressed and the desired compression characteristics of the
197 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
198 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
199 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
200 * be anywhere in the hdr's list.
201 *
202 * The diagram below shows an example of an uncompressed ARC hdr that is
203 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
204 * the last element in the buf list):
205 *
206 *                arc_buf_hdr_t
207 *                +-----------+
208 *                |           |
209 *                |           |
210 *                |           |
211 *                +-----------+
212 * l2arc_buf_hdr_t|           |
213 *                |           |
214 *                +-----------+
215 * l1arc_buf_hdr_t|           |
216 *                |           |                 arc_buf_t    (shared)
217 *                |    b_buf  +------------>+---------+      arc_buf_t
218 *                |           |             |b_next   +---->+---------+
219 *                |  b_pabd   +-+           |---------|     |b_next   +-->NULL
220 *                +-----------+ |           |         |     +---------+
221 *                              |           |b_data   +-+   |         |
222 *                              |           +---------+ |   |b_data   +-+
223 *                              +->+------+             |   +---------+ |
224 *                                 |      |             |               |
225 *                   uncompressed  |      |             |               |
226 *                        data     +------+             |               |
227 *                                    ^                 +->+------+     |
228 *                                    |       uncompressed |      |     |
229 *                                    |           data     |      |     |
230 *                                    |                    +------+     |
231 *                                    +---------------------------------+
232 *
233 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
234 * since the physical block is about to be rewritten. The new data contents
235 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
236 * it may compress the data before writing it to disk. The ARC will be called
237 * with the transformed data and will bcopy the transformed on-disk block into
238 * a newly allocated b_pabd. Writes are always done into buffers which have
239 * either been loaned (and hence are new and don't have other readers) or
240 * buffers which have been released (and hence have their own hdr, if there
241 * were originally other readers of the buf's original hdr). This ensures that
242 * the ARC only needs to update a single buf and its hdr after a write occurs.
243 *
244 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
245 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
246 * that when compressed ARC is enabled that the L2ARC blocks are identical
247 * to the on-disk block in the main data pool. This provides a significant
248 * advantage since the ARC can leverage the bp's checksum when reading from the
249 * L2ARC to determine if the contents are valid. However, if the compressed
250 * ARC is disabled, then the L2ARC's block must be transformed to look
251 * like the physical block in the main data pool before comparing the
252 * checksum and determining its validity.
253 */
254
255#include <sys/spa.h>
256#include <sys/zio.h>
257#include <sys/spa_impl.h>
258#include <sys/zio_compress.h>
259#include <sys/zio_checksum.h>
260#include <sys/zfs_context.h>
261#include <sys/arc.h>
262#include <sys/refcount.h>
263#include <sys/vdev.h>
264#include <sys/vdev_impl.h>
265#include <sys/dsl_pool.h>
266#include <sys/zio_checksum.h>
267#include <sys/multilist.h>
268#include <sys/abd.h>
269#ifdef _KERNEL
270#include <sys/dnlc.h>
271#include <sys/racct.h>
272#endif
273#include <sys/callb.h>
274#include <sys/kstat.h>
275#include <sys/trim_map.h>
276#include <zfs_fletcher.h>
277#include <sys/sdt.h>
278#include <sys/aggsum.h>
279#include <sys/cityhash.h>
280
281#include <machine/vmparam.h>
282
283#ifdef illumos
284#ifndef _KERNEL
285/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
286boolean_t arc_watch = B_FALSE;
287int arc_procfd;
288#endif
289#endif /* illumos */
290
291static kmutex_t		arc_reclaim_lock;
292static kcondvar_t	arc_reclaim_thread_cv;
293static boolean_t	arc_reclaim_thread_exit;
294static kcondvar_t	arc_reclaim_waiters_cv;
295
296static kmutex_t		arc_dnlc_evicts_lock;
297static kcondvar_t	arc_dnlc_evicts_cv;
298static boolean_t	arc_dnlc_evicts_thread_exit;
299
300uint_t arc_reduce_dnlc_percent = 3;
301
302/*
303 * The number of headers to evict in arc_evict_state_impl() before
304 * dropping the sublist lock and evicting from another sublist. A lower
305 * value means we're more likely to evict the "correct" header (i.e. the
306 * oldest header in the arc state), but comes with higher overhead
307 * (i.e. more invocations of arc_evict_state_impl()).
308 */
309int zfs_arc_evict_batch_limit = 10;
310
311/* number of seconds before growing cache again */
312static int		arc_grow_retry = 60;
313
314/* number of milliseconds before attempting a kmem-cache-reap */
315static int		arc_kmem_cache_reap_retry_ms = 1000;
316
317/* shift of arc_c for calculating overflow limit in arc_get_data_impl */
318int		zfs_arc_overflow_shift = 8;
319
320/* shift of arc_c for calculating both min and max arc_p */
321static int		arc_p_min_shift = 4;
322
323/* log2(fraction of arc to reclaim) */
324static int		arc_shrink_shift = 7;
325
326/*
327 * log2(fraction of ARC which must be free to allow growing).
328 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
329 * when reading a new block into the ARC, we will evict an equal-sized block
330 * from the ARC.
331 *
332 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
333 * we will still not allow it to grow.
334 */
335int			arc_no_grow_shift = 5;
336
337
338/*
339 * minimum lifespan of a prefetch block in clock ticks
340 * (initialized in arc_init())
341 */
342static int		arc_min_prefetch_lifespan;
343
344/*
345 * If this percent of memory is free, don't throttle.
346 */
347int arc_lotsfree_percent = 10;
348
349static int arc_dead;
350extern boolean_t zfs_prefetch_disable;
351
352/*
353 * The arc has filled available memory and has now warmed up.
354 */
355static boolean_t arc_warm;
356
357/*
358 * log2 fraction of the zio arena to keep free.
359 */
360int arc_zio_arena_free_shift = 2;
361
362/*
363 * These tunables are for performance analysis.
364 */
365uint64_t zfs_arc_max;
366uint64_t zfs_arc_min;
367uint64_t zfs_arc_meta_limit = 0;
368uint64_t zfs_arc_meta_min = 0;
369int zfs_arc_grow_retry = 0;
370int zfs_arc_shrink_shift = 0;
371int zfs_arc_no_grow_shift = 0;
372int zfs_arc_p_min_shift = 0;
373uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
374u_int zfs_arc_free_target = 0;
375
376/* Absolute min for arc min / max is 16MB. */
377static uint64_t arc_abs_min = 16 << 20;
378
379boolean_t zfs_compressed_arc_enabled = B_TRUE;
380
381static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
382static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
383static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS);
384static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS);
385static int sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS);
386
387#if defined(__FreeBSD__) && defined(_KERNEL)
388static void
389arc_free_target_init(void *unused __unused)
390{
391
392	zfs_arc_free_target = vm_cnt.v_free_target;
393}
394SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
395    arc_free_target_init, NULL);
396
397TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
398TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
399TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
400TUNABLE_INT("vfs.zfs.arc_grow_retry", &zfs_arc_grow_retry);
401TUNABLE_INT("vfs.zfs.arc_no_grow_shift", &zfs_arc_no_grow_shift);
402SYSCTL_DECL(_vfs_zfs);
403SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN,
404    0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size");
405SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN,
406    0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size");
407SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_no_grow_shift, CTLTYPE_U32 | CTLFLAG_RWTUN,
408    0, sizeof(uint32_t), sysctl_vfs_zfs_arc_no_grow_shift, "U",
409    "log2(fraction of ARC which must be free to allow growing)");
410SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
411    &zfs_arc_average_blocksize, 0,
412    "ARC average blocksize");
413SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
414    &arc_shrink_shift, 0,
415    "log2(fraction of arc to reclaim)");
416SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_grow_retry, CTLFLAG_RW,
417    &arc_grow_retry, 0,
418    "Wait in seconds before considering growing ARC");
419SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN,
420    &zfs_compressed_arc_enabled, 0, "Enable compressed ARC");
421
422/*
423 * We don't have a tunable for arc_free_target due to the dependency on
424 * pagedaemon initialisation.
425 */
426SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
427    CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
428    sysctl_vfs_zfs_arc_free_target, "IU",
429    "Desired number of free pages below which ARC triggers reclaim");
430
431static int
432sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
433{
434	u_int val;
435	int err;
436
437	val = zfs_arc_free_target;
438	err = sysctl_handle_int(oidp, &val, 0, req);
439	if (err != 0 || req->newptr == NULL)
440		return (err);
441
442	if (val < minfree)
443		return (EINVAL);
444	if (val > vm_cnt.v_page_count)
445		return (EINVAL);
446
447	zfs_arc_free_target = val;
448
449	return (0);
450}
451
452/*
453 * Must be declared here, before the definition of corresponding kstat
454 * macro which uses the same names will confuse the compiler.
455 */
456SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
457    CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
458    sysctl_vfs_zfs_arc_meta_limit, "QU",
459    "ARC metadata limit");
460#endif
461
462/*
463 * Note that buffers can be in one of 6 states:
464 *	ARC_anon	- anonymous (discussed below)
465 *	ARC_mru		- recently used, currently cached
466 *	ARC_mru_ghost	- recentely used, no longer in cache
467 *	ARC_mfu		- frequently used, currently cached
468 *	ARC_mfu_ghost	- frequently used, no longer in cache
469 *	ARC_l2c_only	- exists in L2ARC but not other states
470 * When there are no active references to the buffer, they are
471 * are linked onto a list in one of these arc states.  These are
472 * the only buffers that can be evicted or deleted.  Within each
473 * state there are multiple lists, one for meta-data and one for
474 * non-meta-data.  Meta-data (indirect blocks, blocks of dnodes,
475 * etc.) is tracked separately so that it can be managed more
476 * explicitly: favored over data, limited explicitly.
477 *
478 * Anonymous buffers are buffers that are not associated with
479 * a DVA.  These are buffers that hold dirty block copies
480 * before they are written to stable storage.  By definition,
481 * they are "ref'd" and are considered part of arc_mru
482 * that cannot be freed.  Generally, they will aquire a DVA
483 * as they are written and migrate onto the arc_mru list.
484 *
485 * The ARC_l2c_only state is for buffers that are in the second
486 * level ARC but no longer in any of the ARC_m* lists.  The second
487 * level ARC itself may also contain buffers that are in any of
488 * the ARC_m* states - meaning that a buffer can exist in two
489 * places.  The reason for the ARC_l2c_only state is to keep the
490 * buffer header in the hash table, so that reads that hit the
491 * second level ARC benefit from these fast lookups.
492 */
493
494typedef struct arc_state {
495	/*
496	 * list of evictable buffers
497	 */
498	multilist_t *arcs_list[ARC_BUFC_NUMTYPES];
499	/*
500	 * total amount of evictable data in this state
501	 */
502	refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
503	/*
504	 * total amount of data in this state; this includes: evictable,
505	 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
506	 */
507	refcount_t arcs_size;
508} arc_state_t;
509
510/* The 6 states: */
511static arc_state_t ARC_anon;
512static arc_state_t ARC_mru;
513static arc_state_t ARC_mru_ghost;
514static arc_state_t ARC_mfu;
515static arc_state_t ARC_mfu_ghost;
516static arc_state_t ARC_l2c_only;
517
518typedef struct arc_stats {
519	kstat_named_t arcstat_hits;
520	kstat_named_t arcstat_misses;
521	kstat_named_t arcstat_demand_data_hits;
522	kstat_named_t arcstat_demand_data_misses;
523	kstat_named_t arcstat_demand_metadata_hits;
524	kstat_named_t arcstat_demand_metadata_misses;
525	kstat_named_t arcstat_prefetch_data_hits;
526	kstat_named_t arcstat_prefetch_data_misses;
527	kstat_named_t arcstat_prefetch_metadata_hits;
528	kstat_named_t arcstat_prefetch_metadata_misses;
529	kstat_named_t arcstat_mru_hits;
530	kstat_named_t arcstat_mru_ghost_hits;
531	kstat_named_t arcstat_mfu_hits;
532	kstat_named_t arcstat_mfu_ghost_hits;
533	kstat_named_t arcstat_allocated;
534	kstat_named_t arcstat_deleted;
535	/*
536	 * Number of buffers that could not be evicted because the hash lock
537	 * was held by another thread.  The lock may not necessarily be held
538	 * by something using the same buffer, since hash locks are shared
539	 * by multiple buffers.
540	 */
541	kstat_named_t arcstat_mutex_miss;
542	/*
543	 * Number of buffers skipped when updating the access state due to the
544	 * header having already been released after acquiring the hash lock.
545	 */
546	kstat_named_t arcstat_access_skip;
547	/*
548	 * Number of buffers skipped because they have I/O in progress, are
549	 * indirect prefetch buffers that have not lived long enough, or are
550	 * not from the spa we're trying to evict from.
551	 */
552	kstat_named_t arcstat_evict_skip;
553	/*
554	 * Number of times arc_evict_state() was unable to evict enough
555	 * buffers to reach it's target amount.
556	 */
557	kstat_named_t arcstat_evict_not_enough;
558	kstat_named_t arcstat_evict_l2_cached;
559	kstat_named_t arcstat_evict_l2_eligible;
560	kstat_named_t arcstat_evict_l2_ineligible;
561	kstat_named_t arcstat_evict_l2_skip;
562	kstat_named_t arcstat_hash_elements;
563	kstat_named_t arcstat_hash_elements_max;
564	kstat_named_t arcstat_hash_collisions;
565	kstat_named_t arcstat_hash_chains;
566	kstat_named_t arcstat_hash_chain_max;
567	kstat_named_t arcstat_p;
568	kstat_named_t arcstat_c;
569	kstat_named_t arcstat_c_min;
570	kstat_named_t arcstat_c_max;
571	/* Not updated directly; only synced in arc_kstat_update. */
572	kstat_named_t arcstat_size;
573	/*
574	 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
575	 * Note that the compressed bytes may match the uncompressed bytes
576	 * if the block is either not compressed or compressed arc is disabled.
577	 */
578	kstat_named_t arcstat_compressed_size;
579	/*
580	 * Uncompressed size of the data stored in b_pabd. If compressed
581	 * arc is disabled then this value will be identical to the stat
582	 * above.
583	 */
584	kstat_named_t arcstat_uncompressed_size;
585	/*
586	 * Number of bytes stored in all the arc_buf_t's. This is classified
587	 * as "overhead" since this data is typically short-lived and will
588	 * be evicted from the arc when it becomes unreferenced unless the
589	 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
590	 * values have been set (see comment in dbuf.c for more information).
591	 */
592	kstat_named_t arcstat_overhead_size;
593	/*
594	 * Number of bytes consumed by internal ARC structures necessary
595	 * for tracking purposes; these structures are not actually
596	 * backed by ARC buffers. This includes arc_buf_hdr_t structures
597	 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
598	 * caches), and arc_buf_t structures (allocated via arc_buf_t
599	 * cache).
600	 * Not updated directly; only synced in arc_kstat_update.
601	 */
602	kstat_named_t arcstat_hdr_size;
603	/*
604	 * Number of bytes consumed by ARC buffers of type equal to
605	 * ARC_BUFC_DATA. This is generally consumed by buffers backing
606	 * on disk user data (e.g. plain file contents).
607	 * Not updated directly; only synced in arc_kstat_update.
608	 */
609	kstat_named_t arcstat_data_size;
610	/*
611	 * Number of bytes consumed by ARC buffers of type equal to
612	 * ARC_BUFC_METADATA. This is generally consumed by buffers
613	 * backing on disk data that is used for internal ZFS
614	 * structures (e.g. ZAP, dnode, indirect blocks, etc).
615	 * Not updated directly; only synced in arc_kstat_update.
616	 */
617	kstat_named_t arcstat_metadata_size;
618	/*
619	 * Number of bytes consumed by various buffers and structures
620	 * not actually backed with ARC buffers. This includes bonus
621	 * buffers (allocated directly via zio_buf_* functions),
622	 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
623	 * cache), and dnode_t structures (allocated via dnode_t cache).
624	 * Not updated directly; only synced in arc_kstat_update.
625	 */
626	kstat_named_t arcstat_other_size;
627	/*
628	 * Total number of bytes consumed by ARC buffers residing in the
629	 * arc_anon state. This includes *all* buffers in the arc_anon
630	 * state; e.g. data, metadata, evictable, and unevictable buffers
631	 * are all included in this value.
632	 * Not updated directly; only synced in arc_kstat_update.
633	 */
634	kstat_named_t arcstat_anon_size;
635	/*
636	 * Number of bytes consumed by ARC buffers that meet the
637	 * following criteria: backing buffers of type ARC_BUFC_DATA,
638	 * residing in the arc_anon state, and are eligible for eviction
639	 * (e.g. have no outstanding holds on the buffer).
640	 * Not updated directly; only synced in arc_kstat_update.
641	 */
642	kstat_named_t arcstat_anon_evictable_data;
643	/*
644	 * Number of bytes consumed by ARC buffers that meet the
645	 * following criteria: backing buffers of type ARC_BUFC_METADATA,
646	 * residing in the arc_anon state, and are eligible for eviction
647	 * (e.g. have no outstanding holds on the buffer).
648	 * Not updated directly; only synced in arc_kstat_update.
649	 */
650	kstat_named_t arcstat_anon_evictable_metadata;
651	/*
652	 * Total number of bytes consumed by ARC buffers residing in the
653	 * arc_mru state. This includes *all* buffers in the arc_mru
654	 * state; e.g. data, metadata, evictable, and unevictable buffers
655	 * are all included in this value.
656	 * Not updated directly; only synced in arc_kstat_update.
657	 */
658	kstat_named_t arcstat_mru_size;
659	/*
660	 * Number of bytes consumed by ARC buffers that meet the
661	 * following criteria: backing buffers of type ARC_BUFC_DATA,
662	 * residing in the arc_mru state, and are eligible for eviction
663	 * (e.g. have no outstanding holds on the buffer).
664	 * Not updated directly; only synced in arc_kstat_update.
665	 */
666	kstat_named_t arcstat_mru_evictable_data;
667	/*
668	 * Number of bytes consumed by ARC buffers that meet the
669	 * following criteria: backing buffers of type ARC_BUFC_METADATA,
670	 * residing in the arc_mru state, and are eligible for eviction
671	 * (e.g. have no outstanding holds on the buffer).
672	 * Not updated directly; only synced in arc_kstat_update.
673	 */
674	kstat_named_t arcstat_mru_evictable_metadata;
675	/*
676	 * Total number of bytes that *would have been* consumed by ARC
677	 * buffers in the arc_mru_ghost state. The key thing to note
678	 * here, is the fact that this size doesn't actually indicate
679	 * RAM consumption. The ghost lists only consist of headers and
680	 * don't actually have ARC buffers linked off of these headers.
681	 * Thus, *if* the headers had associated ARC buffers, these
682	 * buffers *would have* consumed this number of bytes.
683	 * Not updated directly; only synced in arc_kstat_update.
684	 */
685	kstat_named_t arcstat_mru_ghost_size;
686	/*
687	 * Number of bytes that *would have been* consumed by ARC
688	 * buffers that are eligible for eviction, of type
689	 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
690	 * Not updated directly; only synced in arc_kstat_update.
691	 */
692	kstat_named_t arcstat_mru_ghost_evictable_data;
693	/*
694	 * Number of bytes that *would have been* consumed by ARC
695	 * buffers that are eligible for eviction, of type
696	 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
697	 * Not updated directly; only synced in arc_kstat_update.
698	 */
699	kstat_named_t arcstat_mru_ghost_evictable_metadata;
700	/*
701	 * Total number of bytes consumed by ARC buffers residing in the
702	 * arc_mfu state. This includes *all* buffers in the arc_mfu
703	 * state; e.g. data, metadata, evictable, and unevictable buffers
704	 * are all included in this value.
705	 * Not updated directly; only synced in arc_kstat_update.
706	 */
707	kstat_named_t arcstat_mfu_size;
708	/*
709	 * Number of bytes consumed by ARC buffers that are eligible for
710	 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
711	 * state.
712	 * Not updated directly; only synced in arc_kstat_update.
713	 */
714	kstat_named_t arcstat_mfu_evictable_data;
715	/*
716	 * Number of bytes consumed by ARC buffers that are eligible for
717	 * eviction, of type ARC_BUFC_METADATA, and reside in the
718	 * arc_mfu state.
719	 * Not updated directly; only synced in arc_kstat_update.
720	 */
721	kstat_named_t arcstat_mfu_evictable_metadata;
722	/*
723	 * Total number of bytes that *would have been* consumed by ARC
724	 * buffers in the arc_mfu_ghost state. See the comment above
725	 * arcstat_mru_ghost_size for more details.
726	 * Not updated directly; only synced in arc_kstat_update.
727	 */
728	kstat_named_t arcstat_mfu_ghost_size;
729	/*
730	 * Number of bytes that *would have been* consumed by ARC
731	 * buffers that are eligible for eviction, of type
732	 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
733	 * Not updated directly; only synced in arc_kstat_update.
734	 */
735	kstat_named_t arcstat_mfu_ghost_evictable_data;
736	/*
737	 * Number of bytes that *would have been* consumed by ARC
738	 * buffers that are eligible for eviction, of type
739	 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
740	 * Not updated directly; only synced in arc_kstat_update.
741	 */
742	kstat_named_t arcstat_mfu_ghost_evictable_metadata;
743	kstat_named_t arcstat_l2_hits;
744	kstat_named_t arcstat_l2_misses;
745	kstat_named_t arcstat_l2_feeds;
746	kstat_named_t arcstat_l2_rw_clash;
747	kstat_named_t arcstat_l2_read_bytes;
748	kstat_named_t arcstat_l2_write_bytes;
749	kstat_named_t arcstat_l2_writes_sent;
750	kstat_named_t arcstat_l2_writes_done;
751	kstat_named_t arcstat_l2_writes_error;
752	kstat_named_t arcstat_l2_writes_lock_retry;
753	kstat_named_t arcstat_l2_evict_lock_retry;
754	kstat_named_t arcstat_l2_evict_reading;
755	kstat_named_t arcstat_l2_evict_l1cached;
756	kstat_named_t arcstat_l2_free_on_write;
757	kstat_named_t arcstat_l2_abort_lowmem;
758	kstat_named_t arcstat_l2_cksum_bad;
759	kstat_named_t arcstat_l2_io_error;
760	kstat_named_t arcstat_l2_lsize;
761	kstat_named_t arcstat_l2_psize;
762	/* Not updated directly; only synced in arc_kstat_update. */
763	kstat_named_t arcstat_l2_hdr_size;
764	kstat_named_t arcstat_l2_write_trylock_fail;
765	kstat_named_t arcstat_l2_write_passed_headroom;
766	kstat_named_t arcstat_l2_write_spa_mismatch;
767	kstat_named_t arcstat_l2_write_in_l2;
768	kstat_named_t arcstat_l2_write_hdr_io_in_progress;
769	kstat_named_t arcstat_l2_write_not_cacheable;
770	kstat_named_t arcstat_l2_write_full;
771	kstat_named_t arcstat_l2_write_buffer_iter;
772	kstat_named_t arcstat_l2_write_pios;
773	kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
774	kstat_named_t arcstat_l2_write_buffer_list_iter;
775	kstat_named_t arcstat_l2_write_buffer_list_null_iter;
776	kstat_named_t arcstat_memory_throttle_count;
777	/* Not updated directly; only synced in arc_kstat_update. */
778	kstat_named_t arcstat_meta_used;
779	kstat_named_t arcstat_meta_limit;
780	kstat_named_t arcstat_meta_max;
781	kstat_named_t arcstat_meta_min;
782	kstat_named_t arcstat_sync_wait_for_async;
783	kstat_named_t arcstat_demand_hit_predictive_prefetch;
784} arc_stats_t;
785
786static arc_stats_t arc_stats = {
787	{ "hits",			KSTAT_DATA_UINT64 },
788	{ "misses",			KSTAT_DATA_UINT64 },
789	{ "demand_data_hits",		KSTAT_DATA_UINT64 },
790	{ "demand_data_misses",		KSTAT_DATA_UINT64 },
791	{ "demand_metadata_hits",	KSTAT_DATA_UINT64 },
792	{ "demand_metadata_misses",	KSTAT_DATA_UINT64 },
793	{ "prefetch_data_hits",		KSTAT_DATA_UINT64 },
794	{ "prefetch_data_misses",	KSTAT_DATA_UINT64 },
795	{ "prefetch_metadata_hits",	KSTAT_DATA_UINT64 },
796	{ "prefetch_metadata_misses",	KSTAT_DATA_UINT64 },
797	{ "mru_hits",			KSTAT_DATA_UINT64 },
798	{ "mru_ghost_hits",		KSTAT_DATA_UINT64 },
799	{ "mfu_hits",			KSTAT_DATA_UINT64 },
800	{ "mfu_ghost_hits",		KSTAT_DATA_UINT64 },
801	{ "allocated",			KSTAT_DATA_UINT64 },
802	{ "deleted",			KSTAT_DATA_UINT64 },
803	{ "mutex_miss",			KSTAT_DATA_UINT64 },
804	{ "access_skip",		KSTAT_DATA_UINT64 },
805	{ "evict_skip",			KSTAT_DATA_UINT64 },
806	{ "evict_not_enough",		KSTAT_DATA_UINT64 },
807	{ "evict_l2_cached",		KSTAT_DATA_UINT64 },
808	{ "evict_l2_eligible",		KSTAT_DATA_UINT64 },
809	{ "evict_l2_ineligible",	KSTAT_DATA_UINT64 },
810	{ "evict_l2_skip",		KSTAT_DATA_UINT64 },
811	{ "hash_elements",		KSTAT_DATA_UINT64 },
812	{ "hash_elements_max",		KSTAT_DATA_UINT64 },
813	{ "hash_collisions",		KSTAT_DATA_UINT64 },
814	{ "hash_chains",		KSTAT_DATA_UINT64 },
815	{ "hash_chain_max",		KSTAT_DATA_UINT64 },
816	{ "p",				KSTAT_DATA_UINT64 },
817	{ "c",				KSTAT_DATA_UINT64 },
818	{ "c_min",			KSTAT_DATA_UINT64 },
819	{ "c_max",			KSTAT_DATA_UINT64 },
820	{ "size",			KSTAT_DATA_UINT64 },
821	{ "compressed_size",		KSTAT_DATA_UINT64 },
822	{ "uncompressed_size",		KSTAT_DATA_UINT64 },
823	{ "overhead_size",		KSTAT_DATA_UINT64 },
824	{ "hdr_size",			KSTAT_DATA_UINT64 },
825	{ "data_size",			KSTAT_DATA_UINT64 },
826	{ "metadata_size",		KSTAT_DATA_UINT64 },
827	{ "other_size",			KSTAT_DATA_UINT64 },
828	{ "anon_size",			KSTAT_DATA_UINT64 },
829	{ "anon_evictable_data",	KSTAT_DATA_UINT64 },
830	{ "anon_evictable_metadata",	KSTAT_DATA_UINT64 },
831	{ "mru_size",			KSTAT_DATA_UINT64 },
832	{ "mru_evictable_data",		KSTAT_DATA_UINT64 },
833	{ "mru_evictable_metadata",	KSTAT_DATA_UINT64 },
834	{ "mru_ghost_size",		KSTAT_DATA_UINT64 },
835	{ "mru_ghost_evictable_data",	KSTAT_DATA_UINT64 },
836	{ "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
837	{ "mfu_size",			KSTAT_DATA_UINT64 },
838	{ "mfu_evictable_data",		KSTAT_DATA_UINT64 },
839	{ "mfu_evictable_metadata",	KSTAT_DATA_UINT64 },
840	{ "mfu_ghost_size",		KSTAT_DATA_UINT64 },
841	{ "mfu_ghost_evictable_data",	KSTAT_DATA_UINT64 },
842	{ "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
843	{ "l2_hits",			KSTAT_DATA_UINT64 },
844	{ "l2_misses",			KSTAT_DATA_UINT64 },
845	{ "l2_feeds",			KSTAT_DATA_UINT64 },
846	{ "l2_rw_clash",		KSTAT_DATA_UINT64 },
847	{ "l2_read_bytes",		KSTAT_DATA_UINT64 },
848	{ "l2_write_bytes",		KSTAT_DATA_UINT64 },
849	{ "l2_writes_sent",		KSTAT_DATA_UINT64 },
850	{ "l2_writes_done",		KSTAT_DATA_UINT64 },
851	{ "l2_writes_error",		KSTAT_DATA_UINT64 },
852	{ "l2_writes_lock_retry",	KSTAT_DATA_UINT64 },
853	{ "l2_evict_lock_retry",	KSTAT_DATA_UINT64 },
854	{ "l2_evict_reading",		KSTAT_DATA_UINT64 },
855	{ "l2_evict_l1cached",		KSTAT_DATA_UINT64 },
856	{ "l2_free_on_write",		KSTAT_DATA_UINT64 },
857	{ "l2_abort_lowmem",		KSTAT_DATA_UINT64 },
858	{ "l2_cksum_bad",		KSTAT_DATA_UINT64 },
859	{ "l2_io_error",		KSTAT_DATA_UINT64 },
860	{ "l2_size",			KSTAT_DATA_UINT64 },
861	{ "l2_asize",			KSTAT_DATA_UINT64 },
862	{ "l2_hdr_size",		KSTAT_DATA_UINT64 },
863	{ "l2_write_trylock_fail",	KSTAT_DATA_UINT64 },
864	{ "l2_write_passed_headroom",	KSTAT_DATA_UINT64 },
865	{ "l2_write_spa_mismatch",	KSTAT_DATA_UINT64 },
866	{ "l2_write_in_l2",		KSTAT_DATA_UINT64 },
867	{ "l2_write_io_in_progress",	KSTAT_DATA_UINT64 },
868	{ "l2_write_not_cacheable",	KSTAT_DATA_UINT64 },
869	{ "l2_write_full",		KSTAT_DATA_UINT64 },
870	{ "l2_write_buffer_iter",	KSTAT_DATA_UINT64 },
871	{ "l2_write_pios",		KSTAT_DATA_UINT64 },
872	{ "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
873	{ "l2_write_buffer_list_iter",	KSTAT_DATA_UINT64 },
874	{ "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
875	{ "memory_throttle_count",	KSTAT_DATA_UINT64 },
876	{ "arc_meta_used",		KSTAT_DATA_UINT64 },
877	{ "arc_meta_limit",		KSTAT_DATA_UINT64 },
878	{ "arc_meta_max",		KSTAT_DATA_UINT64 },
879	{ "arc_meta_min",		KSTAT_DATA_UINT64 },
880	{ "sync_wait_for_async",	KSTAT_DATA_UINT64 },
881	{ "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
882};
883
884#define	ARCSTAT(stat)	(arc_stats.stat.value.ui64)
885
886#define	ARCSTAT_INCR(stat, val) \
887	atomic_add_64(&arc_stats.stat.value.ui64, (val))
888
889#define	ARCSTAT_BUMP(stat)	ARCSTAT_INCR(stat, 1)
890#define	ARCSTAT_BUMPDOWN(stat)	ARCSTAT_INCR(stat, -1)
891
892#define	ARCSTAT_MAX(stat, val) {					\
893	uint64_t m;							\
894	while ((val) > (m = arc_stats.stat.value.ui64) &&		\
895	    (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val))))	\
896		continue;						\
897}
898
899#define	ARCSTAT_MAXSTAT(stat) \
900	ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
901
902/*
903 * We define a macro to allow ARC hits/misses to be easily broken down by
904 * two separate conditions, giving a total of four different subtypes for
905 * each of hits and misses (so eight statistics total).
906 */
907#define	ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
908	if (cond1) {							\
909		if (cond2) {						\
910			ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
911		} else {						\
912			ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
913		}							\
914	} else {							\
915		if (cond2) {						\
916			ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
917		} else {						\
918			ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
919		}							\
920	}
921
922kstat_t			*arc_ksp;
923static arc_state_t	*arc_anon;
924static arc_state_t	*arc_mru;
925static arc_state_t	*arc_mru_ghost;
926static arc_state_t	*arc_mfu;
927static arc_state_t	*arc_mfu_ghost;
928static arc_state_t	*arc_l2c_only;
929
930/*
931 * There are several ARC variables that are critical to export as kstats --
932 * but we don't want to have to grovel around in the kstat whenever we wish to
933 * manipulate them.  For these variables, we therefore define them to be in
934 * terms of the statistic variable.  This assures that we are not introducing
935 * the possibility of inconsistency by having shadow copies of the variables,
936 * while still allowing the code to be readable.
937 */
938#define	arc_p		ARCSTAT(arcstat_p)	/* target size of MRU */
939#define	arc_c		ARCSTAT(arcstat_c)	/* target size of cache */
940#define	arc_c_min	ARCSTAT(arcstat_c_min)	/* min target cache size */
941#define	arc_c_max	ARCSTAT(arcstat_c_max)	/* max target cache size */
942#define	arc_meta_limit	ARCSTAT(arcstat_meta_limit) /* max size for metadata */
943#define	arc_meta_min	ARCSTAT(arcstat_meta_min) /* min size for metadata */
944#define	arc_meta_max	ARCSTAT(arcstat_meta_max) /* max size of metadata */
945
946/* compressed size of entire arc */
947#define	arc_compressed_size	ARCSTAT(arcstat_compressed_size)
948/* uncompressed size of entire arc */
949#define	arc_uncompressed_size	ARCSTAT(arcstat_uncompressed_size)
950/* number of bytes in the arc from arc_buf_t's */
951#define	arc_overhead_size	ARCSTAT(arcstat_overhead_size)
952
953/*
954 * There are also some ARC variables that we want to export, but that are
955 * updated so often that having the canonical representation be the statistic
956 * variable causes a performance bottleneck. We want to use aggsum_t's for these
957 * instead, but still be able to export the kstat in the same way as before.
958 * The solution is to always use the aggsum version, except in the kstat update
959 * callback.
960 */
961aggsum_t arc_size;
962aggsum_t arc_meta_used;
963aggsum_t astat_data_size;
964aggsum_t astat_metadata_size;
965aggsum_t astat_hdr_size;
966aggsum_t astat_other_size;
967aggsum_t astat_l2_hdr_size;
968
969static int		arc_no_grow;	/* Don't try to grow cache size */
970static uint64_t		arc_tempreserve;
971static uint64_t		arc_loaned_bytes;
972
973typedef struct arc_callback arc_callback_t;
974
975struct arc_callback {
976	void			*acb_private;
977	arc_done_func_t		*acb_done;
978	arc_buf_t		*acb_buf;
979	boolean_t		acb_compressed;
980	zio_t			*acb_zio_dummy;
981	arc_callback_t		*acb_next;
982};
983
984typedef struct arc_write_callback arc_write_callback_t;
985
986struct arc_write_callback {
987	void		*awcb_private;
988	arc_done_func_t	*awcb_ready;
989	arc_done_func_t	*awcb_children_ready;
990	arc_done_func_t	*awcb_physdone;
991	arc_done_func_t	*awcb_done;
992	arc_buf_t	*awcb_buf;
993};
994
995/*
996 * ARC buffers are separated into multiple structs as a memory saving measure:
997 *   - Common fields struct, always defined, and embedded within it:
998 *       - L2-only fields, always allocated but undefined when not in L2ARC
999 *       - L1-only fields, only allocated when in L1ARC
1000 *
1001 *           Buffer in L1                     Buffer only in L2
1002 *    +------------------------+          +------------------------+
1003 *    | arc_buf_hdr_t          |          | arc_buf_hdr_t          |
1004 *    |                        |          |                        |
1005 *    |                        |          |                        |
1006 *    |                        |          |                        |
1007 *    +------------------------+          +------------------------+
1008 *    | l2arc_buf_hdr_t        |          | l2arc_buf_hdr_t        |
1009 *    | (undefined if L1-only) |          |                        |
1010 *    +------------------------+          +------------------------+
1011 *    | l1arc_buf_hdr_t        |
1012 *    |                        |
1013 *    |                        |
1014 *    |                        |
1015 *    |                        |
1016 *    +------------------------+
1017 *
1018 * Because it's possible for the L2ARC to become extremely large, we can wind
1019 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
1020 * is minimized by only allocating the fields necessary for an L1-cached buffer
1021 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
1022 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
1023 * words in pointers. arc_hdr_realloc() is used to switch a header between
1024 * these two allocation states.
1025 */
1026typedef struct l1arc_buf_hdr {
1027	kmutex_t		b_freeze_lock;
1028	zio_cksum_t		*b_freeze_cksum;
1029#ifdef ZFS_DEBUG
1030	/*
1031	 * Used for debugging with kmem_flags - by allocating and freeing
1032	 * b_thawed when the buffer is thawed, we get a record of the stack
1033	 * trace that thawed it.
1034	 */
1035	void			*b_thawed;
1036#endif
1037
1038	arc_buf_t		*b_buf;
1039	uint32_t		b_bufcnt;
1040	/* for waiting on writes to complete */
1041	kcondvar_t		b_cv;
1042	uint8_t			b_byteswap;
1043
1044	/* protected by arc state mutex */
1045	arc_state_t		*b_state;
1046	multilist_node_t	b_arc_node;
1047
1048	/* updated atomically */
1049	clock_t			b_arc_access;
1050
1051	/* self protecting */
1052	refcount_t		b_refcnt;
1053
1054	arc_callback_t		*b_acb;
1055	abd_t			*b_pabd;
1056} l1arc_buf_hdr_t;
1057
1058typedef struct l2arc_dev l2arc_dev_t;
1059
1060typedef struct l2arc_buf_hdr {
1061	/* protected by arc_buf_hdr mutex */
1062	l2arc_dev_t		*b_dev;		/* L2ARC device */
1063	uint64_t		b_daddr;	/* disk address, offset byte */
1064
1065	list_node_t		b_l2node;
1066} l2arc_buf_hdr_t;
1067
1068struct arc_buf_hdr {
1069	/* protected by hash lock */
1070	dva_t			b_dva;
1071	uint64_t		b_birth;
1072
1073	arc_buf_contents_t	b_type;
1074	arc_buf_hdr_t		*b_hash_next;
1075	arc_flags_t		b_flags;
1076
1077	/*
1078	 * This field stores the size of the data buffer after
1079	 * compression, and is set in the arc's zio completion handlers.
1080	 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
1081	 *
1082	 * While the block pointers can store up to 32MB in their psize
1083	 * field, we can only store up to 32MB minus 512B. This is due
1084	 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
1085	 * a field of zeros represents 512B in the bp). We can't use a
1086	 * bias of 1 since we need to reserve a psize of zero, here, to
1087	 * represent holes and embedded blocks.
1088	 *
1089	 * This isn't a problem in practice, since the maximum size of a
1090	 * buffer is limited to 16MB, so we never need to store 32MB in
1091	 * this field. Even in the upstream illumos code base, the
1092	 * maximum size of a buffer is limited to 16MB.
1093	 */
1094	uint16_t		b_psize;
1095
1096	/*
1097	 * This field stores the size of the data buffer before
1098	 * compression, and cannot change once set. It is in units
1099	 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
1100	 */
1101	uint16_t		b_lsize;	/* immutable */
1102	uint64_t		b_spa;		/* immutable */
1103
1104	/* L2ARC fields. Undefined when not in L2ARC. */
1105	l2arc_buf_hdr_t		b_l2hdr;
1106	/* L1ARC fields. Undefined when in l2arc_only state */
1107	l1arc_buf_hdr_t		b_l1hdr;
1108};
1109
1110#if defined(__FreeBSD__) && defined(_KERNEL)
1111static int
1112sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
1113{
1114	uint64_t val;
1115	int err;
1116
1117	val = arc_meta_limit;
1118	err = sysctl_handle_64(oidp, &val, 0, req);
1119	if (err != 0 || req->newptr == NULL)
1120		return (err);
1121
1122        if (val <= 0 || val > arc_c_max)
1123		return (EINVAL);
1124
1125	arc_meta_limit = val;
1126	return (0);
1127}
1128
1129static int
1130sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS)
1131{
1132	uint32_t val;
1133	int err;
1134
1135	val = arc_no_grow_shift;
1136	err = sysctl_handle_32(oidp, &val, 0, req);
1137	if (err != 0 || req->newptr == NULL)
1138		return (err);
1139
1140        if (val >= arc_shrink_shift)
1141		return (EINVAL);
1142
1143	arc_no_grow_shift = val;
1144	return (0);
1145}
1146
1147static int
1148sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
1149{
1150	uint64_t val;
1151	int err;
1152
1153	val = zfs_arc_max;
1154	err = sysctl_handle_64(oidp, &val, 0, req);
1155	if (err != 0 || req->newptr == NULL)
1156		return (err);
1157
1158	if (zfs_arc_max == 0) {
1159		/* Loader tunable so blindly set */
1160		zfs_arc_max = val;
1161		return (0);
1162	}
1163
1164	if (val < arc_abs_min || val > kmem_size())
1165		return (EINVAL);
1166	if (val < arc_c_min)
1167		return (EINVAL);
1168	if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
1169		return (EINVAL);
1170
1171	arc_c_max = val;
1172
1173	arc_c = arc_c_max;
1174        arc_p = (arc_c >> 1);
1175
1176	if (zfs_arc_meta_limit == 0) {
1177		/* limit meta-data to 1/4 of the arc capacity */
1178		arc_meta_limit = arc_c_max / 4;
1179	}
1180
1181	/* if kmem_flags are set, lets try to use less memory */
1182	if (kmem_debugging())
1183		arc_c = arc_c / 2;
1184
1185	zfs_arc_max = arc_c;
1186
1187	return (0);
1188}
1189
1190static int
1191sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
1192{
1193	uint64_t val;
1194	int err;
1195
1196	val = zfs_arc_min;
1197	err = sysctl_handle_64(oidp, &val, 0, req);
1198	if (err != 0 || req->newptr == NULL)
1199		return (err);
1200
1201	if (zfs_arc_min == 0) {
1202		/* Loader tunable so blindly set */
1203		zfs_arc_min = val;
1204		return (0);
1205	}
1206
1207	if (val < arc_abs_min || val > arc_c_max)
1208		return (EINVAL);
1209
1210	arc_c_min = val;
1211
1212	if (zfs_arc_meta_min == 0)
1213                arc_meta_min = arc_c_min / 2;
1214
1215	if (arc_c < arc_c_min)
1216                arc_c = arc_c_min;
1217
1218	zfs_arc_min = arc_c_min;
1219
1220	return (0);
1221}
1222#endif
1223
1224#define	GHOST_STATE(state)	\
1225	((state) == arc_mru_ghost || (state) == arc_mfu_ghost ||	\
1226	(state) == arc_l2c_only)
1227
1228#define	HDR_IN_HASH_TABLE(hdr)	((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1229#define	HDR_IO_IN_PROGRESS(hdr)	((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1230#define	HDR_IO_ERROR(hdr)	((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1231#define	HDR_PREFETCH(hdr)	((hdr)->b_flags & ARC_FLAG_PREFETCH)
1232#define	HDR_COMPRESSION_ENABLED(hdr)	\
1233	((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1234
1235#define	HDR_L2CACHE(hdr)	((hdr)->b_flags & ARC_FLAG_L2CACHE)
1236#define	HDR_L2_READING(hdr)	\
1237	(((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) &&	\
1238	((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1239#define	HDR_L2_WRITING(hdr)	((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1240#define	HDR_L2_EVICTED(hdr)	((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1241#define	HDR_L2_WRITE_HEAD(hdr)	((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1242#define	HDR_SHARED_DATA(hdr)	((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1243
1244#define	HDR_ISTYPE_METADATA(hdr)	\
1245	((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1246#define	HDR_ISTYPE_DATA(hdr)	(!HDR_ISTYPE_METADATA(hdr))
1247
1248#define	HDR_HAS_L1HDR(hdr)	((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1249#define	HDR_HAS_L2HDR(hdr)	((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1250
1251/* For storing compression mode in b_flags */
1252#define	HDR_COMPRESS_OFFSET	(highbit64(ARC_FLAG_COMPRESS_0) - 1)
1253
1254#define	HDR_GET_COMPRESS(hdr)	((enum zio_compress)BF32_GET((hdr)->b_flags, \
1255	HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1256#define	HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1257	HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1258
1259#define	ARC_BUF_LAST(buf)	((buf)->b_next == NULL)
1260#define	ARC_BUF_SHARED(buf)	((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1261#define	ARC_BUF_COMPRESSED(buf)	((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1262
1263/*
1264 * Other sizes
1265 */
1266
1267#define	HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1268#define	HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1269
1270/*
1271 * Hash table routines
1272 */
1273
1274#define	HT_LOCK_PAD	CACHE_LINE_SIZE
1275
1276struct ht_lock {
1277	kmutex_t	ht_lock;
1278#ifdef _KERNEL
1279	unsigned char	pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1280#endif
1281};
1282
1283#define	BUF_LOCKS 256
1284typedef struct buf_hash_table {
1285	uint64_t ht_mask;
1286	arc_buf_hdr_t **ht_table;
1287	struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1288} buf_hash_table_t;
1289
1290static buf_hash_table_t buf_hash_table;
1291
1292#define	BUF_HASH_INDEX(spa, dva, birth) \
1293	(buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1294#define	BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1295#define	BUF_HASH_LOCK(idx)	(&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1296#define	HDR_LOCK(hdr) \
1297	(BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1298
1299uint64_t zfs_crc64_table[256];
1300
1301/*
1302 * Level 2 ARC
1303 */
1304
1305#define	L2ARC_WRITE_SIZE	(8 * 1024 * 1024)	/* initial write max */
1306#define	L2ARC_HEADROOM		2			/* num of writes */
1307/*
1308 * If we discover during ARC scan any buffers to be compressed, we boost
1309 * our headroom for the next scanning cycle by this percentage multiple.
1310 */
1311#define	L2ARC_HEADROOM_BOOST	200
1312#define	L2ARC_FEED_SECS		1		/* caching interval secs */
1313#define	L2ARC_FEED_MIN_MS	200		/* min caching interval ms */
1314
1315#define	l2arc_writes_sent	ARCSTAT(arcstat_l2_writes_sent)
1316#define	l2arc_writes_done	ARCSTAT(arcstat_l2_writes_done)
1317
1318/* L2ARC Performance Tunables */
1319uint64_t l2arc_write_max = L2ARC_WRITE_SIZE;	/* default max write size */
1320uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE;	/* extra write during warmup */
1321uint64_t l2arc_headroom = L2ARC_HEADROOM;	/* number of dev writes */
1322uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1323uint64_t l2arc_feed_secs = L2ARC_FEED_SECS;	/* interval seconds */
1324uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS;	/* min interval milliseconds */
1325boolean_t l2arc_noprefetch = B_TRUE;		/* don't cache prefetch bufs */
1326boolean_t l2arc_feed_again = B_TRUE;		/* turbo warmup */
1327boolean_t l2arc_norw = B_TRUE;			/* no reads during writes */
1328
1329SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1330    &l2arc_write_max, 0, "max write size");
1331SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1332    &l2arc_write_boost, 0, "extra write during warmup");
1333SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1334    &l2arc_headroom, 0, "number of dev writes");
1335SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1336    &l2arc_feed_secs, 0, "interval seconds");
1337SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1338    &l2arc_feed_min_ms, 0, "min interval milliseconds");
1339
1340SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1341    &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1342SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1343    &l2arc_feed_again, 0, "turbo warmup");
1344SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1345    &l2arc_norw, 0, "no reads during writes");
1346
1347SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1348    &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1349SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD,
1350    &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1351    "size of anonymous state");
1352SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD,
1353    &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1354    "size of anonymous state");
1355
1356SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1357    &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1358SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD,
1359    &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1360    "size of metadata in mru state");
1361SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD,
1362    &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1363    "size of data in mru state");
1364
1365SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1366    &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1367SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD,
1368    &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1369    "size of metadata in mru ghost state");
1370SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD,
1371    &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1372    "size of data in mru ghost state");
1373
1374SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1375    &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1376SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD,
1377    &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1378    "size of metadata in mfu state");
1379SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD,
1380    &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1381    "size of data in mfu state");
1382
1383SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1384    &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1385SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD,
1386    &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1387    "size of metadata in mfu ghost state");
1388SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD,
1389    &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1390    "size of data in mfu ghost state");
1391
1392SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1393    &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1394
1395/*
1396 * L2ARC Internals
1397 */
1398struct l2arc_dev {
1399	vdev_t			*l2ad_vdev;	/* vdev */
1400	spa_t			*l2ad_spa;	/* spa */
1401	uint64_t		l2ad_hand;	/* next write location */
1402	uint64_t		l2ad_start;	/* first addr on device */
1403	uint64_t		l2ad_end;	/* last addr on device */
1404	boolean_t		l2ad_first;	/* first sweep through */
1405	boolean_t		l2ad_writing;	/* currently writing */
1406	kmutex_t		l2ad_mtx;	/* lock for buffer list */
1407	list_t			l2ad_buflist;	/* buffer list */
1408	list_node_t		l2ad_node;	/* device list node */
1409	refcount_t		l2ad_alloc;	/* allocated bytes */
1410};
1411
1412static list_t L2ARC_dev_list;			/* device list */
1413static list_t *l2arc_dev_list;			/* device list pointer */
1414static kmutex_t l2arc_dev_mtx;			/* device list mutex */
1415static l2arc_dev_t *l2arc_dev_last;		/* last device used */
1416static list_t L2ARC_free_on_write;		/* free after write buf list */
1417static list_t *l2arc_free_on_write;		/* free after write list ptr */
1418static kmutex_t l2arc_free_on_write_mtx;	/* mutex for list */
1419static uint64_t l2arc_ndev;			/* number of devices */
1420
1421typedef struct l2arc_read_callback {
1422	arc_buf_hdr_t		*l2rcb_hdr;		/* read header */
1423	blkptr_t		l2rcb_bp;		/* original blkptr */
1424	zbookmark_phys_t	l2rcb_zb;		/* original bookmark */
1425	int			l2rcb_flags;		/* original flags */
1426	abd_t			*l2rcb_abd;		/* temporary buffer */
1427} l2arc_read_callback_t;
1428
1429typedef struct l2arc_write_callback {
1430	l2arc_dev_t	*l2wcb_dev;		/* device info */
1431	arc_buf_hdr_t	*l2wcb_head;		/* head of write buflist */
1432} l2arc_write_callback_t;
1433
1434typedef struct l2arc_data_free {
1435	/* protected by l2arc_free_on_write_mtx */
1436	abd_t		*l2df_abd;
1437	size_t		l2df_size;
1438	arc_buf_contents_t l2df_type;
1439	list_node_t	l2df_list_node;
1440} l2arc_data_free_t;
1441
1442static kmutex_t l2arc_feed_thr_lock;
1443static kcondvar_t l2arc_feed_thr_cv;
1444static uint8_t l2arc_thread_exit;
1445
1446static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1447static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1448static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1449static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1450static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1451static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1452static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1453static void arc_hdr_alloc_pabd(arc_buf_hdr_t *);
1454static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1455static boolean_t arc_is_overflowing();
1456static void arc_buf_watch(arc_buf_t *);
1457
1458static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1459static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1460static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1461static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1462
1463static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1464static void l2arc_read_done(zio_t *);
1465
1466static void
1467l2arc_trim(const arc_buf_hdr_t *hdr)
1468{
1469	l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1470
1471	ASSERT(HDR_HAS_L2HDR(hdr));
1472	ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1473
1474	if (HDR_GET_PSIZE(hdr) != 0) {
1475		trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1476		    HDR_GET_PSIZE(hdr), 0);
1477	}
1478}
1479
1480/*
1481 * We use Cityhash for this. It's fast, and has good hash properties without
1482 * requiring any large static buffers.
1483 */
1484static uint64_t
1485buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1486{
1487	return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
1488}
1489
1490#define	HDR_EMPTY(hdr)						\
1491	((hdr)->b_dva.dva_word[0] == 0 &&			\
1492	(hdr)->b_dva.dva_word[1] == 0)
1493
1494#define	HDR_EQUAL(spa, dva, birth, hdr)				\
1495	((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) &&	\
1496	((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) &&	\
1497	((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1498
1499static void
1500buf_discard_identity(arc_buf_hdr_t *hdr)
1501{
1502	hdr->b_dva.dva_word[0] = 0;
1503	hdr->b_dva.dva_word[1] = 0;
1504	hdr->b_birth = 0;
1505}
1506
1507static arc_buf_hdr_t *
1508buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1509{
1510	const dva_t *dva = BP_IDENTITY(bp);
1511	uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1512	uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1513	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1514	arc_buf_hdr_t *hdr;
1515
1516	mutex_enter(hash_lock);
1517	for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1518	    hdr = hdr->b_hash_next) {
1519		if (HDR_EQUAL(spa, dva, birth, hdr)) {
1520			*lockp = hash_lock;
1521			return (hdr);
1522		}
1523	}
1524	mutex_exit(hash_lock);
1525	*lockp = NULL;
1526	return (NULL);
1527}
1528
1529/*
1530 * Insert an entry into the hash table.  If there is already an element
1531 * equal to elem in the hash table, then the already existing element
1532 * will be returned and the new element will not be inserted.
1533 * Otherwise returns NULL.
1534 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1535 */
1536static arc_buf_hdr_t *
1537buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1538{
1539	uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1540	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1541	arc_buf_hdr_t *fhdr;
1542	uint32_t i;
1543
1544	ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1545	ASSERT(hdr->b_birth != 0);
1546	ASSERT(!HDR_IN_HASH_TABLE(hdr));
1547
1548	if (lockp != NULL) {
1549		*lockp = hash_lock;
1550		mutex_enter(hash_lock);
1551	} else {
1552		ASSERT(MUTEX_HELD(hash_lock));
1553	}
1554
1555	for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1556	    fhdr = fhdr->b_hash_next, i++) {
1557		if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1558			return (fhdr);
1559	}
1560
1561	hdr->b_hash_next = buf_hash_table.ht_table[idx];
1562	buf_hash_table.ht_table[idx] = hdr;
1563	arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1564
1565	/* collect some hash table performance data */
1566	if (i > 0) {
1567		ARCSTAT_BUMP(arcstat_hash_collisions);
1568		if (i == 1)
1569			ARCSTAT_BUMP(arcstat_hash_chains);
1570
1571		ARCSTAT_MAX(arcstat_hash_chain_max, i);
1572	}
1573
1574	ARCSTAT_BUMP(arcstat_hash_elements);
1575	ARCSTAT_MAXSTAT(arcstat_hash_elements);
1576
1577	return (NULL);
1578}
1579
1580static void
1581buf_hash_remove(arc_buf_hdr_t *hdr)
1582{
1583	arc_buf_hdr_t *fhdr, **hdrp;
1584	uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1585
1586	ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1587	ASSERT(HDR_IN_HASH_TABLE(hdr));
1588
1589	hdrp = &buf_hash_table.ht_table[idx];
1590	while ((fhdr = *hdrp) != hdr) {
1591		ASSERT3P(fhdr, !=, NULL);
1592		hdrp = &fhdr->b_hash_next;
1593	}
1594	*hdrp = hdr->b_hash_next;
1595	hdr->b_hash_next = NULL;
1596	arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1597
1598	/* collect some hash table performance data */
1599	ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1600
1601	if (buf_hash_table.ht_table[idx] &&
1602	    buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1603		ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1604}
1605
1606/*
1607 * Global data structures and functions for the buf kmem cache.
1608 */
1609static kmem_cache_t *hdr_full_cache;
1610static kmem_cache_t *hdr_l2only_cache;
1611static kmem_cache_t *buf_cache;
1612
1613static void
1614buf_fini(void)
1615{
1616	int i;
1617
1618	kmem_free(buf_hash_table.ht_table,
1619	    (buf_hash_table.ht_mask + 1) * sizeof (void *));
1620	for (i = 0; i < BUF_LOCKS; i++)
1621		mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1622	kmem_cache_destroy(hdr_full_cache);
1623	kmem_cache_destroy(hdr_l2only_cache);
1624	kmem_cache_destroy(buf_cache);
1625}
1626
1627/*
1628 * Constructor callback - called when the cache is empty
1629 * and a new buf is requested.
1630 */
1631/* ARGSUSED */
1632static int
1633hdr_full_cons(void *vbuf, void *unused, int kmflag)
1634{
1635	arc_buf_hdr_t *hdr = vbuf;
1636
1637	bzero(hdr, HDR_FULL_SIZE);
1638	cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1639	refcount_create(&hdr->b_l1hdr.b_refcnt);
1640	mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1641	multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1642	arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1643
1644	return (0);
1645}
1646
1647/* ARGSUSED */
1648static int
1649hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1650{
1651	arc_buf_hdr_t *hdr = vbuf;
1652
1653	bzero(hdr, HDR_L2ONLY_SIZE);
1654	arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1655
1656	return (0);
1657}
1658
1659/* ARGSUSED */
1660static int
1661buf_cons(void *vbuf, void *unused, int kmflag)
1662{
1663	arc_buf_t *buf = vbuf;
1664
1665	bzero(buf, sizeof (arc_buf_t));
1666	mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1667	arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1668
1669	return (0);
1670}
1671
1672/*
1673 * Destructor callback - called when a cached buf is
1674 * no longer required.
1675 */
1676/* ARGSUSED */
1677static void
1678hdr_full_dest(void *vbuf, void *unused)
1679{
1680	arc_buf_hdr_t *hdr = vbuf;
1681
1682	ASSERT(HDR_EMPTY(hdr));
1683	cv_destroy(&hdr->b_l1hdr.b_cv);
1684	refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1685	mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1686	ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1687	arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1688}
1689
1690/* ARGSUSED */
1691static void
1692hdr_l2only_dest(void *vbuf, void *unused)
1693{
1694	arc_buf_hdr_t *hdr = vbuf;
1695
1696	ASSERT(HDR_EMPTY(hdr));
1697	arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1698}
1699
1700/* ARGSUSED */
1701static void
1702buf_dest(void *vbuf, void *unused)
1703{
1704	arc_buf_t *buf = vbuf;
1705
1706	mutex_destroy(&buf->b_evict_lock);
1707	arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1708}
1709
1710/*
1711 * Reclaim callback -- invoked when memory is low.
1712 */
1713/* ARGSUSED */
1714static void
1715hdr_recl(void *unused)
1716{
1717	dprintf("hdr_recl called\n");
1718	/*
1719	 * umem calls the reclaim func when we destroy the buf cache,
1720	 * which is after we do arc_fini().
1721	 */
1722	if (!arc_dead)
1723		cv_signal(&arc_reclaim_thread_cv);
1724}
1725
1726static void
1727buf_init(void)
1728{
1729	uint64_t *ct;
1730	uint64_t hsize = 1ULL << 12;
1731	int i, j;
1732
1733	/*
1734	 * The hash table is big enough to fill all of physical memory
1735	 * with an average block size of zfs_arc_average_blocksize (default 8K).
1736	 * By default, the table will take up
1737	 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1738	 */
1739	while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1740		hsize <<= 1;
1741retry:
1742	buf_hash_table.ht_mask = hsize - 1;
1743	buf_hash_table.ht_table =
1744	    kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1745	if (buf_hash_table.ht_table == NULL) {
1746		ASSERT(hsize > (1ULL << 8));
1747		hsize >>= 1;
1748		goto retry;
1749	}
1750
1751	hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1752	    0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1753	hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1754	    HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1755	    NULL, NULL, 0);
1756	buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1757	    0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1758
1759	for (i = 0; i < 256; i++)
1760		for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1761			*ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1762
1763	for (i = 0; i < BUF_LOCKS; i++) {
1764		mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1765		    NULL, MUTEX_DEFAULT, NULL);
1766	}
1767}
1768
1769/*
1770 * This is the size that the buf occupies in memory. If the buf is compressed,
1771 * it will correspond to the compressed size. You should use this method of
1772 * getting the buf size unless you explicitly need the logical size.
1773 */
1774int32_t
1775arc_buf_size(arc_buf_t *buf)
1776{
1777	return (ARC_BUF_COMPRESSED(buf) ?
1778	    HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1779}
1780
1781int32_t
1782arc_buf_lsize(arc_buf_t *buf)
1783{
1784	return (HDR_GET_LSIZE(buf->b_hdr));
1785}
1786
1787enum zio_compress
1788arc_get_compression(arc_buf_t *buf)
1789{
1790	return (ARC_BUF_COMPRESSED(buf) ?
1791	    HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1792}
1793
1794#define	ARC_MINTIME	(hz>>4) /* 62 ms */
1795
1796static inline boolean_t
1797arc_buf_is_shared(arc_buf_t *buf)
1798{
1799	boolean_t shared = (buf->b_data != NULL &&
1800	    buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1801	    abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1802	    buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1803	IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1804	IMPLY(shared, ARC_BUF_SHARED(buf));
1805	IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1806
1807	/*
1808	 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1809	 * already being shared" requirement prevents us from doing that.
1810	 */
1811
1812	return (shared);
1813}
1814
1815/*
1816 * Free the checksum associated with this header. If there is no checksum, this
1817 * is a no-op.
1818 */
1819static inline void
1820arc_cksum_free(arc_buf_hdr_t *hdr)
1821{
1822	ASSERT(HDR_HAS_L1HDR(hdr));
1823	mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1824	if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1825		kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1826		hdr->b_l1hdr.b_freeze_cksum = NULL;
1827	}
1828	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1829}
1830
1831/*
1832 * Return true iff at least one of the bufs on hdr is not compressed.
1833 */
1834static boolean_t
1835arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1836{
1837	for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1838		if (!ARC_BUF_COMPRESSED(b)) {
1839			return (B_TRUE);
1840		}
1841	}
1842	return (B_FALSE);
1843}
1844
1845/*
1846 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1847 * matches the checksum that is stored in the hdr. If there is no checksum,
1848 * or if the buf is compressed, this is a no-op.
1849 */
1850static void
1851arc_cksum_verify(arc_buf_t *buf)
1852{
1853	arc_buf_hdr_t *hdr = buf->b_hdr;
1854	zio_cksum_t zc;
1855
1856	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1857		return;
1858
1859	if (ARC_BUF_COMPRESSED(buf)) {
1860		ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1861		    arc_hdr_has_uncompressed_buf(hdr));
1862		return;
1863	}
1864
1865	ASSERT(HDR_HAS_L1HDR(hdr));
1866
1867	mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1868	if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1869		mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1870		return;
1871	}
1872
1873	fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1874	if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1875		panic("buffer modified while frozen!");
1876	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1877}
1878
1879static boolean_t
1880arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1881{
1882	enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1883	boolean_t valid_cksum;
1884
1885	ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1886	VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1887
1888	/*
1889	 * We rely on the blkptr's checksum to determine if the block
1890	 * is valid or not. When compressed arc is enabled, the l2arc
1891	 * writes the block to the l2arc just as it appears in the pool.
1892	 * This allows us to use the blkptr's checksum to validate the
1893	 * data that we just read off of the l2arc without having to store
1894	 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1895	 * arc is disabled, then the data written to the l2arc is always
1896	 * uncompressed and won't match the block as it exists in the main
1897	 * pool. When this is the case, we must first compress it if it is
1898	 * compressed on the main pool before we can validate the checksum.
1899	 */
1900	if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1901		ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1902		uint64_t lsize = HDR_GET_LSIZE(hdr);
1903		uint64_t csize;
1904
1905		abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
1906		csize = zio_compress_data(compress, zio->io_abd,
1907		    abd_to_buf(cdata), lsize);
1908
1909		ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1910		if (csize < HDR_GET_PSIZE(hdr)) {
1911			/*
1912			 * Compressed blocks are always a multiple of the
1913			 * smallest ashift in the pool. Ideally, we would
1914			 * like to round up the csize to the next
1915			 * spa_min_ashift but that value may have changed
1916			 * since the block was last written. Instead,
1917			 * we rely on the fact that the hdr's psize
1918			 * was set to the psize of the block when it was
1919			 * last written. We set the csize to that value
1920			 * and zero out any part that should not contain
1921			 * data.
1922			 */
1923			abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
1924			csize = HDR_GET_PSIZE(hdr);
1925		}
1926		zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
1927	}
1928
1929	/*
1930	 * Block pointers always store the checksum for the logical data.
1931	 * If the block pointer has the gang bit set, then the checksum
1932	 * it represents is for the reconstituted data and not for an
1933	 * individual gang member. The zio pipeline, however, must be able to
1934	 * determine the checksum of each of the gang constituents so it
1935	 * treats the checksum comparison differently than what we need
1936	 * for l2arc blocks. This prevents us from using the
1937	 * zio_checksum_error() interface directly. Instead we must call the
1938	 * zio_checksum_error_impl() so that we can ensure the checksum is
1939	 * generated using the correct checksum algorithm and accounts for the
1940	 * logical I/O size and not just a gang fragment.
1941	 */
1942	valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1943	    BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1944	    zio->io_offset, NULL) == 0);
1945	zio_pop_transforms(zio);
1946	return (valid_cksum);
1947}
1948
1949/*
1950 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1951 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1952 * isn't modified later on. If buf is compressed or there is already a checksum
1953 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1954 */
1955static void
1956arc_cksum_compute(arc_buf_t *buf)
1957{
1958	arc_buf_hdr_t *hdr = buf->b_hdr;
1959
1960	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1961		return;
1962
1963	ASSERT(HDR_HAS_L1HDR(hdr));
1964
1965	mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1966	if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1967		ASSERT(arc_hdr_has_uncompressed_buf(hdr));
1968		mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1969		return;
1970	} else if (ARC_BUF_COMPRESSED(buf)) {
1971		mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1972		return;
1973	}
1974
1975	ASSERT(!ARC_BUF_COMPRESSED(buf));
1976	hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1977	    KM_SLEEP);
1978	fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1979	    hdr->b_l1hdr.b_freeze_cksum);
1980	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1981#ifdef illumos
1982	arc_buf_watch(buf);
1983#endif
1984}
1985
1986#ifdef illumos
1987#ifndef _KERNEL
1988typedef struct procctl {
1989	long cmd;
1990	prwatch_t prwatch;
1991} procctl_t;
1992#endif
1993
1994/* ARGSUSED */
1995static void
1996arc_buf_unwatch(arc_buf_t *buf)
1997{
1998#ifndef _KERNEL
1999	if (arc_watch) {
2000		int result;
2001		procctl_t ctl;
2002		ctl.cmd = PCWATCH;
2003		ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2004		ctl.prwatch.pr_size = 0;
2005		ctl.prwatch.pr_wflags = 0;
2006		result = write(arc_procfd, &ctl, sizeof (ctl));
2007		ASSERT3U(result, ==, sizeof (ctl));
2008	}
2009#endif
2010}
2011
2012/* ARGSUSED */
2013static void
2014arc_buf_watch(arc_buf_t *buf)
2015{
2016#ifndef _KERNEL
2017	if (arc_watch) {
2018		int result;
2019		procctl_t ctl;
2020		ctl.cmd = PCWATCH;
2021		ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2022		ctl.prwatch.pr_size = arc_buf_size(buf);
2023		ctl.prwatch.pr_wflags = WA_WRITE;
2024		result = write(arc_procfd, &ctl, sizeof (ctl));
2025		ASSERT3U(result, ==, sizeof (ctl));
2026	}
2027#endif
2028}
2029#endif /* illumos */
2030
2031static arc_buf_contents_t
2032arc_buf_type(arc_buf_hdr_t *hdr)
2033{
2034	arc_buf_contents_t type;
2035	if (HDR_ISTYPE_METADATA(hdr)) {
2036		type = ARC_BUFC_METADATA;
2037	} else {
2038		type = ARC_BUFC_DATA;
2039	}
2040	VERIFY3U(hdr->b_type, ==, type);
2041	return (type);
2042}
2043
2044boolean_t
2045arc_is_metadata(arc_buf_t *buf)
2046{
2047	return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
2048}
2049
2050static uint32_t
2051arc_bufc_to_flags(arc_buf_contents_t type)
2052{
2053	switch (type) {
2054	case ARC_BUFC_DATA:
2055		/* metadata field is 0 if buffer contains normal data */
2056		return (0);
2057	case ARC_BUFC_METADATA:
2058		return (ARC_FLAG_BUFC_METADATA);
2059	default:
2060		break;
2061	}
2062	panic("undefined ARC buffer type!");
2063	return ((uint32_t)-1);
2064}
2065
2066void
2067arc_buf_thaw(arc_buf_t *buf)
2068{
2069	arc_buf_hdr_t *hdr = buf->b_hdr;
2070
2071	ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2072	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2073
2074	arc_cksum_verify(buf);
2075
2076	/*
2077	 * Compressed buffers do not manipulate the b_freeze_cksum or
2078	 * allocate b_thawed.
2079	 */
2080	if (ARC_BUF_COMPRESSED(buf)) {
2081		ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2082		    arc_hdr_has_uncompressed_buf(hdr));
2083		return;
2084	}
2085
2086	ASSERT(HDR_HAS_L1HDR(hdr));
2087	arc_cksum_free(hdr);
2088
2089	mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
2090#ifdef ZFS_DEBUG
2091	if (zfs_flags & ZFS_DEBUG_MODIFY) {
2092		if (hdr->b_l1hdr.b_thawed != NULL)
2093			kmem_free(hdr->b_l1hdr.b_thawed, 1);
2094		hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
2095	}
2096#endif
2097
2098	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2099
2100#ifdef illumos
2101	arc_buf_unwatch(buf);
2102#endif
2103}
2104
2105void
2106arc_buf_freeze(arc_buf_t *buf)
2107{
2108	arc_buf_hdr_t *hdr = buf->b_hdr;
2109	kmutex_t *hash_lock;
2110
2111	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2112		return;
2113
2114	if (ARC_BUF_COMPRESSED(buf)) {
2115		ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2116		    arc_hdr_has_uncompressed_buf(hdr));
2117		return;
2118	}
2119
2120	hash_lock = HDR_LOCK(hdr);
2121	mutex_enter(hash_lock);
2122
2123	ASSERT(HDR_HAS_L1HDR(hdr));
2124	ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
2125	    hdr->b_l1hdr.b_state == arc_anon);
2126	arc_cksum_compute(buf);
2127	mutex_exit(hash_lock);
2128}
2129
2130/*
2131 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
2132 * the following functions should be used to ensure that the flags are
2133 * updated in a thread-safe way. When manipulating the flags either
2134 * the hash_lock must be held or the hdr must be undiscoverable. This
2135 * ensures that we're not racing with any other threads when updating
2136 * the flags.
2137 */
2138static inline void
2139arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2140{
2141	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2142	hdr->b_flags |= flags;
2143}
2144
2145static inline void
2146arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2147{
2148	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2149	hdr->b_flags &= ~flags;
2150}
2151
2152/*
2153 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
2154 * done in a special way since we have to clear and set bits
2155 * at the same time. Consumers that wish to set the compression bits
2156 * must use this function to ensure that the flags are updated in
2157 * thread-safe manner.
2158 */
2159static void
2160arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
2161{
2162	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2163
2164	/*
2165	 * Holes and embedded blocks will always have a psize = 0 so
2166	 * we ignore the compression of the blkptr and set the
2167	 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
2168	 * Holes and embedded blocks remain anonymous so we don't
2169	 * want to uncompress them. Mark them as uncompressed.
2170	 */
2171	if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
2172		arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2173		HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
2174		ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
2175		ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2176	} else {
2177		arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2178		HDR_SET_COMPRESS(hdr, cmp);
2179		ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
2180		ASSERT(HDR_COMPRESSION_ENABLED(hdr));
2181	}
2182}
2183
2184/*
2185 * Looks for another buf on the same hdr which has the data decompressed, copies
2186 * from it, and returns true. If no such buf exists, returns false.
2187 */
2188static boolean_t
2189arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
2190{
2191	arc_buf_hdr_t *hdr = buf->b_hdr;
2192	boolean_t copied = B_FALSE;
2193
2194	ASSERT(HDR_HAS_L1HDR(hdr));
2195	ASSERT3P(buf->b_data, !=, NULL);
2196	ASSERT(!ARC_BUF_COMPRESSED(buf));
2197
2198	for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
2199	    from = from->b_next) {
2200		/* can't use our own data buffer */
2201		if (from == buf) {
2202			continue;
2203		}
2204
2205		if (!ARC_BUF_COMPRESSED(from)) {
2206			bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
2207			copied = B_TRUE;
2208			break;
2209		}
2210	}
2211
2212	/*
2213	 * There were no decompressed bufs, so there should not be a
2214	 * checksum on the hdr either.
2215	 */
2216	EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
2217
2218	return (copied);
2219}
2220
2221/*
2222 * Given a buf that has a data buffer attached to it, this function will
2223 * efficiently fill the buf with data of the specified compression setting from
2224 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2225 * are already sharing a data buf, no copy is performed.
2226 *
2227 * If the buf is marked as compressed but uncompressed data was requested, this
2228 * will allocate a new data buffer for the buf, remove that flag, and fill the
2229 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2230 * uncompressed data, and (since we haven't added support for it yet) if you
2231 * want compressed data your buf must already be marked as compressed and have
2232 * the correct-sized data buffer.
2233 */
2234static int
2235arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
2236{
2237	arc_buf_hdr_t *hdr = buf->b_hdr;
2238	boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
2239	dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2240
2241	ASSERT3P(buf->b_data, !=, NULL);
2242	IMPLY(compressed, hdr_compressed);
2243	IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2244
2245	if (hdr_compressed == compressed) {
2246		if (!arc_buf_is_shared(buf)) {
2247			abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2248			    arc_buf_size(buf));
2249		}
2250	} else {
2251		ASSERT(hdr_compressed);
2252		ASSERT(!compressed);
2253		ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2254
2255		/*
2256		 * If the buf is sharing its data with the hdr, unlink it and
2257		 * allocate a new data buffer for the buf.
2258		 */
2259		if (arc_buf_is_shared(buf)) {
2260			ASSERT(ARC_BUF_COMPRESSED(buf));
2261
2262			/* We need to give the buf it's own b_data */
2263			buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2264			buf->b_data =
2265			    arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2266			arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2267
2268			/* Previously overhead was 0; just add new overhead */
2269			ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2270		} else if (ARC_BUF_COMPRESSED(buf)) {
2271			/* We need to reallocate the buf's b_data */
2272			arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2273			    buf);
2274			buf->b_data =
2275			    arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2276
2277			/* We increased the size of b_data; update overhead */
2278			ARCSTAT_INCR(arcstat_overhead_size,
2279			    HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2280		}
2281
2282		/*
2283		 * Regardless of the buf's previous compression settings, it
2284		 * should not be compressed at the end of this function.
2285		 */
2286		buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2287
2288		/*
2289		 * Try copying the data from another buf which already has a
2290		 * decompressed version. If that's not possible, it's time to
2291		 * bite the bullet and decompress the data from the hdr.
2292		 */
2293		if (arc_buf_try_copy_decompressed_data(buf)) {
2294			/* Skip byteswapping and checksumming (already done) */
2295			ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2296			return (0);
2297		} else {
2298			int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2299			    hdr->b_l1hdr.b_pabd, buf->b_data,
2300			    HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2301
2302			/*
2303			 * Absent hardware errors or software bugs, this should
2304			 * be impossible, but log it anyway so we can debug it.
2305			 */
2306			if (error != 0) {
2307				zfs_dbgmsg(
2308				    "hdr %p, compress %d, psize %d, lsize %d",
2309				    hdr, HDR_GET_COMPRESS(hdr),
2310				    HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2311				return (SET_ERROR(EIO));
2312			}
2313		}
2314	}
2315
2316	/* Byteswap the buf's data if necessary */
2317	if (bswap != DMU_BSWAP_NUMFUNCS) {
2318		ASSERT(!HDR_SHARED_DATA(hdr));
2319		ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2320		dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2321	}
2322
2323	/* Compute the hdr's checksum if necessary */
2324	arc_cksum_compute(buf);
2325
2326	return (0);
2327}
2328
2329int
2330arc_decompress(arc_buf_t *buf)
2331{
2332	return (arc_buf_fill(buf, B_FALSE));
2333}
2334
2335/*
2336 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
2337 */
2338static uint64_t
2339arc_hdr_size(arc_buf_hdr_t *hdr)
2340{
2341	uint64_t size;
2342
2343	if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2344	    HDR_GET_PSIZE(hdr) > 0) {
2345		size = HDR_GET_PSIZE(hdr);
2346	} else {
2347		ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2348		size = HDR_GET_LSIZE(hdr);
2349	}
2350	return (size);
2351}
2352
2353/*
2354 * Increment the amount of evictable space in the arc_state_t's refcount.
2355 * We account for the space used by the hdr and the arc buf individually
2356 * so that we can add and remove them from the refcount individually.
2357 */
2358static void
2359arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2360{
2361	arc_buf_contents_t type = arc_buf_type(hdr);
2362
2363	ASSERT(HDR_HAS_L1HDR(hdr));
2364
2365	if (GHOST_STATE(state)) {
2366		ASSERT0(hdr->b_l1hdr.b_bufcnt);
2367		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2368		ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2369		(void) refcount_add_many(&state->arcs_esize[type],
2370		    HDR_GET_LSIZE(hdr), hdr);
2371		return;
2372	}
2373
2374	ASSERT(!GHOST_STATE(state));
2375	if (hdr->b_l1hdr.b_pabd != NULL) {
2376		(void) refcount_add_many(&state->arcs_esize[type],
2377		    arc_hdr_size(hdr), hdr);
2378	}
2379	for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2380	    buf = buf->b_next) {
2381		if (arc_buf_is_shared(buf))
2382			continue;
2383		(void) refcount_add_many(&state->arcs_esize[type],
2384		    arc_buf_size(buf), buf);
2385	}
2386}
2387
2388/*
2389 * Decrement the amount of evictable space in the arc_state_t's refcount.
2390 * We account for the space used by the hdr and the arc buf individually
2391 * so that we can add and remove them from the refcount individually.
2392 */
2393static void
2394arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2395{
2396	arc_buf_contents_t type = arc_buf_type(hdr);
2397
2398	ASSERT(HDR_HAS_L1HDR(hdr));
2399
2400	if (GHOST_STATE(state)) {
2401		ASSERT0(hdr->b_l1hdr.b_bufcnt);
2402		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2403		ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2404		(void) refcount_remove_many(&state->arcs_esize[type],
2405		    HDR_GET_LSIZE(hdr), hdr);
2406		return;
2407	}
2408
2409	ASSERT(!GHOST_STATE(state));
2410	if (hdr->b_l1hdr.b_pabd != NULL) {
2411		(void) refcount_remove_many(&state->arcs_esize[type],
2412		    arc_hdr_size(hdr), hdr);
2413	}
2414	for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2415	    buf = buf->b_next) {
2416		if (arc_buf_is_shared(buf))
2417			continue;
2418		(void) refcount_remove_many(&state->arcs_esize[type],
2419		    arc_buf_size(buf), buf);
2420	}
2421}
2422
2423/*
2424 * Add a reference to this hdr indicating that someone is actively
2425 * referencing that memory. When the refcount transitions from 0 to 1,
2426 * we remove it from the respective arc_state_t list to indicate that
2427 * it is not evictable.
2428 */
2429static void
2430add_reference(arc_buf_hdr_t *hdr, void *tag)
2431{
2432	ASSERT(HDR_HAS_L1HDR(hdr));
2433	if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2434		ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2435		ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2436		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2437	}
2438
2439	arc_state_t *state = hdr->b_l1hdr.b_state;
2440
2441	if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2442	    (state != arc_anon)) {
2443		/* We don't use the L2-only state list. */
2444		if (state != arc_l2c_only) {
2445			multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2446			    hdr);
2447			arc_evictable_space_decrement(hdr, state);
2448		}
2449		/* remove the prefetch flag if we get a reference */
2450		arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2451	}
2452}
2453
2454/*
2455 * Remove a reference from this hdr. When the reference transitions from
2456 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2457 * list making it eligible for eviction.
2458 */
2459static int
2460remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2461{
2462	int cnt;
2463	arc_state_t *state = hdr->b_l1hdr.b_state;
2464
2465	ASSERT(HDR_HAS_L1HDR(hdr));
2466	ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2467	ASSERT(!GHOST_STATE(state));
2468
2469	/*
2470	 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2471	 * check to prevent usage of the arc_l2c_only list.
2472	 */
2473	if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2474	    (state != arc_anon)) {
2475		multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2476		ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2477		arc_evictable_space_increment(hdr, state);
2478	}
2479	return (cnt);
2480}
2481
2482/*
2483 * Move the supplied buffer to the indicated state. The hash lock
2484 * for the buffer must be held by the caller.
2485 */
2486static void
2487arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2488    kmutex_t *hash_lock)
2489{
2490	arc_state_t *old_state;
2491	int64_t refcnt;
2492	uint32_t bufcnt;
2493	boolean_t update_old, update_new;
2494	arc_buf_contents_t buftype = arc_buf_type(hdr);
2495
2496	/*
2497	 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2498	 * in arc_read() when bringing a buffer out of the L2ARC.  However, the
2499	 * L1 hdr doesn't always exist when we change state to arc_anon before
2500	 * destroying a header, in which case reallocating to add the L1 hdr is
2501	 * pointless.
2502	 */
2503	if (HDR_HAS_L1HDR(hdr)) {
2504		old_state = hdr->b_l1hdr.b_state;
2505		refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2506		bufcnt = hdr->b_l1hdr.b_bufcnt;
2507		update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL);
2508	} else {
2509		old_state = arc_l2c_only;
2510		refcnt = 0;
2511		bufcnt = 0;
2512		update_old = B_FALSE;
2513	}
2514	update_new = update_old;
2515
2516	ASSERT(MUTEX_HELD(hash_lock));
2517	ASSERT3P(new_state, !=, old_state);
2518	ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2519	ASSERT(old_state != arc_anon || bufcnt <= 1);
2520
2521	/*
2522	 * If this buffer is evictable, transfer it from the
2523	 * old state list to the new state list.
2524	 */
2525	if (refcnt == 0) {
2526		if (old_state != arc_anon && old_state != arc_l2c_only) {
2527			ASSERT(HDR_HAS_L1HDR(hdr));
2528			multilist_remove(old_state->arcs_list[buftype], hdr);
2529
2530			if (GHOST_STATE(old_state)) {
2531				ASSERT0(bufcnt);
2532				ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2533				update_old = B_TRUE;
2534			}
2535			arc_evictable_space_decrement(hdr, old_state);
2536		}
2537		if (new_state != arc_anon && new_state != arc_l2c_only) {
2538
2539			/*
2540			 * An L1 header always exists here, since if we're
2541			 * moving to some L1-cached state (i.e. not l2c_only or
2542			 * anonymous), we realloc the header to add an L1hdr
2543			 * beforehand.
2544			 */
2545			ASSERT(HDR_HAS_L1HDR(hdr));
2546			multilist_insert(new_state->arcs_list[buftype], hdr);
2547
2548			if (GHOST_STATE(new_state)) {
2549				ASSERT0(bufcnt);
2550				ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2551				update_new = B_TRUE;
2552			}
2553			arc_evictable_space_increment(hdr, new_state);
2554		}
2555	}
2556
2557	ASSERT(!HDR_EMPTY(hdr));
2558	if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2559		buf_hash_remove(hdr);
2560
2561	/* adjust state sizes (ignore arc_l2c_only) */
2562
2563	if (update_new && new_state != arc_l2c_only) {
2564		ASSERT(HDR_HAS_L1HDR(hdr));
2565		if (GHOST_STATE(new_state)) {
2566			ASSERT0(bufcnt);
2567
2568			/*
2569			 * When moving a header to a ghost state, we first
2570			 * remove all arc buffers. Thus, we'll have a
2571			 * bufcnt of zero, and no arc buffer to use for
2572			 * the reference. As a result, we use the arc
2573			 * header pointer for the reference.
2574			 */
2575			(void) refcount_add_many(&new_state->arcs_size,
2576			    HDR_GET_LSIZE(hdr), hdr);
2577			ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2578		} else {
2579			uint32_t buffers = 0;
2580
2581			/*
2582			 * Each individual buffer holds a unique reference,
2583			 * thus we must remove each of these references one
2584			 * at a time.
2585			 */
2586			for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2587			    buf = buf->b_next) {
2588				ASSERT3U(bufcnt, !=, 0);
2589				buffers++;
2590
2591				/*
2592				 * When the arc_buf_t is sharing the data
2593				 * block with the hdr, the owner of the
2594				 * reference belongs to the hdr. Only
2595				 * add to the refcount if the arc_buf_t is
2596				 * not shared.
2597				 */
2598				if (arc_buf_is_shared(buf))
2599					continue;
2600
2601				(void) refcount_add_many(&new_state->arcs_size,
2602				    arc_buf_size(buf), buf);
2603			}
2604			ASSERT3U(bufcnt, ==, buffers);
2605
2606			if (hdr->b_l1hdr.b_pabd != NULL) {
2607				(void) refcount_add_many(&new_state->arcs_size,
2608				    arc_hdr_size(hdr), hdr);
2609			} else {
2610				ASSERT(GHOST_STATE(old_state));
2611			}
2612		}
2613	}
2614
2615	if (update_old && old_state != arc_l2c_only) {
2616		ASSERT(HDR_HAS_L1HDR(hdr));
2617		if (GHOST_STATE(old_state)) {
2618			ASSERT0(bufcnt);
2619			ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2620
2621			/*
2622			 * When moving a header off of a ghost state,
2623			 * the header will not contain any arc buffers.
2624			 * We use the arc header pointer for the reference
2625			 * which is exactly what we did when we put the
2626			 * header on the ghost state.
2627			 */
2628
2629			(void) refcount_remove_many(&old_state->arcs_size,
2630			    HDR_GET_LSIZE(hdr), hdr);
2631		} else {
2632			uint32_t buffers = 0;
2633
2634			/*
2635			 * Each individual buffer holds a unique reference,
2636			 * thus we must remove each of these references one
2637			 * at a time.
2638			 */
2639			for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2640			    buf = buf->b_next) {
2641				ASSERT3U(bufcnt, !=, 0);
2642				buffers++;
2643
2644				/*
2645				 * When the arc_buf_t is sharing the data
2646				 * block with the hdr, the owner of the
2647				 * reference belongs to the hdr. Only
2648				 * add to the refcount if the arc_buf_t is
2649				 * not shared.
2650				 */
2651				if (arc_buf_is_shared(buf))
2652					continue;
2653
2654				(void) refcount_remove_many(
2655				    &old_state->arcs_size, arc_buf_size(buf),
2656				    buf);
2657			}
2658			ASSERT3U(bufcnt, ==, buffers);
2659			ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2660			(void) refcount_remove_many(
2661			    &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2662		}
2663	}
2664
2665	if (HDR_HAS_L1HDR(hdr))
2666		hdr->b_l1hdr.b_state = new_state;
2667
2668	/*
2669	 * L2 headers should never be on the L2 state list since they don't
2670	 * have L1 headers allocated.
2671	 */
2672	ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2673	    multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2674}
2675
2676void
2677arc_space_consume(uint64_t space, arc_space_type_t type)
2678{
2679	ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2680
2681	switch (type) {
2682	case ARC_SPACE_DATA:
2683		aggsum_add(&astat_data_size, space);
2684		break;
2685	case ARC_SPACE_META:
2686		aggsum_add(&astat_metadata_size, space);
2687		break;
2688	case ARC_SPACE_OTHER:
2689		aggsum_add(&astat_other_size, space);
2690		break;
2691	case ARC_SPACE_HDRS:
2692		aggsum_add(&astat_hdr_size, space);
2693		break;
2694	case ARC_SPACE_L2HDRS:
2695		aggsum_add(&astat_l2_hdr_size, space);
2696		break;
2697	}
2698
2699	if (type != ARC_SPACE_DATA)
2700		aggsum_add(&arc_meta_used, space);
2701
2702	aggsum_add(&arc_size, space);
2703}
2704
2705void
2706arc_space_return(uint64_t space, arc_space_type_t type)
2707{
2708	ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2709
2710	switch (type) {
2711	case ARC_SPACE_DATA:
2712		aggsum_add(&astat_data_size, -space);
2713		break;
2714	case ARC_SPACE_META:
2715		aggsum_add(&astat_metadata_size, -space);
2716		break;
2717	case ARC_SPACE_OTHER:
2718		aggsum_add(&astat_other_size, -space);
2719		break;
2720	case ARC_SPACE_HDRS:
2721		aggsum_add(&astat_hdr_size, -space);
2722		break;
2723	case ARC_SPACE_L2HDRS:
2724		aggsum_add(&astat_l2_hdr_size, -space);
2725		break;
2726	}
2727
2728	if (type != ARC_SPACE_DATA) {
2729		ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2730		/*
2731		 * We use the upper bound here rather than the precise value
2732		 * because the arc_meta_max value doesn't need to be
2733		 * precise. It's only consumed by humans via arcstats.
2734		 */
2735		if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2736			arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2737		aggsum_add(&arc_meta_used, -space);
2738	}
2739
2740	ASSERT(aggsum_compare(&arc_size, space) >= 0);
2741	aggsum_add(&arc_size, -space);
2742}
2743
2744/*
2745 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2746 * with the hdr's b_pabd.
2747 */
2748static boolean_t
2749arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2750{
2751	/*
2752	 * The criteria for sharing a hdr's data are:
2753	 * 1. the hdr's compression matches the buf's compression
2754	 * 2. the hdr doesn't need to be byteswapped
2755	 * 3. the hdr isn't already being shared
2756	 * 4. the buf is either compressed or it is the last buf in the hdr list
2757	 *
2758	 * Criterion #4 maintains the invariant that shared uncompressed
2759	 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2760	 * might ask, "if a compressed buf is allocated first, won't that be the
2761	 * last thing in the list?", but in that case it's impossible to create
2762	 * a shared uncompressed buf anyway (because the hdr must be compressed
2763	 * to have the compressed buf). You might also think that #3 is
2764	 * sufficient to make this guarantee, however it's possible
2765	 * (specifically in the rare L2ARC write race mentioned in
2766	 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2767	 * is sharable, but wasn't at the time of its allocation. Rather than
2768	 * allow a new shared uncompressed buf to be created and then shuffle
2769	 * the list around to make it the last element, this simply disallows
2770	 * sharing if the new buf isn't the first to be added.
2771	 */
2772	ASSERT3P(buf->b_hdr, ==, hdr);
2773	boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF;
2774	boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2775	return (buf_compressed == hdr_compressed &&
2776	    hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2777	    !HDR_SHARED_DATA(hdr) &&
2778	    (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2779}
2780
2781/*
2782 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2783 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2784 * copy was made successfully, or an error code otherwise.
2785 */
2786static int
2787arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2788    boolean_t fill, arc_buf_t **ret)
2789{
2790	arc_buf_t *buf;
2791
2792	ASSERT(HDR_HAS_L1HDR(hdr));
2793	ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2794	VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2795	    hdr->b_type == ARC_BUFC_METADATA);
2796	ASSERT3P(ret, !=, NULL);
2797	ASSERT3P(*ret, ==, NULL);
2798
2799	buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2800	buf->b_hdr = hdr;
2801	buf->b_data = NULL;
2802	buf->b_next = hdr->b_l1hdr.b_buf;
2803	buf->b_flags = 0;
2804
2805	add_reference(hdr, tag);
2806
2807	/*
2808	 * We're about to change the hdr's b_flags. We must either
2809	 * hold the hash_lock or be undiscoverable.
2810	 */
2811	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2812
2813	/*
2814	 * Only honor requests for compressed bufs if the hdr is actually
2815	 * compressed.
2816	 */
2817	if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
2818		buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2819
2820	/*
2821	 * If the hdr's data can be shared then we share the data buffer and
2822	 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2823	 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2824	 * buffer to store the buf's data.
2825	 *
2826	 * There are two additional restrictions here because we're sharing
2827	 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2828	 * actively involved in an L2ARC write, because if this buf is used by
2829	 * an arc_write() then the hdr's data buffer will be released when the
2830	 * write completes, even though the L2ARC write might still be using it.
2831	 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2832	 * need to be ABD-aware.
2833	 */
2834	boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2835	    abd_is_linear(hdr->b_l1hdr.b_pabd);
2836
2837	/* Set up b_data and sharing */
2838	if (can_share) {
2839		buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2840		buf->b_flags |= ARC_BUF_FLAG_SHARED;
2841		arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2842	} else {
2843		buf->b_data =
2844		    arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2845		ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2846	}
2847	VERIFY3P(buf->b_data, !=, NULL);
2848
2849	hdr->b_l1hdr.b_buf = buf;
2850	hdr->b_l1hdr.b_bufcnt += 1;
2851
2852	/*
2853	 * If the user wants the data from the hdr, we need to either copy or
2854	 * decompress the data.
2855	 */
2856	if (fill) {
2857		return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
2858	}
2859
2860	return (0);
2861}
2862
2863static char *arc_onloan_tag = "onloan";
2864
2865static inline void
2866arc_loaned_bytes_update(int64_t delta)
2867{
2868	atomic_add_64(&arc_loaned_bytes, delta);
2869
2870	/* assert that it did not wrap around */
2871	ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2872}
2873
2874/*
2875 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2876 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2877 * buffers must be returned to the arc before they can be used by the DMU or
2878 * freed.
2879 */
2880arc_buf_t *
2881arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2882{
2883	arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2884	    is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2885
2886	arc_loaned_bytes_update(arc_buf_size(buf));
2887
2888	return (buf);
2889}
2890
2891arc_buf_t *
2892arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2893    enum zio_compress compression_type)
2894{
2895	arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2896	    psize, lsize, compression_type);
2897
2898	arc_loaned_bytes_update(arc_buf_size(buf));
2899
2900	return (buf);
2901}
2902
2903
2904/*
2905 * Return a loaned arc buffer to the arc.
2906 */
2907void
2908arc_return_buf(arc_buf_t *buf, void *tag)
2909{
2910	arc_buf_hdr_t *hdr = buf->b_hdr;
2911
2912	ASSERT3P(buf->b_data, !=, NULL);
2913	ASSERT(HDR_HAS_L1HDR(hdr));
2914	(void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2915	(void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2916
2917	arc_loaned_bytes_update(-arc_buf_size(buf));
2918}
2919
2920/* Detach an arc_buf from a dbuf (tag) */
2921void
2922arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2923{
2924	arc_buf_hdr_t *hdr = buf->b_hdr;
2925
2926	ASSERT3P(buf->b_data, !=, NULL);
2927	ASSERT(HDR_HAS_L1HDR(hdr));
2928	(void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2929	(void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2930
2931	arc_loaned_bytes_update(arc_buf_size(buf));
2932}
2933
2934static void
2935l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2936{
2937	l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2938
2939	df->l2df_abd = abd;
2940	df->l2df_size = size;
2941	df->l2df_type = type;
2942	mutex_enter(&l2arc_free_on_write_mtx);
2943	list_insert_head(l2arc_free_on_write, df);
2944	mutex_exit(&l2arc_free_on_write_mtx);
2945}
2946
2947static void
2948arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
2949{
2950	arc_state_t *state = hdr->b_l1hdr.b_state;
2951	arc_buf_contents_t type = arc_buf_type(hdr);
2952	uint64_t size = arc_hdr_size(hdr);
2953
2954	/* protected by hash lock, if in the hash table */
2955	if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2956		ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2957		ASSERT(state != arc_anon && state != arc_l2c_only);
2958
2959		(void) refcount_remove_many(&state->arcs_esize[type],
2960		    size, hdr);
2961	}
2962	(void) refcount_remove_many(&state->arcs_size, size, hdr);
2963	if (type == ARC_BUFC_METADATA) {
2964		arc_space_return(size, ARC_SPACE_META);
2965	} else {
2966		ASSERT(type == ARC_BUFC_DATA);
2967		arc_space_return(size, ARC_SPACE_DATA);
2968	}
2969
2970	l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2971}
2972
2973/*
2974 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2975 * data buffer, we transfer the refcount ownership to the hdr and update
2976 * the appropriate kstats.
2977 */
2978static void
2979arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2980{
2981	arc_state_t *state = hdr->b_l1hdr.b_state;
2982
2983	ASSERT(arc_can_share(hdr, buf));
2984	ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2985	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2986
2987	/*
2988	 * Start sharing the data buffer. We transfer the
2989	 * refcount ownership to the hdr since it always owns
2990	 * the refcount whenever an arc_buf_t is shared.
2991	 */
2992	refcount_transfer_ownership(&state->arcs_size, buf, hdr);
2993	hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2994	abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2995	    HDR_ISTYPE_METADATA(hdr));
2996	arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2997	buf->b_flags |= ARC_BUF_FLAG_SHARED;
2998
2999	/*
3000	 * Since we've transferred ownership to the hdr we need
3001	 * to increment its compressed and uncompressed kstats and
3002	 * decrement the overhead size.
3003	 */
3004	ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3005	ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3006	ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
3007}
3008
3009static void
3010arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3011{
3012	arc_state_t *state = hdr->b_l1hdr.b_state;
3013
3014	ASSERT(arc_buf_is_shared(buf));
3015	ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3016	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3017
3018	/*
3019	 * We are no longer sharing this buffer so we need
3020	 * to transfer its ownership to the rightful owner.
3021	 */
3022	refcount_transfer_ownership(&state->arcs_size, hdr, buf);
3023	arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3024	abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3025	abd_put(hdr->b_l1hdr.b_pabd);
3026	hdr->b_l1hdr.b_pabd = NULL;
3027	buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3028
3029	/*
3030	 * Since the buffer is no longer shared between
3031	 * the arc buf and the hdr, count it as overhead.
3032	 */
3033	ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3034	ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3035	ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3036}
3037
3038/*
3039 * Remove an arc_buf_t from the hdr's buf list and return the last
3040 * arc_buf_t on the list. If no buffers remain on the list then return
3041 * NULL.
3042 */
3043static arc_buf_t *
3044arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3045{
3046	ASSERT(HDR_HAS_L1HDR(hdr));
3047	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3048
3049	arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3050	arc_buf_t *lastbuf = NULL;
3051
3052	/*
3053	 * Remove the buf from the hdr list and locate the last
3054	 * remaining buffer on the list.
3055	 */
3056	while (*bufp != NULL) {
3057		if (*bufp == buf)
3058			*bufp = buf->b_next;
3059
3060		/*
3061		 * If we've removed a buffer in the middle of
3062		 * the list then update the lastbuf and update
3063		 * bufp.
3064		 */
3065		if (*bufp != NULL) {
3066			lastbuf = *bufp;
3067			bufp = &(*bufp)->b_next;
3068		}
3069	}
3070	buf->b_next = NULL;
3071	ASSERT3P(lastbuf, !=, buf);
3072	IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3073	IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3074	IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3075
3076	return (lastbuf);
3077}
3078
3079/*
3080 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3081 * list and free it.
3082 */
3083static void
3084arc_buf_destroy_impl(arc_buf_t *buf)
3085{
3086	arc_buf_hdr_t *hdr = buf->b_hdr;
3087
3088	/*
3089	 * Free up the data associated with the buf but only if we're not
3090	 * sharing this with the hdr. If we are sharing it with the hdr, the
3091	 * hdr is responsible for doing the free.
3092	 */
3093	if (buf->b_data != NULL) {
3094		/*
3095		 * We're about to change the hdr's b_flags. We must either
3096		 * hold the hash_lock or be undiscoverable.
3097		 */
3098		ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3099
3100		arc_cksum_verify(buf);
3101#ifdef illumos
3102		arc_buf_unwatch(buf);
3103#endif
3104
3105		if (arc_buf_is_shared(buf)) {
3106			arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3107		} else {
3108			uint64_t size = arc_buf_size(buf);
3109			arc_free_data_buf(hdr, buf->b_data, size, buf);
3110			ARCSTAT_INCR(arcstat_overhead_size, -size);
3111		}
3112		buf->b_data = NULL;
3113
3114		ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3115		hdr->b_l1hdr.b_bufcnt -= 1;
3116	}
3117
3118	arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3119
3120	if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3121		/*
3122		 * If the current arc_buf_t is sharing its data buffer with the
3123		 * hdr, then reassign the hdr's b_pabd to share it with the new
3124		 * buffer at the end of the list. The shared buffer is always
3125		 * the last one on the hdr's buffer list.
3126		 *
3127		 * There is an equivalent case for compressed bufs, but since
3128		 * they aren't guaranteed to be the last buf in the list and
3129		 * that is an exceedingly rare case, we just allow that space be
3130		 * wasted temporarily.
3131		 */
3132		if (lastbuf != NULL) {
3133			/* Only one buf can be shared at once */
3134			VERIFY(!arc_buf_is_shared(lastbuf));
3135			/* hdr is uncompressed so can't have compressed buf */
3136			VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3137
3138			ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3139			arc_hdr_free_pabd(hdr);
3140
3141			/*
3142			 * We must setup a new shared block between the
3143			 * last buffer and the hdr. The data would have
3144			 * been allocated by the arc buf so we need to transfer
3145			 * ownership to the hdr since it's now being shared.
3146			 */
3147			arc_share_buf(hdr, lastbuf);
3148		}
3149	} else if (HDR_SHARED_DATA(hdr)) {
3150		/*
3151		 * Uncompressed shared buffers are always at the end
3152		 * of the list. Compressed buffers don't have the
3153		 * same requirements. This makes it hard to
3154		 * simply assert that the lastbuf is shared so
3155		 * we rely on the hdr's compression flags to determine
3156		 * if we have a compressed, shared buffer.
3157		 */
3158		ASSERT3P(lastbuf, !=, NULL);
3159		ASSERT(arc_buf_is_shared(lastbuf) ||
3160		    HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
3161	}
3162
3163	/*
3164	 * Free the checksum if we're removing the last uncompressed buf from
3165	 * this hdr.
3166	 */
3167	if (!arc_hdr_has_uncompressed_buf(hdr)) {
3168		arc_cksum_free(hdr);
3169	}
3170
3171	/* clean up the buf */
3172	buf->b_hdr = NULL;
3173	kmem_cache_free(buf_cache, buf);
3174}
3175
3176static void
3177arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr)
3178{
3179	ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3180	ASSERT(HDR_HAS_L1HDR(hdr));
3181	ASSERT(!HDR_SHARED_DATA(hdr));
3182
3183	ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3184	hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
3185	hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3186	ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3187
3188	ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3189	ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3190}
3191
3192static void
3193arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
3194{
3195	ASSERT(HDR_HAS_L1HDR(hdr));
3196	ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3197
3198	/*
3199	 * If the hdr is currently being written to the l2arc then
3200	 * we defer freeing the data by adding it to the l2arc_free_on_write
3201	 * list. The l2arc will free the data once it's finished
3202	 * writing it to the l2arc device.
3203	 */
3204	if (HDR_L2_WRITING(hdr)) {
3205		arc_hdr_free_on_write(hdr);
3206		ARCSTAT_BUMP(arcstat_l2_free_on_write);
3207	} else {
3208		arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
3209		    arc_hdr_size(hdr), hdr);
3210	}
3211	hdr->b_l1hdr.b_pabd = NULL;
3212	hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3213
3214	ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3215	ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3216}
3217
3218static arc_buf_hdr_t *
3219arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3220    enum zio_compress compression_type, arc_buf_contents_t type)
3221{
3222	arc_buf_hdr_t *hdr;
3223
3224	VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3225
3226	hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3227	ASSERT(HDR_EMPTY(hdr));
3228	ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3229	ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
3230	HDR_SET_PSIZE(hdr, psize);
3231	HDR_SET_LSIZE(hdr, lsize);
3232	hdr->b_spa = spa;
3233	hdr->b_type = type;
3234	hdr->b_flags = 0;
3235	arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3236	arc_hdr_set_compress(hdr, compression_type);
3237
3238	hdr->b_l1hdr.b_state = arc_anon;
3239	hdr->b_l1hdr.b_arc_access = 0;
3240	hdr->b_l1hdr.b_bufcnt = 0;
3241	hdr->b_l1hdr.b_buf = NULL;
3242
3243	/*
3244	 * Allocate the hdr's buffer. This will contain either
3245	 * the compressed or uncompressed data depending on the block
3246	 * it references and compressed arc enablement.
3247	 */
3248	arc_hdr_alloc_pabd(hdr);
3249	ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3250
3251	return (hdr);
3252}
3253
3254/*
3255 * Transition between the two allocation states for the arc_buf_hdr struct.
3256 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3257 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3258 * version is used when a cache buffer is only in the L2ARC in order to reduce
3259 * memory usage.
3260 */
3261static arc_buf_hdr_t *
3262arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3263{
3264	ASSERT(HDR_HAS_L2HDR(hdr));
3265
3266	arc_buf_hdr_t *nhdr;
3267	l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3268
3269	ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3270	    (old == hdr_l2only_cache && new == hdr_full_cache));
3271
3272	nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3273
3274	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3275	buf_hash_remove(hdr);
3276
3277	bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3278
3279	if (new == hdr_full_cache) {
3280		arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3281		/*
3282		 * arc_access and arc_change_state need to be aware that a
3283		 * header has just come out of L2ARC, so we set its state to
3284		 * l2c_only even though it's about to change.
3285		 */
3286		nhdr->b_l1hdr.b_state = arc_l2c_only;
3287
3288		/* Verify previous threads set to NULL before freeing */
3289		ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3290	} else {
3291		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3292		ASSERT0(hdr->b_l1hdr.b_bufcnt);
3293		ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3294
3295		/*
3296		 * If we've reached here, We must have been called from
3297		 * arc_evict_hdr(), as such we should have already been
3298		 * removed from any ghost list we were previously on
3299		 * (which protects us from racing with arc_evict_state),
3300		 * thus no locking is needed during this check.
3301		 */
3302		ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3303
3304		/*
3305		 * A buffer must not be moved into the arc_l2c_only
3306		 * state if it's not finished being written out to the
3307		 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3308		 * might try to be accessed, even though it was removed.
3309		 */
3310		VERIFY(!HDR_L2_WRITING(hdr));
3311		VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3312
3313#ifdef ZFS_DEBUG
3314		if (hdr->b_l1hdr.b_thawed != NULL) {
3315			kmem_free(hdr->b_l1hdr.b_thawed, 1);
3316			hdr->b_l1hdr.b_thawed = NULL;
3317		}
3318#endif
3319
3320		arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3321	}
3322	/*
3323	 * The header has been reallocated so we need to re-insert it into any
3324	 * lists it was on.
3325	 */
3326	(void) buf_hash_insert(nhdr, NULL);
3327
3328	ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3329
3330	mutex_enter(&dev->l2ad_mtx);
3331
3332	/*
3333	 * We must place the realloc'ed header back into the list at
3334	 * the same spot. Otherwise, if it's placed earlier in the list,
3335	 * l2arc_write_buffers() could find it during the function's
3336	 * write phase, and try to write it out to the l2arc.
3337	 */
3338	list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3339	list_remove(&dev->l2ad_buflist, hdr);
3340
3341	mutex_exit(&dev->l2ad_mtx);
3342
3343	/*
3344	 * Since we're using the pointer address as the tag when
3345	 * incrementing and decrementing the l2ad_alloc refcount, we
3346	 * must remove the old pointer (that we're about to destroy) and
3347	 * add the new pointer to the refcount. Otherwise we'd remove
3348	 * the wrong pointer address when calling arc_hdr_destroy() later.
3349	 */
3350
3351	(void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
3352	(void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
3353
3354	buf_discard_identity(hdr);
3355	kmem_cache_free(old, hdr);
3356
3357	return (nhdr);
3358}
3359
3360/*
3361 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3362 * The buf is returned thawed since we expect the consumer to modify it.
3363 */
3364arc_buf_t *
3365arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3366{
3367	arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3368	    ZIO_COMPRESS_OFF, type);
3369	ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3370
3371	arc_buf_t *buf = NULL;
3372	VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf));
3373	arc_buf_thaw(buf);
3374
3375	return (buf);
3376}
3377
3378/*
3379 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3380 * for bufs containing metadata.
3381 */
3382arc_buf_t *
3383arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3384    enum zio_compress compression_type)
3385{
3386	ASSERT3U(lsize, >, 0);
3387	ASSERT3U(lsize, >=, psize);
3388	ASSERT(compression_type > ZIO_COMPRESS_OFF);
3389	ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3390
3391	arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3392	    compression_type, ARC_BUFC_DATA);
3393	ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3394
3395	arc_buf_t *buf = NULL;
3396	VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3397	arc_buf_thaw(buf);
3398	ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3399
3400	if (!arc_buf_is_shared(buf)) {
3401		/*
3402		 * To ensure that the hdr has the correct data in it if we call
3403		 * arc_decompress() on this buf before it's been written to
3404		 * disk, it's easiest if we just set up sharing between the
3405		 * buf and the hdr.
3406		 */
3407		ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3408		arc_hdr_free_pabd(hdr);
3409		arc_share_buf(hdr, buf);
3410	}
3411
3412	return (buf);
3413}
3414
3415static void
3416arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3417{
3418	l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3419	l2arc_dev_t *dev = l2hdr->b_dev;
3420	uint64_t psize = arc_hdr_size(hdr);
3421
3422	ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3423	ASSERT(HDR_HAS_L2HDR(hdr));
3424
3425	list_remove(&dev->l2ad_buflist, hdr);
3426
3427	ARCSTAT_INCR(arcstat_l2_psize, -psize);
3428	ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3429
3430	vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3431
3432	(void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3433	arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3434}
3435
3436static void
3437arc_hdr_destroy(arc_buf_hdr_t *hdr)
3438{
3439	if (HDR_HAS_L1HDR(hdr)) {
3440		ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3441		    hdr->b_l1hdr.b_bufcnt > 0);
3442		ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3443		ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3444	}
3445	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3446	ASSERT(!HDR_IN_HASH_TABLE(hdr));
3447
3448	if (!HDR_EMPTY(hdr))
3449		buf_discard_identity(hdr);
3450
3451	if (HDR_HAS_L2HDR(hdr)) {
3452		l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3453		boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3454
3455		if (!buflist_held)
3456			mutex_enter(&dev->l2ad_mtx);
3457
3458		/*
3459		 * Even though we checked this conditional above, we
3460		 * need to check this again now that we have the
3461		 * l2ad_mtx. This is because we could be racing with
3462		 * another thread calling l2arc_evict() which might have
3463		 * destroyed this header's L2 portion as we were waiting
3464		 * to acquire the l2ad_mtx. If that happens, we don't
3465		 * want to re-destroy the header's L2 portion.
3466		 */
3467		if (HDR_HAS_L2HDR(hdr)) {
3468			l2arc_trim(hdr);
3469			arc_hdr_l2hdr_destroy(hdr);
3470		}
3471
3472		if (!buflist_held)
3473			mutex_exit(&dev->l2ad_mtx);
3474	}
3475
3476	if (HDR_HAS_L1HDR(hdr)) {
3477		arc_cksum_free(hdr);
3478
3479		while (hdr->b_l1hdr.b_buf != NULL)
3480			arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3481
3482#ifdef ZFS_DEBUG
3483		if (hdr->b_l1hdr.b_thawed != NULL) {
3484			kmem_free(hdr->b_l1hdr.b_thawed, 1);
3485			hdr->b_l1hdr.b_thawed = NULL;
3486		}
3487#endif
3488
3489		if (hdr->b_l1hdr.b_pabd != NULL) {
3490			arc_hdr_free_pabd(hdr);
3491		}
3492	}
3493
3494	ASSERT3P(hdr->b_hash_next, ==, NULL);
3495	if (HDR_HAS_L1HDR(hdr)) {
3496		ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3497		ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3498		kmem_cache_free(hdr_full_cache, hdr);
3499	} else {
3500		kmem_cache_free(hdr_l2only_cache, hdr);
3501	}
3502}
3503
3504void
3505arc_buf_destroy(arc_buf_t *buf, void* tag)
3506{
3507	arc_buf_hdr_t *hdr = buf->b_hdr;
3508	kmutex_t *hash_lock = HDR_LOCK(hdr);
3509
3510	if (hdr->b_l1hdr.b_state == arc_anon) {
3511		ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3512		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3513		VERIFY0(remove_reference(hdr, NULL, tag));
3514		arc_hdr_destroy(hdr);
3515		return;
3516	}
3517
3518	mutex_enter(hash_lock);
3519	ASSERT3P(hdr, ==, buf->b_hdr);
3520	ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3521	ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3522	ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3523	ASSERT3P(buf->b_data, !=, NULL);
3524
3525	(void) remove_reference(hdr, hash_lock, tag);
3526	arc_buf_destroy_impl(buf);
3527	mutex_exit(hash_lock);
3528}
3529
3530/*
3531 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3532 * state of the header is dependent on its state prior to entering this
3533 * function. The following transitions are possible:
3534 *
3535 *    - arc_mru -> arc_mru_ghost
3536 *    - arc_mfu -> arc_mfu_ghost
3537 *    - arc_mru_ghost -> arc_l2c_only
3538 *    - arc_mru_ghost -> deleted
3539 *    - arc_mfu_ghost -> arc_l2c_only
3540 *    - arc_mfu_ghost -> deleted
3541 */
3542static int64_t
3543arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3544{
3545	arc_state_t *evicted_state, *state;
3546	int64_t bytes_evicted = 0;
3547
3548	ASSERT(MUTEX_HELD(hash_lock));
3549	ASSERT(HDR_HAS_L1HDR(hdr));
3550
3551	state = hdr->b_l1hdr.b_state;
3552	if (GHOST_STATE(state)) {
3553		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3554		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3555
3556		/*
3557		 * l2arc_write_buffers() relies on a header's L1 portion
3558		 * (i.e. its b_pabd field) during it's write phase.
3559		 * Thus, we cannot push a header onto the arc_l2c_only
3560		 * state (removing it's L1 piece) until the header is
3561		 * done being written to the l2arc.
3562		 */
3563		if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3564			ARCSTAT_BUMP(arcstat_evict_l2_skip);
3565			return (bytes_evicted);
3566		}
3567
3568		ARCSTAT_BUMP(arcstat_deleted);
3569		bytes_evicted += HDR_GET_LSIZE(hdr);
3570
3571		DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3572
3573		ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3574		if (HDR_HAS_L2HDR(hdr)) {
3575			/*
3576			 * This buffer is cached on the 2nd Level ARC;
3577			 * don't destroy the header.
3578			 */
3579			arc_change_state(arc_l2c_only, hdr, hash_lock);
3580			/*
3581			 * dropping from L1+L2 cached to L2-only,
3582			 * realloc to remove the L1 header.
3583			 */
3584			hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3585			    hdr_l2only_cache);
3586		} else {
3587			arc_change_state(arc_anon, hdr, hash_lock);
3588			arc_hdr_destroy(hdr);
3589		}
3590		return (bytes_evicted);
3591	}
3592
3593	ASSERT(state == arc_mru || state == arc_mfu);
3594	evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3595
3596	/* prefetch buffers have a minimum lifespan */
3597	if (HDR_IO_IN_PROGRESS(hdr) ||
3598	    ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3599	    ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3600	    arc_min_prefetch_lifespan)) {
3601		ARCSTAT_BUMP(arcstat_evict_skip);
3602		return (bytes_evicted);
3603	}
3604
3605	ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3606	while (hdr->b_l1hdr.b_buf) {
3607		arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3608		if (!mutex_tryenter(&buf->b_evict_lock)) {
3609			ARCSTAT_BUMP(arcstat_mutex_miss);
3610			break;
3611		}
3612		if (buf->b_data != NULL)
3613			bytes_evicted += HDR_GET_LSIZE(hdr);
3614		mutex_exit(&buf->b_evict_lock);
3615		arc_buf_destroy_impl(buf);
3616	}
3617
3618	if (HDR_HAS_L2HDR(hdr)) {
3619		ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3620	} else {
3621		if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3622			ARCSTAT_INCR(arcstat_evict_l2_eligible,
3623			    HDR_GET_LSIZE(hdr));
3624		} else {
3625			ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3626			    HDR_GET_LSIZE(hdr));
3627		}
3628	}
3629
3630	if (hdr->b_l1hdr.b_bufcnt == 0) {
3631		arc_cksum_free(hdr);
3632
3633		bytes_evicted += arc_hdr_size(hdr);
3634
3635		/*
3636		 * If this hdr is being evicted and has a compressed
3637		 * buffer then we discard it here before we change states.
3638		 * This ensures that the accounting is updated correctly
3639		 * in arc_free_data_impl().
3640		 */
3641		arc_hdr_free_pabd(hdr);
3642
3643		arc_change_state(evicted_state, hdr, hash_lock);
3644		ASSERT(HDR_IN_HASH_TABLE(hdr));
3645		arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3646		DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3647	}
3648
3649	return (bytes_evicted);
3650}
3651
3652static uint64_t
3653arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3654    uint64_t spa, int64_t bytes)
3655{
3656	multilist_sublist_t *mls;
3657	uint64_t bytes_evicted = 0;
3658	arc_buf_hdr_t *hdr;
3659	kmutex_t *hash_lock;
3660	int evict_count = 0;
3661
3662	ASSERT3P(marker, !=, NULL);
3663	IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3664
3665	mls = multilist_sublist_lock(ml, idx);
3666
3667	for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3668	    hdr = multilist_sublist_prev(mls, marker)) {
3669		if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3670		    (evict_count >= zfs_arc_evict_batch_limit))
3671			break;
3672
3673		/*
3674		 * To keep our iteration location, move the marker
3675		 * forward. Since we're not holding hdr's hash lock, we
3676		 * must be very careful and not remove 'hdr' from the
3677		 * sublist. Otherwise, other consumers might mistake the
3678		 * 'hdr' as not being on a sublist when they call the
3679		 * multilist_link_active() function (they all rely on
3680		 * the hash lock protecting concurrent insertions and
3681		 * removals). multilist_sublist_move_forward() was
3682		 * specifically implemented to ensure this is the case
3683		 * (only 'marker' will be removed and re-inserted).
3684		 */
3685		multilist_sublist_move_forward(mls, marker);
3686
3687		/*
3688		 * The only case where the b_spa field should ever be
3689		 * zero, is the marker headers inserted by
3690		 * arc_evict_state(). It's possible for multiple threads
3691		 * to be calling arc_evict_state() concurrently (e.g.
3692		 * dsl_pool_close() and zio_inject_fault()), so we must
3693		 * skip any markers we see from these other threads.
3694		 */
3695		if (hdr->b_spa == 0)
3696			continue;
3697
3698		/* we're only interested in evicting buffers of a certain spa */
3699		if (spa != 0 && hdr->b_spa != spa) {
3700			ARCSTAT_BUMP(arcstat_evict_skip);
3701			continue;
3702		}
3703
3704		hash_lock = HDR_LOCK(hdr);
3705
3706		/*
3707		 * We aren't calling this function from any code path
3708		 * that would already be holding a hash lock, so we're
3709		 * asserting on this assumption to be defensive in case
3710		 * this ever changes. Without this check, it would be
3711		 * possible to incorrectly increment arcstat_mutex_miss
3712		 * below (e.g. if the code changed such that we called
3713		 * this function with a hash lock held).
3714		 */
3715		ASSERT(!MUTEX_HELD(hash_lock));
3716
3717		if (mutex_tryenter(hash_lock)) {
3718			uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3719			mutex_exit(hash_lock);
3720
3721			bytes_evicted += evicted;
3722
3723			/*
3724			 * If evicted is zero, arc_evict_hdr() must have
3725			 * decided to skip this header, don't increment
3726			 * evict_count in this case.
3727			 */
3728			if (evicted != 0)
3729				evict_count++;
3730
3731			/*
3732			 * If arc_size isn't overflowing, signal any
3733			 * threads that might happen to be waiting.
3734			 *
3735			 * For each header evicted, we wake up a single
3736			 * thread. If we used cv_broadcast, we could
3737			 * wake up "too many" threads causing arc_size
3738			 * to significantly overflow arc_c; since
3739			 * arc_get_data_impl() doesn't check for overflow
3740			 * when it's woken up (it doesn't because it's
3741			 * possible for the ARC to be overflowing while
3742			 * full of un-evictable buffers, and the
3743			 * function should proceed in this case).
3744			 *
3745			 * If threads are left sleeping, due to not
3746			 * using cv_broadcast, they will be woken up
3747			 * just before arc_reclaim_thread() sleeps.
3748			 */
3749			mutex_enter(&arc_reclaim_lock);
3750			if (!arc_is_overflowing())
3751				cv_signal(&arc_reclaim_waiters_cv);
3752			mutex_exit(&arc_reclaim_lock);
3753		} else {
3754			ARCSTAT_BUMP(arcstat_mutex_miss);
3755		}
3756	}
3757
3758	multilist_sublist_unlock(mls);
3759
3760	return (bytes_evicted);
3761}
3762
3763/*
3764 * Evict buffers from the given arc state, until we've removed the
3765 * specified number of bytes. Move the removed buffers to the
3766 * appropriate evict state.
3767 *
3768 * This function makes a "best effort". It skips over any buffers
3769 * it can't get a hash_lock on, and so, may not catch all candidates.
3770 * It may also return without evicting as much space as requested.
3771 *
3772 * If bytes is specified using the special value ARC_EVICT_ALL, this
3773 * will evict all available (i.e. unlocked and evictable) buffers from
3774 * the given arc state; which is used by arc_flush().
3775 */
3776static uint64_t
3777arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3778    arc_buf_contents_t type)
3779{
3780	uint64_t total_evicted = 0;
3781	multilist_t *ml = state->arcs_list[type];
3782	int num_sublists;
3783	arc_buf_hdr_t **markers;
3784
3785	IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3786
3787	num_sublists = multilist_get_num_sublists(ml);
3788
3789	/*
3790	 * If we've tried to evict from each sublist, made some
3791	 * progress, but still have not hit the target number of bytes
3792	 * to evict, we want to keep trying. The markers allow us to
3793	 * pick up where we left off for each individual sublist, rather
3794	 * than starting from the tail each time.
3795	 */
3796	markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3797	for (int i = 0; i < num_sublists; i++) {
3798		markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3799
3800		/*
3801		 * A b_spa of 0 is used to indicate that this header is
3802		 * a marker. This fact is used in arc_adjust_type() and
3803		 * arc_evict_state_impl().
3804		 */
3805		markers[i]->b_spa = 0;
3806
3807		multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3808		multilist_sublist_insert_tail(mls, markers[i]);
3809		multilist_sublist_unlock(mls);
3810	}
3811
3812	/*
3813	 * While we haven't hit our target number of bytes to evict, or
3814	 * we're evicting all available buffers.
3815	 */
3816	while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3817		/*
3818		 * Start eviction using a randomly selected sublist,
3819		 * this is to try and evenly balance eviction across all
3820		 * sublists. Always starting at the same sublist
3821		 * (e.g. index 0) would cause evictions to favor certain
3822		 * sublists over others.
3823		 */
3824		int sublist_idx = multilist_get_random_index(ml);
3825		uint64_t scan_evicted = 0;
3826
3827		for (int i = 0; i < num_sublists; i++) {
3828			uint64_t bytes_remaining;
3829			uint64_t bytes_evicted;
3830
3831			if (bytes == ARC_EVICT_ALL)
3832				bytes_remaining = ARC_EVICT_ALL;
3833			else if (total_evicted < bytes)
3834				bytes_remaining = bytes - total_evicted;
3835			else
3836				break;
3837
3838			bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3839			    markers[sublist_idx], spa, bytes_remaining);
3840
3841			scan_evicted += bytes_evicted;
3842			total_evicted += bytes_evicted;
3843
3844			/* we've reached the end, wrap to the beginning */
3845			if (++sublist_idx >= num_sublists)
3846				sublist_idx = 0;
3847		}
3848
3849		/*
3850		 * If we didn't evict anything during this scan, we have
3851		 * no reason to believe we'll evict more during another
3852		 * scan, so break the loop.
3853		 */
3854		if (scan_evicted == 0) {
3855			/* This isn't possible, let's make that obvious */
3856			ASSERT3S(bytes, !=, 0);
3857
3858			/*
3859			 * When bytes is ARC_EVICT_ALL, the only way to
3860			 * break the loop is when scan_evicted is zero.
3861			 * In that case, we actually have evicted enough,
3862			 * so we don't want to increment the kstat.
3863			 */
3864			if (bytes != ARC_EVICT_ALL) {
3865				ASSERT3S(total_evicted, <, bytes);
3866				ARCSTAT_BUMP(arcstat_evict_not_enough);
3867			}
3868
3869			break;
3870		}
3871	}
3872
3873	for (int i = 0; i < num_sublists; i++) {
3874		multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3875		multilist_sublist_remove(mls, markers[i]);
3876		multilist_sublist_unlock(mls);
3877
3878		kmem_cache_free(hdr_full_cache, markers[i]);
3879	}
3880	kmem_free(markers, sizeof (*markers) * num_sublists);
3881
3882	return (total_evicted);
3883}
3884
3885/*
3886 * Flush all "evictable" data of the given type from the arc state
3887 * specified. This will not evict any "active" buffers (i.e. referenced).
3888 *
3889 * When 'retry' is set to B_FALSE, the function will make a single pass
3890 * over the state and evict any buffers that it can. Since it doesn't
3891 * continually retry the eviction, it might end up leaving some buffers
3892 * in the ARC due to lock misses.
3893 *
3894 * When 'retry' is set to B_TRUE, the function will continually retry the
3895 * eviction until *all* evictable buffers have been removed from the
3896 * state. As a result, if concurrent insertions into the state are
3897 * allowed (e.g. if the ARC isn't shutting down), this function might
3898 * wind up in an infinite loop, continually trying to evict buffers.
3899 */
3900static uint64_t
3901arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
3902    boolean_t retry)
3903{
3904	uint64_t evicted = 0;
3905
3906	while (refcount_count(&state->arcs_esize[type]) != 0) {
3907		evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
3908
3909		if (!retry)
3910			break;
3911	}
3912
3913	return (evicted);
3914}
3915
3916/*
3917 * Evict the specified number of bytes from the state specified,
3918 * restricting eviction to the spa and type given. This function
3919 * prevents us from trying to evict more from a state's list than
3920 * is "evictable", and to skip evicting altogether when passed a
3921 * negative value for "bytes". In contrast, arc_evict_state() will
3922 * evict everything it can, when passed a negative value for "bytes".
3923 */
3924static uint64_t
3925arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3926    arc_buf_contents_t type)
3927{
3928	int64_t delta;
3929
3930	if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
3931		delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
3932		return (arc_evict_state(state, spa, delta, type));
3933	}
3934
3935	return (0);
3936}
3937
3938/*
3939 * Evict metadata buffers from the cache, such that arc_meta_used is
3940 * capped by the arc_meta_limit tunable.
3941 */
3942static uint64_t
3943arc_adjust_meta(uint64_t meta_used)
3944{
3945	uint64_t total_evicted = 0;
3946	int64_t target;
3947
3948	/*
3949	 * If we're over the meta limit, we want to evict enough
3950	 * metadata to get back under the meta limit. We don't want to
3951	 * evict so much that we drop the MRU below arc_p, though. If
3952	 * we're over the meta limit more than we're over arc_p, we
3953	 * evict some from the MRU here, and some from the MFU below.
3954	 */
3955	target = MIN((int64_t)(meta_used - arc_meta_limit),
3956	    (int64_t)(refcount_count(&arc_anon->arcs_size) +
3957	    refcount_count(&arc_mru->arcs_size) - arc_p));
3958
3959	total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3960
3961	/*
3962	 * Similar to the above, we want to evict enough bytes to get us
3963	 * below the meta limit, but not so much as to drop us below the
3964	 * space allotted to the MFU (which is defined as arc_c - arc_p).
3965	 */
3966	target = MIN((int64_t)(meta_used - arc_meta_limit),
3967	    (int64_t)(refcount_count(&arc_mfu->arcs_size) -
3968	    (arc_c - arc_p)));
3969
3970	total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3971
3972	return (total_evicted);
3973}
3974
3975/*
3976 * Return the type of the oldest buffer in the given arc state
3977 *
3978 * This function will select a random sublist of type ARC_BUFC_DATA and
3979 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3980 * is compared, and the type which contains the "older" buffer will be
3981 * returned.
3982 */
3983static arc_buf_contents_t
3984arc_adjust_type(arc_state_t *state)
3985{
3986	multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
3987	multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
3988	int data_idx = multilist_get_random_index(data_ml);
3989	int meta_idx = multilist_get_random_index(meta_ml);
3990	multilist_sublist_t *data_mls;
3991	multilist_sublist_t *meta_mls;
3992	arc_buf_contents_t type;
3993	arc_buf_hdr_t *data_hdr;
3994	arc_buf_hdr_t *meta_hdr;
3995
3996	/*
3997	 * We keep the sublist lock until we're finished, to prevent
3998	 * the headers from being destroyed via arc_evict_state().
3999	 */
4000	data_mls = multilist_sublist_lock(data_ml, data_idx);
4001	meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4002
4003	/*
4004	 * These two loops are to ensure we skip any markers that
4005	 * might be at the tail of the lists due to arc_evict_state().
4006	 */
4007
4008	for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4009	    data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4010		if (data_hdr->b_spa != 0)
4011			break;
4012	}
4013
4014	for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4015	    meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4016		if (meta_hdr->b_spa != 0)
4017			break;
4018	}
4019
4020	if (data_hdr == NULL && meta_hdr == NULL) {
4021		type = ARC_BUFC_DATA;
4022	} else if (data_hdr == NULL) {
4023		ASSERT3P(meta_hdr, !=, NULL);
4024		type = ARC_BUFC_METADATA;
4025	} else if (meta_hdr == NULL) {
4026		ASSERT3P(data_hdr, !=, NULL);
4027		type = ARC_BUFC_DATA;
4028	} else {
4029		ASSERT3P(data_hdr, !=, NULL);
4030		ASSERT3P(meta_hdr, !=, NULL);
4031
4032		/* The headers can't be on the sublist without an L1 header */
4033		ASSERT(HDR_HAS_L1HDR(data_hdr));
4034		ASSERT(HDR_HAS_L1HDR(meta_hdr));
4035
4036		if (data_hdr->b_l1hdr.b_arc_access <
4037		    meta_hdr->b_l1hdr.b_arc_access) {
4038			type = ARC_BUFC_DATA;
4039		} else {
4040			type = ARC_BUFC_METADATA;
4041		}
4042	}
4043
4044	multilist_sublist_unlock(meta_mls);
4045	multilist_sublist_unlock(data_mls);
4046
4047	return (type);
4048}
4049
4050/*
4051 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4052 */
4053static uint64_t
4054arc_adjust(void)
4055{
4056	uint64_t total_evicted = 0;
4057	uint64_t bytes;
4058	int64_t target;
4059	uint64_t asize = aggsum_value(&arc_size);
4060	uint64_t ameta = aggsum_value(&arc_meta_used);
4061
4062	/*
4063	 * If we're over arc_meta_limit, we want to correct that before
4064	 * potentially evicting data buffers below.
4065	 */
4066	total_evicted += arc_adjust_meta(ameta);
4067
4068	/*
4069	 * Adjust MRU size
4070	 *
4071	 * If we're over the target cache size, we want to evict enough
4072	 * from the list to get back to our target size. We don't want
4073	 * to evict too much from the MRU, such that it drops below
4074	 * arc_p. So, if we're over our target cache size more than
4075	 * the MRU is over arc_p, we'll evict enough to get back to
4076	 * arc_p here, and then evict more from the MFU below.
4077	 */
4078	target = MIN((int64_t)(asize - arc_c),
4079	    (int64_t)(refcount_count(&arc_anon->arcs_size) +
4080	    refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4081
4082	/*
4083	 * If we're below arc_meta_min, always prefer to evict data.
4084	 * Otherwise, try to satisfy the requested number of bytes to
4085	 * evict from the type which contains older buffers; in an
4086	 * effort to keep newer buffers in the cache regardless of their
4087	 * type. If we cannot satisfy the number of bytes from this
4088	 * type, spill over into the next type.
4089	 */
4090	if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4091	    ameta > arc_meta_min) {
4092		bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4093		total_evicted += bytes;
4094
4095		/*
4096		 * If we couldn't evict our target number of bytes from
4097		 * metadata, we try to get the rest from data.
4098		 */
4099		target -= bytes;
4100
4101		total_evicted +=
4102		    arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4103	} else {
4104		bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4105		total_evicted += bytes;
4106
4107		/*
4108		 * If we couldn't evict our target number of bytes from
4109		 * data, we try to get the rest from metadata.
4110		 */
4111		target -= bytes;
4112
4113		total_evicted +=
4114		    arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4115	}
4116
4117	/*
4118	 * Adjust MFU size
4119	 *
4120	 * Now that we've tried to evict enough from the MRU to get its
4121	 * size back to arc_p, if we're still above the target cache
4122	 * size, we evict the rest from the MFU.
4123	 */
4124	target = asize - arc_c;
4125
4126	if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4127	    ameta > arc_meta_min) {
4128		bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4129		total_evicted += bytes;
4130
4131		/*
4132		 * If we couldn't evict our target number of bytes from
4133		 * metadata, we try to get the rest from data.
4134		 */
4135		target -= bytes;
4136
4137		total_evicted +=
4138		    arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4139	} else {
4140		bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4141		total_evicted += bytes;
4142
4143		/*
4144		 * If we couldn't evict our target number of bytes from
4145		 * data, we try to get the rest from data.
4146		 */
4147		target -= bytes;
4148
4149		total_evicted +=
4150		    arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4151	}
4152
4153	/*
4154	 * Adjust ghost lists
4155	 *
4156	 * In addition to the above, the ARC also defines target values
4157	 * for the ghost lists. The sum of the mru list and mru ghost
4158	 * list should never exceed the target size of the cache, and
4159	 * the sum of the mru list, mfu list, mru ghost list, and mfu
4160	 * ghost list should never exceed twice the target size of the
4161	 * cache. The following logic enforces these limits on the ghost
4162	 * caches, and evicts from them as needed.
4163	 */
4164	target = refcount_count(&arc_mru->arcs_size) +
4165	    refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4166
4167	bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4168	total_evicted += bytes;
4169
4170	target -= bytes;
4171
4172	total_evicted +=
4173	    arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4174
4175	/*
4176	 * We assume the sum of the mru list and mfu list is less than
4177	 * or equal to arc_c (we enforced this above), which means we
4178	 * can use the simpler of the two equations below:
4179	 *
4180	 *	mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4181	 *		    mru ghost + mfu ghost <= arc_c
4182	 */
4183	target = refcount_count(&arc_mru_ghost->arcs_size) +
4184	    refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4185
4186	bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4187	total_evicted += bytes;
4188
4189	target -= bytes;
4190
4191	total_evicted +=
4192	    arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4193
4194	return (total_evicted);
4195}
4196
4197void
4198arc_flush(spa_t *spa, boolean_t retry)
4199{
4200	uint64_t guid = 0;
4201
4202	/*
4203	 * If retry is B_TRUE, a spa must not be specified since we have
4204	 * no good way to determine if all of a spa's buffers have been
4205	 * evicted from an arc state.
4206	 */
4207	ASSERT(!retry || spa == 0);
4208
4209	if (spa != NULL)
4210		guid = spa_load_guid(spa);
4211
4212	(void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4213	(void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4214
4215	(void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4216	(void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4217
4218	(void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4219	(void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4220
4221	(void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4222	(void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4223}
4224
4225void
4226arc_shrink(int64_t to_free)
4227{
4228	uint64_t asize = aggsum_value(&arc_size);
4229	if (arc_c > arc_c_min) {
4230		DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
4231			arc_c_min, uint64_t, arc_p, uint64_t, to_free);
4232		if (arc_c > arc_c_min + to_free)
4233			atomic_add_64(&arc_c, -to_free);
4234		else
4235			arc_c = arc_c_min;
4236
4237		atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4238		if (asize < arc_c)
4239			arc_c = MAX(asize, arc_c_min);
4240		if (arc_p > arc_c)
4241			arc_p = (arc_c >> 1);
4242
4243		DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
4244			arc_p);
4245
4246		ASSERT(arc_c >= arc_c_min);
4247		ASSERT((int64_t)arc_p >= 0);
4248	}
4249
4250	if (asize > arc_c) {
4251		DTRACE_PROBE2(arc__shrink_adjust, uint64_t, asize,
4252			uint64_t, arc_c);
4253		(void) arc_adjust();
4254	}
4255}
4256
4257typedef enum free_memory_reason_t {
4258	FMR_UNKNOWN,
4259	FMR_NEEDFREE,
4260	FMR_LOTSFREE,
4261	FMR_SWAPFS_MINFREE,
4262	FMR_PAGES_PP_MAXIMUM,
4263	FMR_HEAP_ARENA,
4264	FMR_ZIO_ARENA,
4265} free_memory_reason_t;
4266
4267int64_t last_free_memory;
4268free_memory_reason_t last_free_reason;
4269
4270/*
4271 * Additional reserve of pages for pp_reserve.
4272 */
4273int64_t arc_pages_pp_reserve = 64;
4274
4275/*
4276 * Additional reserve of pages for swapfs.
4277 */
4278int64_t arc_swapfs_reserve = 64;
4279
4280/*
4281 * Return the amount of memory that can be consumed before reclaim will be
4282 * needed.  Positive if there is sufficient free memory, negative indicates
4283 * the amount of memory that needs to be freed up.
4284 */
4285static int64_t
4286arc_available_memory(void)
4287{
4288	int64_t lowest = INT64_MAX;
4289	int64_t n;
4290	free_memory_reason_t r = FMR_UNKNOWN;
4291
4292#ifdef _KERNEL
4293#ifdef __FreeBSD__
4294	/*
4295	 * Cooperate with pagedaemon when it's time for it to scan
4296	 * and reclaim some pages.
4297	 */
4298	n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
4299	if (n < lowest) {
4300		lowest = n;
4301		r = FMR_LOTSFREE;
4302	}
4303
4304#else
4305	if (needfree > 0) {
4306		n = PAGESIZE * (-needfree);
4307		if (n < lowest) {
4308			lowest = n;
4309			r = FMR_NEEDFREE;
4310		}
4311	}
4312
4313	/*
4314	 * check that we're out of range of the pageout scanner.  It starts to
4315	 * schedule paging if freemem is less than lotsfree and needfree.
4316	 * lotsfree is the high-water mark for pageout, and needfree is the
4317	 * number of needed free pages.  We add extra pages here to make sure
4318	 * the scanner doesn't start up while we're freeing memory.
4319	 */
4320	n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4321	if (n < lowest) {
4322		lowest = n;
4323		r = FMR_LOTSFREE;
4324	}
4325
4326	/*
4327	 * check to make sure that swapfs has enough space so that anon
4328	 * reservations can still succeed. anon_resvmem() checks that the
4329	 * availrmem is greater than swapfs_minfree, and the number of reserved
4330	 * swap pages.  We also add a bit of extra here just to prevent
4331	 * circumstances from getting really dire.
4332	 */
4333	n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4334	    desfree - arc_swapfs_reserve);
4335	if (n < lowest) {
4336		lowest = n;
4337		r = FMR_SWAPFS_MINFREE;
4338	}
4339
4340
4341	/*
4342	 * Check that we have enough availrmem that memory locking (e.g., via
4343	 * mlock(3C) or memcntl(2)) can still succeed.  (pages_pp_maximum
4344	 * stores the number of pages that cannot be locked; when availrmem
4345	 * drops below pages_pp_maximum, page locking mechanisms such as
4346	 * page_pp_lock() will fail.)
4347	 */
4348	n = PAGESIZE * (availrmem - pages_pp_maximum -
4349	    arc_pages_pp_reserve);
4350	if (n < lowest) {
4351		lowest = n;
4352		r = FMR_PAGES_PP_MAXIMUM;
4353	}
4354
4355#endif	/* __FreeBSD__ */
4356#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
4357	/*
4358	 * If we're on an i386 platform, it's possible that we'll exhaust the
4359	 * kernel heap space before we ever run out of available physical
4360	 * memory.  Most checks of the size of the heap_area compare against
4361	 * tune.t_minarmem, which is the minimum available real memory that we
4362	 * can have in the system.  However, this is generally fixed at 25 pages
4363	 * which is so low that it's useless.  In this comparison, we seek to
4364	 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4365	 * heap is allocated.  (Or, in the calculation, if less than 1/4th is
4366	 * free)
4367	 */
4368	n = uma_avail() - (long)(uma_limit() / 4);
4369	if (n < lowest) {
4370		lowest = n;
4371		r = FMR_HEAP_ARENA;
4372	}
4373#endif
4374
4375	/*
4376	 * If zio data pages are being allocated out of a separate heap segment,
4377	 * then enforce that the size of available vmem for this arena remains
4378	 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4379	 *
4380	 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4381	 * memory (in the zio_arena) free, which can avoid memory
4382	 * fragmentation issues.
4383	 */
4384	if (zio_arena != NULL) {
4385		n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4386		    (vmem_size(zio_arena, VMEM_ALLOC) >>
4387		    arc_zio_arena_free_shift);
4388		if (n < lowest) {
4389			lowest = n;
4390			r = FMR_ZIO_ARENA;
4391		}
4392	}
4393
4394#else	/* _KERNEL */
4395	/* Every 100 calls, free a small amount */
4396	if (spa_get_random(100) == 0)
4397		lowest = -1024;
4398#endif	/* _KERNEL */
4399
4400	last_free_memory = lowest;
4401	last_free_reason = r;
4402	DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
4403	return (lowest);
4404}
4405
4406
4407/*
4408 * Determine if the system is under memory pressure and is asking
4409 * to reclaim memory. A return value of B_TRUE indicates that the system
4410 * is under memory pressure and that the arc should adjust accordingly.
4411 */
4412static boolean_t
4413arc_reclaim_needed(void)
4414{
4415	return (arc_available_memory() < 0);
4416}
4417
4418extern kmem_cache_t	*zio_buf_cache[];
4419extern kmem_cache_t	*zio_data_buf_cache[];
4420extern kmem_cache_t	*range_seg_cache;
4421extern kmem_cache_t	*abd_chunk_cache;
4422
4423static __noinline void
4424arc_kmem_reap_now(void)
4425{
4426	size_t			i;
4427	kmem_cache_t		*prev_cache = NULL;
4428	kmem_cache_t		*prev_data_cache = NULL;
4429
4430	DTRACE_PROBE(arc__kmem_reap_start);
4431#ifdef _KERNEL
4432	if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) {
4433		/*
4434		 * We are exceeding our meta-data cache limit.
4435		 * Purge some DNLC entries to release holds on meta-data.
4436		 */
4437		dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4438	}
4439#if defined(__i386)
4440	/*
4441	 * Reclaim unused memory from all kmem caches.
4442	 */
4443	kmem_reap();
4444#endif
4445#endif
4446
4447	/*
4448	 * If a kmem reap is already active, don't schedule more.  We must
4449	 * check for this because kmem_cache_reap_soon() won't actually
4450	 * block on the cache being reaped (this is to prevent callers from
4451	 * becoming implicitly blocked by a system-wide kmem reap -- which,
4452	 * on a system with many, many full magazines, can take minutes).
4453	 */
4454	if (kmem_cache_reap_active())
4455		return;
4456
4457	for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4458		if (zio_buf_cache[i] != prev_cache) {
4459			prev_cache = zio_buf_cache[i];
4460			kmem_cache_reap_soon(zio_buf_cache[i]);
4461		}
4462		if (zio_data_buf_cache[i] != prev_data_cache) {
4463			prev_data_cache = zio_data_buf_cache[i];
4464			kmem_cache_reap_soon(zio_data_buf_cache[i]);
4465		}
4466	}
4467	kmem_cache_reap_soon(abd_chunk_cache);
4468	kmem_cache_reap_soon(buf_cache);
4469	kmem_cache_reap_soon(hdr_full_cache);
4470	kmem_cache_reap_soon(hdr_l2only_cache);
4471	kmem_cache_reap_soon(range_seg_cache);
4472
4473#ifdef illumos
4474	if (zio_arena != NULL) {
4475		/*
4476		 * Ask the vmem arena to reclaim unused memory from its
4477		 * quantum caches.
4478		 */
4479		vmem_qcache_reap(zio_arena);
4480	}
4481#endif
4482	DTRACE_PROBE(arc__kmem_reap_end);
4483}
4484
4485/*
4486 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4487 * enough data and signal them to proceed. When this happens, the threads in
4488 * arc_get_data_impl() are sleeping while holding the hash lock for their
4489 * particular arc header. Thus, we must be careful to never sleep on a
4490 * hash lock in this thread. This is to prevent the following deadlock:
4491 *
4492 *  - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4493 *    waiting for the reclaim thread to signal it.
4494 *
4495 *  - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4496 *    fails, and goes to sleep forever.
4497 *
4498 * This possible deadlock is avoided by always acquiring a hash lock
4499 * using mutex_tryenter() from arc_reclaim_thread().
4500 */
4501/* ARGSUSED */
4502static void
4503arc_reclaim_thread(void *unused __unused)
4504{
4505	hrtime_t		growtime = 0;
4506	hrtime_t		kmem_reap_time = 0;
4507	callb_cpr_t		cpr;
4508
4509	CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4510
4511	mutex_enter(&arc_reclaim_lock);
4512	while (!arc_reclaim_thread_exit) {
4513		uint64_t evicted = 0;
4514
4515		/*
4516		 * This is necessary in order for the mdb ::arc dcmd to
4517		 * show up to date information. Since the ::arc command
4518		 * does not call the kstat's update function, without
4519		 * this call, the command may show stale stats for the
4520		 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4521		 * with this change, the data might be up to 1 second
4522		 * out of date; but that should suffice. The arc_state_t
4523		 * structures can be queried directly if more accurate
4524		 * information is needed.
4525		 */
4526		if (arc_ksp != NULL)
4527			arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4528
4529		mutex_exit(&arc_reclaim_lock);
4530
4531		/*
4532		 * We call arc_adjust() before (possibly) calling
4533		 * arc_kmem_reap_now(), so that we can wake up
4534		 * arc_get_data_impl() sooner.
4535		 */
4536		evicted = arc_adjust();
4537
4538		int64_t free_memory = arc_available_memory();
4539		if (free_memory < 0) {
4540			hrtime_t curtime = gethrtime();
4541			arc_no_grow = B_TRUE;
4542			arc_warm = B_TRUE;
4543
4544			/*
4545			 * Wait at least zfs_grow_retry (default 60) seconds
4546			 * before considering growing.
4547			 */
4548			growtime = curtime + SEC2NSEC(arc_grow_retry);
4549
4550			/*
4551			 * Wait at least arc_kmem_cache_reap_retry_ms
4552			 * between arc_kmem_reap_now() calls. Without
4553			 * this check it is possible to end up in a
4554			 * situation where we spend lots of time
4555			 * reaping caches, while we're near arc_c_min.
4556			 */
4557			if (curtime >= kmem_reap_time) {
4558				arc_kmem_reap_now();
4559				kmem_reap_time = gethrtime() +
4560				    MSEC2NSEC(arc_kmem_cache_reap_retry_ms);
4561			}
4562
4563			/*
4564			 * If we are still low on memory, shrink the ARC
4565			 * so that we have arc_shrink_min free space.
4566			 */
4567			free_memory = arc_available_memory();
4568
4569			int64_t to_free =
4570			    (arc_c >> arc_shrink_shift) - free_memory;
4571			if (to_free > 0) {
4572#ifdef _KERNEL
4573#ifdef illumos
4574				to_free = MAX(to_free, ptob(needfree));
4575#endif
4576#endif
4577				arc_shrink(to_free);
4578			}
4579		} else if (free_memory < arc_c >> arc_no_grow_shift) {
4580			arc_no_grow = B_TRUE;
4581		} else if (gethrtime() >= growtime) {
4582			arc_no_grow = B_FALSE;
4583		}
4584
4585		mutex_enter(&arc_reclaim_lock);
4586
4587		/*
4588		 * If evicted is zero, we couldn't evict anything via
4589		 * arc_adjust(). This could be due to hash lock
4590		 * collisions, but more likely due to the majority of
4591		 * arc buffers being unevictable. Therefore, even if
4592		 * arc_size is above arc_c, another pass is unlikely to
4593		 * be helpful and could potentially cause us to enter an
4594		 * infinite loop.
4595		 */
4596		if (aggsum_compare(&arc_size, arc_c) <= 0|| evicted == 0) {
4597			/*
4598			 * We're either no longer overflowing, or we
4599			 * can't evict anything more, so we should wake
4600			 * up any threads before we go to sleep.
4601			 */
4602			cv_broadcast(&arc_reclaim_waiters_cv);
4603
4604			/*
4605			 * Block until signaled, or after one second (we
4606			 * might need to perform arc_kmem_reap_now()
4607			 * even if we aren't being signalled)
4608			 */
4609			CALLB_CPR_SAFE_BEGIN(&cpr);
4610			(void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4611			    &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4612			CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4613		}
4614	}
4615
4616	arc_reclaim_thread_exit = B_FALSE;
4617	cv_broadcast(&arc_reclaim_thread_cv);
4618	CALLB_CPR_EXIT(&cpr);		/* drops arc_reclaim_lock */
4619	thread_exit();
4620}
4621
4622static u_int arc_dnlc_evicts_arg;
4623extern struct vfsops zfs_vfsops;
4624
4625static void
4626arc_dnlc_evicts_thread(void *dummy __unused)
4627{
4628	callb_cpr_t cpr;
4629	u_int percent;
4630
4631	CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG);
4632
4633	mutex_enter(&arc_dnlc_evicts_lock);
4634	while (!arc_dnlc_evicts_thread_exit) {
4635		CALLB_CPR_SAFE_BEGIN(&cpr);
4636		(void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
4637		CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock);
4638		if (arc_dnlc_evicts_arg != 0) {
4639			percent = arc_dnlc_evicts_arg;
4640			mutex_exit(&arc_dnlc_evicts_lock);
4641#ifdef _KERNEL
4642			vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops);
4643#endif
4644			mutex_enter(&arc_dnlc_evicts_lock);
4645			/*
4646			 * Clear our token only after vnlru_free()
4647			 * pass is done, to avoid false queueing of
4648			 * the requests.
4649			 */
4650			arc_dnlc_evicts_arg = 0;
4651		}
4652	}
4653	arc_dnlc_evicts_thread_exit = FALSE;
4654	cv_broadcast(&arc_dnlc_evicts_cv);
4655	CALLB_CPR_EXIT(&cpr);
4656	thread_exit();
4657}
4658
4659void
4660dnlc_reduce_cache(void *arg)
4661{
4662	u_int percent;
4663
4664	percent = (u_int)(uintptr_t)arg;
4665	mutex_enter(&arc_dnlc_evicts_lock);
4666	if (arc_dnlc_evicts_arg == 0) {
4667		arc_dnlc_evicts_arg = percent;
4668		cv_broadcast(&arc_dnlc_evicts_cv);
4669	}
4670	mutex_exit(&arc_dnlc_evicts_lock);
4671}
4672
4673/*
4674 * Adapt arc info given the number of bytes we are trying to add and
4675 * the state that we are comming from.  This function is only called
4676 * when we are adding new content to the cache.
4677 */
4678static void
4679arc_adapt(int bytes, arc_state_t *state)
4680{
4681	int mult;
4682	uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4683	int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
4684	int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
4685
4686	if (state == arc_l2c_only)
4687		return;
4688
4689	ASSERT(bytes > 0);
4690	/*
4691	 * Adapt the target size of the MRU list:
4692	 *	- if we just hit in the MRU ghost list, then increase
4693	 *	  the target size of the MRU list.
4694	 *	- if we just hit in the MFU ghost list, then increase
4695	 *	  the target size of the MFU list by decreasing the
4696	 *	  target size of the MRU list.
4697	 */
4698	if (state == arc_mru_ghost) {
4699		mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4700		mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4701
4702		arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4703	} else if (state == arc_mfu_ghost) {
4704		uint64_t delta;
4705
4706		mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4707		mult = MIN(mult, 10);
4708
4709		delta = MIN(bytes * mult, arc_p);
4710		arc_p = MAX(arc_p_min, arc_p - delta);
4711	}
4712	ASSERT((int64_t)arc_p >= 0);
4713
4714	if (arc_reclaim_needed()) {
4715		cv_signal(&arc_reclaim_thread_cv);
4716		return;
4717	}
4718
4719	if (arc_no_grow)
4720		return;
4721
4722	if (arc_c >= arc_c_max)
4723		return;
4724
4725	/*
4726	 * If we're within (2 * maxblocksize) bytes of the target
4727	 * cache size, increment the target cache size
4728	 */
4729	if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >
4730	    0) {
4731		DTRACE_PROBE1(arc__inc_adapt, int, bytes);
4732		atomic_add_64(&arc_c, (int64_t)bytes);
4733		if (arc_c > arc_c_max)
4734			arc_c = arc_c_max;
4735		else if (state == arc_anon)
4736			atomic_add_64(&arc_p, (int64_t)bytes);
4737		if (arc_p > arc_c)
4738			arc_p = arc_c;
4739	}
4740	ASSERT((int64_t)arc_p >= 0);
4741}
4742
4743/*
4744 * Check if arc_size has grown past our upper threshold, determined by
4745 * zfs_arc_overflow_shift.
4746 */
4747static boolean_t
4748arc_is_overflowing(void)
4749{
4750	/* Always allow at least one block of overflow */
4751	uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4752	    arc_c >> zfs_arc_overflow_shift);
4753
4754	/*
4755	 * We just compare the lower bound here for performance reasons. Our
4756	 * primary goals are to make sure that the arc never grows without
4757	 * bound, and that it can reach its maximum size. This check
4758	 * accomplishes both goals. The maximum amount we could run over by is
4759	 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
4760	 * in the ARC. In practice, that's in the tens of MB, which is low
4761	 * enough to be safe.
4762	 */
4763	return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
4764}
4765
4766static abd_t *
4767arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4768{
4769	arc_buf_contents_t type = arc_buf_type(hdr);
4770
4771	arc_get_data_impl(hdr, size, tag);
4772	if (type == ARC_BUFC_METADATA) {
4773		return (abd_alloc(size, B_TRUE));
4774	} else {
4775		ASSERT(type == ARC_BUFC_DATA);
4776		return (abd_alloc(size, B_FALSE));
4777	}
4778}
4779
4780static void *
4781arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4782{
4783	arc_buf_contents_t type = arc_buf_type(hdr);
4784
4785	arc_get_data_impl(hdr, size, tag);
4786	if (type == ARC_BUFC_METADATA) {
4787		return (zio_buf_alloc(size));
4788	} else {
4789		ASSERT(type == ARC_BUFC_DATA);
4790		return (zio_data_buf_alloc(size));
4791	}
4792}
4793
4794/*
4795 * Allocate a block and return it to the caller. If we are hitting the
4796 * hard limit for the cache size, we must sleep, waiting for the eviction
4797 * thread to catch up. If we're past the target size but below the hard
4798 * limit, we'll only signal the reclaim thread and continue on.
4799 */
4800static void
4801arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4802{
4803	arc_state_t *state = hdr->b_l1hdr.b_state;
4804	arc_buf_contents_t type = arc_buf_type(hdr);
4805
4806	arc_adapt(size, state);
4807
4808	/*
4809	 * If arc_size is currently overflowing, and has grown past our
4810	 * upper limit, we must be adding data faster than the evict
4811	 * thread can evict. Thus, to ensure we don't compound the
4812	 * problem by adding more data and forcing arc_size to grow even
4813	 * further past it's target size, we halt and wait for the
4814	 * eviction thread to catch up.
4815	 *
4816	 * It's also possible that the reclaim thread is unable to evict
4817	 * enough buffers to get arc_size below the overflow limit (e.g.
4818	 * due to buffers being un-evictable, or hash lock collisions).
4819	 * In this case, we want to proceed regardless if we're
4820	 * overflowing; thus we don't use a while loop here.
4821	 */
4822	if (arc_is_overflowing()) {
4823		mutex_enter(&arc_reclaim_lock);
4824
4825		/*
4826		 * Now that we've acquired the lock, we may no longer be
4827		 * over the overflow limit, lets check.
4828		 *
4829		 * We're ignoring the case of spurious wake ups. If that
4830		 * were to happen, it'd let this thread consume an ARC
4831		 * buffer before it should have (i.e. before we're under
4832		 * the overflow limit and were signalled by the reclaim
4833		 * thread). As long as that is a rare occurrence, it
4834		 * shouldn't cause any harm.
4835		 */
4836		if (arc_is_overflowing()) {
4837			cv_signal(&arc_reclaim_thread_cv);
4838			cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
4839		}
4840
4841		mutex_exit(&arc_reclaim_lock);
4842	}
4843
4844	VERIFY3U(hdr->b_type, ==, type);
4845	if (type == ARC_BUFC_METADATA) {
4846		arc_space_consume(size, ARC_SPACE_META);
4847	} else {
4848		arc_space_consume(size, ARC_SPACE_DATA);
4849	}
4850
4851	/*
4852	 * Update the state size.  Note that ghost states have a
4853	 * "ghost size" and so don't need to be updated.
4854	 */
4855	if (!GHOST_STATE(state)) {
4856
4857		(void) refcount_add_many(&state->arcs_size, size, tag);
4858
4859		/*
4860		 * If this is reached via arc_read, the link is
4861		 * protected by the hash lock. If reached via
4862		 * arc_buf_alloc, the header should not be accessed by
4863		 * any other thread. And, if reached via arc_read_done,
4864		 * the hash lock will protect it if it's found in the
4865		 * hash table; otherwise no other thread should be
4866		 * trying to [add|remove]_reference it.
4867		 */
4868		if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4869			ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4870			(void) refcount_add_many(&state->arcs_esize[type],
4871			    size, tag);
4872		}
4873
4874		/*
4875		 * If we are growing the cache, and we are adding anonymous
4876		 * data, and we have outgrown arc_p, update arc_p
4877		 */
4878		if (aggsum_compare(&arc_size, arc_c) < 0 &&
4879		    hdr->b_l1hdr.b_state == arc_anon &&
4880		    (refcount_count(&arc_anon->arcs_size) +
4881		    refcount_count(&arc_mru->arcs_size) > arc_p))
4882			arc_p = MIN(arc_c, arc_p + size);
4883	}
4884	ARCSTAT_BUMP(arcstat_allocated);
4885}
4886
4887static void
4888arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
4889{
4890	arc_free_data_impl(hdr, size, tag);
4891	abd_free(abd);
4892}
4893
4894static void
4895arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
4896{
4897	arc_buf_contents_t type = arc_buf_type(hdr);
4898
4899	arc_free_data_impl(hdr, size, tag);
4900	if (type == ARC_BUFC_METADATA) {
4901		zio_buf_free(buf, size);
4902	} else {
4903		ASSERT(type == ARC_BUFC_DATA);
4904		zio_data_buf_free(buf, size);
4905	}
4906}
4907
4908/*
4909 * Free the arc data buffer.
4910 */
4911static void
4912arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4913{
4914	arc_state_t *state = hdr->b_l1hdr.b_state;
4915	arc_buf_contents_t type = arc_buf_type(hdr);
4916
4917	/* protected by hash lock, if in the hash table */
4918	if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4919		ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4920		ASSERT(state != arc_anon && state != arc_l2c_only);
4921
4922		(void) refcount_remove_many(&state->arcs_esize[type],
4923		    size, tag);
4924	}
4925	(void) refcount_remove_many(&state->arcs_size, size, tag);
4926
4927	VERIFY3U(hdr->b_type, ==, type);
4928	if (type == ARC_BUFC_METADATA) {
4929		arc_space_return(size, ARC_SPACE_META);
4930	} else {
4931		ASSERT(type == ARC_BUFC_DATA);
4932		arc_space_return(size, ARC_SPACE_DATA);
4933	}
4934}
4935
4936/*
4937 * This routine is called whenever a buffer is accessed.
4938 * NOTE: the hash lock is dropped in this function.
4939 */
4940static void
4941arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4942{
4943	clock_t now;
4944
4945	ASSERT(MUTEX_HELD(hash_lock));
4946	ASSERT(HDR_HAS_L1HDR(hdr));
4947
4948	if (hdr->b_l1hdr.b_state == arc_anon) {
4949		/*
4950		 * This buffer is not in the cache, and does not
4951		 * appear in our "ghost" list.  Add the new buffer
4952		 * to the MRU state.
4953		 */
4954
4955		ASSERT0(hdr->b_l1hdr.b_arc_access);
4956		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4957		DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4958		arc_change_state(arc_mru, hdr, hash_lock);
4959
4960	} else if (hdr->b_l1hdr.b_state == arc_mru) {
4961		now = ddi_get_lbolt();
4962
4963		/*
4964		 * If this buffer is here because of a prefetch, then either:
4965		 * - clear the flag if this is a "referencing" read
4966		 *   (any subsequent access will bump this into the MFU state).
4967		 * or
4968		 * - move the buffer to the head of the list if this is
4969		 *   another prefetch (to make it less likely to be evicted).
4970		 */
4971		if (HDR_PREFETCH(hdr)) {
4972			if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4973				/* link protected by hash lock */
4974				ASSERT(multilist_link_active(
4975				    &hdr->b_l1hdr.b_arc_node));
4976			} else {
4977				arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4978				ARCSTAT_BUMP(arcstat_mru_hits);
4979			}
4980			hdr->b_l1hdr.b_arc_access = now;
4981			return;
4982		}
4983
4984		/*
4985		 * This buffer has been "accessed" only once so far,
4986		 * but it is still in the cache. Move it to the MFU
4987		 * state.
4988		 */
4989		if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
4990			/*
4991			 * More than 125ms have passed since we
4992			 * instantiated this buffer.  Move it to the
4993			 * most frequently used state.
4994			 */
4995			hdr->b_l1hdr.b_arc_access = now;
4996			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4997			arc_change_state(arc_mfu, hdr, hash_lock);
4998		}
4999		ARCSTAT_BUMP(arcstat_mru_hits);
5000	} else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5001		arc_state_t	*new_state;
5002		/*
5003		 * This buffer has been "accessed" recently, but
5004		 * was evicted from the cache.  Move it to the
5005		 * MFU state.
5006		 */
5007
5008		if (HDR_PREFETCH(hdr)) {
5009			new_state = arc_mru;
5010			if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
5011				arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
5012			DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5013		} else {
5014			new_state = arc_mfu;
5015			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5016		}
5017
5018		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5019		arc_change_state(new_state, hdr, hash_lock);
5020
5021		ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5022	} else if (hdr->b_l1hdr.b_state == arc_mfu) {
5023		/*
5024		 * This buffer has been accessed more than once and is
5025		 * still in the cache.  Keep it in the MFU state.
5026		 *
5027		 * NOTE: an add_reference() that occurred when we did
5028		 * the arc_read() will have kicked this off the list.
5029		 * If it was a prefetch, we will explicitly move it to
5030		 * the head of the list now.
5031		 */
5032		if ((HDR_PREFETCH(hdr)) != 0) {
5033			ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5034			/* link protected by hash_lock */
5035			ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5036		}
5037		ARCSTAT_BUMP(arcstat_mfu_hits);
5038		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5039	} else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5040		arc_state_t	*new_state = arc_mfu;
5041		/*
5042		 * This buffer has been accessed more than once but has
5043		 * been evicted from the cache.  Move it back to the
5044		 * MFU state.
5045		 */
5046
5047		if (HDR_PREFETCH(hdr)) {
5048			/*
5049			 * This is a prefetch access...
5050			 * move this block back to the MRU state.
5051			 */
5052			ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
5053			new_state = arc_mru;
5054		}
5055
5056		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5057		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5058		arc_change_state(new_state, hdr, hash_lock);
5059
5060		ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5061	} else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5062		/*
5063		 * This buffer is on the 2nd Level ARC.
5064		 */
5065
5066		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5067		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5068		arc_change_state(arc_mfu, hdr, hash_lock);
5069	} else {
5070		ASSERT(!"invalid arc state");
5071	}
5072}
5073
5074/*
5075 * This routine is called by dbuf_hold() to update the arc_access() state
5076 * which otherwise would be skipped for entries in the dbuf cache.
5077 */
5078void
5079arc_buf_access(arc_buf_t *buf)
5080{
5081	mutex_enter(&buf->b_evict_lock);
5082	arc_buf_hdr_t *hdr = buf->b_hdr;
5083
5084	/*
5085	 * Avoid taking the hash_lock when possible as an optimization.
5086	 * The header must be checked again under the hash_lock in order
5087	 * to handle the case where it is concurrently being released.
5088	 */
5089	if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5090		mutex_exit(&buf->b_evict_lock);
5091		ARCSTAT_BUMP(arcstat_access_skip);
5092		return;
5093	}
5094
5095	kmutex_t *hash_lock = HDR_LOCK(hdr);
5096	mutex_enter(hash_lock);
5097
5098	if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5099		mutex_exit(hash_lock);
5100		mutex_exit(&buf->b_evict_lock);
5101		ARCSTAT_BUMP(arcstat_access_skip);
5102		return;
5103	}
5104
5105	mutex_exit(&buf->b_evict_lock);
5106
5107	ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5108	    hdr->b_l1hdr.b_state == arc_mfu);
5109
5110	DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5111	arc_access(hdr, hash_lock);
5112	mutex_exit(hash_lock);
5113
5114	ARCSTAT_BUMP(arcstat_hits);
5115	ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5116	    demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5117}
5118
5119/* a generic arc_done_func_t which you can use */
5120/* ARGSUSED */
5121void
5122arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
5123{
5124	if (zio == NULL || zio->io_error == 0)
5125		bcopy(buf->b_data, arg, arc_buf_size(buf));
5126	arc_buf_destroy(buf, arg);
5127}
5128
5129/* a generic arc_done_func_t */
5130void
5131arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
5132{
5133	arc_buf_t **bufp = arg;
5134	if (zio && zio->io_error) {
5135		arc_buf_destroy(buf, arg);
5136		*bufp = NULL;
5137	} else {
5138		*bufp = buf;
5139		ASSERT(buf->b_data);
5140	}
5141}
5142
5143static void
5144arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5145{
5146	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5147		ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5148		ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
5149	} else {
5150		if (HDR_COMPRESSION_ENABLED(hdr)) {
5151			ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
5152			    BP_GET_COMPRESS(bp));
5153		}
5154		ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5155		ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5156	}
5157}
5158
5159static void
5160arc_read_done(zio_t *zio)
5161{
5162	arc_buf_hdr_t	*hdr = zio->io_private;
5163	kmutex_t	*hash_lock = NULL;
5164	arc_callback_t	*callback_list;
5165	arc_callback_t	*acb;
5166	boolean_t	freeable = B_FALSE;
5167	boolean_t	no_zio_error = (zio->io_error == 0);
5168
5169	/*
5170	 * The hdr was inserted into hash-table and removed from lists
5171	 * prior to starting I/O.  We should find this header, since
5172	 * it's in the hash table, and it should be legit since it's
5173	 * not possible to evict it during the I/O.  The only possible
5174	 * reason for it not to be found is if we were freed during the
5175	 * read.
5176	 */
5177	if (HDR_IN_HASH_TABLE(hdr)) {
5178		ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5179		ASSERT3U(hdr->b_dva.dva_word[0], ==,
5180		    BP_IDENTITY(zio->io_bp)->dva_word[0]);
5181		ASSERT3U(hdr->b_dva.dva_word[1], ==,
5182		    BP_IDENTITY(zio->io_bp)->dva_word[1]);
5183
5184		arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
5185		    &hash_lock);
5186
5187		ASSERT((found == hdr &&
5188		    DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5189		    (found == hdr && HDR_L2_READING(hdr)));
5190		ASSERT3P(hash_lock, !=, NULL);
5191	}
5192
5193	if (no_zio_error) {
5194		/* byteswap if necessary */
5195		if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5196			if (BP_GET_LEVEL(zio->io_bp) > 0) {
5197				hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5198			} else {
5199				hdr->b_l1hdr.b_byteswap =
5200				    DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5201			}
5202		} else {
5203			hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5204		}
5205	}
5206
5207	arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5208	if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5209		arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5210
5211	callback_list = hdr->b_l1hdr.b_acb;
5212	ASSERT3P(callback_list, !=, NULL);
5213
5214	if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) {
5215		/*
5216		 * Only call arc_access on anonymous buffers.  This is because
5217		 * if we've issued an I/O for an evicted buffer, we've already
5218		 * called arc_access (to prevent any simultaneous readers from
5219		 * getting confused).
5220		 */
5221		arc_access(hdr, hash_lock);
5222	}
5223
5224	/*
5225	 * If a read request has a callback (i.e. acb_done is not NULL), then we
5226	 * make a buf containing the data according to the parameters which were
5227	 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5228	 * aren't needlessly decompressing the data multiple times.
5229	 */
5230	int callback_cnt = 0;
5231	for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5232		if (!acb->acb_done)
5233			continue;
5234
5235		/* This is a demand read since prefetches don't use callbacks */
5236		callback_cnt++;
5237
5238		int error = arc_buf_alloc_impl(hdr, acb->acb_private,
5239		    acb->acb_compressed, no_zio_error, &acb->acb_buf);
5240		if (no_zio_error) {
5241			zio->io_error = error;
5242		}
5243	}
5244	hdr->b_l1hdr.b_acb = NULL;
5245	arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5246	if (callback_cnt == 0) {
5247		ASSERT(HDR_PREFETCH(hdr));
5248		ASSERT0(hdr->b_l1hdr.b_bufcnt);
5249		ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5250	}
5251
5252	ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5253	    callback_list != NULL);
5254
5255	if (no_zio_error) {
5256		arc_hdr_verify(hdr, zio->io_bp);
5257	} else {
5258		arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5259		if (hdr->b_l1hdr.b_state != arc_anon)
5260			arc_change_state(arc_anon, hdr, hash_lock);
5261		if (HDR_IN_HASH_TABLE(hdr))
5262			buf_hash_remove(hdr);
5263		freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5264	}
5265
5266	/*
5267	 * Broadcast before we drop the hash_lock to avoid the possibility
5268	 * that the hdr (and hence the cv) might be freed before we get to
5269	 * the cv_broadcast().
5270	 */
5271	cv_broadcast(&hdr->b_l1hdr.b_cv);
5272
5273	if (hash_lock != NULL) {
5274		mutex_exit(hash_lock);
5275	} else {
5276		/*
5277		 * This block was freed while we waited for the read to
5278		 * complete.  It has been removed from the hash table and
5279		 * moved to the anonymous state (so that it won't show up
5280		 * in the cache).
5281		 */
5282		ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5283		freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5284	}
5285
5286	/* execute each callback and free its structure */
5287	while ((acb = callback_list) != NULL) {
5288		if (acb->acb_done)
5289			acb->acb_done(zio, acb->acb_buf, acb->acb_private);
5290
5291		if (acb->acb_zio_dummy != NULL) {
5292			acb->acb_zio_dummy->io_error = zio->io_error;
5293			zio_nowait(acb->acb_zio_dummy);
5294		}
5295
5296		callback_list = acb->acb_next;
5297		kmem_free(acb, sizeof (arc_callback_t));
5298	}
5299
5300	if (freeable)
5301		arc_hdr_destroy(hdr);
5302}
5303
5304/*
5305 * "Read" the block at the specified DVA (in bp) via the
5306 * cache.  If the block is found in the cache, invoke the provided
5307 * callback immediately and return.  Note that the `zio' parameter
5308 * in the callback will be NULL in this case, since no IO was
5309 * required.  If the block is not in the cache pass the read request
5310 * on to the spa with a substitute callback function, so that the
5311 * requested block will be added to the cache.
5312 *
5313 * If a read request arrives for a block that has a read in-progress,
5314 * either wait for the in-progress read to complete (and return the
5315 * results); or, if this is a read with a "done" func, add a record
5316 * to the read to invoke the "done" func when the read completes,
5317 * and return; or just return.
5318 *
5319 * arc_read_done() will invoke all the requested "done" functions
5320 * for readers of this block.
5321 */
5322int
5323arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
5324    void *private, zio_priority_t priority, int zio_flags,
5325    arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5326{
5327	arc_buf_hdr_t *hdr = NULL;
5328	kmutex_t *hash_lock = NULL;
5329	zio_t *rzio;
5330	uint64_t guid = spa_load_guid(spa);
5331	boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0;
5332
5333	ASSERT(!BP_IS_EMBEDDED(bp) ||
5334	    BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5335
5336top:
5337	if (!BP_IS_EMBEDDED(bp)) {
5338		/*
5339		 * Embedded BP's have no DVA and require no I/O to "read".
5340		 * Create an anonymous arc buf to back it.
5341		 */
5342		hdr = buf_hash_find(guid, bp, &hash_lock);
5343	}
5344
5345	if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) {
5346		arc_buf_t *buf = NULL;
5347		*arc_flags |= ARC_FLAG_CACHED;
5348
5349		if (HDR_IO_IN_PROGRESS(hdr)) {
5350
5351			if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5352			    priority == ZIO_PRIORITY_SYNC_READ) {
5353				/*
5354				 * This sync read must wait for an
5355				 * in-progress async read (e.g. a predictive
5356				 * prefetch).  Async reads are queued
5357				 * separately at the vdev_queue layer, so
5358				 * this is a form of priority inversion.
5359				 * Ideally, we would "inherit" the demand
5360				 * i/o's priority by moving the i/o from
5361				 * the async queue to the synchronous queue,
5362				 * but there is currently no mechanism to do
5363				 * so.  Track this so that we can evaluate
5364				 * the magnitude of this potential performance
5365				 * problem.
5366				 *
5367				 * Note that if the prefetch i/o is already
5368				 * active (has been issued to the device),
5369				 * the prefetch improved performance, because
5370				 * we issued it sooner than we would have
5371				 * without the prefetch.
5372				 */
5373				DTRACE_PROBE1(arc__sync__wait__for__async,
5374				    arc_buf_hdr_t *, hdr);
5375				ARCSTAT_BUMP(arcstat_sync_wait_for_async);
5376			}
5377			if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5378				arc_hdr_clear_flags(hdr,
5379				    ARC_FLAG_PREDICTIVE_PREFETCH);
5380			}
5381
5382			if (*arc_flags & ARC_FLAG_WAIT) {
5383				cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5384				mutex_exit(hash_lock);
5385				goto top;
5386			}
5387			ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5388
5389			if (done) {
5390				arc_callback_t *acb = NULL;
5391
5392				acb = kmem_zalloc(sizeof (arc_callback_t),
5393				    KM_SLEEP);
5394				acb->acb_done = done;
5395				acb->acb_private = private;
5396				acb->acb_compressed = compressed_read;
5397				if (pio != NULL)
5398					acb->acb_zio_dummy = zio_null(pio,
5399					    spa, NULL, NULL, NULL, zio_flags);
5400
5401				ASSERT3P(acb->acb_done, !=, NULL);
5402				acb->acb_next = hdr->b_l1hdr.b_acb;
5403				hdr->b_l1hdr.b_acb = acb;
5404				mutex_exit(hash_lock);
5405				return (0);
5406			}
5407			mutex_exit(hash_lock);
5408			return (0);
5409		}
5410
5411		ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5412		    hdr->b_l1hdr.b_state == arc_mfu);
5413
5414		if (done) {
5415			if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5416				/*
5417				 * This is a demand read which does not have to
5418				 * wait for i/o because we did a predictive
5419				 * prefetch i/o for it, which has completed.
5420				 */
5421				DTRACE_PROBE1(
5422				    arc__demand__hit__predictive__prefetch,
5423				    arc_buf_hdr_t *, hdr);
5424				ARCSTAT_BUMP(
5425				    arcstat_demand_hit_predictive_prefetch);
5426				arc_hdr_clear_flags(hdr,
5427				    ARC_FLAG_PREDICTIVE_PREFETCH);
5428			}
5429			ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5430
5431			/* Get a buf with the desired data in it. */
5432			VERIFY0(arc_buf_alloc_impl(hdr, private,
5433			    compressed_read, B_TRUE, &buf));
5434		} else if (*arc_flags & ARC_FLAG_PREFETCH &&
5435		    refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5436			arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5437		}
5438		DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5439		arc_access(hdr, hash_lock);
5440		if (*arc_flags & ARC_FLAG_L2CACHE)
5441			arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5442		mutex_exit(hash_lock);
5443		ARCSTAT_BUMP(arcstat_hits);
5444		ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5445		    demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5446		    data, metadata, hits);
5447
5448		if (done)
5449			done(NULL, buf, private);
5450	} else {
5451		uint64_t lsize = BP_GET_LSIZE(bp);
5452		uint64_t psize = BP_GET_PSIZE(bp);
5453		arc_callback_t *acb;
5454		vdev_t *vd = NULL;
5455		uint64_t addr = 0;
5456		boolean_t devw = B_FALSE;
5457		uint64_t size;
5458
5459		if (hdr == NULL) {
5460			/* this block is not in the cache */
5461			arc_buf_hdr_t *exists = NULL;
5462			arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5463			hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5464			    BP_GET_COMPRESS(bp), type);
5465
5466			if (!BP_IS_EMBEDDED(bp)) {
5467				hdr->b_dva = *BP_IDENTITY(bp);
5468				hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5469				exists = buf_hash_insert(hdr, &hash_lock);
5470			}
5471			if (exists != NULL) {
5472				/* somebody beat us to the hash insert */
5473				mutex_exit(hash_lock);
5474				buf_discard_identity(hdr);
5475				arc_hdr_destroy(hdr);
5476				goto top; /* restart the IO request */
5477			}
5478		} else {
5479			/*
5480			 * This block is in the ghost cache. If it was L2-only
5481			 * (and thus didn't have an L1 hdr), we realloc the
5482			 * header to add an L1 hdr.
5483			 */
5484			if (!HDR_HAS_L1HDR(hdr)) {
5485				hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5486				    hdr_full_cache);
5487			}
5488			ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5489			ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5490			ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5491			ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5492			ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5493			ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5494
5495			/*
5496			 * This is a delicate dance that we play here.
5497			 * This hdr is in the ghost list so we access it
5498			 * to move it out of the ghost list before we
5499			 * initiate the read. If it's a prefetch then
5500			 * it won't have a callback so we'll remove the
5501			 * reference that arc_buf_alloc_impl() created. We
5502			 * do this after we've called arc_access() to
5503			 * avoid hitting an assert in remove_reference().
5504			 */
5505			arc_access(hdr, hash_lock);
5506			arc_hdr_alloc_pabd(hdr);
5507		}
5508		ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5509		size = arc_hdr_size(hdr);
5510
5511		/*
5512		 * If compression is enabled on the hdr, then will do
5513		 * RAW I/O and will store the compressed data in the hdr's
5514		 * data block. Otherwise, the hdr's data block will contain
5515		 * the uncompressed data.
5516		 */
5517		if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5518			zio_flags |= ZIO_FLAG_RAW;
5519		}
5520
5521		if (*arc_flags & ARC_FLAG_PREFETCH)
5522			arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5523		if (*arc_flags & ARC_FLAG_L2CACHE)
5524			arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5525		if (BP_GET_LEVEL(bp) > 0)
5526			arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5527		if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5528			arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5529		ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5530
5531		acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5532		acb->acb_done = done;
5533		acb->acb_private = private;
5534		acb->acb_compressed = compressed_read;
5535
5536		ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5537		hdr->b_l1hdr.b_acb = acb;
5538		arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5539
5540		if (HDR_HAS_L2HDR(hdr) &&
5541		    (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5542			devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5543			addr = hdr->b_l2hdr.b_daddr;
5544			/*
5545			 * Lock out L2ARC device removal.
5546			 */
5547			if (vdev_is_dead(vd) ||
5548			    !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5549				vd = NULL;
5550		}
5551
5552		if (priority == ZIO_PRIORITY_ASYNC_READ)
5553			arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5554		else
5555			arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5556
5557		if (hash_lock != NULL)
5558			mutex_exit(hash_lock);
5559
5560		/*
5561		 * At this point, we have a level 1 cache miss.  Try again in
5562		 * L2ARC if possible.
5563		 */
5564		ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5565
5566		DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5567		    uint64_t, lsize, zbookmark_phys_t *, zb);
5568		ARCSTAT_BUMP(arcstat_misses);
5569		ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5570		    demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5571		    data, metadata, misses);
5572#ifdef _KERNEL
5573#ifdef RACCT
5574		if (racct_enable) {
5575			PROC_LOCK(curproc);
5576			racct_add_force(curproc, RACCT_READBPS, size);
5577			racct_add_force(curproc, RACCT_READIOPS, 1);
5578			PROC_UNLOCK(curproc);
5579		}
5580#endif /* RACCT */
5581		curthread->td_ru.ru_inblock++;
5582#endif
5583
5584		if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5585			/*
5586			 * Read from the L2ARC if the following are true:
5587			 * 1. The L2ARC vdev was previously cached.
5588			 * 2. This buffer still has L2ARC metadata.
5589			 * 3. This buffer isn't currently writing to the L2ARC.
5590			 * 4. The L2ARC entry wasn't evicted, which may
5591			 *    also have invalidated the vdev.
5592			 * 5. This isn't prefetch and l2arc_noprefetch is set.
5593			 */
5594			if (HDR_HAS_L2HDR(hdr) &&
5595			    !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5596			    !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5597				l2arc_read_callback_t *cb;
5598				abd_t *abd;
5599				uint64_t asize;
5600
5601				DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5602				ARCSTAT_BUMP(arcstat_l2_hits);
5603
5604				cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5605				    KM_SLEEP);
5606				cb->l2rcb_hdr = hdr;
5607				cb->l2rcb_bp = *bp;
5608				cb->l2rcb_zb = *zb;
5609				cb->l2rcb_flags = zio_flags;
5610
5611				asize = vdev_psize_to_asize(vd, size);
5612				if (asize != size) {
5613					abd = abd_alloc_for_io(asize,
5614					    HDR_ISTYPE_METADATA(hdr));
5615					cb->l2rcb_abd = abd;
5616				} else {
5617					abd = hdr->b_l1hdr.b_pabd;
5618				}
5619
5620				ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5621				    addr + asize <= vd->vdev_psize -
5622				    VDEV_LABEL_END_SIZE);
5623
5624				/*
5625				 * l2arc read.  The SCL_L2ARC lock will be
5626				 * released by l2arc_read_done().
5627				 * Issue a null zio if the underlying buffer
5628				 * was squashed to zero size by compression.
5629				 */
5630				ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
5631				    ZIO_COMPRESS_EMPTY);
5632				rzio = zio_read_phys(pio, vd, addr,
5633				    asize, abd,
5634				    ZIO_CHECKSUM_OFF,
5635				    l2arc_read_done, cb, priority,
5636				    zio_flags | ZIO_FLAG_DONT_CACHE |
5637				    ZIO_FLAG_CANFAIL |
5638				    ZIO_FLAG_DONT_PROPAGATE |
5639				    ZIO_FLAG_DONT_RETRY, B_FALSE);
5640				DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5641				    zio_t *, rzio);
5642				ARCSTAT_INCR(arcstat_l2_read_bytes, size);
5643
5644				if (*arc_flags & ARC_FLAG_NOWAIT) {
5645					zio_nowait(rzio);
5646					return (0);
5647				}
5648
5649				ASSERT(*arc_flags & ARC_FLAG_WAIT);
5650				if (zio_wait(rzio) == 0)
5651					return (0);
5652
5653				/* l2arc read error; goto zio_read() */
5654			} else {
5655				DTRACE_PROBE1(l2arc__miss,
5656				    arc_buf_hdr_t *, hdr);
5657				ARCSTAT_BUMP(arcstat_l2_misses);
5658				if (HDR_L2_WRITING(hdr))
5659					ARCSTAT_BUMP(arcstat_l2_rw_clash);
5660				spa_config_exit(spa, SCL_L2ARC, vd);
5661			}
5662		} else {
5663			if (vd != NULL)
5664				spa_config_exit(spa, SCL_L2ARC, vd);
5665			if (l2arc_ndev != 0) {
5666				DTRACE_PROBE1(l2arc__miss,
5667				    arc_buf_hdr_t *, hdr);
5668				ARCSTAT_BUMP(arcstat_l2_misses);
5669			}
5670		}
5671
5672		rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size,
5673		    arc_read_done, hdr, priority, zio_flags, zb);
5674
5675		if (*arc_flags & ARC_FLAG_WAIT)
5676			return (zio_wait(rzio));
5677
5678		ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5679		zio_nowait(rzio);
5680	}
5681	return (0);
5682}
5683
5684/*
5685 * Notify the arc that a block was freed, and thus will never be used again.
5686 */
5687void
5688arc_freed(spa_t *spa, const blkptr_t *bp)
5689{
5690	arc_buf_hdr_t *hdr;
5691	kmutex_t *hash_lock;
5692	uint64_t guid = spa_load_guid(spa);
5693
5694	ASSERT(!BP_IS_EMBEDDED(bp));
5695
5696	hdr = buf_hash_find(guid, bp, &hash_lock);
5697	if (hdr == NULL)
5698		return;
5699
5700	/*
5701	 * We might be trying to free a block that is still doing I/O
5702	 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5703	 * dmu_sync-ed block). If this block is being prefetched, then it
5704	 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5705	 * until the I/O completes. A block may also have a reference if it is
5706	 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5707	 * have written the new block to its final resting place on disk but
5708	 * without the dedup flag set. This would have left the hdr in the MRU
5709	 * state and discoverable. When the txg finally syncs it detects that
5710	 * the block was overridden in open context and issues an override I/O.
5711	 * Since this is a dedup block, the override I/O will determine if the
5712	 * block is already in the DDT. If so, then it will replace the io_bp
5713	 * with the bp from the DDT and allow the I/O to finish. When the I/O
5714	 * reaches the done callback, dbuf_write_override_done, it will
5715	 * check to see if the io_bp and io_bp_override are identical.
5716	 * If they are not, then it indicates that the bp was replaced with
5717	 * the bp in the DDT and the override bp is freed. This allows
5718	 * us to arrive here with a reference on a block that is being
5719	 * freed. So if we have an I/O in progress, or a reference to
5720	 * this hdr, then we don't destroy the hdr.
5721	 */
5722	if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
5723	    refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
5724		arc_change_state(arc_anon, hdr, hash_lock);
5725		arc_hdr_destroy(hdr);
5726		mutex_exit(hash_lock);
5727	} else {
5728		mutex_exit(hash_lock);
5729	}
5730
5731}
5732
5733/*
5734 * Release this buffer from the cache, making it an anonymous buffer.  This
5735 * must be done after a read and prior to modifying the buffer contents.
5736 * If the buffer has more than one reference, we must make
5737 * a new hdr for the buffer.
5738 */
5739void
5740arc_release(arc_buf_t *buf, void *tag)
5741{
5742	arc_buf_hdr_t *hdr = buf->b_hdr;
5743
5744	/*
5745	 * It would be nice to assert that if it's DMU metadata (level >
5746	 * 0 || it's the dnode file), then it must be syncing context.
5747	 * But we don't know that information at this level.
5748	 */
5749
5750	mutex_enter(&buf->b_evict_lock);
5751
5752	ASSERT(HDR_HAS_L1HDR(hdr));
5753
5754	/*
5755	 * We don't grab the hash lock prior to this check, because if
5756	 * the buffer's header is in the arc_anon state, it won't be
5757	 * linked into the hash table.
5758	 */
5759	if (hdr->b_l1hdr.b_state == arc_anon) {
5760		mutex_exit(&buf->b_evict_lock);
5761		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5762		ASSERT(!HDR_IN_HASH_TABLE(hdr));
5763		ASSERT(!HDR_HAS_L2HDR(hdr));
5764		ASSERT(HDR_EMPTY(hdr));
5765		ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5766		ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
5767		ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
5768
5769		hdr->b_l1hdr.b_arc_access = 0;
5770
5771		/*
5772		 * If the buf is being overridden then it may already
5773		 * have a hdr that is not empty.
5774		 */
5775		buf_discard_identity(hdr);
5776		arc_buf_thaw(buf);
5777
5778		return;
5779	}
5780
5781	kmutex_t *hash_lock = HDR_LOCK(hdr);
5782	mutex_enter(hash_lock);
5783
5784	/*
5785	 * This assignment is only valid as long as the hash_lock is
5786	 * held, we must be careful not to reference state or the
5787	 * b_state field after dropping the lock.
5788	 */
5789	arc_state_t *state = hdr->b_l1hdr.b_state;
5790	ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5791	ASSERT3P(state, !=, arc_anon);
5792
5793	/* this buffer is not on any list */
5794	ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
5795
5796	if (HDR_HAS_L2HDR(hdr)) {
5797		mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5798
5799		/*
5800		 * We have to recheck this conditional again now that
5801		 * we're holding the l2ad_mtx to prevent a race with
5802		 * another thread which might be concurrently calling
5803		 * l2arc_evict(). In that case, l2arc_evict() might have
5804		 * destroyed the header's L2 portion as we were waiting
5805		 * to acquire the l2ad_mtx.
5806		 */
5807		if (HDR_HAS_L2HDR(hdr)) {
5808			l2arc_trim(hdr);
5809			arc_hdr_l2hdr_destroy(hdr);
5810		}
5811
5812		mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5813	}
5814
5815	/*
5816	 * Do we have more than one buf?
5817	 */
5818	if (hdr->b_l1hdr.b_bufcnt > 1) {
5819		arc_buf_hdr_t *nhdr;
5820		uint64_t spa = hdr->b_spa;
5821		uint64_t psize = HDR_GET_PSIZE(hdr);
5822		uint64_t lsize = HDR_GET_LSIZE(hdr);
5823		enum zio_compress compress = HDR_GET_COMPRESS(hdr);
5824		arc_buf_contents_t type = arc_buf_type(hdr);
5825		VERIFY3U(hdr->b_type, ==, type);
5826
5827		ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
5828		(void) remove_reference(hdr, hash_lock, tag);
5829
5830		if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
5831			ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5832			ASSERT(ARC_BUF_LAST(buf));
5833		}
5834
5835		/*
5836		 * Pull the data off of this hdr and attach it to
5837		 * a new anonymous hdr. Also find the last buffer
5838		 * in the hdr's buffer list.
5839		 */
5840		arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
5841		ASSERT3P(lastbuf, !=, NULL);
5842
5843		/*
5844		 * If the current arc_buf_t and the hdr are sharing their data
5845		 * buffer, then we must stop sharing that block.
5846		 */
5847		if (arc_buf_is_shared(buf)) {
5848			VERIFY(!arc_buf_is_shared(lastbuf));
5849
5850			/*
5851			 * First, sever the block sharing relationship between
5852			 * buf and the arc_buf_hdr_t.
5853			 */
5854			arc_unshare_buf(hdr, buf);
5855
5856			/*
5857			 * Now we need to recreate the hdr's b_pabd. Since we
5858			 * have lastbuf handy, we try to share with it, but if
5859			 * we can't then we allocate a new b_pabd and copy the
5860			 * data from buf into it.
5861			 */
5862			if (arc_can_share(hdr, lastbuf)) {
5863				arc_share_buf(hdr, lastbuf);
5864			} else {
5865				arc_hdr_alloc_pabd(hdr);
5866				abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
5867				    buf->b_data, psize);
5868			}
5869			VERIFY3P(lastbuf->b_data, !=, NULL);
5870		} else if (HDR_SHARED_DATA(hdr)) {
5871			/*
5872			 * Uncompressed shared buffers are always at the end
5873			 * of the list. Compressed buffers don't have the
5874			 * same requirements. This makes it hard to
5875			 * simply assert that the lastbuf is shared so
5876			 * we rely on the hdr's compression flags to determine
5877			 * if we have a compressed, shared buffer.
5878			 */
5879			ASSERT(arc_buf_is_shared(lastbuf) ||
5880			    HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
5881			ASSERT(!ARC_BUF_SHARED(buf));
5882		}
5883		ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5884		ASSERT3P(state, !=, arc_l2c_only);
5885
5886		(void) refcount_remove_many(&state->arcs_size,
5887		    arc_buf_size(buf), buf);
5888
5889		if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5890			ASSERT3P(state, !=, arc_l2c_only);
5891			(void) refcount_remove_many(&state->arcs_esize[type],
5892			    arc_buf_size(buf), buf);
5893		}
5894
5895		hdr->b_l1hdr.b_bufcnt -= 1;
5896		arc_cksum_verify(buf);
5897#ifdef illumos
5898		arc_buf_unwatch(buf);
5899#endif
5900
5901		mutex_exit(hash_lock);
5902
5903		/*
5904		 * Allocate a new hdr. The new hdr will contain a b_pabd
5905		 * buffer which will be freed in arc_write().
5906		 */
5907		nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
5908		ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
5909		ASSERT0(nhdr->b_l1hdr.b_bufcnt);
5910		ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
5911		VERIFY3U(nhdr->b_type, ==, type);
5912		ASSERT(!HDR_SHARED_DATA(nhdr));
5913
5914		nhdr->b_l1hdr.b_buf = buf;
5915		nhdr->b_l1hdr.b_bufcnt = 1;
5916		(void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
5917		buf->b_hdr = nhdr;
5918
5919		mutex_exit(&buf->b_evict_lock);
5920		(void) refcount_add_many(&arc_anon->arcs_size,
5921		    arc_buf_size(buf), buf);
5922	} else {
5923		mutex_exit(&buf->b_evict_lock);
5924		ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
5925		/* protected by hash lock, or hdr is on arc_anon */
5926		ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5927		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5928		arc_change_state(arc_anon, hdr, hash_lock);
5929		hdr->b_l1hdr.b_arc_access = 0;
5930		mutex_exit(hash_lock);
5931
5932		buf_discard_identity(hdr);
5933		arc_buf_thaw(buf);
5934	}
5935}
5936
5937int
5938arc_released(arc_buf_t *buf)
5939{
5940	int released;
5941
5942	mutex_enter(&buf->b_evict_lock);
5943	released = (buf->b_data != NULL &&
5944	    buf->b_hdr->b_l1hdr.b_state == arc_anon);
5945	mutex_exit(&buf->b_evict_lock);
5946	return (released);
5947}
5948
5949#ifdef ZFS_DEBUG
5950int
5951arc_referenced(arc_buf_t *buf)
5952{
5953	int referenced;
5954
5955	mutex_enter(&buf->b_evict_lock);
5956	referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
5957	mutex_exit(&buf->b_evict_lock);
5958	return (referenced);
5959}
5960#endif
5961
5962static void
5963arc_write_ready(zio_t *zio)
5964{
5965	arc_write_callback_t *callback = zio->io_private;
5966	arc_buf_t *buf = callback->awcb_buf;
5967	arc_buf_hdr_t *hdr = buf->b_hdr;
5968	uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
5969
5970	ASSERT(HDR_HAS_L1HDR(hdr));
5971	ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
5972	ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
5973
5974	/*
5975	 * If we're reexecuting this zio because the pool suspended, then
5976	 * cleanup any state that was previously set the first time the
5977	 * callback was invoked.
5978	 */
5979	if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
5980		arc_cksum_free(hdr);
5981#ifdef illumos
5982		arc_buf_unwatch(buf);
5983#endif
5984		if (hdr->b_l1hdr.b_pabd != NULL) {
5985			if (arc_buf_is_shared(buf)) {
5986				arc_unshare_buf(hdr, buf);
5987			} else {
5988				arc_hdr_free_pabd(hdr);
5989			}
5990		}
5991	}
5992	ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5993	ASSERT(!HDR_SHARED_DATA(hdr));
5994	ASSERT(!arc_buf_is_shared(buf));
5995
5996	callback->awcb_ready(zio, buf, callback->awcb_private);
5997
5998	if (HDR_IO_IN_PROGRESS(hdr))
5999		ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6000
6001	arc_cksum_compute(buf);
6002	arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6003
6004	enum zio_compress compress;
6005	if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6006		compress = ZIO_COMPRESS_OFF;
6007	} else {
6008		ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
6009		compress = BP_GET_COMPRESS(zio->io_bp);
6010	}
6011	HDR_SET_PSIZE(hdr, psize);
6012	arc_hdr_set_compress(hdr, compress);
6013
6014
6015	/*
6016	 * Fill the hdr with data. If the hdr is compressed, the data we want
6017	 * is available from the zio, otherwise we can take it from the buf.
6018	 *
6019	 * We might be able to share the buf's data with the hdr here. However,
6020	 * doing so would cause the ARC to be full of linear ABDs if we write a
6021	 * lot of shareable data. As a compromise, we check whether scattered
6022	 * ABDs are allowed, and assume that if they are then the user wants
6023	 * the ARC to be primarily filled with them regardless of the data being
6024	 * written. Therefore, if they're allowed then we allocate one and copy
6025	 * the data into it; otherwise, we share the data directly if we can.
6026	 */
6027	if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
6028		arc_hdr_alloc_pabd(hdr);
6029
6030		/*
6031		 * Ideally, we would always copy the io_abd into b_pabd, but the
6032		 * user may have disabled compressed ARC, thus we must check the
6033		 * hdr's compression setting rather than the io_bp's.
6034		 */
6035		if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
6036			ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=,
6037			    ZIO_COMPRESS_OFF);
6038			ASSERT3U(psize, >, 0);
6039
6040			abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6041		} else {
6042			ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6043
6044			abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6045			    arc_buf_size(buf));
6046		}
6047	} else {
6048		ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6049		ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6050		ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6051
6052		arc_share_buf(hdr, buf);
6053	}
6054
6055	arc_hdr_verify(hdr, zio->io_bp);
6056}
6057
6058static void
6059arc_write_children_ready(zio_t *zio)
6060{
6061	arc_write_callback_t *callback = zio->io_private;
6062	arc_buf_t *buf = callback->awcb_buf;
6063
6064	callback->awcb_children_ready(zio, buf, callback->awcb_private);
6065}
6066
6067/*
6068 * The SPA calls this callback for each physical write that happens on behalf
6069 * of a logical write.  See the comment in dbuf_write_physdone() for details.
6070 */
6071static void
6072arc_write_physdone(zio_t *zio)
6073{
6074	arc_write_callback_t *cb = zio->io_private;
6075	if (cb->awcb_physdone != NULL)
6076		cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
6077}
6078
6079static void
6080arc_write_done(zio_t *zio)
6081{
6082	arc_write_callback_t *callback = zio->io_private;
6083	arc_buf_t *buf = callback->awcb_buf;
6084	arc_buf_hdr_t *hdr = buf->b_hdr;
6085
6086	ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6087
6088	if (zio->io_error == 0) {
6089		arc_hdr_verify(hdr, zio->io_bp);
6090
6091		if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6092			buf_discard_identity(hdr);
6093		} else {
6094			hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6095			hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6096		}
6097	} else {
6098		ASSERT(HDR_EMPTY(hdr));
6099	}
6100
6101	/*
6102	 * If the block to be written was all-zero or compressed enough to be
6103	 * embedded in the BP, no write was performed so there will be no
6104	 * dva/birth/checksum.  The buffer must therefore remain anonymous
6105	 * (and uncached).
6106	 */
6107	if (!HDR_EMPTY(hdr)) {
6108		arc_buf_hdr_t *exists;
6109		kmutex_t *hash_lock;
6110
6111		ASSERT3U(zio->io_error, ==, 0);
6112
6113		arc_cksum_verify(buf);
6114
6115		exists = buf_hash_insert(hdr, &hash_lock);
6116		if (exists != NULL) {
6117			/*
6118			 * This can only happen if we overwrite for
6119			 * sync-to-convergence, because we remove
6120			 * buffers from the hash table when we arc_free().
6121			 */
6122			if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6123				if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6124					panic("bad overwrite, hdr=%p exists=%p",
6125					    (void *)hdr, (void *)exists);
6126				ASSERT(refcount_is_zero(
6127				    &exists->b_l1hdr.b_refcnt));
6128				arc_change_state(arc_anon, exists, hash_lock);
6129				mutex_exit(hash_lock);
6130				arc_hdr_destroy(exists);
6131				exists = buf_hash_insert(hdr, &hash_lock);
6132				ASSERT3P(exists, ==, NULL);
6133			} else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6134				/* nopwrite */
6135				ASSERT(zio->io_prop.zp_nopwrite);
6136				if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6137					panic("bad nopwrite, hdr=%p exists=%p",
6138					    (void *)hdr, (void *)exists);
6139			} else {
6140				/* Dedup */
6141				ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6142				ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6143				ASSERT(BP_GET_DEDUP(zio->io_bp));
6144				ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6145			}
6146		}
6147		arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6148		/* if it's not anon, we are doing a scrub */
6149		if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6150			arc_access(hdr, hash_lock);
6151		mutex_exit(hash_lock);
6152	} else {
6153		arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6154	}
6155
6156	ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6157	callback->awcb_done(zio, buf, callback->awcb_private);
6158
6159	abd_put(zio->io_abd);
6160	kmem_free(callback, sizeof (arc_write_callback_t));
6161}
6162
6163zio_t *
6164arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
6165    boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready,
6166    arc_done_func_t *children_ready, arc_done_func_t *physdone,
6167    arc_done_func_t *done, void *private, zio_priority_t priority,
6168    int zio_flags, const zbookmark_phys_t *zb)
6169{
6170	arc_buf_hdr_t *hdr = buf->b_hdr;
6171	arc_write_callback_t *callback;
6172	zio_t *zio;
6173	zio_prop_t localprop = *zp;
6174
6175	ASSERT3P(ready, !=, NULL);
6176	ASSERT3P(done, !=, NULL);
6177	ASSERT(!HDR_IO_ERROR(hdr));
6178	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6179	ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6180	ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6181	if (l2arc)
6182		arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6183	if (ARC_BUF_COMPRESSED(buf)) {
6184		/*
6185		 * We're writing a pre-compressed buffer.  Make the
6186		 * compression algorithm requested by the zio_prop_t match
6187		 * the pre-compressed buffer's compression algorithm.
6188		 */
6189		localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6190
6191		ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6192		zio_flags |= ZIO_FLAG_RAW;
6193	}
6194	callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6195	callback->awcb_ready = ready;
6196	callback->awcb_children_ready = children_ready;
6197	callback->awcb_physdone = physdone;
6198	callback->awcb_done = done;
6199	callback->awcb_private = private;
6200	callback->awcb_buf = buf;
6201
6202	/*
6203	 * The hdr's b_pabd is now stale, free it now. A new data block
6204	 * will be allocated when the zio pipeline calls arc_write_ready().
6205	 */
6206	if (hdr->b_l1hdr.b_pabd != NULL) {
6207		/*
6208		 * If the buf is currently sharing the data block with
6209		 * the hdr then we need to break that relationship here.
6210		 * The hdr will remain with a NULL data pointer and the
6211		 * buf will take sole ownership of the block.
6212		 */
6213		if (arc_buf_is_shared(buf)) {
6214			arc_unshare_buf(hdr, buf);
6215		} else {
6216			arc_hdr_free_pabd(hdr);
6217		}
6218		VERIFY3P(buf->b_data, !=, NULL);
6219		arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6220	}
6221	ASSERT(!arc_buf_is_shared(buf));
6222	ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6223
6224	zio = zio_write(pio, spa, txg, bp,
6225	    abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6226	    HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6227	    (children_ready != NULL) ? arc_write_children_ready : NULL,
6228	    arc_write_physdone, arc_write_done, callback,
6229	    priority, zio_flags, zb);
6230
6231	return (zio);
6232}
6233
6234static int
6235arc_memory_throttle(uint64_t reserve, uint64_t txg)
6236{
6237#ifdef _KERNEL
6238	uint64_t available_memory = ptob(freemem);
6239	static uint64_t page_load = 0;
6240	static uint64_t last_txg = 0;
6241
6242#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
6243	available_memory = MIN(available_memory, uma_avail());
6244#endif
6245
6246	if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
6247		return (0);
6248
6249	if (txg > last_txg) {
6250		last_txg = txg;
6251		page_load = 0;
6252	}
6253	/*
6254	 * If we are in pageout, we know that memory is already tight,
6255	 * the arc is already going to be evicting, so we just want to
6256	 * continue to let page writes occur as quickly as possible.
6257	 */
6258	if (curproc == pageproc) {
6259		if (page_load > MAX(ptob(minfree), available_memory) / 4)
6260			return (SET_ERROR(ERESTART));
6261		/* Note: reserve is inflated, so we deflate */
6262		page_load += reserve / 8;
6263		return (0);
6264	} else if (page_load > 0 && arc_reclaim_needed()) {
6265		/* memory is low, delay before restarting */
6266		ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
6267		return (SET_ERROR(EAGAIN));
6268	}
6269	page_load = 0;
6270#endif
6271	return (0);
6272}
6273
6274void
6275arc_tempreserve_clear(uint64_t reserve)
6276{
6277	atomic_add_64(&arc_tempreserve, -reserve);
6278	ASSERT((int64_t)arc_tempreserve >= 0);
6279}
6280
6281int
6282arc_tempreserve_space(uint64_t reserve, uint64_t txg)
6283{
6284	int error;
6285	uint64_t anon_size;
6286
6287	if (reserve > arc_c/4 && !arc_no_grow) {
6288		arc_c = MIN(arc_c_max, reserve * 4);
6289		DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
6290	}
6291	if (reserve > arc_c)
6292		return (SET_ERROR(ENOMEM));
6293
6294	/*
6295	 * Don't count loaned bufs as in flight dirty data to prevent long
6296	 * network delays from blocking transactions that are ready to be
6297	 * assigned to a txg.
6298	 */
6299
6300	/* assert that it has not wrapped around */
6301	ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6302
6303	anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
6304	    arc_loaned_bytes), 0);
6305
6306	/*
6307	 * Writes will, almost always, require additional memory allocations
6308	 * in order to compress/encrypt/etc the data.  We therefore need to
6309	 * make sure that there is sufficient available memory for this.
6310	 */
6311	error = arc_memory_throttle(reserve, txg);
6312	if (error != 0)
6313		return (error);
6314
6315	/*
6316	 * Throttle writes when the amount of dirty data in the cache
6317	 * gets too large.  We try to keep the cache less than half full
6318	 * of dirty blocks so that our sync times don't grow too large.
6319	 * Note: if two requests come in concurrently, we might let them
6320	 * both succeed, when one of them should fail.  Not a huge deal.
6321	 */
6322
6323	if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
6324	    anon_size > arc_c / 4) {
6325		uint64_t meta_esize =
6326		    refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6327		uint64_t data_esize =
6328		    refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6329		dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6330		    "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6331		    arc_tempreserve >> 10, meta_esize >> 10,
6332		    data_esize >> 10, reserve >> 10, arc_c >> 10);
6333		return (SET_ERROR(ERESTART));
6334	}
6335	atomic_add_64(&arc_tempreserve, reserve);
6336	return (0);
6337}
6338
6339static void
6340arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6341    kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6342{
6343	size->value.ui64 = refcount_count(&state->arcs_size);
6344	evict_data->value.ui64 =
6345	    refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6346	evict_metadata->value.ui64 =
6347	    refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6348}
6349
6350static int
6351arc_kstat_update(kstat_t *ksp, int rw)
6352{
6353	arc_stats_t *as = ksp->ks_data;
6354
6355	if (rw == KSTAT_WRITE) {
6356		return (EACCES);
6357	} else {
6358		arc_kstat_update_state(arc_anon,
6359		    &as->arcstat_anon_size,
6360		    &as->arcstat_anon_evictable_data,
6361		    &as->arcstat_anon_evictable_metadata);
6362		arc_kstat_update_state(arc_mru,
6363		    &as->arcstat_mru_size,
6364		    &as->arcstat_mru_evictable_data,
6365		    &as->arcstat_mru_evictable_metadata);
6366		arc_kstat_update_state(arc_mru_ghost,
6367		    &as->arcstat_mru_ghost_size,
6368		    &as->arcstat_mru_ghost_evictable_data,
6369		    &as->arcstat_mru_ghost_evictable_metadata);
6370		arc_kstat_update_state(arc_mfu,
6371		    &as->arcstat_mfu_size,
6372		    &as->arcstat_mfu_evictable_data,
6373		    &as->arcstat_mfu_evictable_metadata);
6374		arc_kstat_update_state(arc_mfu_ghost,
6375		    &as->arcstat_mfu_ghost_size,
6376		    &as->arcstat_mfu_ghost_evictable_data,
6377		    &as->arcstat_mfu_ghost_evictable_metadata);
6378
6379		ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
6380		ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
6381		ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
6382		ARCSTAT(arcstat_metadata_size) =
6383		    aggsum_value(&astat_metadata_size);
6384		ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
6385		ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size);
6386		ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
6387	}
6388
6389	return (0);
6390}
6391
6392/*
6393 * This function *must* return indices evenly distributed between all
6394 * sublists of the multilist. This is needed due to how the ARC eviction
6395 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6396 * distributed between all sublists and uses this assumption when
6397 * deciding which sublist to evict from and how much to evict from it.
6398 */
6399unsigned int
6400arc_state_multilist_index_func(multilist_t *ml, void *obj)
6401{
6402	arc_buf_hdr_t *hdr = obj;
6403
6404	/*
6405	 * We rely on b_dva to generate evenly distributed index
6406	 * numbers using buf_hash below. So, as an added precaution,
6407	 * let's make sure we never add empty buffers to the arc lists.
6408	 */
6409	ASSERT(!HDR_EMPTY(hdr));
6410
6411	/*
6412	 * The assumption here, is the hash value for a given
6413	 * arc_buf_hdr_t will remain constant throughout it's lifetime
6414	 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
6415	 * Thus, we don't need to store the header's sublist index
6416	 * on insertion, as this index can be recalculated on removal.
6417	 *
6418	 * Also, the low order bits of the hash value are thought to be
6419	 * distributed evenly. Otherwise, in the case that the multilist
6420	 * has a power of two number of sublists, each sublists' usage
6421	 * would not be evenly distributed.
6422	 */
6423	return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
6424	    multilist_get_num_sublists(ml));
6425}
6426
6427#ifdef _KERNEL
6428static eventhandler_tag arc_event_lowmem = NULL;
6429
6430static void
6431arc_lowmem(void *arg __unused, int howto __unused)
6432{
6433
6434	mutex_enter(&arc_reclaim_lock);
6435	DTRACE_PROBE1(arc__needfree, int64_t, ((int64_t)freemem - zfs_arc_free_target) * PAGESIZE);
6436	cv_signal(&arc_reclaim_thread_cv);
6437
6438	/*
6439	 * It is unsafe to block here in arbitrary threads, because we can come
6440	 * here from ARC itself and may hold ARC locks and thus risk a deadlock
6441	 * with ARC reclaim thread.
6442	 */
6443	if (curproc == pageproc)
6444		(void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
6445	mutex_exit(&arc_reclaim_lock);
6446}
6447#endif
6448
6449static void
6450arc_state_init(void)
6451{
6452	arc_anon = &ARC_anon;
6453	arc_mru = &ARC_mru;
6454	arc_mru_ghost = &ARC_mru_ghost;
6455	arc_mfu = &ARC_mfu;
6456	arc_mfu_ghost = &ARC_mfu_ghost;
6457	arc_l2c_only = &ARC_l2c_only;
6458
6459	arc_mru->arcs_list[ARC_BUFC_METADATA] =
6460	    multilist_create(sizeof (arc_buf_hdr_t),
6461	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6462	    arc_state_multilist_index_func);
6463	arc_mru->arcs_list[ARC_BUFC_DATA] =
6464	    multilist_create(sizeof (arc_buf_hdr_t),
6465	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6466	    arc_state_multilist_index_func);
6467	arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
6468	    multilist_create(sizeof (arc_buf_hdr_t),
6469	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6470	    arc_state_multilist_index_func);
6471	arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
6472	    multilist_create(sizeof (arc_buf_hdr_t),
6473	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6474	    arc_state_multilist_index_func);
6475	arc_mfu->arcs_list[ARC_BUFC_METADATA] =
6476	    multilist_create(sizeof (arc_buf_hdr_t),
6477	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6478	    arc_state_multilist_index_func);
6479	arc_mfu->arcs_list[ARC_BUFC_DATA] =
6480	    multilist_create(sizeof (arc_buf_hdr_t),
6481	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6482	    arc_state_multilist_index_func);
6483	arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
6484	    multilist_create(sizeof (arc_buf_hdr_t),
6485	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6486	    arc_state_multilist_index_func);
6487	arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
6488	    multilist_create(sizeof (arc_buf_hdr_t),
6489	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6490	    arc_state_multilist_index_func);
6491	arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
6492	    multilist_create(sizeof (arc_buf_hdr_t),
6493	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6494	    arc_state_multilist_index_func);
6495	arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
6496	    multilist_create(sizeof (arc_buf_hdr_t),
6497	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6498	    arc_state_multilist_index_func);
6499
6500	refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6501	refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6502	refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6503	refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6504	refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6505	refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6506	refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6507	refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6508	refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6509	refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6510	refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6511	refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6512
6513	refcount_create(&arc_anon->arcs_size);
6514	refcount_create(&arc_mru->arcs_size);
6515	refcount_create(&arc_mru_ghost->arcs_size);
6516	refcount_create(&arc_mfu->arcs_size);
6517	refcount_create(&arc_mfu_ghost->arcs_size);
6518	refcount_create(&arc_l2c_only->arcs_size);
6519
6520	aggsum_init(&arc_meta_used, 0);
6521	aggsum_init(&arc_size, 0);
6522	aggsum_init(&astat_data_size, 0);
6523	aggsum_init(&astat_metadata_size, 0);
6524	aggsum_init(&astat_hdr_size, 0);
6525	aggsum_init(&astat_other_size, 0);
6526	aggsum_init(&astat_l2_hdr_size, 0);
6527}
6528
6529static void
6530arc_state_fini(void)
6531{
6532	refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6533	refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6534	refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6535	refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6536	refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6537	refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6538	refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6539	refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6540	refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6541	refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6542	refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6543	refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6544
6545	refcount_destroy(&arc_anon->arcs_size);
6546	refcount_destroy(&arc_mru->arcs_size);
6547	refcount_destroy(&arc_mru_ghost->arcs_size);
6548	refcount_destroy(&arc_mfu->arcs_size);
6549	refcount_destroy(&arc_mfu_ghost->arcs_size);
6550	refcount_destroy(&arc_l2c_only->arcs_size);
6551
6552	multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
6553	multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
6554	multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
6555	multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
6556	multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
6557	multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
6558	multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
6559	multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
6560}
6561
6562uint64_t
6563arc_max_bytes(void)
6564{
6565	return (arc_c_max);
6566}
6567
6568void
6569arc_init(void)
6570{
6571	int i, prefetch_tunable_set = 0;
6572
6573	/*
6574	 * allmem is "all memory that we could possibly use".
6575	 */
6576#ifdef illumos
6577#ifdef _KERNEL
6578	uint64_t allmem = ptob(physmem - swapfs_minfree);
6579#else
6580	uint64_t allmem = (physmem * PAGESIZE) / 2;
6581#endif
6582#else
6583	uint64_t allmem = kmem_size();
6584#endif
6585
6586
6587	mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
6588	cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
6589	cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
6590
6591	mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
6592	cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL);
6593
6594	/* Convert seconds to clock ticks */
6595	arc_min_prefetch_lifespan = 1 * hz;
6596
6597	/* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
6598	arc_c_min = MAX(allmem / 32, arc_abs_min);
6599	/* set max to 5/8 of all memory, or all but 1GB, whichever is more */
6600	if (allmem >= 1 << 30)
6601		arc_c_max = allmem - (1 << 30);
6602	else
6603		arc_c_max = arc_c_min;
6604	arc_c_max = MAX(allmem * 5 / 8, arc_c_max);
6605
6606	/*
6607	 * In userland, there's only the memory pressure that we artificially
6608	 * create (see arc_available_memory()).  Don't let arc_c get too
6609	 * small, because it can cause transactions to be larger than
6610	 * arc_c, causing arc_tempreserve_space() to fail.
6611	 */
6612#ifndef _KERNEL
6613	arc_c_min = arc_c_max / 2;
6614#endif
6615
6616#ifdef _KERNEL
6617	/*
6618	 * Allow the tunables to override our calculations if they are
6619	 * reasonable.
6620	 */
6621	if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) {
6622		arc_c_max = zfs_arc_max;
6623		arc_c_min = MIN(arc_c_min, arc_c_max);
6624	}
6625	if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
6626		arc_c_min = zfs_arc_min;
6627#endif
6628
6629	arc_c = arc_c_max;
6630	arc_p = (arc_c >> 1);
6631
6632	/* limit meta-data to 1/4 of the arc capacity */
6633	arc_meta_limit = arc_c_max / 4;
6634
6635#ifdef _KERNEL
6636	/*
6637	 * Metadata is stored in the kernel's heap.  Don't let us
6638	 * use more than half the heap for the ARC.
6639	 */
6640#ifdef __FreeBSD__
6641	arc_meta_limit = MIN(arc_meta_limit, uma_limit() / 2);
6642#else
6643	arc_meta_limit = MIN(arc_meta_limit,
6644	    vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
6645#endif
6646#endif
6647
6648	/* Allow the tunable to override if it is reasonable */
6649	if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
6650		arc_meta_limit = zfs_arc_meta_limit;
6651
6652	if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
6653		arc_c_min = arc_meta_limit / 2;
6654
6655	if (zfs_arc_meta_min > 0) {
6656		arc_meta_min = zfs_arc_meta_min;
6657	} else {
6658		arc_meta_min = arc_c_min / 2;
6659	}
6660
6661	if (zfs_arc_grow_retry > 0)
6662		arc_grow_retry = zfs_arc_grow_retry;
6663
6664	if (zfs_arc_shrink_shift > 0)
6665		arc_shrink_shift = zfs_arc_shrink_shift;
6666
6667	if (zfs_arc_no_grow_shift > 0)
6668		arc_no_grow_shift = zfs_arc_no_grow_shift;
6669	/*
6670	 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6671	 */
6672	if (arc_no_grow_shift >= arc_shrink_shift)
6673		arc_no_grow_shift = arc_shrink_shift - 1;
6674
6675	if (zfs_arc_p_min_shift > 0)
6676		arc_p_min_shift = zfs_arc_p_min_shift;
6677
6678	/* if kmem_flags are set, lets try to use less memory */
6679	if (kmem_debugging())
6680		arc_c = arc_c / 2;
6681	if (arc_c < arc_c_min)
6682		arc_c = arc_c_min;
6683
6684	zfs_arc_min = arc_c_min;
6685	zfs_arc_max = arc_c_max;
6686
6687	arc_state_init();
6688	buf_init();
6689
6690	arc_reclaim_thread_exit = B_FALSE;
6691	arc_dnlc_evicts_thread_exit = FALSE;
6692
6693	arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
6694	    sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
6695
6696	if (arc_ksp != NULL) {
6697		arc_ksp->ks_data = &arc_stats;
6698		arc_ksp->ks_update = arc_kstat_update;
6699		kstat_install(arc_ksp);
6700	}
6701
6702	(void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
6703	    TS_RUN, minclsyspri);
6704
6705#ifdef _KERNEL
6706	arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
6707	    EVENTHANDLER_PRI_FIRST);
6708#endif
6709
6710	(void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0,
6711	    TS_RUN, minclsyspri);
6712
6713	arc_dead = B_FALSE;
6714	arc_warm = B_FALSE;
6715
6716	/*
6717	 * Calculate maximum amount of dirty data per pool.
6718	 *
6719	 * If it has been set by /etc/system, take that.
6720	 * Otherwise, use a percentage of physical memory defined by
6721	 * zfs_dirty_data_max_percent (default 10%) with a cap at
6722	 * zfs_dirty_data_max_max (default 4GB).
6723	 */
6724	if (zfs_dirty_data_max == 0) {
6725		zfs_dirty_data_max = ptob(physmem) *
6726		    zfs_dirty_data_max_percent / 100;
6727		zfs_dirty_data_max = MIN(zfs_dirty_data_max,
6728		    zfs_dirty_data_max_max);
6729	}
6730
6731#ifdef _KERNEL
6732	if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
6733		prefetch_tunable_set = 1;
6734
6735#ifdef __i386__
6736	if (prefetch_tunable_set == 0) {
6737		printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
6738		    "-- to enable,\n");
6739		printf("            add \"vfs.zfs.prefetch_disable=0\" "
6740		    "to /boot/loader.conf.\n");
6741		zfs_prefetch_disable = 1;
6742	}
6743#else
6744	if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
6745	    prefetch_tunable_set == 0) {
6746		printf("ZFS NOTICE: Prefetch is disabled by default if less "
6747		    "than 4GB of RAM is present;\n"
6748		    "            to enable, add \"vfs.zfs.prefetch_disable=0\" "
6749		    "to /boot/loader.conf.\n");
6750		zfs_prefetch_disable = 1;
6751	}
6752#endif
6753	/* Warn about ZFS memory and address space requirements. */
6754	if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
6755		printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
6756		    "expect unstable behavior.\n");
6757	}
6758	if (allmem < 512 * (1 << 20)) {
6759		printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
6760		    "expect unstable behavior.\n");
6761		printf("             Consider tuning vm.kmem_size and "
6762		    "vm.kmem_size_max\n");
6763		printf("             in /boot/loader.conf.\n");
6764	}
6765#endif
6766}
6767
6768void
6769arc_fini(void)
6770{
6771#ifdef _KERNEL
6772	if (arc_event_lowmem != NULL)
6773		EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
6774#endif
6775
6776	mutex_enter(&arc_reclaim_lock);
6777	arc_reclaim_thread_exit = B_TRUE;
6778	/*
6779	 * The reclaim thread will set arc_reclaim_thread_exit back to
6780	 * B_FALSE when it is finished exiting; we're waiting for that.
6781	 */
6782	while (arc_reclaim_thread_exit) {
6783		cv_signal(&arc_reclaim_thread_cv);
6784		cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
6785	}
6786	mutex_exit(&arc_reclaim_lock);
6787
6788	/* Use B_TRUE to ensure *all* buffers are evicted */
6789	arc_flush(NULL, B_TRUE);
6790
6791	mutex_enter(&arc_dnlc_evicts_lock);
6792	arc_dnlc_evicts_thread_exit = TRUE;
6793	/*
6794	 * The user evicts thread will set arc_user_evicts_thread_exit
6795	 * to FALSE when it is finished exiting; we're waiting for that.
6796	 */
6797	while (arc_dnlc_evicts_thread_exit) {
6798		cv_signal(&arc_dnlc_evicts_cv);
6799		cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
6800	}
6801	mutex_exit(&arc_dnlc_evicts_lock);
6802
6803	arc_dead = B_TRUE;
6804
6805	if (arc_ksp != NULL) {
6806		kstat_delete(arc_ksp);
6807		arc_ksp = NULL;
6808	}
6809
6810	mutex_destroy(&arc_reclaim_lock);
6811	cv_destroy(&arc_reclaim_thread_cv);
6812	cv_destroy(&arc_reclaim_waiters_cv);
6813
6814	mutex_destroy(&arc_dnlc_evicts_lock);
6815	cv_destroy(&arc_dnlc_evicts_cv);
6816
6817	arc_state_fini();
6818	buf_fini();
6819
6820	ASSERT0(arc_loaned_bytes);
6821}
6822
6823/*
6824 * Level 2 ARC
6825 *
6826 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6827 * It uses dedicated storage devices to hold cached data, which are populated
6828 * using large infrequent writes.  The main role of this cache is to boost
6829 * the performance of random read workloads.  The intended L2ARC devices
6830 * include short-stroked disks, solid state disks, and other media with
6831 * substantially faster read latency than disk.
6832 *
6833 *                 +-----------------------+
6834 *                 |         ARC           |
6835 *                 +-----------------------+
6836 *                    |         ^     ^
6837 *                    |         |     |
6838 *      l2arc_feed_thread()    arc_read()
6839 *                    |         |     |
6840 *                    |  l2arc read   |
6841 *                    V         |     |
6842 *               +---------------+    |
6843 *               |     L2ARC     |    |
6844 *               +---------------+    |
6845 *                   |    ^           |
6846 *          l2arc_write() |           |
6847 *                   |    |           |
6848 *                   V    |           |
6849 *                 +-------+      +-------+
6850 *                 | vdev  |      | vdev  |
6851 *                 | cache |      | cache |
6852 *                 +-------+      +-------+
6853 *                 +=========+     .-----.
6854 *                 :  L2ARC  :    |-_____-|
6855 *                 : devices :    | Disks |
6856 *                 +=========+    `-_____-'
6857 *
6858 * Read requests are satisfied from the following sources, in order: