xref: /titanic_51/usr/src/uts/common/fs/zfs/arc.c (revision 8edda6281c84e0632a22f9c8dbf0d6f1558878eb)
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 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 /*
29  * DVA-based Adjustable Replacement Cache
30  *
31  * While much of the theory of operation used here is
32  * based on the self-tuning, low overhead replacement cache
33  * presented by Megiddo and Modha at FAST 2003, there are some
34  * significant differences:
35  *
36  * 1. The Megiddo and Modha model assumes any page is evictable.
37  * Pages in its cache cannot be "locked" into memory.  This makes
38  * the eviction algorithm simple: evict the last page in the list.
39  * This also make the performance characteristics easy to reason
40  * about.  Our cache is not so simple.  At any given moment, some
41  * subset of the blocks in the cache are un-evictable because we
42  * have handed out a reference to them.  Blocks are only evictable
43  * when there are no external references active.  This makes
44  * eviction far more problematic:  we choose to evict the evictable
45  * blocks that are the "lowest" in the list.
46  *
47  * There are times when it is not possible to evict the requested
48  * space.  In these circumstances we are unable to adjust the cache
49  * size.  To prevent the cache growing unbounded at these times we
50  * implement a "cache throttle" that slowes the flow of new data
51  * into the cache until we can make space avaiable.
52  *
53  * 2. The Megiddo and Modha model assumes a fixed cache size.
54  * Pages are evicted when the cache is full and there is a cache
55  * miss.  Our model has a variable sized cache.  It grows with
56  * high use, but also tries to react to memory preasure from the
57  * operating system: decreasing its size when system memory is
58  * tight.
59  *
60  * 3. The Megiddo and Modha model assumes a fixed page size. All
61  * elements of the cache are therefor exactly the same size.  So
62  * when adjusting the cache size following a cache miss, its simply
63  * a matter of choosing a single page to evict.  In our model, we
64  * have variable sized cache blocks (rangeing from 512 bytes to
65  * 128K bytes).  We therefor choose a set of blocks to evict to make
66  * space for a cache miss that approximates as closely as possible
67  * the space used by the new block.
68  *
69  * See also:  "ARC: A Self-Tuning, Low Overhead Replacement Cache"
70  * by N. Megiddo & D. Modha, FAST 2003
71  */
72 
73 /*
74  * The locking model:
75  *
76  * A new reference to a cache buffer can be obtained in two
77  * ways: 1) via a hash table lookup using the DVA as a key,
78  * or 2) via one of the ARC lists.  The arc_read() inerface
79  * uses method 1, while the internal arc algorithms for
80  * adjusting the cache use method 2.  We therefor provide two
81  * types of locks: 1) the hash table lock array, and 2) the
82  * arc list locks.
83  *
84  * Buffers do not have their own mutexs, rather they rely on the
85  * hash table mutexs for the bulk of their protection (i.e. most
86  * fields in the arc_buf_hdr_t are protected by these mutexs).
87  *
88  * buf_hash_find() returns the appropriate mutex (held) when it
89  * locates the requested buffer in the hash table.  It returns
90  * NULL for the mutex if the buffer was not in the table.
91  *
92  * buf_hash_remove() expects the appropriate hash mutex to be
93  * already held before it is invoked.
94  *
95  * Each arc state also has a mutex which is used to protect the
96  * buffer list associated with the state.  When attempting to
97  * obtain a hash table lock while holding an arc list lock you
98  * must use: mutex_tryenter() to avoid deadlock.  Also note that
99  * the active state mutex must be held before the ghost state mutex.
100  *
101  * Arc buffers may have an associated eviction callback function.
102  * This function will be invoked prior to removing the buffer (e.g.
103  * in arc_do_user_evicts()).  Note however that the data associated
104  * with the buffer may be evicted prior to the callback.  The callback
105  * must be made with *no locks held* (to prevent deadlock).  Additionally,
106  * the users of callbacks must ensure that their private data is
107  * protected from simultaneous callbacks from arc_buf_evict()
108  * and arc_do_user_evicts().
109  *
110  * Note that the majority of the performance stats are manipulated
111  * with atomic operations.
112  */
113 
114 #include <sys/spa.h>
115 #include <sys/zio.h>
116 #include <sys/zio_checksum.h>
117 #include <sys/zfs_context.h>
118 #include <sys/arc.h>
119 #include <sys/refcount.h>
120 #ifdef _KERNEL
121 #include <sys/vmsystm.h>
122 #include <vm/anon.h>
123 #include <sys/fs/swapnode.h>
124 #include <sys/dnlc.h>
125 #endif
126 #include <sys/callb.h>
127 #include <sys/kstat.h>
128 
129 static kmutex_t		arc_reclaim_thr_lock;
130 static kcondvar_t	arc_reclaim_thr_cv;	/* used to signal reclaim thr */
131 static uint8_t		arc_thread_exit;
132 
133 #define	ARC_REDUCE_DNLC_PERCENT	3
134 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
135 
136 typedef enum arc_reclaim_strategy {
137 	ARC_RECLAIM_AGGR,		/* Aggressive reclaim strategy */
138 	ARC_RECLAIM_CONS		/* Conservative reclaim strategy */
139 } arc_reclaim_strategy_t;
140 
141 /* number of seconds before growing cache again */
142 static int		arc_grow_retry = 60;
143 
144 /*
145  * minimum lifespan of a prefetch block in clock ticks
146  * (initialized in arc_init())
147  */
148 static int		arc_min_prefetch_lifespan;
149 
150 static int arc_dead;
151 
152 /*
153  * These tunables are for performance analysis.
154  */
155 uint64_t zfs_arc_max;
156 uint64_t zfs_arc_min;
157 uint64_t zfs_arc_meta_limit = 0;
158 
159 /*
160  * Note that buffers can be in one of 5 states:
161  *	ARC_anon	- anonymous (discussed below)
162  *	ARC_mru		- recently used, currently cached
163  *	ARC_mru_ghost	- recentely used, no longer in cache
164  *	ARC_mfu		- frequently used, currently cached
165  *	ARC_mfu_ghost	- frequently used, no longer in cache
166  * When there are no active references to the buffer, they are
167  * are linked onto a list in one of these arc states.  These are
168  * the only buffers that can be evicted or deleted.  Within each
169  * state there are multiple lists, one for meta-data and one for
170  * non-meta-data.  Meta-data (indirect blocks, blocks of dnodes,
171  * etc.) is tracked separately so that it can be managed more
172  * explicitly: favored over data, limited explicitely.
173  *
174  * Anonymous buffers are buffers that are not associated with
175  * a DVA.  These are buffers that hold dirty block copies
176  * before they are written to stable storage.  By definition,
177  * they are "ref'd" and are considered part of arc_mru
178  * that cannot be freed.  Generally, they will aquire a DVA
179  * as they are written and migrate onto the arc_mru list.
180  */
181 
182 typedef struct arc_state {
183 	list_t	arcs_list[ARC_BUFC_NUMTYPES];	/* list of evictable buffers */
184 	uint64_t arcs_lsize[ARC_BUFC_NUMTYPES];	/* amount of evictable data */
185 	uint64_t arcs_size;	/* total amount of data in this state */
186 	kmutex_t arcs_mtx;
187 } arc_state_t;
188 
189 /* The 5 states: */
190 static arc_state_t ARC_anon;
191 static arc_state_t ARC_mru;
192 static arc_state_t ARC_mru_ghost;
193 static arc_state_t ARC_mfu;
194 static arc_state_t ARC_mfu_ghost;
195 
196 typedef struct arc_stats {
197 	kstat_named_t arcstat_hits;
198 	kstat_named_t arcstat_misses;
199 	kstat_named_t arcstat_demand_data_hits;
200 	kstat_named_t arcstat_demand_data_misses;
201 	kstat_named_t arcstat_demand_metadata_hits;
202 	kstat_named_t arcstat_demand_metadata_misses;
203 	kstat_named_t arcstat_prefetch_data_hits;
204 	kstat_named_t arcstat_prefetch_data_misses;
205 	kstat_named_t arcstat_prefetch_metadata_hits;
206 	kstat_named_t arcstat_prefetch_metadata_misses;
207 	kstat_named_t arcstat_mru_hits;
208 	kstat_named_t arcstat_mru_ghost_hits;
209 	kstat_named_t arcstat_mfu_hits;
210 	kstat_named_t arcstat_mfu_ghost_hits;
211 	kstat_named_t arcstat_deleted;
212 	kstat_named_t arcstat_recycle_miss;
213 	kstat_named_t arcstat_mutex_miss;
214 	kstat_named_t arcstat_evict_skip;
215 	kstat_named_t arcstat_hash_elements;
216 	kstat_named_t arcstat_hash_elements_max;
217 	kstat_named_t arcstat_hash_collisions;
218 	kstat_named_t arcstat_hash_chains;
219 	kstat_named_t arcstat_hash_chain_max;
220 	kstat_named_t arcstat_p;
221 	kstat_named_t arcstat_c;
222 	kstat_named_t arcstat_c_min;
223 	kstat_named_t arcstat_c_max;
224 	kstat_named_t arcstat_size;
225 } arc_stats_t;
226 
227 static arc_stats_t arc_stats = {
228 	{ "hits",			KSTAT_DATA_UINT64 },
229 	{ "misses",			KSTAT_DATA_UINT64 },
230 	{ "demand_data_hits",		KSTAT_DATA_UINT64 },
231 	{ "demand_data_misses",		KSTAT_DATA_UINT64 },
232 	{ "demand_metadata_hits",	KSTAT_DATA_UINT64 },
233 	{ "demand_metadata_misses",	KSTAT_DATA_UINT64 },
234 	{ "prefetch_data_hits",		KSTAT_DATA_UINT64 },
235 	{ "prefetch_data_misses",	KSTAT_DATA_UINT64 },
236 	{ "prefetch_metadata_hits",	KSTAT_DATA_UINT64 },
237 	{ "prefetch_metadata_misses",	KSTAT_DATA_UINT64 },
238 	{ "mru_hits",			KSTAT_DATA_UINT64 },
239 	{ "mru_ghost_hits",		KSTAT_DATA_UINT64 },
240 	{ "mfu_hits",			KSTAT_DATA_UINT64 },
241 	{ "mfu_ghost_hits",		KSTAT_DATA_UINT64 },
242 	{ "deleted",			KSTAT_DATA_UINT64 },
243 	{ "recycle_miss",		KSTAT_DATA_UINT64 },
244 	{ "mutex_miss",			KSTAT_DATA_UINT64 },
245 	{ "evict_skip",			KSTAT_DATA_UINT64 },
246 	{ "hash_elements",		KSTAT_DATA_UINT64 },
247 	{ "hash_elements_max",		KSTAT_DATA_UINT64 },
248 	{ "hash_collisions",		KSTAT_DATA_UINT64 },
249 	{ "hash_chains",		KSTAT_DATA_UINT64 },
250 	{ "hash_chain_max",		KSTAT_DATA_UINT64 },
251 	{ "p",				KSTAT_DATA_UINT64 },
252 	{ "c",				KSTAT_DATA_UINT64 },
253 	{ "c_min",			KSTAT_DATA_UINT64 },
254 	{ "c_max",			KSTAT_DATA_UINT64 },
255 	{ "size",			KSTAT_DATA_UINT64 }
256 };
257 
258 #define	ARCSTAT(stat)	(arc_stats.stat.value.ui64)
259 
260 #define	ARCSTAT_INCR(stat, val) \
261 	atomic_add_64(&arc_stats.stat.value.ui64, (val));
262 
263 #define	ARCSTAT_BUMP(stat) 	ARCSTAT_INCR(stat, 1)
264 #define	ARCSTAT_BUMPDOWN(stat)	ARCSTAT_INCR(stat, -1)
265 
266 #define	ARCSTAT_MAX(stat, val) {					\
267 	uint64_t m;							\
268 	while ((val) > (m = arc_stats.stat.value.ui64) &&		\
269 	    (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val))))	\
270 		continue;						\
271 }
272 
273 #define	ARCSTAT_MAXSTAT(stat) \
274 	ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
275 
276 /*
277  * We define a macro to allow ARC hits/misses to be easily broken down by
278  * two separate conditions, giving a total of four different subtypes for
279  * each of hits and misses (so eight statistics total).
280  */
281 #define	ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
282 	if (cond1) {							\
283 		if (cond2) {						\
284 			ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
285 		} else {						\
286 			ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
287 		}							\
288 	} else {							\
289 		if (cond2) {						\
290 			ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
291 		} else {						\
292 			ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
293 		}							\
294 	}
295 
296 kstat_t			*arc_ksp;
297 static arc_state_t 	*arc_anon;
298 static arc_state_t	*arc_mru;
299 static arc_state_t	*arc_mru_ghost;
300 static arc_state_t	*arc_mfu;
301 static arc_state_t	*arc_mfu_ghost;
302 
303 /*
304  * There are several ARC variables that are critical to export as kstats --
305  * but we don't want to have to grovel around in the kstat whenever we wish to
306  * manipulate them.  For these variables, we therefore define them to be in
307  * terms of the statistic variable.  This assures that we are not introducing
308  * the possibility of inconsistency by having shadow copies of the variables,
309  * while still allowing the code to be readable.
310  */
311 #define	arc_size	ARCSTAT(arcstat_size)	/* actual total arc size */
312 #define	arc_p		ARCSTAT(arcstat_p)	/* target size of MRU */
313 #define	arc_c		ARCSTAT(arcstat_c)	/* target size of cache */
314 #define	arc_c_min	ARCSTAT(arcstat_c_min)	/* min target cache size */
315 #define	arc_c_max	ARCSTAT(arcstat_c_max)	/* max target cache size */
316 
317 static int		arc_no_grow;	/* Don't try to grow cache size */
318 static uint64_t		arc_tempreserve;
319 static uint64_t		arc_meta_used;
320 static uint64_t		arc_meta_limit;
321 static uint64_t		arc_meta_max = 0;
322 
323 typedef struct arc_callback arc_callback_t;
324 
325 struct arc_callback {
326 	void			*acb_private;
327 	arc_done_func_t		*acb_done;
328 	arc_byteswap_func_t	*acb_byteswap;
329 	arc_buf_t		*acb_buf;
330 	zio_t			*acb_zio_dummy;
331 	arc_callback_t		*acb_next;
332 };
333 
334 typedef struct arc_write_callback arc_write_callback_t;
335 
336 struct arc_write_callback {
337 	void		*awcb_private;
338 	arc_done_func_t	*awcb_ready;
339 	arc_done_func_t	*awcb_done;
340 	arc_buf_t	*awcb_buf;
341 };
342 
343 struct arc_buf_hdr {
344 	/* protected by hash lock */
345 	dva_t			b_dva;
346 	uint64_t		b_birth;
347 	uint64_t		b_cksum0;
348 
349 	kmutex_t		b_freeze_lock;
350 	zio_cksum_t		*b_freeze_cksum;
351 
352 	arc_buf_hdr_t		*b_hash_next;
353 	arc_buf_t		*b_buf;
354 	uint32_t		b_flags;
355 	uint32_t		b_datacnt;
356 
357 	arc_callback_t		*b_acb;
358 	kcondvar_t		b_cv;
359 
360 	/* immutable */
361 	arc_buf_contents_t	b_type;
362 	uint64_t		b_size;
363 	spa_t			*b_spa;
364 
365 	/* protected by arc state mutex */
366 	arc_state_t		*b_state;
367 	list_node_t		b_arc_node;
368 
369 	/* updated atomically */
370 	clock_t			b_arc_access;
371 
372 	/* self protecting */
373 	refcount_t		b_refcnt;
374 };
375 
376 static arc_buf_t *arc_eviction_list;
377 static kmutex_t arc_eviction_mtx;
378 static arc_buf_hdr_t arc_eviction_hdr;
379 static void arc_get_data_buf(arc_buf_t *buf);
380 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
381 static int arc_evict_needed(arc_buf_contents_t type);
382 
383 #define	GHOST_STATE(state)	\
384 	((state) == arc_mru_ghost || (state) == arc_mfu_ghost)
385 
386 /*
387  * Private ARC flags.  These flags are private ARC only flags that will show up
388  * in b_flags in the arc_hdr_buf_t.  Some flags are publicly declared, and can
389  * be passed in as arc_flags in things like arc_read.  However, these flags
390  * should never be passed and should only be set by ARC code.  When adding new
391  * public flags, make sure not to smash the private ones.
392  */
393 
394 #define	ARC_IN_HASH_TABLE	(1 << 9)	/* this buffer is hashed */
395 #define	ARC_IO_IN_PROGRESS	(1 << 10)	/* I/O in progress for buf */
396 #define	ARC_IO_ERROR		(1 << 11)	/* I/O failed for buf */
397 #define	ARC_FREED_IN_READ	(1 << 12)	/* buf freed while in read */
398 #define	ARC_BUF_AVAILABLE	(1 << 13)	/* block not in active use */
399 #define	ARC_INDIRECT		(1 << 14)	/* this is an indirect block */
400 
401 #define	HDR_IN_HASH_TABLE(hdr)	((hdr)->b_flags & ARC_IN_HASH_TABLE)
402 #define	HDR_IO_IN_PROGRESS(hdr)	((hdr)->b_flags & ARC_IO_IN_PROGRESS)
403 #define	HDR_IO_ERROR(hdr)	((hdr)->b_flags & ARC_IO_ERROR)
404 #define	HDR_FREED_IN_READ(hdr)	((hdr)->b_flags & ARC_FREED_IN_READ)
405 #define	HDR_BUF_AVAILABLE(hdr)	((hdr)->b_flags & ARC_BUF_AVAILABLE)
406 
407 /*
408  * Hash table routines
409  */
410 
411 #define	HT_LOCK_PAD	64
412 
413 struct ht_lock {
414 	kmutex_t	ht_lock;
415 #ifdef _KERNEL
416 	unsigned char	pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
417 #endif
418 };
419 
420 #define	BUF_LOCKS 256
421 typedef struct buf_hash_table {
422 	uint64_t ht_mask;
423 	arc_buf_hdr_t **ht_table;
424 	struct ht_lock ht_locks[BUF_LOCKS];
425 } buf_hash_table_t;
426 
427 static buf_hash_table_t buf_hash_table;
428 
429 #define	BUF_HASH_INDEX(spa, dva, birth) \
430 	(buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
431 #define	BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
432 #define	BUF_HASH_LOCK(idx)	(&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
433 #define	HDR_LOCK(buf) \
434 	(BUF_HASH_LOCK(BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth)))
435 
436 uint64_t zfs_crc64_table[256];
437 
438 static uint64_t
439 buf_hash(spa_t *spa, dva_t *dva, uint64_t birth)
440 {
441 	uintptr_t spav = (uintptr_t)spa;
442 	uint8_t *vdva = (uint8_t *)dva;
443 	uint64_t crc = -1ULL;
444 	int i;
445 
446 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
447 
448 	for (i = 0; i < sizeof (dva_t); i++)
449 		crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
450 
451 	crc ^= (spav>>8) ^ birth;
452 
453 	return (crc);
454 }
455 
456 #define	BUF_EMPTY(buf)						\
457 	((buf)->b_dva.dva_word[0] == 0 &&			\
458 	(buf)->b_dva.dva_word[1] == 0 &&			\
459 	(buf)->b_birth == 0)
460 
461 #define	BUF_EQUAL(spa, dva, birth, buf)				\
462 	((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) &&	\
463 	((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) &&	\
464 	((buf)->b_birth == birth) && ((buf)->b_spa == spa)
465 
466 static arc_buf_hdr_t *
467 buf_hash_find(spa_t *spa, dva_t *dva, uint64_t birth, kmutex_t **lockp)
468 {
469 	uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
470 	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
471 	arc_buf_hdr_t *buf;
472 
473 	mutex_enter(hash_lock);
474 	for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
475 	    buf = buf->b_hash_next) {
476 		if (BUF_EQUAL(spa, dva, birth, buf)) {
477 			*lockp = hash_lock;
478 			return (buf);
479 		}
480 	}
481 	mutex_exit(hash_lock);
482 	*lockp = NULL;
483 	return (NULL);
484 }
485 
486 /*
487  * Insert an entry into the hash table.  If there is already an element
488  * equal to elem in the hash table, then the already existing element
489  * will be returned and the new element will not be inserted.
490  * Otherwise returns NULL.
491  */
492 static arc_buf_hdr_t *
493 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
494 {
495 	uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
496 	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
497 	arc_buf_hdr_t *fbuf;
498 	uint32_t i;
499 
500 	ASSERT(!HDR_IN_HASH_TABLE(buf));
501 	*lockp = hash_lock;
502 	mutex_enter(hash_lock);
503 	for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
504 	    fbuf = fbuf->b_hash_next, i++) {
505 		if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
506 			return (fbuf);
507 	}
508 
509 	buf->b_hash_next = buf_hash_table.ht_table[idx];
510 	buf_hash_table.ht_table[idx] = buf;
511 	buf->b_flags |= ARC_IN_HASH_TABLE;
512 
513 	/* collect some hash table performance data */
514 	if (i > 0) {
515 		ARCSTAT_BUMP(arcstat_hash_collisions);
516 		if (i == 1)
517 			ARCSTAT_BUMP(arcstat_hash_chains);
518 
519 		ARCSTAT_MAX(arcstat_hash_chain_max, i);
520 	}
521 
522 	ARCSTAT_BUMP(arcstat_hash_elements);
523 	ARCSTAT_MAXSTAT(arcstat_hash_elements);
524 
525 	return (NULL);
526 }
527 
528 static void
529 buf_hash_remove(arc_buf_hdr_t *buf)
530 {
531 	arc_buf_hdr_t *fbuf, **bufp;
532 	uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
533 
534 	ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
535 	ASSERT(HDR_IN_HASH_TABLE(buf));
536 
537 	bufp = &buf_hash_table.ht_table[idx];
538 	while ((fbuf = *bufp) != buf) {
539 		ASSERT(fbuf != NULL);
540 		bufp = &fbuf->b_hash_next;
541 	}
542 	*bufp = buf->b_hash_next;
543 	buf->b_hash_next = NULL;
544 	buf->b_flags &= ~ARC_IN_HASH_TABLE;
545 
546 	/* collect some hash table performance data */
547 	ARCSTAT_BUMPDOWN(arcstat_hash_elements);
548 
549 	if (buf_hash_table.ht_table[idx] &&
550 	    buf_hash_table.ht_table[idx]->b_hash_next == NULL)
551 		ARCSTAT_BUMPDOWN(arcstat_hash_chains);
552 }
553 
554 /*
555  * Global data structures and functions for the buf kmem cache.
556  */
557 static kmem_cache_t *hdr_cache;
558 static kmem_cache_t *buf_cache;
559 
560 static void
561 buf_fini(void)
562 {
563 	int i;
564 
565 	kmem_free(buf_hash_table.ht_table,
566 	    (buf_hash_table.ht_mask + 1) * sizeof (void *));
567 	for (i = 0; i < BUF_LOCKS; i++)
568 		mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
569 	kmem_cache_destroy(hdr_cache);
570 	kmem_cache_destroy(buf_cache);
571 }
572 
573 /*
574  * Constructor callback - called when the cache is empty
575  * and a new buf is requested.
576  */
577 /* ARGSUSED */
578 static int
579 hdr_cons(void *vbuf, void *unused, int kmflag)
580 {
581 	arc_buf_hdr_t *buf = vbuf;
582 
583 	bzero(buf, sizeof (arc_buf_hdr_t));
584 	refcount_create(&buf->b_refcnt);
585 	cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
586 	return (0);
587 }
588 
589 /*
590  * Destructor callback - called when a cached buf is
591  * no longer required.
592  */
593 /* ARGSUSED */
594 static void
595 hdr_dest(void *vbuf, void *unused)
596 {
597 	arc_buf_hdr_t *buf = vbuf;
598 
599 	refcount_destroy(&buf->b_refcnt);
600 	cv_destroy(&buf->b_cv);
601 }
602 
603 /*
604  * Reclaim callback -- invoked when memory is low.
605  */
606 /* ARGSUSED */
607 static void
608 hdr_recl(void *unused)
609 {
610 	dprintf("hdr_recl called\n");
611 	/*
612 	 * umem calls the reclaim func when we destroy the buf cache,
613 	 * which is after we do arc_fini().
614 	 */
615 	if (!arc_dead)
616 		cv_signal(&arc_reclaim_thr_cv);
617 }
618 
619 static void
620 buf_init(void)
621 {
622 	uint64_t *ct;
623 	uint64_t hsize = 1ULL << 12;
624 	int i, j;
625 
626 	/*
627 	 * The hash table is big enough to fill all of physical memory
628 	 * with an average 64K block size.  The table will take up
629 	 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
630 	 */
631 	while (hsize * 65536 < physmem * PAGESIZE)
632 		hsize <<= 1;
633 retry:
634 	buf_hash_table.ht_mask = hsize - 1;
635 	buf_hash_table.ht_table =
636 	    kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
637 	if (buf_hash_table.ht_table == NULL) {
638 		ASSERT(hsize > (1ULL << 8));
639 		hsize >>= 1;
640 		goto retry;
641 	}
642 
643 	hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
644 	    0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
645 	buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
646 	    0, NULL, NULL, NULL, NULL, NULL, 0);
647 
648 	for (i = 0; i < 256; i++)
649 		for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
650 			*ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
651 
652 	for (i = 0; i < BUF_LOCKS; i++) {
653 		mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
654 		    NULL, MUTEX_DEFAULT, NULL);
655 	}
656 }
657 
658 #define	ARC_MINTIME	(hz>>4) /* 62 ms */
659 
660 static void
661 arc_cksum_verify(arc_buf_t *buf)
662 {
663 	zio_cksum_t zc;
664 
665 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
666 		return;
667 
668 	mutex_enter(&buf->b_hdr->b_freeze_lock);
669 	if (buf->b_hdr->b_freeze_cksum == NULL ||
670 	    (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
671 		mutex_exit(&buf->b_hdr->b_freeze_lock);
672 		return;
673 	}
674 	fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
675 	if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
676 		panic("buffer modified while frozen!");
677 	mutex_exit(&buf->b_hdr->b_freeze_lock);
678 }
679 
680 static void
681 arc_cksum_compute(arc_buf_t *buf)
682 {
683 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
684 		return;
685 
686 	mutex_enter(&buf->b_hdr->b_freeze_lock);
687 	if (buf->b_hdr->b_freeze_cksum != NULL) {
688 		mutex_exit(&buf->b_hdr->b_freeze_lock);
689 		return;
690 	}
691 	buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
692 	fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
693 	    buf->b_hdr->b_freeze_cksum);
694 	mutex_exit(&buf->b_hdr->b_freeze_lock);
695 }
696 
697 void
698 arc_buf_thaw(arc_buf_t *buf)
699 {
700 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
701 		return;
702 
703 	if (buf->b_hdr->b_state != arc_anon)
704 		panic("modifying non-anon buffer!");
705 	if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
706 		panic("modifying buffer while i/o in progress!");
707 	arc_cksum_verify(buf);
708 	mutex_enter(&buf->b_hdr->b_freeze_lock);
709 	if (buf->b_hdr->b_freeze_cksum != NULL) {
710 		kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
711 		buf->b_hdr->b_freeze_cksum = NULL;
712 	}
713 	mutex_exit(&buf->b_hdr->b_freeze_lock);
714 }
715 
716 void
717 arc_buf_freeze(arc_buf_t *buf)
718 {
719 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
720 		return;
721 
722 	ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
723 	    buf->b_hdr->b_state == arc_anon);
724 	arc_cksum_compute(buf);
725 }
726 
727 static void
728 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
729 {
730 	ASSERT(MUTEX_HELD(hash_lock));
731 
732 	if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
733 	    (ab->b_state != arc_anon)) {
734 		uint64_t delta = ab->b_size * ab->b_datacnt;
735 		list_t *list = &ab->b_state->arcs_list[ab->b_type];
736 		uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
737 
738 		ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
739 		mutex_enter(&ab->b_state->arcs_mtx);
740 		ASSERT(list_link_active(&ab->b_arc_node));
741 		list_remove(list, ab);
742 		if (GHOST_STATE(ab->b_state)) {
743 			ASSERT3U(ab->b_datacnt, ==, 0);
744 			ASSERT3P(ab->b_buf, ==, NULL);
745 			delta = ab->b_size;
746 		}
747 		ASSERT(delta > 0);
748 		ASSERT3U(*size, >=, delta);
749 		atomic_add_64(size, -delta);
750 		mutex_exit(&ab->b_state->arcs_mtx);
751 		/* remove the prefetch flag is we get a reference */
752 		if (ab->b_flags & ARC_PREFETCH)
753 			ab->b_flags &= ~ARC_PREFETCH;
754 	}
755 }
756 
757 static int
758 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
759 {
760 	int cnt;
761 	arc_state_t *state = ab->b_state;
762 
763 	ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
764 	ASSERT(!GHOST_STATE(state));
765 
766 	if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
767 	    (state != arc_anon)) {
768 		uint64_t *size = &state->arcs_lsize[ab->b_type];
769 
770 		ASSERT(!MUTEX_HELD(&state->arcs_mtx));
771 		mutex_enter(&state->arcs_mtx);
772 		ASSERT(!list_link_active(&ab->b_arc_node));
773 		list_insert_head(&state->arcs_list[ab->b_type], ab);
774 		ASSERT(ab->b_datacnt > 0);
775 		atomic_add_64(size, ab->b_size * ab->b_datacnt);
776 		mutex_exit(&state->arcs_mtx);
777 	}
778 	return (cnt);
779 }
780 
781 /*
782  * Move the supplied buffer to the indicated state.  The mutex
783  * for the buffer must be held by the caller.
784  */
785 static void
786 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
787 {
788 	arc_state_t *old_state = ab->b_state;
789 	int64_t refcnt = refcount_count(&ab->b_refcnt);
790 	uint64_t from_delta, to_delta;
791 
792 	ASSERT(MUTEX_HELD(hash_lock));
793 	ASSERT(new_state != old_state);
794 	ASSERT(refcnt == 0 || ab->b_datacnt > 0);
795 	ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
796 
797 	from_delta = to_delta = ab->b_datacnt * ab->b_size;
798 
799 	/*
800 	 * If this buffer is evictable, transfer it from the
801 	 * old state list to the new state list.
802 	 */
803 	if (refcnt == 0) {
804 		if (old_state != arc_anon) {
805 			int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
806 			uint64_t *size = &old_state->arcs_lsize[ab->b_type];
807 
808 			if (use_mutex)
809 				mutex_enter(&old_state->arcs_mtx);
810 
811 			ASSERT(list_link_active(&ab->b_arc_node));
812 			list_remove(&old_state->arcs_list[ab->b_type], ab);
813 
814 			/*
815 			 * If prefetching out of the ghost cache,
816 			 * we will have a non-null datacnt.
817 			 */
818 			if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
819 				/* ghost elements have a ghost size */
820 				ASSERT(ab->b_buf == NULL);
821 				from_delta = ab->b_size;
822 			}
823 			ASSERT3U(*size, >=, from_delta);
824 			atomic_add_64(size, -from_delta);
825 
826 			if (use_mutex)
827 				mutex_exit(&old_state->arcs_mtx);
828 		}
829 		if (new_state != arc_anon) {
830 			int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
831 			uint64_t *size = &new_state->arcs_lsize[ab->b_type];
832 
833 			if (use_mutex)
834 				mutex_enter(&new_state->arcs_mtx);
835 
836 			list_insert_head(&new_state->arcs_list[ab->b_type], ab);
837 
838 			/* ghost elements have a ghost size */
839 			if (GHOST_STATE(new_state)) {
840 				ASSERT(ab->b_datacnt == 0);
841 				ASSERT(ab->b_buf == NULL);
842 				to_delta = ab->b_size;
843 			}
844 			atomic_add_64(size, to_delta);
845 			ASSERT3U(new_state->arcs_size + to_delta, >=, *size);
846 
847 			if (use_mutex)
848 				mutex_exit(&new_state->arcs_mtx);
849 		}
850 	}
851 
852 	ASSERT(!BUF_EMPTY(ab));
853 	if (new_state == arc_anon && old_state != arc_anon) {
854 		buf_hash_remove(ab);
855 	}
856 
857 	/* adjust state sizes */
858 	if (to_delta)
859 		atomic_add_64(&new_state->arcs_size, to_delta);
860 	if (from_delta) {
861 		ASSERT3U(old_state->arcs_size, >=, from_delta);
862 		atomic_add_64(&old_state->arcs_size, -from_delta);
863 	}
864 	ab->b_state = new_state;
865 }
866 
867 void
868 arc_space_consume(uint64_t space)
869 {
870 	atomic_add_64(&arc_meta_used, space);
871 	atomic_add_64(&arc_size, space);
872 }
873 
874 void
875 arc_space_return(uint64_t space)
876 {
877 	ASSERT(arc_meta_used >= space);
878 	if (arc_meta_max < arc_meta_used)
879 		arc_meta_max = arc_meta_used;
880 	atomic_add_64(&arc_meta_used, -space);
881 	ASSERT(arc_size >= space);
882 	atomic_add_64(&arc_size, -space);
883 }
884 
885 void *
886 arc_data_buf_alloc(uint64_t size)
887 {
888 	if (arc_evict_needed(ARC_BUFC_DATA))
889 		cv_signal(&arc_reclaim_thr_cv);
890 	atomic_add_64(&arc_size, size);
891 	return (zio_data_buf_alloc(size));
892 }
893 
894 void
895 arc_data_buf_free(void *buf, uint64_t size)
896 {
897 	zio_data_buf_free(buf, size);
898 	ASSERT(arc_size >= size);
899 	atomic_add_64(&arc_size, -size);
900 }
901 
902 arc_buf_t *
903 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
904 {
905 	arc_buf_hdr_t *hdr;
906 	arc_buf_t *buf;
907 
908 	ASSERT3U(size, >, 0);
909 	hdr = kmem_cache_alloc(hdr_cache, KM_SLEEP);
910 	ASSERT(BUF_EMPTY(hdr));
911 	hdr->b_size = size;
912 	hdr->b_type = type;
913 	hdr->b_spa = spa;
914 	hdr->b_state = arc_anon;
915 	hdr->b_arc_access = 0;
916 	buf = kmem_cache_alloc(buf_cache, KM_SLEEP);
917 	buf->b_hdr = hdr;
918 	buf->b_data = NULL;
919 	buf->b_efunc = NULL;
920 	buf->b_private = NULL;
921 	buf->b_next = NULL;
922 	hdr->b_buf = buf;
923 	arc_get_data_buf(buf);
924 	hdr->b_datacnt = 1;
925 	hdr->b_flags = 0;
926 	ASSERT(refcount_is_zero(&hdr->b_refcnt));
927 	(void) refcount_add(&hdr->b_refcnt, tag);
928 
929 	return (buf);
930 }
931 
932 static arc_buf_t *
933 arc_buf_clone(arc_buf_t *from)
934 {
935 	arc_buf_t *buf;
936 	arc_buf_hdr_t *hdr = from->b_hdr;
937 	uint64_t size = hdr->b_size;
938 
939 	buf = kmem_cache_alloc(buf_cache, KM_SLEEP);
940 	buf->b_hdr = hdr;
941 	buf->b_data = NULL;
942 	buf->b_efunc = NULL;
943 	buf->b_private = NULL;
944 	buf->b_next = hdr->b_buf;
945 	hdr->b_buf = buf;
946 	arc_get_data_buf(buf);
947 	bcopy(from->b_data, buf->b_data, size);
948 	hdr->b_datacnt += 1;
949 	return (buf);
950 }
951 
952 void
953 arc_buf_add_ref(arc_buf_t *buf, void* tag)
954 {
955 	arc_buf_hdr_t *hdr;
956 	kmutex_t *hash_lock;
957 
958 	/*
959 	 * Check to see if this buffer is currently being evicted via
960 	 * arc_do_user_evicts().
961 	 */
962 	mutex_enter(&arc_eviction_mtx);
963 	hdr = buf->b_hdr;
964 	if (hdr == NULL) {
965 		mutex_exit(&arc_eviction_mtx);
966 		return;
967 	}
968 	hash_lock = HDR_LOCK(hdr);
969 	mutex_exit(&arc_eviction_mtx);
970 
971 	mutex_enter(hash_lock);
972 	if (buf->b_data == NULL) {
973 		/*
974 		 * This buffer is evicted.
975 		 */
976 		mutex_exit(hash_lock);
977 		return;
978 	}
979 
980 	ASSERT(buf->b_hdr == hdr);
981 	ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
982 	add_reference(hdr, hash_lock, tag);
983 	arc_access(hdr, hash_lock);
984 	mutex_exit(hash_lock);
985 	ARCSTAT_BUMP(arcstat_hits);
986 	ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
987 	    demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
988 	    data, metadata, hits);
989 }
990 
991 static void
992 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
993 {
994 	arc_buf_t **bufp;
995 
996 	/* free up data associated with the buf */
997 	if (buf->b_data) {
998 		arc_state_t *state = buf->b_hdr->b_state;
999 		uint64_t size = buf->b_hdr->b_size;
1000 		arc_buf_contents_t type = buf->b_hdr->b_type;
1001 
1002 		arc_cksum_verify(buf);
1003 		if (!recycle) {
1004 			if (type == ARC_BUFC_METADATA) {
1005 				zio_buf_free(buf->b_data, size);
1006 				arc_space_return(size);
1007 			} else {
1008 				ASSERT(type == ARC_BUFC_DATA);
1009 				zio_data_buf_free(buf->b_data, size);
1010 				atomic_add_64(&arc_size, -size);
1011 			}
1012 		}
1013 		if (list_link_active(&buf->b_hdr->b_arc_node)) {
1014 			uint64_t *cnt = &state->arcs_lsize[type];
1015 
1016 			ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1017 			ASSERT(state != arc_anon);
1018 
1019 			ASSERT3U(*cnt, >=, size);
1020 			atomic_add_64(cnt, -size);
1021 		}
1022 		ASSERT3U(state->arcs_size, >=, size);
1023 		atomic_add_64(&state->arcs_size, -size);
1024 		buf->b_data = NULL;
1025 		ASSERT(buf->b_hdr->b_datacnt > 0);
1026 		buf->b_hdr->b_datacnt -= 1;
1027 	}
1028 
1029 	/* only remove the buf if requested */
1030 	if (!all)
1031 		return;
1032 
1033 	/* remove the buf from the hdr list */
1034 	for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1035 		continue;
1036 	*bufp = buf->b_next;
1037 
1038 	ASSERT(buf->b_efunc == NULL);
1039 
1040 	/* clean up the buf */
1041 	buf->b_hdr = NULL;
1042 	kmem_cache_free(buf_cache, buf);
1043 }
1044 
1045 static void
1046 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1047 {
1048 	ASSERT(refcount_is_zero(&hdr->b_refcnt));
1049 	ASSERT3P(hdr->b_state, ==, arc_anon);
1050 	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1051 
1052 	if (!BUF_EMPTY(hdr)) {
1053 		ASSERT(!HDR_IN_HASH_TABLE(hdr));
1054 		bzero(&hdr->b_dva, sizeof (dva_t));
1055 		hdr->b_birth = 0;
1056 		hdr->b_cksum0 = 0;
1057 	}
1058 	while (hdr->b_buf) {
1059 		arc_buf_t *buf = hdr->b_buf;
1060 
1061 		if (buf->b_efunc) {
1062 			mutex_enter(&arc_eviction_mtx);
1063 			ASSERT(buf->b_hdr != NULL);
1064 			arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1065 			hdr->b_buf = buf->b_next;
1066 			buf->b_hdr = &arc_eviction_hdr;
1067 			buf->b_next = arc_eviction_list;
1068 			arc_eviction_list = buf;
1069 			mutex_exit(&arc_eviction_mtx);
1070 		} else {
1071 			arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1072 		}
1073 	}
1074 	if (hdr->b_freeze_cksum != NULL) {
1075 		kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1076 		hdr->b_freeze_cksum = NULL;
1077 	}
1078 
1079 	ASSERT(!list_link_active(&hdr->b_arc_node));
1080 	ASSERT3P(hdr->b_hash_next, ==, NULL);
1081 	ASSERT3P(hdr->b_acb, ==, NULL);
1082 	kmem_cache_free(hdr_cache, hdr);
1083 }
1084 
1085 void
1086 arc_buf_free(arc_buf_t *buf, void *tag)
1087 {
1088 	arc_buf_hdr_t *hdr = buf->b_hdr;
1089 	int hashed = hdr->b_state != arc_anon;
1090 
1091 	ASSERT(buf->b_efunc == NULL);
1092 	ASSERT(buf->b_data != NULL);
1093 
1094 	if (hashed) {
1095 		kmutex_t *hash_lock = HDR_LOCK(hdr);
1096 
1097 		mutex_enter(hash_lock);
1098 		(void) remove_reference(hdr, hash_lock, tag);
1099 		if (hdr->b_datacnt > 1)
1100 			arc_buf_destroy(buf, FALSE, TRUE);
1101 		else
1102 			hdr->b_flags |= ARC_BUF_AVAILABLE;
1103 		mutex_exit(hash_lock);
1104 	} else if (HDR_IO_IN_PROGRESS(hdr)) {
1105 		int destroy_hdr;
1106 		/*
1107 		 * We are in the middle of an async write.  Don't destroy
1108 		 * this buffer unless the write completes before we finish
1109 		 * decrementing the reference count.
1110 		 */
1111 		mutex_enter(&arc_eviction_mtx);
1112 		(void) remove_reference(hdr, NULL, tag);
1113 		ASSERT(refcount_is_zero(&hdr->b_refcnt));
1114 		destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1115 		mutex_exit(&arc_eviction_mtx);
1116 		if (destroy_hdr)
1117 			arc_hdr_destroy(hdr);
1118 	} else {
1119 		if (remove_reference(hdr, NULL, tag) > 0) {
1120 			ASSERT(HDR_IO_ERROR(hdr));
1121 			arc_buf_destroy(buf, FALSE, TRUE);
1122 		} else {
1123 			arc_hdr_destroy(hdr);
1124 		}
1125 	}
1126 }
1127 
1128 int
1129 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1130 {
1131 	arc_buf_hdr_t *hdr = buf->b_hdr;
1132 	kmutex_t *hash_lock = HDR_LOCK(hdr);
1133 	int no_callback = (buf->b_efunc == NULL);
1134 
1135 	if (hdr->b_state == arc_anon) {
1136 		arc_buf_free(buf, tag);
1137 		return (no_callback);
1138 	}
1139 
1140 	mutex_enter(hash_lock);
1141 	ASSERT(hdr->b_state != arc_anon);
1142 	ASSERT(buf->b_data != NULL);
1143 
1144 	(void) remove_reference(hdr, hash_lock, tag);
1145 	if (hdr->b_datacnt > 1) {
1146 		if (no_callback)
1147 			arc_buf_destroy(buf, FALSE, TRUE);
1148 	} else if (no_callback) {
1149 		ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1150 		hdr->b_flags |= ARC_BUF_AVAILABLE;
1151 	}
1152 	ASSERT(no_callback || hdr->b_datacnt > 1 ||
1153 	    refcount_is_zero(&hdr->b_refcnt));
1154 	mutex_exit(hash_lock);
1155 	return (no_callback);
1156 }
1157 
1158 int
1159 arc_buf_size(arc_buf_t *buf)
1160 {
1161 	return (buf->b_hdr->b_size);
1162 }
1163 
1164 /*
1165  * Evict buffers from list until we've removed the specified number of
1166  * bytes.  Move the removed buffers to the appropriate evict state.
1167  * If the recycle flag is set, then attempt to "recycle" a buffer:
1168  * - look for a buffer to evict that is `bytes' long.
1169  * - return the data block from this buffer rather than freeing it.
1170  * This flag is used by callers that are trying to make space for a
1171  * new buffer in a full arc cache.
1172  */
1173 static void *
1174 arc_evict(arc_state_t *state, int64_t bytes, boolean_t recycle,
1175     arc_buf_contents_t type)
1176 {
1177 	arc_state_t *evicted_state;
1178 	uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1179 	arc_buf_hdr_t *ab, *ab_prev = NULL;
1180 	list_t *list = &state->arcs_list[type];
1181 	kmutex_t *hash_lock;
1182 	boolean_t have_lock;
1183 	void *stolen = NULL;
1184 
1185 	ASSERT(state == arc_mru || state == arc_mfu);
1186 
1187 	evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1188 
1189 	mutex_enter(&state->arcs_mtx);
1190 	mutex_enter(&evicted_state->arcs_mtx);
1191 
1192 	for (ab = list_tail(list); ab; ab = ab_prev) {
1193 		ab_prev = list_prev(list, ab);
1194 		/* prefetch buffers have a minimum lifespan */
1195 		if (HDR_IO_IN_PROGRESS(ab) ||
1196 		    (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1197 		    lbolt - ab->b_arc_access < arc_min_prefetch_lifespan)) {
1198 			skipped++;
1199 			continue;
1200 		}
1201 		/* "lookahead" for better eviction candidate */
1202 		if (recycle && ab->b_size != bytes &&
1203 		    ab_prev && ab_prev->b_size == bytes)
1204 			continue;
1205 		hash_lock = HDR_LOCK(ab);
1206 		have_lock = MUTEX_HELD(hash_lock);
1207 		if (have_lock || mutex_tryenter(hash_lock)) {
1208 			ASSERT3U(refcount_count(&ab->b_refcnt), ==, 0);
1209 			ASSERT(ab->b_datacnt > 0);
1210 			while (ab->b_buf) {
1211 				arc_buf_t *buf = ab->b_buf;
1212 				if (buf->b_data) {
1213 					bytes_evicted += ab->b_size;
1214 					if (recycle && ab->b_type == type &&
1215 					    ab->b_size == bytes) {
1216 						stolen = buf->b_data;
1217 						recycle = FALSE;
1218 					}
1219 				}
1220 				if (buf->b_efunc) {
1221 					mutex_enter(&arc_eviction_mtx);
1222 					arc_buf_destroy(buf,
1223 					    buf->b_data == stolen, FALSE);
1224 					ab->b_buf = buf->b_next;
1225 					buf->b_hdr = &arc_eviction_hdr;
1226 					buf->b_next = arc_eviction_list;
1227 					arc_eviction_list = buf;
1228 					mutex_exit(&arc_eviction_mtx);
1229 				} else {
1230 					arc_buf_destroy(buf,
1231 					    buf->b_data == stolen, TRUE);
1232 				}
1233 			}
1234 			ASSERT(ab->b_datacnt == 0);
1235 			arc_change_state(evicted_state, ab, hash_lock);
1236 			ASSERT(HDR_IN_HASH_TABLE(ab));
1237 			ab->b_flags = ARC_IN_HASH_TABLE;
1238 			DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1239 			if (!have_lock)
1240 				mutex_exit(hash_lock);
1241 			if (bytes >= 0 && bytes_evicted >= bytes)
1242 				break;
1243 		} else {
1244 			missed += 1;
1245 		}
1246 	}
1247 
1248 	mutex_exit(&evicted_state->arcs_mtx);
1249 	mutex_exit(&state->arcs_mtx);
1250 
1251 	if (bytes_evicted < bytes)
1252 		dprintf("only evicted %lld bytes from %x",
1253 		    (longlong_t)bytes_evicted, state);
1254 
1255 	if (skipped)
1256 		ARCSTAT_INCR(arcstat_evict_skip, skipped);
1257 
1258 	if (missed)
1259 		ARCSTAT_INCR(arcstat_mutex_miss, missed);
1260 
1261 	return (stolen);
1262 }
1263 
1264 /*
1265  * Remove buffers from list until we've removed the specified number of
1266  * bytes.  Destroy the buffers that are removed.
1267  */
1268 static void
1269 arc_evict_ghost(arc_state_t *state, int64_t bytes)
1270 {
1271 	arc_buf_hdr_t *ab, *ab_prev;
1272 	list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1273 	kmutex_t *hash_lock;
1274 	uint64_t bytes_deleted = 0;
1275 	uint64_t bufs_skipped = 0;
1276 
1277 	ASSERT(GHOST_STATE(state));
1278 top:
1279 	mutex_enter(&state->arcs_mtx);
1280 	for (ab = list_tail(list); ab; ab = ab_prev) {
1281 		ab_prev = list_prev(list, ab);
1282 		hash_lock = HDR_LOCK(ab);
1283 		if (mutex_tryenter(hash_lock)) {
1284 			ASSERT(!HDR_IO_IN_PROGRESS(ab));
1285 			ASSERT(ab->b_buf == NULL);
1286 			arc_change_state(arc_anon, ab, hash_lock);
1287 			mutex_exit(hash_lock);
1288 			ARCSTAT_BUMP(arcstat_deleted);
1289 			bytes_deleted += ab->b_size;
1290 			arc_hdr_destroy(ab);
1291 			DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1292 			if (bytes >= 0 && bytes_deleted >= bytes)
1293 				break;
1294 		} else {
1295 			if (bytes < 0) {
1296 				mutex_exit(&state->arcs_mtx);
1297 				mutex_enter(hash_lock);
1298 				mutex_exit(hash_lock);
1299 				goto top;
1300 			}
1301 			bufs_skipped += 1;
1302 		}
1303 	}
1304 	mutex_exit(&state->arcs_mtx);
1305 
1306 	if (list == &state->arcs_list[ARC_BUFC_DATA] &&
1307 	    (bytes < 0 || bytes_deleted < bytes)) {
1308 		list = &state->arcs_list[ARC_BUFC_METADATA];
1309 		goto top;
1310 	}
1311 
1312 	if (bufs_skipped) {
1313 		ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1314 		ASSERT(bytes >= 0);
1315 	}
1316 
1317 	if (bytes_deleted < bytes)
1318 		dprintf("only deleted %lld bytes from %p",
1319 		    (longlong_t)bytes_deleted, state);
1320 }
1321 
1322 static void
1323 arc_adjust(void)
1324 {
1325 	int64_t top_sz, mru_over, arc_over, todelete;
1326 
1327 	top_sz = arc_anon->arcs_size + arc_mru->arcs_size;
1328 
1329 	if (top_sz > arc_p && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
1330 		int64_t toevict =
1331 		    MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], top_sz - arc_p);
1332 		(void) arc_evict(arc_mru, toevict, FALSE, ARC_BUFC_DATA);
1333 		top_sz = arc_anon->arcs_size + arc_mru->arcs_size;
1334 	}
1335 
1336 	if (top_sz > arc_p && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1337 		int64_t toevict =
1338 		    MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], top_sz - arc_p);
1339 		(void) arc_evict(arc_mru, toevict, FALSE, ARC_BUFC_METADATA);
1340 		top_sz = arc_anon->arcs_size + arc_mru->arcs_size;
1341 	}
1342 
1343 	mru_over = top_sz + arc_mru_ghost->arcs_size - arc_c;
1344 
1345 	if (mru_over > 0) {
1346 		if (arc_mru_ghost->arcs_size > 0) {
1347 			todelete = MIN(arc_mru_ghost->arcs_size, mru_over);
1348 			arc_evict_ghost(arc_mru_ghost, todelete);
1349 		}
1350 	}
1351 
1352 	if ((arc_over = arc_size - arc_c) > 0) {
1353 		int64_t tbl_over;
1354 
1355 		if (arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
1356 			int64_t toevict =
1357 			    MIN(arc_mfu->arcs_lsize[ARC_BUFC_DATA], arc_over);
1358 			(void) arc_evict(arc_mfu, toevict, FALSE,
1359 			    ARC_BUFC_DATA);
1360 			arc_over = arc_size - arc_c;
1361 		}
1362 
1363 		if (arc_over > 0 &&
1364 		    arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1365 			int64_t toevict =
1366 			    MIN(arc_mfu->arcs_lsize[ARC_BUFC_METADATA],
1367 			    arc_over);
1368 			(void) arc_evict(arc_mfu, toevict, FALSE,
1369 			    ARC_BUFC_METADATA);
1370 		}
1371 
1372 		tbl_over = arc_size + arc_mru_ghost->arcs_size +
1373 		    arc_mfu_ghost->arcs_size - arc_c * 2;
1374 
1375 		if (tbl_over > 0 && arc_mfu_ghost->arcs_size > 0) {
1376 			todelete = MIN(arc_mfu_ghost->arcs_size, tbl_over);
1377 			arc_evict_ghost(arc_mfu_ghost, todelete);
1378 		}
1379 	}
1380 }
1381 
1382 static void
1383 arc_do_user_evicts(void)
1384 {
1385 	mutex_enter(&arc_eviction_mtx);
1386 	while (arc_eviction_list != NULL) {
1387 		arc_buf_t *buf = arc_eviction_list;
1388 		arc_eviction_list = buf->b_next;
1389 		buf->b_hdr = NULL;
1390 		mutex_exit(&arc_eviction_mtx);
1391 
1392 		if (buf->b_efunc != NULL)
1393 			VERIFY(buf->b_efunc(buf) == 0);
1394 
1395 		buf->b_efunc = NULL;
1396 		buf->b_private = NULL;
1397 		kmem_cache_free(buf_cache, buf);
1398 		mutex_enter(&arc_eviction_mtx);
1399 	}
1400 	mutex_exit(&arc_eviction_mtx);
1401 }
1402 
1403 /*
1404  * Flush all *evictable* data from the cache.
1405  * NOTE: this will not touch "active" (i.e. referenced) data.
1406  */
1407 void
1408 arc_flush(void)
1409 {
1410 	while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA]))
1411 		(void) arc_evict(arc_mru, -1, FALSE, ARC_BUFC_DATA);
1412 	while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA]))
1413 		(void) arc_evict(arc_mru, -1, FALSE, ARC_BUFC_METADATA);
1414 	while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA]))
1415 		(void) arc_evict(arc_mfu, -1, FALSE, ARC_BUFC_DATA);
1416 	while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA]))
1417 		(void) arc_evict(arc_mfu, -1, FALSE, ARC_BUFC_METADATA);
1418 
1419 	arc_evict_ghost(arc_mru_ghost, -1);
1420 	arc_evict_ghost(arc_mfu_ghost, -1);
1421 
1422 	mutex_enter(&arc_reclaim_thr_lock);
1423 	arc_do_user_evicts();
1424 	mutex_exit(&arc_reclaim_thr_lock);
1425 	ASSERT(arc_eviction_list == NULL);
1426 }
1427 
1428 int arc_shrink_shift = 5;		/* log2(fraction of arc to reclaim) */
1429 
1430 void
1431 arc_shrink(void)
1432 {
1433 	if (arc_c > arc_c_min) {
1434 		uint64_t to_free;
1435 
1436 #ifdef _KERNEL
1437 		to_free = MAX(arc_c >> arc_shrink_shift, ptob(needfree));
1438 #else
1439 		to_free = arc_c >> arc_shrink_shift;
1440 #endif
1441 		if (arc_c > arc_c_min + to_free)
1442 			atomic_add_64(&arc_c, -to_free);
1443 		else
1444 			arc_c = arc_c_min;
1445 
1446 		atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
1447 		if (arc_c > arc_size)
1448 			arc_c = MAX(arc_size, arc_c_min);
1449 		if (arc_p > arc_c)
1450 			arc_p = (arc_c >> 1);
1451 		ASSERT(arc_c >= arc_c_min);
1452 		ASSERT((int64_t)arc_p >= 0);
1453 	}
1454 
1455 	if (arc_size > arc_c)
1456 		arc_adjust();
1457 }
1458 
1459 static int
1460 arc_reclaim_needed(void)
1461 {
1462 	uint64_t extra;
1463 
1464 #ifdef _KERNEL
1465 
1466 	if (needfree)
1467 		return (1);
1468 
1469 	/*
1470 	 * take 'desfree' extra pages, so we reclaim sooner, rather than later
1471 	 */
1472 	extra = desfree;
1473 
1474 	/*
1475 	 * check that we're out of range of the pageout scanner.  It starts to
1476 	 * schedule paging if freemem is less than lotsfree and needfree.
1477 	 * lotsfree is the high-water mark for pageout, and needfree is the
1478 	 * number of needed free pages.  We add extra pages here to make sure
1479 	 * the scanner doesn't start up while we're freeing memory.
1480 	 */
1481 	if (freemem < lotsfree + needfree + extra)
1482 		return (1);
1483 
1484 	/*
1485 	 * check to make sure that swapfs has enough space so that anon
1486 	 * reservations can still succeeed. anon_resvmem() checks that the
1487 	 * availrmem is greater than swapfs_minfree, and the number of reserved
1488 	 * swap pages.  We also add a bit of extra here just to prevent
1489 	 * circumstances from getting really dire.
1490 	 */
1491 	if (availrmem < swapfs_minfree + swapfs_reserve + extra)
1492 		return (1);
1493 
1494 #if defined(__i386)
1495 	/*
1496 	 * If we're on an i386 platform, it's possible that we'll exhaust the
1497 	 * kernel heap space before we ever run out of available physical
1498 	 * memory.  Most checks of the size of the heap_area compare against
1499 	 * tune.t_minarmem, which is the minimum available real memory that we
1500 	 * can have in the system.  However, this is generally fixed at 25 pages
1501 	 * which is so low that it's useless.  In this comparison, we seek to
1502 	 * calculate the total heap-size, and reclaim if more than 3/4ths of the
1503 	 * heap is allocated.  (Or, in the caclulation, if less than 1/4th is
1504 	 * free)
1505 	 */
1506 	if (btop(vmem_size(heap_arena, VMEM_FREE)) <
1507 	    (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2))
1508 		return (1);
1509 #endif
1510 
1511 #else
1512 	if (spa_get_random(100) == 0)
1513 		return (1);
1514 #endif
1515 	return (0);
1516 }
1517 
1518 static void
1519 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
1520 {
1521 	size_t			i;
1522 	kmem_cache_t		*prev_cache = NULL;
1523 	kmem_cache_t		*prev_data_cache = NULL;
1524 	extern kmem_cache_t	*zio_buf_cache[];
1525 	extern kmem_cache_t	*zio_data_buf_cache[];
1526 
1527 #ifdef _KERNEL
1528 	if (arc_meta_used >= arc_meta_limit) {
1529 		/*
1530 		 * We are exceeding our meta-data cache limit.
1531 		 * Purge some DNLC entries to release holds on meta-data.
1532 		 */
1533 		dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
1534 	}
1535 #if defined(__i386)
1536 	/*
1537 	 * Reclaim unused memory from all kmem caches.
1538 	 */
1539 	kmem_reap();
1540 #endif
1541 #endif
1542 
1543 	/*
1544 	 * An agressive reclamation will shrink the cache size as well as
1545 	 * reap free buffers from the arc kmem caches.
1546 	 */
1547 	if (strat == ARC_RECLAIM_AGGR)
1548 		arc_shrink();
1549 
1550 	for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
1551 		if (zio_buf_cache[i] != prev_cache) {
1552 			prev_cache = zio_buf_cache[i];
1553 			kmem_cache_reap_now(zio_buf_cache[i]);
1554 		}
1555 		if (zio_data_buf_cache[i] != prev_data_cache) {
1556 			prev_data_cache = zio_data_buf_cache[i];
1557 			kmem_cache_reap_now(zio_data_buf_cache[i]);
1558 		}
1559 	}
1560 	kmem_cache_reap_now(buf_cache);
1561 	kmem_cache_reap_now(hdr_cache);
1562 }
1563 
1564 static void
1565 arc_reclaim_thread(void)
1566 {
1567 	clock_t			growtime = 0;
1568 	arc_reclaim_strategy_t	last_reclaim = ARC_RECLAIM_CONS;
1569 	callb_cpr_t		cpr;
1570 
1571 	CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
1572 
1573 	mutex_enter(&arc_reclaim_thr_lock);
1574 	while (arc_thread_exit == 0) {
1575 		if (arc_reclaim_needed()) {
1576 
1577 			if (arc_no_grow) {
1578 				if (last_reclaim == ARC_RECLAIM_CONS) {
1579 					last_reclaim = ARC_RECLAIM_AGGR;
1580 				} else {
1581 					last_reclaim = ARC_RECLAIM_CONS;
1582 				}
1583 			} else {
1584 				arc_no_grow = TRUE;
1585 				last_reclaim = ARC_RECLAIM_AGGR;
1586 				membar_producer();
1587 			}
1588 
1589 			/* reset the growth delay for every reclaim */
1590 			growtime = lbolt + (arc_grow_retry * hz);
1591 
1592 			arc_kmem_reap_now(last_reclaim);
1593 
1594 		} else if (arc_no_grow && lbolt >= growtime) {
1595 			arc_no_grow = FALSE;
1596 		}
1597 
1598 		if (2 * arc_c < arc_size +
1599 		    arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size)
1600 			arc_adjust();
1601 
1602 		if (arc_eviction_list != NULL)
1603 			arc_do_user_evicts();
1604 
1605 		/* block until needed, or one second, whichever is shorter */
1606 		CALLB_CPR_SAFE_BEGIN(&cpr);
1607 		(void) cv_timedwait(&arc_reclaim_thr_cv,
1608 		    &arc_reclaim_thr_lock, (lbolt + hz));
1609 		CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
1610 	}
1611 
1612 	arc_thread_exit = 0;
1613 	cv_broadcast(&arc_reclaim_thr_cv);
1614 	CALLB_CPR_EXIT(&cpr);		/* drops arc_reclaim_thr_lock */
1615 	thread_exit();
1616 }
1617 
1618 /*
1619  * Adapt arc info given the number of bytes we are trying to add and
1620  * the state that we are comming from.  This function is only called
1621  * when we are adding new content to the cache.
1622  */
1623 static void
1624 arc_adapt(int bytes, arc_state_t *state)
1625 {
1626 	int mult;
1627 
1628 	ASSERT(bytes > 0);
1629 	/*
1630 	 * Adapt the target size of the MRU list:
1631 	 *	- if we just hit in the MRU ghost list, then increase
1632 	 *	  the target size of the MRU list.
1633 	 *	- if we just hit in the MFU ghost list, then increase
1634 	 *	  the target size of the MFU list by decreasing the
1635 	 *	  target size of the MRU list.
1636 	 */
1637 	if (state == arc_mru_ghost) {
1638 		mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
1639 		    1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
1640 
1641 		arc_p = MIN(arc_c, arc_p + bytes * mult);
1642 	} else if (state == arc_mfu_ghost) {
1643 		mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
1644 		    1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
1645 
1646 		arc_p = MAX(0, (int64_t)arc_p - bytes * mult);
1647 	}
1648 	ASSERT((int64_t)arc_p >= 0);
1649 
1650 	if (arc_reclaim_needed()) {
1651 		cv_signal(&arc_reclaim_thr_cv);
1652 		return;
1653 	}
1654 
1655 	if (arc_no_grow)
1656 		return;
1657 
1658 	if (arc_c >= arc_c_max)
1659 		return;
1660 
1661 	/*
1662 	 * If we're within (2 * maxblocksize) bytes of the target
1663 	 * cache size, increment the target cache size
1664 	 */
1665 	if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
1666 		atomic_add_64(&arc_c, (int64_t)bytes);
1667 		if (arc_c > arc_c_max)
1668 			arc_c = arc_c_max;
1669 		else if (state == arc_anon)
1670 			atomic_add_64(&arc_p, (int64_t)bytes);
1671 		if (arc_p > arc_c)
1672 			arc_p = arc_c;
1673 	}
1674 	ASSERT((int64_t)arc_p >= 0);
1675 }
1676 
1677 /*
1678  * Check if the cache has reached its limits and eviction is required
1679  * prior to insert.
1680  */
1681 static int
1682 arc_evict_needed(arc_buf_contents_t type)
1683 {
1684 	if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
1685 		return (1);
1686 
1687 #ifdef _KERNEL
1688 	/*
1689 	 * If zio data pages are being allocated out of a separate heap segment,
1690 	 * then enforce that the size of available vmem for this area remains
1691 	 * above about 1/32nd free.
1692 	 */
1693 	if (type == ARC_BUFC_DATA && zio_arena != NULL &&
1694 	    vmem_size(zio_arena, VMEM_FREE) <
1695 	    (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
1696 		return (1);
1697 #endif
1698 
1699 	if (arc_reclaim_needed())
1700 		return (1);
1701 
1702 	return (arc_size > arc_c);
1703 }
1704 
1705 /*
1706  * The buffer, supplied as the first argument, needs a data block.
1707  * So, if we are at cache max, determine which cache should be victimized.
1708  * We have the following cases:
1709  *
1710  * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
1711  * In this situation if we're out of space, but the resident size of the MFU is
1712  * under the limit, victimize the MFU cache to satisfy this insertion request.
1713  *
1714  * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
1715  * Here, we've used up all of the available space for the MRU, so we need to
1716  * evict from our own cache instead.  Evict from the set of resident MRU
1717  * entries.
1718  *
1719  * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
1720  * c minus p represents the MFU space in the cache, since p is the size of the
1721  * cache that is dedicated to the MRU.  In this situation there's still space on
1722  * the MFU side, so the MRU side needs to be victimized.
1723  *
1724  * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
1725  * MFU's resident set is consuming more space than it has been allotted.  In
1726  * this situation, we must victimize our own cache, the MFU, for this insertion.
1727  */
1728 static void
1729 arc_get_data_buf(arc_buf_t *buf)
1730 {
1731 	arc_state_t		*state = buf->b_hdr->b_state;
1732 	uint64_t		size = buf->b_hdr->b_size;
1733 	arc_buf_contents_t	type = buf->b_hdr->b_type;
1734 
1735 	arc_adapt(size, state);
1736 
1737 	/*
1738 	 * We have not yet reached cache maximum size,
1739 	 * just allocate a new buffer.
1740 	 */
1741 	if (!arc_evict_needed(type)) {
1742 		if (type == ARC_BUFC_METADATA) {
1743 			buf->b_data = zio_buf_alloc(size);
1744 			arc_space_consume(size);
1745 		} else {
1746 			ASSERT(type == ARC_BUFC_DATA);
1747 			buf->b_data = zio_data_buf_alloc(size);
1748 			atomic_add_64(&arc_size, size);
1749 		}
1750 		goto out;
1751 	}
1752 
1753 	/*
1754 	 * If we are prefetching from the mfu ghost list, this buffer
1755 	 * will end up on the mru list; so steal space from there.
1756 	 */
1757 	if (state == arc_mfu_ghost)
1758 		state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
1759 	else if (state == arc_mru_ghost)
1760 		state = arc_mru;
1761 
1762 	if (state == arc_mru || state == arc_anon) {
1763 		uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
1764 		state = (arc_mfu->arcs_lsize[type] > 0 &&
1765 		    arc_p > mru_used) ? arc_mfu : arc_mru;
1766 	} else {
1767 		/* MFU cases */
1768 		uint64_t mfu_space = arc_c - arc_p;
1769 		state =  (arc_mru->arcs_lsize[type] > 0 &&
1770 		    mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
1771 	}
1772 	if ((buf->b_data = arc_evict(state, size, TRUE, type)) == NULL) {
1773 		if (type == ARC_BUFC_METADATA) {
1774 			buf->b_data = zio_buf_alloc(size);
1775 			arc_space_consume(size);
1776 		} else {
1777 			ASSERT(type == ARC_BUFC_DATA);
1778 			buf->b_data = zio_data_buf_alloc(size);
1779 			atomic_add_64(&arc_size, size);
1780 		}
1781 		ARCSTAT_BUMP(arcstat_recycle_miss);
1782 	}
1783 	ASSERT(buf->b_data != NULL);
1784 out:
1785 	/*
1786 	 * Update the state size.  Note that ghost states have a
1787 	 * "ghost size" and so don't need to be updated.
1788 	 */
1789 	if (!GHOST_STATE(buf->b_hdr->b_state)) {
1790 		arc_buf_hdr_t *hdr = buf->b_hdr;
1791 
1792 		atomic_add_64(&hdr->b_state->arcs_size, size);
1793 		if (list_link_active(&hdr->b_arc_node)) {
1794 			ASSERT(refcount_is_zero(&hdr->b_refcnt));
1795 			atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
1796 		}
1797 		/*
1798 		 * If we are growing the cache, and we are adding anonymous
1799 		 * data, and we have outgrown arc_p, update arc_p
1800 		 */
1801 		if (arc_size < arc_c && hdr->b_state == arc_anon &&
1802 		    arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
1803 			arc_p = MIN(arc_c, arc_p + size);
1804 	}
1805 }
1806 
1807 /*
1808  * This routine is called whenever a buffer is accessed.
1809  * NOTE: the hash lock is dropped in this function.
1810  */
1811 static void
1812 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
1813 {
1814 	ASSERT(MUTEX_HELD(hash_lock));
1815 
1816 	if (buf->b_state == arc_anon) {
1817 		/*
1818 		 * This buffer is not in the cache, and does not
1819 		 * appear in our "ghost" list.  Add the new buffer
1820 		 * to the MRU state.
1821 		 */
1822 
1823 		ASSERT(buf->b_arc_access == 0);
1824 		buf->b_arc_access = lbolt;
1825 		DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
1826 		arc_change_state(arc_mru, buf, hash_lock);
1827 
1828 	} else if (buf->b_state == arc_mru) {
1829 		/*
1830 		 * If this buffer is here because of a prefetch, then either:
1831 		 * - clear the flag if this is a "referencing" read
1832 		 *   (any subsequent access will bump this into the MFU state).
1833 		 * or
1834 		 * - move the buffer to the head of the list if this is
1835 		 *   another prefetch (to make it less likely to be evicted).
1836 		 */
1837 		if ((buf->b_flags & ARC_PREFETCH) != 0) {
1838 			if (refcount_count(&buf->b_refcnt) == 0) {
1839 				ASSERT(list_link_active(&buf->b_arc_node));
1840 			} else {
1841 				buf->b_flags &= ~ARC_PREFETCH;
1842 				ARCSTAT_BUMP(arcstat_mru_hits);
1843 			}
1844 			buf->b_arc_access = lbolt;
1845 			return;
1846 		}
1847 
1848 		/*
1849 		 * This buffer has been "accessed" only once so far,
1850 		 * but it is still in the cache. Move it to the MFU
1851 		 * state.
1852 		 */
1853 		if (lbolt > buf->b_arc_access + ARC_MINTIME) {
1854 			/*
1855 			 * More than 125ms have passed since we
1856 			 * instantiated this buffer.  Move it to the
1857 			 * most frequently used state.
1858 			 */
1859 			buf->b_arc_access = lbolt;
1860 			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
1861 			arc_change_state(arc_mfu, buf, hash_lock);
1862 		}
1863 		ARCSTAT_BUMP(arcstat_mru_hits);
1864 	} else if (buf->b_state == arc_mru_ghost) {
1865 		arc_state_t	*new_state;
1866 		/*
1867 		 * This buffer has been "accessed" recently, but
1868 		 * was evicted from the cache.  Move it to the
1869 		 * MFU state.
1870 		 */
1871 
1872 		if (buf->b_flags & ARC_PREFETCH) {
1873 			new_state = arc_mru;
1874 			if (refcount_count(&buf->b_refcnt) > 0)
1875 				buf->b_flags &= ~ARC_PREFETCH;
1876 			DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
1877 		} else {
1878 			new_state = arc_mfu;
1879 			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
1880 		}
1881 
1882 		buf->b_arc_access = lbolt;
1883 		arc_change_state(new_state, buf, hash_lock);
1884 
1885 		ARCSTAT_BUMP(arcstat_mru_ghost_hits);
1886 	} else if (buf->b_state == arc_mfu) {
1887 		/*
1888 		 * This buffer has been accessed more than once and is
1889 		 * still in the cache.  Keep it in the MFU state.
1890 		 *
1891 		 * NOTE: an add_reference() that occurred when we did
1892 		 * the arc_read() will have kicked this off the list.
1893 		 * If it was a prefetch, we will explicitly move it to
1894 		 * the head of the list now.
1895 		 */
1896 		if ((buf->b_flags & ARC_PREFETCH) != 0) {
1897 			ASSERT(refcount_count(&buf->b_refcnt) == 0);
1898 			ASSERT(list_link_active(&buf->b_arc_node));
1899 		}
1900 		ARCSTAT_BUMP(arcstat_mfu_hits);
1901 		buf->b_arc_access = lbolt;
1902 	} else if (buf->b_state == arc_mfu_ghost) {
1903 		arc_state_t	*new_state = arc_mfu;
1904 		/*
1905 		 * This buffer has been accessed more than once but has
1906 		 * been evicted from the cache.  Move it back to the
1907 		 * MFU state.
1908 		 */
1909 
1910 		if (buf->b_flags & ARC_PREFETCH) {
1911 			/*
1912 			 * This is a prefetch access...
1913 			 * move this block back to the MRU state.
1914 			 */
1915 			ASSERT3U(refcount_count(&buf->b_refcnt), ==, 0);
1916 			new_state = arc_mru;
1917 		}
1918 
1919 		buf->b_arc_access = lbolt;
1920 		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
1921 		arc_change_state(new_state, buf, hash_lock);
1922 
1923 		ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
1924 	} else {
1925 		ASSERT(!"invalid arc state");
1926 	}
1927 }
1928 
1929 /* a generic arc_done_func_t which you can use */
1930 /* ARGSUSED */
1931 void
1932 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
1933 {
1934 	bcopy(buf->b_data, arg, buf->b_hdr->b_size);
1935 	VERIFY(arc_buf_remove_ref(buf, arg) == 1);
1936 }
1937 
1938 /* a generic arc_done_func_t */
1939 void
1940 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
1941 {
1942 	arc_buf_t **bufp = arg;
1943 	if (zio && zio->io_error) {
1944 		VERIFY(arc_buf_remove_ref(buf, arg) == 1);
1945 		*bufp = NULL;
1946 	} else {
1947 		*bufp = buf;
1948 	}
1949 }
1950 
1951 static void
1952 arc_read_done(zio_t *zio)
1953 {
1954 	arc_buf_hdr_t	*hdr, *found;
1955 	arc_buf_t	*buf;
1956 	arc_buf_t	*abuf;	/* buffer we're assigning to callback */
1957 	kmutex_t	*hash_lock;
1958 	arc_callback_t	*callback_list, *acb;
1959 	int		freeable = FALSE;
1960 
1961 	buf = zio->io_private;
1962 	hdr = buf->b_hdr;
1963 
1964 	/*
1965 	 * The hdr was inserted into hash-table and removed from lists
1966 	 * prior to starting I/O.  We should find this header, since
1967 	 * it's in the hash table, and it should be legit since it's
1968 	 * not possible to evict it during the I/O.  The only possible
1969 	 * reason for it not to be found is if we were freed during the
1970 	 * read.
1971 	 */
1972 	found = buf_hash_find(zio->io_spa, &hdr->b_dva, hdr->b_birth,
1973 	    &hash_lock);
1974 
1975 	ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
1976 	    (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))));
1977 
1978 	/* byteswap if necessary */
1979 	callback_list = hdr->b_acb;
1980 	ASSERT(callback_list != NULL);
1981 	if (BP_SHOULD_BYTESWAP(zio->io_bp) && callback_list->acb_byteswap)
1982 		callback_list->acb_byteswap(buf->b_data, hdr->b_size);
1983 
1984 	arc_cksum_compute(buf);
1985 
1986 	/* create copies of the data buffer for the callers */
1987 	abuf = buf;
1988 	for (acb = callback_list; acb; acb = acb->acb_next) {
1989 		if (acb->acb_done) {
1990 			if (abuf == NULL)
1991 				abuf = arc_buf_clone(buf);
1992 			acb->acb_buf = abuf;
1993 			abuf = NULL;
1994 		}
1995 	}
1996 	hdr->b_acb = NULL;
1997 	hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
1998 	ASSERT(!HDR_BUF_AVAILABLE(hdr));
1999 	if (abuf == buf)
2000 		hdr->b_flags |= ARC_BUF_AVAILABLE;
2001 
2002 	ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2003 
2004 	if (zio->io_error != 0) {
2005 		hdr->b_flags |= ARC_IO_ERROR;
2006 		if (hdr->b_state != arc_anon)
2007 			arc_change_state(arc_anon, hdr, hash_lock);
2008 		if (HDR_IN_HASH_TABLE(hdr))
2009 			buf_hash_remove(hdr);
2010 		freeable = refcount_is_zero(&hdr->b_refcnt);
2011 		/* convert checksum errors into IO errors */
2012 		if (zio->io_error == ECKSUM)
2013 			zio->io_error = EIO;
2014 	}
2015 
2016 	/*
2017 	 * Broadcast before we drop the hash_lock to avoid the possibility
2018 	 * that the hdr (and hence the cv) might be freed before we get to
2019 	 * the cv_broadcast().
2020 	 */
2021 	cv_broadcast(&hdr->b_cv);
2022 
2023 	if (hash_lock) {
2024 		/*
2025 		 * Only call arc_access on anonymous buffers.  This is because
2026 		 * if we've issued an I/O for an evicted buffer, we've already
2027 		 * called arc_access (to prevent any simultaneous readers from
2028 		 * getting confused).
2029 		 */
2030 		if (zio->io_error == 0 && hdr->b_state == arc_anon)
2031 			arc_access(hdr, hash_lock);
2032 		mutex_exit(hash_lock);
2033 	} else {
2034 		/*
2035 		 * This block was freed while we waited for the read to
2036 		 * complete.  It has been removed from the hash table and
2037 		 * moved to the anonymous state (so that it won't show up
2038 		 * in the cache).
2039 		 */
2040 		ASSERT3P(hdr->b_state, ==, arc_anon);
2041 		freeable = refcount_is_zero(&hdr->b_refcnt);
2042 	}
2043 
2044 	/* execute each callback and free its structure */
2045 	while ((acb = callback_list) != NULL) {
2046 		if (acb->acb_done)
2047 			acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2048 
2049 		if (acb->acb_zio_dummy != NULL) {
2050 			acb->acb_zio_dummy->io_error = zio->io_error;
2051 			zio_nowait(acb->acb_zio_dummy);
2052 		}
2053 
2054 		callback_list = acb->acb_next;
2055 		kmem_free(acb, sizeof (arc_callback_t));
2056 	}
2057 
2058 	if (freeable)
2059 		arc_hdr_destroy(hdr);
2060 }
2061 
2062 /*
2063  * "Read" the block block at the specified DVA (in bp) via the
2064  * cache.  If the block is found in the cache, invoke the provided
2065  * callback immediately and return.  Note that the `zio' parameter
2066  * in the callback will be NULL in this case, since no IO was
2067  * required.  If the block is not in the cache pass the read request
2068  * on to the spa with a substitute callback function, so that the
2069  * requested block will be added to the cache.
2070  *
2071  * If a read request arrives for a block that has a read in-progress,
2072  * either wait for the in-progress read to complete (and return the
2073  * results); or, if this is a read with a "done" func, add a record
2074  * to the read to invoke the "done" func when the read completes,
2075  * and return; or just return.
2076  *
2077  * arc_read_done() will invoke all the requested "done" functions
2078  * for readers of this block.
2079  */
2080 int
2081 arc_read(zio_t *pio, spa_t *spa, blkptr_t *bp, arc_byteswap_func_t *swap,
2082     arc_done_func_t *done, void *private, int priority, int flags,
2083     uint32_t *arc_flags, zbookmark_t *zb)
2084 {
2085 	arc_buf_hdr_t *hdr;
2086 	arc_buf_t *buf;
2087 	kmutex_t *hash_lock;
2088 	zio_t	*rzio;
2089 
2090 top:
2091 	hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock);
2092 	if (hdr && hdr->b_datacnt > 0) {
2093 
2094 		*arc_flags |= ARC_CACHED;
2095 
2096 		if (HDR_IO_IN_PROGRESS(hdr)) {
2097 
2098 			if (*arc_flags & ARC_WAIT) {
2099 				cv_wait(&hdr->b_cv, hash_lock);
2100 				mutex_exit(hash_lock);
2101 				goto top;
2102 			}
2103 			ASSERT(*arc_flags & ARC_NOWAIT);
2104 
2105 			if (done) {
2106 				arc_callback_t	*acb = NULL;
2107 
2108 				acb = kmem_zalloc(sizeof (arc_callback_t),
2109 				    KM_SLEEP);
2110 				acb->acb_done = done;
2111 				acb->acb_private = private;
2112 				acb->acb_byteswap = swap;
2113 				if (pio != NULL)
2114 					acb->acb_zio_dummy = zio_null(pio,
2115 					    spa, NULL, NULL, flags);
2116 
2117 				ASSERT(acb->acb_done != NULL);
2118 				acb->acb_next = hdr->b_acb;
2119 				hdr->b_acb = acb;
2120 				add_reference(hdr, hash_lock, private);
2121 				mutex_exit(hash_lock);
2122 				return (0);
2123 			}
2124 			mutex_exit(hash_lock);
2125 			return (0);
2126 		}
2127 
2128 		ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2129 
2130 		if (done) {
2131 			add_reference(hdr, hash_lock, private);
2132 			/*
2133 			 * If this block is already in use, create a new
2134 			 * copy of the data so that we will be guaranteed
2135 			 * that arc_release() will always succeed.
2136 			 */
2137 			buf = hdr->b_buf;
2138 			ASSERT(buf);
2139 			ASSERT(buf->b_data);
2140 			if (HDR_BUF_AVAILABLE(hdr)) {
2141 				ASSERT(buf->b_efunc == NULL);
2142 				hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2143 			} else {
2144 				buf = arc_buf_clone(buf);
2145 			}
2146 		} else if (*arc_flags & ARC_PREFETCH &&
2147 		    refcount_count(&hdr->b_refcnt) == 0) {
2148 			hdr->b_flags |= ARC_PREFETCH;
2149 		}
2150 		DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2151 		arc_access(hdr, hash_lock);
2152 		mutex_exit(hash_lock);
2153 		ARCSTAT_BUMP(arcstat_hits);
2154 		ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2155 		    demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2156 		    data, metadata, hits);
2157 
2158 		if (done)
2159 			done(NULL, buf, private);
2160 	} else {
2161 		uint64_t size = BP_GET_LSIZE(bp);
2162 		arc_callback_t	*acb;
2163 
2164 		if (hdr == NULL) {
2165 			/* this block is not in the cache */
2166 			arc_buf_hdr_t	*exists;
2167 			arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2168 			buf = arc_buf_alloc(spa, size, private, type);
2169 			hdr = buf->b_hdr;
2170 			hdr->b_dva = *BP_IDENTITY(bp);
2171 			hdr->b_birth = bp->blk_birth;
2172 			hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
2173 			exists = buf_hash_insert(hdr, &hash_lock);
2174 			if (exists) {
2175 				/* somebody beat us to the hash insert */
2176 				mutex_exit(hash_lock);
2177 				bzero(&hdr->b_dva, sizeof (dva_t));
2178 				hdr->b_birth = 0;
2179 				hdr->b_cksum0 = 0;
2180 				(void) arc_buf_remove_ref(buf, private);
2181 				goto top; /* restart the IO request */
2182 			}
2183 			/* if this is a prefetch, we don't have a reference */
2184 			if (*arc_flags & ARC_PREFETCH) {
2185 				(void) remove_reference(hdr, hash_lock,
2186 				    private);
2187 				hdr->b_flags |= ARC_PREFETCH;
2188 			}
2189 			if (BP_GET_LEVEL(bp) > 0)
2190 				hdr->b_flags |= ARC_INDIRECT;
2191 		} else {
2192 			/* this block is in the ghost cache */
2193 			ASSERT(GHOST_STATE(hdr->b_state));
2194 			ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2195 			ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 0);
2196 			ASSERT(hdr->b_buf == NULL);
2197 
2198 			/* if this is a prefetch, we don't have a reference */
2199 			if (*arc_flags & ARC_PREFETCH)
2200 				hdr->b_flags |= ARC_PREFETCH;
2201 			else
2202 				add_reference(hdr, hash_lock, private);
2203 			buf = kmem_cache_alloc(buf_cache, KM_SLEEP);
2204 			buf->b_hdr = hdr;
2205 			buf->b_data = NULL;
2206 			buf->b_efunc = NULL;
2207 			buf->b_private = NULL;
2208 			buf->b_next = NULL;
2209 			hdr->b_buf = buf;
2210 			arc_get_data_buf(buf);
2211 			ASSERT(hdr->b_datacnt == 0);
2212 			hdr->b_datacnt = 1;
2213 
2214 		}
2215 
2216 		acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
2217 		acb->acb_done = done;
2218 		acb->acb_private = private;
2219 		acb->acb_byteswap = swap;
2220 
2221 		ASSERT(hdr->b_acb == NULL);
2222 		hdr->b_acb = acb;
2223 		hdr->b_flags |= ARC_IO_IN_PROGRESS;
2224 
2225 		/*
2226 		 * If the buffer has been evicted, migrate it to a present state
2227 		 * before issuing the I/O.  Once we drop the hash-table lock,
2228 		 * the header will be marked as I/O in progress and have an
2229 		 * attached buffer.  At this point, anybody who finds this
2230 		 * buffer ought to notice that it's legit but has a pending I/O.
2231 		 */
2232 
2233 		if (GHOST_STATE(hdr->b_state))
2234 			arc_access(hdr, hash_lock);
2235 		mutex_exit(hash_lock);
2236 
2237 		ASSERT3U(hdr->b_size, ==, size);
2238 		DTRACE_PROBE3(arc__miss, blkptr_t *, bp, uint64_t, size,
2239 		    zbookmark_t *, zb);
2240 		ARCSTAT_BUMP(arcstat_misses);
2241 		ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2242 		    demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2243 		    data, metadata, misses);
2244 
2245 		rzio = zio_read(pio, spa, bp, buf->b_data, size,
2246 		    arc_read_done, buf, priority, flags, zb);
2247 
2248 		if (*arc_flags & ARC_WAIT)
2249 			return (zio_wait(rzio));
2250 
2251 		ASSERT(*arc_flags & ARC_NOWAIT);
2252 		zio_nowait(rzio);
2253 	}
2254 	return (0);
2255 }
2256 
2257 /*
2258  * arc_read() variant to support pool traversal.  If the block is already
2259  * in the ARC, make a copy of it; otherwise, the caller will do the I/O.
2260  * The idea is that we don't want pool traversal filling up memory, but
2261  * if the ARC already has the data anyway, we shouldn't pay for the I/O.
2262  */
2263 int
2264 arc_tryread(spa_t *spa, blkptr_t *bp, void *data)
2265 {
2266 	arc_buf_hdr_t *hdr;
2267 	kmutex_t *hash_mtx;
2268 	int rc = 0;
2269 
2270 	hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_mtx);
2271 
2272 	if (hdr && hdr->b_datacnt > 0 && !HDR_IO_IN_PROGRESS(hdr)) {
2273 		arc_buf_t *buf = hdr->b_buf;
2274 
2275 		ASSERT(buf);
2276 		while (buf->b_data == NULL) {
2277 			buf = buf->b_next;
2278 			ASSERT(buf);
2279 		}
2280 		bcopy(buf->b_data, data, hdr->b_size);
2281 	} else {
2282 		rc = ENOENT;
2283 	}
2284 
2285 	if (hash_mtx)
2286 		mutex_exit(hash_mtx);
2287 
2288 	return (rc);
2289 }
2290 
2291 void
2292 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
2293 {
2294 	ASSERT(buf->b_hdr != NULL);
2295 	ASSERT(buf->b_hdr->b_state != arc_anon);
2296 	ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
2297 	buf->b_efunc = func;
2298 	buf->b_private = private;
2299 }
2300 
2301 /*
2302  * This is used by the DMU to let the ARC know that a buffer is
2303  * being evicted, so the ARC should clean up.  If this arc buf
2304  * is not yet in the evicted state, it will be put there.
2305  */
2306 int
2307 arc_buf_evict(arc_buf_t *buf)
2308 {
2309 	arc_buf_hdr_t *hdr;
2310 	kmutex_t *hash_lock;
2311 	arc_buf_t **bufp;
2312 
2313 	mutex_enter(&arc_eviction_mtx);
2314 	hdr = buf->b_hdr;
2315 	if (hdr == NULL) {
2316 		/*
2317 		 * We are in arc_do_user_evicts().
2318 		 */
2319 		ASSERT(buf->b_data == NULL);
2320 		mutex_exit(&arc_eviction_mtx);
2321 		return (0);
2322 	}
2323 	hash_lock = HDR_LOCK(hdr);
2324 	mutex_exit(&arc_eviction_mtx);
2325 
2326 	mutex_enter(hash_lock);
2327 
2328 	if (buf->b_data == NULL) {
2329 		/*
2330 		 * We are on the eviction list.
2331 		 */
2332 		mutex_exit(hash_lock);
2333 		mutex_enter(&arc_eviction_mtx);
2334 		if (buf->b_hdr == NULL) {
2335 			/*
2336 			 * We are already in arc_do_user_evicts().
2337 			 */
2338 			mutex_exit(&arc_eviction_mtx);
2339 			return (0);
2340 		} else {
2341 			arc_buf_t copy = *buf; /* structure assignment */
2342 			/*
2343 			 * Process this buffer now
2344 			 * but let arc_do_user_evicts() do the reaping.
2345 			 */
2346 			buf->b_efunc = NULL;
2347 			mutex_exit(&arc_eviction_mtx);
2348 			VERIFY(copy.b_efunc(&copy) == 0);
2349 			return (1);
2350 		}
2351 	}
2352 
2353 	ASSERT(buf->b_hdr == hdr);
2354 	ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
2355 	ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2356 
2357 	/*
2358 	 * Pull this buffer off of the hdr
2359 	 */
2360 	bufp = &hdr->b_buf;
2361 	while (*bufp != buf)
2362 		bufp = &(*bufp)->b_next;
2363 	*bufp = buf->b_next;
2364 
2365 	ASSERT(buf->b_data != NULL);
2366 	arc_buf_destroy(buf, FALSE, FALSE);
2367 
2368 	if (hdr->b_datacnt == 0) {
2369 		arc_state_t *old_state = hdr->b_state;
2370 		arc_state_t *evicted_state;
2371 
2372 		ASSERT(refcount_is_zero(&hdr->b_refcnt));
2373 
2374 		evicted_state =
2375 		    (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
2376 
2377 		mutex_enter(&old_state->arcs_mtx);
2378 		mutex_enter(&evicted_state->arcs_mtx);
2379 
2380 		arc_change_state(evicted_state, hdr, hash_lock);
2381 		ASSERT(HDR_IN_HASH_TABLE(hdr));
2382 		hdr->b_flags = ARC_IN_HASH_TABLE;
2383 
2384 		mutex_exit(&evicted_state->arcs_mtx);
2385 		mutex_exit(&old_state->arcs_mtx);
2386 	}
2387 	mutex_exit(hash_lock);
2388 
2389 	VERIFY(buf->b_efunc(buf) == 0);
2390 	buf->b_efunc = NULL;
2391 	buf->b_private = NULL;
2392 	buf->b_hdr = NULL;
2393 	kmem_cache_free(buf_cache, buf);
2394 	return (1);
2395 }
2396 
2397 /*
2398  * Release this buffer from the cache.  This must be done
2399  * after a read and prior to modifying the buffer contents.
2400  * If the buffer has more than one reference, we must make
2401  * make a new hdr for the buffer.
2402  */
2403 void
2404 arc_release(arc_buf_t *buf, void *tag)
2405 {
2406 	arc_buf_hdr_t *hdr = buf->b_hdr;
2407 	kmutex_t *hash_lock = HDR_LOCK(hdr);
2408 
2409 	/* this buffer is not on any list */
2410 	ASSERT(refcount_count(&hdr->b_refcnt) > 0);
2411 
2412 	if (hdr->b_state == arc_anon) {
2413 		/* this buffer is already released */
2414 		ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 1);
2415 		ASSERT(BUF_EMPTY(hdr));
2416 		ASSERT(buf->b_efunc == NULL);
2417 		arc_buf_thaw(buf);
2418 		return;
2419 	}
2420 
2421 	mutex_enter(hash_lock);
2422 
2423 	/*
2424 	 * Do we have more than one buf?
2425 	 */
2426 	if (hdr->b_buf != buf || buf->b_next != NULL) {
2427 		arc_buf_hdr_t *nhdr;
2428 		arc_buf_t **bufp;
2429 		uint64_t blksz = hdr->b_size;
2430 		spa_t *spa = hdr->b_spa;
2431 		arc_buf_contents_t type = hdr->b_type;
2432 
2433 		ASSERT(hdr->b_datacnt > 1);
2434 		/*
2435 		 * Pull the data off of this buf and attach it to
2436 		 * a new anonymous buf.
2437 		 */
2438 		(void) remove_reference(hdr, hash_lock, tag);
2439 		bufp = &hdr->b_buf;
2440 		while (*bufp != buf)
2441 			bufp = &(*bufp)->b_next;
2442 		*bufp = (*bufp)->b_next;
2443 		buf->b_next = NULL;
2444 
2445 		ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
2446 		atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
2447 		if (refcount_is_zero(&hdr->b_refcnt)) {
2448 			uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
2449 			ASSERT3U(*size, >=, hdr->b_size);
2450 			atomic_add_64(size, -hdr->b_size);
2451 		}
2452 		hdr->b_datacnt -= 1;
2453 		arc_cksum_verify(buf);
2454 
2455 		mutex_exit(hash_lock);
2456 
2457 		nhdr = kmem_cache_alloc(hdr_cache, KM_SLEEP);
2458 		nhdr->b_size = blksz;
2459 		nhdr->b_spa = spa;
2460 		nhdr->b_type = type;
2461 		nhdr->b_buf = buf;
2462 		nhdr->b_state = arc_anon;
2463 		nhdr->b_arc_access = 0;
2464 		nhdr->b_flags = 0;
2465 		nhdr->b_datacnt = 1;
2466 		nhdr->b_freeze_cksum = NULL;
2467 		(void) refcount_add(&nhdr->b_refcnt, tag);
2468 		buf->b_hdr = nhdr;
2469 		atomic_add_64(&arc_anon->arcs_size, blksz);
2470 
2471 		hdr = nhdr;
2472 	} else {
2473 		ASSERT(refcount_count(&hdr->b_refcnt) == 1);
2474 		ASSERT(!list_link_active(&hdr->b_arc_node));
2475 		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2476 		arc_change_state(arc_anon, hdr, hash_lock);
2477 		hdr->b_arc_access = 0;
2478 		mutex_exit(hash_lock);
2479 		bzero(&hdr->b_dva, sizeof (dva_t));
2480 		hdr->b_birth = 0;
2481 		hdr->b_cksum0 = 0;
2482 		arc_buf_thaw(buf);
2483 	}
2484 	buf->b_efunc = NULL;
2485 	buf->b_private = NULL;
2486 }
2487 
2488 int
2489 arc_released(arc_buf_t *buf)
2490 {
2491 	return (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
2492 }
2493 
2494 int
2495 arc_has_callback(arc_buf_t *buf)
2496 {
2497 	return (buf->b_efunc != NULL);
2498 }
2499 
2500 #ifdef ZFS_DEBUG
2501 int
2502 arc_referenced(arc_buf_t *buf)
2503 {
2504 	return (refcount_count(&buf->b_hdr->b_refcnt));
2505 }
2506 #endif
2507 
2508 static void
2509 arc_write_ready(zio_t *zio)
2510 {
2511 	arc_write_callback_t *callback = zio->io_private;
2512 	arc_buf_t *buf = callback->awcb_buf;
2513 
2514 	if (callback->awcb_ready) {
2515 		ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
2516 		callback->awcb_ready(zio, buf, callback->awcb_private);
2517 	}
2518 	arc_cksum_compute(buf);
2519 }
2520 
2521 static void
2522 arc_write_done(zio_t *zio)
2523 {
2524 	arc_write_callback_t *callback = zio->io_private;
2525 	arc_buf_t *buf = callback->awcb_buf;
2526 	arc_buf_hdr_t *hdr = buf->b_hdr;
2527 
2528 	hdr->b_acb = NULL;
2529 
2530 	/* this buffer is on no lists and is not in the hash table */
2531 	ASSERT3P(hdr->b_state, ==, arc_anon);
2532 
2533 	hdr->b_dva = *BP_IDENTITY(zio->io_bp);
2534 	hdr->b_birth = zio->io_bp->blk_birth;
2535 	hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
2536 	/*
2537 	 * If the block to be written was all-zero, we may have
2538 	 * compressed it away.  In this case no write was performed
2539 	 * so there will be no dva/birth-date/checksum.  The buffer
2540 	 * must therefor remain anonymous (and uncached).
2541 	 */
2542 	if (!BUF_EMPTY(hdr)) {
2543 		arc_buf_hdr_t *exists;
2544 		kmutex_t *hash_lock;
2545 
2546 		arc_cksum_verify(buf);
2547 
2548 		exists = buf_hash_insert(hdr, &hash_lock);
2549 		if (exists) {
2550 			/*
2551 			 * This can only happen if we overwrite for
2552 			 * sync-to-convergence, because we remove
2553 			 * buffers from the hash table when we arc_free().
2554 			 */
2555 			ASSERT(DVA_EQUAL(BP_IDENTITY(&zio->io_bp_orig),
2556 			    BP_IDENTITY(zio->io_bp)));
2557 			ASSERT3U(zio->io_bp_orig.blk_birth, ==,
2558 			    zio->io_bp->blk_birth);
2559 
2560 			ASSERT(refcount_is_zero(&exists->b_refcnt));
2561 			arc_change_state(arc_anon, exists, hash_lock);
2562 			mutex_exit(hash_lock);
2563 			arc_hdr_destroy(exists);
2564 			exists = buf_hash_insert(hdr, &hash_lock);
2565 			ASSERT3P(exists, ==, NULL);
2566 		}
2567 		hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2568 		arc_access(hdr, hash_lock);
2569 		mutex_exit(hash_lock);
2570 	} else if (callback->awcb_done == NULL) {
2571 		int destroy_hdr;
2572 		/*
2573 		 * This is an anonymous buffer with no user callback,
2574 		 * destroy it if there are no active references.
2575 		 */
2576 		mutex_enter(&arc_eviction_mtx);
2577 		destroy_hdr = refcount_is_zero(&hdr->b_refcnt);
2578 		hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2579 		mutex_exit(&arc_eviction_mtx);
2580 		if (destroy_hdr)
2581 			arc_hdr_destroy(hdr);
2582 	} else {
2583 		hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2584 	}
2585 
2586 	if (callback->awcb_done) {
2587 		ASSERT(!refcount_is_zero(&hdr->b_refcnt));
2588 		callback->awcb_done(zio, buf, callback->awcb_private);
2589 	}
2590 
2591 	kmem_free(callback, sizeof (arc_write_callback_t));
2592 }
2593 
2594 zio_t *
2595 arc_write(zio_t *pio, spa_t *spa, int checksum, int compress, int ncopies,
2596     uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
2597     arc_done_func_t *ready, arc_done_func_t *done, void *private, int priority,
2598     int flags, zbookmark_t *zb)
2599 {
2600 	arc_buf_hdr_t *hdr = buf->b_hdr;
2601 	arc_write_callback_t *callback;
2602 	zio_t	*zio;
2603 
2604 	/* this is a private buffer - no locking required */
2605 	ASSERT3P(hdr->b_state, ==, arc_anon);
2606 	ASSERT(BUF_EMPTY(hdr));
2607 	ASSERT(!HDR_IO_ERROR(hdr));
2608 	ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
2609 	ASSERT(hdr->b_acb == 0);
2610 	callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
2611 	callback->awcb_ready = ready;
2612 	callback->awcb_done = done;
2613 	callback->awcb_private = private;
2614 	callback->awcb_buf = buf;
2615 	hdr->b_flags |= ARC_IO_IN_PROGRESS;
2616 	zio = zio_write(pio, spa, checksum, compress, ncopies, txg, bp,
2617 	    buf->b_data, hdr->b_size, arc_write_ready, arc_write_done, callback,
2618 	    priority, flags, zb);
2619 
2620 	return (zio);
2621 }
2622 
2623 int
2624 arc_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
2625     zio_done_func_t *done, void *private, uint32_t arc_flags)
2626 {
2627 	arc_buf_hdr_t *ab;
2628 	kmutex_t *hash_lock;
2629 	zio_t	*zio;
2630 
2631 	/*
2632 	 * If this buffer is in the cache, release it, so it
2633 	 * can be re-used.
2634 	 */
2635 	ab = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock);
2636 	if (ab != NULL) {
2637 		/*
2638 		 * The checksum of blocks to free is not always
2639 		 * preserved (eg. on the deadlist).  However, if it is
2640 		 * nonzero, it should match what we have in the cache.
2641 		 */
2642 		ASSERT(bp->blk_cksum.zc_word[0] == 0 ||
2643 		    ab->b_cksum0 == bp->blk_cksum.zc_word[0]);
2644 		if (ab->b_state != arc_anon)
2645 			arc_change_state(arc_anon, ab, hash_lock);
2646 		if (HDR_IO_IN_PROGRESS(ab)) {
2647 			/*
2648 			 * This should only happen when we prefetch.
2649 			 */
2650 			ASSERT(ab->b_flags & ARC_PREFETCH);
2651 			ASSERT3U(ab->b_datacnt, ==, 1);
2652 			ab->b_flags |= ARC_FREED_IN_READ;
2653 			if (HDR_IN_HASH_TABLE(ab))
2654 				buf_hash_remove(ab);
2655 			ab->b_arc_access = 0;
2656 			bzero(&ab->b_dva, sizeof (dva_t));
2657 			ab->b_birth = 0;
2658 			ab->b_cksum0 = 0;
2659 			ab->b_buf->b_efunc = NULL;
2660 			ab->b_buf->b_private = NULL;
2661 			mutex_exit(hash_lock);
2662 		} else if (refcount_is_zero(&ab->b_refcnt)) {
2663 			mutex_exit(hash_lock);
2664 			arc_hdr_destroy(ab);
2665 			ARCSTAT_BUMP(arcstat_deleted);
2666 		} else {
2667 			/*
2668 			 * We still have an active reference on this
2669 			 * buffer.  This can happen, e.g., from
2670 			 * dbuf_unoverride().
2671 			 */
2672 			ASSERT(!HDR_IN_HASH_TABLE(ab));
2673 			ab->b_arc_access = 0;
2674 			bzero(&ab->b_dva, sizeof (dva_t));
2675 			ab->b_birth = 0;
2676 			ab->b_cksum0 = 0;
2677 			ab->b_buf->b_efunc = NULL;
2678 			ab->b_buf->b_private = NULL;
2679 			mutex_exit(hash_lock);
2680 		}
2681 	}
2682 
2683 	zio = zio_free(pio, spa, txg, bp, done, private);
2684 
2685 	if (arc_flags & ARC_WAIT)
2686 		return (zio_wait(zio));
2687 
2688 	ASSERT(arc_flags & ARC_NOWAIT);
2689 	zio_nowait(zio);
2690 
2691 	return (0);
2692 }
2693 
2694 void
2695 arc_tempreserve_clear(uint64_t tempreserve)
2696 {
2697 	atomic_add_64(&arc_tempreserve, -tempreserve);
2698 	ASSERT((int64_t)arc_tempreserve >= 0);
2699 }
2700 
2701 int
2702 arc_tempreserve_space(uint64_t tempreserve)
2703 {
2704 #ifdef ZFS_DEBUG
2705 	/*
2706 	 * Once in a while, fail for no reason.  Everything should cope.
2707 	 */
2708 	if (spa_get_random(10000) == 0) {
2709 		dprintf("forcing random failure\n");
2710 		return (ERESTART);
2711 	}
2712 #endif
2713 	if (tempreserve > arc_c/4 && !arc_no_grow)
2714 		arc_c = MIN(arc_c_max, tempreserve * 4);
2715 	if (tempreserve > arc_c)
2716 		return (ENOMEM);
2717 
2718 	/*
2719 	 * Throttle writes when the amount of dirty data in the cache
2720 	 * gets too large.  We try to keep the cache less than half full
2721 	 * of dirty blocks so that our sync times don't grow too large.
2722 	 * Note: if two requests come in concurrently, we might let them
2723 	 * both succeed, when one of them should fail.  Not a huge deal.
2724 	 *
2725 	 * XXX The limit should be adjusted dynamically to keep the time
2726 	 * to sync a dataset fixed (around 1-5 seconds?).
2727 	 */
2728 
2729 	if (tempreserve + arc_tempreserve + arc_anon->arcs_size > arc_c / 2 &&
2730 	    arc_tempreserve + arc_anon->arcs_size > arc_c / 4) {
2731 		dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
2732 		    "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
2733 		    arc_tempreserve>>10,
2734 		    arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
2735 		    arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
2736 		    tempreserve>>10, arc_c>>10);
2737 		return (ERESTART);
2738 	}
2739 	atomic_add_64(&arc_tempreserve, tempreserve);
2740 	return (0);
2741 }
2742 
2743 void
2744 arc_init(void)
2745 {
2746 	mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
2747 	cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
2748 
2749 	/* Convert seconds to clock ticks */
2750 	arc_min_prefetch_lifespan = 1 * hz;
2751 
2752 	/* Start out with 1/8 of all memory */
2753 	arc_c = physmem * PAGESIZE / 8;
2754 
2755 #ifdef _KERNEL
2756 	/*
2757 	 * On architectures where the physical memory can be larger
2758 	 * than the addressable space (intel in 32-bit mode), we may
2759 	 * need to limit the cache to 1/8 of VM size.
2760 	 */
2761 	arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
2762 #endif
2763 
2764 	/* set min cache to 1/32 of all memory, or 64MB, whichever is more */
2765 	arc_c_min = MAX(arc_c / 4, 64<<20);
2766 	/* set max to 3/4 of all memory, or all but 1GB, whichever is more */
2767 	if (arc_c * 8 >= 1<<30)
2768 		arc_c_max = (arc_c * 8) - (1<<30);
2769 	else
2770 		arc_c_max = arc_c_min;
2771 	arc_c_max = MAX(arc_c * 6, arc_c_max);
2772 
2773 	/*
2774 	 * Allow the tunables to override our calculations if they are
2775 	 * reasonable (ie. over 64MB)
2776 	 */
2777 	if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
2778 		arc_c_max = zfs_arc_max;
2779 	if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
2780 		arc_c_min = zfs_arc_min;
2781 
2782 	arc_c = arc_c_max;
2783 	arc_p = (arc_c >> 1);
2784 
2785 	/* limit meta-data to 1/4 of the arc capacity */
2786 	arc_meta_limit = arc_c_max / 4;
2787 
2788 	/* Allow the tunable to override if it is reasonable */
2789 	if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
2790 		arc_meta_limit = zfs_arc_meta_limit;
2791 
2792 	if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
2793 		arc_c_min = arc_meta_limit / 2;
2794 
2795 	/* if kmem_flags are set, lets try to use less memory */
2796 	if (kmem_debugging())
2797 		arc_c = arc_c / 2;
2798 	if (arc_c < arc_c_min)
2799 		arc_c = arc_c_min;
2800 
2801 	arc_anon = &ARC_anon;
2802 	arc_mru = &ARC_mru;
2803 	arc_mru_ghost = &ARC_mru_ghost;
2804 	arc_mfu = &ARC_mfu;
2805 	arc_mfu_ghost = &ARC_mfu_ghost;
2806 	arc_size = 0;
2807 
2808 	mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
2809 	mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
2810 	mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
2811 	mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
2812 	mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
2813 
2814 	list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
2815 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2816 	list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
2817 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2818 	list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
2819 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2820 	list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
2821 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2822 	list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
2823 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2824 	list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
2825 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2826 	list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
2827 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2828 	list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
2829 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2830 
2831 	buf_init();
2832 
2833 	arc_thread_exit = 0;
2834 	arc_eviction_list = NULL;
2835 	mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
2836 	bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
2837 
2838 	arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
2839 	    sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
2840 
2841 	if (arc_ksp != NULL) {
2842 		arc_ksp->ks_data = &arc_stats;
2843 		kstat_install(arc_ksp);
2844 	}
2845 
2846 	(void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
2847 	    TS_RUN, minclsyspri);
2848 
2849 	arc_dead = FALSE;
2850 }
2851 
2852 void
2853 arc_fini(void)
2854 {
2855 	mutex_enter(&arc_reclaim_thr_lock);
2856 	arc_thread_exit = 1;
2857 	while (arc_thread_exit != 0)
2858 		cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
2859 	mutex_exit(&arc_reclaim_thr_lock);
2860 
2861 	arc_flush();
2862 
2863 	arc_dead = TRUE;
2864 
2865 	if (arc_ksp != NULL) {
2866 		kstat_delete(arc_ksp);
2867 		arc_ksp = NULL;
2868 	}
2869 
2870 	mutex_destroy(&arc_eviction_mtx);
2871 	mutex_destroy(&arc_reclaim_thr_lock);
2872 	cv_destroy(&arc_reclaim_thr_cv);
2873 
2874 	list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
2875 	list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
2876 	list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
2877 	list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
2878 	list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
2879 	list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
2880 	list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
2881 	list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
2882 
2883 	mutex_destroy(&arc_anon->arcs_mtx);
2884 	mutex_destroy(&arc_mru->arcs_mtx);
2885 	mutex_destroy(&arc_mru_ghost->arcs_mtx);
2886 	mutex_destroy(&arc_mfu->arcs_mtx);
2887 	mutex_destroy(&arc_mfu_ghost->arcs_mtx);
2888 
2889 	buf_fini();
2890 }
2891