xref: /titanic_52/usr/src/uts/common/fs/zfs/arc.c (revision 038dc6b3d39a14038605fbb3fc6a9bb76dd27fa8)
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 static void arc_evict_ghost(arc_state_t *state, int64_t bytes);
383 
384 #define	GHOST_STATE(state)	\
385 	((state) == arc_mru_ghost || (state) == arc_mfu_ghost)
386 
387 /*
388  * Private ARC flags.  These flags are private ARC only flags that will show up
389  * in b_flags in the arc_hdr_buf_t.  Some flags are publicly declared, and can
390  * be passed in as arc_flags in things like arc_read.  However, these flags
391  * should never be passed and should only be set by ARC code.  When adding new
392  * public flags, make sure not to smash the private ones.
393  */
394 
395 #define	ARC_IN_HASH_TABLE	(1 << 9)	/* this buffer is hashed */
396 #define	ARC_IO_IN_PROGRESS	(1 << 10)	/* I/O in progress for buf */
397 #define	ARC_IO_ERROR		(1 << 11)	/* I/O failed for buf */
398 #define	ARC_FREED_IN_READ	(1 << 12)	/* buf freed while in read */
399 #define	ARC_BUF_AVAILABLE	(1 << 13)	/* block not in active use */
400 #define	ARC_INDIRECT		(1 << 14)	/* this is an indirect block */
401 
402 #define	HDR_IN_HASH_TABLE(hdr)	((hdr)->b_flags & ARC_IN_HASH_TABLE)
403 #define	HDR_IO_IN_PROGRESS(hdr)	((hdr)->b_flags & ARC_IO_IN_PROGRESS)
404 #define	HDR_IO_ERROR(hdr)	((hdr)->b_flags & ARC_IO_ERROR)
405 #define	HDR_FREED_IN_READ(hdr)	((hdr)->b_flags & ARC_FREED_IN_READ)
406 #define	HDR_BUF_AVAILABLE(hdr)	((hdr)->b_flags & ARC_BUF_AVAILABLE)
407 
408 /*
409  * Hash table routines
410  */
411 
412 #define	HT_LOCK_PAD	64
413 
414 struct ht_lock {
415 	kmutex_t	ht_lock;
416 #ifdef _KERNEL
417 	unsigned char	pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
418 #endif
419 };
420 
421 #define	BUF_LOCKS 256
422 typedef struct buf_hash_table {
423 	uint64_t ht_mask;
424 	arc_buf_hdr_t **ht_table;
425 	struct ht_lock ht_locks[BUF_LOCKS];
426 } buf_hash_table_t;
427 
428 static buf_hash_table_t buf_hash_table;
429 
430 #define	BUF_HASH_INDEX(spa, dva, birth) \
431 	(buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
432 #define	BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
433 #define	BUF_HASH_LOCK(idx)	(&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
434 #define	HDR_LOCK(buf) \
435 	(BUF_HASH_LOCK(BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth)))
436 
437 uint64_t zfs_crc64_table[256];
438 
439 static uint64_t
440 buf_hash(spa_t *spa, dva_t *dva, uint64_t birth)
441 {
442 	uintptr_t spav = (uintptr_t)spa;
443 	uint8_t *vdva = (uint8_t *)dva;
444 	uint64_t crc = -1ULL;
445 	int i;
446 
447 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
448 
449 	for (i = 0; i < sizeof (dva_t); i++)
450 		crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
451 
452 	crc ^= (spav>>8) ^ birth;
453 
454 	return (crc);
455 }
456 
457 #define	BUF_EMPTY(buf)						\
458 	((buf)->b_dva.dva_word[0] == 0 &&			\
459 	(buf)->b_dva.dva_word[1] == 0 &&			\
460 	(buf)->b_birth == 0)
461 
462 #define	BUF_EQUAL(spa, dva, birth, buf)				\
463 	((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) &&	\
464 	((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) &&	\
465 	((buf)->b_birth == birth) && ((buf)->b_spa == spa)
466 
467 static arc_buf_hdr_t *
468 buf_hash_find(spa_t *spa, dva_t *dva, uint64_t birth, kmutex_t **lockp)
469 {
470 	uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
471 	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
472 	arc_buf_hdr_t *buf;
473 
474 	mutex_enter(hash_lock);
475 	for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
476 	    buf = buf->b_hash_next) {
477 		if (BUF_EQUAL(spa, dva, birth, buf)) {
478 			*lockp = hash_lock;
479 			return (buf);
480 		}
481 	}
482 	mutex_exit(hash_lock);
483 	*lockp = NULL;
484 	return (NULL);
485 }
486 
487 /*
488  * Insert an entry into the hash table.  If there is already an element
489  * equal to elem in the hash table, then the already existing element
490  * will be returned and the new element will not be inserted.
491  * Otherwise returns NULL.
492  */
493 static arc_buf_hdr_t *
494 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
495 {
496 	uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
497 	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
498 	arc_buf_hdr_t *fbuf;
499 	uint32_t i;
500 
501 	ASSERT(!HDR_IN_HASH_TABLE(buf));
502 	*lockp = hash_lock;
503 	mutex_enter(hash_lock);
504 	for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
505 	    fbuf = fbuf->b_hash_next, i++) {
506 		if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
507 			return (fbuf);
508 	}
509 
510 	buf->b_hash_next = buf_hash_table.ht_table[idx];
511 	buf_hash_table.ht_table[idx] = buf;
512 	buf->b_flags |= ARC_IN_HASH_TABLE;
513 
514 	/* collect some hash table performance data */
515 	if (i > 0) {
516 		ARCSTAT_BUMP(arcstat_hash_collisions);
517 		if (i == 1)
518 			ARCSTAT_BUMP(arcstat_hash_chains);
519 
520 		ARCSTAT_MAX(arcstat_hash_chain_max, i);
521 	}
522 
523 	ARCSTAT_BUMP(arcstat_hash_elements);
524 	ARCSTAT_MAXSTAT(arcstat_hash_elements);
525 
526 	return (NULL);
527 }
528 
529 static void
530 buf_hash_remove(arc_buf_hdr_t *buf)
531 {
532 	arc_buf_hdr_t *fbuf, **bufp;
533 	uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
534 
535 	ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
536 	ASSERT(HDR_IN_HASH_TABLE(buf));
537 
538 	bufp = &buf_hash_table.ht_table[idx];
539 	while ((fbuf = *bufp) != buf) {
540 		ASSERT(fbuf != NULL);
541 		bufp = &fbuf->b_hash_next;
542 	}
543 	*bufp = buf->b_hash_next;
544 	buf->b_hash_next = NULL;
545 	buf->b_flags &= ~ARC_IN_HASH_TABLE;
546 
547 	/* collect some hash table performance data */
548 	ARCSTAT_BUMPDOWN(arcstat_hash_elements);
549 
550 	if (buf_hash_table.ht_table[idx] &&
551 	    buf_hash_table.ht_table[idx]->b_hash_next == NULL)
552 		ARCSTAT_BUMPDOWN(arcstat_hash_chains);
553 }
554 
555 /*
556  * Global data structures and functions for the buf kmem cache.
557  */
558 static kmem_cache_t *hdr_cache;
559 static kmem_cache_t *buf_cache;
560 
561 static void
562 buf_fini(void)
563 {
564 	int i;
565 
566 	kmem_free(buf_hash_table.ht_table,
567 	    (buf_hash_table.ht_mask + 1) * sizeof (void *));
568 	for (i = 0; i < BUF_LOCKS; i++)
569 		mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
570 	kmem_cache_destroy(hdr_cache);
571 	kmem_cache_destroy(buf_cache);
572 }
573 
574 /*
575  * Constructor callback - called when the cache is empty
576  * and a new buf is requested.
577  */
578 /* ARGSUSED */
579 static int
580 hdr_cons(void *vbuf, void *unused, int kmflag)
581 {
582 	arc_buf_hdr_t *buf = vbuf;
583 
584 	bzero(buf, sizeof (arc_buf_hdr_t));
585 	refcount_create(&buf->b_refcnt);
586 	cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
587 	return (0);
588 }
589 
590 /*
591  * Destructor callback - called when a cached buf is
592  * no longer required.
593  */
594 /* ARGSUSED */
595 static void
596 hdr_dest(void *vbuf, void *unused)
597 {
598 	arc_buf_hdr_t *buf = vbuf;
599 
600 	refcount_destroy(&buf->b_refcnt);
601 	cv_destroy(&buf->b_cv);
602 }
603 
604 /*
605  * Reclaim callback -- invoked when memory is low.
606  */
607 /* ARGSUSED */
608 static void
609 hdr_recl(void *unused)
610 {
611 	dprintf("hdr_recl called\n");
612 	/*
613 	 * umem calls the reclaim func when we destroy the buf cache,
614 	 * which is after we do arc_fini().
615 	 */
616 	if (!arc_dead)
617 		cv_signal(&arc_reclaim_thr_cv);
618 }
619 
620 static void
621 buf_init(void)
622 {
623 	uint64_t *ct;
624 	uint64_t hsize = 1ULL << 12;
625 	int i, j;
626 
627 	/*
628 	 * The hash table is big enough to fill all of physical memory
629 	 * with an average 64K block size.  The table will take up
630 	 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
631 	 */
632 	while (hsize * 65536 < physmem * PAGESIZE)
633 		hsize <<= 1;
634 retry:
635 	buf_hash_table.ht_mask = hsize - 1;
636 	buf_hash_table.ht_table =
637 	    kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
638 	if (buf_hash_table.ht_table == NULL) {
639 		ASSERT(hsize > (1ULL << 8));
640 		hsize >>= 1;
641 		goto retry;
642 	}
643 
644 	hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
645 	    0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
646 	buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
647 	    0, NULL, NULL, NULL, NULL, NULL, 0);
648 
649 	for (i = 0; i < 256; i++)
650 		for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
651 			*ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
652 
653 	for (i = 0; i < BUF_LOCKS; i++) {
654 		mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
655 		    NULL, MUTEX_DEFAULT, NULL);
656 	}
657 }
658 
659 #define	ARC_MINTIME	(hz>>4) /* 62 ms */
660 
661 static void
662 arc_cksum_verify(arc_buf_t *buf)
663 {
664 	zio_cksum_t zc;
665 
666 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
667 		return;
668 
669 	mutex_enter(&buf->b_hdr->b_freeze_lock);
670 	if (buf->b_hdr->b_freeze_cksum == NULL ||
671 	    (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
672 		mutex_exit(&buf->b_hdr->b_freeze_lock);
673 		return;
674 	}
675 	fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
676 	if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
677 		panic("buffer modified while frozen!");
678 	mutex_exit(&buf->b_hdr->b_freeze_lock);
679 }
680 
681 static void
682 arc_cksum_compute(arc_buf_t *buf)
683 {
684 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
685 		return;
686 
687 	mutex_enter(&buf->b_hdr->b_freeze_lock);
688 	if (buf->b_hdr->b_freeze_cksum != NULL) {
689 		mutex_exit(&buf->b_hdr->b_freeze_lock);
690 		return;
691 	}
692 	buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
693 	fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
694 	    buf->b_hdr->b_freeze_cksum);
695 	mutex_exit(&buf->b_hdr->b_freeze_lock);
696 }
697 
698 void
699 arc_buf_thaw(arc_buf_t *buf)
700 {
701 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
702 		return;
703 
704 	if (buf->b_hdr->b_state != arc_anon)
705 		panic("modifying non-anon buffer!");
706 	if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
707 		panic("modifying buffer while i/o in progress!");
708 	arc_cksum_verify(buf);
709 	mutex_enter(&buf->b_hdr->b_freeze_lock);
710 	if (buf->b_hdr->b_freeze_cksum != NULL) {
711 		kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
712 		buf->b_hdr->b_freeze_cksum = NULL;
713 	}
714 	mutex_exit(&buf->b_hdr->b_freeze_lock);
715 }
716 
717 void
718 arc_buf_freeze(arc_buf_t *buf)
719 {
720 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
721 		return;
722 
723 	ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
724 	    buf->b_hdr->b_state == arc_anon);
725 	arc_cksum_compute(buf);
726 }
727 
728 static void
729 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
730 {
731 	ASSERT(MUTEX_HELD(hash_lock));
732 
733 	if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
734 	    (ab->b_state != arc_anon)) {
735 		uint64_t delta = ab->b_size * ab->b_datacnt;
736 		list_t *list = &ab->b_state->arcs_list[ab->b_type];
737 		uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
738 
739 		ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
740 		mutex_enter(&ab->b_state->arcs_mtx);
741 		ASSERT(list_link_active(&ab->b_arc_node));
742 		list_remove(list, ab);
743 		if (GHOST_STATE(ab->b_state)) {
744 			ASSERT3U(ab->b_datacnt, ==, 0);
745 			ASSERT3P(ab->b_buf, ==, NULL);
746 			delta = ab->b_size;
747 		}
748 		ASSERT(delta > 0);
749 		ASSERT3U(*size, >=, delta);
750 		atomic_add_64(size, -delta);
751 		mutex_exit(&ab->b_state->arcs_mtx);
752 		/* remove the prefetch flag is we get a reference */
753 		if (ab->b_flags & ARC_PREFETCH)
754 			ab->b_flags &= ~ARC_PREFETCH;
755 	}
756 }
757 
758 static int
759 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
760 {
761 	int cnt;
762 	arc_state_t *state = ab->b_state;
763 
764 	ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
765 	ASSERT(!GHOST_STATE(state));
766 
767 	if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
768 	    (state != arc_anon)) {
769 		uint64_t *size = &state->arcs_lsize[ab->b_type];
770 
771 		ASSERT(!MUTEX_HELD(&state->arcs_mtx));
772 		mutex_enter(&state->arcs_mtx);
773 		ASSERT(!list_link_active(&ab->b_arc_node));
774 		list_insert_head(&state->arcs_list[ab->b_type], ab);
775 		ASSERT(ab->b_datacnt > 0);
776 		atomic_add_64(size, ab->b_size * ab->b_datacnt);
777 		mutex_exit(&state->arcs_mtx);
778 	}
779 	return (cnt);
780 }
781 
782 /*
783  * Move the supplied buffer to the indicated state.  The mutex
784  * for the buffer must be held by the caller.
785  */
786 static void
787 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
788 {
789 	arc_state_t *old_state = ab->b_state;
790 	int64_t refcnt = refcount_count(&ab->b_refcnt);
791 	uint64_t from_delta, to_delta;
792 
793 	ASSERT(MUTEX_HELD(hash_lock));
794 	ASSERT(new_state != old_state);
795 	ASSERT(refcnt == 0 || ab->b_datacnt > 0);
796 	ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
797 
798 	from_delta = to_delta = ab->b_datacnt * ab->b_size;
799 
800 	/*
801 	 * If this buffer is evictable, transfer it from the
802 	 * old state list to the new state list.
803 	 */
804 	if (refcnt == 0) {
805 		if (old_state != arc_anon) {
806 			int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
807 			uint64_t *size = &old_state->arcs_lsize[ab->b_type];
808 
809 			if (use_mutex)
810 				mutex_enter(&old_state->arcs_mtx);
811 
812 			ASSERT(list_link_active(&ab->b_arc_node));
813 			list_remove(&old_state->arcs_list[ab->b_type], ab);
814 
815 			/*
816 			 * If prefetching out of the ghost cache,
817 			 * we will have a non-null datacnt.
818 			 */
819 			if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
820 				/* ghost elements have a ghost size */
821 				ASSERT(ab->b_buf == NULL);
822 				from_delta = ab->b_size;
823 			}
824 			ASSERT3U(*size, >=, from_delta);
825 			atomic_add_64(size, -from_delta);
826 
827 			if (use_mutex)
828 				mutex_exit(&old_state->arcs_mtx);
829 		}
830 		if (new_state != arc_anon) {
831 			int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
832 			uint64_t *size = &new_state->arcs_lsize[ab->b_type];
833 
834 			if (use_mutex)
835 				mutex_enter(&new_state->arcs_mtx);
836 
837 			list_insert_head(&new_state->arcs_list[ab->b_type], ab);
838 
839 			/* ghost elements have a ghost size */
840 			if (GHOST_STATE(new_state)) {
841 				ASSERT(ab->b_datacnt == 0);
842 				ASSERT(ab->b_buf == NULL);
843 				to_delta = ab->b_size;
844 			}
845 			atomic_add_64(size, to_delta);
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 	/*
1262 	 * We have just evicted some date into the ghost state, make
1263 	 * sure we also adjust the ghost state size if necessary.
1264 	 */
1265 	if (arc_no_grow &&
1266 	    arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1267 		int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1268 		    arc_mru_ghost->arcs_size - arc_c;
1269 
1270 		if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1271 			int64_t todelete =
1272 			    MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1273 			arc_evict_ghost(arc_mru_ghost, todelete);
1274 		} else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1275 			int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1276 			    arc_mru_ghost->arcs_size +
1277 			    arc_mfu_ghost->arcs_size - arc_c);
1278 			arc_evict_ghost(arc_mfu_ghost, todelete);
1279 		}
1280 	}
1281 
1282 	return (stolen);
1283 }
1284 
1285 /*
1286  * Remove buffers from list until we've removed the specified number of
1287  * bytes.  Destroy the buffers that are removed.
1288  */
1289 static void
1290 arc_evict_ghost(arc_state_t *state, int64_t bytes)
1291 {
1292 	arc_buf_hdr_t *ab, *ab_prev;
1293 	list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1294 	kmutex_t *hash_lock;
1295 	uint64_t bytes_deleted = 0;
1296 	uint64_t bufs_skipped = 0;
1297 
1298 	ASSERT(GHOST_STATE(state));
1299 top:
1300 	mutex_enter(&state->arcs_mtx);
1301 	for (ab = list_tail(list); ab; ab = ab_prev) {
1302 		ab_prev = list_prev(list, ab);
1303 		hash_lock = HDR_LOCK(ab);
1304 		if (mutex_tryenter(hash_lock)) {
1305 			ASSERT(!HDR_IO_IN_PROGRESS(ab));
1306 			ASSERT(ab->b_buf == NULL);
1307 			arc_change_state(arc_anon, ab, hash_lock);
1308 			mutex_exit(hash_lock);
1309 			ARCSTAT_BUMP(arcstat_deleted);
1310 			bytes_deleted += ab->b_size;
1311 			arc_hdr_destroy(ab);
1312 			DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1313 			if (bytes >= 0 && bytes_deleted >= bytes)
1314 				break;
1315 		} else {
1316 			if (bytes < 0) {
1317 				mutex_exit(&state->arcs_mtx);
1318 				mutex_enter(hash_lock);
1319 				mutex_exit(hash_lock);
1320 				goto top;
1321 			}
1322 			bufs_skipped += 1;
1323 		}
1324 	}
1325 	mutex_exit(&state->arcs_mtx);
1326 
1327 	if (list == &state->arcs_list[ARC_BUFC_DATA] &&
1328 	    (bytes < 0 || bytes_deleted < bytes)) {
1329 		list = &state->arcs_list[ARC_BUFC_METADATA];
1330 		goto top;
1331 	}
1332 
1333 	if (bufs_skipped) {
1334 		ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1335 		ASSERT(bytes >= 0);
1336 	}
1337 
1338 	if (bytes_deleted < bytes)
1339 		dprintf("only deleted %lld bytes from %p",
1340 		    (longlong_t)bytes_deleted, state);
1341 }
1342 
1343 static void
1344 arc_adjust(void)
1345 {
1346 	int64_t top_sz, mru_over, arc_over, todelete;
1347 
1348 	top_sz = arc_anon->arcs_size + arc_mru->arcs_size;
1349 
1350 	if (top_sz > arc_p && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
1351 		int64_t toevict =
1352 		    MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], top_sz - arc_p);
1353 		(void) arc_evict(arc_mru, toevict, FALSE, ARC_BUFC_DATA);
1354 		top_sz = arc_anon->arcs_size + arc_mru->arcs_size;
1355 	}
1356 
1357 	if (top_sz > arc_p && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1358 		int64_t toevict =
1359 		    MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], top_sz - arc_p);
1360 		(void) arc_evict(arc_mru, toevict, FALSE, ARC_BUFC_METADATA);
1361 		top_sz = arc_anon->arcs_size + arc_mru->arcs_size;
1362 	}
1363 
1364 	mru_over = top_sz + arc_mru_ghost->arcs_size - arc_c;
1365 
1366 	if (mru_over > 0) {
1367 		if (arc_mru_ghost->arcs_size > 0) {
1368 			todelete = MIN(arc_mru_ghost->arcs_size, mru_over);
1369 			arc_evict_ghost(arc_mru_ghost, todelete);
1370 		}
1371 	}
1372 
1373 	if ((arc_over = arc_size - arc_c) > 0) {
1374 		int64_t tbl_over;
1375 
1376 		if (arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
1377 			int64_t toevict =
1378 			    MIN(arc_mfu->arcs_lsize[ARC_BUFC_DATA], arc_over);
1379 			(void) arc_evict(arc_mfu, toevict, FALSE,
1380 			    ARC_BUFC_DATA);
1381 			arc_over = arc_size - arc_c;
1382 		}
1383 
1384 		if (arc_over > 0 &&
1385 		    arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1386 			int64_t toevict =
1387 			    MIN(arc_mfu->arcs_lsize[ARC_BUFC_METADATA],
1388 			    arc_over);
1389 			(void) arc_evict(arc_mfu, toevict, FALSE,
1390 			    ARC_BUFC_METADATA);
1391 		}
1392 
1393 		tbl_over = arc_size + arc_mru_ghost->arcs_size +
1394 		    arc_mfu_ghost->arcs_size - arc_c * 2;
1395 
1396 		if (tbl_over > 0 && arc_mfu_ghost->arcs_size > 0) {
1397 			todelete = MIN(arc_mfu_ghost->arcs_size, tbl_over);
1398 			arc_evict_ghost(arc_mfu_ghost, todelete);
1399 		}
1400 	}
1401 }
1402 
1403 static void
1404 arc_do_user_evicts(void)
1405 {
1406 	mutex_enter(&arc_eviction_mtx);
1407 	while (arc_eviction_list != NULL) {
1408 		arc_buf_t *buf = arc_eviction_list;
1409 		arc_eviction_list = buf->b_next;
1410 		buf->b_hdr = NULL;
1411 		mutex_exit(&arc_eviction_mtx);
1412 
1413 		if (buf->b_efunc != NULL)
1414 			VERIFY(buf->b_efunc(buf) == 0);
1415 
1416 		buf->b_efunc = NULL;
1417 		buf->b_private = NULL;
1418 		kmem_cache_free(buf_cache, buf);
1419 		mutex_enter(&arc_eviction_mtx);
1420 	}
1421 	mutex_exit(&arc_eviction_mtx);
1422 }
1423 
1424 /*
1425  * Flush all *evictable* data from the cache.
1426  * NOTE: this will not touch "active" (i.e. referenced) data.
1427  */
1428 void
1429 arc_flush(void)
1430 {
1431 	while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA]))
1432 		(void) arc_evict(arc_mru, -1, FALSE, ARC_BUFC_DATA);
1433 	while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA]))
1434 		(void) arc_evict(arc_mru, -1, FALSE, ARC_BUFC_METADATA);
1435 	while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA]))
1436 		(void) arc_evict(arc_mfu, -1, FALSE, ARC_BUFC_DATA);
1437 	while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA]))
1438 		(void) arc_evict(arc_mfu, -1, FALSE, ARC_BUFC_METADATA);
1439 
1440 	arc_evict_ghost(arc_mru_ghost, -1);
1441 	arc_evict_ghost(arc_mfu_ghost, -1);
1442 
1443 	mutex_enter(&arc_reclaim_thr_lock);
1444 	arc_do_user_evicts();
1445 	mutex_exit(&arc_reclaim_thr_lock);
1446 	ASSERT(arc_eviction_list == NULL);
1447 }
1448 
1449 int arc_shrink_shift = 5;		/* log2(fraction of arc to reclaim) */
1450 
1451 void
1452 arc_shrink(void)
1453 {
1454 	if (arc_c > arc_c_min) {
1455 		uint64_t to_free;
1456 
1457 #ifdef _KERNEL
1458 		to_free = MAX(arc_c >> arc_shrink_shift, ptob(needfree));
1459 #else
1460 		to_free = arc_c >> arc_shrink_shift;
1461 #endif
1462 		if (arc_c > arc_c_min + to_free)
1463 			atomic_add_64(&arc_c, -to_free);
1464 		else
1465 			arc_c = arc_c_min;
1466 
1467 		atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
1468 		if (arc_c > arc_size)
1469 			arc_c = MAX(arc_size, arc_c_min);
1470 		if (arc_p > arc_c)
1471 			arc_p = (arc_c >> 1);
1472 		ASSERT(arc_c >= arc_c_min);
1473 		ASSERT((int64_t)arc_p >= 0);
1474 	}
1475 
1476 	if (arc_size > arc_c)
1477 		arc_adjust();
1478 }
1479 
1480 static int
1481 arc_reclaim_needed(void)
1482 {
1483 	uint64_t extra;
1484 
1485 #ifdef _KERNEL
1486 
1487 	if (needfree)
1488 		return (1);
1489 
1490 	/*
1491 	 * take 'desfree' extra pages, so we reclaim sooner, rather than later
1492 	 */
1493 	extra = desfree;
1494 
1495 	/*
1496 	 * check that we're out of range of the pageout scanner.  It starts to
1497 	 * schedule paging if freemem is less than lotsfree and needfree.
1498 	 * lotsfree is the high-water mark for pageout, and needfree is the
1499 	 * number of needed free pages.  We add extra pages here to make sure
1500 	 * the scanner doesn't start up while we're freeing memory.
1501 	 */
1502 	if (freemem < lotsfree + needfree + extra)
1503 		return (1);
1504 
1505 	/*
1506 	 * check to make sure that swapfs has enough space so that anon
1507 	 * reservations can still succeeed. anon_resvmem() checks that the
1508 	 * availrmem is greater than swapfs_minfree, and the number of reserved
1509 	 * swap pages.  We also add a bit of extra here just to prevent
1510 	 * circumstances from getting really dire.
1511 	 */
1512 	if (availrmem < swapfs_minfree + swapfs_reserve + extra)
1513 		return (1);
1514 
1515 #if defined(__i386)
1516 	/*
1517 	 * If we're on an i386 platform, it's possible that we'll exhaust the
1518 	 * kernel heap space before we ever run out of available physical
1519 	 * memory.  Most checks of the size of the heap_area compare against
1520 	 * tune.t_minarmem, which is the minimum available real memory that we
1521 	 * can have in the system.  However, this is generally fixed at 25 pages
1522 	 * which is so low that it's useless.  In this comparison, we seek to
1523 	 * calculate the total heap-size, and reclaim if more than 3/4ths of the
1524 	 * heap is allocated.  (Or, in the caclulation, if less than 1/4th is
1525 	 * free)
1526 	 */
1527 	if (btop(vmem_size(heap_arena, VMEM_FREE)) <
1528 	    (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2))
1529 		return (1);
1530 #endif
1531 
1532 #else
1533 	if (spa_get_random(100) == 0)
1534 		return (1);
1535 #endif
1536 	return (0);
1537 }
1538 
1539 static void
1540 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
1541 {
1542 	size_t			i;
1543 	kmem_cache_t		*prev_cache = NULL;
1544 	kmem_cache_t		*prev_data_cache = NULL;
1545 	extern kmem_cache_t	*zio_buf_cache[];
1546 	extern kmem_cache_t	*zio_data_buf_cache[];
1547 
1548 #ifdef _KERNEL
1549 	if (arc_meta_used >= arc_meta_limit) {
1550 		/*
1551 		 * We are exceeding our meta-data cache limit.
1552 		 * Purge some DNLC entries to release holds on meta-data.
1553 		 */
1554 		dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
1555 	}
1556 #if defined(__i386)
1557 	/*
1558 	 * Reclaim unused memory from all kmem caches.
1559 	 */
1560 	kmem_reap();
1561 #endif
1562 #endif
1563 
1564 	/*
1565 	 * An agressive reclamation will shrink the cache size as well as
1566 	 * reap free buffers from the arc kmem caches.
1567 	 */
1568 	if (strat == ARC_RECLAIM_AGGR)
1569 		arc_shrink();
1570 
1571 	for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
1572 		if (zio_buf_cache[i] != prev_cache) {
1573 			prev_cache = zio_buf_cache[i];
1574 			kmem_cache_reap_now(zio_buf_cache[i]);
1575 		}
1576 		if (zio_data_buf_cache[i] != prev_data_cache) {
1577 			prev_data_cache = zio_data_buf_cache[i];
1578 			kmem_cache_reap_now(zio_data_buf_cache[i]);
1579 		}
1580 	}
1581 	kmem_cache_reap_now(buf_cache);
1582 	kmem_cache_reap_now(hdr_cache);
1583 }
1584 
1585 static void
1586 arc_reclaim_thread(void)
1587 {
1588 	clock_t			growtime = 0;
1589 	arc_reclaim_strategy_t	last_reclaim = ARC_RECLAIM_CONS;
1590 	callb_cpr_t		cpr;
1591 
1592 	CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
1593 
1594 	mutex_enter(&arc_reclaim_thr_lock);
1595 	while (arc_thread_exit == 0) {
1596 		if (arc_reclaim_needed()) {
1597 
1598 			if (arc_no_grow) {
1599 				if (last_reclaim == ARC_RECLAIM_CONS) {
1600 					last_reclaim = ARC_RECLAIM_AGGR;
1601 				} else {
1602 					last_reclaim = ARC_RECLAIM_CONS;
1603 				}
1604 			} else {
1605 				arc_no_grow = TRUE;
1606 				last_reclaim = ARC_RECLAIM_AGGR;
1607 				membar_producer();
1608 			}
1609 
1610 			/* reset the growth delay for every reclaim */
1611 			growtime = lbolt + (arc_grow_retry * hz);
1612 
1613 			arc_kmem_reap_now(last_reclaim);
1614 
1615 		} else if (arc_no_grow && lbolt >= growtime) {
1616 			arc_no_grow = FALSE;
1617 		}
1618 
1619 		if (2 * arc_c < arc_size +
1620 		    arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size)
1621 			arc_adjust();
1622 
1623 		if (arc_eviction_list != NULL)
1624 			arc_do_user_evicts();
1625 
1626 		/* block until needed, or one second, whichever is shorter */
1627 		CALLB_CPR_SAFE_BEGIN(&cpr);
1628 		(void) cv_timedwait(&arc_reclaim_thr_cv,
1629 		    &arc_reclaim_thr_lock, (lbolt + hz));
1630 		CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
1631 	}
1632 
1633 	arc_thread_exit = 0;
1634 	cv_broadcast(&arc_reclaim_thr_cv);
1635 	CALLB_CPR_EXIT(&cpr);		/* drops arc_reclaim_thr_lock */
1636 	thread_exit();
1637 }
1638 
1639 /*
1640  * Adapt arc info given the number of bytes we are trying to add and
1641  * the state that we are comming from.  This function is only called
1642  * when we are adding new content to the cache.
1643  */
1644 static void
1645 arc_adapt(int bytes, arc_state_t *state)
1646 {
1647 	int mult;
1648 
1649 	ASSERT(bytes > 0);
1650 	/*
1651 	 * Adapt the target size of the MRU list:
1652 	 *	- if we just hit in the MRU ghost list, then increase
1653 	 *	  the target size of the MRU list.
1654 	 *	- if we just hit in the MFU ghost list, then increase
1655 	 *	  the target size of the MFU list by decreasing the
1656 	 *	  target size of the MRU list.
1657 	 */
1658 	if (state == arc_mru_ghost) {
1659 		mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
1660 		    1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
1661 
1662 		arc_p = MIN(arc_c, arc_p + bytes * mult);
1663 	} else if (state == arc_mfu_ghost) {
1664 		mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
1665 		    1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
1666 
1667 		arc_p = MAX(0, (int64_t)arc_p - bytes * mult);
1668 	}
1669 	ASSERT((int64_t)arc_p >= 0);
1670 
1671 	if (arc_reclaim_needed()) {
1672 		cv_signal(&arc_reclaim_thr_cv);
1673 		return;
1674 	}
1675 
1676 	if (arc_no_grow)
1677 		return;
1678 
1679 	if (arc_c >= arc_c_max)
1680 		return;
1681 
1682 	/*
1683 	 * If we're within (2 * maxblocksize) bytes of the target
1684 	 * cache size, increment the target cache size
1685 	 */
1686 	if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
1687 		atomic_add_64(&arc_c, (int64_t)bytes);
1688 		if (arc_c > arc_c_max)
1689 			arc_c = arc_c_max;
1690 		else if (state == arc_anon)
1691 			atomic_add_64(&arc_p, (int64_t)bytes);
1692 		if (arc_p > arc_c)
1693 			arc_p = arc_c;
1694 	}
1695 	ASSERT((int64_t)arc_p >= 0);
1696 }
1697 
1698 /*
1699  * Check if the cache has reached its limits and eviction is required
1700  * prior to insert.
1701  */
1702 static int
1703 arc_evict_needed(arc_buf_contents_t type)
1704 {
1705 	if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
1706 		return (1);
1707 
1708 #ifdef _KERNEL
1709 	/*
1710 	 * If zio data pages are being allocated out of a separate heap segment,
1711 	 * then enforce that the size of available vmem for this area remains
1712 	 * above about 1/32nd free.
1713 	 */
1714 	if (type == ARC_BUFC_DATA && zio_arena != NULL &&
1715 	    vmem_size(zio_arena, VMEM_FREE) <
1716 	    (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
1717 		return (1);
1718 #endif
1719 
1720 	if (arc_reclaim_needed())
1721 		return (1);
1722 
1723 	return (arc_size > arc_c);
1724 }
1725 
1726 /*
1727  * The buffer, supplied as the first argument, needs a data block.
1728  * So, if we are at cache max, determine which cache should be victimized.
1729  * We have the following cases:
1730  *
1731  * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
1732  * In this situation if we're out of space, but the resident size of the MFU is
1733  * under the limit, victimize the MFU cache to satisfy this insertion request.
1734  *
1735  * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
1736  * Here, we've used up all of the available space for the MRU, so we need to
1737  * evict from our own cache instead.  Evict from the set of resident MRU
1738  * entries.
1739  *
1740  * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
1741  * c minus p represents the MFU space in the cache, since p is the size of the
1742  * cache that is dedicated to the MRU.  In this situation there's still space on
1743  * the MFU side, so the MRU side needs to be victimized.
1744  *
1745  * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
1746  * MFU's resident set is consuming more space than it has been allotted.  In
1747  * this situation, we must victimize our own cache, the MFU, for this insertion.
1748  */
1749 static void
1750 arc_get_data_buf(arc_buf_t *buf)
1751 {
1752 	arc_state_t		*state = buf->b_hdr->b_state;
1753 	uint64_t		size = buf->b_hdr->b_size;
1754 	arc_buf_contents_t	type = buf->b_hdr->b_type;
1755 
1756 	arc_adapt(size, state);
1757 
1758 	/*
1759 	 * We have not yet reached cache maximum size,
1760 	 * just allocate a new buffer.
1761 	 */
1762 	if (!arc_evict_needed(type)) {
1763 		if (type == ARC_BUFC_METADATA) {
1764 			buf->b_data = zio_buf_alloc(size);
1765 			arc_space_consume(size);
1766 		} else {
1767 			ASSERT(type == ARC_BUFC_DATA);
1768 			buf->b_data = zio_data_buf_alloc(size);
1769 			atomic_add_64(&arc_size, size);
1770 		}
1771 		goto out;
1772 	}
1773 
1774 	/*
1775 	 * If we are prefetching from the mfu ghost list, this buffer
1776 	 * will end up on the mru list; so steal space from there.
1777 	 */
1778 	if (state == arc_mfu_ghost)
1779 		state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
1780 	else if (state == arc_mru_ghost)
1781 		state = arc_mru;
1782 
1783 	if (state == arc_mru || state == arc_anon) {
1784 		uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
1785 		state = (arc_mfu->arcs_lsize[type] > 0 &&
1786 		    arc_p > mru_used) ? arc_mfu : arc_mru;
1787 	} else {
1788 		/* MFU cases */
1789 		uint64_t mfu_space = arc_c - arc_p;
1790 		state =  (arc_mru->arcs_lsize[type] > 0 &&
1791 		    mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
1792 	}
1793 	if ((buf->b_data = arc_evict(state, size, TRUE, type)) == NULL) {
1794 		if (type == ARC_BUFC_METADATA) {
1795 			buf->b_data = zio_buf_alloc(size);
1796 			arc_space_consume(size);
1797 		} else {
1798 			ASSERT(type == ARC_BUFC_DATA);
1799 			buf->b_data = zio_data_buf_alloc(size);
1800 			atomic_add_64(&arc_size, size);
1801 		}
1802 		ARCSTAT_BUMP(arcstat_recycle_miss);
1803 	}
1804 	ASSERT(buf->b_data != NULL);
1805 out:
1806 	/*
1807 	 * Update the state size.  Note that ghost states have a
1808 	 * "ghost size" and so don't need to be updated.
1809 	 */
1810 	if (!GHOST_STATE(buf->b_hdr->b_state)) {
1811 		arc_buf_hdr_t *hdr = buf->b_hdr;
1812 
1813 		atomic_add_64(&hdr->b_state->arcs_size, size);
1814 		if (list_link_active(&hdr->b_arc_node)) {
1815 			ASSERT(refcount_is_zero(&hdr->b_refcnt));
1816 			atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
1817 		}
1818 		/*
1819 		 * If we are growing the cache, and we are adding anonymous
1820 		 * data, and we have outgrown arc_p, update arc_p
1821 		 */
1822 		if (arc_size < arc_c && hdr->b_state == arc_anon &&
1823 		    arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
1824 			arc_p = MIN(arc_c, arc_p + size);
1825 	}
1826 }
1827 
1828 /*
1829  * This routine is called whenever a buffer is accessed.
1830  * NOTE: the hash lock is dropped in this function.
1831  */
1832 static void
1833 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
1834 {
1835 	ASSERT(MUTEX_HELD(hash_lock));
1836 
1837 	if (buf->b_state == arc_anon) {
1838 		/*
1839 		 * This buffer is not in the cache, and does not
1840 		 * appear in our "ghost" list.  Add the new buffer
1841 		 * to the MRU state.
1842 		 */
1843 
1844 		ASSERT(buf->b_arc_access == 0);
1845 		buf->b_arc_access = lbolt;
1846 		DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
1847 		arc_change_state(arc_mru, buf, hash_lock);
1848 
1849 	} else if (buf->b_state == arc_mru) {
1850 		/*
1851 		 * If this buffer is here because of a prefetch, then either:
1852 		 * - clear the flag if this is a "referencing" read
1853 		 *   (any subsequent access will bump this into the MFU state).
1854 		 * or
1855 		 * - move the buffer to the head of the list if this is
1856 		 *   another prefetch (to make it less likely to be evicted).
1857 		 */
1858 		if ((buf->b_flags & ARC_PREFETCH) != 0) {
1859 			if (refcount_count(&buf->b_refcnt) == 0) {
1860 				ASSERT(list_link_active(&buf->b_arc_node));
1861 			} else {
1862 				buf->b_flags &= ~ARC_PREFETCH;
1863 				ARCSTAT_BUMP(arcstat_mru_hits);
1864 			}
1865 			buf->b_arc_access = lbolt;
1866 			return;
1867 		}
1868 
1869 		/*
1870 		 * This buffer has been "accessed" only once so far,
1871 		 * but it is still in the cache. Move it to the MFU
1872 		 * state.
1873 		 */
1874 		if (lbolt > buf->b_arc_access + ARC_MINTIME) {
1875 			/*
1876 			 * More than 125ms have passed since we
1877 			 * instantiated this buffer.  Move it to the
1878 			 * most frequently used state.
1879 			 */
1880 			buf->b_arc_access = lbolt;
1881 			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
1882 			arc_change_state(arc_mfu, buf, hash_lock);
1883 		}
1884 		ARCSTAT_BUMP(arcstat_mru_hits);
1885 	} else if (buf->b_state == arc_mru_ghost) {
1886 		arc_state_t	*new_state;
1887 		/*
1888 		 * This buffer has been "accessed" recently, but
1889 		 * was evicted from the cache.  Move it to the
1890 		 * MFU state.
1891 		 */
1892 
1893 		if (buf->b_flags & ARC_PREFETCH) {
1894 			new_state = arc_mru;
1895 			if (refcount_count(&buf->b_refcnt) > 0)
1896 				buf->b_flags &= ~ARC_PREFETCH;
1897 			DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
1898 		} else {
1899 			new_state = arc_mfu;
1900 			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
1901 		}
1902 
1903 		buf->b_arc_access = lbolt;
1904 		arc_change_state(new_state, buf, hash_lock);
1905 
1906 		ARCSTAT_BUMP(arcstat_mru_ghost_hits);
1907 	} else if (buf->b_state == arc_mfu) {
1908 		/*
1909 		 * This buffer has been accessed more than once and is
1910 		 * still in the cache.  Keep it in the MFU state.
1911 		 *
1912 		 * NOTE: an add_reference() that occurred when we did
1913 		 * the arc_read() will have kicked this off the list.
1914 		 * If it was a prefetch, we will explicitly move it to
1915 		 * the head of the list now.
1916 		 */
1917 		if ((buf->b_flags & ARC_PREFETCH) != 0) {
1918 			ASSERT(refcount_count(&buf->b_refcnt) == 0);
1919 			ASSERT(list_link_active(&buf->b_arc_node));
1920 		}
1921 		ARCSTAT_BUMP(arcstat_mfu_hits);
1922 		buf->b_arc_access = lbolt;
1923 	} else if (buf->b_state == arc_mfu_ghost) {
1924 		arc_state_t	*new_state = arc_mfu;
1925 		/*
1926 		 * This buffer has been accessed more than once but has
1927 		 * been evicted from the cache.  Move it back to the
1928 		 * MFU state.
1929 		 */
1930 
1931 		if (buf->b_flags & ARC_PREFETCH) {
1932 			/*
1933 			 * This is a prefetch access...
1934 			 * move this block back to the MRU state.
1935 			 */
1936 			ASSERT3U(refcount_count(&buf->b_refcnt), ==, 0);
1937 			new_state = arc_mru;
1938 		}
1939 
1940 		buf->b_arc_access = lbolt;
1941 		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
1942 		arc_change_state(new_state, buf, hash_lock);
1943 
1944 		ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
1945 	} else {
1946 		ASSERT(!"invalid arc state");
1947 	}
1948 }
1949 
1950 /* a generic arc_done_func_t which you can use */
1951 /* ARGSUSED */
1952 void
1953 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
1954 {
1955 	bcopy(buf->b_data, arg, buf->b_hdr->b_size);
1956 	VERIFY(arc_buf_remove_ref(buf, arg) == 1);
1957 }
1958 
1959 /* a generic arc_done_func_t */
1960 void
1961 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
1962 {
1963 	arc_buf_t **bufp = arg;
1964 	if (zio && zio->io_error) {
1965 		VERIFY(arc_buf_remove_ref(buf, arg) == 1);
1966 		*bufp = NULL;
1967 	} else {
1968 		*bufp = buf;
1969 	}
1970 }
1971 
1972 static void
1973 arc_read_done(zio_t *zio)
1974 {
1975 	arc_buf_hdr_t	*hdr, *found;
1976 	arc_buf_t	*buf;
1977 	arc_buf_t	*abuf;	/* buffer we're assigning to callback */
1978 	kmutex_t	*hash_lock;
1979 	arc_callback_t	*callback_list, *acb;
1980 	int		freeable = FALSE;
1981 
1982 	buf = zio->io_private;
1983 	hdr = buf->b_hdr;
1984 
1985 	/*
1986 	 * The hdr was inserted into hash-table and removed from lists
1987 	 * prior to starting I/O.  We should find this header, since
1988 	 * it's in the hash table, and it should be legit since it's
1989 	 * not possible to evict it during the I/O.  The only possible
1990 	 * reason for it not to be found is if we were freed during the
1991 	 * read.
1992 	 */
1993 	found = buf_hash_find(zio->io_spa, &hdr->b_dva, hdr->b_birth,
1994 	    &hash_lock);
1995 
1996 	ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
1997 	    (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))));
1998 
1999 	/* byteswap if necessary */
2000 	callback_list = hdr->b_acb;
2001 	ASSERT(callback_list != NULL);
2002 	if (BP_SHOULD_BYTESWAP(zio->io_bp) && callback_list->acb_byteswap)
2003 		callback_list->acb_byteswap(buf->b_data, hdr->b_size);
2004 
2005 	arc_cksum_compute(buf);
2006 
2007 	/* create copies of the data buffer for the callers */
2008 	abuf = buf;
2009 	for (acb = callback_list; acb; acb = acb->acb_next) {
2010 		if (acb->acb_done) {
2011 			if (abuf == NULL)
2012 				abuf = arc_buf_clone(buf);
2013 			acb->acb_buf = abuf;
2014 			abuf = NULL;
2015 		}
2016 	}
2017 	hdr->b_acb = NULL;
2018 	hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2019 	ASSERT(!HDR_BUF_AVAILABLE(hdr));
2020 	if (abuf == buf)
2021 		hdr->b_flags |= ARC_BUF_AVAILABLE;
2022 
2023 	ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2024 
2025 	if (zio->io_error != 0) {
2026 		hdr->b_flags |= ARC_IO_ERROR;
2027 		if (hdr->b_state != arc_anon)
2028 			arc_change_state(arc_anon, hdr, hash_lock);
2029 		if (HDR_IN_HASH_TABLE(hdr))
2030 			buf_hash_remove(hdr);
2031 		freeable = refcount_is_zero(&hdr->b_refcnt);
2032 		/* convert checksum errors into IO errors */
2033 		if (zio->io_error == ECKSUM)
2034 			zio->io_error = EIO;
2035 	}
2036 
2037 	/*
2038 	 * Broadcast before we drop the hash_lock to avoid the possibility
2039 	 * that the hdr (and hence the cv) might be freed before we get to
2040 	 * the cv_broadcast().
2041 	 */
2042 	cv_broadcast(&hdr->b_cv);
2043 
2044 	if (hash_lock) {
2045 		/*
2046 		 * Only call arc_access on anonymous buffers.  This is because
2047 		 * if we've issued an I/O for an evicted buffer, we've already
2048 		 * called arc_access (to prevent any simultaneous readers from
2049 		 * getting confused).
2050 		 */
2051 		if (zio->io_error == 0 && hdr->b_state == arc_anon)
2052 			arc_access(hdr, hash_lock);
2053 		mutex_exit(hash_lock);
2054 	} else {
2055 		/*
2056 		 * This block was freed while we waited for the read to
2057 		 * complete.  It has been removed from the hash table and
2058 		 * moved to the anonymous state (so that it won't show up
2059 		 * in the cache).
2060 		 */
2061 		ASSERT3P(hdr->b_state, ==, arc_anon);
2062 		freeable = refcount_is_zero(&hdr->b_refcnt);
2063 	}
2064 
2065 	/* execute each callback and free its structure */
2066 	while ((acb = callback_list) != NULL) {
2067 		if (acb->acb_done)
2068 			acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2069 
2070 		if (acb->acb_zio_dummy != NULL) {
2071 			acb->acb_zio_dummy->io_error = zio->io_error;
2072 			zio_nowait(acb->acb_zio_dummy);
2073 		}
2074 
2075 		callback_list = acb->acb_next;
2076 		kmem_free(acb, sizeof (arc_callback_t));
2077 	}
2078 
2079 	if (freeable)
2080 		arc_hdr_destroy(hdr);
2081 }
2082 
2083 /*
2084  * "Read" the block block at the specified DVA (in bp) via the
2085  * cache.  If the block is found in the cache, invoke the provided
2086  * callback immediately and return.  Note that the `zio' parameter
2087  * in the callback will be NULL in this case, since no IO was
2088  * required.  If the block is not in the cache pass the read request
2089  * on to the spa with a substitute callback function, so that the
2090  * requested block will be added to the cache.
2091  *
2092  * If a read request arrives for a block that has a read in-progress,
2093  * either wait for the in-progress read to complete (and return the
2094  * results); or, if this is a read with a "done" func, add a record
2095  * to the read to invoke the "done" func when the read completes,
2096  * and return; or just return.
2097  *
2098  * arc_read_done() will invoke all the requested "done" functions
2099  * for readers of this block.
2100  */
2101 int
2102 arc_read(zio_t *pio, spa_t *spa, blkptr_t *bp, arc_byteswap_func_t *swap,
2103     arc_done_func_t *done, void *private, int priority, int flags,
2104     uint32_t *arc_flags, zbookmark_t *zb)
2105 {
2106 	arc_buf_hdr_t *hdr;
2107 	arc_buf_t *buf;
2108 	kmutex_t *hash_lock;
2109 	zio_t	*rzio;
2110 
2111 top:
2112 	hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock);
2113 	if (hdr && hdr->b_datacnt > 0) {
2114 
2115 		*arc_flags |= ARC_CACHED;
2116 
2117 		if (HDR_IO_IN_PROGRESS(hdr)) {
2118 
2119 			if (*arc_flags & ARC_WAIT) {
2120 				cv_wait(&hdr->b_cv, hash_lock);
2121 				mutex_exit(hash_lock);
2122 				goto top;
2123 			}
2124 			ASSERT(*arc_flags & ARC_NOWAIT);
2125 
2126 			if (done) {
2127 				arc_callback_t	*acb = NULL;
2128 
2129 				acb = kmem_zalloc(sizeof (arc_callback_t),
2130 				    KM_SLEEP);
2131 				acb->acb_done = done;
2132 				acb->acb_private = private;
2133 				acb->acb_byteswap = swap;
2134 				if (pio != NULL)
2135 					acb->acb_zio_dummy = zio_null(pio,
2136 					    spa, NULL, NULL, flags);
2137 
2138 				ASSERT(acb->acb_done != NULL);
2139 				acb->acb_next = hdr->b_acb;
2140 				hdr->b_acb = acb;
2141 				add_reference(hdr, hash_lock, private);
2142 				mutex_exit(hash_lock);
2143 				return (0);
2144 			}
2145 			mutex_exit(hash_lock);
2146 			return (0);
2147 		}
2148 
2149 		ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2150 
2151 		if (done) {
2152 			add_reference(hdr, hash_lock, private);
2153 			/*
2154 			 * If this block is already in use, create a new
2155 			 * copy of the data so that we will be guaranteed
2156 			 * that arc_release() will always succeed.
2157 			 */
2158 			buf = hdr->b_buf;
2159 			ASSERT(buf);
2160 			ASSERT(buf->b_data);
2161 			if (HDR_BUF_AVAILABLE(hdr)) {
2162 				ASSERT(buf->b_efunc == NULL);
2163 				hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2164 			} else {
2165 				buf = arc_buf_clone(buf);
2166 			}
2167 		} else if (*arc_flags & ARC_PREFETCH &&
2168 		    refcount_count(&hdr->b_refcnt) == 0) {
2169 			hdr->b_flags |= ARC_PREFETCH;
2170 		}
2171 		DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2172 		arc_access(hdr, hash_lock);
2173 		mutex_exit(hash_lock);
2174 		ARCSTAT_BUMP(arcstat_hits);
2175 		ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2176 		    demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2177 		    data, metadata, hits);
2178 
2179 		if (done)
2180 			done(NULL, buf, private);
2181 	} else {
2182 		uint64_t size = BP_GET_LSIZE(bp);
2183 		arc_callback_t	*acb;
2184 
2185 		if (hdr == NULL) {
2186 			/* this block is not in the cache */
2187 			arc_buf_hdr_t	*exists;
2188 			arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2189 			buf = arc_buf_alloc(spa, size, private, type);
2190 			hdr = buf->b_hdr;
2191 			hdr->b_dva = *BP_IDENTITY(bp);
2192 			hdr->b_birth = bp->blk_birth;
2193 			hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
2194 			exists = buf_hash_insert(hdr, &hash_lock);
2195 			if (exists) {
2196 				/* somebody beat us to the hash insert */
2197 				mutex_exit(hash_lock);
2198 				bzero(&hdr->b_dva, sizeof (dva_t));
2199 				hdr->b_birth = 0;
2200 				hdr->b_cksum0 = 0;
2201 				(void) arc_buf_remove_ref(buf, private);
2202 				goto top; /* restart the IO request */
2203 			}
2204 			/* if this is a prefetch, we don't have a reference */
2205 			if (*arc_flags & ARC_PREFETCH) {
2206 				(void) remove_reference(hdr, hash_lock,
2207 				    private);
2208 				hdr->b_flags |= ARC_PREFETCH;
2209 			}
2210 			if (BP_GET_LEVEL(bp) > 0)
2211 				hdr->b_flags |= ARC_INDIRECT;
2212 		} else {
2213 			/* this block is in the ghost cache */
2214 			ASSERT(GHOST_STATE(hdr->b_state));
2215 			ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2216 			ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 0);
2217 			ASSERT(hdr->b_buf == NULL);
2218 
2219 			/* if this is a prefetch, we don't have a reference */
2220 			if (*arc_flags & ARC_PREFETCH)
2221 				hdr->b_flags |= ARC_PREFETCH;
2222 			else
2223 				add_reference(hdr, hash_lock, private);
2224 			buf = kmem_cache_alloc(buf_cache, KM_SLEEP);
2225 			buf->b_hdr = hdr;
2226 			buf->b_data = NULL;
2227 			buf->b_efunc = NULL;
2228 			buf->b_private = NULL;
2229 			buf->b_next = NULL;
2230 			hdr->b_buf = buf;
2231 			arc_get_data_buf(buf);
2232 			ASSERT(hdr->b_datacnt == 0);
2233 			hdr->b_datacnt = 1;
2234 
2235 		}
2236 
2237 		acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
2238 		acb->acb_done = done;
2239 		acb->acb_private = private;
2240 		acb->acb_byteswap = swap;
2241 
2242 		ASSERT(hdr->b_acb == NULL);
2243 		hdr->b_acb = acb;
2244 		hdr->b_flags |= ARC_IO_IN_PROGRESS;
2245 
2246 		/*
2247 		 * If the buffer has been evicted, migrate it to a present state
2248 		 * before issuing the I/O.  Once we drop the hash-table lock,
2249 		 * the header will be marked as I/O in progress and have an
2250 		 * attached buffer.  At this point, anybody who finds this
2251 		 * buffer ought to notice that it's legit but has a pending I/O.
2252 		 */
2253 
2254 		if (GHOST_STATE(hdr->b_state))
2255 			arc_access(hdr, hash_lock);
2256 		mutex_exit(hash_lock);
2257 
2258 		ASSERT3U(hdr->b_size, ==, size);
2259 		DTRACE_PROBE3(arc__miss, blkptr_t *, bp, uint64_t, size,
2260 		    zbookmark_t *, zb);
2261 		ARCSTAT_BUMP(arcstat_misses);
2262 		ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2263 		    demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2264 		    data, metadata, misses);
2265 
2266 		rzio = zio_read(pio, spa, bp, buf->b_data, size,
2267 		    arc_read_done, buf, priority, flags, zb);
2268 
2269 		if (*arc_flags & ARC_WAIT)
2270 			return (zio_wait(rzio));
2271 
2272 		ASSERT(*arc_flags & ARC_NOWAIT);
2273 		zio_nowait(rzio);
2274 	}
2275 	return (0);
2276 }
2277 
2278 /*
2279  * arc_read() variant to support pool traversal.  If the block is already
2280  * in the ARC, make a copy of it; otherwise, the caller will do the I/O.
2281  * The idea is that we don't want pool traversal filling up memory, but
2282  * if the ARC already has the data anyway, we shouldn't pay for the I/O.
2283  */
2284 int
2285 arc_tryread(spa_t *spa, blkptr_t *bp, void *data)
2286 {
2287 	arc_buf_hdr_t *hdr;
2288 	kmutex_t *hash_mtx;
2289 	int rc = 0;
2290 
2291 	hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_mtx);
2292 
2293 	if (hdr && hdr->b_datacnt > 0 && !HDR_IO_IN_PROGRESS(hdr)) {
2294 		arc_buf_t *buf = hdr->b_buf;
2295 
2296 		ASSERT(buf);
2297 		while (buf->b_data == NULL) {
2298 			buf = buf->b_next;
2299 			ASSERT(buf);
2300 		}
2301 		bcopy(buf->b_data, data, hdr->b_size);
2302 	} else {
2303 		rc = ENOENT;
2304 	}
2305 
2306 	if (hash_mtx)
2307 		mutex_exit(hash_mtx);
2308 
2309 	return (rc);
2310 }
2311 
2312 void
2313 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
2314 {
2315 	ASSERT(buf->b_hdr != NULL);
2316 	ASSERT(buf->b_hdr->b_state != arc_anon);
2317 	ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
2318 	buf->b_efunc = func;
2319 	buf->b_private = private;
2320 }
2321 
2322 /*
2323  * This is used by the DMU to let the ARC know that a buffer is
2324  * being evicted, so the ARC should clean up.  If this arc buf
2325  * is not yet in the evicted state, it will be put there.
2326  */
2327 int
2328 arc_buf_evict(arc_buf_t *buf)
2329 {
2330 	arc_buf_hdr_t *hdr;
2331 	kmutex_t *hash_lock;
2332 	arc_buf_t **bufp;
2333 
2334 	mutex_enter(&arc_eviction_mtx);
2335 	hdr = buf->b_hdr;
2336 	if (hdr == NULL) {
2337 		/*
2338 		 * We are in arc_do_user_evicts().
2339 		 */
2340 		ASSERT(buf->b_data == NULL);
2341 		mutex_exit(&arc_eviction_mtx);
2342 		return (0);
2343 	}
2344 	hash_lock = HDR_LOCK(hdr);
2345 	mutex_exit(&arc_eviction_mtx);
2346 
2347 	mutex_enter(hash_lock);
2348 
2349 	if (buf->b_data == NULL) {
2350 		/*
2351 		 * We are on the eviction list.
2352 		 */
2353 		mutex_exit(hash_lock);
2354 		mutex_enter(&arc_eviction_mtx);
2355 		if (buf->b_hdr == NULL) {
2356 			/*
2357 			 * We are already in arc_do_user_evicts().
2358 			 */
2359 			mutex_exit(&arc_eviction_mtx);
2360 			return (0);
2361 		} else {
2362 			arc_buf_t copy = *buf; /* structure assignment */
2363 			/*
2364 			 * Process this buffer now
2365 			 * but let arc_do_user_evicts() do the reaping.
2366 			 */
2367 			buf->b_efunc = NULL;
2368 			mutex_exit(&arc_eviction_mtx);
2369 			VERIFY(copy.b_efunc(&copy) == 0);
2370 			return (1);
2371 		}
2372 	}
2373 
2374 	ASSERT(buf->b_hdr == hdr);
2375 	ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
2376 	ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2377 
2378 	/*
2379 	 * Pull this buffer off of the hdr
2380 	 */
2381 	bufp = &hdr->b_buf;
2382 	while (*bufp != buf)
2383 		bufp = &(*bufp)->b_next;
2384 	*bufp = buf->b_next;
2385 
2386 	ASSERT(buf->b_data != NULL);
2387 	arc_buf_destroy(buf, FALSE, FALSE);
2388 
2389 	if (hdr->b_datacnt == 0) {
2390 		arc_state_t *old_state = hdr->b_state;
2391 		arc_state_t *evicted_state;
2392 
2393 		ASSERT(refcount_is_zero(&hdr->b_refcnt));
2394 
2395 		evicted_state =
2396 		    (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
2397 
2398 		mutex_enter(&old_state->arcs_mtx);
2399 		mutex_enter(&evicted_state->arcs_mtx);
2400 
2401 		arc_change_state(evicted_state, hdr, hash_lock);
2402 		ASSERT(HDR_IN_HASH_TABLE(hdr));
2403 		hdr->b_flags = ARC_IN_HASH_TABLE;
2404 
2405 		mutex_exit(&evicted_state->arcs_mtx);
2406 		mutex_exit(&old_state->arcs_mtx);
2407 	}
2408 	mutex_exit(hash_lock);
2409 
2410 	VERIFY(buf->b_efunc(buf) == 0);
2411 	buf->b_efunc = NULL;
2412 	buf->b_private = NULL;
2413 	buf->b_hdr = NULL;
2414 	kmem_cache_free(buf_cache, buf);
2415 	return (1);
2416 }
2417 
2418 /*
2419  * Release this buffer from the cache.  This must be done
2420  * after a read and prior to modifying the buffer contents.
2421  * If the buffer has more than one reference, we must make
2422  * make a new hdr for the buffer.
2423  */
2424 void
2425 arc_release(arc_buf_t *buf, void *tag)
2426 {
2427 	arc_buf_hdr_t *hdr = buf->b_hdr;
2428 	kmutex_t *hash_lock = HDR_LOCK(hdr);
2429 
2430 	/* this buffer is not on any list */
2431 	ASSERT(refcount_count(&hdr->b_refcnt) > 0);
2432 
2433 	if (hdr->b_state == arc_anon) {
2434 		/* this buffer is already released */
2435 		ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 1);
2436 		ASSERT(BUF_EMPTY(hdr));
2437 		ASSERT(buf->b_efunc == NULL);
2438 		arc_buf_thaw(buf);
2439 		return;
2440 	}
2441 
2442 	mutex_enter(hash_lock);
2443 
2444 	/*
2445 	 * Do we have more than one buf?
2446 	 */
2447 	if (hdr->b_buf != buf || buf->b_next != NULL) {
2448 		arc_buf_hdr_t *nhdr;
2449 		arc_buf_t **bufp;
2450 		uint64_t blksz = hdr->b_size;
2451 		spa_t *spa = hdr->b_spa;
2452 		arc_buf_contents_t type = hdr->b_type;
2453 
2454 		ASSERT(hdr->b_datacnt > 1);
2455 		/*
2456 		 * Pull the data off of this buf and attach it to
2457 		 * a new anonymous buf.
2458 		 */
2459 		(void) remove_reference(hdr, hash_lock, tag);
2460 		bufp = &hdr->b_buf;
2461 		while (*bufp != buf)
2462 			bufp = &(*bufp)->b_next;
2463 		*bufp = (*bufp)->b_next;
2464 		buf->b_next = NULL;
2465 
2466 		ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
2467 		atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
2468 		if (refcount_is_zero(&hdr->b_refcnt)) {
2469 			uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
2470 			ASSERT3U(*size, >=, hdr->b_size);
2471 			atomic_add_64(size, -hdr->b_size);
2472 		}
2473 		hdr->b_datacnt -= 1;
2474 		arc_cksum_verify(buf);
2475 
2476 		mutex_exit(hash_lock);
2477 
2478 		nhdr = kmem_cache_alloc(hdr_cache, KM_SLEEP);
2479 		nhdr->b_size = blksz;
2480 		nhdr->b_spa = spa;
2481 		nhdr->b_type = type;
2482 		nhdr->b_buf = buf;
2483 		nhdr->b_state = arc_anon;
2484 		nhdr->b_arc_access = 0;
2485 		nhdr->b_flags = 0;
2486 		nhdr->b_datacnt = 1;
2487 		nhdr->b_freeze_cksum = NULL;
2488 		(void) refcount_add(&nhdr->b_refcnt, tag);
2489 		buf->b_hdr = nhdr;
2490 		atomic_add_64(&arc_anon->arcs_size, blksz);
2491 
2492 		hdr = nhdr;
2493 	} else {
2494 		ASSERT(refcount_count(&hdr->b_refcnt) == 1);
2495 		ASSERT(!list_link_active(&hdr->b_arc_node));
2496 		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2497 		arc_change_state(arc_anon, hdr, hash_lock);
2498 		hdr->b_arc_access = 0;
2499 		mutex_exit(hash_lock);
2500 		bzero(&hdr->b_dva, sizeof (dva_t));
2501 		hdr->b_birth = 0;
2502 		hdr->b_cksum0 = 0;
2503 		arc_buf_thaw(buf);
2504 	}
2505 	buf->b_efunc = NULL;
2506 	buf->b_private = NULL;
2507 }
2508 
2509 int
2510 arc_released(arc_buf_t *buf)
2511 {
2512 	return (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
2513 }
2514 
2515 int
2516 arc_has_callback(arc_buf_t *buf)
2517 {
2518 	return (buf->b_efunc != NULL);
2519 }
2520 
2521 #ifdef ZFS_DEBUG
2522 int
2523 arc_referenced(arc_buf_t *buf)
2524 {
2525 	return (refcount_count(&buf->b_hdr->b_refcnt));
2526 }
2527 #endif
2528 
2529 static void
2530 arc_write_ready(zio_t *zio)
2531 {
2532 	arc_write_callback_t *callback = zio->io_private;
2533 	arc_buf_t *buf = callback->awcb_buf;
2534 
2535 	if (callback->awcb_ready) {
2536 		ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
2537 		callback->awcb_ready(zio, buf, callback->awcb_private);
2538 	}
2539 	arc_cksum_compute(buf);
2540 }
2541 
2542 static void
2543 arc_write_done(zio_t *zio)
2544 {
2545 	arc_write_callback_t *callback = zio->io_private;
2546 	arc_buf_t *buf = callback->awcb_buf;
2547 	arc_buf_hdr_t *hdr = buf->b_hdr;
2548 
2549 	hdr->b_acb = NULL;
2550 
2551 	/* this buffer is on no lists and is not in the hash table */
2552 	ASSERT3P(hdr->b_state, ==, arc_anon);
2553 
2554 	hdr->b_dva = *BP_IDENTITY(zio->io_bp);
2555 	hdr->b_birth = zio->io_bp->blk_birth;
2556 	hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
2557 	/*
2558 	 * If the block to be written was all-zero, we may have
2559 	 * compressed it away.  In this case no write was performed
2560 	 * so there will be no dva/birth-date/checksum.  The buffer
2561 	 * must therefor remain anonymous (and uncached).
2562 	 */
2563 	if (!BUF_EMPTY(hdr)) {
2564 		arc_buf_hdr_t *exists;
2565 		kmutex_t *hash_lock;
2566 
2567 		arc_cksum_verify(buf);
2568 
2569 		exists = buf_hash_insert(hdr, &hash_lock);
2570 		if (exists) {
2571 			/*
2572 			 * This can only happen if we overwrite for
2573 			 * sync-to-convergence, because we remove
2574 			 * buffers from the hash table when we arc_free().
2575 			 */
2576 			ASSERT(DVA_EQUAL(BP_IDENTITY(&zio->io_bp_orig),
2577 			    BP_IDENTITY(zio->io_bp)));
2578 			ASSERT3U(zio->io_bp_orig.blk_birth, ==,
2579 			    zio->io_bp->blk_birth);
2580 
2581 			ASSERT(refcount_is_zero(&exists->b_refcnt));
2582 			arc_change_state(arc_anon, exists, hash_lock);
2583 			mutex_exit(hash_lock);
2584 			arc_hdr_destroy(exists);
2585 			exists = buf_hash_insert(hdr, &hash_lock);
2586 			ASSERT3P(exists, ==, NULL);
2587 		}
2588 		hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2589 		arc_access(hdr, hash_lock);
2590 		mutex_exit(hash_lock);
2591 	} else if (callback->awcb_done == NULL) {
2592 		int destroy_hdr;
2593 		/*
2594 		 * This is an anonymous buffer with no user callback,
2595 		 * destroy it if there are no active references.
2596 		 */
2597 		mutex_enter(&arc_eviction_mtx);
2598 		destroy_hdr = refcount_is_zero(&hdr->b_refcnt);
2599 		hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2600 		mutex_exit(&arc_eviction_mtx);
2601 		if (destroy_hdr)
2602 			arc_hdr_destroy(hdr);
2603 	} else {
2604 		hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2605 	}
2606 
2607 	if (callback->awcb_done) {
2608 		ASSERT(!refcount_is_zero(&hdr->b_refcnt));
2609 		callback->awcb_done(zio, buf, callback->awcb_private);
2610 	}
2611 
2612 	kmem_free(callback, sizeof (arc_write_callback_t));
2613 }
2614 
2615 zio_t *
2616 arc_write(zio_t *pio, spa_t *spa, int checksum, int compress, int ncopies,
2617     uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
2618     arc_done_func_t *ready, arc_done_func_t *done, void *private, int priority,
2619     int flags, zbookmark_t *zb)
2620 {
2621 	arc_buf_hdr_t *hdr = buf->b_hdr;
2622 	arc_write_callback_t *callback;
2623 	zio_t	*zio;
2624 
2625 	/* this is a private buffer - no locking required */
2626 	ASSERT3P(hdr->b_state, ==, arc_anon);
2627 	ASSERT(BUF_EMPTY(hdr));
2628 	ASSERT(!HDR_IO_ERROR(hdr));
2629 	ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
2630 	ASSERT(hdr->b_acb == 0);
2631 	callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
2632 	callback->awcb_ready = ready;
2633 	callback->awcb_done = done;
2634 	callback->awcb_private = private;
2635 	callback->awcb_buf = buf;
2636 	hdr->b_flags |= ARC_IO_IN_PROGRESS;
2637 	zio = zio_write(pio, spa, checksum, compress, ncopies, txg, bp,
2638 	    buf->b_data, hdr->b_size, arc_write_ready, arc_write_done, callback,
2639 	    priority, flags, zb);
2640 
2641 	return (zio);
2642 }
2643 
2644 int
2645 arc_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
2646     zio_done_func_t *done, void *private, uint32_t arc_flags)
2647 {
2648 	arc_buf_hdr_t *ab;
2649 	kmutex_t *hash_lock;
2650 	zio_t	*zio;
2651 
2652 	/*
2653 	 * If this buffer is in the cache, release it, so it
2654 	 * can be re-used.
2655 	 */
2656 	ab = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock);
2657 	if (ab != NULL) {
2658 		/*
2659 		 * The checksum of blocks to free is not always
2660 		 * preserved (eg. on the deadlist).  However, if it is
2661 		 * nonzero, it should match what we have in the cache.
2662 		 */
2663 		ASSERT(bp->blk_cksum.zc_word[0] == 0 ||
2664 		    ab->b_cksum0 == bp->blk_cksum.zc_word[0]);
2665 		if (ab->b_state != arc_anon)
2666 			arc_change_state(arc_anon, ab, hash_lock);
2667 		if (HDR_IO_IN_PROGRESS(ab)) {
2668 			/*
2669 			 * This should only happen when we prefetch.
2670 			 */
2671 			ASSERT(ab->b_flags & ARC_PREFETCH);
2672 			ASSERT3U(ab->b_datacnt, ==, 1);
2673 			ab->b_flags |= ARC_FREED_IN_READ;
2674 			if (HDR_IN_HASH_TABLE(ab))
2675 				buf_hash_remove(ab);
2676 			ab->b_arc_access = 0;
2677 			bzero(&ab->b_dva, sizeof (dva_t));
2678 			ab->b_birth = 0;
2679 			ab->b_cksum0 = 0;
2680 			ab->b_buf->b_efunc = NULL;
2681 			ab->b_buf->b_private = NULL;
2682 			mutex_exit(hash_lock);
2683 		} else if (refcount_is_zero(&ab->b_refcnt)) {
2684 			mutex_exit(hash_lock);
2685 			arc_hdr_destroy(ab);
2686 			ARCSTAT_BUMP(arcstat_deleted);
2687 		} else {
2688 			/*
2689 			 * We still have an active reference on this
2690 			 * buffer.  This can happen, e.g., from
2691 			 * dbuf_unoverride().
2692 			 */
2693 			ASSERT(!HDR_IN_HASH_TABLE(ab));
2694 			ab->b_arc_access = 0;
2695 			bzero(&ab->b_dva, sizeof (dva_t));
2696 			ab->b_birth = 0;
2697 			ab->b_cksum0 = 0;
2698 			ab->b_buf->b_efunc = NULL;
2699 			ab->b_buf->b_private = NULL;
2700 			mutex_exit(hash_lock);
2701 		}
2702 	}
2703 
2704 	zio = zio_free(pio, spa, txg, bp, done, private);
2705 
2706 	if (arc_flags & ARC_WAIT)
2707 		return (zio_wait(zio));
2708 
2709 	ASSERT(arc_flags & ARC_NOWAIT);
2710 	zio_nowait(zio);
2711 
2712 	return (0);
2713 }
2714 
2715 void
2716 arc_tempreserve_clear(uint64_t tempreserve)
2717 {
2718 	atomic_add_64(&arc_tempreserve, -tempreserve);
2719 	ASSERT((int64_t)arc_tempreserve >= 0);
2720 }
2721 
2722 int
2723 arc_tempreserve_space(uint64_t tempreserve)
2724 {
2725 #ifdef ZFS_DEBUG
2726 	/*
2727 	 * Once in a while, fail for no reason.  Everything should cope.
2728 	 */
2729 	if (spa_get_random(10000) == 0) {
2730 		dprintf("forcing random failure\n");
2731 		return (ERESTART);
2732 	}
2733 #endif
2734 	if (tempreserve > arc_c/4 && !arc_no_grow)
2735 		arc_c = MIN(arc_c_max, tempreserve * 4);
2736 	if (tempreserve > arc_c)
2737 		return (ENOMEM);
2738 
2739 	/*
2740 	 * Throttle writes when the amount of dirty data in the cache
2741 	 * gets too large.  We try to keep the cache less than half full
2742 	 * of dirty blocks so that our sync times don't grow too large.
2743 	 * Note: if two requests come in concurrently, we might let them
2744 	 * both succeed, when one of them should fail.  Not a huge deal.
2745 	 *
2746 	 * XXX The limit should be adjusted dynamically to keep the time
2747 	 * to sync a dataset fixed (around 1-5 seconds?).
2748 	 */
2749 
2750 	if (tempreserve + arc_tempreserve + arc_anon->arcs_size > arc_c / 2 &&
2751 	    arc_tempreserve + arc_anon->arcs_size > arc_c / 4) {
2752 		dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
2753 		    "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
2754 		    arc_tempreserve>>10,
2755 		    arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
2756 		    arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
2757 		    tempreserve>>10, arc_c>>10);
2758 		return (ERESTART);
2759 	}
2760 	atomic_add_64(&arc_tempreserve, tempreserve);
2761 	return (0);
2762 }
2763 
2764 void
2765 arc_init(void)
2766 {
2767 	mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
2768 	cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
2769 
2770 	/* Convert seconds to clock ticks */
2771 	arc_min_prefetch_lifespan = 1 * hz;
2772 
2773 	/* Start out with 1/8 of all memory */
2774 	arc_c = physmem * PAGESIZE / 8;
2775 
2776 #ifdef _KERNEL
2777 	/*
2778 	 * On architectures where the physical memory can be larger
2779 	 * than the addressable space (intel in 32-bit mode), we may
2780 	 * need to limit the cache to 1/8 of VM size.
2781 	 */
2782 	arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
2783 #endif
2784 
2785 	/* set min cache to 1/32 of all memory, or 64MB, whichever is more */
2786 	arc_c_min = MAX(arc_c / 4, 64<<20);
2787 	/* set max to 3/4 of all memory, or all but 1GB, whichever is more */
2788 	if (arc_c * 8 >= 1<<30)
2789 		arc_c_max = (arc_c * 8) - (1<<30);
2790 	else
2791 		arc_c_max = arc_c_min;
2792 	arc_c_max = MAX(arc_c * 6, arc_c_max);
2793 
2794 	/*
2795 	 * Allow the tunables to override our calculations if they are
2796 	 * reasonable (ie. over 64MB)
2797 	 */
2798 	if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
2799 		arc_c_max = zfs_arc_max;
2800 	if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
2801 		arc_c_min = zfs_arc_min;
2802 
2803 	arc_c = arc_c_max;
2804 	arc_p = (arc_c >> 1);
2805 
2806 	/* limit meta-data to 1/4 of the arc capacity */
2807 	arc_meta_limit = arc_c_max / 4;
2808 
2809 	/* Allow the tunable to override if it is reasonable */
2810 	if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
2811 		arc_meta_limit = zfs_arc_meta_limit;
2812 
2813 	if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
2814 		arc_c_min = arc_meta_limit / 2;
2815 
2816 	/* if kmem_flags are set, lets try to use less memory */
2817 	if (kmem_debugging())
2818 		arc_c = arc_c / 2;
2819 	if (arc_c < arc_c_min)
2820 		arc_c = arc_c_min;
2821 
2822 	arc_anon = &ARC_anon;
2823 	arc_mru = &ARC_mru;
2824 	arc_mru_ghost = &ARC_mru_ghost;
2825 	arc_mfu = &ARC_mfu;
2826 	arc_mfu_ghost = &ARC_mfu_ghost;
2827 	arc_size = 0;
2828 
2829 	mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
2830 	mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
2831 	mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
2832 	mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
2833 	mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
2834 
2835 	list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
2836 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2837 	list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
2838 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2839 	list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
2840 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2841 	list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
2842 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2843 	list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
2844 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2845 	list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
2846 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2847 	list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
2848 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2849 	list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
2850 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2851 
2852 	buf_init();
2853 
2854 	arc_thread_exit = 0;
2855 	arc_eviction_list = NULL;
2856 	mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
2857 	bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
2858 
2859 	arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
2860 	    sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
2861 
2862 	if (arc_ksp != NULL) {
2863 		arc_ksp->ks_data = &arc_stats;
2864 		kstat_install(arc_ksp);
2865 	}
2866 
2867 	(void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
2868 	    TS_RUN, minclsyspri);
2869 
2870 	arc_dead = FALSE;
2871 }
2872 
2873 void
2874 arc_fini(void)
2875 {
2876 	mutex_enter(&arc_reclaim_thr_lock);
2877 	arc_thread_exit = 1;
2878 	while (arc_thread_exit != 0)
2879 		cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
2880 	mutex_exit(&arc_reclaim_thr_lock);
2881 
2882 	arc_flush();
2883 
2884 	arc_dead = TRUE;
2885 
2886 	if (arc_ksp != NULL) {
2887 		kstat_delete(arc_ksp);
2888 		arc_ksp = NULL;
2889 	}
2890 
2891 	mutex_destroy(&arc_eviction_mtx);
2892 	mutex_destroy(&arc_reclaim_thr_lock);
2893 	cv_destroy(&arc_reclaim_thr_cv);
2894 
2895 	list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
2896 	list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
2897 	list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
2898 	list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
2899 	list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
2900 	list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
2901 	list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
2902 	list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
2903 
2904 	mutex_destroy(&arc_anon->arcs_mtx);
2905 	mutex_destroy(&arc_mru->arcs_mtx);
2906 	mutex_destroy(&arc_mru_ghost->arcs_mtx);
2907 	mutex_destroy(&arc_mfu->arcs_mtx);
2908 	mutex_destroy(&arc_mfu_ghost->arcs_mtx);
2909 
2910 	buf_fini();
2911 }
2912