xref: /titanic_41/usr/src/uts/common/fs/zfs/arc.c (revision e5ba14ff435beeefdaa2e6649e175c74afe02c76)
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 	mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
588 	return (0);
589 }
590 
591 /*
592  * Destructor callback - called when a cached buf is
593  * no longer required.
594  */
595 /* ARGSUSED */
596 static void
597 hdr_dest(void *vbuf, void *unused)
598 {
599 	arc_buf_hdr_t *buf = vbuf;
600 
601 	refcount_destroy(&buf->b_refcnt);
602 	cv_destroy(&buf->b_cv);
603 	mutex_destroy(&buf->b_freeze_lock);
604 }
605 
606 /*
607  * Reclaim callback -- invoked when memory is low.
608  */
609 /* ARGSUSED */
610 static void
611 hdr_recl(void *unused)
612 {
613 	dprintf("hdr_recl called\n");
614 	/*
615 	 * umem calls the reclaim func when we destroy the buf cache,
616 	 * which is after we do arc_fini().
617 	 */
618 	if (!arc_dead)
619 		cv_signal(&arc_reclaim_thr_cv);
620 }
621 
622 static void
623 buf_init(void)
624 {
625 	uint64_t *ct;
626 	uint64_t hsize = 1ULL << 12;
627 	int i, j;
628 
629 	/*
630 	 * The hash table is big enough to fill all of physical memory
631 	 * with an average 64K block size.  The table will take up
632 	 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
633 	 */
634 	while (hsize * 65536 < physmem * PAGESIZE)
635 		hsize <<= 1;
636 retry:
637 	buf_hash_table.ht_mask = hsize - 1;
638 	buf_hash_table.ht_table =
639 	    kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
640 	if (buf_hash_table.ht_table == NULL) {
641 		ASSERT(hsize > (1ULL << 8));
642 		hsize >>= 1;
643 		goto retry;
644 	}
645 
646 	hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
647 	    0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
648 	buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
649 	    0, NULL, NULL, NULL, NULL, NULL, 0);
650 
651 	for (i = 0; i < 256; i++)
652 		for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
653 			*ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
654 
655 	for (i = 0; i < BUF_LOCKS; i++) {
656 		mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
657 		    NULL, MUTEX_DEFAULT, NULL);
658 	}
659 }
660 
661 #define	ARC_MINTIME	(hz>>4) /* 62 ms */
662 
663 static void
664 arc_cksum_verify(arc_buf_t *buf)
665 {
666 	zio_cksum_t zc;
667 
668 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
669 		return;
670 
671 	mutex_enter(&buf->b_hdr->b_freeze_lock);
672 	if (buf->b_hdr->b_freeze_cksum == NULL ||
673 	    (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
674 		mutex_exit(&buf->b_hdr->b_freeze_lock);
675 		return;
676 	}
677 	fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
678 	if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
679 		panic("buffer modified while frozen!");
680 	mutex_exit(&buf->b_hdr->b_freeze_lock);
681 }
682 
683 static void
684 arc_cksum_compute(arc_buf_t *buf)
685 {
686 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
687 		return;
688 
689 	mutex_enter(&buf->b_hdr->b_freeze_lock);
690 	if (buf->b_hdr->b_freeze_cksum != NULL) {
691 		mutex_exit(&buf->b_hdr->b_freeze_lock);
692 		return;
693 	}
694 	buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
695 	fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
696 	    buf->b_hdr->b_freeze_cksum);
697 	mutex_exit(&buf->b_hdr->b_freeze_lock);
698 }
699 
700 void
701 arc_buf_thaw(arc_buf_t *buf)
702 {
703 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
704 		return;
705 
706 	if (buf->b_hdr->b_state != arc_anon)
707 		panic("modifying non-anon buffer!");
708 	if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
709 		panic("modifying buffer while i/o in progress!");
710 	arc_cksum_verify(buf);
711 	mutex_enter(&buf->b_hdr->b_freeze_lock);
712 	if (buf->b_hdr->b_freeze_cksum != NULL) {
713 		kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
714 		buf->b_hdr->b_freeze_cksum = NULL;
715 	}
716 	mutex_exit(&buf->b_hdr->b_freeze_lock);
717 }
718 
719 void
720 arc_buf_freeze(arc_buf_t *buf)
721 {
722 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
723 		return;
724 
725 	ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
726 	    buf->b_hdr->b_state == arc_anon);
727 	arc_cksum_compute(buf);
728 }
729 
730 static void
731 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
732 {
733 	ASSERT(MUTEX_HELD(hash_lock));
734 
735 	if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
736 	    (ab->b_state != arc_anon)) {
737 		uint64_t delta = ab->b_size * ab->b_datacnt;
738 		list_t *list = &ab->b_state->arcs_list[ab->b_type];
739 		uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
740 
741 		ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
742 		mutex_enter(&ab->b_state->arcs_mtx);
743 		ASSERT(list_link_active(&ab->b_arc_node));
744 		list_remove(list, ab);
745 		if (GHOST_STATE(ab->b_state)) {
746 			ASSERT3U(ab->b_datacnt, ==, 0);
747 			ASSERT3P(ab->b_buf, ==, NULL);
748 			delta = ab->b_size;
749 		}
750 		ASSERT(delta > 0);
751 		ASSERT3U(*size, >=, delta);
752 		atomic_add_64(size, -delta);
753 		mutex_exit(&ab->b_state->arcs_mtx);
754 		/* remove the prefetch flag is we get a reference */
755 		if (ab->b_flags & ARC_PREFETCH)
756 			ab->b_flags &= ~ARC_PREFETCH;
757 	}
758 }
759 
760 static int
761 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
762 {
763 	int cnt;
764 	arc_state_t *state = ab->b_state;
765 
766 	ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
767 	ASSERT(!GHOST_STATE(state));
768 
769 	if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
770 	    (state != arc_anon)) {
771 		uint64_t *size = &state->arcs_lsize[ab->b_type];
772 
773 		ASSERT(!MUTEX_HELD(&state->arcs_mtx));
774 		mutex_enter(&state->arcs_mtx);
775 		ASSERT(!list_link_active(&ab->b_arc_node));
776 		list_insert_head(&state->arcs_list[ab->b_type], ab);
777 		ASSERT(ab->b_datacnt > 0);
778 		atomic_add_64(size, ab->b_size * ab->b_datacnt);
779 		mutex_exit(&state->arcs_mtx);
780 	}
781 	return (cnt);
782 }
783 
784 /*
785  * Move the supplied buffer to the indicated state.  The mutex
786  * for the buffer must be held by the caller.
787  */
788 static void
789 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
790 {
791 	arc_state_t *old_state = ab->b_state;
792 	int64_t refcnt = refcount_count(&ab->b_refcnt);
793 	uint64_t from_delta, to_delta;
794 
795 	ASSERT(MUTEX_HELD(hash_lock));
796 	ASSERT(new_state != old_state);
797 	ASSERT(refcnt == 0 || ab->b_datacnt > 0);
798 	ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
799 
800 	from_delta = to_delta = ab->b_datacnt * ab->b_size;
801 
802 	/*
803 	 * If this buffer is evictable, transfer it from the
804 	 * old state list to the new state list.
805 	 */
806 	if (refcnt == 0) {
807 		if (old_state != arc_anon) {
808 			int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
809 			uint64_t *size = &old_state->arcs_lsize[ab->b_type];
810 
811 			if (use_mutex)
812 				mutex_enter(&old_state->arcs_mtx);
813 
814 			ASSERT(list_link_active(&ab->b_arc_node));
815 			list_remove(&old_state->arcs_list[ab->b_type], ab);
816 
817 			/*
818 			 * If prefetching out of the ghost cache,
819 			 * we will have a non-null datacnt.
820 			 */
821 			if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
822 				/* ghost elements have a ghost size */
823 				ASSERT(ab->b_buf == NULL);
824 				from_delta = ab->b_size;
825 			}
826 			ASSERT3U(*size, >=, from_delta);
827 			atomic_add_64(size, -from_delta);
828 
829 			if (use_mutex)
830 				mutex_exit(&old_state->arcs_mtx);
831 		}
832 		if (new_state != arc_anon) {
833 			int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
834 			uint64_t *size = &new_state->arcs_lsize[ab->b_type];
835 
836 			if (use_mutex)
837 				mutex_enter(&new_state->arcs_mtx);
838 
839 			list_insert_head(&new_state->arcs_list[ab->b_type], ab);
840 
841 			/* ghost elements have a ghost size */
842 			if (GHOST_STATE(new_state)) {
843 				ASSERT(ab->b_datacnt == 0);
844 				ASSERT(ab->b_buf == NULL);
845 				to_delta = ab->b_size;
846 			}
847 			atomic_add_64(size, to_delta);
848 
849 			if (use_mutex)
850 				mutex_exit(&new_state->arcs_mtx);
851 		}
852 	}
853 
854 	ASSERT(!BUF_EMPTY(ab));
855 	if (new_state == arc_anon && old_state != arc_anon) {
856 		buf_hash_remove(ab);
857 	}
858 
859 	/* adjust state sizes */
860 	if (to_delta)
861 		atomic_add_64(&new_state->arcs_size, to_delta);
862 	if (from_delta) {
863 		ASSERT3U(old_state->arcs_size, >=, from_delta);
864 		atomic_add_64(&old_state->arcs_size, -from_delta);
865 	}
866 	ab->b_state = new_state;
867 }
868 
869 void
870 arc_space_consume(uint64_t space)
871 {
872 	atomic_add_64(&arc_meta_used, space);
873 	atomic_add_64(&arc_size, space);
874 }
875 
876 void
877 arc_space_return(uint64_t space)
878 {
879 	ASSERT(arc_meta_used >= space);
880 	if (arc_meta_max < arc_meta_used)
881 		arc_meta_max = arc_meta_used;
882 	atomic_add_64(&arc_meta_used, -space);
883 	ASSERT(arc_size >= space);
884 	atomic_add_64(&arc_size, -space);
885 }
886 
887 void *
888 arc_data_buf_alloc(uint64_t size)
889 {
890 	if (arc_evict_needed(ARC_BUFC_DATA))
891 		cv_signal(&arc_reclaim_thr_cv);
892 	atomic_add_64(&arc_size, size);
893 	return (zio_data_buf_alloc(size));
894 }
895 
896 void
897 arc_data_buf_free(void *buf, uint64_t size)
898 {
899 	zio_data_buf_free(buf, size);
900 	ASSERT(arc_size >= size);
901 	atomic_add_64(&arc_size, -size);
902 }
903 
904 arc_buf_t *
905 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
906 {
907 	arc_buf_hdr_t *hdr;
908 	arc_buf_t *buf;
909 
910 	ASSERT3U(size, >, 0);
911 	hdr = kmem_cache_alloc(hdr_cache, KM_SLEEP);
912 	ASSERT(BUF_EMPTY(hdr));
913 	hdr->b_size = size;
914 	hdr->b_type = type;
915 	hdr->b_spa = spa;
916 	hdr->b_state = arc_anon;
917 	hdr->b_arc_access = 0;
918 	buf = kmem_cache_alloc(buf_cache, KM_SLEEP);
919 	buf->b_hdr = hdr;
920 	buf->b_data = NULL;
921 	buf->b_efunc = NULL;
922 	buf->b_private = NULL;
923 	buf->b_next = NULL;
924 	hdr->b_buf = buf;
925 	arc_get_data_buf(buf);
926 	hdr->b_datacnt = 1;
927 	hdr->b_flags = 0;
928 	ASSERT(refcount_is_zero(&hdr->b_refcnt));
929 	(void) refcount_add(&hdr->b_refcnt, tag);
930 
931 	return (buf);
932 }
933 
934 static arc_buf_t *
935 arc_buf_clone(arc_buf_t *from)
936 {
937 	arc_buf_t *buf;
938 	arc_buf_hdr_t *hdr = from->b_hdr;
939 	uint64_t size = hdr->b_size;
940 
941 	buf = kmem_cache_alloc(buf_cache, KM_SLEEP);
942 	buf->b_hdr = hdr;
943 	buf->b_data = NULL;
944 	buf->b_efunc = NULL;
945 	buf->b_private = NULL;
946 	buf->b_next = hdr->b_buf;
947 	hdr->b_buf = buf;
948 	arc_get_data_buf(buf);
949 	bcopy(from->b_data, buf->b_data, size);
950 	hdr->b_datacnt += 1;
951 	return (buf);
952 }
953 
954 void
955 arc_buf_add_ref(arc_buf_t *buf, void* tag)
956 {
957 	arc_buf_hdr_t *hdr;
958 	kmutex_t *hash_lock;
959 
960 	/*
961 	 * Check to see if this buffer is currently being evicted via
962 	 * arc_do_user_evicts().
963 	 */
964 	mutex_enter(&arc_eviction_mtx);
965 	hdr = buf->b_hdr;
966 	if (hdr == NULL) {
967 		mutex_exit(&arc_eviction_mtx);
968 		return;
969 	}
970 	hash_lock = HDR_LOCK(hdr);
971 	mutex_exit(&arc_eviction_mtx);
972 
973 	mutex_enter(hash_lock);
974 	if (buf->b_data == NULL) {
975 		/*
976 		 * This buffer is evicted.
977 		 */
978 		mutex_exit(hash_lock);
979 		return;
980 	}
981 
982 	ASSERT(buf->b_hdr == hdr);
983 	ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
984 	add_reference(hdr, hash_lock, tag);
985 	arc_access(hdr, hash_lock);
986 	mutex_exit(hash_lock);
987 	ARCSTAT_BUMP(arcstat_hits);
988 	ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
989 	    demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
990 	    data, metadata, hits);
991 }
992 
993 static void
994 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
995 {
996 	arc_buf_t **bufp;
997 
998 	/* free up data associated with the buf */
999 	if (buf->b_data) {
1000 		arc_state_t *state = buf->b_hdr->b_state;
1001 		uint64_t size = buf->b_hdr->b_size;
1002 		arc_buf_contents_t type = buf->b_hdr->b_type;
1003 
1004 		arc_cksum_verify(buf);
1005 		if (!recycle) {
1006 			if (type == ARC_BUFC_METADATA) {
1007 				zio_buf_free(buf->b_data, size);
1008 				arc_space_return(size);
1009 			} else {
1010 				ASSERT(type == ARC_BUFC_DATA);
1011 				zio_data_buf_free(buf->b_data, size);
1012 				atomic_add_64(&arc_size, -size);
1013 			}
1014 		}
1015 		if (list_link_active(&buf->b_hdr->b_arc_node)) {
1016 			uint64_t *cnt = &state->arcs_lsize[type];
1017 
1018 			ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1019 			ASSERT(state != arc_anon);
1020 
1021 			ASSERT3U(*cnt, >=, size);
1022 			atomic_add_64(cnt, -size);
1023 		}
1024 		ASSERT3U(state->arcs_size, >=, size);
1025 		atomic_add_64(&state->arcs_size, -size);
1026 		buf->b_data = NULL;
1027 		ASSERT(buf->b_hdr->b_datacnt > 0);
1028 		buf->b_hdr->b_datacnt -= 1;
1029 	}
1030 
1031 	/* only remove the buf if requested */
1032 	if (!all)
1033 		return;
1034 
1035 	/* remove the buf from the hdr list */
1036 	for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1037 		continue;
1038 	*bufp = buf->b_next;
1039 
1040 	ASSERT(buf->b_efunc == NULL);
1041 
1042 	/* clean up the buf */
1043 	buf->b_hdr = NULL;
1044 	kmem_cache_free(buf_cache, buf);
1045 }
1046 
1047 static void
1048 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1049 {
1050 	ASSERT(refcount_is_zero(&hdr->b_refcnt));
1051 	ASSERT3P(hdr->b_state, ==, arc_anon);
1052 	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1053 
1054 	if (!BUF_EMPTY(hdr)) {
1055 		ASSERT(!HDR_IN_HASH_TABLE(hdr));
1056 		bzero(&hdr->b_dva, sizeof (dva_t));
1057 		hdr->b_birth = 0;
1058 		hdr->b_cksum0 = 0;
1059 	}
1060 	while (hdr->b_buf) {
1061 		arc_buf_t *buf = hdr->b_buf;
1062 
1063 		if (buf->b_efunc) {
1064 			mutex_enter(&arc_eviction_mtx);
1065 			ASSERT(buf->b_hdr != NULL);
1066 			arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1067 			hdr->b_buf = buf->b_next;
1068 			buf->b_hdr = &arc_eviction_hdr;
1069 			buf->b_next = arc_eviction_list;
1070 			arc_eviction_list = buf;
1071 			mutex_exit(&arc_eviction_mtx);
1072 		} else {
1073 			arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1074 		}
1075 	}
1076 	if (hdr->b_freeze_cksum != NULL) {
1077 		kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1078 		hdr->b_freeze_cksum = NULL;
1079 	}
1080 
1081 	ASSERT(!list_link_active(&hdr->b_arc_node));
1082 	ASSERT3P(hdr->b_hash_next, ==, NULL);
1083 	ASSERT3P(hdr->b_acb, ==, NULL);
1084 	kmem_cache_free(hdr_cache, hdr);
1085 }
1086 
1087 void
1088 arc_buf_free(arc_buf_t *buf, void *tag)
1089 {
1090 	arc_buf_hdr_t *hdr = buf->b_hdr;
1091 	int hashed = hdr->b_state != arc_anon;
1092 
1093 	ASSERT(buf->b_efunc == NULL);
1094 	ASSERT(buf->b_data != NULL);
1095 
1096 	if (hashed) {
1097 		kmutex_t *hash_lock = HDR_LOCK(hdr);
1098 
1099 		mutex_enter(hash_lock);
1100 		(void) remove_reference(hdr, hash_lock, tag);
1101 		if (hdr->b_datacnt > 1)
1102 			arc_buf_destroy(buf, FALSE, TRUE);
1103 		else
1104 			hdr->b_flags |= ARC_BUF_AVAILABLE;
1105 		mutex_exit(hash_lock);
1106 	} else if (HDR_IO_IN_PROGRESS(hdr)) {
1107 		int destroy_hdr;
1108 		/*
1109 		 * We are in the middle of an async write.  Don't destroy
1110 		 * this buffer unless the write completes before we finish
1111 		 * decrementing the reference count.
1112 		 */
1113 		mutex_enter(&arc_eviction_mtx);
1114 		(void) remove_reference(hdr, NULL, tag);
1115 		ASSERT(refcount_is_zero(&hdr->b_refcnt));
1116 		destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1117 		mutex_exit(&arc_eviction_mtx);
1118 		if (destroy_hdr)
1119 			arc_hdr_destroy(hdr);
1120 	} else {
1121 		if (remove_reference(hdr, NULL, tag) > 0) {
1122 			ASSERT(HDR_IO_ERROR(hdr));
1123 			arc_buf_destroy(buf, FALSE, TRUE);
1124 		} else {
1125 			arc_hdr_destroy(hdr);
1126 		}
1127 	}
1128 }
1129 
1130 int
1131 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1132 {
1133 	arc_buf_hdr_t *hdr = buf->b_hdr;
1134 	kmutex_t *hash_lock = HDR_LOCK(hdr);
1135 	int no_callback = (buf->b_efunc == NULL);
1136 
1137 	if (hdr->b_state == arc_anon) {
1138 		arc_buf_free(buf, tag);
1139 		return (no_callback);
1140 	}
1141 
1142 	mutex_enter(hash_lock);
1143 	ASSERT(hdr->b_state != arc_anon);
1144 	ASSERT(buf->b_data != NULL);
1145 
1146 	(void) remove_reference(hdr, hash_lock, tag);
1147 	if (hdr->b_datacnt > 1) {
1148 		if (no_callback)
1149 			arc_buf_destroy(buf, FALSE, TRUE);
1150 	} else if (no_callback) {
1151 		ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1152 		hdr->b_flags |= ARC_BUF_AVAILABLE;
1153 	}
1154 	ASSERT(no_callback || hdr->b_datacnt > 1 ||
1155 	    refcount_is_zero(&hdr->b_refcnt));
1156 	mutex_exit(hash_lock);
1157 	return (no_callback);
1158 }
1159 
1160 int
1161 arc_buf_size(arc_buf_t *buf)
1162 {
1163 	return (buf->b_hdr->b_size);
1164 }
1165 
1166 /*
1167  * Evict buffers from list until we've removed the specified number of
1168  * bytes.  Move the removed buffers to the appropriate evict state.
1169  * If the recycle flag is set, then attempt to "recycle" a buffer:
1170  * - look for a buffer to evict that is `bytes' long.
1171  * - return the data block from this buffer rather than freeing it.
1172  * This flag is used by callers that are trying to make space for a
1173  * new buffer in a full arc cache.
1174  */
1175 static void *
1176 arc_evict(arc_state_t *state, int64_t bytes, boolean_t recycle,
1177     arc_buf_contents_t type)
1178 {
1179 	arc_state_t *evicted_state;
1180 	uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1181 	arc_buf_hdr_t *ab, *ab_prev = NULL;
1182 	list_t *list = &state->arcs_list[type];
1183 	kmutex_t *hash_lock;
1184 	boolean_t have_lock;
1185 	void *stolen = NULL;
1186 
1187 	ASSERT(state == arc_mru || state == arc_mfu);
1188 
1189 	evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1190 
1191 	mutex_enter(&state->arcs_mtx);
1192 	mutex_enter(&evicted_state->arcs_mtx);
1193 
1194 	for (ab = list_tail(list); ab; ab = ab_prev) {
1195 		ab_prev = list_prev(list, ab);
1196 		/* prefetch buffers have a minimum lifespan */
1197 		if (HDR_IO_IN_PROGRESS(ab) ||
1198 		    (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1199 		    lbolt - ab->b_arc_access < arc_min_prefetch_lifespan)) {
1200 			skipped++;
1201 			continue;
1202 		}
1203 		/* "lookahead" for better eviction candidate */
1204 		if (recycle && ab->b_size != bytes &&
1205 		    ab_prev && ab_prev->b_size == bytes)
1206 			continue;
1207 		hash_lock = HDR_LOCK(ab);
1208 		have_lock = MUTEX_HELD(hash_lock);
1209 		if (have_lock || mutex_tryenter(hash_lock)) {
1210 			ASSERT3U(refcount_count(&ab->b_refcnt), ==, 0);
1211 			ASSERT(ab->b_datacnt > 0);
1212 			while (ab->b_buf) {
1213 				arc_buf_t *buf = ab->b_buf;
1214 				if (buf->b_data) {
1215 					bytes_evicted += ab->b_size;
1216 					if (recycle && ab->b_type == type &&
1217 					    ab->b_size == bytes) {
1218 						stolen = buf->b_data;
1219 						recycle = FALSE;
1220 					}
1221 				}
1222 				if (buf->b_efunc) {
1223 					mutex_enter(&arc_eviction_mtx);
1224 					arc_buf_destroy(buf,
1225 					    buf->b_data == stolen, FALSE);
1226 					ab->b_buf = buf->b_next;
1227 					buf->b_hdr = &arc_eviction_hdr;
1228 					buf->b_next = arc_eviction_list;
1229 					arc_eviction_list = buf;
1230 					mutex_exit(&arc_eviction_mtx);
1231 				} else {
1232 					arc_buf_destroy(buf,
1233 					    buf->b_data == stolen, TRUE);
1234 				}
1235 			}
1236 			ASSERT(ab->b_datacnt == 0);
1237 			arc_change_state(evicted_state, ab, hash_lock);
1238 			ASSERT(HDR_IN_HASH_TABLE(ab));
1239 			ab->b_flags = ARC_IN_HASH_TABLE;
1240 			DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1241 			if (!have_lock)
1242 				mutex_exit(hash_lock);
1243 			if (bytes >= 0 && bytes_evicted >= bytes)
1244 				break;
1245 		} else {
1246 			missed += 1;
1247 		}
1248 	}
1249 
1250 	mutex_exit(&evicted_state->arcs_mtx);
1251 	mutex_exit(&state->arcs_mtx);
1252 
1253 	if (bytes_evicted < bytes)
1254 		dprintf("only evicted %lld bytes from %x",
1255 		    (longlong_t)bytes_evicted, state);
1256 
1257 	if (skipped)
1258 		ARCSTAT_INCR(arcstat_evict_skip, skipped);
1259 
1260 	if (missed)
1261 		ARCSTAT_INCR(arcstat_mutex_miss, missed);
1262 
1263 	/*
1264 	 * We have just evicted some date into the ghost state, make
1265 	 * sure we also adjust the ghost state size if necessary.
1266 	 */
1267 	if (arc_no_grow &&
1268 	    arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1269 		int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1270 		    arc_mru_ghost->arcs_size - arc_c;
1271 
1272 		if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1273 			int64_t todelete =
1274 			    MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1275 			arc_evict_ghost(arc_mru_ghost, todelete);
1276 		} else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1277 			int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1278 			    arc_mru_ghost->arcs_size +
1279 			    arc_mfu_ghost->arcs_size - arc_c);
1280 			arc_evict_ghost(arc_mfu_ghost, todelete);
1281 		}
1282 	}
1283 
1284 	return (stolen);
1285 }
1286 
1287 /*
1288  * Remove buffers from list until we've removed the specified number of
1289  * bytes.  Destroy the buffers that are removed.
1290  */
1291 static void
1292 arc_evict_ghost(arc_state_t *state, int64_t bytes)
1293 {
1294 	arc_buf_hdr_t *ab, *ab_prev;
1295 	list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1296 	kmutex_t *hash_lock;
1297 	uint64_t bytes_deleted = 0;
1298 	uint64_t bufs_skipped = 0;
1299 
1300 	ASSERT(GHOST_STATE(state));
1301 top:
1302 	mutex_enter(&state->arcs_mtx);
1303 	for (ab = list_tail(list); ab; ab = ab_prev) {
1304 		ab_prev = list_prev(list, ab);
1305 		hash_lock = HDR_LOCK(ab);
1306 		if (mutex_tryenter(hash_lock)) {
1307 			ASSERT(!HDR_IO_IN_PROGRESS(ab));
1308 			ASSERT(ab->b_buf == NULL);
1309 			arc_change_state(arc_anon, ab, hash_lock);
1310 			mutex_exit(hash_lock);
1311 			ARCSTAT_BUMP(arcstat_deleted);
1312 			bytes_deleted += ab->b_size;
1313 			arc_hdr_destroy(ab);
1314 			DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1315 			if (bytes >= 0 && bytes_deleted >= bytes)
1316 				break;
1317 		} else {
1318 			if (bytes < 0) {
1319 				mutex_exit(&state->arcs_mtx);
1320 				mutex_enter(hash_lock);
1321 				mutex_exit(hash_lock);
1322 				goto top;
1323 			}
1324 			bufs_skipped += 1;
1325 		}
1326 	}
1327 	mutex_exit(&state->arcs_mtx);
1328 
1329 	if (list == &state->arcs_list[ARC_BUFC_DATA] &&
1330 	    (bytes < 0 || bytes_deleted < bytes)) {
1331 		list = &state->arcs_list[ARC_BUFC_METADATA];
1332 		goto top;
1333 	}
1334 
1335 	if (bufs_skipped) {
1336 		ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1337 		ASSERT(bytes >= 0);
1338 	}
1339 
1340 	if (bytes_deleted < bytes)
1341 		dprintf("only deleted %lld bytes from %p",
1342 		    (longlong_t)bytes_deleted, state);
1343 }
1344 
1345 static void
1346 arc_adjust(void)
1347 {
1348 	int64_t top_sz, mru_over, arc_over, todelete;
1349 
1350 	top_sz = arc_anon->arcs_size + arc_mru->arcs_size;
1351 
1352 	if (top_sz > arc_p && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
1353 		int64_t toevict =
1354 		    MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], top_sz - arc_p);
1355 		(void) arc_evict(arc_mru, toevict, FALSE, ARC_BUFC_DATA);
1356 		top_sz = arc_anon->arcs_size + arc_mru->arcs_size;
1357 	}
1358 
1359 	if (top_sz > arc_p && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1360 		int64_t toevict =
1361 		    MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], top_sz - arc_p);
1362 		(void) arc_evict(arc_mru, toevict, FALSE, ARC_BUFC_METADATA);
1363 		top_sz = arc_anon->arcs_size + arc_mru->arcs_size;
1364 	}
1365 
1366 	mru_over = top_sz + arc_mru_ghost->arcs_size - arc_c;
1367 
1368 	if (mru_over > 0) {
1369 		if (arc_mru_ghost->arcs_size > 0) {
1370 			todelete = MIN(arc_mru_ghost->arcs_size, mru_over);
1371 			arc_evict_ghost(arc_mru_ghost, todelete);
1372 		}
1373 	}
1374 
1375 	if ((arc_over = arc_size - arc_c) > 0) {
1376 		int64_t tbl_over;
1377 
1378 		if (arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
1379 			int64_t toevict =
1380 			    MIN(arc_mfu->arcs_lsize[ARC_BUFC_DATA], arc_over);
1381 			(void) arc_evict(arc_mfu, toevict, FALSE,
1382 			    ARC_BUFC_DATA);
1383 			arc_over = arc_size - arc_c;
1384 		}
1385 
1386 		if (arc_over > 0 &&
1387 		    arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1388 			int64_t toevict =
1389 			    MIN(arc_mfu->arcs_lsize[ARC_BUFC_METADATA],
1390 			    arc_over);
1391 			(void) arc_evict(arc_mfu, toevict, FALSE,
1392 			    ARC_BUFC_METADATA);
1393 		}
1394 
1395 		tbl_over = arc_size + arc_mru_ghost->arcs_size +
1396 		    arc_mfu_ghost->arcs_size - arc_c * 2;
1397 
1398 		if (tbl_over > 0 && arc_mfu_ghost->arcs_size > 0) {
1399 			todelete = MIN(arc_mfu_ghost->arcs_size, tbl_over);
1400 			arc_evict_ghost(arc_mfu_ghost, todelete);
1401 		}
1402 	}
1403 }
1404 
1405 static void
1406 arc_do_user_evicts(void)
1407 {
1408 	mutex_enter(&arc_eviction_mtx);
1409 	while (arc_eviction_list != NULL) {
1410 		arc_buf_t *buf = arc_eviction_list;
1411 		arc_eviction_list = buf->b_next;
1412 		buf->b_hdr = NULL;
1413 		mutex_exit(&arc_eviction_mtx);
1414 
1415 		if (buf->b_efunc != NULL)
1416 			VERIFY(buf->b_efunc(buf) == 0);
1417 
1418 		buf->b_efunc = NULL;
1419 		buf->b_private = NULL;
1420 		kmem_cache_free(buf_cache, buf);
1421 		mutex_enter(&arc_eviction_mtx);
1422 	}
1423 	mutex_exit(&arc_eviction_mtx);
1424 }
1425 
1426 /*
1427  * Flush all *evictable* data from the cache.
1428  * NOTE: this will not touch "active" (i.e. referenced) data.
1429  */
1430 void
1431 arc_flush(void)
1432 {
1433 	while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA]))
1434 		(void) arc_evict(arc_mru, -1, FALSE, ARC_BUFC_DATA);
1435 	while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA]))
1436 		(void) arc_evict(arc_mru, -1, FALSE, ARC_BUFC_METADATA);
1437 	while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA]))
1438 		(void) arc_evict(arc_mfu, -1, FALSE, ARC_BUFC_DATA);
1439 	while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA]))
1440 		(void) arc_evict(arc_mfu, -1, FALSE, ARC_BUFC_METADATA);
1441 
1442 	arc_evict_ghost(arc_mru_ghost, -1);
1443 	arc_evict_ghost(arc_mfu_ghost, -1);
1444 
1445 	mutex_enter(&arc_reclaim_thr_lock);
1446 	arc_do_user_evicts();
1447 	mutex_exit(&arc_reclaim_thr_lock);
1448 	ASSERT(arc_eviction_list == NULL);
1449 }
1450 
1451 int arc_shrink_shift = 5;		/* log2(fraction of arc to reclaim) */
1452 
1453 void
1454 arc_shrink(void)
1455 {
1456 	if (arc_c > arc_c_min) {
1457 		uint64_t to_free;
1458 
1459 #ifdef _KERNEL
1460 		to_free = MAX(arc_c >> arc_shrink_shift, ptob(needfree));
1461 #else
1462 		to_free = arc_c >> arc_shrink_shift;
1463 #endif
1464 		if (arc_c > arc_c_min + to_free)
1465 			atomic_add_64(&arc_c, -to_free);
1466 		else
1467 			arc_c = arc_c_min;
1468 
1469 		atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
1470 		if (arc_c > arc_size)
1471 			arc_c = MAX(arc_size, arc_c_min);
1472 		if (arc_p > arc_c)
1473 			arc_p = (arc_c >> 1);
1474 		ASSERT(arc_c >= arc_c_min);
1475 		ASSERT((int64_t)arc_p >= 0);
1476 	}
1477 
1478 	if (arc_size > arc_c)
1479 		arc_adjust();
1480 }
1481 
1482 static int
1483 arc_reclaim_needed(void)
1484 {
1485 	uint64_t extra;
1486 
1487 #ifdef _KERNEL
1488 
1489 	if (needfree)
1490 		return (1);
1491 
1492 	/*
1493 	 * take 'desfree' extra pages, so we reclaim sooner, rather than later
1494 	 */
1495 	extra = desfree;
1496 
1497 	/*
1498 	 * check that we're out of range of the pageout scanner.  It starts to
1499 	 * schedule paging if freemem is less than lotsfree and needfree.
1500 	 * lotsfree is the high-water mark for pageout, and needfree is the
1501 	 * number of needed free pages.  We add extra pages here to make sure
1502 	 * the scanner doesn't start up while we're freeing memory.
1503 	 */
1504 	if (freemem < lotsfree + needfree + extra)
1505 		return (1);
1506 
1507 	/*
1508 	 * check to make sure that swapfs has enough space so that anon
1509 	 * reservations can still succeeed. anon_resvmem() checks that the
1510 	 * availrmem is greater than swapfs_minfree, and the number of reserved
1511 	 * swap pages.  We also add a bit of extra here just to prevent
1512 	 * circumstances from getting really dire.
1513 	 */
1514 	if (availrmem < swapfs_minfree + swapfs_reserve + extra)
1515 		return (1);
1516 
1517 #if defined(__i386)
1518 	/*
1519 	 * If we're on an i386 platform, it's possible that we'll exhaust the
1520 	 * kernel heap space before we ever run out of available physical
1521 	 * memory.  Most checks of the size of the heap_area compare against
1522 	 * tune.t_minarmem, which is the minimum available real memory that we
1523 	 * can have in the system.  However, this is generally fixed at 25 pages
1524 	 * which is so low that it's useless.  In this comparison, we seek to
1525 	 * calculate the total heap-size, and reclaim if more than 3/4ths of the
1526 	 * heap is allocated.  (Or, in the caclulation, if less than 1/4th is
1527 	 * free)
1528 	 */
1529 	if (btop(vmem_size(heap_arena, VMEM_FREE)) <
1530 	    (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2))
1531 		return (1);
1532 #endif
1533 
1534 #else
1535 	if (spa_get_random(100) == 0)
1536 		return (1);
1537 #endif
1538 	return (0);
1539 }
1540 
1541 static void
1542 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
1543 {
1544 	size_t			i;
1545 	kmem_cache_t		*prev_cache = NULL;
1546 	kmem_cache_t		*prev_data_cache = NULL;
1547 	extern kmem_cache_t	*zio_buf_cache[];
1548 	extern kmem_cache_t	*zio_data_buf_cache[];
1549 
1550 #ifdef _KERNEL
1551 	if (arc_meta_used >= arc_meta_limit) {
1552 		/*
1553 		 * We are exceeding our meta-data cache limit.
1554 		 * Purge some DNLC entries to release holds on meta-data.
1555 		 */
1556 		dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
1557 	}
1558 #if defined(__i386)
1559 	/*
1560 	 * Reclaim unused memory from all kmem caches.
1561 	 */
1562 	kmem_reap();
1563 #endif
1564 #endif
1565 
1566 	/*
1567 	 * An agressive reclamation will shrink the cache size as well as
1568 	 * reap free buffers from the arc kmem caches.
1569 	 */
1570 	if (strat == ARC_RECLAIM_AGGR)
1571 		arc_shrink();
1572 
1573 	for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
1574 		if (zio_buf_cache[i] != prev_cache) {
1575 			prev_cache = zio_buf_cache[i];
1576 			kmem_cache_reap_now(zio_buf_cache[i]);
1577 		}
1578 		if (zio_data_buf_cache[i] != prev_data_cache) {
1579 			prev_data_cache = zio_data_buf_cache[i];
1580 			kmem_cache_reap_now(zio_data_buf_cache[i]);
1581 		}
1582 	}
1583 	kmem_cache_reap_now(buf_cache);
1584 	kmem_cache_reap_now(hdr_cache);
1585 }
1586 
1587 static void
1588 arc_reclaim_thread(void)
1589 {
1590 	clock_t			growtime = 0;
1591 	arc_reclaim_strategy_t	last_reclaim = ARC_RECLAIM_CONS;
1592 	callb_cpr_t		cpr;
1593 
1594 	CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
1595 
1596 	mutex_enter(&arc_reclaim_thr_lock);
1597 	while (arc_thread_exit == 0) {
1598 		if (arc_reclaim_needed()) {
1599 
1600 			if (arc_no_grow) {
1601 				if (last_reclaim == ARC_RECLAIM_CONS) {
1602 					last_reclaim = ARC_RECLAIM_AGGR;
1603 				} else {
1604 					last_reclaim = ARC_RECLAIM_CONS;
1605 				}
1606 			} else {
1607 				arc_no_grow = TRUE;
1608 				last_reclaim = ARC_RECLAIM_AGGR;
1609 				membar_producer();
1610 			}
1611 
1612 			/* reset the growth delay for every reclaim */
1613 			growtime = lbolt + (arc_grow_retry * hz);
1614 
1615 			arc_kmem_reap_now(last_reclaim);
1616 
1617 		} else if (arc_no_grow && lbolt >= growtime) {
1618 			arc_no_grow = FALSE;
1619 		}
1620 
1621 		if (2 * arc_c < arc_size +
1622 		    arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size)
1623 			arc_adjust();
1624 
1625 		if (arc_eviction_list != NULL)
1626 			arc_do_user_evicts();
1627 
1628 		/* block until needed, or one second, whichever is shorter */
1629 		CALLB_CPR_SAFE_BEGIN(&cpr);
1630 		(void) cv_timedwait(&arc_reclaim_thr_cv,
1631 		    &arc_reclaim_thr_lock, (lbolt + hz));
1632 		CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
1633 	}
1634 
1635 	arc_thread_exit = 0;
1636 	cv_broadcast(&arc_reclaim_thr_cv);
1637 	CALLB_CPR_EXIT(&cpr);		/* drops arc_reclaim_thr_lock */
1638 	thread_exit();
1639 }
1640 
1641 /*
1642  * Adapt arc info given the number of bytes we are trying to add and
1643  * the state that we are comming from.  This function is only called
1644  * when we are adding new content to the cache.
1645  */
1646 static void
1647 arc_adapt(int bytes, arc_state_t *state)
1648 {
1649 	int mult;
1650 
1651 	ASSERT(bytes > 0);
1652 	/*
1653 	 * Adapt the target size of the MRU list:
1654 	 *	- if we just hit in the MRU ghost list, then increase
1655 	 *	  the target size of the MRU list.
1656 	 *	- if we just hit in the MFU ghost list, then increase
1657 	 *	  the target size of the MFU list by decreasing the
1658 	 *	  target size of the MRU list.
1659 	 */
1660 	if (state == arc_mru_ghost) {
1661 		mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
1662 		    1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
1663 
1664 		arc_p = MIN(arc_c, arc_p + bytes * mult);
1665 	} else if (state == arc_mfu_ghost) {
1666 		mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
1667 		    1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
1668 
1669 		arc_p = MAX(0, (int64_t)arc_p - bytes * mult);
1670 	}
1671 	ASSERT((int64_t)arc_p >= 0);
1672 
1673 	if (arc_reclaim_needed()) {
1674 		cv_signal(&arc_reclaim_thr_cv);
1675 		return;
1676 	}
1677 
1678 	if (arc_no_grow)
1679 		return;
1680 
1681 	if (arc_c >= arc_c_max)
1682 		return;
1683 
1684 	/*
1685 	 * If we're within (2 * maxblocksize) bytes of the target
1686 	 * cache size, increment the target cache size
1687 	 */
1688 	if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
1689 		atomic_add_64(&arc_c, (int64_t)bytes);
1690 		if (arc_c > arc_c_max)
1691 			arc_c = arc_c_max;
1692 		else if (state == arc_anon)
1693 			atomic_add_64(&arc_p, (int64_t)bytes);
1694 		if (arc_p > arc_c)
1695 			arc_p = arc_c;
1696 	}
1697 	ASSERT((int64_t)arc_p >= 0);
1698 }
1699 
1700 /*
1701  * Check if the cache has reached its limits and eviction is required
1702  * prior to insert.
1703  */
1704 static int
1705 arc_evict_needed(arc_buf_contents_t type)
1706 {
1707 	if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
1708 		return (1);
1709 
1710 #ifdef _KERNEL
1711 	/*
1712 	 * If zio data pages are being allocated out of a separate heap segment,
1713 	 * then enforce that the size of available vmem for this area remains
1714 	 * above about 1/32nd free.
1715 	 */
1716 	if (type == ARC_BUFC_DATA && zio_arena != NULL &&
1717 	    vmem_size(zio_arena, VMEM_FREE) <
1718 	    (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
1719 		return (1);
1720 #endif
1721 
1722 	if (arc_reclaim_needed())
1723 		return (1);
1724 
1725 	return (arc_size > arc_c);
1726 }
1727 
1728 /*
1729  * The buffer, supplied as the first argument, needs a data block.
1730  * So, if we are at cache max, determine which cache should be victimized.
1731  * We have the following cases:
1732  *
1733  * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
1734  * In this situation if we're out of space, but the resident size of the MFU is
1735  * under the limit, victimize the MFU cache to satisfy this insertion request.
1736  *
1737  * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
1738  * Here, we've used up all of the available space for the MRU, so we need to
1739  * evict from our own cache instead.  Evict from the set of resident MRU
1740  * entries.
1741  *
1742  * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
1743  * c minus p represents the MFU space in the cache, since p is the size of the
1744  * cache that is dedicated to the MRU.  In this situation there's still space on
1745  * the MFU side, so the MRU side needs to be victimized.
1746  *
1747  * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
1748  * MFU's resident set is consuming more space than it has been allotted.  In
1749  * this situation, we must victimize our own cache, the MFU, for this insertion.
1750  */
1751 static void
1752 arc_get_data_buf(arc_buf_t *buf)
1753 {
1754 	arc_state_t		*state = buf->b_hdr->b_state;
1755 	uint64_t		size = buf->b_hdr->b_size;
1756 	arc_buf_contents_t	type = buf->b_hdr->b_type;
1757 
1758 	arc_adapt(size, state);
1759 
1760 	/*
1761 	 * We have not yet reached cache maximum size,
1762 	 * just allocate a new buffer.
1763 	 */
1764 	if (!arc_evict_needed(type)) {
1765 		if (type == ARC_BUFC_METADATA) {
1766 			buf->b_data = zio_buf_alloc(size);
1767 			arc_space_consume(size);
1768 		} else {
1769 			ASSERT(type == ARC_BUFC_DATA);
1770 			buf->b_data = zio_data_buf_alloc(size);
1771 			atomic_add_64(&arc_size, size);
1772 		}
1773 		goto out;
1774 	}
1775 
1776 	/*
1777 	 * If we are prefetching from the mfu ghost list, this buffer
1778 	 * will end up on the mru list; so steal space from there.
1779 	 */
1780 	if (state == arc_mfu_ghost)
1781 		state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
1782 	else if (state == arc_mru_ghost)
1783 		state = arc_mru;
1784 
1785 	if (state == arc_mru || state == arc_anon) {
1786 		uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
1787 		state = (arc_mfu->arcs_lsize[type] > 0 &&
1788 		    arc_p > mru_used) ? arc_mfu : arc_mru;
1789 	} else {
1790 		/* MFU cases */
1791 		uint64_t mfu_space = arc_c - arc_p;
1792 		state =  (arc_mru->arcs_lsize[type] > 0 &&
1793 		    mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
1794 	}
1795 	if ((buf->b_data = arc_evict(state, size, TRUE, type)) == NULL) {
1796 		if (type == ARC_BUFC_METADATA) {
1797 			buf->b_data = zio_buf_alloc(size);
1798 			arc_space_consume(size);
1799 		} else {
1800 			ASSERT(type == ARC_BUFC_DATA);
1801 			buf->b_data = zio_data_buf_alloc(size);
1802 			atomic_add_64(&arc_size, size);
1803 		}
1804 		ARCSTAT_BUMP(arcstat_recycle_miss);
1805 	}
1806 	ASSERT(buf->b_data != NULL);
1807 out:
1808 	/*
1809 	 * Update the state size.  Note that ghost states have a
1810 	 * "ghost size" and so don't need to be updated.
1811 	 */
1812 	if (!GHOST_STATE(buf->b_hdr->b_state)) {
1813 		arc_buf_hdr_t *hdr = buf->b_hdr;
1814 
1815 		atomic_add_64(&hdr->b_state->arcs_size, size);
1816 		if (list_link_active(&hdr->b_arc_node)) {
1817 			ASSERT(refcount_is_zero(&hdr->b_refcnt));
1818 			atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
1819 		}
1820 		/*
1821 		 * If we are growing the cache, and we are adding anonymous
1822 		 * data, and we have outgrown arc_p, update arc_p
1823 		 */
1824 		if (arc_size < arc_c && hdr->b_state == arc_anon &&
1825 		    arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
1826 			arc_p = MIN(arc_c, arc_p + size);
1827 	}
1828 }
1829 
1830 /*
1831  * This routine is called whenever a buffer is accessed.
1832  * NOTE: the hash lock is dropped in this function.
1833  */
1834 static void
1835 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
1836 {
1837 	ASSERT(MUTEX_HELD(hash_lock));
1838 
1839 	if (buf->b_state == arc_anon) {
1840 		/*
1841 		 * This buffer is not in the cache, and does not
1842 		 * appear in our "ghost" list.  Add the new buffer
1843 		 * to the MRU state.
1844 		 */
1845 
1846 		ASSERT(buf->b_arc_access == 0);
1847 		buf->b_arc_access = lbolt;
1848 		DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
1849 		arc_change_state(arc_mru, buf, hash_lock);
1850 
1851 	} else if (buf->b_state == arc_mru) {
1852 		/*
1853 		 * If this buffer is here because of a prefetch, then either:
1854 		 * - clear the flag if this is a "referencing" read
1855 		 *   (any subsequent access will bump this into the MFU state).
1856 		 * or
1857 		 * - move the buffer to the head of the list if this is
1858 		 *   another prefetch (to make it less likely to be evicted).
1859 		 */
1860 		if ((buf->b_flags & ARC_PREFETCH) != 0) {
1861 			if (refcount_count(&buf->b_refcnt) == 0) {
1862 				ASSERT(list_link_active(&buf->b_arc_node));
1863 			} else {
1864 				buf->b_flags &= ~ARC_PREFETCH;
1865 				ARCSTAT_BUMP(arcstat_mru_hits);
1866 			}
1867 			buf->b_arc_access = lbolt;
1868 			return;
1869 		}
1870 
1871 		/*
1872 		 * This buffer has been "accessed" only once so far,
1873 		 * but it is still in the cache. Move it to the MFU
1874 		 * state.
1875 		 */
1876 		if (lbolt > buf->b_arc_access + ARC_MINTIME) {
1877 			/*
1878 			 * More than 125ms have passed since we
1879 			 * instantiated this buffer.  Move it to the
1880 			 * most frequently used state.
1881 			 */
1882 			buf->b_arc_access = lbolt;
1883 			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
1884 			arc_change_state(arc_mfu, buf, hash_lock);
1885 		}
1886 		ARCSTAT_BUMP(arcstat_mru_hits);
1887 	} else if (buf->b_state == arc_mru_ghost) {
1888 		arc_state_t	*new_state;
1889 		/*
1890 		 * This buffer has been "accessed" recently, but
1891 		 * was evicted from the cache.  Move it to the
1892 		 * MFU state.
1893 		 */
1894 
1895 		if (buf->b_flags & ARC_PREFETCH) {
1896 			new_state = arc_mru;
1897 			if (refcount_count(&buf->b_refcnt) > 0)
1898 				buf->b_flags &= ~ARC_PREFETCH;
1899 			DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
1900 		} else {
1901 			new_state = arc_mfu;
1902 			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
1903 		}
1904 
1905 		buf->b_arc_access = lbolt;
1906 		arc_change_state(new_state, buf, hash_lock);
1907 
1908 		ARCSTAT_BUMP(arcstat_mru_ghost_hits);
1909 	} else if (buf->b_state == arc_mfu) {
1910 		/*
1911 		 * This buffer has been accessed more than once and is
1912 		 * still in the cache.  Keep it in the MFU state.
1913 		 *
1914 		 * NOTE: an add_reference() that occurred when we did
1915 		 * the arc_read() will have kicked this off the list.
1916 		 * If it was a prefetch, we will explicitly move it to
1917 		 * the head of the list now.
1918 		 */
1919 		if ((buf->b_flags & ARC_PREFETCH) != 0) {
1920 			ASSERT(refcount_count(&buf->b_refcnt) == 0);
1921 			ASSERT(list_link_active(&buf->b_arc_node));
1922 		}
1923 		ARCSTAT_BUMP(arcstat_mfu_hits);
1924 		buf->b_arc_access = lbolt;
1925 	} else if (buf->b_state == arc_mfu_ghost) {
1926 		arc_state_t	*new_state = arc_mfu;
1927 		/*
1928 		 * This buffer has been accessed more than once but has
1929 		 * been evicted from the cache.  Move it back to the
1930 		 * MFU state.
1931 		 */
1932 
1933 		if (buf->b_flags & ARC_PREFETCH) {
1934 			/*
1935 			 * This is a prefetch access...
1936 			 * move this block back to the MRU state.
1937 			 */
1938 			ASSERT3U(refcount_count(&buf->b_refcnt), ==, 0);
1939 			new_state = arc_mru;
1940 		}
1941 
1942 		buf->b_arc_access = lbolt;
1943 		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
1944 		arc_change_state(new_state, buf, hash_lock);
1945 
1946 		ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
1947 	} else {
1948 		ASSERT(!"invalid arc state");
1949 	}
1950 }
1951 
1952 /* a generic arc_done_func_t which you can use */
1953 /* ARGSUSED */
1954 void
1955 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
1956 {
1957 	bcopy(buf->b_data, arg, buf->b_hdr->b_size);
1958 	VERIFY(arc_buf_remove_ref(buf, arg) == 1);
1959 }
1960 
1961 /* a generic arc_done_func_t */
1962 void
1963 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
1964 {
1965 	arc_buf_t **bufp = arg;
1966 	if (zio && zio->io_error) {
1967 		VERIFY(arc_buf_remove_ref(buf, arg) == 1);
1968 		*bufp = NULL;
1969 	} else {
1970 		*bufp = buf;
1971 	}
1972 }
1973 
1974 static void
1975 arc_read_done(zio_t *zio)
1976 {
1977 	arc_buf_hdr_t	*hdr, *found;
1978 	arc_buf_t	*buf;
1979 	arc_buf_t	*abuf;	/* buffer we're assigning to callback */
1980 	kmutex_t	*hash_lock;
1981 	arc_callback_t	*callback_list, *acb;
1982 	int		freeable = FALSE;
1983 
1984 	buf = zio->io_private;
1985 	hdr = buf->b_hdr;
1986 
1987 	/*
1988 	 * The hdr was inserted into hash-table and removed from lists
1989 	 * prior to starting I/O.  We should find this header, since
1990 	 * it's in the hash table, and it should be legit since it's
1991 	 * not possible to evict it during the I/O.  The only possible
1992 	 * reason for it not to be found is if we were freed during the
1993 	 * read.
1994 	 */
1995 	found = buf_hash_find(zio->io_spa, &hdr->b_dva, hdr->b_birth,
1996 	    &hash_lock);
1997 
1998 	ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
1999 	    (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))));
2000 
2001 	/* byteswap if necessary */
2002 	callback_list = hdr->b_acb;
2003 	ASSERT(callback_list != NULL);
2004 	if (BP_SHOULD_BYTESWAP(zio->io_bp) && callback_list->acb_byteswap)
2005 		callback_list->acb_byteswap(buf->b_data, hdr->b_size);
2006 
2007 	arc_cksum_compute(buf);
2008 
2009 	/* create copies of the data buffer for the callers */
2010 	abuf = buf;
2011 	for (acb = callback_list; acb; acb = acb->acb_next) {
2012 		if (acb->acb_done) {
2013 			if (abuf == NULL)
2014 				abuf = arc_buf_clone(buf);
2015 			acb->acb_buf = abuf;
2016 			abuf = NULL;
2017 		}
2018 	}
2019 	hdr->b_acb = NULL;
2020 	hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2021 	ASSERT(!HDR_BUF_AVAILABLE(hdr));
2022 	if (abuf == buf)
2023 		hdr->b_flags |= ARC_BUF_AVAILABLE;
2024 
2025 	ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2026 
2027 	if (zio->io_error != 0) {
2028 		hdr->b_flags |= ARC_IO_ERROR;
2029 		if (hdr->b_state != arc_anon)
2030 			arc_change_state(arc_anon, hdr, hash_lock);
2031 		if (HDR_IN_HASH_TABLE(hdr))
2032 			buf_hash_remove(hdr);
2033 		freeable = refcount_is_zero(&hdr->b_refcnt);
2034 		/* convert checksum errors into IO errors */
2035 		if (zio->io_error == ECKSUM)
2036 			zio->io_error = EIO;
2037 	}
2038 
2039 	/*
2040 	 * Broadcast before we drop the hash_lock to avoid the possibility
2041 	 * that the hdr (and hence the cv) might be freed before we get to
2042 	 * the cv_broadcast().
2043 	 */
2044 	cv_broadcast(&hdr->b_cv);
2045 
2046 	if (hash_lock) {
2047 		/*
2048 		 * Only call arc_access on anonymous buffers.  This is because
2049 		 * if we've issued an I/O for an evicted buffer, we've already
2050 		 * called arc_access (to prevent any simultaneous readers from
2051 		 * getting confused).
2052 		 */
2053 		if (zio->io_error == 0 && hdr->b_state == arc_anon)
2054 			arc_access(hdr, hash_lock);
2055 		mutex_exit(hash_lock);
2056 	} else {
2057 		/*
2058 		 * This block was freed while we waited for the read to
2059 		 * complete.  It has been removed from the hash table and
2060 		 * moved to the anonymous state (so that it won't show up
2061 		 * in the cache).
2062 		 */
2063 		ASSERT3P(hdr->b_state, ==, arc_anon);
2064 		freeable = refcount_is_zero(&hdr->b_refcnt);
2065 	}
2066 
2067 	/* execute each callback and free its structure */
2068 	while ((acb = callback_list) != NULL) {
2069 		if (acb->acb_done)
2070 			acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2071 
2072 		if (acb->acb_zio_dummy != NULL) {
2073 			acb->acb_zio_dummy->io_error = zio->io_error;
2074 			zio_nowait(acb->acb_zio_dummy);
2075 		}
2076 
2077 		callback_list = acb->acb_next;
2078 		kmem_free(acb, sizeof (arc_callback_t));
2079 	}
2080 
2081 	if (freeable)
2082 		arc_hdr_destroy(hdr);
2083 }
2084 
2085 /*
2086  * "Read" the block block at the specified DVA (in bp) via the
2087  * cache.  If the block is found in the cache, invoke the provided
2088  * callback immediately and return.  Note that the `zio' parameter
2089  * in the callback will be NULL in this case, since no IO was
2090  * required.  If the block is not in the cache pass the read request
2091  * on to the spa with a substitute callback function, so that the
2092  * requested block will be added to the cache.
2093  *
2094  * If a read request arrives for a block that has a read in-progress,
2095  * either wait for the in-progress read to complete (and return the
2096  * results); or, if this is a read with a "done" func, add a record
2097  * to the read to invoke the "done" func when the read completes,
2098  * and return; or just return.
2099  *
2100  * arc_read_done() will invoke all the requested "done" functions
2101  * for readers of this block.
2102  */
2103 int
2104 arc_read(zio_t *pio, spa_t *spa, blkptr_t *bp, arc_byteswap_func_t *swap,
2105     arc_done_func_t *done, void *private, int priority, int flags,
2106     uint32_t *arc_flags, zbookmark_t *zb)
2107 {
2108 	arc_buf_hdr_t *hdr;
2109 	arc_buf_t *buf;
2110 	kmutex_t *hash_lock;
2111 	zio_t	*rzio;
2112 
2113 top:
2114 	hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock);
2115 	if (hdr && hdr->b_datacnt > 0) {
2116 
2117 		*arc_flags |= ARC_CACHED;
2118 
2119 		if (HDR_IO_IN_PROGRESS(hdr)) {
2120 
2121 			if (*arc_flags & ARC_WAIT) {
2122 				cv_wait(&hdr->b_cv, hash_lock);
2123 				mutex_exit(hash_lock);
2124 				goto top;
2125 			}
2126 			ASSERT(*arc_flags & ARC_NOWAIT);
2127 
2128 			if (done) {
2129 				arc_callback_t	*acb = NULL;
2130 
2131 				acb = kmem_zalloc(sizeof (arc_callback_t),
2132 				    KM_SLEEP);
2133 				acb->acb_done = done;
2134 				acb->acb_private = private;
2135 				acb->acb_byteswap = swap;
2136 				if (pio != NULL)
2137 					acb->acb_zio_dummy = zio_null(pio,
2138 					    spa, NULL, NULL, flags);
2139 
2140 				ASSERT(acb->acb_done != NULL);
2141 				acb->acb_next = hdr->b_acb;
2142 				hdr->b_acb = acb;
2143 				add_reference(hdr, hash_lock, private);
2144 				mutex_exit(hash_lock);
2145 				return (0);
2146 			}
2147 			mutex_exit(hash_lock);
2148 			return (0);
2149 		}
2150 
2151 		ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2152 
2153 		if (done) {
2154 			add_reference(hdr, hash_lock, private);
2155 			/*
2156 			 * If this block is already in use, create a new
2157 			 * copy of the data so that we will be guaranteed
2158 			 * that arc_release() will always succeed.
2159 			 */
2160 			buf = hdr->b_buf;
2161 			ASSERT(buf);
2162 			ASSERT(buf->b_data);
2163 			if (HDR_BUF_AVAILABLE(hdr)) {
2164 				ASSERT(buf->b_efunc == NULL);
2165 				hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2166 			} else {
2167 				buf = arc_buf_clone(buf);
2168 			}
2169 		} else if (*arc_flags & ARC_PREFETCH &&
2170 		    refcount_count(&hdr->b_refcnt) == 0) {
2171 			hdr->b_flags |= ARC_PREFETCH;
2172 		}
2173 		DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2174 		arc_access(hdr, hash_lock);
2175 		mutex_exit(hash_lock);
2176 		ARCSTAT_BUMP(arcstat_hits);
2177 		ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2178 		    demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2179 		    data, metadata, hits);
2180 
2181 		if (done)
2182 			done(NULL, buf, private);
2183 	} else {
2184 		uint64_t size = BP_GET_LSIZE(bp);
2185 		arc_callback_t	*acb;
2186 
2187 		if (hdr == NULL) {
2188 			/* this block is not in the cache */
2189 			arc_buf_hdr_t	*exists;
2190 			arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2191 			buf = arc_buf_alloc(spa, size, private, type);
2192 			hdr = buf->b_hdr;
2193 			hdr->b_dva = *BP_IDENTITY(bp);
2194 			hdr->b_birth = bp->blk_birth;
2195 			hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
2196 			exists = buf_hash_insert(hdr, &hash_lock);
2197 			if (exists) {
2198 				/* somebody beat us to the hash insert */
2199 				mutex_exit(hash_lock);
2200 				bzero(&hdr->b_dva, sizeof (dva_t));
2201 				hdr->b_birth = 0;
2202 				hdr->b_cksum0 = 0;
2203 				(void) arc_buf_remove_ref(buf, private);
2204 				goto top; /* restart the IO request */
2205 			}
2206 			/* if this is a prefetch, we don't have a reference */
2207 			if (*arc_flags & ARC_PREFETCH) {
2208 				(void) remove_reference(hdr, hash_lock,
2209 				    private);
2210 				hdr->b_flags |= ARC_PREFETCH;
2211 			}
2212 			if (BP_GET_LEVEL(bp) > 0)
2213 				hdr->b_flags |= ARC_INDIRECT;
2214 		} else {
2215 			/* this block is in the ghost cache */
2216 			ASSERT(GHOST_STATE(hdr->b_state));
2217 			ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2218 			ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 0);
2219 			ASSERT(hdr->b_buf == NULL);
2220 
2221 			/* if this is a prefetch, we don't have a reference */
2222 			if (*arc_flags & ARC_PREFETCH)
2223 				hdr->b_flags |= ARC_PREFETCH;
2224 			else
2225 				add_reference(hdr, hash_lock, private);
2226 			buf = kmem_cache_alloc(buf_cache, KM_SLEEP);
2227 			buf->b_hdr = hdr;
2228 			buf->b_data = NULL;
2229 			buf->b_efunc = NULL;
2230 			buf->b_private = NULL;
2231 			buf->b_next = NULL;
2232 			hdr->b_buf = buf;
2233 			arc_get_data_buf(buf);
2234 			ASSERT(hdr->b_datacnt == 0);
2235 			hdr->b_datacnt = 1;
2236 
2237 		}
2238 
2239 		acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
2240 		acb->acb_done = done;
2241 		acb->acb_private = private;
2242 		acb->acb_byteswap = swap;
2243 
2244 		ASSERT(hdr->b_acb == NULL);
2245 		hdr->b_acb = acb;
2246 		hdr->b_flags |= ARC_IO_IN_PROGRESS;
2247 
2248 		/*
2249 		 * If the buffer has been evicted, migrate it to a present state
2250 		 * before issuing the I/O.  Once we drop the hash-table lock,
2251 		 * the header will be marked as I/O in progress and have an
2252 		 * attached buffer.  At this point, anybody who finds this
2253 		 * buffer ought to notice that it's legit but has a pending I/O.
2254 		 */
2255 
2256 		if (GHOST_STATE(hdr->b_state))
2257 			arc_access(hdr, hash_lock);
2258 		mutex_exit(hash_lock);
2259 
2260 		ASSERT3U(hdr->b_size, ==, size);
2261 		DTRACE_PROBE3(arc__miss, blkptr_t *, bp, uint64_t, size,
2262 		    zbookmark_t *, zb);
2263 		ARCSTAT_BUMP(arcstat_misses);
2264 		ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2265 		    demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2266 		    data, metadata, misses);
2267 
2268 		rzio = zio_read(pio, spa, bp, buf->b_data, size,
2269 		    arc_read_done, buf, priority, flags, zb);
2270 
2271 		if (*arc_flags & ARC_WAIT)
2272 			return (zio_wait(rzio));
2273 
2274 		ASSERT(*arc_flags & ARC_NOWAIT);
2275 		zio_nowait(rzio);
2276 	}
2277 	return (0);
2278 }
2279 
2280 /*
2281  * arc_read() variant to support pool traversal.  If the block is already
2282  * in the ARC, make a copy of it; otherwise, the caller will do the I/O.
2283  * The idea is that we don't want pool traversal filling up memory, but
2284  * if the ARC already has the data anyway, we shouldn't pay for the I/O.
2285  */
2286 int
2287 arc_tryread(spa_t *spa, blkptr_t *bp, void *data)
2288 {
2289 	arc_buf_hdr_t *hdr;
2290 	kmutex_t *hash_mtx;
2291 	int rc = 0;
2292 
2293 	hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_mtx);
2294 
2295 	if (hdr && hdr->b_datacnt > 0 && !HDR_IO_IN_PROGRESS(hdr)) {
2296 		arc_buf_t *buf = hdr->b_buf;
2297 
2298 		ASSERT(buf);
2299 		while (buf->b_data == NULL) {
2300 			buf = buf->b_next;
2301 			ASSERT(buf);
2302 		}
2303 		bcopy(buf->b_data, data, hdr->b_size);
2304 	} else {
2305 		rc = ENOENT;
2306 	}
2307 
2308 	if (hash_mtx)
2309 		mutex_exit(hash_mtx);
2310 
2311 	return (rc);
2312 }
2313 
2314 void
2315 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
2316 {
2317 	ASSERT(buf->b_hdr != NULL);
2318 	ASSERT(buf->b_hdr->b_state != arc_anon);
2319 	ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
2320 	buf->b_efunc = func;
2321 	buf->b_private = private;
2322 }
2323 
2324 /*
2325  * This is used by the DMU to let the ARC know that a buffer is
2326  * being evicted, so the ARC should clean up.  If this arc buf
2327  * is not yet in the evicted state, it will be put there.
2328  */
2329 int
2330 arc_buf_evict(arc_buf_t *buf)
2331 {
2332 	arc_buf_hdr_t *hdr;
2333 	kmutex_t *hash_lock;
2334 	arc_buf_t **bufp;
2335 
2336 	mutex_enter(&arc_eviction_mtx);
2337 	hdr = buf->b_hdr;
2338 	if (hdr == NULL) {
2339 		/*
2340 		 * We are in arc_do_user_evicts().
2341 		 */
2342 		ASSERT(buf->b_data == NULL);
2343 		mutex_exit(&arc_eviction_mtx);
2344 		return (0);
2345 	}
2346 	hash_lock = HDR_LOCK(hdr);
2347 	mutex_exit(&arc_eviction_mtx);
2348 
2349 	mutex_enter(hash_lock);
2350 
2351 	if (buf->b_data == NULL) {
2352 		/*
2353 		 * We are on the eviction list.
2354 		 */
2355 		mutex_exit(hash_lock);
2356 		mutex_enter(&arc_eviction_mtx);
2357 		if (buf->b_hdr == NULL) {
2358 			/*
2359 			 * We are already in arc_do_user_evicts().
2360 			 */
2361 			mutex_exit(&arc_eviction_mtx);
2362 			return (0);
2363 		} else {
2364 			arc_buf_t copy = *buf; /* structure assignment */
2365 			/*
2366 			 * Process this buffer now
2367 			 * but let arc_do_user_evicts() do the reaping.
2368 			 */
2369 			buf->b_efunc = NULL;
2370 			mutex_exit(&arc_eviction_mtx);
2371 			VERIFY(copy.b_efunc(&copy) == 0);
2372 			return (1);
2373 		}
2374 	}
2375 
2376 	ASSERT(buf->b_hdr == hdr);
2377 	ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
2378 	ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2379 
2380 	/*
2381 	 * Pull this buffer off of the hdr
2382 	 */
2383 	bufp = &hdr->b_buf;
2384 	while (*bufp != buf)
2385 		bufp = &(*bufp)->b_next;
2386 	*bufp = buf->b_next;
2387 
2388 	ASSERT(buf->b_data != NULL);
2389 	arc_buf_destroy(buf, FALSE, FALSE);
2390 
2391 	if (hdr->b_datacnt == 0) {
2392 		arc_state_t *old_state = hdr->b_state;
2393 		arc_state_t *evicted_state;
2394 
2395 		ASSERT(refcount_is_zero(&hdr->b_refcnt));
2396 
2397 		evicted_state =
2398 		    (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
2399 
2400 		mutex_enter(&old_state->arcs_mtx);
2401 		mutex_enter(&evicted_state->arcs_mtx);
2402 
2403 		arc_change_state(evicted_state, hdr, hash_lock);
2404 		ASSERT(HDR_IN_HASH_TABLE(hdr));
2405 		hdr->b_flags = ARC_IN_HASH_TABLE;
2406 
2407 		mutex_exit(&evicted_state->arcs_mtx);
2408 		mutex_exit(&old_state->arcs_mtx);
2409 	}
2410 	mutex_exit(hash_lock);
2411 
2412 	VERIFY(buf->b_efunc(buf) == 0);
2413 	buf->b_efunc = NULL;
2414 	buf->b_private = NULL;
2415 	buf->b_hdr = NULL;
2416 	kmem_cache_free(buf_cache, buf);
2417 	return (1);
2418 }
2419 
2420 /*
2421  * Release this buffer from the cache.  This must be done
2422  * after a read and prior to modifying the buffer contents.
2423  * If the buffer has more than one reference, we must make
2424  * make a new hdr for the buffer.
2425  */
2426 void
2427 arc_release(arc_buf_t *buf, void *tag)
2428 {
2429 	arc_buf_hdr_t *hdr = buf->b_hdr;
2430 	kmutex_t *hash_lock = HDR_LOCK(hdr);
2431 
2432 	/* this buffer is not on any list */
2433 	ASSERT(refcount_count(&hdr->b_refcnt) > 0);
2434 
2435 	if (hdr->b_state == arc_anon) {
2436 		/* this buffer is already released */
2437 		ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 1);
2438 		ASSERT(BUF_EMPTY(hdr));
2439 		ASSERT(buf->b_efunc == NULL);
2440 		arc_buf_thaw(buf);
2441 		return;
2442 	}
2443 
2444 	mutex_enter(hash_lock);
2445 
2446 	/*
2447 	 * Do we have more than one buf?
2448 	 */
2449 	if (hdr->b_buf != buf || buf->b_next != NULL) {
2450 		arc_buf_hdr_t *nhdr;
2451 		arc_buf_t **bufp;
2452 		uint64_t blksz = hdr->b_size;
2453 		spa_t *spa = hdr->b_spa;
2454 		arc_buf_contents_t type = hdr->b_type;
2455 
2456 		ASSERT(hdr->b_datacnt > 1);
2457 		/*
2458 		 * Pull the data off of this buf and attach it to
2459 		 * a new anonymous buf.
2460 		 */
2461 		(void) remove_reference(hdr, hash_lock, tag);
2462 		bufp = &hdr->b_buf;
2463 		while (*bufp != buf)
2464 			bufp = &(*bufp)->b_next;
2465 		*bufp = (*bufp)->b_next;
2466 		buf->b_next = NULL;
2467 
2468 		ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
2469 		atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
2470 		if (refcount_is_zero(&hdr->b_refcnt)) {
2471 			uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
2472 			ASSERT3U(*size, >=, hdr->b_size);
2473 			atomic_add_64(size, -hdr->b_size);
2474 		}
2475 		hdr->b_datacnt -= 1;
2476 		arc_cksum_verify(buf);
2477 
2478 		mutex_exit(hash_lock);
2479 
2480 		nhdr = kmem_cache_alloc(hdr_cache, KM_SLEEP);
2481 		nhdr->b_size = blksz;
2482 		nhdr->b_spa = spa;
2483 		nhdr->b_type = type;
2484 		nhdr->b_buf = buf;
2485 		nhdr->b_state = arc_anon;
2486 		nhdr->b_arc_access = 0;
2487 		nhdr->b_flags = 0;
2488 		nhdr->b_datacnt = 1;
2489 		nhdr->b_freeze_cksum = NULL;
2490 		(void) refcount_add(&nhdr->b_refcnt, tag);
2491 		buf->b_hdr = nhdr;
2492 		atomic_add_64(&arc_anon->arcs_size, blksz);
2493 
2494 		hdr = nhdr;
2495 	} else {
2496 		ASSERT(refcount_count(&hdr->b_refcnt) == 1);
2497 		ASSERT(!list_link_active(&hdr->b_arc_node));
2498 		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2499 		arc_change_state(arc_anon, hdr, hash_lock);
2500 		hdr->b_arc_access = 0;
2501 		mutex_exit(hash_lock);
2502 		bzero(&hdr->b_dva, sizeof (dva_t));
2503 		hdr->b_birth = 0;
2504 		hdr->b_cksum0 = 0;
2505 		arc_buf_thaw(buf);
2506 	}
2507 	buf->b_efunc = NULL;
2508 	buf->b_private = NULL;
2509 }
2510 
2511 int
2512 arc_released(arc_buf_t *buf)
2513 {
2514 	return (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
2515 }
2516 
2517 int
2518 arc_has_callback(arc_buf_t *buf)
2519 {
2520 	return (buf->b_efunc != NULL);
2521 }
2522 
2523 #ifdef ZFS_DEBUG
2524 int
2525 arc_referenced(arc_buf_t *buf)
2526 {
2527 	return (refcount_count(&buf->b_hdr->b_refcnt));
2528 }
2529 #endif
2530 
2531 static void
2532 arc_write_ready(zio_t *zio)
2533 {
2534 	arc_write_callback_t *callback = zio->io_private;
2535 	arc_buf_t *buf = callback->awcb_buf;
2536 
2537 	if (callback->awcb_ready) {
2538 		ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
2539 		callback->awcb_ready(zio, buf, callback->awcb_private);
2540 	}
2541 	arc_cksum_compute(buf);
2542 }
2543 
2544 static void
2545 arc_write_done(zio_t *zio)
2546 {
2547 	arc_write_callback_t *callback = zio->io_private;
2548 	arc_buf_t *buf = callback->awcb_buf;
2549 	arc_buf_hdr_t *hdr = buf->b_hdr;
2550 
2551 	hdr->b_acb = NULL;
2552 
2553 	/* this buffer is on no lists and is not in the hash table */
2554 	ASSERT3P(hdr->b_state, ==, arc_anon);
2555 
2556 	hdr->b_dva = *BP_IDENTITY(zio->io_bp);
2557 	hdr->b_birth = zio->io_bp->blk_birth;
2558 	hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
2559 	/*
2560 	 * If the block to be written was all-zero, we may have
2561 	 * compressed it away.  In this case no write was performed
2562 	 * so there will be no dva/birth-date/checksum.  The buffer
2563 	 * must therefor remain anonymous (and uncached).
2564 	 */
2565 	if (!BUF_EMPTY(hdr)) {
2566 		arc_buf_hdr_t *exists;
2567 		kmutex_t *hash_lock;
2568 
2569 		arc_cksum_verify(buf);
2570 
2571 		exists = buf_hash_insert(hdr, &hash_lock);
2572 		if (exists) {
2573 			/*
2574 			 * This can only happen if we overwrite for
2575 			 * sync-to-convergence, because we remove
2576 			 * buffers from the hash table when we arc_free().
2577 			 */
2578 			ASSERT(DVA_EQUAL(BP_IDENTITY(&zio->io_bp_orig),
2579 			    BP_IDENTITY(zio->io_bp)));
2580 			ASSERT3U(zio->io_bp_orig.blk_birth, ==,
2581 			    zio->io_bp->blk_birth);
2582 
2583 			ASSERT(refcount_is_zero(&exists->b_refcnt));
2584 			arc_change_state(arc_anon, exists, hash_lock);
2585 			mutex_exit(hash_lock);
2586 			arc_hdr_destroy(exists);
2587 			exists = buf_hash_insert(hdr, &hash_lock);
2588 			ASSERT3P(exists, ==, NULL);
2589 		}
2590 		hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2591 		arc_access(hdr, hash_lock);
2592 		mutex_exit(hash_lock);
2593 	} else if (callback->awcb_done == NULL) {
2594 		int destroy_hdr;
2595 		/*
2596 		 * This is an anonymous buffer with no user callback,
2597 		 * destroy it if there are no active references.
2598 		 */
2599 		mutex_enter(&arc_eviction_mtx);
2600 		destroy_hdr = refcount_is_zero(&hdr->b_refcnt);
2601 		hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2602 		mutex_exit(&arc_eviction_mtx);
2603 		if (destroy_hdr)
2604 			arc_hdr_destroy(hdr);
2605 	} else {
2606 		hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2607 	}
2608 
2609 	if (callback->awcb_done) {
2610 		ASSERT(!refcount_is_zero(&hdr->b_refcnt));
2611 		callback->awcb_done(zio, buf, callback->awcb_private);
2612 	}
2613 
2614 	kmem_free(callback, sizeof (arc_write_callback_t));
2615 }
2616 
2617 zio_t *
2618 arc_write(zio_t *pio, spa_t *spa, int checksum, int compress, int ncopies,
2619     uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
2620     arc_done_func_t *ready, arc_done_func_t *done, void *private, int priority,
2621     int flags, zbookmark_t *zb)
2622 {
2623 	arc_buf_hdr_t *hdr = buf->b_hdr;
2624 	arc_write_callback_t *callback;
2625 	zio_t	*zio;
2626 
2627 	/* this is a private buffer - no locking required */
2628 	ASSERT3P(hdr->b_state, ==, arc_anon);
2629 	ASSERT(BUF_EMPTY(hdr));
2630 	ASSERT(!HDR_IO_ERROR(hdr));
2631 	ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
2632 	ASSERT(hdr->b_acb == 0);
2633 	callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
2634 	callback->awcb_ready = ready;
2635 	callback->awcb_done = done;
2636 	callback->awcb_private = private;
2637 	callback->awcb_buf = buf;
2638 	hdr->b_flags |= ARC_IO_IN_PROGRESS;
2639 	zio = zio_write(pio, spa, checksum, compress, ncopies, txg, bp,
2640 	    buf->b_data, hdr->b_size, arc_write_ready, arc_write_done, callback,
2641 	    priority, flags, zb);
2642 
2643 	return (zio);
2644 }
2645 
2646 int
2647 arc_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
2648     zio_done_func_t *done, void *private, uint32_t arc_flags)
2649 {
2650 	arc_buf_hdr_t *ab;
2651 	kmutex_t *hash_lock;
2652 	zio_t	*zio;
2653 
2654 	/*
2655 	 * If this buffer is in the cache, release it, so it
2656 	 * can be re-used.
2657 	 */
2658 	ab = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock);
2659 	if (ab != NULL) {
2660 		/*
2661 		 * The checksum of blocks to free is not always
2662 		 * preserved (eg. on the deadlist).  However, if it is
2663 		 * nonzero, it should match what we have in the cache.
2664 		 */
2665 		ASSERT(bp->blk_cksum.zc_word[0] == 0 ||
2666 		    ab->b_cksum0 == bp->blk_cksum.zc_word[0]);
2667 		if (ab->b_state != arc_anon)
2668 			arc_change_state(arc_anon, ab, hash_lock);
2669 		if (HDR_IO_IN_PROGRESS(ab)) {
2670 			/*
2671 			 * This should only happen when we prefetch.
2672 			 */
2673 			ASSERT(ab->b_flags & ARC_PREFETCH);
2674 			ASSERT3U(ab->b_datacnt, ==, 1);
2675 			ab->b_flags |= ARC_FREED_IN_READ;
2676 			if (HDR_IN_HASH_TABLE(ab))
2677 				buf_hash_remove(ab);
2678 			ab->b_arc_access = 0;
2679 			bzero(&ab->b_dva, sizeof (dva_t));
2680 			ab->b_birth = 0;
2681 			ab->b_cksum0 = 0;
2682 			ab->b_buf->b_efunc = NULL;
2683 			ab->b_buf->b_private = NULL;
2684 			mutex_exit(hash_lock);
2685 		} else if (refcount_is_zero(&ab->b_refcnt)) {
2686 			mutex_exit(hash_lock);
2687 			arc_hdr_destroy(ab);
2688 			ARCSTAT_BUMP(arcstat_deleted);
2689 		} else {
2690 			/*
2691 			 * We still have an active reference on this
2692 			 * buffer.  This can happen, e.g., from
2693 			 * dbuf_unoverride().
2694 			 */
2695 			ASSERT(!HDR_IN_HASH_TABLE(ab));
2696 			ab->b_arc_access = 0;
2697 			bzero(&ab->b_dva, sizeof (dva_t));
2698 			ab->b_birth = 0;
2699 			ab->b_cksum0 = 0;
2700 			ab->b_buf->b_efunc = NULL;
2701 			ab->b_buf->b_private = NULL;
2702 			mutex_exit(hash_lock);
2703 		}
2704 	}
2705 
2706 	zio = zio_free(pio, spa, txg, bp, done, private);
2707 
2708 	if (arc_flags & ARC_WAIT)
2709 		return (zio_wait(zio));
2710 
2711 	ASSERT(arc_flags & ARC_NOWAIT);
2712 	zio_nowait(zio);
2713 
2714 	return (0);
2715 }
2716 
2717 void
2718 arc_tempreserve_clear(uint64_t tempreserve)
2719 {
2720 	atomic_add_64(&arc_tempreserve, -tempreserve);
2721 	ASSERT((int64_t)arc_tempreserve >= 0);
2722 }
2723 
2724 int
2725 arc_tempreserve_space(uint64_t tempreserve)
2726 {
2727 #ifdef ZFS_DEBUG
2728 	/*
2729 	 * Once in a while, fail for no reason.  Everything should cope.
2730 	 */
2731 	if (spa_get_random(10000) == 0) {
2732 		dprintf("forcing random failure\n");
2733 		return (ERESTART);
2734 	}
2735 #endif
2736 	if (tempreserve > arc_c/4 && !arc_no_grow)
2737 		arc_c = MIN(arc_c_max, tempreserve * 4);
2738 	if (tempreserve > arc_c)
2739 		return (ENOMEM);
2740 
2741 	/*
2742 	 * Throttle writes when the amount of dirty data in the cache
2743 	 * gets too large.  We try to keep the cache less than half full
2744 	 * of dirty blocks so that our sync times don't grow too large.
2745 	 * Note: if two requests come in concurrently, we might let them
2746 	 * both succeed, when one of them should fail.  Not a huge deal.
2747 	 *
2748 	 * XXX The limit should be adjusted dynamically to keep the time
2749 	 * to sync a dataset fixed (around 1-5 seconds?).
2750 	 */
2751 
2752 	if (tempreserve + arc_tempreserve + arc_anon->arcs_size > arc_c / 2 &&
2753 	    arc_tempreserve + arc_anon->arcs_size > arc_c / 4) {
2754 		dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
2755 		    "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
2756 		    arc_tempreserve>>10,
2757 		    arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
2758 		    arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
2759 		    tempreserve>>10, arc_c>>10);
2760 		return (ERESTART);
2761 	}
2762 	atomic_add_64(&arc_tempreserve, tempreserve);
2763 	return (0);
2764 }
2765 
2766 void
2767 arc_init(void)
2768 {
2769 	mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
2770 	cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
2771 
2772 	/* Convert seconds to clock ticks */
2773 	arc_min_prefetch_lifespan = 1 * hz;
2774 
2775 	/* Start out with 1/8 of all memory */
2776 	arc_c = physmem * PAGESIZE / 8;
2777 
2778 #ifdef _KERNEL
2779 	/*
2780 	 * On architectures where the physical memory can be larger
2781 	 * than the addressable space (intel in 32-bit mode), we may
2782 	 * need to limit the cache to 1/8 of VM size.
2783 	 */
2784 	arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
2785 #endif
2786 
2787 	/* set min cache to 1/32 of all memory, or 64MB, whichever is more */
2788 	arc_c_min = MAX(arc_c / 4, 64<<20);
2789 	/* set max to 3/4 of all memory, or all but 1GB, whichever is more */
2790 	if (arc_c * 8 >= 1<<30)
2791 		arc_c_max = (arc_c * 8) - (1<<30);
2792 	else
2793 		arc_c_max = arc_c_min;
2794 	arc_c_max = MAX(arc_c * 6, arc_c_max);
2795 
2796 	/*
2797 	 * Allow the tunables to override our calculations if they are
2798 	 * reasonable (ie. over 64MB)
2799 	 */
2800 	if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
2801 		arc_c_max = zfs_arc_max;
2802 	if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
2803 		arc_c_min = zfs_arc_min;
2804 
2805 	arc_c = arc_c_max;
2806 	arc_p = (arc_c >> 1);
2807 
2808 	/* limit meta-data to 1/4 of the arc capacity */
2809 	arc_meta_limit = arc_c_max / 4;
2810 
2811 	/* Allow the tunable to override if it is reasonable */
2812 	if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
2813 		arc_meta_limit = zfs_arc_meta_limit;
2814 
2815 	if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
2816 		arc_c_min = arc_meta_limit / 2;
2817 
2818 	/* if kmem_flags are set, lets try to use less memory */
2819 	if (kmem_debugging())
2820 		arc_c = arc_c / 2;
2821 	if (arc_c < arc_c_min)
2822 		arc_c = arc_c_min;
2823 
2824 	arc_anon = &ARC_anon;
2825 	arc_mru = &ARC_mru;
2826 	arc_mru_ghost = &ARC_mru_ghost;
2827 	arc_mfu = &ARC_mfu;
2828 	arc_mfu_ghost = &ARC_mfu_ghost;
2829 	arc_size = 0;
2830 
2831 	mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
2832 	mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
2833 	mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
2834 	mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
2835 	mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
2836 
2837 	list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
2838 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2839 	list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
2840 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2841 	list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
2842 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2843 	list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
2844 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2845 	list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
2846 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2847 	list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
2848 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2849 	list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
2850 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2851 	list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
2852 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
2853 
2854 	buf_init();
2855 
2856 	arc_thread_exit = 0;
2857 	arc_eviction_list = NULL;
2858 	mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
2859 	bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
2860 
2861 	arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
2862 	    sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
2863 
2864 	if (arc_ksp != NULL) {
2865 		arc_ksp->ks_data = &arc_stats;
2866 		kstat_install(arc_ksp);
2867 	}
2868 
2869 	(void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
2870 	    TS_RUN, minclsyspri);
2871 
2872 	arc_dead = FALSE;
2873 }
2874 
2875 void
2876 arc_fini(void)
2877 {
2878 	mutex_enter(&arc_reclaim_thr_lock);
2879 	arc_thread_exit = 1;
2880 	while (arc_thread_exit != 0)
2881 		cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
2882 	mutex_exit(&arc_reclaim_thr_lock);
2883 
2884 	arc_flush();
2885 
2886 	arc_dead = TRUE;
2887 
2888 	if (arc_ksp != NULL) {
2889 		kstat_delete(arc_ksp);
2890 		arc_ksp = NULL;
2891 	}
2892 
2893 	mutex_destroy(&arc_eviction_mtx);
2894 	mutex_destroy(&arc_reclaim_thr_lock);
2895 	cv_destroy(&arc_reclaim_thr_cv);
2896 
2897 	list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
2898 	list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
2899 	list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
2900 	list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
2901 	list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
2902 	list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
2903 	list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
2904 	list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
2905 
2906 	mutex_destroy(&arc_anon->arcs_mtx);
2907 	mutex_destroy(&arc_mru->arcs_mtx);
2908 	mutex_destroy(&arc_mru_ghost->arcs_mtx);
2909 	mutex_destroy(&arc_mfu->arcs_mtx);
2910 	mutex_destroy(&arc_mfu_ghost->arcs_mtx);
2911 
2912 	buf_fini();
2913 }
2914