xref: /illumos-gate/usr/src/uts/common/fs/zfs/arc.c (revision e14bb3258d05c1b1077e2db7cf77088924e56919)
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 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*
27  * DVA-based Adjustable Replacement Cache
28  *
29  * While much of the theory of operation used here is
30  * based on the self-tuning, low overhead replacement cache
31  * presented by Megiddo and Modha at FAST 2003, there are some
32  * significant differences:
33  *
34  * 1. The Megiddo and Modha model assumes any page is evictable.
35  * Pages in its cache cannot be "locked" into memory.  This makes
36  * the eviction algorithm simple: evict the last page in the list.
37  * This also make the performance characteristics easy to reason
38  * about.  Our cache is not so simple.  At any given moment, some
39  * subset of the blocks in the cache are un-evictable because we
40  * have handed out a reference to them.  Blocks are only evictable
41  * when there are no external references active.  This makes
42  * eviction far more problematic:  we choose to evict the evictable
43  * blocks that are the "lowest" in the list.
44  *
45  * There are times when it is not possible to evict the requested
46  * space.  In these circumstances we are unable to adjust the cache
47  * size.  To prevent the cache growing unbounded at these times we
48  * implement a "cache throttle" that slows the flow of new data
49  * into the cache until we can make space available.
50  *
51  * 2. The Megiddo and Modha model assumes a fixed cache size.
52  * Pages are evicted when the cache is full and there is a cache
53  * miss.  Our model has a variable sized cache.  It grows with
54  * high use, but also tries to react to memory pressure from the
55  * operating system: decreasing its size when system memory is
56  * tight.
57  *
58  * 3. The Megiddo and Modha model assumes a fixed page size. All
59  * elements of the cache are therefor exactly the same size.  So
60  * when adjusting the cache size following a cache miss, its simply
61  * a matter of choosing a single page to evict.  In our model, we
62  * have variable sized cache blocks (rangeing from 512 bytes to
63  * 128K bytes).  We therefor choose a set of blocks to evict to make
64  * space for a cache miss that approximates as closely as possible
65  * the space used by the new block.
66  *
67  * See also:  "ARC: A Self-Tuning, Low Overhead Replacement Cache"
68  * by N. Megiddo & D. Modha, FAST 2003
69  */
70 
71 /*
72  * The locking model:
73  *
74  * A new reference to a cache buffer can be obtained in two
75  * ways: 1) via a hash table lookup using the DVA as a key,
76  * or 2) via one of the ARC lists.  The arc_read() interface
77  * uses method 1, while the internal arc algorithms for
78  * adjusting the cache use method 2.  We therefor provide two
79  * types of locks: 1) the hash table lock array, and 2) the
80  * arc list locks.
81  *
82  * Buffers do not have their own mutexs, rather they rely on the
83  * hash table mutexs for the bulk of their protection (i.e. most
84  * fields in the arc_buf_hdr_t are protected by these mutexs).
85  *
86  * buf_hash_find() returns the appropriate mutex (held) when it
87  * locates the requested buffer in the hash table.  It returns
88  * NULL for the mutex if the buffer was not in the table.
89  *
90  * buf_hash_remove() expects the appropriate hash mutex to be
91  * already held before it is invoked.
92  *
93  * Each arc state also has a mutex which is used to protect the
94  * buffer list associated with the state.  When attempting to
95  * obtain a hash table lock while holding an arc list lock you
96  * must use: mutex_tryenter() to avoid deadlock.  Also note that
97  * the active state mutex must be held before the ghost state mutex.
98  *
99  * Arc buffers may have an associated eviction callback function.
100  * This function will be invoked prior to removing the buffer (e.g.
101  * in arc_do_user_evicts()).  Note however that the data associated
102  * with the buffer may be evicted prior to the callback.  The callback
103  * must be made with *no locks held* (to prevent deadlock).  Additionally,
104  * the users of callbacks must ensure that their private data is
105  * protected from simultaneous callbacks from arc_buf_evict()
106  * and arc_do_user_evicts().
107  *
108  * Note that the majority of the performance stats are manipulated
109  * with atomic operations.
110  *
111  * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
112  *
113  *	- L2ARC buflist creation
114  *	- L2ARC buflist eviction
115  *	- L2ARC write completion, which walks L2ARC buflists
116  *	- ARC header destruction, as it removes from L2ARC buflists
117  *	- ARC header release, as it removes from L2ARC buflists
118  */
119 
120 #include <sys/spa.h>
121 #include <sys/zio.h>
122 #include <sys/zio_checksum.h>
123 #include <sys/zfs_context.h>
124 #include <sys/arc.h>
125 #include <sys/refcount.h>
126 #include <sys/vdev.h>
127 #ifdef _KERNEL
128 #include <sys/vmsystm.h>
129 #include <vm/anon.h>
130 #include <sys/fs/swapnode.h>
131 #include <sys/dnlc.h>
132 #endif
133 #include <sys/callb.h>
134 #include <sys/kstat.h>
135 
136 static kmutex_t		arc_reclaim_thr_lock;
137 static kcondvar_t	arc_reclaim_thr_cv;	/* used to signal reclaim thr */
138 static uint8_t		arc_thread_exit;
139 
140 extern int zfs_write_limit_shift;
141 extern uint64_t zfs_write_limit_max;
142 extern kmutex_t zfs_write_limit_lock;
143 
144 #define	ARC_REDUCE_DNLC_PERCENT	3
145 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
146 
147 typedef enum arc_reclaim_strategy {
148 	ARC_RECLAIM_AGGR,		/* Aggressive reclaim strategy */
149 	ARC_RECLAIM_CONS		/* Conservative reclaim strategy */
150 } arc_reclaim_strategy_t;
151 
152 /* number of seconds before growing cache again */
153 static int		arc_grow_retry = 60;
154 
155 /*
156  * minimum lifespan of a prefetch block in clock ticks
157  * (initialized in arc_init())
158  */
159 static int		arc_min_prefetch_lifespan;
160 
161 static int arc_dead;
162 
163 /*
164  * The arc has filled available memory and has now warmed up.
165  */
166 static boolean_t arc_warm;
167 
168 /*
169  * These tunables are for performance analysis.
170  */
171 uint64_t zfs_arc_max;
172 uint64_t zfs_arc_min;
173 uint64_t zfs_arc_meta_limit = 0;
174 int zfs_mdcomp_disable = 0;
175 
176 /*
177  * Note that buffers can be in one of 6 states:
178  *	ARC_anon	- anonymous (discussed below)
179  *	ARC_mru		- recently used, currently cached
180  *	ARC_mru_ghost	- recentely used, no longer in cache
181  *	ARC_mfu		- frequently used, currently cached
182  *	ARC_mfu_ghost	- frequently used, no longer in cache
183  *	ARC_l2c_only	- exists in L2ARC but not other states
184  * When there are no active references to the buffer, they are
185  * are linked onto a list in one of these arc states.  These are
186  * the only buffers that can be evicted or deleted.  Within each
187  * state there are multiple lists, one for meta-data and one for
188  * non-meta-data.  Meta-data (indirect blocks, blocks of dnodes,
189  * etc.) is tracked separately so that it can be managed more
190  * explicitly: favored over data, limited explicitly.
191  *
192  * Anonymous buffers are buffers that are not associated with
193  * a DVA.  These are buffers that hold dirty block copies
194  * before they are written to stable storage.  By definition,
195  * they are "ref'd" and are considered part of arc_mru
196  * that cannot be freed.  Generally, they will aquire a DVA
197  * as they are written and migrate onto the arc_mru list.
198  *
199  * The ARC_l2c_only state is for buffers that are in the second
200  * level ARC but no longer in any of the ARC_m* lists.  The second
201  * level ARC itself may also contain buffers that are in any of
202  * the ARC_m* states - meaning that a buffer can exist in two
203  * places.  The reason for the ARC_l2c_only state is to keep the
204  * buffer header in the hash table, so that reads that hit the
205  * second level ARC benefit from these fast lookups.
206  */
207 
208 typedef struct arc_state {
209 	list_t	arcs_list[ARC_BUFC_NUMTYPES];	/* list of evictable buffers */
210 	uint64_t arcs_lsize[ARC_BUFC_NUMTYPES];	/* amount of evictable data */
211 	uint64_t arcs_size;	/* total amount of data in this state */
212 	kmutex_t arcs_mtx;
213 } arc_state_t;
214 
215 /* The 6 states: */
216 static arc_state_t ARC_anon;
217 static arc_state_t ARC_mru;
218 static arc_state_t ARC_mru_ghost;
219 static arc_state_t ARC_mfu;
220 static arc_state_t ARC_mfu_ghost;
221 static arc_state_t ARC_l2c_only;
222 
223 typedef struct arc_stats {
224 	kstat_named_t arcstat_hits;
225 	kstat_named_t arcstat_misses;
226 	kstat_named_t arcstat_demand_data_hits;
227 	kstat_named_t arcstat_demand_data_misses;
228 	kstat_named_t arcstat_demand_metadata_hits;
229 	kstat_named_t arcstat_demand_metadata_misses;
230 	kstat_named_t arcstat_prefetch_data_hits;
231 	kstat_named_t arcstat_prefetch_data_misses;
232 	kstat_named_t arcstat_prefetch_metadata_hits;
233 	kstat_named_t arcstat_prefetch_metadata_misses;
234 	kstat_named_t arcstat_mru_hits;
235 	kstat_named_t arcstat_mru_ghost_hits;
236 	kstat_named_t arcstat_mfu_hits;
237 	kstat_named_t arcstat_mfu_ghost_hits;
238 	kstat_named_t arcstat_deleted;
239 	kstat_named_t arcstat_recycle_miss;
240 	kstat_named_t arcstat_mutex_miss;
241 	kstat_named_t arcstat_evict_skip;
242 	kstat_named_t arcstat_hash_elements;
243 	kstat_named_t arcstat_hash_elements_max;
244 	kstat_named_t arcstat_hash_collisions;
245 	kstat_named_t arcstat_hash_chains;
246 	kstat_named_t arcstat_hash_chain_max;
247 	kstat_named_t arcstat_p;
248 	kstat_named_t arcstat_c;
249 	kstat_named_t arcstat_c_min;
250 	kstat_named_t arcstat_c_max;
251 	kstat_named_t arcstat_size;
252 	kstat_named_t arcstat_hdr_size;
253 	kstat_named_t arcstat_l2_hits;
254 	kstat_named_t arcstat_l2_misses;
255 	kstat_named_t arcstat_l2_feeds;
256 	kstat_named_t arcstat_l2_rw_clash;
257 	kstat_named_t arcstat_l2_writes_sent;
258 	kstat_named_t arcstat_l2_writes_done;
259 	kstat_named_t arcstat_l2_writes_error;
260 	kstat_named_t arcstat_l2_writes_hdr_miss;
261 	kstat_named_t arcstat_l2_evict_lock_retry;
262 	kstat_named_t arcstat_l2_evict_reading;
263 	kstat_named_t arcstat_l2_free_on_write;
264 	kstat_named_t arcstat_l2_abort_lowmem;
265 	kstat_named_t arcstat_l2_cksum_bad;
266 	kstat_named_t arcstat_l2_io_error;
267 	kstat_named_t arcstat_l2_size;
268 	kstat_named_t arcstat_l2_hdr_size;
269 	kstat_named_t arcstat_memory_throttle_count;
270 } arc_stats_t;
271 
272 static arc_stats_t arc_stats = {
273 	{ "hits",			KSTAT_DATA_UINT64 },
274 	{ "misses",			KSTAT_DATA_UINT64 },
275 	{ "demand_data_hits",		KSTAT_DATA_UINT64 },
276 	{ "demand_data_misses",		KSTAT_DATA_UINT64 },
277 	{ "demand_metadata_hits",	KSTAT_DATA_UINT64 },
278 	{ "demand_metadata_misses",	KSTAT_DATA_UINT64 },
279 	{ "prefetch_data_hits",		KSTAT_DATA_UINT64 },
280 	{ "prefetch_data_misses",	KSTAT_DATA_UINT64 },
281 	{ "prefetch_metadata_hits",	KSTAT_DATA_UINT64 },
282 	{ "prefetch_metadata_misses",	KSTAT_DATA_UINT64 },
283 	{ "mru_hits",			KSTAT_DATA_UINT64 },
284 	{ "mru_ghost_hits",		KSTAT_DATA_UINT64 },
285 	{ "mfu_hits",			KSTAT_DATA_UINT64 },
286 	{ "mfu_ghost_hits",		KSTAT_DATA_UINT64 },
287 	{ "deleted",			KSTAT_DATA_UINT64 },
288 	{ "recycle_miss",		KSTAT_DATA_UINT64 },
289 	{ "mutex_miss",			KSTAT_DATA_UINT64 },
290 	{ "evict_skip",			KSTAT_DATA_UINT64 },
291 	{ "hash_elements",		KSTAT_DATA_UINT64 },
292 	{ "hash_elements_max",		KSTAT_DATA_UINT64 },
293 	{ "hash_collisions",		KSTAT_DATA_UINT64 },
294 	{ "hash_chains",		KSTAT_DATA_UINT64 },
295 	{ "hash_chain_max",		KSTAT_DATA_UINT64 },
296 	{ "p",				KSTAT_DATA_UINT64 },
297 	{ "c",				KSTAT_DATA_UINT64 },
298 	{ "c_min",			KSTAT_DATA_UINT64 },
299 	{ "c_max",			KSTAT_DATA_UINT64 },
300 	{ "size",			KSTAT_DATA_UINT64 },
301 	{ "hdr_size",			KSTAT_DATA_UINT64 },
302 	{ "l2_hits",			KSTAT_DATA_UINT64 },
303 	{ "l2_misses",			KSTAT_DATA_UINT64 },
304 	{ "l2_feeds",			KSTAT_DATA_UINT64 },
305 	{ "l2_rw_clash",		KSTAT_DATA_UINT64 },
306 	{ "l2_writes_sent",		KSTAT_DATA_UINT64 },
307 	{ "l2_writes_done",		KSTAT_DATA_UINT64 },
308 	{ "l2_writes_error",		KSTAT_DATA_UINT64 },
309 	{ "l2_writes_hdr_miss",		KSTAT_DATA_UINT64 },
310 	{ "l2_evict_lock_retry",	KSTAT_DATA_UINT64 },
311 	{ "l2_evict_reading",		KSTAT_DATA_UINT64 },
312 	{ "l2_free_on_write",		KSTAT_DATA_UINT64 },
313 	{ "l2_abort_lowmem",		KSTAT_DATA_UINT64 },
314 	{ "l2_cksum_bad",		KSTAT_DATA_UINT64 },
315 	{ "l2_io_error",		KSTAT_DATA_UINT64 },
316 	{ "l2_size",			KSTAT_DATA_UINT64 },
317 	{ "l2_hdr_size",		KSTAT_DATA_UINT64 },
318 	{ "memory_throttle_count",	KSTAT_DATA_UINT64 }
319 };
320 
321 #define	ARCSTAT(stat)	(arc_stats.stat.value.ui64)
322 
323 #define	ARCSTAT_INCR(stat, val) \
324 	atomic_add_64(&arc_stats.stat.value.ui64, (val));
325 
326 #define	ARCSTAT_BUMP(stat) 	ARCSTAT_INCR(stat, 1)
327 #define	ARCSTAT_BUMPDOWN(stat)	ARCSTAT_INCR(stat, -1)
328 
329 #define	ARCSTAT_MAX(stat, val) {					\
330 	uint64_t m;							\
331 	while ((val) > (m = arc_stats.stat.value.ui64) &&		\
332 	    (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val))))	\
333 		continue;						\
334 }
335 
336 #define	ARCSTAT_MAXSTAT(stat) \
337 	ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
338 
339 /*
340  * We define a macro to allow ARC hits/misses to be easily broken down by
341  * two separate conditions, giving a total of four different subtypes for
342  * each of hits and misses (so eight statistics total).
343  */
344 #define	ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
345 	if (cond1) {							\
346 		if (cond2) {						\
347 			ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
348 		} else {						\
349 			ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
350 		}							\
351 	} else {							\
352 		if (cond2) {						\
353 			ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
354 		} else {						\
355 			ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
356 		}							\
357 	}
358 
359 kstat_t			*arc_ksp;
360 static arc_state_t 	*arc_anon;
361 static arc_state_t	*arc_mru;
362 static arc_state_t	*arc_mru_ghost;
363 static arc_state_t	*arc_mfu;
364 static arc_state_t	*arc_mfu_ghost;
365 static arc_state_t	*arc_l2c_only;
366 
367 /*
368  * There are several ARC variables that are critical to export as kstats --
369  * but we don't want to have to grovel around in the kstat whenever we wish to
370  * manipulate them.  For these variables, we therefore define them to be in
371  * terms of the statistic variable.  This assures that we are not introducing
372  * the possibility of inconsistency by having shadow copies of the variables,
373  * while still allowing the code to be readable.
374  */
375 #define	arc_size	ARCSTAT(arcstat_size)	/* actual total arc size */
376 #define	arc_p		ARCSTAT(arcstat_p)	/* target size of MRU */
377 #define	arc_c		ARCSTAT(arcstat_c)	/* target size of cache */
378 #define	arc_c_min	ARCSTAT(arcstat_c_min)	/* min target cache size */
379 #define	arc_c_max	ARCSTAT(arcstat_c_max)	/* max target cache size */
380 
381 static int		arc_no_grow;	/* Don't try to grow cache size */
382 static uint64_t		arc_tempreserve;
383 static uint64_t		arc_meta_used;
384 static uint64_t		arc_meta_limit;
385 static uint64_t		arc_meta_max = 0;
386 
387 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
388 
389 typedef struct arc_callback arc_callback_t;
390 
391 struct arc_callback {
392 	void			*acb_private;
393 	arc_done_func_t		*acb_done;
394 	arc_buf_t		*acb_buf;
395 	zio_t			*acb_zio_dummy;
396 	arc_callback_t		*acb_next;
397 };
398 
399 typedef struct arc_write_callback arc_write_callback_t;
400 
401 struct arc_write_callback {
402 	void		*awcb_private;
403 	arc_done_func_t	*awcb_ready;
404 	arc_done_func_t	*awcb_done;
405 	arc_buf_t	*awcb_buf;
406 };
407 
408 struct arc_buf_hdr {
409 	/* protected by hash lock */
410 	dva_t			b_dva;
411 	uint64_t		b_birth;
412 	uint64_t		b_cksum0;
413 
414 	kmutex_t		b_freeze_lock;
415 	zio_cksum_t		*b_freeze_cksum;
416 
417 	arc_buf_hdr_t		*b_hash_next;
418 	arc_buf_t		*b_buf;
419 	uint32_t		b_flags;
420 	uint32_t		b_datacnt;
421 
422 	arc_callback_t		*b_acb;
423 	kcondvar_t		b_cv;
424 
425 	/* immutable */
426 	arc_buf_contents_t	b_type;
427 	uint64_t		b_size;
428 	spa_t			*b_spa;
429 
430 	/* protected by arc state mutex */
431 	arc_state_t		*b_state;
432 	list_node_t		b_arc_node;
433 
434 	/* updated atomically */
435 	clock_t			b_arc_access;
436 
437 	/* self protecting */
438 	refcount_t		b_refcnt;
439 
440 	l2arc_buf_hdr_t		*b_l2hdr;
441 	list_node_t		b_l2node;
442 };
443 
444 static arc_buf_t *arc_eviction_list;
445 static kmutex_t arc_eviction_mtx;
446 static arc_buf_hdr_t arc_eviction_hdr;
447 static void arc_get_data_buf(arc_buf_t *buf);
448 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
449 static int arc_evict_needed(arc_buf_contents_t type);
450 static void arc_evict_ghost(arc_state_t *state, spa_t *spa, int64_t bytes);
451 
452 #define	GHOST_STATE(state)	\
453 	((state) == arc_mru_ghost || (state) == arc_mfu_ghost ||	\
454 	(state) == arc_l2c_only)
455 
456 /*
457  * Private ARC flags.  These flags are private ARC only flags that will show up
458  * in b_flags in the arc_hdr_buf_t.  Some flags are publicly declared, and can
459  * be passed in as arc_flags in things like arc_read.  However, these flags
460  * should never be passed and should only be set by ARC code.  When adding new
461  * public flags, make sure not to smash the private ones.
462  */
463 
464 #define	ARC_IN_HASH_TABLE	(1 << 9)	/* this buffer is hashed */
465 #define	ARC_IO_IN_PROGRESS	(1 << 10)	/* I/O in progress for buf */
466 #define	ARC_IO_ERROR		(1 << 11)	/* I/O failed for buf */
467 #define	ARC_FREED_IN_READ	(1 << 12)	/* buf freed while in read */
468 #define	ARC_BUF_AVAILABLE	(1 << 13)	/* block not in active use */
469 #define	ARC_INDIRECT		(1 << 14)	/* this is an indirect block */
470 #define	ARC_FREE_IN_PROGRESS	(1 << 15)	/* hdr about to be freed */
471 #define	ARC_L2_WRITING		(1 << 16)	/* L2ARC write in progress */
472 #define	ARC_L2_EVICTED		(1 << 17)	/* evicted during I/O */
473 #define	ARC_L2_WRITE_HEAD	(1 << 18)	/* head of write list */
474 #define	ARC_STORED		(1 << 19)	/* has been store()d to */
475 
476 #define	HDR_IN_HASH_TABLE(hdr)	((hdr)->b_flags & ARC_IN_HASH_TABLE)
477 #define	HDR_IO_IN_PROGRESS(hdr)	((hdr)->b_flags & ARC_IO_IN_PROGRESS)
478 #define	HDR_IO_ERROR(hdr)	((hdr)->b_flags & ARC_IO_ERROR)
479 #define	HDR_FREED_IN_READ(hdr)	((hdr)->b_flags & ARC_FREED_IN_READ)
480 #define	HDR_BUF_AVAILABLE(hdr)	((hdr)->b_flags & ARC_BUF_AVAILABLE)
481 #define	HDR_FREE_IN_PROGRESS(hdr)	((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
482 #define	HDR_L2CACHE(hdr)	((hdr)->b_flags & ARC_L2CACHE)
483 #define	HDR_L2_READING(hdr)	((hdr)->b_flags & ARC_IO_IN_PROGRESS &&	\
484 				    (hdr)->b_l2hdr != NULL)
485 #define	HDR_L2_WRITING(hdr)	((hdr)->b_flags & ARC_L2_WRITING)
486 #define	HDR_L2_EVICTED(hdr)	((hdr)->b_flags & ARC_L2_EVICTED)
487 #define	HDR_L2_WRITE_HEAD(hdr)	((hdr)->b_flags & ARC_L2_WRITE_HEAD)
488 
489 /*
490  * Other sizes
491  */
492 
493 #define	HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
494 #define	L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
495 
496 /*
497  * Hash table routines
498  */
499 
500 #define	HT_LOCK_PAD	64
501 
502 struct ht_lock {
503 	kmutex_t	ht_lock;
504 #ifdef _KERNEL
505 	unsigned char	pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
506 #endif
507 };
508 
509 #define	BUF_LOCKS 256
510 typedef struct buf_hash_table {
511 	uint64_t ht_mask;
512 	arc_buf_hdr_t **ht_table;
513 	struct ht_lock ht_locks[BUF_LOCKS];
514 } buf_hash_table_t;
515 
516 static buf_hash_table_t buf_hash_table;
517 
518 #define	BUF_HASH_INDEX(spa, dva, birth) \
519 	(buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
520 #define	BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
521 #define	BUF_HASH_LOCK(idx)	(&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
522 #define	HDR_LOCK(buf) \
523 	(BUF_HASH_LOCK(BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth)))
524 
525 uint64_t zfs_crc64_table[256];
526 
527 /*
528  * Level 2 ARC
529  */
530 
531 #define	L2ARC_WRITE_SIZE	(8 * 1024 * 1024)	/* initial write max */
532 #define	L2ARC_HEADROOM		4		/* num of writes */
533 #define	L2ARC_FEED_SECS		1		/* caching interval */
534 
535 #define	l2arc_writes_sent	ARCSTAT(arcstat_l2_writes_sent)
536 #define	l2arc_writes_done	ARCSTAT(arcstat_l2_writes_done)
537 
538 /*
539  * L2ARC Performance Tunables
540  */
541 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE;	/* default max write size */
542 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE;	/* extra write during warmup */
543 uint64_t l2arc_headroom = L2ARC_HEADROOM;	/* number of dev writes */
544 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS;	/* interval seconds */
545 boolean_t l2arc_noprefetch = B_TRUE;		/* don't cache prefetch bufs */
546 
547 /*
548  * L2ARC Internals
549  */
550 typedef struct l2arc_dev {
551 	vdev_t			*l2ad_vdev;	/* vdev */
552 	spa_t			*l2ad_spa;	/* spa */
553 	uint64_t		l2ad_hand;	/* next write location */
554 	uint64_t		l2ad_write;	/* desired write size, bytes */
555 	uint64_t		l2ad_boost;	/* warmup write boost, bytes */
556 	uint64_t		l2ad_start;	/* first addr on device */
557 	uint64_t		l2ad_end;	/* last addr on device */
558 	uint64_t		l2ad_evict;	/* last addr eviction reached */
559 	boolean_t		l2ad_first;	/* first sweep through */
560 	list_t			*l2ad_buflist;	/* buffer list */
561 	list_node_t		l2ad_node;	/* device list node */
562 } l2arc_dev_t;
563 
564 static list_t L2ARC_dev_list;			/* device list */
565 static list_t *l2arc_dev_list;			/* device list pointer */
566 static kmutex_t l2arc_dev_mtx;			/* device list mutex */
567 static l2arc_dev_t *l2arc_dev_last;		/* last device used */
568 static kmutex_t l2arc_buflist_mtx;		/* mutex for all buflists */
569 static list_t L2ARC_free_on_write;		/* free after write buf list */
570 static list_t *l2arc_free_on_write;		/* free after write list ptr */
571 static kmutex_t l2arc_free_on_write_mtx;	/* mutex for list */
572 static uint64_t l2arc_ndev;			/* number of devices */
573 
574 typedef struct l2arc_read_callback {
575 	arc_buf_t	*l2rcb_buf;		/* read buffer */
576 	spa_t		*l2rcb_spa;		/* spa */
577 	blkptr_t	l2rcb_bp;		/* original blkptr */
578 	zbookmark_t	l2rcb_zb;		/* original bookmark */
579 	int		l2rcb_flags;		/* original flags */
580 } l2arc_read_callback_t;
581 
582 typedef struct l2arc_write_callback {
583 	l2arc_dev_t	*l2wcb_dev;		/* device info */
584 	arc_buf_hdr_t	*l2wcb_head;		/* head of write buflist */
585 } l2arc_write_callback_t;
586 
587 struct l2arc_buf_hdr {
588 	/* protected by arc_buf_hdr  mutex */
589 	l2arc_dev_t	*b_dev;			/* L2ARC device */
590 	daddr_t		b_daddr;		/* disk address, offset byte */
591 };
592 
593 typedef struct l2arc_data_free {
594 	/* protected by l2arc_free_on_write_mtx */
595 	void		*l2df_data;
596 	size_t		l2df_size;
597 	void		(*l2df_func)(void *, size_t);
598 	list_node_t	l2df_list_node;
599 } l2arc_data_free_t;
600 
601 static kmutex_t l2arc_feed_thr_lock;
602 static kcondvar_t l2arc_feed_thr_cv;
603 static uint8_t l2arc_thread_exit;
604 
605 static void l2arc_read_done(zio_t *zio);
606 static void l2arc_hdr_stat_add(void);
607 static void l2arc_hdr_stat_remove(void);
608 
609 static uint64_t
610 buf_hash(spa_t *spa, const dva_t *dva, uint64_t birth)
611 {
612 	uintptr_t spav = (uintptr_t)spa;
613 	uint8_t *vdva = (uint8_t *)dva;
614 	uint64_t crc = -1ULL;
615 	int i;
616 
617 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
618 
619 	for (i = 0; i < sizeof (dva_t); i++)
620 		crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
621 
622 	crc ^= (spav>>8) ^ birth;
623 
624 	return (crc);
625 }
626 
627 #define	BUF_EMPTY(buf)						\
628 	((buf)->b_dva.dva_word[0] == 0 &&			\
629 	(buf)->b_dva.dva_word[1] == 0 &&			\
630 	(buf)->b_birth == 0)
631 
632 #define	BUF_EQUAL(spa, dva, birth, buf)				\
633 	((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) &&	\
634 	((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) &&	\
635 	((buf)->b_birth == birth) && ((buf)->b_spa == spa)
636 
637 static arc_buf_hdr_t *
638 buf_hash_find(spa_t *spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
639 {
640 	uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
641 	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
642 	arc_buf_hdr_t *buf;
643 
644 	mutex_enter(hash_lock);
645 	for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
646 	    buf = buf->b_hash_next) {
647 		if (BUF_EQUAL(spa, dva, birth, buf)) {
648 			*lockp = hash_lock;
649 			return (buf);
650 		}
651 	}
652 	mutex_exit(hash_lock);
653 	*lockp = NULL;
654 	return (NULL);
655 }
656 
657 /*
658  * Insert an entry into the hash table.  If there is already an element
659  * equal to elem in the hash table, then the already existing element
660  * will be returned and the new element will not be inserted.
661  * Otherwise returns NULL.
662  */
663 static arc_buf_hdr_t *
664 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
665 {
666 	uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
667 	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
668 	arc_buf_hdr_t *fbuf;
669 	uint32_t i;
670 
671 	ASSERT(!HDR_IN_HASH_TABLE(buf));
672 	*lockp = hash_lock;
673 	mutex_enter(hash_lock);
674 	for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
675 	    fbuf = fbuf->b_hash_next, i++) {
676 		if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
677 			return (fbuf);
678 	}
679 
680 	buf->b_hash_next = buf_hash_table.ht_table[idx];
681 	buf_hash_table.ht_table[idx] = buf;
682 	buf->b_flags |= ARC_IN_HASH_TABLE;
683 
684 	/* collect some hash table performance data */
685 	if (i > 0) {
686 		ARCSTAT_BUMP(arcstat_hash_collisions);
687 		if (i == 1)
688 			ARCSTAT_BUMP(arcstat_hash_chains);
689 
690 		ARCSTAT_MAX(arcstat_hash_chain_max, i);
691 	}
692 
693 	ARCSTAT_BUMP(arcstat_hash_elements);
694 	ARCSTAT_MAXSTAT(arcstat_hash_elements);
695 
696 	return (NULL);
697 }
698 
699 static void
700 buf_hash_remove(arc_buf_hdr_t *buf)
701 {
702 	arc_buf_hdr_t *fbuf, **bufp;
703 	uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
704 
705 	ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
706 	ASSERT(HDR_IN_HASH_TABLE(buf));
707 
708 	bufp = &buf_hash_table.ht_table[idx];
709 	while ((fbuf = *bufp) != buf) {
710 		ASSERT(fbuf != NULL);
711 		bufp = &fbuf->b_hash_next;
712 	}
713 	*bufp = buf->b_hash_next;
714 	buf->b_hash_next = NULL;
715 	buf->b_flags &= ~ARC_IN_HASH_TABLE;
716 
717 	/* collect some hash table performance data */
718 	ARCSTAT_BUMPDOWN(arcstat_hash_elements);
719 
720 	if (buf_hash_table.ht_table[idx] &&
721 	    buf_hash_table.ht_table[idx]->b_hash_next == NULL)
722 		ARCSTAT_BUMPDOWN(arcstat_hash_chains);
723 }
724 
725 /*
726  * Global data structures and functions for the buf kmem cache.
727  */
728 static kmem_cache_t *hdr_cache;
729 static kmem_cache_t *buf_cache;
730 
731 static void
732 buf_fini(void)
733 {
734 	int i;
735 
736 	kmem_free(buf_hash_table.ht_table,
737 	    (buf_hash_table.ht_mask + 1) * sizeof (void *));
738 	for (i = 0; i < BUF_LOCKS; i++)
739 		mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
740 	kmem_cache_destroy(hdr_cache);
741 	kmem_cache_destroy(buf_cache);
742 }
743 
744 /*
745  * Constructor callback - called when the cache is empty
746  * and a new buf is requested.
747  */
748 /* ARGSUSED */
749 static int
750 hdr_cons(void *vbuf, void *unused, int kmflag)
751 {
752 	arc_buf_hdr_t *buf = vbuf;
753 
754 	bzero(buf, sizeof (arc_buf_hdr_t));
755 	refcount_create(&buf->b_refcnt);
756 	cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
757 	mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
758 
759 	ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
760 	return (0);
761 }
762 
763 /* ARGSUSED */
764 static int
765 buf_cons(void *vbuf, void *unused, int kmflag)
766 {
767 	arc_buf_t *buf = vbuf;
768 
769 	bzero(buf, sizeof (arc_buf_t));
770 	rw_init(&buf->b_lock, NULL, RW_DEFAULT, NULL);
771 	return (0);
772 }
773 
774 /*
775  * Destructor callback - called when a cached buf is
776  * no longer required.
777  */
778 /* ARGSUSED */
779 static void
780 hdr_dest(void *vbuf, void *unused)
781 {
782 	arc_buf_hdr_t *buf = vbuf;
783 
784 	refcount_destroy(&buf->b_refcnt);
785 	cv_destroy(&buf->b_cv);
786 	mutex_destroy(&buf->b_freeze_lock);
787 
788 	ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
789 }
790 
791 /* ARGSUSED */
792 static void
793 buf_dest(void *vbuf, void *unused)
794 {
795 	arc_buf_t *buf = vbuf;
796 
797 	rw_destroy(&buf->b_lock);
798 }
799 
800 /*
801  * Reclaim callback -- invoked when memory is low.
802  */
803 /* ARGSUSED */
804 static void
805 hdr_recl(void *unused)
806 {
807 	dprintf("hdr_recl called\n");
808 	/*
809 	 * umem calls the reclaim func when we destroy the buf cache,
810 	 * which is after we do arc_fini().
811 	 */
812 	if (!arc_dead)
813 		cv_signal(&arc_reclaim_thr_cv);
814 }
815 
816 static void
817 buf_init(void)
818 {
819 	uint64_t *ct;
820 	uint64_t hsize = 1ULL << 12;
821 	int i, j;
822 
823 	/*
824 	 * The hash table is big enough to fill all of physical memory
825 	 * with an average 64K block size.  The table will take up
826 	 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
827 	 */
828 	while (hsize * 65536 < physmem * PAGESIZE)
829 		hsize <<= 1;
830 retry:
831 	buf_hash_table.ht_mask = hsize - 1;
832 	buf_hash_table.ht_table =
833 	    kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
834 	if (buf_hash_table.ht_table == NULL) {
835 		ASSERT(hsize > (1ULL << 8));
836 		hsize >>= 1;
837 		goto retry;
838 	}
839 
840 	hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
841 	    0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
842 	buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
843 	    0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
844 
845 	for (i = 0; i < 256; i++)
846 		for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
847 			*ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
848 
849 	for (i = 0; i < BUF_LOCKS; i++) {
850 		mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
851 		    NULL, MUTEX_DEFAULT, NULL);
852 	}
853 }
854 
855 #define	ARC_MINTIME	(hz>>4) /* 62 ms */
856 
857 static void
858 arc_cksum_verify(arc_buf_t *buf)
859 {
860 	zio_cksum_t zc;
861 
862 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
863 		return;
864 
865 	mutex_enter(&buf->b_hdr->b_freeze_lock);
866 	if (buf->b_hdr->b_freeze_cksum == NULL ||
867 	    (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
868 		mutex_exit(&buf->b_hdr->b_freeze_lock);
869 		return;
870 	}
871 	fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
872 	if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
873 		panic("buffer modified while frozen!");
874 	mutex_exit(&buf->b_hdr->b_freeze_lock);
875 }
876 
877 static int
878 arc_cksum_equal(arc_buf_t *buf)
879 {
880 	zio_cksum_t zc;
881 	int equal;
882 
883 	mutex_enter(&buf->b_hdr->b_freeze_lock);
884 	fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
885 	equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
886 	mutex_exit(&buf->b_hdr->b_freeze_lock);
887 
888 	return (equal);
889 }
890 
891 static void
892 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
893 {
894 	if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
895 		return;
896 
897 	mutex_enter(&buf->b_hdr->b_freeze_lock);
898 	if (buf->b_hdr->b_freeze_cksum != NULL) {
899 		mutex_exit(&buf->b_hdr->b_freeze_lock);
900 		return;
901 	}
902 	buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
903 	fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
904 	    buf->b_hdr->b_freeze_cksum);
905 	mutex_exit(&buf->b_hdr->b_freeze_lock);
906 }
907 
908 void
909 arc_buf_thaw(arc_buf_t *buf)
910 {
911 	if (zfs_flags & ZFS_DEBUG_MODIFY) {
912 		if (buf->b_hdr->b_state != arc_anon)
913 			panic("modifying non-anon buffer!");
914 		if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
915 			panic("modifying buffer while i/o in progress!");
916 		arc_cksum_verify(buf);
917 	}
918 
919 	mutex_enter(&buf->b_hdr->b_freeze_lock);
920 	if (buf->b_hdr->b_freeze_cksum != NULL) {
921 		kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
922 		buf->b_hdr->b_freeze_cksum = NULL;
923 	}
924 	mutex_exit(&buf->b_hdr->b_freeze_lock);
925 }
926 
927 void
928 arc_buf_freeze(arc_buf_t *buf)
929 {
930 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
931 		return;
932 
933 	ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
934 	    buf->b_hdr->b_state == arc_anon);
935 	arc_cksum_compute(buf, B_FALSE);
936 }
937 
938 static void
939 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
940 {
941 	ASSERT(MUTEX_HELD(hash_lock));
942 
943 	if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
944 	    (ab->b_state != arc_anon)) {
945 		uint64_t delta = ab->b_size * ab->b_datacnt;
946 		list_t *list = &ab->b_state->arcs_list[ab->b_type];
947 		uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
948 
949 		ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
950 		mutex_enter(&ab->b_state->arcs_mtx);
951 		ASSERT(list_link_active(&ab->b_arc_node));
952 		list_remove(list, ab);
953 		if (GHOST_STATE(ab->b_state)) {
954 			ASSERT3U(ab->b_datacnt, ==, 0);
955 			ASSERT3P(ab->b_buf, ==, NULL);
956 			delta = ab->b_size;
957 		}
958 		ASSERT(delta > 0);
959 		ASSERT3U(*size, >=, delta);
960 		atomic_add_64(size, -delta);
961 		mutex_exit(&ab->b_state->arcs_mtx);
962 		/* remove the prefetch flag if we get a reference */
963 		if (ab->b_flags & ARC_PREFETCH)
964 			ab->b_flags &= ~ARC_PREFETCH;
965 	}
966 }
967 
968 static int
969 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
970 {
971 	int cnt;
972 	arc_state_t *state = ab->b_state;
973 
974 	ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
975 	ASSERT(!GHOST_STATE(state));
976 
977 	if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
978 	    (state != arc_anon)) {
979 		uint64_t *size = &state->arcs_lsize[ab->b_type];
980 
981 		ASSERT(!MUTEX_HELD(&state->arcs_mtx));
982 		mutex_enter(&state->arcs_mtx);
983 		ASSERT(!list_link_active(&ab->b_arc_node));
984 		list_insert_head(&state->arcs_list[ab->b_type], ab);
985 		ASSERT(ab->b_datacnt > 0);
986 		atomic_add_64(size, ab->b_size * ab->b_datacnt);
987 		mutex_exit(&state->arcs_mtx);
988 	}
989 	return (cnt);
990 }
991 
992 /*
993  * Move the supplied buffer to the indicated state.  The mutex
994  * for the buffer must be held by the caller.
995  */
996 static void
997 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
998 {
999 	arc_state_t *old_state = ab->b_state;
1000 	int64_t refcnt = refcount_count(&ab->b_refcnt);
1001 	uint64_t from_delta, to_delta;
1002 
1003 	ASSERT(MUTEX_HELD(hash_lock));
1004 	ASSERT(new_state != old_state);
1005 	ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1006 	ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1007 
1008 	from_delta = to_delta = ab->b_datacnt * ab->b_size;
1009 
1010 	/*
1011 	 * If this buffer is evictable, transfer it from the
1012 	 * old state list to the new state list.
1013 	 */
1014 	if (refcnt == 0) {
1015 		if (old_state != arc_anon) {
1016 			int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1017 			uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1018 
1019 			if (use_mutex)
1020 				mutex_enter(&old_state->arcs_mtx);
1021 
1022 			ASSERT(list_link_active(&ab->b_arc_node));
1023 			list_remove(&old_state->arcs_list[ab->b_type], ab);
1024 
1025 			/*
1026 			 * If prefetching out of the ghost cache,
1027 			 * we will have a non-null datacnt.
1028 			 */
1029 			if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1030 				/* ghost elements have a ghost size */
1031 				ASSERT(ab->b_buf == NULL);
1032 				from_delta = ab->b_size;
1033 			}
1034 			ASSERT3U(*size, >=, from_delta);
1035 			atomic_add_64(size, -from_delta);
1036 
1037 			if (use_mutex)
1038 				mutex_exit(&old_state->arcs_mtx);
1039 		}
1040 		if (new_state != arc_anon) {
1041 			int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1042 			uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1043 
1044 			if (use_mutex)
1045 				mutex_enter(&new_state->arcs_mtx);
1046 
1047 			list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1048 
1049 			/* ghost elements have a ghost size */
1050 			if (GHOST_STATE(new_state)) {
1051 				ASSERT(ab->b_datacnt == 0);
1052 				ASSERT(ab->b_buf == NULL);
1053 				to_delta = ab->b_size;
1054 			}
1055 			atomic_add_64(size, to_delta);
1056 
1057 			if (use_mutex)
1058 				mutex_exit(&new_state->arcs_mtx);
1059 		}
1060 	}
1061 
1062 	ASSERT(!BUF_EMPTY(ab));
1063 	if (new_state == arc_anon) {
1064 		buf_hash_remove(ab);
1065 	}
1066 
1067 	/* adjust state sizes */
1068 	if (to_delta)
1069 		atomic_add_64(&new_state->arcs_size, to_delta);
1070 	if (from_delta) {
1071 		ASSERT3U(old_state->arcs_size, >=, from_delta);
1072 		atomic_add_64(&old_state->arcs_size, -from_delta);
1073 	}
1074 	ab->b_state = new_state;
1075 
1076 	/* adjust l2arc hdr stats */
1077 	if (new_state == arc_l2c_only)
1078 		l2arc_hdr_stat_add();
1079 	else if (old_state == arc_l2c_only)
1080 		l2arc_hdr_stat_remove();
1081 }
1082 
1083 void
1084 arc_space_consume(uint64_t space)
1085 {
1086 	atomic_add_64(&arc_meta_used, space);
1087 	atomic_add_64(&arc_size, space);
1088 }
1089 
1090 void
1091 arc_space_return(uint64_t space)
1092 {
1093 	ASSERT(arc_meta_used >= space);
1094 	if (arc_meta_max < arc_meta_used)
1095 		arc_meta_max = arc_meta_used;
1096 	atomic_add_64(&arc_meta_used, -space);
1097 	ASSERT(arc_size >= space);
1098 	atomic_add_64(&arc_size, -space);
1099 }
1100 
1101 void *
1102 arc_data_buf_alloc(uint64_t size)
1103 {
1104 	if (arc_evict_needed(ARC_BUFC_DATA))
1105 		cv_signal(&arc_reclaim_thr_cv);
1106 	atomic_add_64(&arc_size, size);
1107 	return (zio_data_buf_alloc(size));
1108 }
1109 
1110 void
1111 arc_data_buf_free(void *buf, uint64_t size)
1112 {
1113 	zio_data_buf_free(buf, size);
1114 	ASSERT(arc_size >= size);
1115 	atomic_add_64(&arc_size, -size);
1116 }
1117 
1118 arc_buf_t *
1119 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1120 {
1121 	arc_buf_hdr_t *hdr;
1122 	arc_buf_t *buf;
1123 
1124 	ASSERT3U(size, >, 0);
1125 	hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1126 	ASSERT(BUF_EMPTY(hdr));
1127 	hdr->b_size = size;
1128 	hdr->b_type = type;
1129 	hdr->b_spa = spa;
1130 	hdr->b_state = arc_anon;
1131 	hdr->b_arc_access = 0;
1132 	buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1133 	buf->b_hdr = hdr;
1134 	buf->b_data = NULL;
1135 	buf->b_efunc = NULL;
1136 	buf->b_private = NULL;
1137 	buf->b_next = NULL;
1138 	hdr->b_buf = buf;
1139 	arc_get_data_buf(buf);
1140 	hdr->b_datacnt = 1;
1141 	hdr->b_flags = 0;
1142 	ASSERT(refcount_is_zero(&hdr->b_refcnt));
1143 	(void) refcount_add(&hdr->b_refcnt, tag);
1144 
1145 	return (buf);
1146 }
1147 
1148 static arc_buf_t *
1149 arc_buf_clone(arc_buf_t *from)
1150 {
1151 	arc_buf_t *buf;
1152 	arc_buf_hdr_t *hdr = from->b_hdr;
1153 	uint64_t size = hdr->b_size;
1154 
1155 	buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1156 	buf->b_hdr = hdr;
1157 	buf->b_data = NULL;
1158 	buf->b_efunc = NULL;
1159 	buf->b_private = NULL;
1160 	buf->b_next = hdr->b_buf;
1161 	hdr->b_buf = buf;
1162 	arc_get_data_buf(buf);
1163 	bcopy(from->b_data, buf->b_data, size);
1164 	hdr->b_datacnt += 1;
1165 	return (buf);
1166 }
1167 
1168 void
1169 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1170 {
1171 	arc_buf_hdr_t *hdr;
1172 	kmutex_t *hash_lock;
1173 
1174 	/*
1175 	 * Check to see if this buffer is evicted.  Callers
1176 	 * must verify b_data != NULL to know if the add_ref
1177 	 * was successful.
1178 	 */
1179 	rw_enter(&buf->b_lock, RW_READER);
1180 	if (buf->b_data == NULL) {
1181 		rw_exit(&buf->b_lock);
1182 		return;
1183 	}
1184 	hdr = buf->b_hdr;
1185 	ASSERT(hdr != NULL);
1186 	hash_lock = HDR_LOCK(hdr);
1187 	mutex_enter(hash_lock);
1188 	rw_exit(&buf->b_lock);
1189 
1190 	ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1191 	add_reference(hdr, hash_lock, tag);
1192 	arc_access(hdr, hash_lock);
1193 	mutex_exit(hash_lock);
1194 	ARCSTAT_BUMP(arcstat_hits);
1195 	ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1196 	    demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1197 	    data, metadata, hits);
1198 }
1199 
1200 /*
1201  * Free the arc data buffer.  If it is an l2arc write in progress,
1202  * the buffer is placed on l2arc_free_on_write to be freed later.
1203  */
1204 static void
1205 arc_buf_data_free(arc_buf_hdr_t *hdr, void (*free_func)(void *, size_t),
1206     void *data, size_t size)
1207 {
1208 	if (HDR_L2_WRITING(hdr)) {
1209 		l2arc_data_free_t *df;
1210 		df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1211 		df->l2df_data = data;
1212 		df->l2df_size = size;
1213 		df->l2df_func = free_func;
1214 		mutex_enter(&l2arc_free_on_write_mtx);
1215 		list_insert_head(l2arc_free_on_write, df);
1216 		mutex_exit(&l2arc_free_on_write_mtx);
1217 		ARCSTAT_BUMP(arcstat_l2_free_on_write);
1218 	} else {
1219 		free_func(data, size);
1220 	}
1221 }
1222 
1223 static void
1224 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1225 {
1226 	arc_buf_t **bufp;
1227 
1228 	/* free up data associated with the buf */
1229 	if (buf->b_data) {
1230 		arc_state_t *state = buf->b_hdr->b_state;
1231 		uint64_t size = buf->b_hdr->b_size;
1232 		arc_buf_contents_t type = buf->b_hdr->b_type;
1233 
1234 		arc_cksum_verify(buf);
1235 		if (!recycle) {
1236 			if (type == ARC_BUFC_METADATA) {
1237 				arc_buf_data_free(buf->b_hdr, zio_buf_free,
1238 				    buf->b_data, size);
1239 				arc_space_return(size);
1240 			} else {
1241 				ASSERT(type == ARC_BUFC_DATA);
1242 				arc_buf_data_free(buf->b_hdr,
1243 				    zio_data_buf_free, buf->b_data, size);
1244 				atomic_add_64(&arc_size, -size);
1245 			}
1246 		}
1247 		if (list_link_active(&buf->b_hdr->b_arc_node)) {
1248 			uint64_t *cnt = &state->arcs_lsize[type];
1249 
1250 			ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1251 			ASSERT(state != arc_anon);
1252 
1253 			ASSERT3U(*cnt, >=, size);
1254 			atomic_add_64(cnt, -size);
1255 		}
1256 		ASSERT3U(state->arcs_size, >=, size);
1257 		atomic_add_64(&state->arcs_size, -size);
1258 		buf->b_data = NULL;
1259 		ASSERT(buf->b_hdr->b_datacnt > 0);
1260 		buf->b_hdr->b_datacnt -= 1;
1261 	}
1262 
1263 	/* only remove the buf if requested */
1264 	if (!all)
1265 		return;
1266 
1267 	/* remove the buf from the hdr list */
1268 	for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1269 		continue;
1270 	*bufp = buf->b_next;
1271 
1272 	ASSERT(buf->b_efunc == NULL);
1273 
1274 	/* clean up the buf */
1275 	buf->b_hdr = NULL;
1276 	kmem_cache_free(buf_cache, buf);
1277 }
1278 
1279 static void
1280 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1281 {
1282 	ASSERT(refcount_is_zero(&hdr->b_refcnt));
1283 	ASSERT3P(hdr->b_state, ==, arc_anon);
1284 	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1285 	ASSERT(!(hdr->b_flags & ARC_STORED));
1286 
1287 	if (hdr->b_l2hdr != NULL) {
1288 		if (!MUTEX_HELD(&l2arc_buflist_mtx)) {
1289 			/*
1290 			 * To prevent arc_free() and l2arc_evict() from
1291 			 * attempting to free the same buffer at the same time,
1292 			 * a FREE_IN_PROGRESS flag is given to arc_free() to
1293 			 * give it priority.  l2arc_evict() can't destroy this
1294 			 * header while we are waiting on l2arc_buflist_mtx.
1295 			 *
1296 			 * The hdr may be removed from l2ad_buflist before we
1297 			 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1298 			 */
1299 			mutex_enter(&l2arc_buflist_mtx);
1300 			if (hdr->b_l2hdr != NULL) {
1301 				list_remove(hdr->b_l2hdr->b_dev->l2ad_buflist,
1302 				    hdr);
1303 			}
1304 			mutex_exit(&l2arc_buflist_mtx);
1305 		} else {
1306 			list_remove(hdr->b_l2hdr->b_dev->l2ad_buflist, hdr);
1307 		}
1308 		ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1309 		kmem_free(hdr->b_l2hdr, sizeof (l2arc_buf_hdr_t));
1310 		if (hdr->b_state == arc_l2c_only)
1311 			l2arc_hdr_stat_remove();
1312 		hdr->b_l2hdr = NULL;
1313 	}
1314 
1315 	if (!BUF_EMPTY(hdr)) {
1316 		ASSERT(!HDR_IN_HASH_TABLE(hdr));
1317 		bzero(&hdr->b_dva, sizeof (dva_t));
1318 		hdr->b_birth = 0;
1319 		hdr->b_cksum0 = 0;
1320 	}
1321 	while (hdr->b_buf) {
1322 		arc_buf_t *buf = hdr->b_buf;
1323 
1324 		if (buf->b_efunc) {
1325 			mutex_enter(&arc_eviction_mtx);
1326 			rw_enter(&buf->b_lock, RW_WRITER);
1327 			ASSERT(buf->b_hdr != NULL);
1328 			arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1329 			hdr->b_buf = buf->b_next;
1330 			buf->b_hdr = &arc_eviction_hdr;
1331 			buf->b_next = arc_eviction_list;
1332 			arc_eviction_list = buf;
1333 			rw_exit(&buf->b_lock);
1334 			mutex_exit(&arc_eviction_mtx);
1335 		} else {
1336 			arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1337 		}
1338 	}
1339 	if (hdr->b_freeze_cksum != NULL) {
1340 		kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1341 		hdr->b_freeze_cksum = NULL;
1342 	}
1343 
1344 	ASSERT(!list_link_active(&hdr->b_arc_node));
1345 	ASSERT3P(hdr->b_hash_next, ==, NULL);
1346 	ASSERT3P(hdr->b_acb, ==, NULL);
1347 	kmem_cache_free(hdr_cache, hdr);
1348 }
1349 
1350 void
1351 arc_buf_free(arc_buf_t *buf, void *tag)
1352 {
1353 	arc_buf_hdr_t *hdr = buf->b_hdr;
1354 	int hashed = hdr->b_state != arc_anon;
1355 
1356 	ASSERT(buf->b_efunc == NULL);
1357 	ASSERT(buf->b_data != NULL);
1358 
1359 	if (hashed) {
1360 		kmutex_t *hash_lock = HDR_LOCK(hdr);
1361 
1362 		mutex_enter(hash_lock);
1363 		(void) remove_reference(hdr, hash_lock, tag);
1364 		if (hdr->b_datacnt > 1)
1365 			arc_buf_destroy(buf, FALSE, TRUE);
1366 		else
1367 			hdr->b_flags |= ARC_BUF_AVAILABLE;
1368 		mutex_exit(hash_lock);
1369 	} else if (HDR_IO_IN_PROGRESS(hdr)) {
1370 		int destroy_hdr;
1371 		/*
1372 		 * We are in the middle of an async write.  Don't destroy
1373 		 * this buffer unless the write completes before we finish
1374 		 * decrementing the reference count.
1375 		 */
1376 		mutex_enter(&arc_eviction_mtx);
1377 		(void) remove_reference(hdr, NULL, tag);
1378 		ASSERT(refcount_is_zero(&hdr->b_refcnt));
1379 		destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1380 		mutex_exit(&arc_eviction_mtx);
1381 		if (destroy_hdr)
1382 			arc_hdr_destroy(hdr);
1383 	} else {
1384 		if (remove_reference(hdr, NULL, tag) > 0) {
1385 			ASSERT(HDR_IO_ERROR(hdr));
1386 			arc_buf_destroy(buf, FALSE, TRUE);
1387 		} else {
1388 			arc_hdr_destroy(hdr);
1389 		}
1390 	}
1391 }
1392 
1393 int
1394 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1395 {
1396 	arc_buf_hdr_t *hdr = buf->b_hdr;
1397 	kmutex_t *hash_lock = HDR_LOCK(hdr);
1398 	int no_callback = (buf->b_efunc == NULL);
1399 
1400 	if (hdr->b_state == arc_anon) {
1401 		arc_buf_free(buf, tag);
1402 		return (no_callback);
1403 	}
1404 
1405 	mutex_enter(hash_lock);
1406 	ASSERT(hdr->b_state != arc_anon);
1407 	ASSERT(buf->b_data != NULL);
1408 
1409 	(void) remove_reference(hdr, hash_lock, tag);
1410 	if (hdr->b_datacnt > 1) {
1411 		if (no_callback)
1412 			arc_buf_destroy(buf, FALSE, TRUE);
1413 	} else if (no_callback) {
1414 		ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1415 		hdr->b_flags |= ARC_BUF_AVAILABLE;
1416 	}
1417 	ASSERT(no_callback || hdr->b_datacnt > 1 ||
1418 	    refcount_is_zero(&hdr->b_refcnt));
1419 	mutex_exit(hash_lock);
1420 	return (no_callback);
1421 }
1422 
1423 int
1424 arc_buf_size(arc_buf_t *buf)
1425 {
1426 	return (buf->b_hdr->b_size);
1427 }
1428 
1429 /*
1430  * Evict buffers from list until we've removed the specified number of
1431  * bytes.  Move the removed buffers to the appropriate evict state.
1432  * If the recycle flag is set, then attempt to "recycle" a buffer:
1433  * - look for a buffer to evict that is `bytes' long.
1434  * - return the data block from this buffer rather than freeing it.
1435  * This flag is used by callers that are trying to make space for a
1436  * new buffer in a full arc cache.
1437  *
1438  * This function makes a "best effort".  It skips over any buffers
1439  * it can't get a hash_lock on, and so may not catch all candidates.
1440  * It may also return without evicting as much space as requested.
1441  */
1442 static void *
1443 arc_evict(arc_state_t *state, spa_t *spa, int64_t bytes, boolean_t recycle,
1444     arc_buf_contents_t type)
1445 {
1446 	arc_state_t *evicted_state;
1447 	uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1448 	arc_buf_hdr_t *ab, *ab_prev = NULL;
1449 	list_t *list = &state->arcs_list[type];
1450 	kmutex_t *hash_lock;
1451 	boolean_t have_lock;
1452 	void *stolen = NULL;
1453 
1454 	ASSERT(state == arc_mru || state == arc_mfu);
1455 
1456 	evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1457 
1458 	mutex_enter(&state->arcs_mtx);
1459 	mutex_enter(&evicted_state->arcs_mtx);
1460 
1461 	for (ab = list_tail(list); ab; ab = ab_prev) {
1462 		ab_prev = list_prev(list, ab);
1463 		/* prefetch buffers have a minimum lifespan */
1464 		if (HDR_IO_IN_PROGRESS(ab) ||
1465 		    (spa && ab->b_spa != spa) ||
1466 		    (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1467 		    lbolt - ab->b_arc_access < arc_min_prefetch_lifespan)) {
1468 			skipped++;
1469 			continue;
1470 		}
1471 		/* "lookahead" for better eviction candidate */
1472 		if (recycle && ab->b_size != bytes &&
1473 		    ab_prev && ab_prev->b_size == bytes)
1474 			continue;
1475 		hash_lock = HDR_LOCK(ab);
1476 		have_lock = MUTEX_HELD(hash_lock);
1477 		if (have_lock || mutex_tryenter(hash_lock)) {
1478 			ASSERT3U(refcount_count(&ab->b_refcnt), ==, 0);
1479 			ASSERT(ab->b_datacnt > 0);
1480 			while (ab->b_buf) {
1481 				arc_buf_t *buf = ab->b_buf;
1482 				if (!rw_tryenter(&buf->b_lock, RW_WRITER)) {
1483 					missed += 1;
1484 					break;
1485 				}
1486 				if (buf->b_data) {
1487 					bytes_evicted += ab->b_size;
1488 					if (recycle && ab->b_type == type &&
1489 					    ab->b_size == bytes &&
1490 					    !HDR_L2_WRITING(ab)) {
1491 						stolen = buf->b_data;
1492 						recycle = FALSE;
1493 					}
1494 				}
1495 				if (buf->b_efunc) {
1496 					mutex_enter(&arc_eviction_mtx);
1497 					arc_buf_destroy(buf,
1498 					    buf->b_data == stolen, FALSE);
1499 					ab->b_buf = buf->b_next;
1500 					buf->b_hdr = &arc_eviction_hdr;
1501 					buf->b_next = arc_eviction_list;
1502 					arc_eviction_list = buf;
1503 					mutex_exit(&arc_eviction_mtx);
1504 					rw_exit(&buf->b_lock);
1505 				} else {
1506 					rw_exit(&buf->b_lock);
1507 					arc_buf_destroy(buf,
1508 					    buf->b_data == stolen, TRUE);
1509 				}
1510 			}
1511 			if (ab->b_datacnt == 0) {
1512 				arc_change_state(evicted_state, ab, hash_lock);
1513 				ASSERT(HDR_IN_HASH_TABLE(ab));
1514 				ab->b_flags |= ARC_IN_HASH_TABLE;
1515 				ab->b_flags &= ~ARC_BUF_AVAILABLE;
1516 				DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1517 			}
1518 			if (!have_lock)
1519 				mutex_exit(hash_lock);
1520 			if (bytes >= 0 && bytes_evicted >= bytes)
1521 				break;
1522 		} else {
1523 			missed += 1;
1524 		}
1525 	}
1526 
1527 	mutex_exit(&evicted_state->arcs_mtx);
1528 	mutex_exit(&state->arcs_mtx);
1529 
1530 	if (bytes_evicted < bytes)
1531 		dprintf("only evicted %lld bytes from %x",
1532 		    (longlong_t)bytes_evicted, state);
1533 
1534 	if (skipped)
1535 		ARCSTAT_INCR(arcstat_evict_skip, skipped);
1536 
1537 	if (missed)
1538 		ARCSTAT_INCR(arcstat_mutex_miss, missed);
1539 
1540 	/*
1541 	 * We have just evicted some date into the ghost state, make
1542 	 * sure we also adjust the ghost state size if necessary.
1543 	 */
1544 	if (arc_no_grow &&
1545 	    arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1546 		int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1547 		    arc_mru_ghost->arcs_size - arc_c;
1548 
1549 		if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1550 			int64_t todelete =
1551 			    MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1552 			arc_evict_ghost(arc_mru_ghost, NULL, todelete);
1553 		} else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1554 			int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1555 			    arc_mru_ghost->arcs_size +
1556 			    arc_mfu_ghost->arcs_size - arc_c);
1557 			arc_evict_ghost(arc_mfu_ghost, NULL, todelete);
1558 		}
1559 	}
1560 
1561 	return (stolen);
1562 }
1563 
1564 /*
1565  * Remove buffers from list until we've removed the specified number of
1566  * bytes.  Destroy the buffers that are removed.
1567  */
1568 static void
1569 arc_evict_ghost(arc_state_t *state, spa_t *spa, int64_t bytes)
1570 {
1571 	arc_buf_hdr_t *ab, *ab_prev;
1572 	list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1573 	kmutex_t *hash_lock;
1574 	uint64_t bytes_deleted = 0;
1575 	uint64_t bufs_skipped = 0;
1576 
1577 	ASSERT(GHOST_STATE(state));
1578 top:
1579 	mutex_enter(&state->arcs_mtx);
1580 	for (ab = list_tail(list); ab; ab = ab_prev) {
1581 		ab_prev = list_prev(list, ab);
1582 		if (spa && ab->b_spa != spa)
1583 			continue;
1584 		hash_lock = HDR_LOCK(ab);
1585 		if (mutex_tryenter(hash_lock)) {
1586 			ASSERT(!HDR_IO_IN_PROGRESS(ab));
1587 			ASSERT(ab->b_buf == NULL);
1588 			ARCSTAT_BUMP(arcstat_deleted);
1589 			bytes_deleted += ab->b_size;
1590 
1591 			if (ab->b_l2hdr != NULL) {
1592 				/*
1593 				 * This buffer is cached on the 2nd Level ARC;
1594 				 * don't destroy the header.
1595 				 */
1596 				arc_change_state(arc_l2c_only, ab, hash_lock);
1597 				mutex_exit(hash_lock);
1598 			} else {
1599 				arc_change_state(arc_anon, ab, hash_lock);
1600 				mutex_exit(hash_lock);
1601 				arc_hdr_destroy(ab);
1602 			}
1603 
1604 			DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1605 			if (bytes >= 0 && bytes_deleted >= bytes)
1606 				break;
1607 		} else {
1608 			if (bytes < 0) {
1609 				mutex_exit(&state->arcs_mtx);
1610 				mutex_enter(hash_lock);
1611 				mutex_exit(hash_lock);
1612 				goto top;
1613 			}
1614 			bufs_skipped += 1;
1615 		}
1616 	}
1617 	mutex_exit(&state->arcs_mtx);
1618 
1619 	if (list == &state->arcs_list[ARC_BUFC_DATA] &&
1620 	    (bytes < 0 || bytes_deleted < bytes)) {
1621 		list = &state->arcs_list[ARC_BUFC_METADATA];
1622 		goto top;
1623 	}
1624 
1625 	if (bufs_skipped) {
1626 		ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1627 		ASSERT(bytes >= 0);
1628 	}
1629 
1630 	if (bytes_deleted < bytes)
1631 		dprintf("only deleted %lld bytes from %p",
1632 		    (longlong_t)bytes_deleted, state);
1633 }
1634 
1635 static void
1636 arc_adjust(void)
1637 {
1638 	int64_t top_sz, mru_over, arc_over, todelete;
1639 
1640 	top_sz = arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used;
1641 
1642 	if (top_sz > arc_p && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
1643 		int64_t toevict =
1644 		    MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], top_sz - arc_p);
1645 		(void) arc_evict(arc_mru, NULL, toevict, FALSE, ARC_BUFC_DATA);
1646 		top_sz = arc_anon->arcs_size + arc_mru->arcs_size;
1647 	}
1648 
1649 	if (top_sz > arc_p && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1650 		int64_t toevict =
1651 		    MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], top_sz - arc_p);
1652 		(void) arc_evict(arc_mru, NULL, toevict, FALSE,
1653 		    ARC_BUFC_METADATA);
1654 		top_sz = arc_anon->arcs_size + arc_mru->arcs_size;
1655 	}
1656 
1657 	mru_over = top_sz + arc_mru_ghost->arcs_size - arc_c;
1658 
1659 	if (mru_over > 0) {
1660 		if (arc_mru_ghost->arcs_size > 0) {
1661 			todelete = MIN(arc_mru_ghost->arcs_size, mru_over);
1662 			arc_evict_ghost(arc_mru_ghost, NULL, todelete);
1663 		}
1664 	}
1665 
1666 	if ((arc_over = arc_size - arc_c) > 0) {
1667 		int64_t tbl_over;
1668 
1669 		if (arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
1670 			int64_t toevict =
1671 			    MIN(arc_mfu->arcs_lsize[ARC_BUFC_DATA], arc_over);
1672 			(void) arc_evict(arc_mfu, NULL, toevict, FALSE,
1673 			    ARC_BUFC_DATA);
1674 			arc_over = arc_size - arc_c;
1675 		}
1676 
1677 		if (arc_over > 0 &&
1678 		    arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1679 			int64_t toevict =
1680 			    MIN(arc_mfu->arcs_lsize[ARC_BUFC_METADATA],
1681 			    arc_over);
1682 			(void) arc_evict(arc_mfu, NULL, toevict, FALSE,
1683 			    ARC_BUFC_METADATA);
1684 		}
1685 
1686 		tbl_over = arc_size + arc_mru_ghost->arcs_size +
1687 		    arc_mfu_ghost->arcs_size - arc_c * 2;
1688 
1689 		if (tbl_over > 0 && arc_mfu_ghost->arcs_size > 0) {
1690 			todelete = MIN(arc_mfu_ghost->arcs_size, tbl_over);
1691 			arc_evict_ghost(arc_mfu_ghost, NULL, todelete);
1692 		}
1693 	}
1694 }
1695 
1696 static void
1697 arc_do_user_evicts(void)
1698 {
1699 	mutex_enter(&arc_eviction_mtx);
1700 	while (arc_eviction_list != NULL) {
1701 		arc_buf_t *buf = arc_eviction_list;
1702 		arc_eviction_list = buf->b_next;
1703 		rw_enter(&buf->b_lock, RW_WRITER);
1704 		buf->b_hdr = NULL;
1705 		rw_exit(&buf->b_lock);
1706 		mutex_exit(&arc_eviction_mtx);
1707 
1708 		if (buf->b_efunc != NULL)
1709 			VERIFY(buf->b_efunc(buf) == 0);
1710 
1711 		buf->b_efunc = NULL;
1712 		buf->b_private = NULL;
1713 		kmem_cache_free(buf_cache, buf);
1714 		mutex_enter(&arc_eviction_mtx);
1715 	}
1716 	mutex_exit(&arc_eviction_mtx);
1717 }
1718 
1719 /*
1720  * Flush all *evictable* data from the cache for the given spa.
1721  * NOTE: this will not touch "active" (i.e. referenced) data.
1722  */
1723 void
1724 arc_flush(spa_t *spa)
1725 {
1726 	while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
1727 		(void) arc_evict(arc_mru, spa, -1, FALSE, ARC_BUFC_DATA);
1728 		if (spa)
1729 			break;
1730 	}
1731 	while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
1732 		(void) arc_evict(arc_mru, spa, -1, FALSE, ARC_BUFC_METADATA);
1733 		if (spa)
1734 			break;
1735 	}
1736 	while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
1737 		(void) arc_evict(arc_mfu, spa, -1, FALSE, ARC_BUFC_DATA);
1738 		if (spa)
1739 			break;
1740 	}
1741 	while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
1742 		(void) arc_evict(arc_mfu, spa, -1, FALSE, ARC_BUFC_METADATA);
1743 		if (spa)
1744 			break;
1745 	}
1746 
1747 	arc_evict_ghost(arc_mru_ghost, spa, -1);
1748 	arc_evict_ghost(arc_mfu_ghost, spa, -1);
1749 
1750 	mutex_enter(&arc_reclaim_thr_lock);
1751 	arc_do_user_evicts();
1752 	mutex_exit(&arc_reclaim_thr_lock);
1753 	ASSERT(spa || arc_eviction_list == NULL);
1754 }
1755 
1756 int arc_shrink_shift = 5;		/* log2(fraction of arc to reclaim) */
1757 
1758 void
1759 arc_shrink(void)
1760 {
1761 	if (arc_c > arc_c_min) {
1762 		uint64_t to_free;
1763 
1764 #ifdef _KERNEL
1765 		to_free = MAX(arc_c >> arc_shrink_shift, ptob(needfree));
1766 #else
1767 		to_free = arc_c >> arc_shrink_shift;
1768 #endif
1769 		if (arc_c > arc_c_min + to_free)
1770 			atomic_add_64(&arc_c, -to_free);
1771 		else
1772 			arc_c = arc_c_min;
1773 
1774 		atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
1775 		if (arc_c > arc_size)
1776 			arc_c = MAX(arc_size, arc_c_min);
1777 		if (arc_p > arc_c)
1778 			arc_p = (arc_c >> 1);
1779 		ASSERT(arc_c >= arc_c_min);
1780 		ASSERT((int64_t)arc_p >= 0);
1781 	}
1782 
1783 	if (arc_size > arc_c)
1784 		arc_adjust();
1785 }
1786 
1787 static int
1788 arc_reclaim_needed(void)
1789 {
1790 	uint64_t extra;
1791 
1792 #ifdef _KERNEL
1793 
1794 	if (needfree)
1795 		return (1);
1796 
1797 	/*
1798 	 * take 'desfree' extra pages, so we reclaim sooner, rather than later
1799 	 */
1800 	extra = desfree;
1801 
1802 	/*
1803 	 * check that we're out of range of the pageout scanner.  It starts to
1804 	 * schedule paging if freemem is less than lotsfree and needfree.
1805 	 * lotsfree is the high-water mark for pageout, and needfree is the
1806 	 * number of needed free pages.  We add extra pages here to make sure
1807 	 * the scanner doesn't start up while we're freeing memory.
1808 	 */
1809 	if (freemem < lotsfree + needfree + extra)
1810 		return (1);
1811 
1812 	/*
1813 	 * check to make sure that swapfs has enough space so that anon
1814 	 * reservations can still succeed. anon_resvmem() checks that the
1815 	 * availrmem is greater than swapfs_minfree, and the number of reserved
1816 	 * swap pages.  We also add a bit of extra here just to prevent
1817 	 * circumstances from getting really dire.
1818 	 */
1819 	if (availrmem < swapfs_minfree + swapfs_reserve + extra)
1820 		return (1);
1821 
1822 #if defined(__i386)
1823 	/*
1824 	 * If we're on an i386 platform, it's possible that we'll exhaust the
1825 	 * kernel heap space before we ever run out of available physical
1826 	 * memory.  Most checks of the size of the heap_area compare against
1827 	 * tune.t_minarmem, which is the minimum available real memory that we
1828 	 * can have in the system.  However, this is generally fixed at 25 pages
1829 	 * which is so low that it's useless.  In this comparison, we seek to
1830 	 * calculate the total heap-size, and reclaim if more than 3/4ths of the
1831 	 * heap is allocated.  (Or, in the calculation, if less than 1/4th is
1832 	 * free)
1833 	 */
1834 	if (btop(vmem_size(heap_arena, VMEM_FREE)) <
1835 	    (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2))
1836 		return (1);
1837 #endif
1838 
1839 #else
1840 	if (spa_get_random(100) == 0)
1841 		return (1);
1842 #endif
1843 	return (0);
1844 }
1845 
1846 static void
1847 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
1848 {
1849 	size_t			i;
1850 	kmem_cache_t		*prev_cache = NULL;
1851 	kmem_cache_t		*prev_data_cache = NULL;
1852 	extern kmem_cache_t	*zio_buf_cache[];
1853 	extern kmem_cache_t	*zio_data_buf_cache[];
1854 
1855 #ifdef _KERNEL
1856 	if (arc_meta_used >= arc_meta_limit) {
1857 		/*
1858 		 * We are exceeding our meta-data cache limit.
1859 		 * Purge some DNLC entries to release holds on meta-data.
1860 		 */
1861 		dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
1862 	}
1863 #if defined(__i386)
1864 	/*
1865 	 * Reclaim unused memory from all kmem caches.
1866 	 */
1867 	kmem_reap();
1868 #endif
1869 #endif
1870 
1871 	/*
1872 	 * An aggressive reclamation will shrink the cache size as well as
1873 	 * reap free buffers from the arc kmem caches.
1874 	 */
1875 	if (strat == ARC_RECLAIM_AGGR)
1876 		arc_shrink();
1877 
1878 	for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
1879 		if (zio_buf_cache[i] != prev_cache) {
1880 			prev_cache = zio_buf_cache[i];
1881 			kmem_cache_reap_now(zio_buf_cache[i]);
1882 		}
1883 		if (zio_data_buf_cache[i] != prev_data_cache) {
1884 			prev_data_cache = zio_data_buf_cache[i];
1885 			kmem_cache_reap_now(zio_data_buf_cache[i]);
1886 		}
1887 	}
1888 	kmem_cache_reap_now(buf_cache);
1889 	kmem_cache_reap_now(hdr_cache);
1890 }
1891 
1892 static void
1893 arc_reclaim_thread(void)
1894 {
1895 	clock_t			growtime = 0;
1896 	arc_reclaim_strategy_t	last_reclaim = ARC_RECLAIM_CONS;
1897 	callb_cpr_t		cpr;
1898 
1899 	CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
1900 
1901 	mutex_enter(&arc_reclaim_thr_lock);
1902 	while (arc_thread_exit == 0) {
1903 		if (arc_reclaim_needed()) {
1904 
1905 			if (arc_no_grow) {
1906 				if (last_reclaim == ARC_RECLAIM_CONS) {
1907 					last_reclaim = ARC_RECLAIM_AGGR;
1908 				} else {
1909 					last_reclaim = ARC_RECLAIM_CONS;
1910 				}
1911 			} else {
1912 				arc_no_grow = TRUE;
1913 				last_reclaim = ARC_RECLAIM_AGGR;
1914 				membar_producer();
1915 			}
1916 
1917 			/* reset the growth delay for every reclaim */
1918 			growtime = lbolt + (arc_grow_retry * hz);
1919 
1920 			arc_kmem_reap_now(last_reclaim);
1921 			arc_warm = B_TRUE;
1922 
1923 		} else if (arc_no_grow && lbolt >= growtime) {
1924 			arc_no_grow = FALSE;
1925 		}
1926 
1927 		if (2 * arc_c < arc_size +
1928 		    arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size)
1929 			arc_adjust();
1930 
1931 		if (arc_eviction_list != NULL)
1932 			arc_do_user_evicts();
1933 
1934 		/* block until needed, or one second, whichever is shorter */
1935 		CALLB_CPR_SAFE_BEGIN(&cpr);
1936 		(void) cv_timedwait(&arc_reclaim_thr_cv,
1937 		    &arc_reclaim_thr_lock, (lbolt + hz));
1938 		CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
1939 	}
1940 
1941 	arc_thread_exit = 0;
1942 	cv_broadcast(&arc_reclaim_thr_cv);
1943 	CALLB_CPR_EXIT(&cpr);		/* drops arc_reclaim_thr_lock */
1944 	thread_exit();
1945 }
1946 
1947 /*
1948  * Adapt arc info given the number of bytes we are trying to add and
1949  * the state that we are comming from.  This function is only called
1950  * when we are adding new content to the cache.
1951  */
1952 static void
1953 arc_adapt(int bytes, arc_state_t *state)
1954 {
1955 	int mult;
1956 
1957 	if (state == arc_l2c_only)
1958 		return;
1959 
1960 	ASSERT(bytes > 0);
1961 	/*
1962 	 * Adapt the target size of the MRU list:
1963 	 *	- if we just hit in the MRU ghost list, then increase
1964 	 *	  the target size of the MRU list.
1965 	 *	- if we just hit in the MFU ghost list, then increase
1966 	 *	  the target size of the MFU list by decreasing the
1967 	 *	  target size of the MRU list.
1968 	 */
1969 	if (state == arc_mru_ghost) {
1970 		mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
1971 		    1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
1972 
1973 		arc_p = MIN(arc_c, arc_p + bytes * mult);
1974 	} else if (state == arc_mfu_ghost) {
1975 		mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
1976 		    1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
1977 
1978 		arc_p = MAX(0, (int64_t)arc_p - bytes * mult);
1979 	}
1980 	ASSERT((int64_t)arc_p >= 0);
1981 
1982 	if (arc_reclaim_needed()) {
1983 		cv_signal(&arc_reclaim_thr_cv);
1984 		return;
1985 	}
1986 
1987 	if (arc_no_grow)
1988 		return;
1989 
1990 	if (arc_c >= arc_c_max)
1991 		return;
1992 
1993 	/*
1994 	 * If we're within (2 * maxblocksize) bytes of the target
1995 	 * cache size, increment the target cache size
1996 	 */
1997 	if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
1998 		atomic_add_64(&arc_c, (int64_t)bytes);
1999 		if (arc_c > arc_c_max)
2000 			arc_c = arc_c_max;
2001 		else if (state == arc_anon)
2002 			atomic_add_64(&arc_p, (int64_t)bytes);
2003 		if (arc_p > arc_c)
2004 			arc_p = arc_c;
2005 	}
2006 	ASSERT((int64_t)arc_p >= 0);
2007 }
2008 
2009 /*
2010  * Check if the cache has reached its limits and eviction is required
2011  * prior to insert.
2012  */
2013 static int
2014 arc_evict_needed(arc_buf_contents_t type)
2015 {
2016 	if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2017 		return (1);
2018 
2019 #ifdef _KERNEL
2020 	/*
2021 	 * If zio data pages are being allocated out of a separate heap segment,
2022 	 * then enforce that the size of available vmem for this area remains
2023 	 * above about 1/32nd free.
2024 	 */
2025 	if (type == ARC_BUFC_DATA && zio_arena != NULL &&
2026 	    vmem_size(zio_arena, VMEM_FREE) <
2027 	    (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
2028 		return (1);
2029 #endif
2030 
2031 	if (arc_reclaim_needed())
2032 		return (1);
2033 
2034 	return (arc_size > arc_c);
2035 }
2036 
2037 /*
2038  * The buffer, supplied as the first argument, needs a data block.
2039  * So, if we are at cache max, determine which cache should be victimized.
2040  * We have the following cases:
2041  *
2042  * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2043  * In this situation if we're out of space, but the resident size of the MFU is
2044  * under the limit, victimize the MFU cache to satisfy this insertion request.
2045  *
2046  * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2047  * Here, we've used up all of the available space for the MRU, so we need to
2048  * evict from our own cache instead.  Evict from the set of resident MRU
2049  * entries.
2050  *
2051  * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2052  * c minus p represents the MFU space in the cache, since p is the size of the
2053  * cache that is dedicated to the MRU.  In this situation there's still space on
2054  * the MFU side, so the MRU side needs to be victimized.
2055  *
2056  * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2057  * MFU's resident set is consuming more space than it has been allotted.  In
2058  * this situation, we must victimize our own cache, the MFU, for this insertion.
2059  */
2060 static void
2061 arc_get_data_buf(arc_buf_t *buf)
2062 {
2063 	arc_state_t		*state = buf->b_hdr->b_state;
2064 	uint64_t		size = buf->b_hdr->b_size;
2065 	arc_buf_contents_t	type = buf->b_hdr->b_type;
2066 
2067 	arc_adapt(size, state);
2068 
2069 	/*
2070 	 * We have not yet reached cache maximum size,
2071 	 * just allocate a new buffer.
2072 	 */
2073 	if (!arc_evict_needed(type)) {
2074 		if (type == ARC_BUFC_METADATA) {
2075 			buf->b_data = zio_buf_alloc(size);
2076 			arc_space_consume(size);
2077 		} else {
2078 			ASSERT(type == ARC_BUFC_DATA);
2079 			buf->b_data = zio_data_buf_alloc(size);
2080 			atomic_add_64(&arc_size, size);
2081 		}
2082 		goto out;
2083 	}
2084 
2085 	/*
2086 	 * If we are prefetching from the mfu ghost list, this buffer
2087 	 * will end up on the mru list; so steal space from there.
2088 	 */
2089 	if (state == arc_mfu_ghost)
2090 		state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2091 	else if (state == arc_mru_ghost)
2092 		state = arc_mru;
2093 
2094 	if (state == arc_mru || state == arc_anon) {
2095 		uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2096 		state = (arc_mfu->arcs_lsize[type] > 0 &&
2097 		    arc_p > mru_used) ? arc_mfu : arc_mru;
2098 	} else {
2099 		/* MFU cases */
2100 		uint64_t mfu_space = arc_c - arc_p;
2101 		state =  (arc_mru->arcs_lsize[type] > 0 &&
2102 		    mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2103 	}
2104 	if ((buf->b_data = arc_evict(state, NULL, size, TRUE, type)) == NULL) {
2105 		if (type == ARC_BUFC_METADATA) {
2106 			buf->b_data = zio_buf_alloc(size);
2107 			arc_space_consume(size);
2108 		} else {
2109 			ASSERT(type == ARC_BUFC_DATA);
2110 			buf->b_data = zio_data_buf_alloc(size);
2111 			atomic_add_64(&arc_size, size);
2112 		}
2113 		ARCSTAT_BUMP(arcstat_recycle_miss);
2114 	}
2115 	ASSERT(buf->b_data != NULL);
2116 out:
2117 	/*
2118 	 * Update the state size.  Note that ghost states have a
2119 	 * "ghost size" and so don't need to be updated.
2120 	 */
2121 	if (!GHOST_STATE(buf->b_hdr->b_state)) {
2122 		arc_buf_hdr_t *hdr = buf->b_hdr;
2123 
2124 		atomic_add_64(&hdr->b_state->arcs_size, size);
2125 		if (list_link_active(&hdr->b_arc_node)) {
2126 			ASSERT(refcount_is_zero(&hdr->b_refcnt));
2127 			atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2128 		}
2129 		/*
2130 		 * If we are growing the cache, and we are adding anonymous
2131 		 * data, and we have outgrown arc_p, update arc_p
2132 		 */
2133 		if (arc_size < arc_c && hdr->b_state == arc_anon &&
2134 		    arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2135 			arc_p = MIN(arc_c, arc_p + size);
2136 	}
2137 }
2138 
2139 /*
2140  * This routine is called whenever a buffer is accessed.
2141  * NOTE: the hash lock is dropped in this function.
2142  */
2143 static void
2144 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2145 {
2146 	ASSERT(MUTEX_HELD(hash_lock));
2147 
2148 	if (buf->b_state == arc_anon) {
2149 		/*
2150 		 * This buffer is not in the cache, and does not
2151 		 * appear in our "ghost" list.  Add the new buffer
2152 		 * to the MRU state.
2153 		 */
2154 
2155 		ASSERT(buf->b_arc_access == 0);
2156 		buf->b_arc_access = lbolt;
2157 		DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2158 		arc_change_state(arc_mru, buf, hash_lock);
2159 
2160 	} else if (buf->b_state == arc_mru) {
2161 		/*
2162 		 * If this buffer is here because of a prefetch, then either:
2163 		 * - clear the flag if this is a "referencing" read
2164 		 *   (any subsequent access will bump this into the MFU state).
2165 		 * or
2166 		 * - move the buffer to the head of the list if this is
2167 		 *   another prefetch (to make it less likely to be evicted).
2168 		 */
2169 		if ((buf->b_flags & ARC_PREFETCH) != 0) {
2170 			if (refcount_count(&buf->b_refcnt) == 0) {
2171 				ASSERT(list_link_active(&buf->b_arc_node));
2172 			} else {
2173 				buf->b_flags &= ~ARC_PREFETCH;
2174 				ARCSTAT_BUMP(arcstat_mru_hits);
2175 			}
2176 			buf->b_arc_access = lbolt;
2177 			return;
2178 		}
2179 
2180 		/*
2181 		 * This buffer has been "accessed" only once so far,
2182 		 * but it is still in the cache. Move it to the MFU
2183 		 * state.
2184 		 */
2185 		if (lbolt > buf->b_arc_access + ARC_MINTIME) {
2186 			/*
2187 			 * More than 125ms have passed since we
2188 			 * instantiated this buffer.  Move it to the
2189 			 * most frequently used state.
2190 			 */
2191 			buf->b_arc_access = lbolt;
2192 			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2193 			arc_change_state(arc_mfu, buf, hash_lock);
2194 		}
2195 		ARCSTAT_BUMP(arcstat_mru_hits);
2196 	} else if (buf->b_state == arc_mru_ghost) {
2197 		arc_state_t	*new_state;
2198 		/*
2199 		 * This buffer has been "accessed" recently, but
2200 		 * was evicted from the cache.  Move it to the
2201 		 * MFU state.
2202 		 */
2203 
2204 		if (buf->b_flags & ARC_PREFETCH) {
2205 			new_state = arc_mru;
2206 			if (refcount_count(&buf->b_refcnt) > 0)
2207 				buf->b_flags &= ~ARC_PREFETCH;
2208 			DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2209 		} else {
2210 			new_state = arc_mfu;
2211 			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2212 		}
2213 
2214 		buf->b_arc_access = lbolt;
2215 		arc_change_state(new_state, buf, hash_lock);
2216 
2217 		ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2218 	} else if (buf->b_state == arc_mfu) {
2219 		/*
2220 		 * This buffer has been accessed more than once and is
2221 		 * still in the cache.  Keep it in the MFU state.
2222 		 *
2223 		 * NOTE: an add_reference() that occurred when we did
2224 		 * the arc_read() will have kicked this off the list.
2225 		 * If it was a prefetch, we will explicitly move it to
2226 		 * the head of the list now.
2227 		 */
2228 		if ((buf->b_flags & ARC_PREFETCH) != 0) {
2229 			ASSERT(refcount_count(&buf->b_refcnt) == 0);
2230 			ASSERT(list_link_active(&buf->b_arc_node));
2231 		}
2232 		ARCSTAT_BUMP(arcstat_mfu_hits);
2233 		buf->b_arc_access = lbolt;
2234 	} else if (buf->b_state == arc_mfu_ghost) {
2235 		arc_state_t	*new_state = arc_mfu;
2236 		/*
2237 		 * This buffer has been accessed more than once but has
2238 		 * been evicted from the cache.  Move it back to the
2239 		 * MFU state.
2240 		 */
2241 
2242 		if (buf->b_flags & ARC_PREFETCH) {
2243 			/*
2244 			 * This is a prefetch access...
2245 			 * move this block back to the MRU state.
2246 			 */
2247 			ASSERT3U(refcount_count(&buf->b_refcnt), ==, 0);
2248 			new_state = arc_mru;
2249 		}
2250 
2251 		buf->b_arc_access = lbolt;
2252 		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2253 		arc_change_state(new_state, buf, hash_lock);
2254 
2255 		ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2256 	} else if (buf->b_state == arc_l2c_only) {
2257 		/*
2258 		 * This buffer is on the 2nd Level ARC.
2259 		 */
2260 
2261 		buf->b_arc_access = lbolt;
2262 		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2263 		arc_change_state(arc_mfu, buf, hash_lock);
2264 	} else {
2265 		ASSERT(!"invalid arc state");
2266 	}
2267 }
2268 
2269 /* a generic arc_done_func_t which you can use */
2270 /* ARGSUSED */
2271 void
2272 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2273 {
2274 	bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2275 	VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2276 }
2277 
2278 /* a generic arc_done_func_t */
2279 void
2280 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2281 {
2282 	arc_buf_t **bufp = arg;
2283 	if (zio && zio->io_error) {
2284 		VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2285 		*bufp = NULL;
2286 	} else {
2287 		*bufp = buf;
2288 	}
2289 }
2290 
2291 static void
2292 arc_read_done(zio_t *zio)
2293 {
2294 	arc_buf_hdr_t	*hdr, *found;
2295 	arc_buf_t	*buf;
2296 	arc_buf_t	*abuf;	/* buffer we're assigning to callback */
2297 	kmutex_t	*hash_lock;
2298 	arc_callback_t	*callback_list, *acb;
2299 	int		freeable = FALSE;
2300 
2301 	buf = zio->io_private;
2302 	hdr = buf->b_hdr;
2303 
2304 	/*
2305 	 * The hdr was inserted into hash-table and removed from lists
2306 	 * prior to starting I/O.  We should find this header, since
2307 	 * it's in the hash table, and it should be legit since it's
2308 	 * not possible to evict it during the I/O.  The only possible
2309 	 * reason for it not to be found is if we were freed during the
2310 	 * read.
2311 	 */
2312 	found = buf_hash_find(zio->io_spa, &hdr->b_dva, hdr->b_birth,
2313 	    &hash_lock);
2314 
2315 	ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2316 	    (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2317 	    (found == hdr && HDR_L2_READING(hdr)));
2318 
2319 	hdr->b_flags &= ~ARC_L2_EVICTED;
2320 	if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2321 		hdr->b_flags &= ~ARC_L2CACHE;
2322 
2323 	/* byteswap if necessary */
2324 	callback_list = hdr->b_acb;
2325 	ASSERT(callback_list != NULL);
2326 	if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
2327 		arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2328 		    byteswap_uint64_array :
2329 		    dmu_ot[BP_GET_TYPE(zio->io_bp)].ot_byteswap;
2330 		func(buf->b_data, hdr->b_size);
2331 	}
2332 
2333 	arc_cksum_compute(buf, B_FALSE);
2334 
2335 	/* create copies of the data buffer for the callers */
2336 	abuf = buf;
2337 	for (acb = callback_list; acb; acb = acb->acb_next) {
2338 		if (acb->acb_done) {
2339 			if (abuf == NULL)
2340 				abuf = arc_buf_clone(buf);
2341 			acb->acb_buf = abuf;
2342 			abuf = NULL;
2343 		}
2344 	}
2345 	hdr->b_acb = NULL;
2346 	hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2347 	ASSERT(!HDR_BUF_AVAILABLE(hdr));
2348 	if (abuf == buf)
2349 		hdr->b_flags |= ARC_BUF_AVAILABLE;
2350 
2351 	ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2352 
2353 	if (zio->io_error != 0) {
2354 		hdr->b_flags |= ARC_IO_ERROR;
2355 		if (hdr->b_state != arc_anon)
2356 			arc_change_state(arc_anon, hdr, hash_lock);
2357 		if (HDR_IN_HASH_TABLE(hdr))
2358 			buf_hash_remove(hdr);
2359 		freeable = refcount_is_zero(&hdr->b_refcnt);
2360 	}
2361 
2362 	/*
2363 	 * Broadcast before we drop the hash_lock to avoid the possibility
2364 	 * that the hdr (and hence the cv) might be freed before we get to
2365 	 * the cv_broadcast().
2366 	 */
2367 	cv_broadcast(&hdr->b_cv);
2368 
2369 	if (hash_lock) {
2370 		/*
2371 		 * Only call arc_access on anonymous buffers.  This is because
2372 		 * if we've issued an I/O for an evicted buffer, we've already
2373 		 * called arc_access (to prevent any simultaneous readers from
2374 		 * getting confused).
2375 		 */
2376 		if (zio->io_error == 0 && hdr->b_state == arc_anon)
2377 			arc_access(hdr, hash_lock);
2378 		mutex_exit(hash_lock);
2379 	} else {
2380 		/*
2381 		 * This block was freed while we waited for the read to
2382 		 * complete.  It has been removed from the hash table and
2383 		 * moved to the anonymous state (so that it won't show up
2384 		 * in the cache).
2385 		 */
2386 		ASSERT3P(hdr->b_state, ==, arc_anon);
2387 		freeable = refcount_is_zero(&hdr->b_refcnt);
2388 	}
2389 
2390 	/* execute each callback and free its structure */
2391 	while ((acb = callback_list) != NULL) {
2392 		if (acb->acb_done)
2393 			acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2394 
2395 		if (acb->acb_zio_dummy != NULL) {
2396 			acb->acb_zio_dummy->io_error = zio->io_error;
2397 			zio_nowait(acb->acb_zio_dummy);
2398 		}
2399 
2400 		callback_list = acb->acb_next;
2401 		kmem_free(acb, sizeof (arc_callback_t));
2402 	}
2403 
2404 	if (freeable)
2405 		arc_hdr_destroy(hdr);
2406 }
2407 
2408 /*
2409  * "Read" the block block at the specified DVA (in bp) via the
2410  * cache.  If the block is found in the cache, invoke the provided
2411  * callback immediately and return.  Note that the `zio' parameter
2412  * in the callback will be NULL in this case, since no IO was
2413  * required.  If the block is not in the cache pass the read request
2414  * on to the spa with a substitute callback function, so that the
2415  * requested block will be added to the cache.
2416  *
2417  * If a read request arrives for a block that has a read in-progress,
2418  * either wait for the in-progress read to complete (and return the
2419  * results); or, if this is a read with a "done" func, add a record
2420  * to the read to invoke the "done" func when the read completes,
2421  * and return; or just return.
2422  *
2423  * arc_read_done() will invoke all the requested "done" functions
2424  * for readers of this block.
2425  *
2426  * Normal callers should use arc_read and pass the arc buffer and offset
2427  * for the bp.  But if you know you don't need locking, you can use
2428  * arc_read_bp.
2429  */
2430 int
2431 arc_read(zio_t *pio, spa_t *spa, blkptr_t *bp, arc_buf_t *pbuf,
2432     arc_done_func_t *done, void *private, int priority, int zio_flags,
2433     uint32_t *arc_flags, const zbookmark_t *zb)
2434 {
2435 	int err;
2436 	arc_buf_hdr_t *hdr = pbuf->b_hdr;
2437 
2438 	ASSERT(!refcount_is_zero(&pbuf->b_hdr->b_refcnt));
2439 	ASSERT3U((char *)bp - (char *)pbuf->b_data, <, pbuf->b_hdr->b_size);
2440 	rw_enter(&pbuf->b_lock, RW_READER);
2441 
2442 	err = arc_read_nolock(pio, spa, bp, done, private, priority,
2443 	    zio_flags, arc_flags, zb);
2444 
2445 	ASSERT3P(hdr, ==, pbuf->b_hdr);
2446 	rw_exit(&pbuf->b_lock);
2447 	return (err);
2448 }
2449 
2450 int
2451 arc_read_nolock(zio_t *pio, spa_t *spa, blkptr_t *bp,
2452     arc_done_func_t *done, void *private, int priority, int zio_flags,
2453     uint32_t *arc_flags, const zbookmark_t *zb)
2454 {
2455 	arc_buf_hdr_t *hdr;
2456 	arc_buf_t *buf;
2457 	kmutex_t *hash_lock;
2458 	zio_t *rzio;
2459 
2460 top:
2461 	hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock);
2462 	if (hdr && hdr->b_datacnt > 0) {
2463 
2464 		*arc_flags |= ARC_CACHED;
2465 
2466 		if (HDR_IO_IN_PROGRESS(hdr)) {
2467 
2468 			if (*arc_flags & ARC_WAIT) {
2469 				cv_wait(&hdr->b_cv, hash_lock);
2470 				mutex_exit(hash_lock);
2471 				goto top;
2472 			}
2473 			ASSERT(*arc_flags & ARC_NOWAIT);
2474 
2475 			if (done) {
2476 				arc_callback_t	*acb = NULL;
2477 
2478 				acb = kmem_zalloc(sizeof (arc_callback_t),
2479 				    KM_SLEEP);
2480 				acb->acb_done = done;
2481 				acb->acb_private = private;
2482 				if (pio != NULL)
2483 					acb->acb_zio_dummy = zio_null(pio,
2484 					    spa, NULL, NULL, zio_flags);
2485 
2486 				ASSERT(acb->acb_done != NULL);
2487 				acb->acb_next = hdr->b_acb;
2488 				hdr->b_acb = acb;
2489 				add_reference(hdr, hash_lock, private);
2490 				mutex_exit(hash_lock);
2491 				return (0);
2492 			}
2493 			mutex_exit(hash_lock);
2494 			return (0);
2495 		}
2496 
2497 		ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2498 
2499 		if (done) {
2500 			add_reference(hdr, hash_lock, private);
2501 			/*
2502 			 * If this block is already in use, create a new
2503 			 * copy of the data so that we will be guaranteed
2504 			 * that arc_release() will always succeed.
2505 			 */
2506 			buf = hdr->b_buf;
2507 			ASSERT(buf);
2508 			ASSERT(buf->b_data);
2509 			if (HDR_BUF_AVAILABLE(hdr)) {
2510 				ASSERT(buf->b_efunc == NULL);
2511 				hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2512 			} else {
2513 				buf = arc_buf_clone(buf);
2514 			}
2515 		} else if (*arc_flags & ARC_PREFETCH &&
2516 		    refcount_count(&hdr->b_refcnt) == 0) {
2517 			hdr->b_flags |= ARC_PREFETCH;
2518 		}
2519 		DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2520 		arc_access(hdr, hash_lock);
2521 		if (*arc_flags & ARC_L2CACHE)
2522 			hdr->b_flags |= ARC_L2CACHE;
2523 		mutex_exit(hash_lock);
2524 		ARCSTAT_BUMP(arcstat_hits);
2525 		ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2526 		    demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2527 		    data, metadata, hits);
2528 
2529 		if (done)
2530 			done(NULL, buf, private);
2531 	} else {
2532 		uint64_t size = BP_GET_LSIZE(bp);
2533 		arc_callback_t	*acb;
2534 		vdev_t *vd = NULL;
2535 		daddr_t addr;
2536 
2537 		if (hdr == NULL) {
2538 			/* this block is not in the cache */
2539 			arc_buf_hdr_t	*exists;
2540 			arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2541 			buf = arc_buf_alloc(spa, size, private, type);
2542 			hdr = buf->b_hdr;
2543 			hdr->b_dva = *BP_IDENTITY(bp);
2544 			hdr->b_birth = bp->blk_birth;
2545 			hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
2546 			exists = buf_hash_insert(hdr, &hash_lock);
2547 			if (exists) {
2548 				/* somebody beat us to the hash insert */
2549 				mutex_exit(hash_lock);
2550 				bzero(&hdr->b_dva, sizeof (dva_t));
2551 				hdr->b_birth = 0;
2552 				hdr->b_cksum0 = 0;
2553 				(void) arc_buf_remove_ref(buf, private);
2554 				goto top; /* restart the IO request */
2555 			}
2556 			/* if this is a prefetch, we don't have a reference */
2557 			if (*arc_flags & ARC_PREFETCH) {
2558 				(void) remove_reference(hdr, hash_lock,
2559 				    private);
2560 				hdr->b_flags |= ARC_PREFETCH;
2561 			}
2562 			if (*arc_flags & ARC_L2CACHE)
2563 				hdr->b_flags |= ARC_L2CACHE;
2564 			if (BP_GET_LEVEL(bp) > 0)
2565 				hdr->b_flags |= ARC_INDIRECT;
2566 		} else {
2567 			/* this block is in the ghost cache */
2568 			ASSERT(GHOST_STATE(hdr->b_state));
2569 			ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2570 			ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 0);
2571 			ASSERT(hdr->b_buf == NULL);
2572 
2573 			/* if this is a prefetch, we don't have a reference */
2574 			if (*arc_flags & ARC_PREFETCH)
2575 				hdr->b_flags |= ARC_PREFETCH;
2576 			else
2577 				add_reference(hdr, hash_lock, private);
2578 			if (*arc_flags & ARC_L2CACHE)
2579 				hdr->b_flags |= ARC_L2CACHE;
2580 			buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2581 			buf->b_hdr = hdr;
2582 			buf->b_data = NULL;
2583 			buf->b_efunc = NULL;
2584 			buf->b_private = NULL;
2585 			buf->b_next = NULL;
2586 			hdr->b_buf = buf;
2587 			arc_get_data_buf(buf);
2588 			ASSERT(hdr->b_datacnt == 0);
2589 			hdr->b_datacnt = 1;
2590 
2591 		}
2592 
2593 		acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
2594 		acb->acb_done = done;
2595 		acb->acb_private = private;
2596 
2597 		ASSERT(hdr->b_acb == NULL);
2598 		hdr->b_acb = acb;
2599 		hdr->b_flags |= ARC_IO_IN_PROGRESS;
2600 
2601 		/*
2602 		 * If the buffer has been evicted, migrate it to a present state
2603 		 * before issuing the I/O.  Once we drop the hash-table lock,
2604 		 * the header will be marked as I/O in progress and have an
2605 		 * attached buffer.  At this point, anybody who finds this
2606 		 * buffer ought to notice that it's legit but has a pending I/O.
2607 		 */
2608 
2609 		if (GHOST_STATE(hdr->b_state))
2610 			arc_access(hdr, hash_lock);
2611 
2612 		if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
2613 		    (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
2614 			addr = hdr->b_l2hdr->b_daddr;
2615 			/*
2616 			 * Lock out device removal.
2617 			 */
2618 			if (vdev_is_dead(vd) ||
2619 			    !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
2620 				vd = NULL;
2621 		}
2622 
2623 		mutex_exit(hash_lock);
2624 
2625 		ASSERT3U(hdr->b_size, ==, size);
2626 		DTRACE_PROBE3(arc__miss, blkptr_t *, bp, uint64_t, size,
2627 		    zbookmark_t *, zb);
2628 		ARCSTAT_BUMP(arcstat_misses);
2629 		ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2630 		    demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2631 		    data, metadata, misses);
2632 
2633 		if (vd != NULL) {
2634 			/*
2635 			 * Read from the L2ARC if the following are true:
2636 			 * 1. The L2ARC vdev was previously cached.
2637 			 * 2. This buffer still has L2ARC metadata.
2638 			 * 3. This buffer isn't currently writing to the L2ARC.
2639 			 * 4. The L2ARC entry wasn't evicted, which may
2640 			 *    also have invalidated the vdev.
2641 			 */
2642 			if (hdr->b_l2hdr != NULL &&
2643 			    !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr)) {
2644 				l2arc_read_callback_t *cb;
2645 
2646 				DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
2647 				ARCSTAT_BUMP(arcstat_l2_hits);
2648 
2649 				cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
2650 				    KM_SLEEP);
2651 				cb->l2rcb_buf = buf;
2652 				cb->l2rcb_spa = spa;
2653 				cb->l2rcb_bp = *bp;
2654 				cb->l2rcb_zb = *zb;
2655 				cb->l2rcb_flags = zio_flags;
2656 
2657 				/*
2658 				 * l2arc read.  The SCL_L2ARC lock will be
2659 				 * released by l2arc_read_done().
2660 				 */
2661 				rzio = zio_read_phys(pio, vd, addr, size,
2662 				    buf->b_data, ZIO_CHECKSUM_OFF,
2663 				    l2arc_read_done, cb, priority, zio_flags |
2664 				    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
2665 				    ZIO_FLAG_DONT_PROPAGATE |
2666 				    ZIO_FLAG_DONT_RETRY, B_FALSE);
2667 				DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
2668 				    zio_t *, rzio);
2669 
2670 				if (*arc_flags & ARC_NOWAIT) {
2671 					zio_nowait(rzio);
2672 					return (0);
2673 				}
2674 
2675 				ASSERT(*arc_flags & ARC_WAIT);
2676 				if (zio_wait(rzio) == 0)
2677 					return (0);
2678 
2679 				/* l2arc read error; goto zio_read() */
2680 			} else {
2681 				DTRACE_PROBE1(l2arc__miss,
2682 				    arc_buf_hdr_t *, hdr);
2683 				ARCSTAT_BUMP(arcstat_l2_misses);
2684 				if (HDR_L2_WRITING(hdr))
2685 					ARCSTAT_BUMP(arcstat_l2_rw_clash);
2686 				spa_config_exit(spa, SCL_L2ARC, vd);
2687 			}
2688 		}
2689 
2690 		rzio = zio_read(pio, spa, bp, buf->b_data, size,
2691 		    arc_read_done, buf, priority, zio_flags, zb);
2692 
2693 		if (*arc_flags & ARC_WAIT)
2694 			return (zio_wait(rzio));
2695 
2696 		ASSERT(*arc_flags & ARC_NOWAIT);
2697 		zio_nowait(rzio);
2698 	}
2699 	return (0);
2700 }
2701 
2702 /*
2703  * arc_read() variant to support pool traversal.  If the block is already
2704  * in the ARC, make a copy of it; otherwise, the caller will do the I/O.
2705  * The idea is that we don't want pool traversal filling up memory, but
2706  * if the ARC already has the data anyway, we shouldn't pay for the I/O.
2707  */
2708 int
2709 arc_tryread(spa_t *spa, blkptr_t *bp, void *data)
2710 {
2711 	arc_buf_hdr_t *hdr;
2712 	kmutex_t *hash_mtx;
2713 	int rc = 0;
2714 
2715 	hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_mtx);
2716 
2717 	if (hdr && hdr->b_datacnt > 0 && !HDR_IO_IN_PROGRESS(hdr)) {
2718 		arc_buf_t *buf = hdr->b_buf;
2719 
2720 		ASSERT(buf);
2721 		while (buf->b_data == NULL) {
2722 			buf = buf->b_next;
2723 			ASSERT(buf);
2724 		}
2725 		bcopy(buf->b_data, data, hdr->b_size);
2726 	} else {
2727 		rc = ENOENT;
2728 	}
2729 
2730 	if (hash_mtx)
2731 		mutex_exit(hash_mtx);
2732 
2733 	return (rc);
2734 }
2735 
2736 void
2737 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
2738 {
2739 	ASSERT(buf->b_hdr != NULL);
2740 	ASSERT(buf->b_hdr->b_state != arc_anon);
2741 	ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
2742 	buf->b_efunc = func;
2743 	buf->b_private = private;
2744 }
2745 
2746 /*
2747  * This is used by the DMU to let the ARC know that a buffer is
2748  * being evicted, so the ARC should clean up.  If this arc buf
2749  * is not yet in the evicted state, it will be put there.
2750  */
2751 int
2752 arc_buf_evict(arc_buf_t *buf)
2753 {
2754 	arc_buf_hdr_t *hdr;
2755 	kmutex_t *hash_lock;
2756 	arc_buf_t **bufp;
2757 
2758 	rw_enter(&buf->b_lock, RW_WRITER);
2759 	hdr = buf->b_hdr;
2760 	if (hdr == NULL) {
2761 		/*
2762 		 * We are in arc_do_user_evicts().
2763 		 */
2764 		ASSERT(buf->b_data == NULL);
2765 		rw_exit(&buf->b_lock);
2766 		return (0);
2767 	} else if (buf->b_data == NULL) {
2768 		arc_buf_t copy = *buf; /* structure assignment */
2769 		/*
2770 		 * We are on the eviction list; process this buffer now
2771 		 * but let arc_do_user_evicts() do the reaping.
2772 		 */
2773 		buf->b_efunc = NULL;
2774 		rw_exit(&buf->b_lock);
2775 		VERIFY(copy.b_efunc(&copy) == 0);
2776 		return (1);
2777 	}
2778 	hash_lock = HDR_LOCK(hdr);
2779 	mutex_enter(hash_lock);
2780 
2781 	ASSERT(buf->b_hdr == hdr);
2782 	ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
2783 	ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2784 
2785 	/*
2786 	 * Pull this buffer off of the hdr
2787 	 */
2788 	bufp = &hdr->b_buf;
2789 	while (*bufp != buf)
2790 		bufp = &(*bufp)->b_next;
2791 	*bufp = buf->b_next;
2792 
2793 	ASSERT(buf->b_data != NULL);
2794 	arc_buf_destroy(buf, FALSE, FALSE);
2795 
2796 	if (hdr->b_datacnt == 0) {
2797 		arc_state_t *old_state = hdr->b_state;
2798 		arc_state_t *evicted_state;
2799 
2800 		ASSERT(refcount_is_zero(&hdr->b_refcnt));
2801 
2802 		evicted_state =
2803 		    (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
2804 
2805 		mutex_enter(&old_state->arcs_mtx);
2806 		mutex_enter(&evicted_state->arcs_mtx);
2807 
2808 		arc_change_state(evicted_state, hdr, hash_lock);
2809 		ASSERT(HDR_IN_HASH_TABLE(hdr));
2810 		hdr->b_flags |= ARC_IN_HASH_TABLE;
2811 		hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2812 
2813 		mutex_exit(&evicted_state->arcs_mtx);
2814 		mutex_exit(&old_state->arcs_mtx);
2815 	}
2816 	mutex_exit(hash_lock);
2817 	rw_exit(&buf->b_lock);
2818 
2819 	VERIFY(buf->b_efunc(buf) == 0);
2820 	buf->b_efunc = NULL;
2821 	buf->b_private = NULL;
2822 	buf->b_hdr = NULL;
2823 	kmem_cache_free(buf_cache, buf);
2824 	return (1);
2825 }
2826 
2827 /*
2828  * Release this buffer from the cache.  This must be done
2829  * after a read and prior to modifying the buffer contents.
2830  * If the buffer has more than one reference, we must make
2831  * a new hdr for the buffer.
2832  */
2833 void
2834 arc_release(arc_buf_t *buf, void *tag)
2835 {
2836 	arc_buf_hdr_t *hdr;
2837 	kmutex_t *hash_lock;
2838 	l2arc_buf_hdr_t *l2hdr;
2839 	uint64_t buf_size;
2840 
2841 	rw_enter(&buf->b_lock, RW_WRITER);
2842 	hdr = buf->b_hdr;
2843 
2844 	/* this buffer is not on any list */
2845 	ASSERT(refcount_count(&hdr->b_refcnt) > 0);
2846 	ASSERT(!(hdr->b_flags & ARC_STORED));
2847 
2848 	if (hdr->b_state == arc_anon) {
2849 		/* this buffer is already released */
2850 		ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 1);
2851 		ASSERT(BUF_EMPTY(hdr));
2852 		ASSERT(buf->b_efunc == NULL);
2853 		arc_buf_thaw(buf);
2854 		rw_exit(&buf->b_lock);
2855 		return;
2856 	}
2857 
2858 	hash_lock = HDR_LOCK(hdr);
2859 	mutex_enter(hash_lock);
2860 
2861 	l2hdr = hdr->b_l2hdr;
2862 	if (l2hdr) {
2863 		mutex_enter(&l2arc_buflist_mtx);
2864 		hdr->b_l2hdr = NULL;
2865 		buf_size = hdr->b_size;
2866 	}
2867 
2868 	/*
2869 	 * Do we have more than one buf?
2870 	 */
2871 	if (hdr->b_datacnt > 1) {
2872 		arc_buf_hdr_t *nhdr;
2873 		arc_buf_t **bufp;
2874 		uint64_t blksz = hdr->b_size;
2875 		spa_t *spa = hdr->b_spa;
2876 		arc_buf_contents_t type = hdr->b_type;
2877 		uint32_t flags = hdr->b_flags;
2878 
2879 		ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
2880 		/*
2881 		 * Pull the data off of this buf and attach it to
2882 		 * a new anonymous buf.
2883 		 */
2884 		(void) remove_reference(hdr, hash_lock, tag);
2885 		bufp = &hdr->b_buf;
2886 		while (*bufp != buf)
2887 			bufp = &(*bufp)->b_next;
2888 		*bufp = (*bufp)->b_next;
2889 		buf->b_next = NULL;
2890 
2891 		ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
2892 		atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
2893 		if (refcount_is_zero(&hdr->b_refcnt)) {
2894 			uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
2895 			ASSERT3U(*size, >=, hdr->b_size);
2896 			atomic_add_64(size, -hdr->b_size);
2897 		}
2898 		hdr->b_datacnt -= 1;
2899 		arc_cksum_verify(buf);
2900 
2901 		mutex_exit(hash_lock);
2902 
2903 		nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
2904 		nhdr->b_size = blksz;
2905 		nhdr->b_spa = spa;
2906 		nhdr->b_type = type;
2907 		nhdr->b_buf = buf;
2908 		nhdr->b_state = arc_anon;
2909 		nhdr->b_arc_access = 0;
2910 		nhdr->b_flags = flags & ARC_L2_WRITING;
2911 		nhdr->b_l2hdr = NULL;
2912 		nhdr->b_datacnt = 1;
2913 		nhdr->b_freeze_cksum = NULL;
2914 		(void) refcount_add(&nhdr->b_refcnt, tag);
2915 		buf->b_hdr = nhdr;
2916 		rw_exit(&buf->b_lock);
2917 		atomic_add_64(&arc_anon->arcs_size, blksz);
2918 	} else {
2919 		rw_exit(&buf->b_lock);
2920 		ASSERT(refcount_count(&hdr->b_refcnt) == 1);
2921 		ASSERT(!list_link_active(&hdr->b_arc_node));
2922 		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2923 		arc_change_state(arc_anon, hdr, hash_lock);
2924 		hdr->b_arc_access = 0;
2925 		mutex_exit(hash_lock);
2926 
2927 		bzero(&hdr->b_dva, sizeof (dva_t));
2928 		hdr->b_birth = 0;
2929 		hdr->b_cksum0 = 0;
2930 		arc_buf_thaw(buf);
2931 	}
2932 	buf->b_efunc = NULL;
2933 	buf->b_private = NULL;
2934 
2935 	if (l2hdr) {
2936 		list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
2937 		kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
2938 		ARCSTAT_INCR(arcstat_l2_size, -buf_size);
2939 		mutex_exit(&l2arc_buflist_mtx);
2940 	}
2941 }
2942 
2943 int
2944 arc_released(arc_buf_t *buf)
2945 {
2946 	int released;
2947 
2948 	rw_enter(&buf->b_lock, RW_READER);
2949 	released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
2950 	rw_exit(&buf->b_lock);
2951 	return (released);
2952 }
2953 
2954 int
2955 arc_has_callback(arc_buf_t *buf)
2956 {
2957 	int callback;
2958 
2959 	rw_enter(&buf->b_lock, RW_READER);
2960 	callback = (buf->b_efunc != NULL);
2961 	rw_exit(&buf->b_lock);
2962 	return (callback);
2963 }
2964 
2965 #ifdef ZFS_DEBUG
2966 int
2967 arc_referenced(arc_buf_t *buf)
2968 {
2969 	int referenced;
2970 
2971 	rw_enter(&buf->b_lock, RW_READER);
2972 	referenced = (refcount_count(&buf->b_hdr->b_refcnt));
2973 	rw_exit(&buf->b_lock);
2974 	return (referenced);
2975 }
2976 #endif
2977 
2978 static void
2979 arc_write_ready(zio_t *zio)
2980 {
2981 	arc_write_callback_t *callback = zio->io_private;
2982 	arc_buf_t *buf = callback->awcb_buf;
2983 	arc_buf_hdr_t *hdr = buf->b_hdr;
2984 
2985 	ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
2986 	callback->awcb_ready(zio, buf, callback->awcb_private);
2987 
2988 	/*
2989 	 * If the IO is already in progress, then this is a re-write
2990 	 * attempt, so we need to thaw and re-compute the cksum.
2991 	 * It is the responsibility of the callback to handle the
2992 	 * accounting for any re-write attempt.
2993 	 */
2994 	if (HDR_IO_IN_PROGRESS(hdr)) {
2995 		mutex_enter(&hdr->b_freeze_lock);
2996 		if (hdr->b_freeze_cksum != NULL) {
2997 			kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
2998 			hdr->b_freeze_cksum = NULL;
2999 		}
3000 		mutex_exit(&hdr->b_freeze_lock);
3001 	}
3002 	arc_cksum_compute(buf, B_FALSE);
3003 	hdr->b_flags |= ARC_IO_IN_PROGRESS;
3004 }
3005 
3006 static void
3007 arc_write_done(zio_t *zio)
3008 {
3009 	arc_write_callback_t *callback = zio->io_private;
3010 	arc_buf_t *buf = callback->awcb_buf;
3011 	arc_buf_hdr_t *hdr = buf->b_hdr;
3012 
3013 	hdr->b_acb = NULL;
3014 
3015 	hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3016 	hdr->b_birth = zio->io_bp->blk_birth;
3017 	hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3018 	/*
3019 	 * If the block to be written was all-zero, we may have
3020 	 * compressed it away.  In this case no write was performed
3021 	 * so there will be no dva/birth-date/checksum.  The buffer
3022 	 * must therefor remain anonymous (and uncached).
3023 	 */
3024 	if (!BUF_EMPTY(hdr)) {
3025 		arc_buf_hdr_t *exists;
3026 		kmutex_t *hash_lock;
3027 
3028 		arc_cksum_verify(buf);
3029 
3030 		exists = buf_hash_insert(hdr, &hash_lock);
3031 		if (exists) {
3032 			/*
3033 			 * This can only happen if we overwrite for
3034 			 * sync-to-convergence, because we remove
3035 			 * buffers from the hash table when we arc_free().
3036 			 */
3037 			ASSERT(zio->io_flags & ZIO_FLAG_IO_REWRITE);
3038 			ASSERT(DVA_EQUAL(BP_IDENTITY(&zio->io_bp_orig),
3039 			    BP_IDENTITY(zio->io_bp)));
3040 			ASSERT3U(zio->io_bp_orig.blk_birth, ==,
3041 			    zio->io_bp->blk_birth);
3042 
3043 			ASSERT(refcount_is_zero(&exists->b_refcnt));
3044 			arc_change_state(arc_anon, exists, hash_lock);
3045 			mutex_exit(hash_lock);
3046 			arc_hdr_destroy(exists);
3047 			exists = buf_hash_insert(hdr, &hash_lock);
3048 			ASSERT3P(exists, ==, NULL);
3049 		}
3050 		hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3051 		/* if it's not anon, we are doing a scrub */
3052 		if (hdr->b_state == arc_anon)
3053 			arc_access(hdr, hash_lock);
3054 		mutex_exit(hash_lock);
3055 	} else if (callback->awcb_done == NULL) {
3056 		int destroy_hdr;
3057 		/*
3058 		 * This is an anonymous buffer with no user callback,
3059 		 * destroy it if there are no active references.
3060 		 */
3061 		mutex_enter(&arc_eviction_mtx);
3062 		destroy_hdr = refcount_is_zero(&hdr->b_refcnt);
3063 		hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3064 		mutex_exit(&arc_eviction_mtx);
3065 		if (destroy_hdr)
3066 			arc_hdr_destroy(hdr);
3067 	} else {
3068 		hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3069 	}
3070 	hdr->b_flags &= ~ARC_STORED;
3071 
3072 	if (callback->awcb_done) {
3073 		ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3074 		callback->awcb_done(zio, buf, callback->awcb_private);
3075 	}
3076 
3077 	kmem_free(callback, sizeof (arc_write_callback_t));
3078 }
3079 
3080 static void
3081 write_policy(spa_t *spa, const writeprops_t *wp, zio_prop_t *zp)
3082 {
3083 	boolean_t ismd = (wp->wp_level > 0 || dmu_ot[wp->wp_type].ot_metadata);
3084 
3085 	/* Determine checksum setting */
3086 	if (ismd) {
3087 		/*
3088 		 * Metadata always gets checksummed.  If the data
3089 		 * checksum is multi-bit correctable, and it's not a
3090 		 * ZBT-style checksum, then it's suitable for metadata
3091 		 * as well.  Otherwise, the metadata checksum defaults
3092 		 * to fletcher4.
3093 		 */
3094 		if (zio_checksum_table[wp->wp_oschecksum].ci_correctable &&
3095 		    !zio_checksum_table[wp->wp_oschecksum].ci_zbt)
3096 			zp->zp_checksum = wp->wp_oschecksum;
3097 		else
3098 			zp->zp_checksum = ZIO_CHECKSUM_FLETCHER_4;
3099 	} else {
3100 		zp->zp_checksum = zio_checksum_select(wp->wp_dnchecksum,
3101 		    wp->wp_oschecksum);
3102 	}
3103 
3104 	/* Determine compression setting */
3105 	if (ismd) {
3106 		/*
3107 		 * XXX -- we should design a compression algorithm
3108 		 * that specializes in arrays of bps.
3109 		 */
3110 		zp->zp_compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
3111 		    ZIO_COMPRESS_LZJB;
3112 	} else {
3113 		zp->zp_compress = zio_compress_select(wp->wp_dncompress,
3114 		    wp->wp_oscompress);
3115 	}
3116 
3117 	zp->zp_type = wp->wp_type;
3118 	zp->zp_level = wp->wp_level;
3119 	zp->zp_ndvas = MIN(wp->wp_copies + ismd, spa_max_replication(spa));
3120 }
3121 
3122 zio_t *
3123 arc_write(zio_t *pio, spa_t *spa, const writeprops_t *wp,
3124     boolean_t l2arc, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
3125     arc_done_func_t *ready, arc_done_func_t *done, void *private, int priority,
3126     int zio_flags, const zbookmark_t *zb)
3127 {
3128 	arc_buf_hdr_t *hdr = buf->b_hdr;
3129 	arc_write_callback_t *callback;
3130 	zio_t *zio;
3131 	zio_prop_t zp;
3132 
3133 	ASSERT(ready != NULL);
3134 	ASSERT(!HDR_IO_ERROR(hdr));
3135 	ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3136 	ASSERT(hdr->b_acb == 0);
3137 	if (l2arc)
3138 		hdr->b_flags |= ARC_L2CACHE;
3139 	callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3140 	callback->awcb_ready = ready;
3141 	callback->awcb_done = done;
3142 	callback->awcb_private = private;
3143 	callback->awcb_buf = buf;
3144 
3145 	write_policy(spa, wp, &zp);
3146 	zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, &zp,
3147 	    arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3148 
3149 	return (zio);
3150 }
3151 
3152 int
3153 arc_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
3154     zio_done_func_t *done, void *private, uint32_t arc_flags)
3155 {
3156 	arc_buf_hdr_t *ab;
3157 	kmutex_t *hash_lock;
3158 	zio_t	*zio;
3159 
3160 	/*
3161 	 * If this buffer is in the cache, release it, so it
3162 	 * can be re-used.
3163 	 */
3164 	ab = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock);
3165 	if (ab != NULL) {
3166 		/*
3167 		 * The checksum of blocks to free is not always
3168 		 * preserved (eg. on the deadlist).  However, if it is
3169 		 * nonzero, it should match what we have in the cache.
3170 		 */
3171 		ASSERT(bp->blk_cksum.zc_word[0] == 0 ||
3172 		    bp->blk_cksum.zc_word[0] == ab->b_cksum0 ||
3173 		    bp->blk_fill == BLK_FILL_ALREADY_FREED);
3174 
3175 		if (ab->b_state != arc_anon)
3176 			arc_change_state(arc_anon, ab, hash_lock);
3177 		if (HDR_IO_IN_PROGRESS(ab)) {
3178 			/*
3179 			 * This should only happen when we prefetch.
3180 			 */
3181 			ASSERT(ab->b_flags & ARC_PREFETCH);
3182 			ASSERT3U(ab->b_datacnt, ==, 1);
3183 			ab->b_flags |= ARC_FREED_IN_READ;
3184 			if (HDR_IN_HASH_TABLE(ab))
3185 				buf_hash_remove(ab);
3186 			ab->b_arc_access = 0;
3187 			bzero(&ab->b_dva, sizeof (dva_t));
3188 			ab->b_birth = 0;
3189 			ab->b_cksum0 = 0;
3190 			ab->b_buf->b_efunc = NULL;
3191 			ab->b_buf->b_private = NULL;
3192 			mutex_exit(hash_lock);
3193 		} else if (refcount_is_zero(&ab->b_refcnt)) {
3194 			ab->b_flags |= ARC_FREE_IN_PROGRESS;
3195 			mutex_exit(hash_lock);
3196 			arc_hdr_destroy(ab);
3197 			ARCSTAT_BUMP(arcstat_deleted);
3198 		} else {
3199 			/*
3200 			 * We still have an active reference on this
3201 			 * buffer.  This can happen, e.g., from
3202 			 * dbuf_unoverride().
3203 			 */
3204 			ASSERT(!HDR_IN_HASH_TABLE(ab));
3205 			ab->b_arc_access = 0;
3206 			bzero(&ab->b_dva, sizeof (dva_t));
3207 			ab->b_birth = 0;
3208 			ab->b_cksum0 = 0;
3209 			ab->b_buf->b_efunc = NULL;
3210 			ab->b_buf->b_private = NULL;
3211 			mutex_exit(hash_lock);
3212 		}
3213 	}
3214 
3215 	zio = zio_free(pio, spa, txg, bp, done, private, ZIO_FLAG_MUSTSUCCEED);
3216 
3217 	if (arc_flags & ARC_WAIT)
3218 		return (zio_wait(zio));
3219 
3220 	ASSERT(arc_flags & ARC_NOWAIT);
3221 	zio_nowait(zio);
3222 
3223 	return (0);
3224 }
3225 
3226 static int
3227 arc_memory_throttle(uint64_t reserve, uint64_t txg)
3228 {
3229 #ifdef _KERNEL
3230 	uint64_t inflight_data = arc_anon->arcs_size;
3231 	uint64_t available_memory = ptob(freemem);
3232 	static uint64_t page_load = 0;
3233 	static uint64_t last_txg = 0;
3234 
3235 #if defined(__i386)
3236 	available_memory =
3237 	    MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3238 #endif
3239 	if (available_memory >= zfs_write_limit_max)
3240 		return (0);
3241 
3242 	if (txg > last_txg) {
3243 		last_txg = txg;
3244 		page_load = 0;
3245 	}
3246 	/*
3247 	 * If we are in pageout, we know that memory is already tight,
3248 	 * the arc is already going to be evicting, so we just want to
3249 	 * continue to let page writes occur as quickly as possible.
3250 	 */
3251 	if (curproc == proc_pageout) {
3252 		if (page_load > MAX(ptob(minfree), available_memory) / 4)
3253 			return (ERESTART);
3254 		/* Note: reserve is inflated, so we deflate */
3255 		page_load += reserve / 8;
3256 		return (0);
3257 	} else if (page_load > 0 && arc_reclaim_needed()) {
3258 		/* memory is low, delay before restarting */
3259 		ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3260 		return (EAGAIN);
3261 	}
3262 	page_load = 0;
3263 
3264 	if (arc_size > arc_c_min) {
3265 		uint64_t evictable_memory =
3266 		    arc_mru->arcs_lsize[ARC_BUFC_DATA] +
3267 		    arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
3268 		    arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
3269 		    arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
3270 		available_memory += MIN(evictable_memory, arc_size - arc_c_min);
3271 	}
3272 
3273 	if (inflight_data > available_memory / 4) {
3274 		ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3275 		return (ERESTART);
3276 	}
3277 #endif
3278 	return (0);
3279 }
3280 
3281 void
3282 arc_tempreserve_clear(uint64_t reserve)
3283 {
3284 	atomic_add_64(&arc_tempreserve, -reserve);
3285 	ASSERT((int64_t)arc_tempreserve >= 0);
3286 }
3287 
3288 int
3289 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3290 {
3291 	int error;
3292 
3293 #ifdef ZFS_DEBUG
3294 	/*
3295 	 * Once in a while, fail for no reason.  Everything should cope.
3296 	 */
3297 	if (spa_get_random(10000) == 0) {
3298 		dprintf("forcing random failure\n");
3299 		return (ERESTART);
3300 	}
3301 #endif
3302 	if (reserve > arc_c/4 && !arc_no_grow)
3303 		arc_c = MIN(arc_c_max, reserve * 4);
3304 	if (reserve > arc_c)
3305 		return (ENOMEM);
3306 
3307 	/*
3308 	 * Writes will, almost always, require additional memory allocations
3309 	 * in order to compress/encrypt/etc the data.  We therefor need to
3310 	 * make sure that there is sufficient available memory for this.
3311 	 */
3312 	if (error = arc_memory_throttle(reserve, txg))
3313 		return (error);
3314 
3315 	/*
3316 	 * Throttle writes when the amount of dirty data in the cache
3317 	 * gets too large.  We try to keep the cache less than half full
3318 	 * of dirty blocks so that our sync times don't grow too large.
3319 	 * Note: if two requests come in concurrently, we might let them
3320 	 * both succeed, when one of them should fail.  Not a huge deal.
3321 	 */
3322 	if (reserve + arc_tempreserve + arc_anon->arcs_size > arc_c / 2 &&
3323 	    arc_anon->arcs_size > arc_c / 4) {
3324 		dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3325 		    "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3326 		    arc_tempreserve>>10,
3327 		    arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3328 		    arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3329 		    reserve>>10, arc_c>>10);
3330 		return (ERESTART);
3331 	}
3332 	atomic_add_64(&arc_tempreserve, reserve);
3333 	return (0);
3334 }
3335 
3336 void
3337 arc_init(void)
3338 {
3339 	mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3340 	cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3341 
3342 	/* Convert seconds to clock ticks */
3343 	arc_min_prefetch_lifespan = 1 * hz;
3344 
3345 	/* Start out with 1/8 of all memory */
3346 	arc_c = physmem * PAGESIZE / 8;
3347 
3348 #ifdef _KERNEL
3349 	/*
3350 	 * On architectures where the physical memory can be larger
3351 	 * than the addressable space (intel in 32-bit mode), we may
3352 	 * need to limit the cache to 1/8 of VM size.
3353 	 */
3354 	arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3355 #endif
3356 
3357 	/* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3358 	arc_c_min = MAX(arc_c / 4, 64<<20);
3359 	/* set max to 3/4 of all memory, or all but 1GB, whichever is more */
3360 	if (arc_c * 8 >= 1<<30)
3361 		arc_c_max = (arc_c * 8) - (1<<30);
3362 	else
3363 		arc_c_max = arc_c_min;
3364 	arc_c_max = MAX(arc_c * 6, arc_c_max);
3365 
3366 	/*
3367 	 * Allow the tunables to override our calculations if they are
3368 	 * reasonable (ie. over 64MB)
3369 	 */
3370 	if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3371 		arc_c_max = zfs_arc_max;
3372 	if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3373 		arc_c_min = zfs_arc_min;
3374 
3375 	arc_c = arc_c_max;
3376 	arc_p = (arc_c >> 1);
3377 
3378 	/* limit meta-data to 1/4 of the arc capacity */
3379 	arc_meta_limit = arc_c_max / 4;
3380 
3381 	/* Allow the tunable to override if it is reasonable */
3382 	if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3383 		arc_meta_limit = zfs_arc_meta_limit;
3384 
3385 	if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3386 		arc_c_min = arc_meta_limit / 2;
3387 
3388 	/* if kmem_flags are set, lets try to use less memory */
3389 	if (kmem_debugging())
3390 		arc_c = arc_c / 2;
3391 	if (arc_c < arc_c_min)
3392 		arc_c = arc_c_min;
3393 
3394 	arc_anon = &ARC_anon;
3395 	arc_mru = &ARC_mru;
3396 	arc_mru_ghost = &ARC_mru_ghost;
3397 	arc_mfu = &ARC_mfu;
3398 	arc_mfu_ghost = &ARC_mfu_ghost;
3399 	arc_l2c_only = &ARC_l2c_only;
3400 	arc_size = 0;
3401 
3402 	mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3403 	mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3404 	mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3405 	mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3406 	mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3407 	mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3408 
3409 	list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3410 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3411 	list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3412 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3413 	list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3414 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3415 	list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3416 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3417 	list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3418 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3419 	list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3420 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3421 	list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3422 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3423 	list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3424 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3425 	list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3426 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3427 	list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3428 	    sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3429 
3430 	buf_init();
3431 
3432 	arc_thread_exit = 0;
3433 	arc_eviction_list = NULL;
3434 	mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3435 	bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3436 
3437 	arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3438 	    sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3439 
3440 	if (arc_ksp != NULL) {
3441 		arc_ksp->ks_data = &arc_stats;
3442 		kstat_install(arc_ksp);
3443 	}
3444 
3445 	(void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3446 	    TS_RUN, minclsyspri);
3447 
3448 	arc_dead = FALSE;
3449 	arc_warm = B_FALSE;
3450 
3451 	if (zfs_write_limit_max == 0)
3452 		zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3453 	else
3454 		zfs_write_limit_shift = 0;
3455 	mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3456 }
3457 
3458 void
3459 arc_fini(void)
3460 {
3461 	mutex_enter(&arc_reclaim_thr_lock);
3462 	arc_thread_exit = 1;
3463 	while (arc_thread_exit != 0)
3464 		cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3465 	mutex_exit(&arc_reclaim_thr_lock);
3466 
3467 	arc_flush(NULL);
3468 
3469 	arc_dead = TRUE;
3470 
3471 	if (arc_ksp != NULL) {
3472 		kstat_delete(arc_ksp);
3473 		arc_ksp = NULL;
3474 	}
3475 
3476 	mutex_destroy(&arc_eviction_mtx);
3477 	mutex_destroy(&arc_reclaim_thr_lock);
3478 	cv_destroy(&arc_reclaim_thr_cv);
3479 
3480 	list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3481 	list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3482 	list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3483 	list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3484 	list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3485 	list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3486 	list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3487 	list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3488 
3489 	mutex_destroy(&arc_anon->arcs_mtx);
3490 	mutex_destroy(&arc_mru->arcs_mtx);
3491 	mutex_destroy(&arc_mru_ghost->arcs_mtx);
3492 	mutex_destroy(&arc_mfu->arcs_mtx);
3493 	mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3494 
3495 	mutex_destroy(&zfs_write_limit_lock);
3496 
3497 	buf_fini();
3498 }
3499 
3500 /*
3501  * Level 2 ARC
3502  *
3503  * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3504  * It uses dedicated storage devices to hold cached data, which are populated
3505  * using large infrequent writes.  The main role of this cache is to boost
3506  * the performance of random read workloads.  The intended L2ARC devices
3507  * include short-stroked disks, solid state disks, and other media with
3508  * substantially faster read latency than disk.
3509  *
3510  *                 +-----------------------+
3511  *                 |         ARC           |
3512  *                 +-----------------------+
3513  *                    |         ^     ^
3514  *                    |         |     |
3515  *      l2arc_feed_thread()    arc_read()
3516  *                    |         |     |
3517  *                    |  l2arc read   |
3518  *                    V         |     |
3519  *               +---------------+    |
3520  *               |     L2ARC     |    |
3521  *               +---------------+    |
3522  *                   |    ^           |
3523  *          l2arc_write() |           |
3524  *                   |    |           |
3525  *                   V    |           |
3526  *                 +-------+      +-------+
3527  *                 | vdev  |      | vdev  |
3528  *                 | cache |      | cache |
3529  *                 +-------+      +-------+
3530  *                 +=========+     .-----.
3531  *                 :  L2ARC  :    |-_____-|
3532  *                 : devices :    | Disks |
3533  *                 +=========+    `-_____-'
3534  *
3535  * Read requests are satisfied from the following sources, in order:
3536  *
3537  *	1) ARC
3538  *	2) vdev cache of L2ARC devices
3539  *	3) L2ARC devices
3540  *	4) vdev cache of disks
3541  *	5) disks
3542  *
3543  * Some L2ARC device types exhibit extremely slow write performance.
3544  * To accommodate for this there are some significant differences between
3545  * the L2ARC and traditional cache design:
3546  *
3547  * 1. There is no eviction path from the ARC to the L2ARC.  Evictions from
3548  * the ARC behave as usual, freeing buffers and placing headers on ghost
3549  * lists.  The ARC does not send buffers to the L2ARC during eviction as
3550  * this would add inflated write latencies for all ARC memory pressure.
3551  *
3552  * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
3553  * It does this by periodically scanning buffers from the eviction-end of
3554  * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
3555  * not already there.  It scans until a headroom of buffers is satisfied,
3556  * which itself is a buffer for ARC eviction.  The thread that does this is
3557  * l2arc_feed_thread(), illustrated below; example sizes are included to
3558  * provide a better sense of ratio than this diagram:
3559  *
3560  *	       head -->                        tail
3561  *	        +---------------------+----------+
3562  *	ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->.   # already on L2ARC
3563  *	        +---------------------+----------+   |   o L2ARC eligible
3564  *	ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->|   : ARC buffer
3565  *	        +---------------------+----------+   |
3566  *	             15.9 Gbytes      ^ 32 Mbytes    |
3567  *	                           headroom          |
3568  *	                                      l2arc_feed_thread()
3569  *	                                             |
3570  *	                 l2arc write hand <--[oooo]--'
3571  *	                         |           8 Mbyte
3572  *	                         |          write max
3573  *	                         V
3574  *		  +==============================+
3575  *	L2ARC dev |####|#|###|###|    |####| ... |
3576  *	          +==============================+
3577  *	                     32 Gbytes
3578  *
3579  * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
3580  * evicted, then the L2ARC has cached a buffer much sooner than it probably
3581  * needed to, potentially wasting L2ARC device bandwidth and storage.  It is
3582  * safe to say that this is an uncommon case, since buffers at the end of
3583  * the ARC lists have moved there due to inactivity.
3584  *
3585  * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
3586  * then the L2ARC simply misses copying some buffers.  This serves as a
3587  * pressure valve to prevent heavy read workloads from both stalling the ARC
3588  * with waits and clogging the L2ARC with writes.  This also helps prevent
3589  * the potential for the L2ARC to churn if it attempts to cache content too
3590  * quickly, such as during backups of the entire pool.
3591  *
3592  * 5. After system boot and before the ARC has filled main memory, there are
3593  * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
3594  * lists can remain mostly static.  Instead of searching from tail of these
3595  * lists as pictured, the l2arc_feed_thread() will search from the list heads
3596  * for eligible buffers, greatly increasing its chance of finding them.
3597  *
3598  * The L2ARC device write speed is also boosted during this time so that
3599  * the L2ARC warms up faster.  Since there have been no ARC evictions yet,
3600  * there are no L2ARC reads, and no fear of degrading read performance
3601  * through increased writes.
3602  *
3603  * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
3604  * the vdev queue can aggregate them into larger and fewer writes.  Each
3605  * device is written to in a rotor fashion, sweeping writes through
3606  * available space then repeating.
3607  *
3608  * 7. The L2ARC does not store dirty content.  It never needs to flush
3609  * write buffers back to disk based storage.
3610  *
3611  * 8. If an ARC buffer is written (and dirtied) which also exists in the
3612  * L2ARC, the now stale L2ARC buffer is immediately dropped.
3613  *
3614  * The performance of the L2ARC can be tweaked by a number of tunables, which
3615  * may be necessary for different workloads:
3616  *
3617  *	l2arc_write_max		max write bytes per interval
3618  *	l2arc_write_boost	extra write bytes during device warmup
3619  *	l2arc_noprefetch	skip caching prefetched buffers
3620  *	l2arc_headroom		number of max device writes to precache
3621  *	l2arc_feed_secs		seconds between L2ARC writing
3622  *
3623  * Tunables may be removed or added as future performance improvements are
3624  * integrated, and also may become zpool properties.
3625  */
3626 
3627 static void
3628 l2arc_hdr_stat_add(void)
3629 {
3630 	ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
3631 	ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
3632 }
3633 
3634 static void
3635 l2arc_hdr_stat_remove(void)
3636 {
3637 	ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
3638 	ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
3639 }
3640 
3641 /*
3642  * Cycle through L2ARC devices.  This is how L2ARC load balances.
3643  * If a device is returned, this also returns holding the spa config lock.
3644  */
3645 static l2arc_dev_t *
3646 l2arc_dev_get_next(void)
3647 {
3648 	l2arc_dev_t *first, *next = NULL;
3649 
3650 	/*
3651 	 * Lock out the removal of spas (spa_namespace_lock), then removal
3652 	 * of cache devices (l2arc_dev_mtx).  Once a device has been selected,
3653 	 * both locks will be dropped and a spa config lock held instead.
3654 	 */
3655 	mutex_enter(&spa_namespace_lock);
3656 	mutex_enter(&l2arc_dev_mtx);
3657 
3658 	/* if there are no vdevs, there is nothing to do */
3659 	if (l2arc_ndev == 0)
3660 		goto out;
3661 
3662 	first = NULL;
3663 	next = l2arc_dev_last;
3664 	do {
3665 		/* loop around the list looking for a non-faulted vdev */
3666 		if (next == NULL) {
3667 			next = list_head(l2arc_dev_list);
3668 		} else {
3669 			next = list_next(l2arc_dev_list, next);
3670 			if (next == NULL)
3671 				next = list_head(l2arc_dev_list);
3672 		}
3673 
3674 		/* if we have come back to the start, bail out */
3675 		if (first == NULL)
3676 			first = next;
3677 		else if (next == first)
3678 			break;
3679 
3680 	} while (vdev_is_dead(next->l2ad_vdev));
3681 
3682 	/* if we were unable to find any usable vdevs, return NULL */
3683 	if (vdev_is_dead(next->l2ad_vdev))
3684 		next = NULL;
3685 
3686 	l2arc_dev_last = next;
3687 
3688 out:
3689 	mutex_exit(&l2arc_dev_mtx);
3690 
3691 	/*
3692 	 * Grab the config lock to prevent the 'next' device from being
3693 	 * removed while we are writing to it.
3694 	 */
3695 	if (next != NULL)
3696 		spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
3697 	mutex_exit(&spa_namespace_lock);
3698 
3699 	return (next);
3700 }
3701 
3702 /*
3703  * Free buffers that were tagged for destruction.
3704  */
3705 static void
3706 l2arc_do_free_on_write()
3707 {
3708 	list_t *buflist;
3709 	l2arc_data_free_t *df, *df_prev;
3710 
3711 	mutex_enter(&l2arc_free_on_write_mtx);
3712 	buflist = l2arc_free_on_write;
3713 
3714 	for (df = list_tail(buflist); df; df = df_prev) {
3715 		df_prev = list_prev(buflist, df);
3716 		ASSERT(df->l2df_data != NULL);
3717 		ASSERT(df->l2df_func != NULL);
3718 		df->l2df_func(df->l2df_data, df->l2df_size);
3719 		list_remove(buflist, df);
3720 		kmem_free(df, sizeof (l2arc_data_free_t));
3721 	}
3722 
3723 	mutex_exit(&l2arc_free_on_write_mtx);
3724 }
3725 
3726 /*
3727  * A write to a cache device has completed.  Update all headers to allow
3728  * reads from these buffers to begin.
3729  */
3730 static void
3731 l2arc_write_done(zio_t *zio)
3732 {
3733 	l2arc_write_callback_t *cb;
3734 	l2arc_dev_t *dev;
3735 	list_t *buflist;
3736 	arc_buf_hdr_t *head, *ab, *ab_prev;
3737 	l2arc_buf_hdr_t *abl2;
3738 	kmutex_t *hash_lock;
3739 
3740 	cb = zio->io_private;
3741 	ASSERT(cb != NULL);
3742 	dev = cb->l2wcb_dev;
3743 	ASSERT(dev != NULL);
3744 	head = cb->l2wcb_head;
3745 	ASSERT(head != NULL);
3746 	buflist = dev->l2ad_buflist;
3747 	ASSERT(buflist != NULL);
3748 	DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
3749 	    l2arc_write_callback_t *, cb);
3750 
3751 	if (zio->io_error != 0)
3752 		ARCSTAT_BUMP(arcstat_l2_writes_error);
3753 
3754 	mutex_enter(&l2arc_buflist_mtx);
3755 
3756 	/*
3757 	 * All writes completed, or an error was hit.
3758 	 */
3759 	for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
3760 		ab_prev = list_prev(buflist, ab);
3761 
3762 		hash_lock = HDR_LOCK(ab);
3763 		if (!mutex_tryenter(hash_lock)) {
3764 			/*
3765 			 * This buffer misses out.  It may be in a stage
3766 			 * of eviction.  Its ARC_L2_WRITING flag will be
3767 			 * left set, denying reads to this buffer.
3768 			 */
3769 			ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
3770 			continue;
3771 		}
3772 
3773 		if (zio->io_error != 0) {
3774 			/*
3775 			 * Error - drop L2ARC entry.
3776 			 */
3777 			list_remove(buflist, ab);
3778 			abl2 = ab->b_l2hdr;
3779 			ab->b_l2hdr = NULL;
3780 			kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
3781 			ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
3782 		}
3783 
3784 		/*
3785 		 * Allow ARC to begin reads to this L2ARC entry.
3786 		 */
3787 		ab->b_flags &= ~ARC_L2_WRITING;
3788 
3789 		mutex_exit(hash_lock);
3790 	}
3791 
3792 	atomic_inc_64(&l2arc_writes_done);
3793 	list_remove(buflist, head);
3794 	kmem_cache_free(hdr_cache, head);
3795 	mutex_exit(&l2arc_buflist_mtx);
3796 
3797 	l2arc_do_free_on_write();
3798 
3799 	kmem_free(cb, sizeof (l2arc_write_callback_t));
3800 }
3801 
3802 /*
3803  * A read to a cache device completed.  Validate buffer contents before
3804  * handing over to the regular ARC routines.
3805  */
3806 static void
3807 l2arc_read_done(zio_t *zio)
3808 {
3809 	l2arc_read_callback_t *cb;
3810 	arc_buf_hdr_t *hdr;
3811 	arc_buf_t *buf;
3812 	kmutex_t *hash_lock;
3813 	int equal;
3814 
3815 	ASSERT(zio->io_vd != NULL);
3816 	ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
3817 
3818 	spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
3819 
3820 	cb = zio->io_private;
3821 	ASSERT(cb != NULL);
3822 	buf = cb->l2rcb_buf;
3823 	ASSERT(buf != NULL);
3824 	hdr = buf->b_hdr;
3825 	ASSERT(hdr != NULL);
3826 
3827 	hash_lock = HDR_LOCK(hdr);
3828 	mutex_enter(hash_lock);
3829 
3830 	/*
3831 	 * Check this survived the L2ARC journey.
3832 	 */
3833 	equal = arc_cksum_equal(buf);
3834 	if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
3835 		mutex_exit(hash_lock);
3836 		zio->io_private = buf;
3837 		zio->io_bp_copy = cb->l2rcb_bp;	/* XXX fix in L2ARC 2.0	*/
3838 		zio->io_bp = &zio->io_bp_copy;	/* XXX fix in L2ARC 2.0	*/
3839 		arc_read_done(zio);
3840 	} else {
3841 		mutex_exit(hash_lock);
3842 		/*
3843 		 * Buffer didn't survive caching.  Increment stats and
3844 		 * reissue to the original storage device.
3845 		 */
3846 		if (zio->io_error != 0) {
3847 			ARCSTAT_BUMP(arcstat_l2_io_error);
3848 		} else {
3849 			zio->io_error = EIO;
3850 		}
3851 		if (!equal)
3852 			ARCSTAT_BUMP(arcstat_l2_cksum_bad);
3853 
3854 		/*
3855 		 * If there's no waiter, issue an async i/o to the primary
3856 		 * storage now.  If there *is* a waiter, the caller must
3857 		 * issue the i/o in a context where it's OK to block.
3858 		 */
3859 		if (zio->io_waiter == NULL)
3860 			zio_nowait(zio_read(zio->io_parent,
3861 			    cb->l2rcb_spa, &cb->l2rcb_bp,
3862 			    buf->b_data, zio->io_size, arc_read_done, buf,
3863 			    zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
3864 	}
3865 
3866 	kmem_free(cb, sizeof (l2arc_read_callback_t));
3867 }
3868 
3869 /*
3870  * This is the list priority from which the L2ARC will search for pages to
3871  * cache.  This is used within loops (0..3) to cycle through lists in the
3872  * desired order.  This order can have a significant effect on cache
3873  * performance.
3874  *
3875  * Currently the metadata lists are hit first, MFU then MRU, followed by
3876  * the data lists.  This function returns a locked list, and also returns
3877  * the lock pointer.
3878  */
3879 static list_t *
3880 l2arc_list_locked(int list_num, kmutex_t **lock)
3881 {
3882 	list_t *list;
3883 
3884 	ASSERT(list_num >= 0 && list_num <= 3);
3885 
3886 	switch (list_num) {
3887 	case 0:
3888 		list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
3889 		*lock = &arc_mfu->arcs_mtx;
3890 		break;
3891 	case 1:
3892 		list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
3893 		*lock = &arc_mru->arcs_mtx;
3894 		break;
3895 	case 2:
3896 		list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
3897 		*lock = &arc_mfu->arcs_mtx;
3898 		break;
3899 	case 3:
3900 		list = &arc_mru->arcs_list[ARC_BUFC_DATA];
3901 		*lock = &arc_mru->arcs_mtx;
3902 		break;
3903 	}
3904 
3905 	ASSERT(!(MUTEX_HELD(*lock)));
3906 	mutex_enter(*lock);
3907 	return (list);
3908 }
3909 
3910 /*
3911  * Evict buffers from the device write hand to the distance specified in
3912  * bytes.  This distance may span populated buffers, it may span nothing.
3913  * This is clearing a region on the L2ARC device ready for writing.
3914  * If the 'all' boolean is set, every buffer is evicted.
3915  */
3916 static void
3917 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
3918 {
3919 	list_t *buflist;
3920 	l2arc_buf_hdr_t *abl2;
3921 	arc_buf_hdr_t *ab, *ab_prev;
3922 	kmutex_t *hash_lock;
3923 	uint64_t taddr;
3924 
3925 	buflist = dev->l2ad_buflist;
3926 
3927 	if (buflist == NULL)
3928 		return;
3929 
3930 	if (!all && dev->l2ad_first) {
3931 		/*
3932 		 * This is the first sweep through the device.  There is
3933 		 * nothing to evict.
3934 		 */
3935 		return;
3936 	}
3937 
3938 	if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
3939 		/*
3940 		 * When nearing the end of the device, evict to the end
3941 		 * before the device write hand jumps to the start.
3942 		 */
3943 		taddr = dev->l2ad_end;
3944 	} else {
3945 		taddr = dev->l2ad_hand + distance;
3946 	}
3947 	DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
3948 	    uint64_t, taddr, boolean_t, all);
3949 
3950 top:
3951 	mutex_enter(&l2arc_buflist_mtx);
3952 	for (ab = list_tail(buflist); ab; ab = ab_prev) {
3953 		ab_prev = list_prev(buflist, ab);
3954 
3955 		hash_lock = HDR_LOCK(ab);
3956 		if (!mutex_tryenter(hash_lock)) {
3957 			/*
3958 			 * Missed the hash lock.  Retry.
3959 			 */
3960 			ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
3961 			mutex_exit(&l2arc_buflist_mtx);
3962 			mutex_enter(hash_lock);
3963 			mutex_exit(hash_lock);
3964 			goto top;
3965 		}
3966 
3967 		if (HDR_L2_WRITE_HEAD(ab)) {
3968 			/*
3969 			 * We hit a write head node.  Leave it for
3970 			 * l2arc_write_done().
3971 			 */
3972 			list_remove(buflist, ab);
3973 			mutex_exit(hash_lock);
3974 			continue;
3975 		}
3976 
3977 		if (!all && ab->b_l2hdr != NULL &&
3978 		    (ab->b_l2hdr->b_daddr > taddr ||
3979 		    ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
3980 			/*
3981 			 * We've evicted to the target address,
3982 			 * or the end of the device.
3983 			 */
3984 			mutex_exit(hash_lock);
3985 			break;
3986 		}
3987 
3988 		if (HDR_FREE_IN_PROGRESS(ab)) {
3989 			/*
3990 			 * Already on the path to destruction.
3991 			 */
3992 			mutex_exit(hash_lock);
3993 			continue;
3994 		}
3995 
3996 		if (ab->b_state == arc_l2c_only) {
3997 			ASSERT(!HDR_L2_READING(ab));
3998 			/*
3999 			 * This doesn't exist in the ARC.  Destroy.
4000 			 * arc_hdr_destroy() will call list_remove()
4001 			 * and decrement arcstat_l2_size.
4002 			 */
4003 			arc_change_state(arc_anon, ab, hash_lock);
4004 			arc_hdr_destroy(ab);
4005 		} else {
4006 			/*
4007 			 * Invalidate issued or about to be issued
4008 			 * reads, since we may be about to write
4009 			 * over this location.
4010 			 */
4011 			if (HDR_L2_READING(ab)) {
4012 				ARCSTAT_BUMP(arcstat_l2_evict_reading);
4013 				ab->b_flags |= ARC_L2_EVICTED;
4014 			}
4015 
4016 			/*
4017 			 * Tell ARC this no longer exists in L2ARC.
4018 			 */
4019 			if (ab->b_l2hdr != NULL) {
4020 				abl2 = ab->b_l2hdr;
4021 				ab->b_l2hdr = NULL;
4022 				kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4023 				ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4024 			}
4025 			list_remove(buflist, ab);
4026 
4027 			/*
4028 			 * This may have been leftover after a
4029 			 * failed write.
4030 			 */
4031 			ab->b_flags &= ~ARC_L2_WRITING;
4032 		}
4033 		mutex_exit(hash_lock);
4034 	}
4035 	mutex_exit(&l2arc_buflist_mtx);
4036 
4037 	spa_l2cache_space_update(dev->l2ad_vdev, 0, -(taddr - dev->l2ad_evict));
4038 	dev->l2ad_evict = taddr;
4039 }
4040 
4041 /*
4042  * Find and write ARC buffers to the L2ARC device.
4043  *
4044  * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4045  * for reading until they have completed writing.
4046  */
4047 static void
4048 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
4049 {
4050 	arc_buf_hdr_t *ab, *ab_prev, *head;
4051 	l2arc_buf_hdr_t *hdrl2;
4052 	list_t *list;
4053 	uint64_t passed_sz, write_sz, buf_sz, headroom;
4054 	void *buf_data;
4055 	kmutex_t *hash_lock, *list_lock;
4056 	boolean_t have_lock, full;
4057 	l2arc_write_callback_t *cb;
4058 	zio_t *pio, *wzio;
4059 
4060 	ASSERT(dev->l2ad_vdev != NULL);
4061 
4062 	pio = NULL;
4063 	write_sz = 0;
4064 	full = B_FALSE;
4065 	head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4066 	head->b_flags |= ARC_L2_WRITE_HEAD;
4067 
4068 	/*
4069 	 * Copy buffers for L2ARC writing.
4070 	 */
4071 	mutex_enter(&l2arc_buflist_mtx);
4072 	for (int try = 0; try <= 3; try++) {
4073 		list = l2arc_list_locked(try, &list_lock);
4074 		passed_sz = 0;
4075 
4076 		/*
4077 		 * L2ARC fast warmup.
4078 		 *
4079 		 * Until the ARC is warm and starts to evict, read from the
4080 		 * head of the ARC lists rather than the tail.
4081 		 */
4082 		headroom = target_sz * l2arc_headroom;
4083 		if (arc_warm == B_FALSE)
4084 			ab = list_head(list);
4085 		else
4086 			ab = list_tail(list);
4087 
4088 		for (; ab; ab = ab_prev) {
4089 			if (arc_warm == B_FALSE)
4090 				ab_prev = list_next(list, ab);
4091 			else
4092 				ab_prev = list_prev(list, ab);
4093 
4094 			hash_lock = HDR_LOCK(ab);
4095 			have_lock = MUTEX_HELD(hash_lock);
4096 			if (!have_lock && !mutex_tryenter(hash_lock)) {
4097 				/*
4098 				 * Skip this buffer rather than waiting.
4099 				 */
4100 				continue;
4101 			}
4102 
4103 			passed_sz += ab->b_size;
4104 			if (passed_sz > headroom) {
4105 				/*
4106 				 * Searched too far.
4107 				 */
4108 				mutex_exit(hash_lock);
4109 				break;
4110 			}
4111 
4112 			if (ab->b_spa != spa) {
4113 				mutex_exit(hash_lock);
4114 				continue;
4115 			}
4116 
4117 			if (ab->b_l2hdr != NULL) {
4118 				/*
4119 				 * Already in L2ARC.
4120 				 */
4121 				mutex_exit(hash_lock);
4122 				continue;
4123 			}
4124 
4125 			if (HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab)) {
4126 				mutex_exit(hash_lock);
4127 				continue;
4128 			}
4129 
4130 			if ((write_sz + ab->b_size) > target_sz) {
4131 				full = B_TRUE;
4132 				mutex_exit(hash_lock);
4133 				break;
4134 			}
4135 
4136 			if (ab->b_buf == NULL) {
4137 				DTRACE_PROBE1(l2arc__buf__null, void *, ab);
4138 				mutex_exit(hash_lock);
4139 				continue;
4140 			}
4141 
4142 			if (pio == NULL) {
4143 				/*
4144 				 * Insert a dummy header on the buflist so
4145 				 * l2arc_write_done() can find where the
4146 				 * write buffers begin without searching.
4147 				 */
4148 				list_insert_head(dev->l2ad_buflist, head);
4149 
4150 				cb = kmem_alloc(
4151 				    sizeof (l2arc_write_callback_t), KM_SLEEP);
4152 				cb->l2wcb_dev = dev;
4153 				cb->l2wcb_head = head;
4154 				pio = zio_root(spa, l2arc_write_done, cb,
4155 				    ZIO_FLAG_CANFAIL);
4156 			}
4157 
4158 			/*
4159 			 * Create and add a new L2ARC header.
4160 			 */
4161 			hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4162 			hdrl2->b_dev = dev;
4163 			hdrl2->b_daddr = dev->l2ad_hand;
4164 
4165 			ab->b_flags |= ARC_L2_WRITING;
4166 			ab->b_l2hdr = hdrl2;
4167 			list_insert_head(dev->l2ad_buflist, ab);
4168 			buf_data = ab->b_buf->b_data;
4169 			buf_sz = ab->b_size;
4170 
4171 			/*
4172 			 * Compute and store the buffer cksum before
4173 			 * writing.  On debug the cksum is verified first.
4174 			 */
4175 			arc_cksum_verify(ab->b_buf);
4176 			arc_cksum_compute(ab->b_buf, B_TRUE);
4177 
4178 			mutex_exit(hash_lock);
4179 
4180 			wzio = zio_write_phys(pio, dev->l2ad_vdev,
4181 			    dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4182 			    NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4183 			    ZIO_FLAG_CANFAIL, B_FALSE);
4184 
4185 			DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4186 			    zio_t *, wzio);
4187 			(void) zio_nowait(wzio);
4188 
4189 			/*
4190 			 * Keep the clock hand suitably device-aligned.
4191 			 */
4192 			buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4193 
4194 			write_sz += buf_sz;
4195 			dev->l2ad_hand += buf_sz;
4196 		}
4197 
4198 		mutex_exit(list_lock);
4199 
4200 		if (full == B_TRUE)
4201 			break;
4202 	}
4203 	mutex_exit(&l2arc_buflist_mtx);
4204 
4205 	if (pio == NULL) {
4206 		ASSERT3U(write_sz, ==, 0);
4207 		kmem_cache_free(hdr_cache, head);
4208 		return;
4209 	}
4210 
4211 	ASSERT3U(write_sz, <=, target_sz);
4212 	ARCSTAT_BUMP(arcstat_l2_writes_sent);
4213 	ARCSTAT_INCR(arcstat_l2_size, write_sz);
4214 	spa_l2cache_space_update(dev->l2ad_vdev, 0, write_sz);
4215 
4216 	/*
4217 	 * Bump device hand to the device start if it is approaching the end.
4218 	 * l2arc_evict() will already have evicted ahead for this case.
4219 	 */
4220 	if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4221 		spa_l2cache_space_update(dev->l2ad_vdev, 0,
4222 		    dev->l2ad_end - dev->l2ad_hand);
4223 		dev->l2ad_hand = dev->l2ad_start;
4224 		dev->l2ad_evict = dev->l2ad_start;
4225 		dev->l2ad_first = B_FALSE;
4226 	}
4227 
4228 	(void) zio_wait(pio);
4229 }
4230 
4231 /*
4232  * This thread feeds the L2ARC at regular intervals.  This is the beating
4233  * heart of the L2ARC.
4234  */
4235 static void
4236 l2arc_feed_thread(void)
4237 {
4238 	callb_cpr_t cpr;
4239 	l2arc_dev_t *dev;
4240 	spa_t *spa;
4241 	uint64_t size;
4242 
4243 	CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4244 
4245 	mutex_enter(&l2arc_feed_thr_lock);
4246 
4247 	while (l2arc_thread_exit == 0) {
4248 		/*
4249 		 * Pause for l2arc_feed_secs seconds between writes.
4250 		 */
4251 		CALLB_CPR_SAFE_BEGIN(&cpr);
4252 		(void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
4253 		    lbolt + (hz * l2arc_feed_secs));
4254 		CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4255 
4256 		/*
4257 		 * Quick check for L2ARC devices.
4258 		 */
4259 		mutex_enter(&l2arc_dev_mtx);
4260 		if (l2arc_ndev == 0) {
4261 			mutex_exit(&l2arc_dev_mtx);
4262 			continue;
4263 		}
4264 		mutex_exit(&l2arc_dev_mtx);
4265 
4266 		/*
4267 		 * This selects the next l2arc device to write to, and in
4268 		 * doing so the next spa to feed from: dev->l2ad_spa.   This
4269 		 * will return NULL if there are now no l2arc devices or if
4270 		 * they are all faulted.
4271 		 *
4272 		 * If a device is returned, its spa's config lock is also
4273 		 * held to prevent device removal.  l2arc_dev_get_next()
4274 		 * will grab and release l2arc_dev_mtx.
4275 		 */
4276 		if ((dev = l2arc_dev_get_next()) == NULL)
4277 			continue;
4278 
4279 		spa = dev->l2ad_spa;
4280 		ASSERT(spa != NULL);
4281 
4282 		/*
4283 		 * Avoid contributing to memory pressure.
4284 		 */
4285 		if (arc_reclaim_needed()) {
4286 			ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4287 			spa_config_exit(spa, SCL_L2ARC, dev);
4288 			continue;
4289 		}
4290 
4291 		ARCSTAT_BUMP(arcstat_l2_feeds);
4292 
4293 		size = dev->l2ad_write;
4294 		if (arc_warm == B_FALSE)
4295 			size += dev->l2ad_boost;
4296 
4297 		/*
4298 		 * Evict L2ARC buffers that will be overwritten.
4299 		 */
4300 		l2arc_evict(dev, size, B_FALSE);
4301 
4302 		/*
4303 		 * Write ARC buffers.
4304 		 */
4305 		l2arc_write_buffers(spa, dev, size);
4306 		spa_config_exit(spa, SCL_L2ARC, dev);
4307 	}
4308 
4309 	l2arc_thread_exit = 0;
4310 	cv_broadcast(&l2arc_feed_thr_cv);
4311 	CALLB_CPR_EXIT(&cpr);		/* drops l2arc_feed_thr_lock */
4312 	thread_exit();
4313 }
4314 
4315 boolean_t
4316 l2arc_vdev_present(vdev_t *vd)
4317 {
4318 	l2arc_dev_t *dev;
4319 
4320 	mutex_enter(&l2arc_dev_mtx);
4321 	for (dev = list_head(l2arc_dev_list); dev != NULL;
4322 	    dev = list_next(l2arc_dev_list, dev)) {
4323 		if (dev->l2ad_vdev == vd)
4324 			break;
4325 	}
4326 	mutex_exit(&l2arc_dev_mtx);
4327 
4328 	return (dev != NULL);
4329 }
4330 
4331 /*
4332  * Add a vdev for use by the L2ARC.  By this point the spa has already
4333  * validated the vdev and opened it.
4334  */
4335 void
4336 l2arc_add_vdev(spa_t *spa, vdev_t *vd, uint64_t start, uint64_t end)
4337 {
4338 	l2arc_dev_t *adddev;
4339 
4340 	ASSERT(!l2arc_vdev_present(vd));
4341 
4342 	/*
4343 	 * Create a new l2arc device entry.
4344 	 */
4345 	adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
4346 	adddev->l2ad_spa = spa;
4347 	adddev->l2ad_vdev = vd;
4348 	adddev->l2ad_write = l2arc_write_max;
4349 	adddev->l2ad_boost = l2arc_write_boost;
4350 	adddev->l2ad_start = start;
4351 	adddev->l2ad_end = end;
4352 	adddev->l2ad_hand = adddev->l2ad_start;
4353 	adddev->l2ad_evict = adddev->l2ad_start;
4354 	adddev->l2ad_first = B_TRUE;
4355 	ASSERT3U(adddev->l2ad_write, >, 0);
4356 
4357 	/*
4358 	 * This is a list of all ARC buffers that are still valid on the
4359 	 * device.
4360 	 */
4361 	adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
4362 	list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
4363 	    offsetof(arc_buf_hdr_t, b_l2node));
4364 
4365 	spa_l2cache_space_update(vd, adddev->l2ad_end - adddev->l2ad_hand, 0);
4366 
4367 	/*
4368 	 * Add device to global list
4369 	 */
4370 	mutex_enter(&l2arc_dev_mtx);
4371 	list_insert_head(l2arc_dev_list, adddev);
4372 	atomic_inc_64(&l2arc_ndev);
4373 	mutex_exit(&l2arc_dev_mtx);
4374 }
4375 
4376 /*
4377  * Remove a vdev from the L2ARC.
4378  */
4379 void
4380 l2arc_remove_vdev(vdev_t *vd)
4381 {
4382 	l2arc_dev_t *dev, *nextdev, *remdev = NULL;
4383 
4384 	/*
4385 	 * Find the device by vdev
4386 	 */
4387 	mutex_enter(&l2arc_dev_mtx);
4388 	for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
4389 		nextdev = list_next(l2arc_dev_list, dev);
4390 		if (vd == dev->l2ad_vdev) {
4391 			remdev = dev;
4392 			break;
4393 		}
4394 	}
4395 	ASSERT(remdev != NULL);
4396 
4397 	/*
4398 	 * Remove device from global list
4399 	 */
4400 	list_remove(l2arc_dev_list, remdev);
4401 	l2arc_dev_last = NULL;		/* may have been invalidated */
4402 	atomic_dec_64(&l2arc_ndev);
4403 	mutex_exit(&l2arc_dev_mtx);
4404 
4405 	/*
4406 	 * Clear all buflists and ARC references.  L2ARC device flush.
4407 	 */
4408 	l2arc_evict(remdev, 0, B_TRUE);
4409 	list_destroy(remdev->l2ad_buflist);
4410 	kmem_free(remdev->l2ad_buflist, sizeof (list_t));
4411 	kmem_free(remdev, sizeof (l2arc_dev_t));
4412 }
4413 
4414 void
4415 l2arc_init(void)
4416 {
4417 	l2arc_thread_exit = 0;
4418 	l2arc_ndev = 0;
4419 	l2arc_writes_sent = 0;
4420 	l2arc_writes_done = 0;
4421 
4422 	mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
4423 	cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
4424 	mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
4425 	mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
4426 	mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
4427 
4428 	l2arc_dev_list = &L2ARC_dev_list;
4429 	l2arc_free_on_write = &L2ARC_free_on_write;
4430 	list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
4431 	    offsetof(l2arc_dev_t, l2ad_node));
4432 	list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
4433 	    offsetof(l2arc_data_free_t, l2df_list_node));
4434 }
4435 
4436 void
4437 l2arc_fini(void)
4438 {
4439 	/*
4440 	 * This is called from dmu_fini(), which is called from spa_fini();
4441 	 * Because of this, we can assume that all l2arc devices have
4442 	 * already been removed when the pools themselves were removed.
4443 	 */
4444 
4445 	l2arc_do_free_on_write();
4446 
4447 	mutex_destroy(&l2arc_feed_thr_lock);
4448 	cv_destroy(&l2arc_feed_thr_cv);
4449 	mutex_destroy(&l2arc_dev_mtx);
4450 	mutex_destroy(&l2arc_buflist_mtx);
4451 	mutex_destroy(&l2arc_free_on_write_mtx);
4452 
4453 	list_destroy(l2arc_dev_list);
4454 	list_destroy(l2arc_free_on_write);
4455 }
4456 
4457 void
4458 l2arc_start(void)
4459 {
4460 	if (!(spa_mode & FWRITE))
4461 		return;
4462 
4463 	(void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
4464 	    TS_RUN, minclsyspri);
4465 }
4466 
4467 void
4468 l2arc_stop(void)
4469 {
4470 	if (!(spa_mode & FWRITE))
4471 		return;
4472 
4473 	mutex_enter(&l2arc_feed_thr_lock);
4474 	cv_signal(&l2arc_feed_thr_cv);	/* kick thread out of startup */
4475 	l2arc_thread_exit = 1;
4476 	while (l2arc_thread_exit != 0)
4477 		cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
4478 	mutex_exit(&l2arc_feed_thr_lock);
4479 }
4480