xref: /titanic_41/usr/src/uts/common/fs/zfs/arc.c (revision 18ba57035593bcafd166b6540f5881f7ecfca174)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2012, Joyent, Inc. All rights reserved.
24  * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
25  * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26  * Copyright 2014 Nexenta Systems, Inc.  All rights reserved.
27  */
28 
29 /*
30  * DVA-based Adjustable Replacement Cache
31  *
32  * While much of the theory of operation used here is
33  * based on the self-tuning, low overhead replacement cache
34  * presented by Megiddo and Modha at FAST 2003, there are some
35  * significant differences:
36  *
37  * 1. The Megiddo and Modha model assumes any page is evictable.
38  * Pages in its cache cannot be "locked" into memory.  This makes
39  * the eviction algorithm simple: evict the last page in the list.
40  * This also make the performance characteristics easy to reason
41  * about.  Our cache is not so simple.  At any given moment, some
42  * subset of the blocks in the cache are un-evictable because we
43  * have handed out a reference to them.  Blocks are only evictable
44  * when there are no external references active.  This makes
45  * eviction far more problematic:  we choose to evict the evictable
46  * blocks that are the "lowest" in the list.
47  *
48  * There are times when it is not possible to evict the requested
49  * space.  In these circumstances we are unable to adjust the cache
50  * size.  To prevent the cache growing unbounded at these times we
51  * implement a "cache throttle" that slows the flow of new data
52  * into the cache until we can make space available.
53  *
54  * 2. The Megiddo and Modha model assumes a fixed cache size.
55  * Pages are evicted when the cache is full and there is a cache
56  * miss.  Our model has a variable sized cache.  It grows with
57  * high use, but also tries to react to memory pressure from the
58  * operating system: decreasing its size when system memory is
59  * tight.
60  *
61  * 3. The Megiddo and Modha model assumes a fixed page size. All
62  * elements of the cache are therefore exactly the same size.  So
63  * when adjusting the cache size following a cache miss, its simply
64  * a matter of choosing a single page to evict.  In our model, we
65  * have variable sized cache blocks (rangeing from 512 bytes to
66  * 128K bytes).  We therefore choose a set of blocks to evict to make
67  * space for a cache miss that approximates as closely as possible
68  * the space used by the new block.
69  *
70  * See also:  "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71  * by N. Megiddo & D. Modha, FAST 2003
72  */
73 
74 /*
75  * The locking model:
76  *
77  * A new reference to a cache buffer can be obtained in two
78  * ways: 1) via a hash table lookup using the DVA as a key,
79  * or 2) via one of the ARC lists.  The arc_read() interface
80  * uses method 1, while the internal arc algorithms for
81  * adjusting the cache use method 2.  We therefore provide two
82  * types of locks: 1) the hash table lock array, and 2) the
83  * arc list locks.
84  *
85  * Buffers do not have their own mutexes, rather they rely on the
86  * hash table mutexes for the bulk of their protection (i.e. most
87  * fields in the arc_buf_hdr_t are protected by these mutexes).
88  *
89  * buf_hash_find() returns the appropriate mutex (held) when it
90  * locates the requested buffer in the hash table.  It returns
91  * NULL for the mutex if the buffer was not in the table.
92  *
93  * buf_hash_remove() expects the appropriate hash mutex to be
94  * already held before it is invoked.
95  *
96  * Each arc state also has a mutex which is used to protect the
97  * buffer list associated with the state.  When attempting to
98  * obtain a hash table lock while holding an arc list lock you
99  * must use: mutex_tryenter() to avoid deadlock.  Also note that
100  * the active state mutex must be held before the ghost state mutex.
101  *
102  * Arc buffers may have an associated eviction callback function.
103  * This function will be invoked prior to removing the buffer (e.g.
104  * in arc_do_user_evicts()).  Note however that the data associated
105  * with the buffer may be evicted prior to the callback.  The callback
106  * must be made with *no locks held* (to prevent deadlock).  Additionally,
107  * the users of callbacks must ensure that their private data is
108  * protected from simultaneous callbacks from arc_clear_callback()
109  * and arc_do_user_evicts().
110  *
111  * Note that the majority of the performance stats are manipulated
112  * with atomic operations.
113  *
114  * The L2ARC uses the l2ad_mtx on each vdev for the following:
115  *
116  *	- L2ARC buflist creation
117  *	- L2ARC buflist eviction
118  *	- L2ARC write completion, which walks L2ARC buflists
119  *	- ARC header destruction, as it removes from L2ARC buflists
120  *	- ARC header release, as it removes from L2ARC buflists
121  */
122 
123 #include <sys/spa.h>
124 #include <sys/zio.h>
125 #include <sys/zio_compress.h>
126 #include <sys/zfs_context.h>
127 #include <sys/arc.h>
128 #include <sys/refcount.h>
129 #include <sys/vdev.h>
130 #include <sys/vdev_impl.h>
131 #include <sys/dsl_pool.h>
132 #include <sys/multilist.h>
133 #ifdef _KERNEL
134 #include <sys/vmsystm.h>
135 #include <vm/anon.h>
136 #include <sys/fs/swapnode.h>
137 #include <sys/dnlc.h>
138 #endif
139 #include <sys/callb.h>
140 #include <sys/kstat.h>
141 #include <zfs_fletcher.h>
142 
143 #ifndef _KERNEL
144 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
145 boolean_t arc_watch = B_FALSE;
146 int arc_procfd;
147 #endif
148 
149 static kmutex_t		arc_reclaim_lock;
150 static kcondvar_t	arc_reclaim_thread_cv;
151 static boolean_t	arc_reclaim_thread_exit;
152 static kcondvar_t	arc_reclaim_waiters_cv;
153 
154 static kmutex_t		arc_user_evicts_lock;
155 static kcondvar_t	arc_user_evicts_cv;
156 static boolean_t	arc_user_evicts_thread_exit;
157 
158 uint_t arc_reduce_dnlc_percent = 3;
159 
160 /*
161  * The number of headers to evict in arc_evict_state_impl() before
162  * dropping the sublist lock and evicting from another sublist. A lower
163  * value means we're more likely to evict the "correct" header (i.e. the
164  * oldest header in the arc state), but comes with higher overhead
165  * (i.e. more invocations of arc_evict_state_impl()).
166  */
167 int zfs_arc_evict_batch_limit = 10;
168 
169 /*
170  * The number of sublists used for each of the arc state lists. If this
171  * is not set to a suitable value by the user, it will be configured to
172  * the number of CPUs on the system in arc_init().
173  */
174 int zfs_arc_num_sublists_per_state = 0;
175 
176 /* number of seconds before growing cache again */
177 static int		arc_grow_retry = 60;
178 
179 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */
180 int		zfs_arc_overflow_shift = 8;
181 
182 /* shift of arc_c for calculating both min and max arc_p */
183 static int		arc_p_min_shift = 4;
184 
185 /* log2(fraction of arc to reclaim) */
186 static int		arc_shrink_shift = 7;
187 
188 /*
189  * log2(fraction of ARC which must be free to allow growing).
190  * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
191  * when reading a new block into the ARC, we will evict an equal-sized block
192  * from the ARC.
193  *
194  * This must be less than arc_shrink_shift, so that when we shrink the ARC,
195  * we will still not allow it to grow.
196  */
197 int			arc_no_grow_shift = 5;
198 
199 
200 /*
201  * minimum lifespan of a prefetch block in clock ticks
202  * (initialized in arc_init())
203  */
204 static int		arc_min_prefetch_lifespan;
205 
206 /*
207  * If this percent of memory is free, don't throttle.
208  */
209 int arc_lotsfree_percent = 10;
210 
211 static int arc_dead;
212 
213 /*
214  * The arc has filled available memory and has now warmed up.
215  */
216 static boolean_t arc_warm;
217 
218 /*
219  * These tunables are for performance analysis.
220  */
221 uint64_t zfs_arc_max;
222 uint64_t zfs_arc_min;
223 uint64_t zfs_arc_meta_limit = 0;
224 uint64_t zfs_arc_meta_min = 0;
225 int zfs_arc_grow_retry = 0;
226 int zfs_arc_shrink_shift = 0;
227 int zfs_arc_p_min_shift = 0;
228 int zfs_disable_dup_eviction = 0;
229 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
230 
231 /*
232  * Note that buffers can be in one of 6 states:
233  *	ARC_anon	- anonymous (discussed below)
234  *	ARC_mru		- recently used, currently cached
235  *	ARC_mru_ghost	- recentely used, no longer in cache
236  *	ARC_mfu		- frequently used, currently cached
237  *	ARC_mfu_ghost	- frequently used, no longer in cache
238  *	ARC_l2c_only	- exists in L2ARC but not other states
239  * When there are no active references to the buffer, they are
240  * are linked onto a list in one of these arc states.  These are
241  * the only buffers that can be evicted or deleted.  Within each
242  * state there are multiple lists, one for meta-data and one for
243  * non-meta-data.  Meta-data (indirect blocks, blocks of dnodes,
244  * etc.) is tracked separately so that it can be managed more
245  * explicitly: favored over data, limited explicitly.
246  *
247  * Anonymous buffers are buffers that are not associated with
248  * a DVA.  These are buffers that hold dirty block copies
249  * before they are written to stable storage.  By definition,
250  * they are "ref'd" and are considered part of arc_mru
251  * that cannot be freed.  Generally, they will aquire a DVA
252  * as they are written and migrate onto the arc_mru list.
253  *
254  * The ARC_l2c_only state is for buffers that are in the second
255  * level ARC but no longer in any of the ARC_m* lists.  The second
256  * level ARC itself may also contain buffers that are in any of
257  * the ARC_m* states - meaning that a buffer can exist in two
258  * places.  The reason for the ARC_l2c_only state is to keep the
259  * buffer header in the hash table, so that reads that hit the
260  * second level ARC benefit from these fast lookups.
261  */
262 
263 typedef struct arc_state {
264 	/*
265 	 * list of evictable buffers
266 	 */
267 	multilist_t arcs_list[ARC_BUFC_NUMTYPES];
268 	/*
269 	 * total amount of evictable data in this state
270 	 */
271 	uint64_t arcs_lsize[ARC_BUFC_NUMTYPES];
272 	/*
273 	 * total amount of data in this state; this includes: evictable,
274 	 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
275 	 */
276 	refcount_t arcs_size;
277 } arc_state_t;
278 
279 /* The 6 states: */
280 static arc_state_t ARC_anon;
281 static arc_state_t ARC_mru;
282 static arc_state_t ARC_mru_ghost;
283 static arc_state_t ARC_mfu;
284 static arc_state_t ARC_mfu_ghost;
285 static arc_state_t ARC_l2c_only;
286 
287 typedef struct arc_stats {
288 	kstat_named_t arcstat_hits;
289 	kstat_named_t arcstat_misses;
290 	kstat_named_t arcstat_demand_data_hits;
291 	kstat_named_t arcstat_demand_data_misses;
292 	kstat_named_t arcstat_demand_metadata_hits;
293 	kstat_named_t arcstat_demand_metadata_misses;
294 	kstat_named_t arcstat_prefetch_data_hits;
295 	kstat_named_t arcstat_prefetch_data_misses;
296 	kstat_named_t arcstat_prefetch_metadata_hits;
297 	kstat_named_t arcstat_prefetch_metadata_misses;
298 	kstat_named_t arcstat_mru_hits;
299 	kstat_named_t arcstat_mru_ghost_hits;
300 	kstat_named_t arcstat_mfu_hits;
301 	kstat_named_t arcstat_mfu_ghost_hits;
302 	kstat_named_t arcstat_deleted;
303 	/*
304 	 * Number of buffers that could not be evicted because the hash lock
305 	 * was held by another thread.  The lock may not necessarily be held
306 	 * by something using the same buffer, since hash locks are shared
307 	 * by multiple buffers.
308 	 */
309 	kstat_named_t arcstat_mutex_miss;
310 	/*
311 	 * Number of buffers skipped because they have I/O in progress, are
312 	 * indrect prefetch buffers that have not lived long enough, or are
313 	 * not from the spa we're trying to evict from.
314 	 */
315 	kstat_named_t arcstat_evict_skip;
316 	/*
317 	 * Number of times arc_evict_state() was unable to evict enough
318 	 * buffers to reach it's target amount.
319 	 */
320 	kstat_named_t arcstat_evict_not_enough;
321 	kstat_named_t arcstat_evict_l2_cached;
322 	kstat_named_t arcstat_evict_l2_eligible;
323 	kstat_named_t arcstat_evict_l2_ineligible;
324 	kstat_named_t arcstat_evict_l2_skip;
325 	kstat_named_t arcstat_hash_elements;
326 	kstat_named_t arcstat_hash_elements_max;
327 	kstat_named_t arcstat_hash_collisions;
328 	kstat_named_t arcstat_hash_chains;
329 	kstat_named_t arcstat_hash_chain_max;
330 	kstat_named_t arcstat_p;
331 	kstat_named_t arcstat_c;
332 	kstat_named_t arcstat_c_min;
333 	kstat_named_t arcstat_c_max;
334 	kstat_named_t arcstat_size;
335 	/*
336 	 * Number of bytes consumed by internal ARC structures necessary
337 	 * for tracking purposes; these structures are not actually
338 	 * backed by ARC buffers. This includes arc_buf_hdr_t structures
339 	 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
340 	 * caches), and arc_buf_t structures (allocated via arc_buf_t
341 	 * cache).
342 	 */
343 	kstat_named_t arcstat_hdr_size;
344 	/*
345 	 * Number of bytes consumed by ARC buffers of type equal to
346 	 * ARC_BUFC_DATA. This is generally consumed by buffers backing
347 	 * on disk user data (e.g. plain file contents).
348 	 */
349 	kstat_named_t arcstat_data_size;
350 	/*
351 	 * Number of bytes consumed by ARC buffers of type equal to
352 	 * ARC_BUFC_METADATA. This is generally consumed by buffers
353 	 * backing on disk data that is used for internal ZFS
354 	 * structures (e.g. ZAP, dnode, indirect blocks, etc).
355 	 */
356 	kstat_named_t arcstat_metadata_size;
357 	/*
358 	 * Number of bytes consumed by various buffers and structures
359 	 * not actually backed with ARC buffers. This includes bonus
360 	 * buffers (allocated directly via zio_buf_* functions),
361 	 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
362 	 * cache), and dnode_t structures (allocated via dnode_t cache).
363 	 */
364 	kstat_named_t arcstat_other_size;
365 	/*
366 	 * Total number of bytes consumed by ARC buffers residing in the
367 	 * arc_anon state. This includes *all* buffers in the arc_anon
368 	 * state; e.g. data, metadata, evictable, and unevictable buffers
369 	 * are all included in this value.
370 	 */
371 	kstat_named_t arcstat_anon_size;
372 	/*
373 	 * Number of bytes consumed by ARC buffers that meet the
374 	 * following criteria: backing buffers of type ARC_BUFC_DATA,
375 	 * residing in the arc_anon state, and are eligible for eviction
376 	 * (e.g. have no outstanding holds on the buffer).
377 	 */
378 	kstat_named_t arcstat_anon_evictable_data;
379 	/*
380 	 * Number of bytes consumed by ARC buffers that meet the
381 	 * following criteria: backing buffers of type ARC_BUFC_METADATA,
382 	 * residing in the arc_anon state, and are eligible for eviction
383 	 * (e.g. have no outstanding holds on the buffer).
384 	 */
385 	kstat_named_t arcstat_anon_evictable_metadata;
386 	/*
387 	 * Total number of bytes consumed by ARC buffers residing in the
388 	 * arc_mru state. This includes *all* buffers in the arc_mru
389 	 * state; e.g. data, metadata, evictable, and unevictable buffers
390 	 * are all included in this value.
391 	 */
392 	kstat_named_t arcstat_mru_size;
393 	/*
394 	 * Number of bytes consumed by ARC buffers that meet the
395 	 * following criteria: backing buffers of type ARC_BUFC_DATA,
396 	 * residing in the arc_mru state, and are eligible for eviction
397 	 * (e.g. have no outstanding holds on the buffer).
398 	 */
399 	kstat_named_t arcstat_mru_evictable_data;
400 	/*
401 	 * Number of bytes consumed by ARC buffers that meet the
402 	 * following criteria: backing buffers of type ARC_BUFC_METADATA,
403 	 * residing in the arc_mru state, and are eligible for eviction
404 	 * (e.g. have no outstanding holds on the buffer).
405 	 */
406 	kstat_named_t arcstat_mru_evictable_metadata;
407 	/*
408 	 * Total number of bytes that *would have been* consumed by ARC
409 	 * buffers in the arc_mru_ghost state. The key thing to note
410 	 * here, is the fact that this size doesn't actually indicate
411 	 * RAM consumption. The ghost lists only consist of headers and
412 	 * don't actually have ARC buffers linked off of these headers.
413 	 * Thus, *if* the headers had associated ARC buffers, these
414 	 * buffers *would have* consumed this number of bytes.
415 	 */
416 	kstat_named_t arcstat_mru_ghost_size;
417 	/*
418 	 * Number of bytes that *would have been* consumed by ARC
419 	 * buffers that are eligible for eviction, of type
420 	 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
421 	 */
422 	kstat_named_t arcstat_mru_ghost_evictable_data;
423 	/*
424 	 * Number of bytes that *would have been* consumed by ARC
425 	 * buffers that are eligible for eviction, of type
426 	 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
427 	 */
428 	kstat_named_t arcstat_mru_ghost_evictable_metadata;
429 	/*
430 	 * Total number of bytes consumed by ARC buffers residing in the
431 	 * arc_mfu state. This includes *all* buffers in the arc_mfu
432 	 * state; e.g. data, metadata, evictable, and unevictable buffers
433 	 * are all included in this value.
434 	 */
435 	kstat_named_t arcstat_mfu_size;
436 	/*
437 	 * Number of bytes consumed by ARC buffers that are eligible for
438 	 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
439 	 * state.
440 	 */
441 	kstat_named_t arcstat_mfu_evictable_data;
442 	/*
443 	 * Number of bytes consumed by ARC buffers that are eligible for
444 	 * eviction, of type ARC_BUFC_METADATA, and reside in the
445 	 * arc_mfu state.
446 	 */
447 	kstat_named_t arcstat_mfu_evictable_metadata;
448 	/*
449 	 * Total number of bytes that *would have been* consumed by ARC
450 	 * buffers in the arc_mfu_ghost state. See the comment above
451 	 * arcstat_mru_ghost_size for more details.
452 	 */
453 	kstat_named_t arcstat_mfu_ghost_size;
454 	/*
455 	 * Number of bytes that *would have been* consumed by ARC
456 	 * buffers that are eligible for eviction, of type
457 	 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
458 	 */
459 	kstat_named_t arcstat_mfu_ghost_evictable_data;
460 	/*
461 	 * Number of bytes that *would have been* consumed by ARC
462 	 * buffers that are eligible for eviction, of type
463 	 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
464 	 */
465 	kstat_named_t arcstat_mfu_ghost_evictable_metadata;
466 	kstat_named_t arcstat_l2_hits;
467 	kstat_named_t arcstat_l2_misses;
468 	kstat_named_t arcstat_l2_feeds;
469 	kstat_named_t arcstat_l2_rw_clash;
470 	kstat_named_t arcstat_l2_read_bytes;
471 	kstat_named_t arcstat_l2_write_bytes;
472 	kstat_named_t arcstat_l2_writes_sent;
473 	kstat_named_t arcstat_l2_writes_done;
474 	kstat_named_t arcstat_l2_writes_error;
475 	kstat_named_t arcstat_l2_writes_lock_retry;
476 	kstat_named_t arcstat_l2_evict_lock_retry;
477 	kstat_named_t arcstat_l2_evict_reading;
478 	kstat_named_t arcstat_l2_evict_l1cached;
479 	kstat_named_t arcstat_l2_free_on_write;
480 	kstat_named_t arcstat_l2_cdata_free_on_write;
481 	kstat_named_t arcstat_l2_abort_lowmem;
482 	kstat_named_t arcstat_l2_cksum_bad;
483 	kstat_named_t arcstat_l2_io_error;
484 	kstat_named_t arcstat_l2_size;
485 	kstat_named_t arcstat_l2_asize;
486 	kstat_named_t arcstat_l2_hdr_size;
487 	kstat_named_t arcstat_l2_compress_successes;
488 	kstat_named_t arcstat_l2_compress_zeros;
489 	kstat_named_t arcstat_l2_compress_failures;
490 	kstat_named_t arcstat_memory_throttle_count;
491 	kstat_named_t arcstat_duplicate_buffers;
492 	kstat_named_t arcstat_duplicate_buffers_size;
493 	kstat_named_t arcstat_duplicate_reads;
494 	kstat_named_t arcstat_meta_used;
495 	kstat_named_t arcstat_meta_limit;
496 	kstat_named_t arcstat_meta_max;
497 	kstat_named_t arcstat_meta_min;
498 } arc_stats_t;
499 
500 static arc_stats_t arc_stats = {
501 	{ "hits",			KSTAT_DATA_UINT64 },
502 	{ "misses",			KSTAT_DATA_UINT64 },
503 	{ "demand_data_hits",		KSTAT_DATA_UINT64 },
504 	{ "demand_data_misses",		KSTAT_DATA_UINT64 },
505 	{ "demand_metadata_hits",	KSTAT_DATA_UINT64 },
506 	{ "demand_metadata_misses",	KSTAT_DATA_UINT64 },
507 	{ "prefetch_data_hits",		KSTAT_DATA_UINT64 },
508 	{ "prefetch_data_misses",	KSTAT_DATA_UINT64 },
509 	{ "prefetch_metadata_hits",	KSTAT_DATA_UINT64 },
510 	{ "prefetch_metadata_misses",	KSTAT_DATA_UINT64 },
511 	{ "mru_hits",			KSTAT_DATA_UINT64 },
512 	{ "mru_ghost_hits",		KSTAT_DATA_UINT64 },
513 	{ "mfu_hits",			KSTAT_DATA_UINT64 },
514 	{ "mfu_ghost_hits",		KSTAT_DATA_UINT64 },
515 	{ "deleted",			KSTAT_DATA_UINT64 },
516 	{ "mutex_miss",			KSTAT_DATA_UINT64 },
517 	{ "evict_skip",			KSTAT_DATA_UINT64 },
518 	{ "evict_not_enough",		KSTAT_DATA_UINT64 },
519 	{ "evict_l2_cached",		KSTAT_DATA_UINT64 },
520 	{ "evict_l2_eligible",		KSTAT_DATA_UINT64 },
521 	{ "evict_l2_ineligible",	KSTAT_DATA_UINT64 },
522 	{ "evict_l2_skip",		KSTAT_DATA_UINT64 },
523 	{ "hash_elements",		KSTAT_DATA_UINT64 },
524 	{ "hash_elements_max",		KSTAT_DATA_UINT64 },
525 	{ "hash_collisions",		KSTAT_DATA_UINT64 },
526 	{ "hash_chains",		KSTAT_DATA_UINT64 },
527 	{ "hash_chain_max",		KSTAT_DATA_UINT64 },
528 	{ "p",				KSTAT_DATA_UINT64 },
529 	{ "c",				KSTAT_DATA_UINT64 },
530 	{ "c_min",			KSTAT_DATA_UINT64 },
531 	{ "c_max",			KSTAT_DATA_UINT64 },
532 	{ "size",			KSTAT_DATA_UINT64 },
533 	{ "hdr_size",			KSTAT_DATA_UINT64 },
534 	{ "data_size",			KSTAT_DATA_UINT64 },
535 	{ "metadata_size",		KSTAT_DATA_UINT64 },
536 	{ "other_size",			KSTAT_DATA_UINT64 },
537 	{ "anon_size",			KSTAT_DATA_UINT64 },
538 	{ "anon_evictable_data",	KSTAT_DATA_UINT64 },
539 	{ "anon_evictable_metadata",	KSTAT_DATA_UINT64 },
540 	{ "mru_size",			KSTAT_DATA_UINT64 },
541 	{ "mru_evictable_data",		KSTAT_DATA_UINT64 },
542 	{ "mru_evictable_metadata",	KSTAT_DATA_UINT64 },
543 	{ "mru_ghost_size",		KSTAT_DATA_UINT64 },
544 	{ "mru_ghost_evictable_data",	KSTAT_DATA_UINT64 },
545 	{ "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
546 	{ "mfu_size",			KSTAT_DATA_UINT64 },
547 	{ "mfu_evictable_data",		KSTAT_DATA_UINT64 },
548 	{ "mfu_evictable_metadata",	KSTAT_DATA_UINT64 },
549 	{ "mfu_ghost_size",		KSTAT_DATA_UINT64 },
550 	{ "mfu_ghost_evictable_data",	KSTAT_DATA_UINT64 },
551 	{ "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
552 	{ "l2_hits",			KSTAT_DATA_UINT64 },
553 	{ "l2_misses",			KSTAT_DATA_UINT64 },
554 	{ "l2_feeds",			KSTAT_DATA_UINT64 },
555 	{ "l2_rw_clash",		KSTAT_DATA_UINT64 },
556 	{ "l2_read_bytes",		KSTAT_DATA_UINT64 },
557 	{ "l2_write_bytes",		KSTAT_DATA_UINT64 },
558 	{ "l2_writes_sent",		KSTAT_DATA_UINT64 },
559 	{ "l2_writes_done",		KSTAT_DATA_UINT64 },
560 	{ "l2_writes_error",		KSTAT_DATA_UINT64 },
561 	{ "l2_writes_lock_retry",	KSTAT_DATA_UINT64 },
562 	{ "l2_evict_lock_retry",	KSTAT_DATA_UINT64 },
563 	{ "l2_evict_reading",		KSTAT_DATA_UINT64 },
564 	{ "l2_evict_l1cached",		KSTAT_DATA_UINT64 },
565 	{ "l2_free_on_write",		KSTAT_DATA_UINT64 },
566 	{ "l2_cdata_free_on_write",	KSTAT_DATA_UINT64 },
567 	{ "l2_abort_lowmem",		KSTAT_DATA_UINT64 },
568 	{ "l2_cksum_bad",		KSTAT_DATA_UINT64 },
569 	{ "l2_io_error",		KSTAT_DATA_UINT64 },
570 	{ "l2_size",			KSTAT_DATA_UINT64 },
571 	{ "l2_asize",			KSTAT_DATA_UINT64 },
572 	{ "l2_hdr_size",		KSTAT_DATA_UINT64 },
573 	{ "l2_compress_successes",	KSTAT_DATA_UINT64 },
574 	{ "l2_compress_zeros",		KSTAT_DATA_UINT64 },
575 	{ "l2_compress_failures",	KSTAT_DATA_UINT64 },
576 	{ "memory_throttle_count",	KSTAT_DATA_UINT64 },
577 	{ "duplicate_buffers",		KSTAT_DATA_UINT64 },
578 	{ "duplicate_buffers_size",	KSTAT_DATA_UINT64 },
579 	{ "duplicate_reads",		KSTAT_DATA_UINT64 },
580 	{ "arc_meta_used",		KSTAT_DATA_UINT64 },
581 	{ "arc_meta_limit",		KSTAT_DATA_UINT64 },
582 	{ "arc_meta_max",		KSTAT_DATA_UINT64 },
583 	{ "arc_meta_min",		KSTAT_DATA_UINT64 }
584 };
585 
586 #define	ARCSTAT(stat)	(arc_stats.stat.value.ui64)
587 
588 #define	ARCSTAT_INCR(stat, val) \
589 	atomic_add_64(&arc_stats.stat.value.ui64, (val))
590 
591 #define	ARCSTAT_BUMP(stat)	ARCSTAT_INCR(stat, 1)
592 #define	ARCSTAT_BUMPDOWN(stat)	ARCSTAT_INCR(stat, -1)
593 
594 #define	ARCSTAT_MAX(stat, val) {					\
595 	uint64_t m;							\
596 	while ((val) > (m = arc_stats.stat.value.ui64) &&		\
597 	    (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val))))	\
598 		continue;						\
599 }
600 
601 #define	ARCSTAT_MAXSTAT(stat) \
602 	ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
603 
604 /*
605  * We define a macro to allow ARC hits/misses to be easily broken down by
606  * two separate conditions, giving a total of four different subtypes for
607  * each of hits and misses (so eight statistics total).
608  */
609 #define	ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
610 	if (cond1) {							\
611 		if (cond2) {						\
612 			ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
613 		} else {						\
614 			ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
615 		}							\
616 	} else {							\
617 		if (cond2) {						\
618 			ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
619 		} else {						\
620 			ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
621 		}							\
622 	}
623 
624 kstat_t			*arc_ksp;
625 static arc_state_t	*arc_anon;
626 static arc_state_t	*arc_mru;
627 static arc_state_t	*arc_mru_ghost;
628 static arc_state_t	*arc_mfu;
629 static arc_state_t	*arc_mfu_ghost;
630 static arc_state_t	*arc_l2c_only;
631 
632 /*
633  * There are several ARC variables that are critical to export as kstats --
634  * but we don't want to have to grovel around in the kstat whenever we wish to
635  * manipulate them.  For these variables, we therefore define them to be in
636  * terms of the statistic variable.  This assures that we are not introducing
637  * the possibility of inconsistency by having shadow copies of the variables,
638  * while still allowing the code to be readable.
639  */
640 #define	arc_size	ARCSTAT(arcstat_size)	/* actual total arc size */
641 #define	arc_p		ARCSTAT(arcstat_p)	/* target size of MRU */
642 #define	arc_c		ARCSTAT(arcstat_c)	/* target size of cache */
643 #define	arc_c_min	ARCSTAT(arcstat_c_min)	/* min target cache size */
644 #define	arc_c_max	ARCSTAT(arcstat_c_max)	/* max target cache size */
645 #define	arc_meta_limit	ARCSTAT(arcstat_meta_limit) /* max size for metadata */
646 #define	arc_meta_min	ARCSTAT(arcstat_meta_min) /* min size for metadata */
647 #define	arc_meta_used	ARCSTAT(arcstat_meta_used) /* size of metadata */
648 #define	arc_meta_max	ARCSTAT(arcstat_meta_max) /* max size of metadata */
649 
650 #define	L2ARC_IS_VALID_COMPRESS(_c_) \
651 	((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
652 
653 static int		arc_no_grow;	/* Don't try to grow cache size */
654 static uint64_t		arc_tempreserve;
655 static uint64_t		arc_loaned_bytes;
656 
657 typedef struct arc_callback arc_callback_t;
658 
659 struct arc_callback {
660 	void			*acb_private;
661 	arc_done_func_t		*acb_done;
662 	arc_buf_t		*acb_buf;
663 	zio_t			*acb_zio_dummy;
664 	arc_callback_t		*acb_next;
665 };
666 
667 typedef struct arc_write_callback arc_write_callback_t;
668 
669 struct arc_write_callback {
670 	void		*awcb_private;
671 	arc_done_func_t	*awcb_ready;
672 	arc_done_func_t	*awcb_physdone;
673 	arc_done_func_t	*awcb_done;
674 	arc_buf_t	*awcb_buf;
675 };
676 
677 /*
678  * ARC buffers are separated into multiple structs as a memory saving measure:
679  *   - Common fields struct, always defined, and embedded within it:
680  *       - L2-only fields, always allocated but undefined when not in L2ARC
681  *       - L1-only fields, only allocated when in L1ARC
682  *
683  *           Buffer in L1                     Buffer only in L2
684  *    +------------------------+          +------------------------+
685  *    | arc_buf_hdr_t          |          | arc_buf_hdr_t          |
686  *    |                        |          |                        |
687  *    |                        |          |                        |
688  *    |                        |          |                        |
689  *    +------------------------+          +------------------------+
690  *    | l2arc_buf_hdr_t        |          | l2arc_buf_hdr_t        |
691  *    | (undefined if L1-only) |          |                        |
692  *    +------------------------+          +------------------------+
693  *    | l1arc_buf_hdr_t        |
694  *    |                        |
695  *    |                        |
696  *    |                        |
697  *    |                        |
698  *    +------------------------+
699  *
700  * Because it's possible for the L2ARC to become extremely large, we can wind
701  * up eating a lot of memory in L2ARC buffer headers, so the size of a header
702  * is minimized by only allocating the fields necessary for an L1-cached buffer
703  * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
704  * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
705  * words in pointers. arc_hdr_realloc() is used to switch a header between
706  * these two allocation states.
707  */
708 typedef struct l1arc_buf_hdr {
709 	kmutex_t		b_freeze_lock;
710 #ifdef ZFS_DEBUG
711 	/*
712 	 * used for debugging wtih kmem_flags - by allocating and freeing
713 	 * b_thawed when the buffer is thawed, we get a record of the stack
714 	 * trace that thawed it.
715 	 */
716 	void			*b_thawed;
717 #endif
718 
719 	arc_buf_t		*b_buf;
720 	uint32_t		b_datacnt;
721 	/* for waiting on writes to complete */
722 	kcondvar_t		b_cv;
723 
724 	/* protected by arc state mutex */
725 	arc_state_t		*b_state;
726 	multilist_node_t	b_arc_node;
727 
728 	/* updated atomically */
729 	clock_t			b_arc_access;
730 
731 	/* self protecting */
732 	refcount_t		b_refcnt;
733 
734 	arc_callback_t		*b_acb;
735 	/* temporary buffer holder for in-flight compressed data */
736 	void			*b_tmp_cdata;
737 } l1arc_buf_hdr_t;
738 
739 typedef struct l2arc_dev l2arc_dev_t;
740 
741 typedef struct l2arc_buf_hdr {
742 	/* protected by arc_buf_hdr mutex */
743 	l2arc_dev_t		*b_dev;		/* L2ARC device */
744 	uint64_t		b_daddr;	/* disk address, offset byte */
745 	/* real alloc'd buffer size depending on b_compress applied */
746 	int32_t			b_asize;
747 
748 	list_node_t		b_l2node;
749 } l2arc_buf_hdr_t;
750 
751 struct arc_buf_hdr {
752 	/* protected by hash lock */
753 	dva_t			b_dva;
754 	uint64_t		b_birth;
755 	/*
756 	 * Even though this checksum is only set/verified when a buffer is in
757 	 * the L1 cache, it needs to be in the set of common fields because it
758 	 * must be preserved from the time before a buffer is written out to
759 	 * L2ARC until after it is read back in.
760 	 */
761 	zio_cksum_t		*b_freeze_cksum;
762 
763 	arc_buf_hdr_t		*b_hash_next;
764 	arc_flags_t		b_flags;
765 
766 	/* immutable */
767 	int32_t			b_size;
768 	uint64_t		b_spa;
769 
770 	/* L2ARC fields. Undefined when not in L2ARC. */
771 	l2arc_buf_hdr_t		b_l2hdr;
772 	/* L1ARC fields. Undefined when in l2arc_only state */
773 	l1arc_buf_hdr_t		b_l1hdr;
774 };
775 
776 static arc_buf_t *arc_eviction_list;
777 static arc_buf_hdr_t arc_eviction_hdr;
778 
779 #define	GHOST_STATE(state)	\
780 	((state) == arc_mru_ghost || (state) == arc_mfu_ghost ||	\
781 	(state) == arc_l2c_only)
782 
783 #define	HDR_IN_HASH_TABLE(hdr)	((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
784 #define	HDR_IO_IN_PROGRESS(hdr)	((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
785 #define	HDR_IO_ERROR(hdr)	((hdr)->b_flags & ARC_FLAG_IO_ERROR)
786 #define	HDR_PREFETCH(hdr)	((hdr)->b_flags & ARC_FLAG_PREFETCH)
787 #define	HDR_FREED_IN_READ(hdr)	((hdr)->b_flags & ARC_FLAG_FREED_IN_READ)
788 #define	HDR_BUF_AVAILABLE(hdr)	((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE)
789 
790 #define	HDR_L2CACHE(hdr)	((hdr)->b_flags & ARC_FLAG_L2CACHE)
791 #define	HDR_L2COMPRESS(hdr)	((hdr)->b_flags & ARC_FLAG_L2COMPRESS)
792 #define	HDR_L2_READING(hdr)	\
793 	    (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) &&	\
794 	    ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
795 #define	HDR_L2_WRITING(hdr)	((hdr)->b_flags & ARC_FLAG_L2_WRITING)
796 #define	HDR_L2_EVICTED(hdr)	((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
797 #define	HDR_L2_WRITE_HEAD(hdr)	((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
798 
799 #define	HDR_ISTYPE_METADATA(hdr)	\
800 	    ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
801 #define	HDR_ISTYPE_DATA(hdr)	(!HDR_ISTYPE_METADATA(hdr))
802 
803 #define	HDR_HAS_L1HDR(hdr)	((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
804 #define	HDR_HAS_L2HDR(hdr)	((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
805 
806 /* For storing compression mode in b_flags */
807 #define	HDR_COMPRESS_OFFSET	24
808 #define	HDR_COMPRESS_NBITS	7
809 
810 #define	HDR_GET_COMPRESS(hdr)	((enum zio_compress)BF32_GET(hdr->b_flags, \
811 	    HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS))
812 #define	HDR_SET_COMPRESS(hdr, cmp) BF32_SET(hdr->b_flags, \
813 	    HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS, (cmp))
814 
815 /*
816  * Other sizes
817  */
818 
819 #define	HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
820 #define	HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
821 
822 /*
823  * Hash table routines
824  */
825 
826 #define	HT_LOCK_PAD	64
827 
828 struct ht_lock {
829 	kmutex_t	ht_lock;
830 #ifdef _KERNEL
831 	unsigned char	pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
832 #endif
833 };
834 
835 #define	BUF_LOCKS 256
836 typedef struct buf_hash_table {
837 	uint64_t ht_mask;
838 	arc_buf_hdr_t **ht_table;
839 	struct ht_lock ht_locks[BUF_LOCKS];
840 } buf_hash_table_t;
841 
842 static buf_hash_table_t buf_hash_table;
843 
844 #define	BUF_HASH_INDEX(spa, dva, birth) \
845 	(buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
846 #define	BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
847 #define	BUF_HASH_LOCK(idx)	(&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
848 #define	HDR_LOCK(hdr) \
849 	(BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
850 
851 uint64_t zfs_crc64_table[256];
852 
853 /*
854  * Level 2 ARC
855  */
856 
857 #define	L2ARC_WRITE_SIZE	(8 * 1024 * 1024)	/* initial write max */
858 #define	L2ARC_HEADROOM		2			/* num of writes */
859 /*
860  * If we discover during ARC scan any buffers to be compressed, we boost
861  * our headroom for the next scanning cycle by this percentage multiple.
862  */
863 #define	L2ARC_HEADROOM_BOOST	200
864 #define	L2ARC_FEED_SECS		1		/* caching interval secs */
865 #define	L2ARC_FEED_MIN_MS	200		/* min caching interval ms */
866 
867 /*
868  * Used to distinguish headers that are being process by
869  * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk
870  * address. This can happen when the header is added to the l2arc's list
871  * of buffers to write in the first stage of l2arc_write_buffers(), but
872  * has not yet been written out which happens in the second stage of
873  * l2arc_write_buffers().
874  */
875 #define	L2ARC_ADDR_UNSET	((uint64_t)(-1))
876 
877 #define	l2arc_writes_sent	ARCSTAT(arcstat_l2_writes_sent)
878 #define	l2arc_writes_done	ARCSTAT(arcstat_l2_writes_done)
879 
880 /* L2ARC Performance Tunables */
881 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE;	/* default max write size */
882 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE;	/* extra write during warmup */
883 uint64_t l2arc_headroom = L2ARC_HEADROOM;	/* number of dev writes */
884 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
885 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS;	/* interval seconds */
886 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS;	/* min interval milliseconds */
887 boolean_t l2arc_noprefetch = B_TRUE;		/* don't cache prefetch bufs */
888 boolean_t l2arc_feed_again = B_TRUE;		/* turbo warmup */
889 boolean_t l2arc_norw = B_TRUE;			/* no reads during writes */
890 
891 /*
892  * L2ARC Internals
893  */
894 struct l2arc_dev {
895 	vdev_t			*l2ad_vdev;	/* vdev */
896 	spa_t			*l2ad_spa;	/* spa */
897 	uint64_t		l2ad_hand;	/* next write location */
898 	uint64_t		l2ad_start;	/* first addr on device */
899 	uint64_t		l2ad_end;	/* last addr on device */
900 	boolean_t		l2ad_first;	/* first sweep through */
901 	boolean_t		l2ad_writing;	/* currently writing */
902 	kmutex_t		l2ad_mtx;	/* lock for buffer list */
903 	list_t			l2ad_buflist;	/* buffer list */
904 	list_node_t		l2ad_node;	/* device list node */
905 	refcount_t		l2ad_alloc;	/* allocated bytes */
906 };
907 
908 static list_t L2ARC_dev_list;			/* device list */
909 static list_t *l2arc_dev_list;			/* device list pointer */
910 static kmutex_t l2arc_dev_mtx;			/* device list mutex */
911 static l2arc_dev_t *l2arc_dev_last;		/* last device used */
912 static list_t L2ARC_free_on_write;		/* free after write buf list */
913 static list_t *l2arc_free_on_write;		/* free after write list ptr */
914 static kmutex_t l2arc_free_on_write_mtx;	/* mutex for list */
915 static uint64_t l2arc_ndev;			/* number of devices */
916 
917 typedef struct l2arc_read_callback {
918 	arc_buf_t		*l2rcb_buf;		/* read buffer */
919 	spa_t			*l2rcb_spa;		/* spa */
920 	blkptr_t		l2rcb_bp;		/* original blkptr */
921 	zbookmark_phys_t	l2rcb_zb;		/* original bookmark */
922 	int			l2rcb_flags;		/* original flags */
923 	enum zio_compress	l2rcb_compress;		/* applied compress */
924 } l2arc_read_callback_t;
925 
926 typedef struct l2arc_write_callback {
927 	l2arc_dev_t	*l2wcb_dev;		/* device info */
928 	arc_buf_hdr_t	*l2wcb_head;		/* head of write buflist */
929 } l2arc_write_callback_t;
930 
931 typedef struct l2arc_data_free {
932 	/* protected by l2arc_free_on_write_mtx */
933 	void		*l2df_data;
934 	size_t		l2df_size;
935 	void		(*l2df_func)(void *, size_t);
936 	list_node_t	l2df_list_node;
937 } l2arc_data_free_t;
938 
939 static kmutex_t l2arc_feed_thr_lock;
940 static kcondvar_t l2arc_feed_thr_cv;
941 static uint8_t l2arc_thread_exit;
942 
943 static void arc_get_data_buf(arc_buf_t *);
944 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
945 static boolean_t arc_is_overflowing();
946 static void arc_buf_watch(arc_buf_t *);
947 
948 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
949 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
950 
951 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
952 static void l2arc_read_done(zio_t *);
953 
954 static boolean_t l2arc_compress_buf(arc_buf_hdr_t *);
955 static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress);
956 static void l2arc_release_cdata_buf(arc_buf_hdr_t *);
957 
958 static uint64_t
959 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
960 {
961 	uint8_t *vdva = (uint8_t *)dva;
962 	uint64_t crc = -1ULL;
963 	int i;
964 
965 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
966 
967 	for (i = 0; i < sizeof (dva_t); i++)
968 		crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
969 
970 	crc ^= (spa>>8) ^ birth;
971 
972 	return (crc);
973 }
974 
975 #define	BUF_EMPTY(buf)						\
976 	((buf)->b_dva.dva_word[0] == 0 &&			\
977 	(buf)->b_dva.dva_word[1] == 0)
978 
979 #define	BUF_EQUAL(spa, dva, birth, buf)				\
980 	((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) &&	\
981 	((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) &&	\
982 	((buf)->b_birth == birth) && ((buf)->b_spa == spa)
983 
984 static void
985 buf_discard_identity(arc_buf_hdr_t *hdr)
986 {
987 	hdr->b_dva.dva_word[0] = 0;
988 	hdr->b_dva.dva_word[1] = 0;
989 	hdr->b_birth = 0;
990 }
991 
992 static arc_buf_hdr_t *
993 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
994 {
995 	const dva_t *dva = BP_IDENTITY(bp);
996 	uint64_t birth = BP_PHYSICAL_BIRTH(bp);
997 	uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
998 	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
999 	arc_buf_hdr_t *hdr;
1000 
1001 	mutex_enter(hash_lock);
1002 	for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1003 	    hdr = hdr->b_hash_next) {
1004 		if (BUF_EQUAL(spa, dva, birth, hdr)) {
1005 			*lockp = hash_lock;
1006 			return (hdr);
1007 		}
1008 	}
1009 	mutex_exit(hash_lock);
1010 	*lockp = NULL;
1011 	return (NULL);
1012 }
1013 
1014 /*
1015  * Insert an entry into the hash table.  If there is already an element
1016  * equal to elem in the hash table, then the already existing element
1017  * will be returned and the new element will not be inserted.
1018  * Otherwise returns NULL.
1019  * If lockp == NULL, the caller is assumed to already hold the hash lock.
1020  */
1021 static arc_buf_hdr_t *
1022 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1023 {
1024 	uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1025 	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1026 	arc_buf_hdr_t *fhdr;
1027 	uint32_t i;
1028 
1029 	ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1030 	ASSERT(hdr->b_birth != 0);
1031 	ASSERT(!HDR_IN_HASH_TABLE(hdr));
1032 
1033 	if (lockp != NULL) {
1034 		*lockp = hash_lock;
1035 		mutex_enter(hash_lock);
1036 	} else {
1037 		ASSERT(MUTEX_HELD(hash_lock));
1038 	}
1039 
1040 	for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1041 	    fhdr = fhdr->b_hash_next, i++) {
1042 		if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1043 			return (fhdr);
1044 	}
1045 
1046 	hdr->b_hash_next = buf_hash_table.ht_table[idx];
1047 	buf_hash_table.ht_table[idx] = hdr;
1048 	hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
1049 
1050 	/* collect some hash table performance data */
1051 	if (i > 0) {
1052 		ARCSTAT_BUMP(arcstat_hash_collisions);
1053 		if (i == 1)
1054 			ARCSTAT_BUMP(arcstat_hash_chains);
1055 
1056 		ARCSTAT_MAX(arcstat_hash_chain_max, i);
1057 	}
1058 
1059 	ARCSTAT_BUMP(arcstat_hash_elements);
1060 	ARCSTAT_MAXSTAT(arcstat_hash_elements);
1061 
1062 	return (NULL);
1063 }
1064 
1065 static void
1066 buf_hash_remove(arc_buf_hdr_t *hdr)
1067 {
1068 	arc_buf_hdr_t *fhdr, **hdrp;
1069 	uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1070 
1071 	ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1072 	ASSERT(HDR_IN_HASH_TABLE(hdr));
1073 
1074 	hdrp = &buf_hash_table.ht_table[idx];
1075 	while ((fhdr = *hdrp) != hdr) {
1076 		ASSERT(fhdr != NULL);
1077 		hdrp = &fhdr->b_hash_next;
1078 	}
1079 	*hdrp = hdr->b_hash_next;
1080 	hdr->b_hash_next = NULL;
1081 	hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE;
1082 
1083 	/* collect some hash table performance data */
1084 	ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1085 
1086 	if (buf_hash_table.ht_table[idx] &&
1087 	    buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1088 		ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1089 }
1090 
1091 /*
1092  * Global data structures and functions for the buf kmem cache.
1093  */
1094 static kmem_cache_t *hdr_full_cache;
1095 static kmem_cache_t *hdr_l2only_cache;
1096 static kmem_cache_t *buf_cache;
1097 
1098 static void
1099 buf_fini(void)
1100 {
1101 	int i;
1102 
1103 	kmem_free(buf_hash_table.ht_table,
1104 	    (buf_hash_table.ht_mask + 1) * sizeof (void *));
1105 	for (i = 0; i < BUF_LOCKS; i++)
1106 		mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1107 	kmem_cache_destroy(hdr_full_cache);
1108 	kmem_cache_destroy(hdr_l2only_cache);
1109 	kmem_cache_destroy(buf_cache);
1110 }
1111 
1112 /*
1113  * Constructor callback - called when the cache is empty
1114  * and a new buf is requested.
1115  */
1116 /* ARGSUSED */
1117 static int
1118 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1119 {
1120 	arc_buf_hdr_t *hdr = vbuf;
1121 
1122 	bzero(hdr, HDR_FULL_SIZE);
1123 	cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1124 	refcount_create(&hdr->b_l1hdr.b_refcnt);
1125 	mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1126 	multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1127 	arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1128 
1129 	return (0);
1130 }
1131 
1132 /* ARGSUSED */
1133 static int
1134 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1135 {
1136 	arc_buf_hdr_t *hdr = vbuf;
1137 
1138 	bzero(hdr, HDR_L2ONLY_SIZE);
1139 	arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1140 
1141 	return (0);
1142 }
1143 
1144 /* ARGSUSED */
1145 static int
1146 buf_cons(void *vbuf, void *unused, int kmflag)
1147 {
1148 	arc_buf_t *buf = vbuf;
1149 
1150 	bzero(buf, sizeof (arc_buf_t));
1151 	mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1152 	arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1153 
1154 	return (0);
1155 }
1156 
1157 /*
1158  * Destructor callback - called when a cached buf is
1159  * no longer required.
1160  */
1161 /* ARGSUSED */
1162 static void
1163 hdr_full_dest(void *vbuf, void *unused)
1164 {
1165 	arc_buf_hdr_t *hdr = vbuf;
1166 
1167 	ASSERT(BUF_EMPTY(hdr));
1168 	cv_destroy(&hdr->b_l1hdr.b_cv);
1169 	refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1170 	mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1171 	ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1172 	arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1173 }
1174 
1175 /* ARGSUSED */
1176 static void
1177 hdr_l2only_dest(void *vbuf, void *unused)
1178 {
1179 	arc_buf_hdr_t *hdr = vbuf;
1180 
1181 	ASSERT(BUF_EMPTY(hdr));
1182 	arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1183 }
1184 
1185 /* ARGSUSED */
1186 static void
1187 buf_dest(void *vbuf, void *unused)
1188 {
1189 	arc_buf_t *buf = vbuf;
1190 
1191 	mutex_destroy(&buf->b_evict_lock);
1192 	arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1193 }
1194 
1195 /*
1196  * Reclaim callback -- invoked when memory is low.
1197  */
1198 /* ARGSUSED */
1199 static void
1200 hdr_recl(void *unused)
1201 {
1202 	dprintf("hdr_recl called\n");
1203 	/*
1204 	 * umem calls the reclaim func when we destroy the buf cache,
1205 	 * which is after we do arc_fini().
1206 	 */
1207 	if (!arc_dead)
1208 		cv_signal(&arc_reclaim_thread_cv);
1209 }
1210 
1211 static void
1212 buf_init(void)
1213 {
1214 	uint64_t *ct;
1215 	uint64_t hsize = 1ULL << 12;
1216 	int i, j;
1217 
1218 	/*
1219 	 * The hash table is big enough to fill all of physical memory
1220 	 * with an average block size of zfs_arc_average_blocksize (default 8K).
1221 	 * By default, the table will take up
1222 	 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1223 	 */
1224 	while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
1225 		hsize <<= 1;
1226 retry:
1227 	buf_hash_table.ht_mask = hsize - 1;
1228 	buf_hash_table.ht_table =
1229 	    kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1230 	if (buf_hash_table.ht_table == NULL) {
1231 		ASSERT(hsize > (1ULL << 8));
1232 		hsize >>= 1;
1233 		goto retry;
1234 	}
1235 
1236 	hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1237 	    0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1238 	hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1239 	    HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1240 	    NULL, NULL, 0);
1241 	buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1242 	    0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1243 
1244 	for (i = 0; i < 256; i++)
1245 		for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1246 			*ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1247 
1248 	for (i = 0; i < BUF_LOCKS; i++) {
1249 		mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1250 		    NULL, MUTEX_DEFAULT, NULL);
1251 	}
1252 }
1253 
1254 /*
1255  * Transition between the two allocation states for the arc_buf_hdr struct.
1256  * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
1257  * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
1258  * version is used when a cache buffer is only in the L2ARC in order to reduce
1259  * memory usage.
1260  */
1261 static arc_buf_hdr_t *
1262 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
1263 {
1264 	ASSERT(HDR_HAS_L2HDR(hdr));
1265 
1266 	arc_buf_hdr_t *nhdr;
1267 	l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1268 
1269 	ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
1270 	    (old == hdr_l2only_cache && new == hdr_full_cache));
1271 
1272 	nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
1273 
1274 	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
1275 	buf_hash_remove(hdr);
1276 
1277 	bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
1278 
1279 	if (new == hdr_full_cache) {
1280 		nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1281 		/*
1282 		 * arc_access and arc_change_state need to be aware that a
1283 		 * header has just come out of L2ARC, so we set its state to
1284 		 * l2c_only even though it's about to change.
1285 		 */
1286 		nhdr->b_l1hdr.b_state = arc_l2c_only;
1287 
1288 		/* Verify previous threads set to NULL before freeing */
1289 		ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1290 	} else {
1291 		ASSERT(hdr->b_l1hdr.b_buf == NULL);
1292 		ASSERT0(hdr->b_l1hdr.b_datacnt);
1293 
1294 		/*
1295 		 * If we've reached here, We must have been called from
1296 		 * arc_evict_hdr(), as such we should have already been
1297 		 * removed from any ghost list we were previously on
1298 		 * (which protects us from racing with arc_evict_state),
1299 		 * thus no locking is needed during this check.
1300 		 */
1301 		ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1302 
1303 		/*
1304 		 * A buffer must not be moved into the arc_l2c_only
1305 		 * state if it's not finished being written out to the
1306 		 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field
1307 		 * might try to be accessed, even though it was removed.
1308 		 */
1309 		VERIFY(!HDR_L2_WRITING(hdr));
1310 		VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1311 
1312 		nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR;
1313 	}
1314 	/*
1315 	 * The header has been reallocated so we need to re-insert it into any
1316 	 * lists it was on.
1317 	 */
1318 	(void) buf_hash_insert(nhdr, NULL);
1319 
1320 	ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
1321 
1322 	mutex_enter(&dev->l2ad_mtx);
1323 
1324 	/*
1325 	 * We must place the realloc'ed header back into the list at
1326 	 * the same spot. Otherwise, if it's placed earlier in the list,
1327 	 * l2arc_write_buffers() could find it during the function's
1328 	 * write phase, and try to write it out to the l2arc.
1329 	 */
1330 	list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
1331 	list_remove(&dev->l2ad_buflist, hdr);
1332 
1333 	mutex_exit(&dev->l2ad_mtx);
1334 
1335 	/*
1336 	 * Since we're using the pointer address as the tag when
1337 	 * incrementing and decrementing the l2ad_alloc refcount, we
1338 	 * must remove the old pointer (that we're about to destroy) and
1339 	 * add the new pointer to the refcount. Otherwise we'd remove
1340 	 * the wrong pointer address when calling arc_hdr_destroy() later.
1341 	 */
1342 
1343 	(void) refcount_remove_many(&dev->l2ad_alloc,
1344 	    hdr->b_l2hdr.b_asize, hdr);
1345 
1346 	(void) refcount_add_many(&dev->l2ad_alloc,
1347 	    nhdr->b_l2hdr.b_asize, nhdr);
1348 
1349 	buf_discard_identity(hdr);
1350 	hdr->b_freeze_cksum = NULL;
1351 	kmem_cache_free(old, hdr);
1352 
1353 	return (nhdr);
1354 }
1355 
1356 
1357 #define	ARC_MINTIME	(hz>>4) /* 62 ms */
1358 
1359 static void
1360 arc_cksum_verify(arc_buf_t *buf)
1361 {
1362 	zio_cksum_t zc;
1363 
1364 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1365 		return;
1366 
1367 	mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1368 	if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) {
1369 		mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1370 		return;
1371 	}
1372 	fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1373 	if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1374 		panic("buffer modified while frozen!");
1375 	mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1376 }
1377 
1378 static int
1379 arc_cksum_equal(arc_buf_t *buf)
1380 {
1381 	zio_cksum_t zc;
1382 	int equal;
1383 
1384 	mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1385 	fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1386 	equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1387 	mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1388 
1389 	return (equal);
1390 }
1391 
1392 static void
1393 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1394 {
1395 	if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1396 		return;
1397 
1398 	mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1399 	if (buf->b_hdr->b_freeze_cksum != NULL) {
1400 		mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1401 		return;
1402 	}
1403 	buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1404 	fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1405 	    buf->b_hdr->b_freeze_cksum);
1406 	mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1407 	arc_buf_watch(buf);
1408 }
1409 
1410 #ifndef _KERNEL
1411 typedef struct procctl {
1412 	long cmd;
1413 	prwatch_t prwatch;
1414 } procctl_t;
1415 #endif
1416 
1417 /* ARGSUSED */
1418 static void
1419 arc_buf_unwatch(arc_buf_t *buf)
1420 {
1421 #ifndef _KERNEL
1422 	if (arc_watch) {
1423 		int result;
1424 		procctl_t ctl;
1425 		ctl.cmd = PCWATCH;
1426 		ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1427 		ctl.prwatch.pr_size = 0;
1428 		ctl.prwatch.pr_wflags = 0;
1429 		result = write(arc_procfd, &ctl, sizeof (ctl));
1430 		ASSERT3U(result, ==, sizeof (ctl));
1431 	}
1432 #endif
1433 }
1434 
1435 /* ARGSUSED */
1436 static void
1437 arc_buf_watch(arc_buf_t *buf)
1438 {
1439 #ifndef _KERNEL
1440 	if (arc_watch) {
1441 		int result;
1442 		procctl_t ctl;
1443 		ctl.cmd = PCWATCH;
1444 		ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1445 		ctl.prwatch.pr_size = buf->b_hdr->b_size;
1446 		ctl.prwatch.pr_wflags = WA_WRITE;
1447 		result = write(arc_procfd, &ctl, sizeof (ctl));
1448 		ASSERT3U(result, ==, sizeof (ctl));
1449 	}
1450 #endif
1451 }
1452 
1453 static arc_buf_contents_t
1454 arc_buf_type(arc_buf_hdr_t *hdr)
1455 {
1456 	if (HDR_ISTYPE_METADATA(hdr)) {
1457 		return (ARC_BUFC_METADATA);
1458 	} else {
1459 		return (ARC_BUFC_DATA);
1460 	}
1461 }
1462 
1463 static uint32_t
1464 arc_bufc_to_flags(arc_buf_contents_t type)
1465 {
1466 	switch (type) {
1467 	case ARC_BUFC_DATA:
1468 		/* metadata field is 0 if buffer contains normal data */
1469 		return (0);
1470 	case ARC_BUFC_METADATA:
1471 		return (ARC_FLAG_BUFC_METADATA);
1472 	default:
1473 		break;
1474 	}
1475 	panic("undefined ARC buffer type!");
1476 	return ((uint32_t)-1);
1477 }
1478 
1479 void
1480 arc_buf_thaw(arc_buf_t *buf)
1481 {
1482 	if (zfs_flags & ZFS_DEBUG_MODIFY) {
1483 		if (buf->b_hdr->b_l1hdr.b_state != arc_anon)
1484 			panic("modifying non-anon buffer!");
1485 		if (HDR_IO_IN_PROGRESS(buf->b_hdr))
1486 			panic("modifying buffer while i/o in progress!");
1487 		arc_cksum_verify(buf);
1488 	}
1489 
1490 	mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1491 	if (buf->b_hdr->b_freeze_cksum != NULL) {
1492 		kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1493 		buf->b_hdr->b_freeze_cksum = NULL;
1494 	}
1495 
1496 #ifdef ZFS_DEBUG
1497 	if (zfs_flags & ZFS_DEBUG_MODIFY) {
1498 		if (buf->b_hdr->b_l1hdr.b_thawed != NULL)
1499 			kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1);
1500 		buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1501 	}
1502 #endif
1503 
1504 	mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1505 
1506 	arc_buf_unwatch(buf);
1507 }
1508 
1509 void
1510 arc_buf_freeze(arc_buf_t *buf)
1511 {
1512 	kmutex_t *hash_lock;
1513 
1514 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1515 		return;
1516 
1517 	hash_lock = HDR_LOCK(buf->b_hdr);
1518 	mutex_enter(hash_lock);
1519 
1520 	ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1521 	    buf->b_hdr->b_l1hdr.b_state == arc_anon);
1522 	arc_cksum_compute(buf, B_FALSE);
1523 	mutex_exit(hash_lock);
1524 
1525 }
1526 
1527 static void
1528 add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1529 {
1530 	ASSERT(HDR_HAS_L1HDR(hdr));
1531 	ASSERT(MUTEX_HELD(hash_lock));
1532 	arc_state_t *state = hdr->b_l1hdr.b_state;
1533 
1534 	if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
1535 	    (state != arc_anon)) {
1536 		/* We don't use the L2-only state list. */
1537 		if (state != arc_l2c_only) {
1538 			arc_buf_contents_t type = arc_buf_type(hdr);
1539 			uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt;
1540 			multilist_t *list = &state->arcs_list[type];
1541 			uint64_t *size = &state->arcs_lsize[type];
1542 
1543 			multilist_remove(list, hdr);
1544 
1545 			if (GHOST_STATE(state)) {
1546 				ASSERT0(hdr->b_l1hdr.b_datacnt);
1547 				ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
1548 				delta = hdr->b_size;
1549 			}
1550 			ASSERT(delta > 0);
1551 			ASSERT3U(*size, >=, delta);
1552 			atomic_add_64(size, -delta);
1553 		}
1554 		/* remove the prefetch flag if we get a reference */
1555 		hdr->b_flags &= ~ARC_FLAG_PREFETCH;
1556 	}
1557 }
1558 
1559 static int
1560 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1561 {
1562 	int cnt;
1563 	arc_state_t *state = hdr->b_l1hdr.b_state;
1564 
1565 	ASSERT(HDR_HAS_L1HDR(hdr));
1566 	ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1567 	ASSERT(!GHOST_STATE(state));
1568 
1569 	/*
1570 	 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1571 	 * check to prevent usage of the arc_l2c_only list.
1572 	 */
1573 	if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
1574 	    (state != arc_anon)) {
1575 		arc_buf_contents_t type = arc_buf_type(hdr);
1576 		multilist_t *list = &state->arcs_list[type];
1577 		uint64_t *size = &state->arcs_lsize[type];
1578 
1579 		multilist_insert(list, hdr);
1580 
1581 		ASSERT(hdr->b_l1hdr.b_datacnt > 0);
1582 		atomic_add_64(size, hdr->b_size *
1583 		    hdr->b_l1hdr.b_datacnt);
1584 	}
1585 	return (cnt);
1586 }
1587 
1588 /*
1589  * Move the supplied buffer to the indicated state. The hash lock
1590  * for the buffer must be held by the caller.
1591  */
1592 static void
1593 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
1594     kmutex_t *hash_lock)
1595 {
1596 	arc_state_t *old_state;
1597 	int64_t refcnt;
1598 	uint32_t datacnt;
1599 	uint64_t from_delta, to_delta;
1600 	arc_buf_contents_t buftype = arc_buf_type(hdr);
1601 
1602 	/*
1603 	 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
1604 	 * in arc_read() when bringing a buffer out of the L2ARC.  However, the
1605 	 * L1 hdr doesn't always exist when we change state to arc_anon before
1606 	 * destroying a header, in which case reallocating to add the L1 hdr is
1607 	 * pointless.
1608 	 */
1609 	if (HDR_HAS_L1HDR(hdr)) {
1610 		old_state = hdr->b_l1hdr.b_state;
1611 		refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
1612 		datacnt = hdr->b_l1hdr.b_datacnt;
1613 	} else {
1614 		old_state = arc_l2c_only;
1615 		refcnt = 0;
1616 		datacnt = 0;
1617 	}
1618 
1619 	ASSERT(MUTEX_HELD(hash_lock));
1620 	ASSERT3P(new_state, !=, old_state);
1621 	ASSERT(refcnt == 0 || datacnt > 0);
1622 	ASSERT(!GHOST_STATE(new_state) || datacnt == 0);
1623 	ASSERT(old_state != arc_anon || datacnt <= 1);
1624 
1625 	from_delta = to_delta = datacnt * hdr->b_size;
1626 
1627 	/*
1628 	 * If this buffer is evictable, transfer it from the
1629 	 * old state list to the new state list.
1630 	 */
1631 	if (refcnt == 0) {
1632 		if (old_state != arc_anon && old_state != arc_l2c_only) {
1633 			uint64_t *size = &old_state->arcs_lsize[buftype];
1634 
1635 			ASSERT(HDR_HAS_L1HDR(hdr));
1636 			multilist_remove(&old_state->arcs_list[buftype], hdr);
1637 
1638 			/*
1639 			 * If prefetching out of the ghost cache,
1640 			 * we will have a non-zero datacnt.
1641 			 */
1642 			if (GHOST_STATE(old_state) && datacnt == 0) {
1643 				/* ghost elements have a ghost size */
1644 				ASSERT(hdr->b_l1hdr.b_buf == NULL);
1645 				from_delta = hdr->b_size;
1646 			}
1647 			ASSERT3U(*size, >=, from_delta);
1648 			atomic_add_64(size, -from_delta);
1649 		}
1650 		if (new_state != arc_anon && new_state != arc_l2c_only) {
1651 			uint64_t *size = &new_state->arcs_lsize[buftype];
1652 
1653 			/*
1654 			 * An L1 header always exists here, since if we're
1655 			 * moving to some L1-cached state (i.e. not l2c_only or
1656 			 * anonymous), we realloc the header to add an L1hdr
1657 			 * beforehand.
1658 			 */
1659 			ASSERT(HDR_HAS_L1HDR(hdr));
1660 			multilist_insert(&new_state->arcs_list[buftype], hdr);
1661 
1662 			/* ghost elements have a ghost size */
1663 			if (GHOST_STATE(new_state)) {
1664 				ASSERT0(datacnt);
1665 				ASSERT(hdr->b_l1hdr.b_buf == NULL);
1666 				to_delta = hdr->b_size;
1667 			}
1668 			atomic_add_64(size, to_delta);
1669 		}
1670 	}
1671 
1672 	ASSERT(!BUF_EMPTY(hdr));
1673 	if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
1674 		buf_hash_remove(hdr);
1675 
1676 	/* adjust state sizes (ignore arc_l2c_only) */
1677 
1678 	if (to_delta && new_state != arc_l2c_only) {
1679 		ASSERT(HDR_HAS_L1HDR(hdr));
1680 		if (GHOST_STATE(new_state)) {
1681 			ASSERT0(datacnt);
1682 
1683 			/*
1684 			 * We moving a header to a ghost state, we first
1685 			 * remove all arc buffers. Thus, we'll have a
1686 			 * datacnt of zero, and no arc buffer to use for
1687 			 * the reference. As a result, we use the arc
1688 			 * header pointer for the reference.
1689 			 */
1690 			(void) refcount_add_many(&new_state->arcs_size,
1691 			    hdr->b_size, hdr);
1692 		} else {
1693 			ASSERT3U(datacnt, !=, 0);
1694 
1695 			/*
1696 			 * Each individual buffer holds a unique reference,
1697 			 * thus we must remove each of these references one
1698 			 * at a time.
1699 			 */
1700 			for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1701 			    buf = buf->b_next) {
1702 				(void) refcount_add_many(&new_state->arcs_size,
1703 				    hdr->b_size, buf);
1704 			}
1705 		}
1706 	}
1707 
1708 	if (from_delta && old_state != arc_l2c_only) {
1709 		ASSERT(HDR_HAS_L1HDR(hdr));
1710 		if (GHOST_STATE(old_state)) {
1711 			/*
1712 			 * When moving a header off of a ghost state,
1713 			 * there's the possibility for datacnt to be
1714 			 * non-zero. This is because we first add the
1715 			 * arc buffer to the header prior to changing
1716 			 * the header's state. Since we used the header
1717 			 * for the reference when putting the header on
1718 			 * the ghost state, we must balance that and use
1719 			 * the header when removing off the ghost state
1720 			 * (even though datacnt is non zero).
1721 			 */
1722 
1723 			IMPLY(datacnt == 0, new_state == arc_anon ||
1724 			    new_state == arc_l2c_only);
1725 
1726 			(void) refcount_remove_many(&old_state->arcs_size,
1727 			    hdr->b_size, hdr);
1728 		} else {
1729 			ASSERT3P(datacnt, !=, 0);
1730 
1731 			/*
1732 			 * Each individual buffer holds a unique reference,
1733 			 * thus we must remove each of these references one
1734 			 * at a time.
1735 			 */
1736 			for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1737 			    buf = buf->b_next) {
1738 				(void) refcount_remove_many(
1739 				    &old_state->arcs_size, hdr->b_size, buf);
1740 			}
1741 		}
1742 	}
1743 
1744 	if (HDR_HAS_L1HDR(hdr))
1745 		hdr->b_l1hdr.b_state = new_state;
1746 
1747 	/*
1748 	 * L2 headers should never be on the L2 state list since they don't
1749 	 * have L1 headers allocated.
1750 	 */
1751 	ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
1752 	    multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
1753 }
1754 
1755 void
1756 arc_space_consume(uint64_t space, arc_space_type_t type)
1757 {
1758 	ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1759 
1760 	switch (type) {
1761 	case ARC_SPACE_DATA:
1762 		ARCSTAT_INCR(arcstat_data_size, space);
1763 		break;
1764 	case ARC_SPACE_META:
1765 		ARCSTAT_INCR(arcstat_metadata_size, space);
1766 		break;
1767 	case ARC_SPACE_OTHER:
1768 		ARCSTAT_INCR(arcstat_other_size, space);
1769 		break;
1770 	case ARC_SPACE_HDRS:
1771 		ARCSTAT_INCR(arcstat_hdr_size, space);
1772 		break;
1773 	case ARC_SPACE_L2HDRS:
1774 		ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1775 		break;
1776 	}
1777 
1778 	if (type != ARC_SPACE_DATA)
1779 		ARCSTAT_INCR(arcstat_meta_used, space);
1780 
1781 	atomic_add_64(&arc_size, space);
1782 }
1783 
1784 void
1785 arc_space_return(uint64_t space, arc_space_type_t type)
1786 {
1787 	ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1788 
1789 	switch (type) {
1790 	case ARC_SPACE_DATA:
1791 		ARCSTAT_INCR(arcstat_data_size, -space);
1792 		break;
1793 	case ARC_SPACE_META:
1794 		ARCSTAT_INCR(arcstat_metadata_size, -space);
1795 		break;
1796 	case ARC_SPACE_OTHER:
1797 		ARCSTAT_INCR(arcstat_other_size, -space);
1798 		break;
1799 	case ARC_SPACE_HDRS:
1800 		ARCSTAT_INCR(arcstat_hdr_size, -space);
1801 		break;
1802 	case ARC_SPACE_L2HDRS:
1803 		ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1804 		break;
1805 	}
1806 
1807 	if (type != ARC_SPACE_DATA) {
1808 		ASSERT(arc_meta_used >= space);
1809 		if (arc_meta_max < arc_meta_used)
1810 			arc_meta_max = arc_meta_used;
1811 		ARCSTAT_INCR(arcstat_meta_used, -space);
1812 	}
1813 
1814 	ASSERT(arc_size >= space);
1815 	atomic_add_64(&arc_size, -space);
1816 }
1817 
1818 arc_buf_t *
1819 arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type)
1820 {
1821 	arc_buf_hdr_t *hdr;
1822 	arc_buf_t *buf;
1823 
1824 	ASSERT3U(size, >, 0);
1825 	hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
1826 	ASSERT(BUF_EMPTY(hdr));
1827 	ASSERT3P(hdr->b_freeze_cksum, ==, NULL);
1828 	hdr->b_size = size;
1829 	hdr->b_spa = spa_load_guid(spa);
1830 
1831 	buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1832 	buf->b_hdr = hdr;
1833 	buf->b_data = NULL;
1834 	buf->b_efunc = NULL;
1835 	buf->b_private = NULL;
1836 	buf->b_next = NULL;
1837 
1838 	hdr->b_flags = arc_bufc_to_flags(type);
1839 	hdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1840 
1841 	hdr->b_l1hdr.b_buf = buf;
1842 	hdr->b_l1hdr.b_state = arc_anon;
1843 	hdr->b_l1hdr.b_arc_access = 0;
1844 	hdr->b_l1hdr.b_datacnt = 1;
1845 	hdr->b_l1hdr.b_tmp_cdata = NULL;
1846 
1847 	arc_get_data_buf(buf);
1848 	ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
1849 	(void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
1850 
1851 	return (buf);
1852 }
1853 
1854 static char *arc_onloan_tag = "onloan";
1855 
1856 /*
1857  * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1858  * flight data by arc_tempreserve_space() until they are "returned". Loaned
1859  * buffers must be returned to the arc before they can be used by the DMU or
1860  * freed.
1861  */
1862 arc_buf_t *
1863 arc_loan_buf(spa_t *spa, int size)
1864 {
1865 	arc_buf_t *buf;
1866 
1867 	buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1868 
1869 	atomic_add_64(&arc_loaned_bytes, size);
1870 	return (buf);
1871 }
1872 
1873 /*
1874  * Return a loaned arc buffer to the arc.
1875  */
1876 void
1877 arc_return_buf(arc_buf_t *buf, void *tag)
1878 {
1879 	arc_buf_hdr_t *hdr = buf->b_hdr;
1880 
1881 	ASSERT(buf->b_data != NULL);
1882 	ASSERT(HDR_HAS_L1HDR(hdr));
1883 	(void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
1884 	(void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
1885 
1886 	atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1887 }
1888 
1889 /* Detach an arc_buf from a dbuf (tag) */
1890 void
1891 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1892 {
1893 	arc_buf_hdr_t *hdr = buf->b_hdr;
1894 
1895 	ASSERT(buf->b_data != NULL);
1896 	ASSERT(HDR_HAS_L1HDR(hdr));
1897 	(void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
1898 	(void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
1899 	buf->b_efunc = NULL;
1900 	buf->b_private = NULL;
1901 
1902 	atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1903 }
1904 
1905 static arc_buf_t *
1906 arc_buf_clone(arc_buf_t *from)
1907 {
1908 	arc_buf_t *buf;
1909 	arc_buf_hdr_t *hdr = from->b_hdr;
1910 	uint64_t size = hdr->b_size;
1911 
1912 	ASSERT(HDR_HAS_L1HDR(hdr));
1913 	ASSERT(hdr->b_l1hdr.b_state != arc_anon);
1914 
1915 	buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1916 	buf->b_hdr = hdr;
1917 	buf->b_data = NULL;
1918 	buf->b_efunc = NULL;
1919 	buf->b_private = NULL;
1920 	buf->b_next = hdr->b_l1hdr.b_buf;
1921 	hdr->b_l1hdr.b_buf = buf;
1922 	arc_get_data_buf(buf);
1923 	bcopy(from->b_data, buf->b_data, size);
1924 
1925 	/*
1926 	 * This buffer already exists in the arc so create a duplicate
1927 	 * copy for the caller.  If the buffer is associated with user data
1928 	 * then track the size and number of duplicates.  These stats will be
1929 	 * updated as duplicate buffers are created and destroyed.
1930 	 */
1931 	if (HDR_ISTYPE_DATA(hdr)) {
1932 		ARCSTAT_BUMP(arcstat_duplicate_buffers);
1933 		ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1934 	}
1935 	hdr->b_l1hdr.b_datacnt += 1;
1936 	return (buf);
1937 }
1938 
1939 void
1940 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1941 {
1942 	arc_buf_hdr_t *hdr;
1943 	kmutex_t *hash_lock;
1944 
1945 	/*
1946 	 * Check to see if this buffer is evicted.  Callers
1947 	 * must verify b_data != NULL to know if the add_ref
1948 	 * was successful.
1949 	 */
1950 	mutex_enter(&buf->b_evict_lock);
1951 	if (buf->b_data == NULL) {
1952 		mutex_exit(&buf->b_evict_lock);
1953 		return;
1954 	}
1955 	hash_lock = HDR_LOCK(buf->b_hdr);
1956 	mutex_enter(hash_lock);
1957 	hdr = buf->b_hdr;
1958 	ASSERT(HDR_HAS_L1HDR(hdr));
1959 	ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1960 	mutex_exit(&buf->b_evict_lock);
1961 
1962 	ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
1963 	    hdr->b_l1hdr.b_state == arc_mfu);
1964 
1965 	add_reference(hdr, hash_lock, tag);
1966 	DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1967 	arc_access(hdr, hash_lock);
1968 	mutex_exit(hash_lock);
1969 	ARCSTAT_BUMP(arcstat_hits);
1970 	ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
1971 	    demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
1972 	    data, metadata, hits);
1973 }
1974 
1975 static void
1976 arc_buf_free_on_write(void *data, size_t size,
1977     void (*free_func)(void *, size_t))
1978 {
1979 	l2arc_data_free_t *df;
1980 
1981 	df = kmem_alloc(sizeof (*df), KM_SLEEP);
1982 	df->l2df_data = data;
1983 	df->l2df_size = size;
1984 	df->l2df_func = free_func;
1985 	mutex_enter(&l2arc_free_on_write_mtx);
1986 	list_insert_head(l2arc_free_on_write, df);
1987 	mutex_exit(&l2arc_free_on_write_mtx);
1988 }
1989 
1990 /*
1991  * Free the arc data buffer.  If it is an l2arc write in progress,
1992  * the buffer is placed on l2arc_free_on_write to be freed later.
1993  */
1994 static void
1995 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1996 {
1997 	arc_buf_hdr_t *hdr = buf->b_hdr;
1998 
1999 	if (HDR_L2_WRITING(hdr)) {
2000 		arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func);
2001 		ARCSTAT_BUMP(arcstat_l2_free_on_write);
2002 	} else {
2003 		free_func(buf->b_data, hdr->b_size);
2004 	}
2005 }
2006 
2007 static void
2008 arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr)
2009 {
2010 	ASSERT(HDR_HAS_L2HDR(hdr));
2011 	ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx));
2012 
2013 	/*
2014 	 * The b_tmp_cdata field is linked off of the b_l1hdr, so if
2015 	 * that doesn't exist, the header is in the arc_l2c_only state,
2016 	 * and there isn't anything to free (it's already been freed).
2017 	 */
2018 	if (!HDR_HAS_L1HDR(hdr))
2019 		return;
2020 
2021 	/*
2022 	 * The header isn't being written to the l2arc device, thus it
2023 	 * shouldn't have a b_tmp_cdata to free.
2024 	 */
2025 	if (!HDR_L2_WRITING(hdr)) {
2026 		ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2027 		return;
2028 	}
2029 
2030 	/*
2031 	 * The header does not have compression enabled. This can be due
2032 	 * to the buffer not being compressible, or because we're
2033 	 * freeing the buffer before the second phase of
2034 	 * l2arc_write_buffer() has started (which does the compression
2035 	 * step). In either case, b_tmp_cdata does not point to a
2036 	 * separately compressed buffer, so there's nothing to free (it
2037 	 * points to the same buffer as the arc_buf_t's b_data field).
2038 	 */
2039 	if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) {
2040 		hdr->b_l1hdr.b_tmp_cdata = NULL;
2041 		return;
2042 	}
2043 
2044 	/*
2045 	 * There's nothing to free since the buffer was all zero's and
2046 	 * compressed to a zero length buffer.
2047 	 */
2048 	if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_EMPTY) {
2049 		ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2050 		return;
2051 	}
2052 
2053 	ASSERT(L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr)));
2054 
2055 	arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata,
2056 	    hdr->b_size, zio_data_buf_free);
2057 
2058 	ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write);
2059 	hdr->b_l1hdr.b_tmp_cdata = NULL;
2060 }
2061 
2062 /*
2063  * Free up buf->b_data and if 'remove' is set, then pull the
2064  * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
2065  */
2066 static void
2067 arc_buf_destroy(arc_buf_t *buf, boolean_t remove)
2068 {
2069 	arc_buf_t **bufp;
2070 
2071 	/* free up data associated with the buf */
2072 	if (buf->b_data != NULL) {
2073 		arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
2074 		uint64_t size = buf->b_hdr->b_size;
2075 		arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
2076 
2077 		arc_cksum_verify(buf);
2078 		arc_buf_unwatch(buf);
2079 
2080 		if (type == ARC_BUFC_METADATA) {
2081 			arc_buf_data_free(buf, zio_buf_free);
2082 			arc_space_return(size, ARC_SPACE_META);
2083 		} else {
2084 			ASSERT(type == ARC_BUFC_DATA);
2085 			arc_buf_data_free(buf, zio_data_buf_free);
2086 			arc_space_return(size, ARC_SPACE_DATA);
2087 		}
2088 
2089 		/* protected by hash lock, if in the hash table */
2090 		if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) {
2091 			uint64_t *cnt = &state->arcs_lsize[type];
2092 
2093 			ASSERT(refcount_is_zero(
2094 			    &buf->b_hdr->b_l1hdr.b_refcnt));
2095 			ASSERT(state != arc_anon && state != arc_l2c_only);
2096 
2097 			ASSERT3U(*cnt, >=, size);
2098 			atomic_add_64(cnt, -size);
2099 		}
2100 
2101 		(void) refcount_remove_many(&state->arcs_size, size, buf);
2102 		buf->b_data = NULL;
2103 
2104 		/*
2105 		 * If we're destroying a duplicate buffer make sure
2106 		 * that the appropriate statistics are updated.
2107 		 */
2108 		if (buf->b_hdr->b_l1hdr.b_datacnt > 1 &&
2109 		    HDR_ISTYPE_DATA(buf->b_hdr)) {
2110 			ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
2111 			ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
2112 		}
2113 		ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0);
2114 		buf->b_hdr->b_l1hdr.b_datacnt -= 1;
2115 	}
2116 
2117 	/* only remove the buf if requested */
2118 	if (!remove)
2119 		return;
2120 
2121 	/* remove the buf from the hdr list */
2122 	for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf;
2123 	    bufp = &(*bufp)->b_next)
2124 		continue;
2125 	*bufp = buf->b_next;
2126 	buf->b_next = NULL;
2127 
2128 	ASSERT(buf->b_efunc == NULL);
2129 
2130 	/* clean up the buf */
2131 	buf->b_hdr = NULL;
2132 	kmem_cache_free(buf_cache, buf);
2133 }
2134 
2135 static void
2136 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
2137 {
2138 	l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
2139 	l2arc_dev_t *dev = l2hdr->b_dev;
2140 
2141 	ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
2142 	ASSERT(HDR_HAS_L2HDR(hdr));
2143 
2144 	list_remove(&dev->l2ad_buflist, hdr);
2145 
2146 	/*
2147 	 * We don't want to leak the b_tmp_cdata buffer that was
2148 	 * allocated in l2arc_write_buffers()
2149 	 */
2150 	arc_buf_l2_cdata_free(hdr);
2151 
2152 	/*
2153 	 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then
2154 	 * this header is being processed by l2arc_write_buffers() (i.e.
2155 	 * it's in the first stage of l2arc_write_buffers()).
2156 	 * Re-affirming that truth here, just to serve as a reminder. If
2157 	 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or
2158 	 * may not have its HDR_L2_WRITING flag set. (the write may have
2159 	 * completed, in which case HDR_L2_WRITING will be false and the
2160 	 * b_daddr field will point to the address of the buffer on disk).
2161 	 */
2162 	IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr));
2163 
2164 	/*
2165 	 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with
2166 	 * l2arc_write_buffers(). Since we've just removed this header
2167 	 * from the l2arc buffer list, this header will never reach the
2168 	 * second stage of l2arc_write_buffers(), which increments the
2169 	 * accounting stats for this header. Thus, we must be careful
2170 	 * not to decrement them for this header either.
2171 	 */
2172 	if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) {
2173 		ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
2174 		ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
2175 
2176 		vdev_space_update(dev->l2ad_vdev,
2177 		    -l2hdr->b_asize, 0, 0);
2178 
2179 		(void) refcount_remove_many(&dev->l2ad_alloc,
2180 		    l2hdr->b_asize, hdr);
2181 	}
2182 
2183 	hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
2184 }
2185 
2186 static void
2187 arc_hdr_destroy(arc_buf_hdr_t *hdr)
2188 {
2189 	if (HDR_HAS_L1HDR(hdr)) {
2190 		ASSERT(hdr->b_l1hdr.b_buf == NULL ||
2191 		    hdr->b_l1hdr.b_datacnt > 0);
2192 		ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2193 		ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2194 	}
2195 	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2196 	ASSERT(!HDR_IN_HASH_TABLE(hdr));
2197 
2198 	if (HDR_HAS_L2HDR(hdr)) {
2199 		l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2200 		boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
2201 
2202 		if (!buflist_held)
2203 			mutex_enter(&dev->l2ad_mtx);
2204 
2205 		/*
2206 		 * Even though we checked this conditional above, we
2207 		 * need to check this again now that we have the
2208 		 * l2ad_mtx. This is because we could be racing with
2209 		 * another thread calling l2arc_evict() which might have
2210 		 * destroyed this header's L2 portion as we were waiting
2211 		 * to acquire the l2ad_mtx. If that happens, we don't
2212 		 * want to re-destroy the header's L2 portion.
2213 		 */
2214 		if (HDR_HAS_L2HDR(hdr))
2215 			arc_hdr_l2hdr_destroy(hdr);
2216 
2217 		if (!buflist_held)
2218 			mutex_exit(&dev->l2ad_mtx);
2219 	}
2220 
2221 	if (!BUF_EMPTY(hdr))
2222 		buf_discard_identity(hdr);
2223 
2224 	if (hdr->b_freeze_cksum != NULL) {
2225 		kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
2226 		hdr->b_freeze_cksum = NULL;
2227 	}
2228 
2229 	if (HDR_HAS_L1HDR(hdr)) {
2230 		while (hdr->b_l1hdr.b_buf) {
2231 			arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2232 
2233 			if (buf->b_efunc != NULL) {
2234 				mutex_enter(&arc_user_evicts_lock);
2235 				mutex_enter(&buf->b_evict_lock);
2236 				ASSERT(buf->b_hdr != NULL);
2237 				arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE);
2238 				hdr->b_l1hdr.b_buf = buf->b_next;
2239 				buf->b_hdr = &arc_eviction_hdr;
2240 				buf->b_next = arc_eviction_list;
2241 				arc_eviction_list = buf;
2242 				mutex_exit(&buf->b_evict_lock);
2243 				cv_signal(&arc_user_evicts_cv);
2244 				mutex_exit(&arc_user_evicts_lock);
2245 			} else {
2246 				arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE);
2247 			}
2248 		}
2249 #ifdef ZFS_DEBUG
2250 		if (hdr->b_l1hdr.b_thawed != NULL) {
2251 			kmem_free(hdr->b_l1hdr.b_thawed, 1);
2252 			hdr->b_l1hdr.b_thawed = NULL;
2253 		}
2254 #endif
2255 	}
2256 
2257 	ASSERT3P(hdr->b_hash_next, ==, NULL);
2258 	if (HDR_HAS_L1HDR(hdr)) {
2259 		ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
2260 		ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
2261 		kmem_cache_free(hdr_full_cache, hdr);
2262 	} else {
2263 		kmem_cache_free(hdr_l2only_cache, hdr);
2264 	}
2265 }
2266 
2267 void
2268 arc_buf_free(arc_buf_t *buf, void *tag)
2269 {
2270 	arc_buf_hdr_t *hdr = buf->b_hdr;
2271 	int hashed = hdr->b_l1hdr.b_state != arc_anon;
2272 
2273 	ASSERT(buf->b_efunc == NULL);
2274 	ASSERT(buf->b_data != NULL);
2275 
2276 	if (hashed) {
2277 		kmutex_t *hash_lock = HDR_LOCK(hdr);
2278 
2279 		mutex_enter(hash_lock);
2280 		hdr = buf->b_hdr;
2281 		ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2282 
2283 		(void) remove_reference(hdr, hash_lock, tag);
2284 		if (hdr->b_l1hdr.b_datacnt > 1) {
2285 			arc_buf_destroy(buf, TRUE);
2286 		} else {
2287 			ASSERT(buf == hdr->b_l1hdr.b_buf);
2288 			ASSERT(buf->b_efunc == NULL);
2289 			hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2290 		}
2291 		mutex_exit(hash_lock);
2292 	} else if (HDR_IO_IN_PROGRESS(hdr)) {
2293 		int destroy_hdr;
2294 		/*
2295 		 * We are in the middle of an async write.  Don't destroy
2296 		 * this buffer unless the write completes before we finish
2297 		 * decrementing the reference count.
2298 		 */
2299 		mutex_enter(&arc_user_evicts_lock);
2300 		(void) remove_reference(hdr, NULL, tag);
2301 		ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2302 		destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
2303 		mutex_exit(&arc_user_evicts_lock);
2304 		if (destroy_hdr)
2305 			arc_hdr_destroy(hdr);
2306 	} else {
2307 		if (remove_reference(hdr, NULL, tag) > 0)
2308 			arc_buf_destroy(buf, TRUE);
2309 		else
2310 			arc_hdr_destroy(hdr);
2311 	}
2312 }
2313 
2314 boolean_t
2315 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
2316 {
2317 	arc_buf_hdr_t *hdr = buf->b_hdr;
2318 	kmutex_t *hash_lock = HDR_LOCK(hdr);
2319 	boolean_t no_callback = (buf->b_efunc == NULL);
2320 
2321 	if (hdr->b_l1hdr.b_state == arc_anon) {
2322 		ASSERT(hdr->b_l1hdr.b_datacnt == 1);
2323 		arc_buf_free(buf, tag);
2324 		return (no_callback);
2325 	}
2326 
2327 	mutex_enter(hash_lock);
2328 	hdr = buf->b_hdr;
2329 	ASSERT(hdr->b_l1hdr.b_datacnt > 0);
2330 	ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2331 	ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2332 	ASSERT(buf->b_data != NULL);
2333 
2334 	(void) remove_reference(hdr, hash_lock, tag);
2335 	if (hdr->b_l1hdr.b_datacnt > 1) {
2336 		if (no_callback)
2337 			arc_buf_destroy(buf, TRUE);
2338 	} else if (no_callback) {
2339 		ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL);
2340 		ASSERT(buf->b_efunc == NULL);
2341 		hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2342 	}
2343 	ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 ||
2344 	    refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2345 	mutex_exit(hash_lock);
2346 	return (no_callback);
2347 }
2348 
2349 int32_t
2350 arc_buf_size(arc_buf_t *buf)
2351 {
2352 	return (buf->b_hdr->b_size);
2353 }
2354 
2355 /*
2356  * Called from the DMU to determine if the current buffer should be
2357  * evicted. In order to ensure proper locking, the eviction must be initiated
2358  * from the DMU. Return true if the buffer is associated with user data and
2359  * duplicate buffers still exist.
2360  */
2361 boolean_t
2362 arc_buf_eviction_needed(arc_buf_t *buf)
2363 {
2364 	arc_buf_hdr_t *hdr;
2365 	boolean_t evict_needed = B_FALSE;
2366 
2367 	if (zfs_disable_dup_eviction)
2368 		return (B_FALSE);
2369 
2370 	mutex_enter(&buf->b_evict_lock);
2371 	hdr = buf->b_hdr;
2372 	if (hdr == NULL) {
2373 		/*
2374 		 * We are in arc_do_user_evicts(); let that function
2375 		 * perform the eviction.
2376 		 */
2377 		ASSERT(buf->b_data == NULL);
2378 		mutex_exit(&buf->b_evict_lock);
2379 		return (B_FALSE);
2380 	} else if (buf->b_data == NULL) {
2381 		/*
2382 		 * We have already been added to the arc eviction list;
2383 		 * recommend eviction.
2384 		 */
2385 		ASSERT3P(hdr, ==, &arc_eviction_hdr);
2386 		mutex_exit(&buf->b_evict_lock);
2387 		return (B_TRUE);
2388 	}
2389 
2390 	if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr))
2391 		evict_needed = B_TRUE;
2392 
2393 	mutex_exit(&buf->b_evict_lock);
2394 	return (evict_needed);
2395 }
2396 
2397 /*
2398  * Evict the arc_buf_hdr that is provided as a parameter. The resultant
2399  * state of the header is dependent on it's state prior to entering this
2400  * function. The following transitions are possible:
2401  *
2402  *    - arc_mru -> arc_mru_ghost
2403  *    - arc_mfu -> arc_mfu_ghost
2404  *    - arc_mru_ghost -> arc_l2c_only
2405  *    - arc_mru_ghost -> deleted
2406  *    - arc_mfu_ghost -> arc_l2c_only
2407  *    - arc_mfu_ghost -> deleted
2408  */
2409 static int64_t
2410 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
2411 {
2412 	arc_state_t *evicted_state, *state;
2413 	int64_t bytes_evicted = 0;
2414 
2415 	ASSERT(MUTEX_HELD(hash_lock));
2416 	ASSERT(HDR_HAS_L1HDR(hdr));
2417 
2418 	state = hdr->b_l1hdr.b_state;
2419 	if (GHOST_STATE(state)) {
2420 		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2421 		ASSERT(hdr->b_l1hdr.b_buf == NULL);
2422 
2423 		/*
2424 		 * l2arc_write_buffers() relies on a header's L1 portion
2425 		 * (i.e. it's b_tmp_cdata field) during it's write phase.
2426 		 * Thus, we cannot push a header onto the arc_l2c_only
2427 		 * state (removing it's L1 piece) until the header is
2428 		 * done being written to the l2arc.
2429 		 */
2430 		if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
2431 			ARCSTAT_BUMP(arcstat_evict_l2_skip);
2432 			return (bytes_evicted);
2433 		}
2434 
2435 		ARCSTAT_BUMP(arcstat_deleted);
2436 		bytes_evicted += hdr->b_size;
2437 
2438 		DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
2439 
2440 		if (HDR_HAS_L2HDR(hdr)) {
2441 			/*
2442 			 * This buffer is cached on the 2nd Level ARC;
2443 			 * don't destroy the header.
2444 			 */
2445 			arc_change_state(arc_l2c_only, hdr, hash_lock);
2446 			/*
2447 			 * dropping from L1+L2 cached to L2-only,
2448 			 * realloc to remove the L1 header.
2449 			 */
2450 			hdr = arc_hdr_realloc(hdr, hdr_full_cache,
2451 			    hdr_l2only_cache);
2452 		} else {
2453 			arc_change_state(arc_anon, hdr, hash_lock);
2454 			arc_hdr_destroy(hdr);
2455 		}
2456 		return (bytes_evicted);
2457 	}
2458 
2459 	ASSERT(state == arc_mru || state == arc_mfu);
2460 	evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
2461 
2462 	/* prefetch buffers have a minimum lifespan */
2463 	if (HDR_IO_IN_PROGRESS(hdr) ||
2464 	    ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
2465 	    ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
2466 	    arc_min_prefetch_lifespan)) {
2467 		ARCSTAT_BUMP(arcstat_evict_skip);
2468 		return (bytes_evicted);
2469 	}
2470 
2471 	ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
2472 	ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0);
2473 	while (hdr->b_l1hdr.b_buf) {
2474 		arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2475 		if (!mutex_tryenter(&buf->b_evict_lock)) {
2476 			ARCSTAT_BUMP(arcstat_mutex_miss);
2477 			break;
2478 		}
2479 		if (buf->b_data != NULL)
2480 			bytes_evicted += hdr->b_size;
2481 		if (buf->b_efunc != NULL) {
2482 			mutex_enter(&arc_user_evicts_lock);
2483 			arc_buf_destroy(buf, FALSE);
2484 			hdr->b_l1hdr.b_buf = buf->b_next;
2485 			buf->b_hdr = &arc_eviction_hdr;
2486 			buf->b_next = arc_eviction_list;
2487 			arc_eviction_list = buf;
2488 			cv_signal(&arc_user_evicts_cv);
2489 			mutex_exit(&arc_user_evicts_lock);
2490 			mutex_exit(&buf->b_evict_lock);
2491 		} else {
2492 			mutex_exit(&buf->b_evict_lock);
2493 			arc_buf_destroy(buf, TRUE);
2494 		}
2495 	}
2496 
2497 	if (HDR_HAS_L2HDR(hdr)) {
2498 		ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size);
2499 	} else {
2500 		if (l2arc_write_eligible(hdr->b_spa, hdr))
2501 			ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size);
2502 		else
2503 			ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size);
2504 	}
2505 
2506 	if (hdr->b_l1hdr.b_datacnt == 0) {
2507 		arc_change_state(evicted_state, hdr, hash_lock);
2508 		ASSERT(HDR_IN_HASH_TABLE(hdr));
2509 		hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
2510 		hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
2511 		DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
2512 	}
2513 
2514 	return (bytes_evicted);
2515 }
2516 
2517 static uint64_t
2518 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
2519     uint64_t spa, int64_t bytes)
2520 {
2521 	multilist_sublist_t *mls;
2522 	uint64_t bytes_evicted = 0;
2523 	arc_buf_hdr_t *hdr;
2524 	kmutex_t *hash_lock;
2525 	int evict_count = 0;
2526 
2527 	ASSERT3P(marker, !=, NULL);
2528 	IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2529 
2530 	mls = multilist_sublist_lock(ml, idx);
2531 
2532 	for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
2533 	    hdr = multilist_sublist_prev(mls, marker)) {
2534 		if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
2535 		    (evict_count >= zfs_arc_evict_batch_limit))
2536 			break;
2537 
2538 		/*
2539 		 * To keep our iteration location, move the marker
2540 		 * forward. Since we're not holding hdr's hash lock, we
2541 		 * must be very careful and not remove 'hdr' from the
2542 		 * sublist. Otherwise, other consumers might mistake the
2543 		 * 'hdr' as not being on a sublist when they call the
2544 		 * multilist_link_active() function (they all rely on
2545 		 * the hash lock protecting concurrent insertions and
2546 		 * removals). multilist_sublist_move_forward() was
2547 		 * specifically implemented to ensure this is the case
2548 		 * (only 'marker' will be removed and re-inserted).
2549 		 */
2550 		multilist_sublist_move_forward(mls, marker);
2551 
2552 		/*
2553 		 * The only case where the b_spa field should ever be
2554 		 * zero, is the marker headers inserted by
2555 		 * arc_evict_state(). It's possible for multiple threads
2556 		 * to be calling arc_evict_state() concurrently (e.g.
2557 		 * dsl_pool_close() and zio_inject_fault()), so we must
2558 		 * skip any markers we see from these other threads.
2559 		 */
2560 		if (hdr->b_spa == 0)
2561 			continue;
2562 
2563 		/* we're only interested in evicting buffers of a certain spa */
2564 		if (spa != 0 && hdr->b_spa != spa) {
2565 			ARCSTAT_BUMP(arcstat_evict_skip);
2566 			continue;
2567 		}
2568 
2569 		hash_lock = HDR_LOCK(hdr);
2570 
2571 		/*
2572 		 * We aren't calling this function from any code path
2573 		 * that would already be holding a hash lock, so we're
2574 		 * asserting on this assumption to be defensive in case
2575 		 * this ever changes. Without this check, it would be
2576 		 * possible to incorrectly increment arcstat_mutex_miss
2577 		 * below (e.g. if the code changed such that we called
2578 		 * this function with a hash lock held).
2579 		 */
2580 		ASSERT(!MUTEX_HELD(hash_lock));
2581 
2582 		if (mutex_tryenter(hash_lock)) {
2583 			uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
2584 			mutex_exit(hash_lock);
2585 
2586 			bytes_evicted += evicted;
2587 
2588 			/*
2589 			 * If evicted is zero, arc_evict_hdr() must have
2590 			 * decided to skip this header, don't increment
2591 			 * evict_count in this case.
2592 			 */
2593 			if (evicted != 0)
2594 				evict_count++;
2595 
2596 			/*
2597 			 * If arc_size isn't overflowing, signal any
2598 			 * threads that might happen to be waiting.
2599 			 *
2600 			 * For each header evicted, we wake up a single
2601 			 * thread. If we used cv_broadcast, we could
2602 			 * wake up "too many" threads causing arc_size
2603 			 * to significantly overflow arc_c; since
2604 			 * arc_get_data_buf() doesn't check for overflow
2605 			 * when it's woken up (it doesn't because it's
2606 			 * possible for the ARC to be overflowing while
2607 			 * full of un-evictable buffers, and the
2608 			 * function should proceed in this case).
2609 			 *
2610 			 * If threads are left sleeping, due to not
2611 			 * using cv_broadcast, they will be woken up
2612 			 * just before arc_reclaim_thread() sleeps.
2613 			 */
2614 			mutex_enter(&arc_reclaim_lock);
2615 			if (!arc_is_overflowing())
2616 				cv_signal(&arc_reclaim_waiters_cv);
2617 			mutex_exit(&arc_reclaim_lock);
2618 		} else {
2619 			ARCSTAT_BUMP(arcstat_mutex_miss);
2620 		}
2621 	}
2622 
2623 	multilist_sublist_unlock(mls);
2624 
2625 	return (bytes_evicted);
2626 }
2627 
2628 /*
2629  * Evict buffers from the given arc state, until we've removed the
2630  * specified number of bytes. Move the removed buffers to the
2631  * appropriate evict state.
2632  *
2633  * This function makes a "best effort". It skips over any buffers
2634  * it can't get a hash_lock on, and so, may not catch all candidates.
2635  * It may also return without evicting as much space as requested.
2636  *
2637  * If bytes is specified using the special value ARC_EVICT_ALL, this
2638  * will evict all available (i.e. unlocked and evictable) buffers from
2639  * the given arc state; which is used by arc_flush().
2640  */
2641 static uint64_t
2642 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
2643     arc_buf_contents_t type)
2644 {
2645 	uint64_t total_evicted = 0;
2646 	multilist_t *ml = &state->arcs_list[type];
2647 	int num_sublists;
2648 	arc_buf_hdr_t **markers;
2649 
2650 	IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2651 
2652 	num_sublists = multilist_get_num_sublists(ml);
2653 
2654 	/*
2655 	 * If we've tried to evict from each sublist, made some
2656 	 * progress, but still have not hit the target number of bytes
2657 	 * to evict, we want to keep trying. The markers allow us to
2658 	 * pick up where we left off for each individual sublist, rather
2659 	 * than starting from the tail each time.
2660 	 */
2661 	markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
2662 	for (int i = 0; i < num_sublists; i++) {
2663 		markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
2664 
2665 		/*
2666 		 * A b_spa of 0 is used to indicate that this header is
2667 		 * a marker. This fact is used in arc_adjust_type() and
2668 		 * arc_evict_state_impl().
2669 		 */
2670 		markers[i]->b_spa = 0;
2671 
2672 		multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2673 		multilist_sublist_insert_tail(mls, markers[i]);
2674 		multilist_sublist_unlock(mls);
2675 	}
2676 
2677 	/*
2678 	 * While we haven't hit our target number of bytes to evict, or
2679 	 * we're evicting all available buffers.
2680 	 */
2681 	while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
2682 		/*
2683 		 * Start eviction using a randomly selected sublist,
2684 		 * this is to try and evenly balance eviction across all
2685 		 * sublists. Always starting at the same sublist
2686 		 * (e.g. index 0) would cause evictions to favor certain
2687 		 * sublists over others.
2688 		 */
2689 		int sublist_idx = multilist_get_random_index(ml);
2690 		uint64_t scan_evicted = 0;
2691 
2692 		for (int i = 0; i < num_sublists; i++) {
2693 			uint64_t bytes_remaining;
2694 			uint64_t bytes_evicted;
2695 
2696 			if (bytes == ARC_EVICT_ALL)
2697 				bytes_remaining = ARC_EVICT_ALL;
2698 			else if (total_evicted < bytes)
2699 				bytes_remaining = bytes - total_evicted;
2700 			else
2701 				break;
2702 
2703 			bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
2704 			    markers[sublist_idx], spa, bytes_remaining);
2705 
2706 			scan_evicted += bytes_evicted;
2707 			total_evicted += bytes_evicted;
2708 
2709 			/* we've reached the end, wrap to the beginning */
2710 			if (++sublist_idx >= num_sublists)
2711 				sublist_idx = 0;
2712 		}
2713 
2714 		/*
2715 		 * If we didn't evict anything during this scan, we have
2716 		 * no reason to believe we'll evict more during another
2717 		 * scan, so break the loop.
2718 		 */
2719 		if (scan_evicted == 0) {
2720 			/* This isn't possible, let's make that obvious */
2721 			ASSERT3S(bytes, !=, 0);
2722 
2723 			/*
2724 			 * When bytes is ARC_EVICT_ALL, the only way to
2725 			 * break the loop is when scan_evicted is zero.
2726 			 * In that case, we actually have evicted enough,
2727 			 * so we don't want to increment the kstat.
2728 			 */
2729 			if (bytes != ARC_EVICT_ALL) {
2730 				ASSERT3S(total_evicted, <, bytes);
2731 				ARCSTAT_BUMP(arcstat_evict_not_enough);
2732 			}
2733 
2734 			break;
2735 		}
2736 	}
2737 
2738 	for (int i = 0; i < num_sublists; i++) {
2739 		multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2740 		multilist_sublist_remove(mls, markers[i]);
2741 		multilist_sublist_unlock(mls);
2742 
2743 		kmem_cache_free(hdr_full_cache, markers[i]);
2744 	}
2745 	kmem_free(markers, sizeof (*markers) * num_sublists);
2746 
2747 	return (total_evicted);
2748 }
2749 
2750 /*
2751  * Flush all "evictable" data of the given type from the arc state
2752  * specified. This will not evict any "active" buffers (i.e. referenced).
2753  *
2754  * When 'retry' is set to FALSE, the function will make a single pass
2755  * over the state and evict any buffers that it can. Since it doesn't
2756  * continually retry the eviction, it might end up leaving some buffers
2757  * in the ARC due to lock misses.
2758  *
2759  * When 'retry' is set to TRUE, the function will continually retry the
2760  * eviction until *all* evictable buffers have been removed from the
2761  * state. As a result, if concurrent insertions into the state are
2762  * allowed (e.g. if the ARC isn't shutting down), this function might
2763  * wind up in an infinite loop, continually trying to evict buffers.
2764  */
2765 static uint64_t
2766 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
2767     boolean_t retry)
2768 {
2769 	uint64_t evicted = 0;
2770 
2771 	while (state->arcs_lsize[type] != 0) {
2772 		evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
2773 
2774 		if (!retry)
2775 			break;
2776 	}
2777 
2778 	return (evicted);
2779 }
2780 
2781 /*
2782  * Evict the specified number of bytes from the state specified,
2783  * restricting eviction to the spa and type given. This function
2784  * prevents us from trying to evict more from a state's list than
2785  * is "evictable", and to skip evicting altogether when passed a
2786  * negative value for "bytes". In contrast, arc_evict_state() will
2787  * evict everything it can, when passed a negative value for "bytes".
2788  */
2789 static uint64_t
2790 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
2791     arc_buf_contents_t type)
2792 {
2793 	int64_t delta;
2794 
2795 	if (bytes > 0 && state->arcs_lsize[type] > 0) {
2796 		delta = MIN(state->arcs_lsize[type], bytes);
2797 		return (arc_evict_state(state, spa, delta, type));
2798 	}
2799 
2800 	return (0);
2801 }
2802 
2803 /*
2804  * Evict metadata buffers from the cache, such that arc_meta_used is
2805  * capped by the arc_meta_limit tunable.
2806  */
2807 static uint64_t
2808 arc_adjust_meta(void)
2809 {
2810 	uint64_t total_evicted = 0;
2811 	int64_t target;
2812 
2813 	/*
2814 	 * If we're over the meta limit, we want to evict enough
2815 	 * metadata to get back under the meta limit. We don't want to
2816 	 * evict so much that we drop the MRU below arc_p, though. If
2817 	 * we're over the meta limit more than we're over arc_p, we
2818 	 * evict some from the MRU here, and some from the MFU below.
2819 	 */
2820 	target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
2821 	    (int64_t)(refcount_count(&arc_anon->arcs_size) +
2822 	    refcount_count(&arc_mru->arcs_size) - arc_p));
2823 
2824 	total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
2825 
2826 	/*
2827 	 * Similar to the above, we want to evict enough bytes to get us
2828 	 * below the meta limit, but not so much as to drop us below the
2829 	 * space alloted to the MFU (which is defined as arc_c - arc_p).
2830 	 */
2831 	target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
2832 	    (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
2833 
2834 	total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
2835 
2836 	return (total_evicted);
2837 }
2838 
2839 /*
2840  * Return the type of the oldest buffer in the given arc state
2841  *
2842  * This function will select a random sublist of type ARC_BUFC_DATA and
2843  * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
2844  * is compared, and the type which contains the "older" buffer will be
2845  * returned.
2846  */
2847 static arc_buf_contents_t
2848 arc_adjust_type(arc_state_t *state)
2849 {
2850 	multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
2851 	multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
2852 	int data_idx = multilist_get_random_index(data_ml);
2853 	int meta_idx = multilist_get_random_index(meta_ml);
2854 	multilist_sublist_t *data_mls;
2855 	multilist_sublist_t *meta_mls;
2856 	arc_buf_contents_t type;
2857 	arc_buf_hdr_t *data_hdr;
2858 	arc_buf_hdr_t *meta_hdr;
2859 
2860 	/*
2861 	 * We keep the sublist lock until we're finished, to prevent
2862 	 * the headers from being destroyed via arc_evict_state().
2863 	 */
2864 	data_mls = multilist_sublist_lock(data_ml, data_idx);
2865 	meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
2866 
2867 	/*
2868 	 * These two loops are to ensure we skip any markers that
2869 	 * might be at the tail of the lists due to arc_evict_state().
2870 	 */
2871 
2872 	for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
2873 	    data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
2874 		if (data_hdr->b_spa != 0)
2875 			break;
2876 	}
2877 
2878 	for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
2879 	    meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
2880 		if (meta_hdr->b_spa != 0)
2881 			break;
2882 	}
2883 
2884 	if (data_hdr == NULL && meta_hdr == NULL) {
2885 		type = ARC_BUFC_DATA;
2886 	} else if (data_hdr == NULL) {
2887 		ASSERT3P(meta_hdr, !=, NULL);
2888 		type = ARC_BUFC_METADATA;
2889 	} else if (meta_hdr == NULL) {
2890 		ASSERT3P(data_hdr, !=, NULL);
2891 		type = ARC_BUFC_DATA;
2892 	} else {
2893 		ASSERT3P(data_hdr, !=, NULL);
2894 		ASSERT3P(meta_hdr, !=, NULL);
2895 
2896 		/* The headers can't be on the sublist without an L1 header */
2897 		ASSERT(HDR_HAS_L1HDR(data_hdr));
2898 		ASSERT(HDR_HAS_L1HDR(meta_hdr));
2899 
2900 		if (data_hdr->b_l1hdr.b_arc_access <
2901 		    meta_hdr->b_l1hdr.b_arc_access) {
2902 			type = ARC_BUFC_DATA;
2903 		} else {
2904 			type = ARC_BUFC_METADATA;
2905 		}
2906 	}
2907 
2908 	multilist_sublist_unlock(meta_mls);
2909 	multilist_sublist_unlock(data_mls);
2910 
2911 	return (type);
2912 }
2913 
2914 /*
2915  * Evict buffers from the cache, such that arc_size is capped by arc_c.
2916  */
2917 static uint64_t
2918 arc_adjust(void)
2919 {
2920 	uint64_t total_evicted = 0;
2921 	uint64_t bytes;
2922 	int64_t target;
2923 
2924 	/*
2925 	 * If we're over arc_meta_limit, we want to correct that before
2926 	 * potentially evicting data buffers below.
2927 	 */
2928 	total_evicted += arc_adjust_meta();
2929 
2930 	/*
2931 	 * Adjust MRU size
2932 	 *
2933 	 * If we're over the target cache size, we want to evict enough
2934 	 * from the list to get back to our target size. We don't want
2935 	 * to evict too much from the MRU, such that it drops below
2936 	 * arc_p. So, if we're over our target cache size more than
2937 	 * the MRU is over arc_p, we'll evict enough to get back to
2938 	 * arc_p here, and then evict more from the MFU below.
2939 	 */
2940 	target = MIN((int64_t)(arc_size - arc_c),
2941 	    (int64_t)(refcount_count(&arc_anon->arcs_size) +
2942 	    refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
2943 
2944 	/*
2945 	 * If we're below arc_meta_min, always prefer to evict data.
2946 	 * Otherwise, try to satisfy the requested number of bytes to
2947 	 * evict from the type which contains older buffers; in an
2948 	 * effort to keep newer buffers in the cache regardless of their
2949 	 * type. If we cannot satisfy the number of bytes from this
2950 	 * type, spill over into the next type.
2951 	 */
2952 	if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
2953 	    arc_meta_used > arc_meta_min) {
2954 		bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
2955 		total_evicted += bytes;
2956 
2957 		/*
2958 		 * If we couldn't evict our target number of bytes from
2959 		 * metadata, we try to get the rest from data.
2960 		 */
2961 		target -= bytes;
2962 
2963 		total_evicted +=
2964 		    arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
2965 	} else {
2966 		bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
2967 		total_evicted += bytes;
2968 
2969 		/*
2970 		 * If we couldn't evict our target number of bytes from
2971 		 * data, we try to get the rest from metadata.
2972 		 */
2973 		target -= bytes;
2974 
2975 		total_evicted +=
2976 		    arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
2977 	}
2978 
2979 	/*
2980 	 * Adjust MFU size
2981 	 *
2982 	 * Now that we've tried to evict enough from the MRU to get its
2983 	 * size back to arc_p, if we're still above the target cache
2984 	 * size, we evict the rest from the MFU.
2985 	 */
2986 	target = arc_size - arc_c;
2987 
2988 	if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
2989 	    arc_meta_used > arc_meta_min) {
2990 		bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
2991 		total_evicted += bytes;
2992 
2993 		/*
2994 		 * If we couldn't evict our target number of bytes from
2995 		 * metadata, we try to get the rest from data.
2996 		 */
2997 		target -= bytes;
2998 
2999 		total_evicted +=
3000 		    arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3001 	} else {
3002 		bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3003 		total_evicted += bytes;
3004 
3005 		/*
3006 		 * If we couldn't evict our target number of bytes from
3007 		 * data, we try to get the rest from data.
3008 		 */
3009 		target -= bytes;
3010 
3011 		total_evicted +=
3012 		    arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3013 	}
3014 
3015 	/*
3016 	 * Adjust ghost lists
3017 	 *
3018 	 * In addition to the above, the ARC also defines target values
3019 	 * for the ghost lists. The sum of the mru list and mru ghost
3020 	 * list should never exceed the target size of the cache, and
3021 	 * the sum of the mru list, mfu list, mru ghost list, and mfu
3022 	 * ghost list should never exceed twice the target size of the
3023 	 * cache. The following logic enforces these limits on the ghost
3024 	 * caches, and evicts from them as needed.
3025 	 */
3026 	target = refcount_count(&arc_mru->arcs_size) +
3027 	    refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3028 
3029 	bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3030 	total_evicted += bytes;
3031 
3032 	target -= bytes;
3033 
3034 	total_evicted +=
3035 	    arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3036 
3037 	/*
3038 	 * We assume the sum of the mru list and mfu list is less than
3039 	 * or equal to arc_c (we enforced this above), which means we
3040 	 * can use the simpler of the two equations below:
3041 	 *
3042 	 *	mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3043 	 *		    mru ghost + mfu ghost <= arc_c
3044 	 */
3045 	target = refcount_count(&arc_mru_ghost->arcs_size) +
3046 	    refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3047 
3048 	bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3049 	total_evicted += bytes;
3050 
3051 	target -= bytes;
3052 
3053 	total_evicted +=
3054 	    arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3055 
3056 	return (total_evicted);
3057 }
3058 
3059 static void
3060 arc_do_user_evicts(void)
3061 {
3062 	mutex_enter(&arc_user_evicts_lock);
3063 	while (arc_eviction_list != NULL) {
3064 		arc_buf_t *buf = arc_eviction_list;
3065 		arc_eviction_list = buf->b_next;
3066 		mutex_enter(&buf->b_evict_lock);
3067 		buf->b_hdr = NULL;
3068 		mutex_exit(&buf->b_evict_lock);
3069 		mutex_exit(&arc_user_evicts_lock);
3070 
3071 		if (buf->b_efunc != NULL)
3072 			VERIFY0(buf->b_efunc(buf->b_private));
3073 
3074 		buf->b_efunc = NULL;
3075 		buf->b_private = NULL;
3076 		kmem_cache_free(buf_cache, buf);
3077 		mutex_enter(&arc_user_evicts_lock);
3078 	}
3079 	mutex_exit(&arc_user_evicts_lock);
3080 }
3081 
3082 void
3083 arc_flush(spa_t *spa, boolean_t retry)
3084 {
3085 	uint64_t guid = 0;
3086 
3087 	/*
3088 	 * If retry is TRUE, a spa must not be specified since we have
3089 	 * no good way to determine if all of a spa's buffers have been
3090 	 * evicted from an arc state.
3091 	 */
3092 	ASSERT(!retry || spa == 0);
3093 
3094 	if (spa != NULL)
3095 		guid = spa_load_guid(spa);
3096 
3097 	(void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3098 	(void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3099 
3100 	(void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3101 	(void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3102 
3103 	(void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3104 	(void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3105 
3106 	(void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3107 	(void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3108 
3109 	arc_do_user_evicts();
3110 	ASSERT(spa || arc_eviction_list == NULL);
3111 }
3112 
3113 void
3114 arc_shrink(int64_t to_free)
3115 {
3116 	if (arc_c > arc_c_min) {
3117 
3118 		if (arc_c > arc_c_min + to_free)
3119 			atomic_add_64(&arc_c, -to_free);
3120 		else
3121 			arc_c = arc_c_min;
3122 
3123 		atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3124 		if (arc_c > arc_size)
3125 			arc_c = MAX(arc_size, arc_c_min);
3126 		if (arc_p > arc_c)
3127 			arc_p = (arc_c >> 1);
3128 		ASSERT(arc_c >= arc_c_min);
3129 		ASSERT((int64_t)arc_p >= 0);
3130 	}
3131 
3132 	if (arc_size > arc_c)
3133 		(void) arc_adjust();
3134 }
3135 
3136 typedef enum free_memory_reason_t {
3137 	FMR_UNKNOWN,
3138 	FMR_NEEDFREE,
3139 	FMR_LOTSFREE,
3140 	FMR_SWAPFS_MINFREE,
3141 	FMR_PAGES_PP_MAXIMUM,
3142 	FMR_HEAP_ARENA,
3143 	FMR_ZIO_ARENA,
3144 } free_memory_reason_t;
3145 
3146 int64_t last_free_memory;
3147 free_memory_reason_t last_free_reason;
3148 
3149 /*
3150  * Additional reserve of pages for pp_reserve.
3151  */
3152 int64_t arc_pages_pp_reserve = 64;
3153 
3154 /*
3155  * Additional reserve of pages for swapfs.
3156  */
3157 int64_t arc_swapfs_reserve = 64;
3158 
3159 /*
3160  * Return the amount of memory that can be consumed before reclaim will be
3161  * needed.  Positive if there is sufficient free memory, negative indicates
3162  * the amount of memory that needs to be freed up.
3163  */
3164 static int64_t
3165 arc_available_memory(void)
3166 {
3167 	int64_t lowest = INT64_MAX;
3168 	int64_t n;
3169 	free_memory_reason_t r = FMR_UNKNOWN;
3170 
3171 #ifdef _KERNEL
3172 	if (needfree > 0) {
3173 		n = PAGESIZE * (-needfree);
3174 		if (n < lowest) {
3175 			lowest = n;
3176 			r = FMR_NEEDFREE;
3177 		}
3178 	}
3179 
3180 	/*
3181 	 * check that we're out of range of the pageout scanner.  It starts to
3182 	 * schedule paging if freemem is less than lotsfree and needfree.
3183 	 * lotsfree is the high-water mark for pageout, and needfree is the
3184 	 * number of needed free pages.  We add extra pages here to make sure
3185 	 * the scanner doesn't start up while we're freeing memory.
3186 	 */
3187 	n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
3188 	if (n < lowest) {
3189 		lowest = n;
3190 		r = FMR_LOTSFREE;
3191 	}
3192 
3193 	/*
3194 	 * check to make sure that swapfs has enough space so that anon
3195 	 * reservations can still succeed. anon_resvmem() checks that the
3196 	 * availrmem is greater than swapfs_minfree, and the number of reserved
3197 	 * swap pages.  We also add a bit of extra here just to prevent
3198 	 * circumstances from getting really dire.
3199 	 */
3200 	n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
3201 	    desfree - arc_swapfs_reserve);
3202 	if (n < lowest) {
3203 		lowest = n;
3204 		r = FMR_SWAPFS_MINFREE;
3205 	}
3206 
3207 
3208 	/*
3209 	 * Check that we have enough availrmem that memory locking (e.g., via
3210 	 * mlock(3C) or memcntl(2)) can still succeed.  (pages_pp_maximum
3211 	 * stores the number of pages that cannot be locked; when availrmem
3212 	 * drops below pages_pp_maximum, page locking mechanisms such as
3213 	 * page_pp_lock() will fail.)
3214 	 */
3215 	n = PAGESIZE * (availrmem - pages_pp_maximum -
3216 	    arc_pages_pp_reserve);
3217 	if (n < lowest) {
3218 		lowest = n;
3219 		r = FMR_PAGES_PP_MAXIMUM;
3220 	}
3221 
3222 #if defined(__i386)
3223 	/*
3224 	 * If we're on an i386 platform, it's possible that we'll exhaust the
3225 	 * kernel heap space before we ever run out of available physical
3226 	 * memory.  Most checks of the size of the heap_area compare against
3227 	 * tune.t_minarmem, which is the minimum available real memory that we
3228 	 * can have in the system.  However, this is generally fixed at 25 pages
3229 	 * which is so low that it's useless.  In this comparison, we seek to
3230 	 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3231 	 * heap is allocated.  (Or, in the calculation, if less than 1/4th is
3232 	 * free)
3233 	 */
3234 	n = vmem_size(heap_arena, VMEM_FREE) -
3235 	    (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
3236 	if (n < lowest) {
3237 		lowest = n;
3238 		r = FMR_HEAP_ARENA;
3239 	}
3240 #endif
3241 
3242 	/*
3243 	 * If zio data pages are being allocated out of a separate heap segment,
3244 	 * then enforce that the size of available vmem for this arena remains
3245 	 * above about 1/16th free.
3246 	 *
3247 	 * Note: The 1/16th arena free requirement was put in place
3248 	 * to aggressively evict memory from the arc in order to avoid
3249 	 * memory fragmentation issues.
3250 	 */
3251 	if (zio_arena != NULL) {
3252 		n = vmem_size(zio_arena, VMEM_FREE) -
3253 		    (vmem_size(zio_arena, VMEM_ALLOC) >> 4);
3254 		if (n < lowest) {
3255 			lowest = n;
3256 			r = FMR_ZIO_ARENA;
3257 		}
3258 	}
3259 #else
3260 	/* Every 100 calls, free a small amount */
3261 	if (spa_get_random(100) == 0)
3262 		lowest = -1024;
3263 #endif
3264 
3265 	last_free_memory = lowest;
3266 	last_free_reason = r;
3267 
3268 	return (lowest);
3269 }
3270 
3271 
3272 /*
3273  * Determine if the system is under memory pressure and is asking
3274  * to reclaim memory. A return value of TRUE indicates that the system
3275  * is under memory pressure and that the arc should adjust accordingly.
3276  */
3277 static boolean_t
3278 arc_reclaim_needed(void)
3279 {
3280 	return (arc_available_memory() < 0);
3281 }
3282 
3283 static void
3284 arc_kmem_reap_now(void)
3285 {
3286 	size_t			i;
3287 	kmem_cache_t		*prev_cache = NULL;
3288 	kmem_cache_t		*prev_data_cache = NULL;
3289 	extern kmem_cache_t	*zio_buf_cache[];
3290 	extern kmem_cache_t	*zio_data_buf_cache[];
3291 	extern kmem_cache_t	*range_seg_cache;
3292 
3293 #ifdef _KERNEL
3294 	if (arc_meta_used >= arc_meta_limit) {
3295 		/*
3296 		 * We are exceeding our meta-data cache limit.
3297 		 * Purge some DNLC entries to release holds on meta-data.
3298 		 */
3299 		dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
3300 	}
3301 #if defined(__i386)
3302 	/*
3303 	 * Reclaim unused memory from all kmem caches.
3304 	 */
3305 	kmem_reap();
3306 #endif
3307 #endif
3308 
3309 	for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
3310 		if (zio_buf_cache[i] != prev_cache) {
3311 			prev_cache = zio_buf_cache[i];
3312 			kmem_cache_reap_now(zio_buf_cache[i]);
3313 		}
3314 		if (zio_data_buf_cache[i] != prev_data_cache) {
3315 			prev_data_cache = zio_data_buf_cache[i];
3316 			kmem_cache_reap_now(zio_data_buf_cache[i]);
3317 		}
3318 	}
3319 	kmem_cache_reap_now(buf_cache);
3320 	kmem_cache_reap_now(hdr_full_cache);
3321 	kmem_cache_reap_now(hdr_l2only_cache);
3322 	kmem_cache_reap_now(range_seg_cache);
3323 
3324 	if (zio_arena != NULL) {
3325 		/*
3326 		 * Ask the vmem arena to reclaim unused memory from its
3327 		 * quantum caches.
3328 		 */
3329 		vmem_qcache_reap(zio_arena);
3330 	}
3331 }
3332 
3333 /*
3334  * Threads can block in arc_get_data_buf() waiting for this thread to evict
3335  * enough data and signal them to proceed. When this happens, the threads in
3336  * arc_get_data_buf() are sleeping while holding the hash lock for their
3337  * particular arc header. Thus, we must be careful to never sleep on a
3338  * hash lock in this thread. This is to prevent the following deadlock:
3339  *
3340  *  - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
3341  *    waiting for the reclaim thread to signal it.
3342  *
3343  *  - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
3344  *    fails, and goes to sleep forever.
3345  *
3346  * This possible deadlock is avoided by always acquiring a hash lock
3347  * using mutex_tryenter() from arc_reclaim_thread().
3348  */
3349 static void
3350 arc_reclaim_thread(void)
3351 {
3352 	clock_t			growtime = 0;
3353 	callb_cpr_t		cpr;
3354 
3355 	CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
3356 
3357 	mutex_enter(&arc_reclaim_lock);
3358 	while (!arc_reclaim_thread_exit) {
3359 		int64_t free_memory = arc_available_memory();
3360 		uint64_t evicted = 0;
3361 
3362 		mutex_exit(&arc_reclaim_lock);
3363 
3364 		if (free_memory < 0) {
3365 
3366 			arc_no_grow = B_TRUE;
3367 			arc_warm = B_TRUE;
3368 
3369 			/*
3370 			 * Wait at least zfs_grow_retry (default 60) seconds
3371 			 * before considering growing.
3372 			 */
3373 			growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
3374 
3375 			arc_kmem_reap_now();
3376 
3377 			/*
3378 			 * If we are still low on memory, shrink the ARC
3379 			 * so that we have arc_shrink_min free space.
3380 			 */
3381 			free_memory = arc_available_memory();
3382 
3383 			int64_t to_free =
3384 			    (arc_c >> arc_shrink_shift) - free_memory;
3385 			if (to_free > 0) {
3386 #ifdef _KERNEL
3387 				to_free = MAX(to_free, ptob(needfree));
3388 #endif
3389 				arc_shrink(to_free);
3390 			}
3391 		} else if (free_memory < arc_c >> arc_no_grow_shift) {
3392 			arc_no_grow = B_TRUE;
3393 		} else if (ddi_get_lbolt() >= growtime) {
3394 			arc_no_grow = B_FALSE;
3395 		}
3396 
3397 		evicted = arc_adjust();
3398 
3399 		mutex_enter(&arc_reclaim_lock);
3400 
3401 		/*
3402 		 * If evicted is zero, we couldn't evict anything via
3403 		 * arc_adjust(). This could be due to hash lock
3404 		 * collisions, but more likely due to the majority of
3405 		 * arc buffers being unevictable. Therefore, even if
3406 		 * arc_size is above arc_c, another pass is unlikely to
3407 		 * be helpful and could potentially cause us to enter an
3408 		 * infinite loop.
3409 		 */
3410 		if (arc_size <= arc_c || evicted == 0) {
3411 			/*
3412 			 * We're either no longer overflowing, or we
3413 			 * can't evict anything more, so we should wake
3414 			 * up any threads before we go to sleep.
3415 			 */
3416 			cv_broadcast(&arc_reclaim_waiters_cv);
3417 
3418 			/*
3419 			 * Block until signaled, or after one second (we
3420 			 * might need to perform arc_kmem_reap_now()
3421 			 * even if we aren't being signalled)
3422 			 */
3423 			CALLB_CPR_SAFE_BEGIN(&cpr);
3424 			(void) cv_timedwait(&arc_reclaim_thread_cv,
3425 			    &arc_reclaim_lock, ddi_get_lbolt() + hz);
3426 			CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
3427 		}
3428 	}
3429 
3430 	arc_reclaim_thread_exit = FALSE;
3431 	cv_broadcast(&arc_reclaim_thread_cv);
3432 	CALLB_CPR_EXIT(&cpr);		/* drops arc_reclaim_lock */
3433 	thread_exit();
3434 }
3435 
3436 static void
3437 arc_user_evicts_thread(void)
3438 {
3439 	callb_cpr_t cpr;
3440 
3441 	CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG);
3442 
3443 	mutex_enter(&arc_user_evicts_lock);
3444 	while (!arc_user_evicts_thread_exit) {
3445 		mutex_exit(&arc_user_evicts_lock);
3446 
3447 		arc_do_user_evicts();
3448 
3449 		/*
3450 		 * This is necessary in order for the mdb ::arc dcmd to
3451 		 * show up to date information. Since the ::arc command
3452 		 * does not call the kstat's update function, without
3453 		 * this call, the command may show stale stats for the
3454 		 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
3455 		 * with this change, the data might be up to 1 second
3456 		 * out of date; but that should suffice. The arc_state_t
3457 		 * structures can be queried directly if more accurate
3458 		 * information is needed.
3459 		 */
3460 		if (arc_ksp != NULL)
3461 			arc_ksp->ks_update(arc_ksp, KSTAT_READ);
3462 
3463 		mutex_enter(&arc_user_evicts_lock);
3464 
3465 		/*
3466 		 * Block until signaled, or after one second (we need to
3467 		 * call the arc's kstat update function regularly).
3468 		 */
3469 		CALLB_CPR_SAFE_BEGIN(&cpr);
3470 		(void) cv_timedwait(&arc_user_evicts_cv,
3471 		    &arc_user_evicts_lock, ddi_get_lbolt() + hz);
3472 		CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock);
3473 	}
3474 
3475 	arc_user_evicts_thread_exit = FALSE;
3476 	cv_broadcast(&arc_user_evicts_cv);
3477 	CALLB_CPR_EXIT(&cpr);		/* drops arc_user_evicts_lock */
3478 	thread_exit();
3479 }
3480 
3481 /*
3482  * Adapt arc info given the number of bytes we are trying to add and
3483  * the state that we are comming from.  This function is only called
3484  * when we are adding new content to the cache.
3485  */
3486 static void
3487 arc_adapt(int bytes, arc_state_t *state)
3488 {
3489 	int mult;
3490 	uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
3491 	int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
3492 	int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
3493 
3494 	if (state == arc_l2c_only)
3495 		return;
3496 
3497 	ASSERT(bytes > 0);
3498 	/*
3499 	 * Adapt the target size of the MRU list:
3500 	 *	- if we just hit in the MRU ghost list, then increase
3501 	 *	  the target size of the MRU list.
3502 	 *	- if we just hit in the MFU ghost list, then increase
3503 	 *	  the target size of the MFU list by decreasing the
3504 	 *	  target size of the MRU list.
3505 	 */
3506 	if (state == arc_mru_ghost) {
3507 		mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
3508 		mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
3509 
3510 		arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
3511 	} else if (state == arc_mfu_ghost) {
3512 		uint64_t delta;
3513 
3514 		mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
3515 		mult = MIN(mult, 10);
3516 
3517 		delta = MIN(bytes * mult, arc_p);
3518 		arc_p = MAX(arc_p_min, arc_p - delta);
3519 	}
3520 	ASSERT((int64_t)arc_p >= 0);
3521 
3522 	if (arc_reclaim_needed()) {
3523 		cv_signal(&arc_reclaim_thread_cv);
3524 		return;
3525 	}
3526 
3527 	if (arc_no_grow)
3528 		return;
3529 
3530 	if (arc_c >= arc_c_max)
3531 		return;
3532 
3533 	/*
3534 	 * If we're within (2 * maxblocksize) bytes of the target
3535 	 * cache size, increment the target cache size
3536 	 */
3537 	if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
3538 		atomic_add_64(&arc_c, (int64_t)bytes);
3539 		if (arc_c > arc_c_max)
3540 			arc_c = arc_c_max;
3541 		else if (state == arc_anon)
3542 			atomic_add_64(&arc_p, (int64_t)bytes);
3543 		if (arc_p > arc_c)
3544 			arc_p = arc_c;
3545 	}
3546 	ASSERT((int64_t)arc_p >= 0);
3547 }
3548 
3549 /*
3550  * Check if arc_size has grown past our upper threshold, determined by
3551  * zfs_arc_overflow_shift.
3552  */
3553 static boolean_t
3554 arc_is_overflowing(void)
3555 {
3556 	/* Always allow at least one block of overflow */
3557 	uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
3558 	    arc_c >> zfs_arc_overflow_shift);
3559 
3560 	return (arc_size >= arc_c + overflow);
3561 }
3562 
3563 /*
3564  * The buffer, supplied as the first argument, needs a data block. If we
3565  * are hitting the hard limit for the cache size, we must sleep, waiting
3566  * for the eviction thread to catch up. If we're past the target size
3567  * but below the hard limit, we'll only signal the reclaim thread and
3568  * continue on.
3569  */
3570 static void
3571 arc_get_data_buf(arc_buf_t *buf)
3572 {
3573 	arc_state_t		*state = buf->b_hdr->b_l1hdr.b_state;
3574 	uint64_t		size = buf->b_hdr->b_size;
3575 	arc_buf_contents_t	type = arc_buf_type(buf->b_hdr);
3576 
3577 	arc_adapt(size, state);
3578 
3579 	/*
3580 	 * If arc_size is currently overflowing, and has grown past our
3581 	 * upper limit, we must be adding data faster than the evict
3582 	 * thread can evict. Thus, to ensure we don't compound the
3583 	 * problem by adding more data and forcing arc_size to grow even
3584 	 * further past it's target size, we halt and wait for the
3585 	 * eviction thread to catch up.
3586 	 *
3587 	 * It's also possible that the reclaim thread is unable to evict
3588 	 * enough buffers to get arc_size below the overflow limit (e.g.
3589 	 * due to buffers being un-evictable, or hash lock collisions).
3590 	 * In this case, we want to proceed regardless if we're
3591 	 * overflowing; thus we don't use a while loop here.
3592 	 */
3593 	if (arc_is_overflowing()) {
3594 		mutex_enter(&arc_reclaim_lock);
3595 
3596 		/*
3597 		 * Now that we've acquired the lock, we may no longer be
3598 		 * over the overflow limit, lets check.
3599 		 *
3600 		 * We're ignoring the case of spurious wake ups. If that
3601 		 * were to happen, it'd let this thread consume an ARC
3602 		 * buffer before it should have (i.e. before we're under
3603 		 * the overflow limit and were signalled by the reclaim
3604 		 * thread). As long as that is a rare occurrence, it
3605 		 * shouldn't cause any harm.
3606 		 */
3607 		if (arc_is_overflowing()) {
3608 			cv_signal(&arc_reclaim_thread_cv);
3609 			cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
3610 		}
3611 
3612 		mutex_exit(&arc_reclaim_lock);
3613 	}
3614 
3615 	if (type == ARC_BUFC_METADATA) {
3616 		buf->b_data = zio_buf_alloc(size);
3617 		arc_space_consume(size, ARC_SPACE_META);
3618 	} else {
3619 		ASSERT(type == ARC_BUFC_DATA);
3620 		buf->b_data = zio_data_buf_alloc(size);
3621 		arc_space_consume(size, ARC_SPACE_DATA);
3622 	}
3623 
3624 	/*
3625 	 * Update the state size.  Note that ghost states have a
3626 	 * "ghost size" and so don't need to be updated.
3627 	 */
3628 	if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) {
3629 		arc_buf_hdr_t *hdr = buf->b_hdr;
3630 		arc_state_t *state = hdr->b_l1hdr.b_state;
3631 
3632 		(void) refcount_add_many(&state->arcs_size, size, buf);
3633 
3634 		/*
3635 		 * If this is reached via arc_read, the link is
3636 		 * protected by the hash lock. If reached via
3637 		 * arc_buf_alloc, the header should not be accessed by
3638 		 * any other thread. And, if reached via arc_read_done,
3639 		 * the hash lock will protect it if it's found in the
3640 		 * hash table; otherwise no other thread should be
3641 		 * trying to [add|remove]_reference it.
3642 		 */
3643 		if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3644 			ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3645 			atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type],
3646 			    size);
3647 		}
3648 		/*
3649 		 * If we are growing the cache, and we are adding anonymous
3650 		 * data, and we have outgrown arc_p, update arc_p
3651 		 */
3652 		if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
3653 		    (refcount_count(&arc_anon->arcs_size) +
3654 		    refcount_count(&arc_mru->arcs_size) > arc_p))
3655 			arc_p = MIN(arc_c, arc_p + size);
3656 	}
3657 }
3658 
3659 /*
3660  * This routine is called whenever a buffer is accessed.
3661  * NOTE: the hash lock is dropped in this function.
3662  */
3663 static void
3664 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3665 {
3666 	clock_t now;
3667 
3668 	ASSERT(MUTEX_HELD(hash_lock));
3669 	ASSERT(HDR_HAS_L1HDR(hdr));
3670 
3671 	if (hdr->b_l1hdr.b_state == arc_anon) {
3672 		/*
3673 		 * This buffer is not in the cache, and does not
3674 		 * appear in our "ghost" list.  Add the new buffer
3675 		 * to the MRU state.
3676 		 */
3677 
3678 		ASSERT0(hdr->b_l1hdr.b_arc_access);
3679 		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3680 		DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3681 		arc_change_state(arc_mru, hdr, hash_lock);
3682 
3683 	} else if (hdr->b_l1hdr.b_state == arc_mru) {
3684 		now = ddi_get_lbolt();
3685 
3686 		/*
3687 		 * If this buffer is here because of a prefetch, then either:
3688 		 * - clear the flag if this is a "referencing" read
3689 		 *   (any subsequent access will bump this into the MFU state).
3690 		 * or
3691 		 * - move the buffer to the head of the list if this is
3692 		 *   another prefetch (to make it less likely to be evicted).
3693 		 */
3694 		if (HDR_PREFETCH(hdr)) {
3695 			if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
3696 				/* link protected by hash lock */
3697 				ASSERT(multilist_link_active(
3698 				    &hdr->b_l1hdr.b_arc_node));
3699 			} else {
3700 				hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3701 				ARCSTAT_BUMP(arcstat_mru_hits);
3702 			}
3703 			hdr->b_l1hdr.b_arc_access = now;
3704 			return;
3705 		}
3706 
3707 		/*
3708 		 * This buffer has been "accessed" only once so far,
3709 		 * but it is still in the cache. Move it to the MFU
3710 		 * state.
3711 		 */
3712 		if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
3713 			/*
3714 			 * More than 125ms have passed since we
3715 			 * instantiated this buffer.  Move it to the
3716 			 * most frequently used state.
3717 			 */
3718 			hdr->b_l1hdr.b_arc_access = now;
3719 			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3720 			arc_change_state(arc_mfu, hdr, hash_lock);
3721 		}
3722 		ARCSTAT_BUMP(arcstat_mru_hits);
3723 	} else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
3724 		arc_state_t	*new_state;
3725 		/*
3726 		 * This buffer has been "accessed" recently, but
3727 		 * was evicted from the cache.  Move it to the
3728 		 * MFU state.
3729 		 */
3730 
3731 		if (HDR_PREFETCH(hdr)) {
3732 			new_state = arc_mru;
3733 			if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
3734 				hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3735 			DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3736 		} else {
3737 			new_state = arc_mfu;
3738 			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3739 		}
3740 
3741 		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3742 		arc_change_state(new_state, hdr, hash_lock);
3743 
3744 		ARCSTAT_BUMP(arcstat_mru_ghost_hits);
3745 	} else if (hdr->b_l1hdr.b_state == arc_mfu) {
3746 		/*
3747 		 * This buffer has been accessed more than once and is
3748 		 * still in the cache.  Keep it in the MFU state.
3749 		 *
3750 		 * NOTE: an add_reference() that occurred when we did
3751 		 * the arc_read() will have kicked this off the list.
3752 		 * If it was a prefetch, we will explicitly move it to
3753 		 * the head of the list now.
3754 		 */
3755 		if ((HDR_PREFETCH(hdr)) != 0) {
3756 			ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3757 			/* link protected by hash_lock */
3758 			ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3759 		}
3760 		ARCSTAT_BUMP(arcstat_mfu_hits);
3761 		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3762 	} else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
3763 		arc_state_t	*new_state = arc_mfu;
3764 		/*
3765 		 * This buffer has been accessed more than once but has
3766 		 * been evicted from the cache.  Move it back to the
3767 		 * MFU state.
3768 		 */
3769 
3770 		if (HDR_PREFETCH(hdr)) {
3771 			/*
3772 			 * This is a prefetch access...
3773 			 * move this block back to the MRU state.
3774 			 */
3775 			ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3776 			new_state = arc_mru;
3777 		}
3778 
3779 		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3780 		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3781 		arc_change_state(new_state, hdr, hash_lock);
3782 
3783 		ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
3784 	} else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
3785 		/*
3786 		 * This buffer is on the 2nd Level ARC.
3787 		 */
3788 
3789 		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3790 		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3791 		arc_change_state(arc_mfu, hdr, hash_lock);
3792 	} else {
3793 		ASSERT(!"invalid arc state");
3794 	}
3795 }
3796 
3797 /* a generic arc_done_func_t which you can use */
3798 /* ARGSUSED */
3799 void
3800 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
3801 {
3802 	if (zio == NULL || zio->io_error == 0)
3803 		bcopy(buf->b_data, arg, buf->b_hdr->b_size);
3804 	VERIFY(arc_buf_remove_ref(buf, arg));
3805 }
3806 
3807 /* a generic arc_done_func_t */
3808 void
3809 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
3810 {
3811 	arc_buf_t **bufp = arg;
3812 	if (zio && zio->io_error) {
3813 		VERIFY(arc_buf_remove_ref(buf, arg));
3814 		*bufp = NULL;
3815 	} else {
3816 		*bufp = buf;
3817 		ASSERT(buf->b_data);
3818 	}
3819 }
3820 
3821 static void
3822 arc_read_done(zio_t *zio)
3823 {
3824 	arc_buf_hdr_t	*hdr;
3825 	arc_buf_t	*buf;
3826 	arc_buf_t	*abuf;	/* buffer we're assigning to callback */
3827 	kmutex_t	*hash_lock = NULL;
3828 	arc_callback_t	*callback_list, *acb;
3829 	int		freeable = FALSE;
3830 
3831 	buf = zio->io_private;
3832 	hdr = buf->b_hdr;
3833 
3834 	/*
3835 	 * The hdr was inserted into hash-table and removed from lists
3836 	 * prior to starting I/O.  We should find this header, since
3837 	 * it's in the hash table, and it should be legit since it's
3838 	 * not possible to evict it during the I/O.  The only possible
3839 	 * reason for it not to be found is if we were freed during the
3840 	 * read.
3841 	 */
3842 	if (HDR_IN_HASH_TABLE(hdr)) {
3843 		ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
3844 		ASSERT3U(hdr->b_dva.dva_word[0], ==,
3845 		    BP_IDENTITY(zio->io_bp)->dva_word[0]);
3846 		ASSERT3U(hdr->b_dva.dva_word[1], ==,
3847 		    BP_IDENTITY(zio->io_bp)->dva_word[1]);
3848 
3849 		arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
3850 		    &hash_lock);
3851 
3852 		ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) &&
3853 		    hash_lock == NULL) ||
3854 		    (found == hdr &&
3855 		    DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
3856 		    (found == hdr && HDR_L2_READING(hdr)));
3857 	}
3858 
3859 	hdr->b_flags &= ~ARC_FLAG_L2_EVICTED;
3860 	if (l2arc_noprefetch && HDR_PREFETCH(hdr))
3861 		hdr->b_flags &= ~ARC_FLAG_L2CACHE;
3862 
3863 	/* byteswap if necessary */
3864 	callback_list = hdr->b_l1hdr.b_acb;
3865 	ASSERT(callback_list != NULL);
3866 	if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
3867 		dmu_object_byteswap_t bswap =
3868 		    DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
3869 		arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
3870 		    byteswap_uint64_array :
3871 		    dmu_ot_byteswap[bswap].ob_func;
3872 		func(buf->b_data, hdr->b_size);
3873 	}
3874 
3875 	arc_cksum_compute(buf, B_FALSE);
3876 	arc_buf_watch(buf);
3877 
3878 	if (hash_lock && zio->io_error == 0 &&
3879 	    hdr->b_l1hdr.b_state == arc_anon) {
3880 		/*
3881 		 * Only call arc_access on anonymous buffers.  This is because
3882 		 * if we've issued an I/O for an evicted buffer, we've already
3883 		 * called arc_access (to prevent any simultaneous readers from
3884 		 * getting confused).
3885 		 */
3886 		arc_access(hdr, hash_lock);
3887 	}
3888 
3889 	/* create copies of the data buffer for the callers */
3890 	abuf = buf;
3891 	for (acb = callback_list; acb; acb = acb->acb_next) {
3892 		if (acb->acb_done) {
3893 			if (abuf == NULL) {
3894 				ARCSTAT_BUMP(arcstat_duplicate_reads);
3895 				abuf = arc_buf_clone(buf);
3896 			}
3897 			acb->acb_buf = abuf;
3898 			abuf = NULL;
3899 		}
3900 	}
3901 	hdr->b_l1hdr.b_acb = NULL;
3902 	hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
3903 	ASSERT(!HDR_BUF_AVAILABLE(hdr));
3904 	if (abuf == buf) {
3905 		ASSERT(buf->b_efunc == NULL);
3906 		ASSERT(hdr->b_l1hdr.b_datacnt == 1);
3907 		hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
3908 	}
3909 
3910 	ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
3911 	    callback_list != NULL);
3912 
3913 	if (zio->io_error != 0) {
3914 		hdr->b_flags |= ARC_FLAG_IO_ERROR;
3915 		if (hdr->b_l1hdr.b_state != arc_anon)
3916 			arc_change_state(arc_anon, hdr, hash_lock);
3917 		if (HDR_IN_HASH_TABLE(hdr))
3918 			buf_hash_remove(hdr);
3919 		freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
3920 	}
3921 
3922 	/*
3923 	 * Broadcast before we drop the hash_lock to avoid the possibility
3924 	 * that the hdr (and hence the cv) might be freed before we get to
3925 	 * the cv_broadcast().
3926 	 */
3927 	cv_broadcast(&hdr->b_l1hdr.b_cv);
3928 
3929 	if (hash_lock != NULL) {
3930 		mutex_exit(hash_lock);
3931 	} else {
3932 		/*
3933 		 * This block was freed while we waited for the read to
3934 		 * complete.  It has been removed from the hash table and
3935 		 * moved to the anonymous state (so that it won't show up
3936 		 * in the cache).
3937 		 */
3938 		ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3939 		freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
3940 	}
3941 
3942 	/* execute each callback and free its structure */
3943 	while ((acb = callback_list) != NULL) {
3944 		if (acb->acb_done)
3945 			acb->acb_done(zio, acb->acb_buf, acb->acb_private);
3946 
3947 		if (acb->acb_zio_dummy != NULL) {
3948 			acb->acb_zio_dummy->io_error = zio->io_error;
3949 			zio_nowait(acb->acb_zio_dummy);
3950 		}
3951 
3952 		callback_list = acb->acb_next;
3953 		kmem_free(acb, sizeof (arc_callback_t));
3954 	}
3955 
3956 	if (freeable)
3957 		arc_hdr_destroy(hdr);
3958 }
3959 
3960 /*
3961  * "Read" the block at the specified DVA (in bp) via the
3962  * cache.  If the block is found in the cache, invoke the provided
3963  * callback immediately and return.  Note that the `zio' parameter
3964  * in the callback will be NULL in this case, since no IO was
3965  * required.  If the block is not in the cache pass the read request
3966  * on to the spa with a substitute callback function, so that the
3967  * requested block will be added to the cache.
3968  *
3969  * If a read request arrives for a block that has a read in-progress,
3970  * either wait for the in-progress read to complete (and return the
3971  * results); or, if this is a read with a "done" func, add a record
3972  * to the read to invoke the "done" func when the read completes,
3973  * and return; or just return.
3974  *
3975  * arc_read_done() will invoke all the requested "done" functions
3976  * for readers of this block.
3977  */
3978 int
3979 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
3980     void *private, zio_priority_t priority, int zio_flags,
3981     arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
3982 {
3983 	arc_buf_hdr_t *hdr = NULL;
3984 	arc_buf_t *buf = NULL;
3985 	kmutex_t *hash_lock = NULL;
3986 	zio_t *rzio;
3987 	uint64_t guid = spa_load_guid(spa);
3988 
3989 	ASSERT(!BP_IS_EMBEDDED(bp) ||
3990 	    BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
3991 
3992 top:
3993 	if (!BP_IS_EMBEDDED(bp)) {
3994 		/*
3995 		 * Embedded BP's have no DVA and require no I/O to "read".
3996 		 * Create an anonymous arc buf to back it.
3997 		 */
3998 		hdr = buf_hash_find(guid, bp, &hash_lock);
3999 	}
4000 
4001 	if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) {
4002 
4003 		*arc_flags |= ARC_FLAG_CACHED;
4004 
4005 		if (HDR_IO_IN_PROGRESS(hdr)) {
4006 
4007 			if (*arc_flags & ARC_FLAG_WAIT) {
4008 				cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
4009 				mutex_exit(hash_lock);
4010 				goto top;
4011 			}
4012 			ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4013 
4014 			if (done) {
4015 				arc_callback_t	*acb = NULL;
4016 
4017 				acb = kmem_zalloc(sizeof (arc_callback_t),
4018 				    KM_SLEEP);
4019 				acb->acb_done = done;
4020 				acb->acb_private = private;
4021 				if (pio != NULL)
4022 					acb->acb_zio_dummy = zio_null(pio,
4023 					    spa, NULL, NULL, NULL, zio_flags);
4024 
4025 				ASSERT(acb->acb_done != NULL);
4026 				acb->acb_next = hdr->b_l1hdr.b_acb;
4027 				hdr->b_l1hdr.b_acb = acb;
4028 				add_reference(hdr, hash_lock, private);
4029 				mutex_exit(hash_lock);
4030 				return (0);
4031 			}
4032 			mutex_exit(hash_lock);
4033 			return (0);
4034 		}
4035 
4036 		ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4037 		    hdr->b_l1hdr.b_state == arc_mfu);
4038 
4039 		if (done) {
4040 			add_reference(hdr, hash_lock, private);
4041 			/*
4042 			 * If this block is already in use, create a new
4043 			 * copy of the data so that we will be guaranteed
4044 			 * that arc_release() will always succeed.
4045 			 */
4046 			buf = hdr->b_l1hdr.b_buf;
4047 			ASSERT(buf);
4048 			ASSERT(buf->b_data);
4049 			if (HDR_BUF_AVAILABLE(hdr)) {
4050 				ASSERT(buf->b_efunc == NULL);
4051 				hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4052 			} else {
4053 				buf = arc_buf_clone(buf);
4054 			}
4055 
4056 		} else if (*arc_flags & ARC_FLAG_PREFETCH &&
4057 		    refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4058 			hdr->b_flags |= ARC_FLAG_PREFETCH;
4059 		}
4060 		DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4061 		arc_access(hdr, hash_lock);
4062 		if (*arc_flags & ARC_FLAG_L2CACHE)
4063 			hdr->b_flags |= ARC_FLAG_L2CACHE;
4064 		if (*arc_flags & ARC_FLAG_L2COMPRESS)
4065 			hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4066 		mutex_exit(hash_lock);
4067 		ARCSTAT_BUMP(arcstat_hits);
4068 		ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4069 		    demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4070 		    data, metadata, hits);
4071 
4072 		if (done)
4073 			done(NULL, buf, private);
4074 	} else {
4075 		uint64_t size = BP_GET_LSIZE(bp);
4076 		arc_callback_t *acb;
4077 		vdev_t *vd = NULL;
4078 		uint64_t addr = 0;
4079 		boolean_t devw = B_FALSE;
4080 		enum zio_compress b_compress = ZIO_COMPRESS_OFF;
4081 		int32_t b_asize = 0;
4082 
4083 		if (hdr == NULL) {
4084 			/* this block is not in the cache */
4085 			arc_buf_hdr_t *exists = NULL;
4086 			arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
4087 			buf = arc_buf_alloc(spa, size, private, type);
4088 			hdr = buf->b_hdr;
4089 			if (!BP_IS_EMBEDDED(bp)) {
4090 				hdr->b_dva = *BP_IDENTITY(bp);
4091 				hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
4092 				exists = buf_hash_insert(hdr, &hash_lock);
4093 			}
4094 			if (exists != NULL) {
4095 				/* somebody beat us to the hash insert */
4096 				mutex_exit(hash_lock);
4097 				buf_discard_identity(hdr);
4098 				(void) arc_buf_remove_ref(buf, private);
4099 				goto top; /* restart the IO request */
4100 			}
4101 
4102 			/* if this is a prefetch, we don't have a reference */
4103 			if (*arc_flags & ARC_FLAG_PREFETCH) {
4104 				(void) remove_reference(hdr, hash_lock,
4105 				    private);
4106 				hdr->b_flags |= ARC_FLAG_PREFETCH;
4107 			}
4108 			if (*arc_flags & ARC_FLAG_L2CACHE)
4109 				hdr->b_flags |= ARC_FLAG_L2CACHE;
4110 			if (*arc_flags & ARC_FLAG_L2COMPRESS)
4111 				hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4112 			if (BP_GET_LEVEL(bp) > 0)
4113 				hdr->b_flags |= ARC_FLAG_INDIRECT;
4114 		} else {
4115 			/*
4116 			 * This block is in the ghost cache. If it was L2-only
4117 			 * (and thus didn't have an L1 hdr), we realloc the
4118 			 * header to add an L1 hdr.
4119 			 */
4120 			if (!HDR_HAS_L1HDR(hdr)) {
4121 				hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
4122 				    hdr_full_cache);
4123 			}
4124 
4125 			ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
4126 			ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4127 			ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4128 			ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
4129 
4130 			/* if this is a prefetch, we don't have a reference */
4131 			if (*arc_flags & ARC_FLAG_PREFETCH)
4132 				hdr->b_flags |= ARC_FLAG_PREFETCH;
4133 			else
4134 				add_reference(hdr, hash_lock, private);
4135 			if (*arc_flags & ARC_FLAG_L2CACHE)
4136 				hdr->b_flags |= ARC_FLAG_L2CACHE;
4137 			if (*arc_flags & ARC_FLAG_L2COMPRESS)
4138 				hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4139 			buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
4140 			buf->b_hdr = hdr;
4141 			buf->b_data = NULL;
4142 			buf->b_efunc = NULL;
4143 			buf->b_private = NULL;
4144 			buf->b_next = NULL;
4145 			hdr->b_l1hdr.b_buf = buf;
4146 			ASSERT0(hdr->b_l1hdr.b_datacnt);
4147 			hdr->b_l1hdr.b_datacnt = 1;
4148 			arc_get_data_buf(buf);
4149 			arc_access(hdr, hash_lock);
4150 		}
4151 
4152 		ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
4153 
4154 		acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
4155 		acb->acb_done = done;
4156 		acb->acb_private = private;
4157 
4158 		ASSERT(hdr->b_l1hdr.b_acb == NULL);
4159 		hdr->b_l1hdr.b_acb = acb;
4160 		hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4161 
4162 		if (HDR_HAS_L2HDR(hdr) &&
4163 		    (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
4164 			devw = hdr->b_l2hdr.b_dev->l2ad_writing;
4165 			addr = hdr->b_l2hdr.b_daddr;
4166 			b_compress = HDR_GET_COMPRESS(hdr);
4167 			b_asize = hdr->b_l2hdr.b_asize;
4168 			/*
4169 			 * Lock out device removal.
4170 			 */
4171 			if (vdev_is_dead(vd) ||
4172 			    !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
4173 				vd = NULL;
4174 		}
4175 
4176 		if (hash_lock != NULL)
4177 			mutex_exit(hash_lock);
4178 
4179 		/*
4180 		 * At this point, we have a level 1 cache miss.  Try again in
4181 		 * L2ARC if possible.
4182 		 */
4183 		ASSERT3U(hdr->b_size, ==, size);
4184 		DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
4185 		    uint64_t, size, zbookmark_phys_t *, zb);
4186 		ARCSTAT_BUMP(arcstat_misses);
4187 		ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4188 		    demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4189 		    data, metadata, misses);
4190 
4191 		if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
4192 			/*
4193 			 * Read from the L2ARC if the following are true:
4194 			 * 1. The L2ARC vdev was previously cached.
4195 			 * 2. This buffer still has L2ARC metadata.
4196 			 * 3. This buffer isn't currently writing to the L2ARC.
4197 			 * 4. The L2ARC entry wasn't evicted, which may
4198 			 *    also have invalidated the vdev.
4199 			 * 5. This isn't prefetch and l2arc_noprefetch is set.
4200 			 */
4201 			if (HDR_HAS_L2HDR(hdr) &&
4202 			    !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
4203 			    !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
4204 				l2arc_read_callback_t *cb;
4205 
4206 				DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
4207 				ARCSTAT_BUMP(arcstat_l2_hits);
4208 
4209 				cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
4210 				    KM_SLEEP);
4211 				cb->l2rcb_buf = buf;
4212 				cb->l2rcb_spa = spa;
4213 				cb->l2rcb_bp = *bp;
4214 				cb->l2rcb_zb = *zb;
4215 				cb->l2rcb_flags = zio_flags;
4216 				cb->l2rcb_compress = b_compress;
4217 
4218 				ASSERT(addr >= VDEV_LABEL_START_SIZE &&
4219 				    addr + size < vd->vdev_psize -
4220 				    VDEV_LABEL_END_SIZE);
4221 
4222 				/*
4223 				 * l2arc read.  The SCL_L2ARC lock will be
4224 				 * released by l2arc_read_done().
4225 				 * Issue a null zio if the underlying buffer
4226 				 * was squashed to zero size by compression.
4227 				 */
4228 				if (b_compress == ZIO_COMPRESS_EMPTY) {
4229 					rzio = zio_null(pio, spa, vd,
4230 					    l2arc_read_done, cb,
4231 					    zio_flags | ZIO_FLAG_DONT_CACHE |
4232 					    ZIO_FLAG_CANFAIL |
4233 					    ZIO_FLAG_DONT_PROPAGATE |
4234 					    ZIO_FLAG_DONT_RETRY);
4235 				} else {
4236 					rzio = zio_read_phys(pio, vd, addr,
4237 					    b_asize, buf->b_data,
4238 					    ZIO_CHECKSUM_OFF,
4239 					    l2arc_read_done, cb, priority,
4240 					    zio_flags | ZIO_FLAG_DONT_CACHE |
4241 					    ZIO_FLAG_CANFAIL |
4242 					    ZIO_FLAG_DONT_PROPAGATE |
4243 					    ZIO_FLAG_DONT_RETRY, B_FALSE);
4244 				}
4245 				DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
4246 				    zio_t *, rzio);
4247 				ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize);
4248 
4249 				if (*arc_flags & ARC_FLAG_NOWAIT) {
4250 					zio_nowait(rzio);
4251 					return (0);
4252 				}
4253 
4254 				ASSERT(*arc_flags & ARC_FLAG_WAIT);
4255 				if (zio_wait(rzio) == 0)
4256 					return (0);
4257 
4258 				/* l2arc read error; goto zio_read() */
4259 			} else {
4260 				DTRACE_PROBE1(l2arc__miss,
4261 				    arc_buf_hdr_t *, hdr);
4262 				ARCSTAT_BUMP(arcstat_l2_misses);
4263 				if (HDR_L2_WRITING(hdr))
4264 					ARCSTAT_BUMP(arcstat_l2_rw_clash);
4265 				spa_config_exit(spa, SCL_L2ARC, vd);
4266 			}
4267 		} else {
4268 			if (vd != NULL)
4269 				spa_config_exit(spa, SCL_L2ARC, vd);
4270 			if (l2arc_ndev != 0) {
4271 				DTRACE_PROBE1(l2arc__miss,
4272 				    arc_buf_hdr_t *, hdr);
4273 				ARCSTAT_BUMP(arcstat_l2_misses);
4274 			}
4275 		}
4276 
4277 		rzio = zio_read(pio, spa, bp, buf->b_data, size,
4278 		    arc_read_done, buf, priority, zio_flags, zb);
4279 
4280 		if (*arc_flags & ARC_FLAG_WAIT)
4281 			return (zio_wait(rzio));
4282 
4283 		ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4284 		zio_nowait(rzio);
4285 	}
4286 	return (0);
4287 }
4288 
4289 void
4290 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
4291 {
4292 	ASSERT(buf->b_hdr != NULL);
4293 	ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon);
4294 	ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) ||
4295 	    func == NULL);
4296 	ASSERT(buf->b_efunc == NULL);
4297 	ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
4298 
4299 	buf->b_efunc = func;
4300 	buf->b_private = private;
4301 }
4302 
4303 /*
4304  * Notify the arc that a block was freed, and thus will never be used again.
4305  */
4306 void
4307 arc_freed(spa_t *spa, const blkptr_t *bp)
4308 {
4309 	arc_buf_hdr_t *hdr;
4310 	kmutex_t *hash_lock;
4311 	uint64_t guid = spa_load_guid(spa);
4312 
4313 	ASSERT(!BP_IS_EMBEDDED(bp));
4314 
4315 	hdr = buf_hash_find(guid, bp, &hash_lock);
4316 	if (hdr == NULL)
4317 		return;
4318 	if (HDR_BUF_AVAILABLE(hdr)) {
4319 		arc_buf_t *buf = hdr->b_l1hdr.b_buf;
4320 		add_reference(hdr, hash_lock, FTAG);
4321 		hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4322 		mutex_exit(hash_lock);
4323 
4324 		arc_release(buf, FTAG);
4325 		(void) arc_buf_remove_ref(buf, FTAG);
4326 	} else {
4327 		mutex_exit(hash_lock);
4328 	}
4329 
4330 }
4331 
4332 /*
4333  * Clear the user eviction callback set by arc_set_callback(), first calling
4334  * it if it exists.  Because the presence of a callback keeps an arc_buf cached
4335  * clearing the callback may result in the arc_buf being destroyed.  However,
4336  * it will not result in the *last* arc_buf being destroyed, hence the data
4337  * will remain cached in the ARC. We make a copy of the arc buffer here so
4338  * that we can process the callback without holding any locks.
4339  *
4340  * It's possible that the callback is already in the process of being cleared
4341  * by another thread.  In this case we can not clear the callback.
4342  *
4343  * Returns B_TRUE if the callback was successfully called and cleared.
4344  */
4345 boolean_t
4346 arc_clear_callback(arc_buf_t *buf)
4347 {
4348 	arc_buf_hdr_t *hdr;
4349 	kmutex_t *hash_lock;
4350 	arc_evict_func_t *efunc = buf->b_efunc;
4351 	void *private = buf->b_private;
4352 
4353 	mutex_enter(&buf->b_evict_lock);
4354 	hdr = buf->b_hdr;
4355 	if (hdr == NULL) {
4356 		/*
4357 		 * We are in arc_do_user_evicts().
4358 		 */
4359 		ASSERT(buf->b_data == NULL);
4360 		mutex_exit(&buf->b_evict_lock);
4361 		return (B_FALSE);
4362 	} else if (buf->b_data == NULL) {
4363 		/*
4364 		 * We are on the eviction list; process this buffer now
4365 		 * but let arc_do_user_evicts() do the reaping.
4366 		 */
4367 		buf->b_efunc = NULL;
4368 		mutex_exit(&buf->b_evict_lock);
4369 		VERIFY0(efunc(private));
4370 		return (B_TRUE);
4371 	}
4372 	hash_lock = HDR_LOCK(hdr);
4373 	mutex_enter(hash_lock);
4374 	hdr = buf->b_hdr;
4375 	ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4376 
4377 	ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <,
4378 	    hdr->b_l1hdr.b_datacnt);
4379 	ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4380 	    hdr->b_l1hdr.b_state == arc_mfu);
4381 
4382 	buf->b_efunc = NULL;
4383 	buf->b_private = NULL;
4384 
4385 	if (hdr->b_l1hdr.b_datacnt > 1) {
4386 		mutex_exit(&buf->b_evict_lock);
4387 		arc_buf_destroy(buf, TRUE);
4388 	} else {
4389 		ASSERT(buf == hdr->b_l1hdr.b_buf);
4390 		hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
4391 		mutex_exit(&buf->b_evict_lock);
4392 	}
4393 
4394 	mutex_exit(hash_lock);
4395 	VERIFY0(efunc(private));
4396 	return (B_TRUE);
4397 }
4398 
4399 /*
4400  * Release this buffer from the cache, making it an anonymous buffer.  This
4401  * must be done after a read and prior to modifying the buffer contents.
4402  * If the buffer has more than one reference, we must make
4403  * a new hdr for the buffer.
4404  */
4405 void
4406 arc_release(arc_buf_t *buf, void *tag)
4407 {
4408 	arc_buf_hdr_t *hdr = buf->b_hdr;
4409 
4410 	/*
4411 	 * It would be nice to assert that if it's DMU metadata (level >
4412 	 * 0 || it's the dnode file), then it must be syncing context.
4413 	 * But we don't know that information at this level.
4414 	 */
4415 
4416 	mutex_enter(&buf->b_evict_lock);
4417 
4418 	ASSERT(HDR_HAS_L1HDR(hdr));
4419 
4420 	/*
4421 	 * We don't grab the hash lock prior to this check, because if
4422 	 * the buffer's header is in the arc_anon state, it won't be
4423 	 * linked into the hash table.
4424 	 */
4425 	if (hdr->b_l1hdr.b_state == arc_anon) {
4426 		mutex_exit(&buf->b_evict_lock);
4427 		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4428 		ASSERT(!HDR_IN_HASH_TABLE(hdr));
4429 		ASSERT(!HDR_HAS_L2HDR(hdr));
4430 		ASSERT(BUF_EMPTY(hdr));
4431 
4432 		ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1);
4433 		ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
4434 		ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
4435 
4436 		ASSERT3P(buf->b_efunc, ==, NULL);
4437 		ASSERT3P(buf->b_private, ==, NULL);
4438 
4439 		hdr->b_l1hdr.b_arc_access = 0;
4440 		arc_buf_thaw(buf);
4441 
4442 		return;
4443 	}
4444 
4445 	kmutex_t *hash_lock = HDR_LOCK(hdr);
4446 	mutex_enter(hash_lock);
4447 
4448 	/*
4449 	 * This assignment is only valid as long as the hash_lock is
4450 	 * held, we must be careful not to reference state or the
4451 	 * b_state field after dropping the lock.
4452 	 */
4453 	arc_state_t *state = hdr->b_l1hdr.b_state;
4454 	ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4455 	ASSERT3P(state, !=, arc_anon);
4456 
4457 	/* this buffer is not on any list */
4458 	ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0);
4459 
4460 	if (HDR_HAS_L2HDR(hdr)) {
4461 		mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4462 
4463 		/*
4464 		 * We have to recheck this conditional again now that
4465 		 * we're holding the l2ad_mtx to prevent a race with
4466 		 * another thread which might be concurrently calling
4467 		 * l2arc_evict(). In that case, l2arc_evict() might have
4468 		 * destroyed the header's L2 portion as we were waiting
4469 		 * to acquire the l2ad_mtx.
4470 		 */
4471 		if (HDR_HAS_L2HDR(hdr))
4472 			arc_hdr_l2hdr_destroy(hdr);
4473 
4474 		mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4475 	}
4476 
4477 	/*
4478 	 * Do we have more than one buf?
4479 	 */
4480 	if (hdr->b_l1hdr.b_datacnt > 1) {
4481 		arc_buf_hdr_t *nhdr;
4482 		arc_buf_t **bufp;
4483 		uint64_t blksz = hdr->b_size;
4484 		uint64_t spa = hdr->b_spa;
4485 		arc_buf_contents_t type = arc_buf_type(hdr);
4486 		uint32_t flags = hdr->b_flags;
4487 
4488 		ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
4489 		/*
4490 		 * Pull the data off of this hdr and attach it to
4491 		 * a new anonymous hdr.
4492 		 */
4493 		(void) remove_reference(hdr, hash_lock, tag);
4494 		bufp = &hdr->b_l1hdr.b_buf;
4495 		while (*bufp != buf)
4496 			bufp = &(*bufp)->b_next;
4497 		*bufp = buf->b_next;
4498 		buf->b_next = NULL;
4499 
4500 		ASSERT3P(state, !=, arc_l2c_only);
4501 
4502 		(void) refcount_remove_many(
4503 		    &state->arcs_size, hdr->b_size, buf);
4504 
4505 		if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
4506 			ASSERT3P(state, !=, arc_l2c_only);
4507 			uint64_t *size = &state->arcs_lsize[type];
4508 			ASSERT3U(*size, >=, hdr->b_size);
4509 			atomic_add_64(size, -hdr->b_size);
4510 		}
4511 
4512 		/*
4513 		 * We're releasing a duplicate user data buffer, update
4514 		 * our statistics accordingly.
4515 		 */
4516 		if (HDR_ISTYPE_DATA(hdr)) {
4517 			ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
4518 			ARCSTAT_INCR(arcstat_duplicate_buffers_size,
4519 			    -hdr->b_size);
4520 		}
4521 		hdr->b_l1hdr.b_datacnt -= 1;
4522 		arc_cksum_verify(buf);
4523 		arc_buf_unwatch(buf);
4524 
4525 		mutex_exit(hash_lock);
4526 
4527 		nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
4528 		nhdr->b_size = blksz;
4529 		nhdr->b_spa = spa;
4530 
4531 		nhdr->b_flags = flags & ARC_FLAG_L2_WRITING;
4532 		nhdr->b_flags |= arc_bufc_to_flags(type);
4533 		nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
4534 
4535 		nhdr->b_l1hdr.b_buf = buf;
4536 		nhdr->b_l1hdr.b_datacnt = 1;
4537 		nhdr->b_l1hdr.b_state = arc_anon;
4538 		nhdr->b_l1hdr.b_arc_access = 0;
4539 		nhdr->b_l1hdr.b_tmp_cdata = NULL;
4540 		nhdr->b_freeze_cksum = NULL;
4541 
4542 		(void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
4543 		buf->b_hdr = nhdr;
4544 		mutex_exit(&buf->b_evict_lock);
4545 		(void) refcount_add_many(&arc_anon->arcs_size, blksz, buf);
4546 	} else {
4547 		mutex_exit(&buf->b_evict_lock);
4548 		ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
4549 		/* protected by hash lock, or hdr is on arc_anon */
4550 		ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4551 		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4552 		arc_change_state(arc_anon, hdr, hash_lock);
4553 		hdr->b_l1hdr.b_arc_access = 0;
4554 		mutex_exit(hash_lock);
4555 
4556 		buf_discard_identity(hdr);
4557 		arc_buf_thaw(buf);
4558 	}
4559 	buf->b_efunc = NULL;
4560 	buf->b_private = NULL;
4561 }
4562 
4563 int
4564 arc_released(arc_buf_t *buf)
4565 {
4566 	int released;
4567 
4568 	mutex_enter(&buf->b_evict_lock);
4569 	released = (buf->b_data != NULL &&
4570 	    buf->b_hdr->b_l1hdr.b_state == arc_anon);
4571 	mutex_exit(&buf->b_evict_lock);
4572 	return (released);
4573 }
4574 
4575 #ifdef ZFS_DEBUG
4576 int
4577 arc_referenced(arc_buf_t *buf)
4578 {
4579 	int referenced;
4580 
4581 	mutex_enter(&buf->b_evict_lock);
4582 	referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
4583 	mutex_exit(&buf->b_evict_lock);
4584 	return (referenced);
4585 }
4586 #endif
4587 
4588 static void
4589 arc_write_ready(zio_t *zio)
4590 {
4591 	arc_write_callback_t *callback = zio->io_private;
4592 	arc_buf_t *buf = callback->awcb_buf;
4593 	arc_buf_hdr_t *hdr = buf->b_hdr;
4594 
4595 	ASSERT(HDR_HAS_L1HDR(hdr));
4596 	ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
4597 	ASSERT(hdr->b_l1hdr.b_datacnt > 0);
4598 	callback->awcb_ready(zio, buf, callback->awcb_private);
4599 
4600 	/*
4601 	 * If the IO is already in progress, then this is a re-write
4602 	 * attempt, so we need to thaw and re-compute the cksum.
4603 	 * It is the responsibility of the callback to handle the
4604 	 * accounting for any re-write attempt.
4605 	 */
4606 	if (HDR_IO_IN_PROGRESS(hdr)) {
4607 		mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
4608 		if (hdr->b_freeze_cksum != NULL) {
4609 			kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
4610 			hdr->b_freeze_cksum = NULL;
4611 		}
4612 		mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
4613 	}
4614 	arc_cksum_compute(buf, B_FALSE);
4615 	hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4616 }
4617 
4618 /*
4619  * The SPA calls this callback for each physical write that happens on behalf
4620  * of a logical write.  See the comment in dbuf_write_physdone() for details.
4621  */
4622 static void
4623 arc_write_physdone(zio_t *zio)
4624 {
4625 	arc_write_callback_t *cb = zio->io_private;
4626 	if (cb->awcb_physdone != NULL)
4627 		cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
4628 }
4629 
4630 static void
4631 arc_write_done(zio_t *zio)
4632 {
4633 	arc_write_callback_t *callback = zio->io_private;
4634 	arc_buf_t *buf = callback->awcb_buf;
4635 	arc_buf_hdr_t *hdr = buf->b_hdr;
4636 
4637 	ASSERT(hdr->b_l1hdr.b_acb == NULL);
4638 
4639 	if (zio->io_error == 0) {
4640 		if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
4641 			buf_discard_identity(hdr);
4642 		} else {
4643 			hdr->b_dva = *BP_IDENTITY(zio->io_bp);
4644 			hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
4645 		}
4646 	} else {
4647 		ASSERT(BUF_EMPTY(hdr));
4648 	}
4649 
4650 	/*
4651 	 * If the block to be written was all-zero or compressed enough to be
4652 	 * embedded in the BP, no write was performed so there will be no
4653 	 * dva/birth/checksum.  The buffer must therefore remain anonymous
4654 	 * (and uncached).
4655 	 */
4656 	if (!BUF_EMPTY(hdr)) {
4657 		arc_buf_hdr_t *exists;
4658 		kmutex_t *hash_lock;
4659 
4660 		ASSERT(zio->io_error == 0);
4661 
4662 		arc_cksum_verify(buf);
4663 
4664 		exists = buf_hash_insert(hdr, &hash_lock);
4665 		if (exists != NULL) {
4666 			/*
4667 			 * This can only happen if we overwrite for
4668 			 * sync-to-convergence, because we remove
4669 			 * buffers from the hash table when we arc_free().
4670 			 */
4671 			if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
4672 				if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
4673 					panic("bad overwrite, hdr=%p exists=%p",
4674 					    (void *)hdr, (void *)exists);
4675 				ASSERT(refcount_is_zero(
4676 				    &exists->b_l1hdr.b_refcnt));
4677 				arc_change_state(arc_anon, exists, hash_lock);
4678 				mutex_exit(hash_lock);
4679 				arc_hdr_destroy(exists);
4680 				exists = buf_hash_insert(hdr, &hash_lock);
4681 				ASSERT3P(exists, ==, NULL);
4682 			} else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
4683 				/* nopwrite */
4684 				ASSERT(zio->io_prop.zp_nopwrite);
4685 				if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
4686 					panic("bad nopwrite, hdr=%p exists=%p",
4687 					    (void *)hdr, (void *)exists);
4688 			} else {
4689 				/* Dedup */
4690 				ASSERT(hdr->b_l1hdr.b_datacnt == 1);
4691 				ASSERT(hdr->b_l1hdr.b_state == arc_anon);
4692 				ASSERT(BP_GET_DEDUP(zio->io_bp));
4693 				ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
4694 			}
4695 		}
4696 		hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4697 		/* if it's not anon, we are doing a scrub */
4698 		if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
4699 			arc_access(hdr, hash_lock);
4700 		mutex_exit(hash_lock);
4701 	} else {
4702 		hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4703 	}
4704 
4705 	ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4706 	callback->awcb_done(zio, buf, callback->awcb_private);
4707 
4708 	kmem_free(callback, sizeof (arc_write_callback_t));
4709 }
4710 
4711 zio_t *
4712 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
4713     blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
4714     const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
4715     arc_done_func_t *done, void *private, zio_priority_t priority,
4716     int zio_flags, const zbookmark_phys_t *zb)
4717 {
4718 	arc_buf_hdr_t *hdr = buf->b_hdr;
4719 	arc_write_callback_t *callback;
4720 	zio_t *zio;
4721 
4722 	ASSERT(ready != NULL);
4723 	ASSERT(done != NULL);
4724 	ASSERT(!HDR_IO_ERROR(hdr));
4725 	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4726 	ASSERT(hdr->b_l1hdr.b_acb == NULL);
4727 	ASSERT(hdr->b_l1hdr.b_datacnt > 0);
4728 	if (l2arc)
4729 		hdr->b_flags |= ARC_FLAG_L2CACHE;
4730 	if (l2arc_compress)
4731 		hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4732 	callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
4733 	callback->awcb_ready = ready;
4734 	callback->awcb_physdone = physdone;
4735 	callback->awcb_done = done;
4736 	callback->awcb_private = private;
4737 	callback->awcb_buf = buf;
4738 
4739 	zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
4740 	    arc_write_ready, arc_write_physdone, arc_write_done, callback,
4741 	    priority, zio_flags, zb);
4742 
4743 	return (zio);
4744 }
4745 
4746 static int
4747 arc_memory_throttle(uint64_t reserve, uint64_t txg)
4748 {
4749 #ifdef _KERNEL
4750 	uint64_t available_memory = ptob(freemem);
4751 	static uint64_t page_load = 0;
4752 	static uint64_t last_txg = 0;
4753 
4754 #if defined(__i386)
4755 	available_memory =
4756 	    MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
4757 #endif
4758 
4759 	if (freemem > physmem * arc_lotsfree_percent / 100)
4760 		return (0);
4761 
4762 	if (txg > last_txg) {
4763 		last_txg = txg;
4764 		page_load = 0;
4765 	}
4766 	/*
4767 	 * If we are in pageout, we know that memory is already tight,
4768 	 * the arc is already going to be evicting, so we just want to
4769 	 * continue to let page writes occur as quickly as possible.
4770 	 */
4771 	if (curproc == proc_pageout) {
4772 		if (page_load > MAX(ptob(minfree), available_memory) / 4)
4773 			return (SET_ERROR(ERESTART));
4774 		/* Note: reserve is inflated, so we deflate */
4775 		page_load += reserve / 8;
4776 		return (0);
4777 	} else if (page_load > 0 && arc_reclaim_needed()) {
4778 		/* memory is low, delay before restarting */
4779 		ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
4780 		return (SET_ERROR(EAGAIN));
4781 	}
4782 	page_load = 0;
4783 #endif
4784 	return (0);
4785 }
4786 
4787 void
4788 arc_tempreserve_clear(uint64_t reserve)
4789 {
4790 	atomic_add_64(&arc_tempreserve, -reserve);
4791 	ASSERT((int64_t)arc_tempreserve >= 0);
4792 }
4793 
4794 int
4795 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
4796 {
4797 	int error;
4798 	uint64_t anon_size;
4799 
4800 	if (reserve > arc_c/4 && !arc_no_grow)
4801 		arc_c = MIN(arc_c_max, reserve * 4);
4802 	if (reserve > arc_c)
4803 		return (SET_ERROR(ENOMEM));
4804 
4805 	/*
4806 	 * Don't count loaned bufs as in flight dirty data to prevent long
4807 	 * network delays from blocking transactions that are ready to be
4808 	 * assigned to a txg.
4809 	 */
4810 	anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
4811 	    arc_loaned_bytes), 0);
4812 
4813 	/*
4814 	 * Writes will, almost always, require additional memory allocations
4815 	 * in order to compress/encrypt/etc the data.  We therefore need to
4816 	 * make sure that there is sufficient available memory for this.
4817 	 */
4818 	error = arc_memory_throttle(reserve, txg);
4819 	if (error != 0)
4820 		return (error);
4821 
4822 	/*
4823 	 * Throttle writes when the amount of dirty data in the cache
4824 	 * gets too large.  We try to keep the cache less than half full
4825 	 * of dirty blocks so that our sync times don't grow too large.
4826 	 * Note: if two requests come in concurrently, we might let them
4827 	 * both succeed, when one of them should fail.  Not a huge deal.
4828 	 */
4829 
4830 	if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
4831 	    anon_size > arc_c / 4) {
4832 		dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
4833 		    "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
4834 		    arc_tempreserve>>10,
4835 		    arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
4836 		    arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
4837 		    reserve>>10, arc_c>>10);
4838 		return (SET_ERROR(ERESTART));
4839 	}
4840 	atomic_add_64(&arc_tempreserve, reserve);
4841 	return (0);
4842 }
4843 
4844 static void
4845 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
4846     kstat_named_t *evict_data, kstat_named_t *evict_metadata)
4847 {
4848 	size->value.ui64 = refcount_count(&state->arcs_size);
4849 	evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
4850 	evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
4851 }
4852 
4853 static int
4854 arc_kstat_update(kstat_t *ksp, int rw)
4855 {
4856 	arc_stats_t *as = ksp->ks_data;
4857 
4858 	if (rw == KSTAT_WRITE) {
4859 		return (EACCES);
4860 	} else {
4861 		arc_kstat_update_state(arc_anon,
4862 		    &as->arcstat_anon_size,
4863 		    &as->arcstat_anon_evictable_data,
4864 		    &as->arcstat_anon_evictable_metadata);
4865 		arc_kstat_update_state(arc_mru,
4866 		    &as->arcstat_mru_size,
4867 		    &as->arcstat_mru_evictable_data,
4868 		    &as->arcstat_mru_evictable_metadata);
4869 		arc_kstat_update_state(arc_mru_ghost,
4870 		    &as->arcstat_mru_ghost_size,
4871 		    &as->arcstat_mru_ghost_evictable_data,
4872 		    &as->arcstat_mru_ghost_evictable_metadata);
4873 		arc_kstat_update_state(arc_mfu,
4874 		    &as->arcstat_mfu_size,
4875 		    &as->arcstat_mfu_evictable_data,
4876 		    &as->arcstat_mfu_evictable_metadata);
4877 		arc_kstat_update_state(arc_mfu_ghost,
4878 		    &as->arcstat_mfu_ghost_size,
4879 		    &as->arcstat_mfu_ghost_evictable_data,
4880 		    &as->arcstat_mfu_ghost_evictable_metadata);
4881 	}
4882 
4883 	return (0);
4884 }
4885 
4886 /*
4887  * This function *must* return indices evenly distributed between all
4888  * sublists of the multilist. This is needed due to how the ARC eviction
4889  * code is laid out; arc_evict_state() assumes ARC buffers are evenly
4890  * distributed between all sublists and uses this assumption when
4891  * deciding which sublist to evict from and how much to evict from it.
4892  */
4893 unsigned int
4894 arc_state_multilist_index_func(multilist_t *ml, void *obj)
4895 {
4896 	arc_buf_hdr_t *hdr = obj;
4897 
4898 	/*
4899 	 * We rely on b_dva to generate evenly distributed index
4900 	 * numbers using buf_hash below. So, as an added precaution,
4901 	 * let's make sure we never add empty buffers to the arc lists.
4902 	 */
4903 	ASSERT(!BUF_EMPTY(hdr));
4904 
4905 	/*
4906 	 * The assumption here, is the hash value for a given
4907 	 * arc_buf_hdr_t will remain constant throughout it's lifetime
4908 	 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
4909 	 * Thus, we don't need to store the header's sublist index
4910 	 * on insertion, as this index can be recalculated on removal.
4911 	 *
4912 	 * Also, the low order bits of the hash value are thought to be
4913 	 * distributed evenly. Otherwise, in the case that the multilist
4914 	 * has a power of two number of sublists, each sublists' usage
4915 	 * would not be evenly distributed.
4916 	 */
4917 	return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
4918 	    multilist_get_num_sublists(ml));
4919 }
4920 
4921 void
4922 arc_init(void)
4923 {
4924 	/*
4925 	 * allmem is "all memory that we could possibly use".
4926 	 */
4927 #ifdef _KERNEL
4928 	uint64_t allmem = ptob(physmem - swapfs_minfree);
4929 #else
4930 	uint64_t allmem = (physmem * PAGESIZE) / 2;
4931 #endif
4932 
4933 	mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
4934 	cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
4935 	cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
4936 
4937 	mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
4938 	cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL);
4939 
4940 	/* Convert seconds to clock ticks */
4941 	arc_min_prefetch_lifespan = 1 * hz;
4942 
4943 	/* Start out with 1/8 of all memory */
4944 	arc_c = allmem / 8;
4945 
4946 #ifdef _KERNEL
4947 	/*
4948 	 * On architectures where the physical memory can be larger
4949 	 * than the addressable space (intel in 32-bit mode), we may
4950 	 * need to limit the cache to 1/8 of VM size.
4951 	 */
4952 	arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
4953 #endif
4954 
4955 	/* set min cache to 1/32 of all memory, or 64MB, whichever is more */
4956 	arc_c_min = MAX(allmem / 32, 64 << 20);
4957 	/* set max to 3/4 of all memory, or all but 1GB, whichever is more */
4958 	if (allmem >= 1 << 30)
4959 		arc_c_max = allmem - (1 << 30);
4960 	else
4961 		arc_c_max = arc_c_min;
4962 	arc_c_max = MAX(allmem * 3 / 4, arc_c_max);
4963 
4964 	/*
4965 	 * Allow the tunables to override our calculations if they are
4966 	 * reasonable (ie. over 64MB)
4967 	 */
4968 	if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem)
4969 		arc_c_max = zfs_arc_max;
4970 	if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max)
4971 		arc_c_min = zfs_arc_min;
4972 
4973 	arc_c = arc_c_max;
4974 	arc_p = (arc_c >> 1);
4975 
4976 	/* limit meta-data to 1/4 of the arc capacity */
4977 	arc_meta_limit = arc_c_max / 4;
4978 
4979 	/* Allow the tunable to override if it is reasonable */
4980 	if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
4981 		arc_meta_limit = zfs_arc_meta_limit;
4982 
4983 	if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
4984 		arc_c_min = arc_meta_limit / 2;
4985 
4986 	if (zfs_arc_meta_min > 0) {
4987 		arc_meta_min = zfs_arc_meta_min;
4988 	} else {
4989 		arc_meta_min = arc_c_min / 2;
4990 	}
4991 
4992 	if (zfs_arc_grow_retry > 0)
4993 		arc_grow_retry = zfs_arc_grow_retry;
4994 
4995 	if (zfs_arc_shrink_shift > 0)
4996 		arc_shrink_shift = zfs_arc_shrink_shift;
4997 
4998 	/*
4999 	 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
5000 	 */
5001 	if (arc_no_grow_shift >= arc_shrink_shift)
5002 		arc_no_grow_shift = arc_shrink_shift - 1;
5003 
5004 	if (zfs_arc_p_min_shift > 0)
5005 		arc_p_min_shift = zfs_arc_p_min_shift;
5006 
5007 	if (zfs_arc_num_sublists_per_state < 1)
5008 		zfs_arc_num_sublists_per_state = MAX(boot_ncpus, 1);
5009 
5010 	/* if kmem_flags are set, lets try to use less memory */
5011 	if (kmem_debugging())
5012 		arc_c = arc_c / 2;
5013 	if (arc_c < arc_c_min)
5014 		arc_c = arc_c_min;
5015 
5016 	arc_anon = &ARC_anon;
5017 	arc_mru = &ARC_mru;
5018 	arc_mru_ghost = &ARC_mru_ghost;
5019 	arc_mfu = &ARC_mfu;
5020 	arc_mfu_ghost = &ARC_mfu_ghost;
5021 	arc_l2c_only = &ARC_l2c_only;
5022 	arc_size = 0;
5023 
5024 	multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
5025 	    sizeof (arc_buf_hdr_t),
5026 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5027 	    zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5028 	multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
5029 	    sizeof (arc_buf_hdr_t),
5030 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5031 	    zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5032 	multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
5033 	    sizeof (arc_buf_hdr_t),
5034 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5035 	    zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5036 	multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
5037 	    sizeof (arc_buf_hdr_t),
5038 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5039 	    zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5040 	multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
5041 	    sizeof (arc_buf_hdr_t),
5042 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5043 	    zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5044 	multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
5045 	    sizeof (arc_buf_hdr_t),
5046 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5047 	    zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5048 	multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
5049 	    sizeof (arc_buf_hdr_t),
5050 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5051 	    zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5052 	multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
5053 	    sizeof (arc_buf_hdr_t),
5054 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5055 	    zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5056 	multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
5057 	    sizeof (arc_buf_hdr_t),
5058 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5059 	    zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5060 	multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
5061 	    sizeof (arc_buf_hdr_t),
5062 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5063 	    zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5064 
5065 	refcount_create(&arc_anon->arcs_size);
5066 	refcount_create(&arc_mru->arcs_size);
5067 	refcount_create(&arc_mru_ghost->arcs_size);
5068 	refcount_create(&arc_mfu->arcs_size);
5069 	refcount_create(&arc_mfu_ghost->arcs_size);
5070 	refcount_create(&arc_l2c_only->arcs_size);
5071 
5072 	buf_init();
5073 
5074 	arc_reclaim_thread_exit = FALSE;
5075 	arc_user_evicts_thread_exit = FALSE;
5076 	arc_eviction_list = NULL;
5077 	bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
5078 
5079 	arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
5080 	    sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
5081 
5082 	if (arc_ksp != NULL) {
5083 		arc_ksp->ks_data = &arc_stats;
5084 		arc_ksp->ks_update = arc_kstat_update;
5085 		kstat_install(arc_ksp);
5086 	}
5087 
5088 	(void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
5089 	    TS_RUN, minclsyspri);
5090 
5091 	(void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0,
5092 	    TS_RUN, minclsyspri);
5093 
5094 	arc_dead = FALSE;
5095 	arc_warm = B_FALSE;
5096 
5097 	/*
5098 	 * Calculate maximum amount of dirty data per pool.
5099 	 *
5100 	 * If it has been set by /etc/system, take that.
5101 	 * Otherwise, use a percentage of physical memory defined by
5102 	 * zfs_dirty_data_max_percent (default 10%) with a cap at
5103 	 * zfs_dirty_data_max_max (default 4GB).
5104 	 */
5105 	if (zfs_dirty_data_max == 0) {
5106 		zfs_dirty_data_max = physmem * PAGESIZE *
5107 		    zfs_dirty_data_max_percent / 100;
5108 		zfs_dirty_data_max = MIN(zfs_dirty_data_max,
5109 		    zfs_dirty_data_max_max);
5110 	}
5111 }
5112 
5113 void
5114 arc_fini(void)
5115 {
5116 	mutex_enter(&arc_reclaim_lock);
5117 	arc_reclaim_thread_exit = TRUE;
5118 	/*
5119 	 * The reclaim thread will set arc_reclaim_thread_exit back to
5120 	 * FALSE when it is finished exiting; we're waiting for that.
5121 	 */
5122 	while (arc_reclaim_thread_exit) {
5123 		cv_signal(&arc_reclaim_thread_cv);
5124 		cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
5125 	}
5126 	mutex_exit(&arc_reclaim_lock);
5127 
5128 	mutex_enter(&arc_user_evicts_lock);
5129 	arc_user_evicts_thread_exit = TRUE;
5130 	/*
5131 	 * The user evicts thread will set arc_user_evicts_thread_exit
5132 	 * to FALSE when it is finished exiting; we're waiting for that.
5133 	 */
5134 	while (arc_user_evicts_thread_exit) {
5135 		cv_signal(&arc_user_evicts_cv);
5136 		cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock);
5137 	}
5138 	mutex_exit(&arc_user_evicts_lock);
5139 
5140 	/* Use TRUE to ensure *all* buffers are evicted */
5141 	arc_flush(NULL, TRUE);
5142 
5143 	arc_dead = TRUE;
5144 
5145 	if (arc_ksp != NULL) {
5146 		kstat_delete(arc_ksp);
5147 		arc_ksp = NULL;
5148 	}
5149 
5150 	mutex_destroy(&arc_reclaim_lock);
5151 	cv_destroy(&arc_reclaim_thread_cv);
5152 	cv_destroy(&arc_reclaim_waiters_cv);
5153 
5154 	mutex_destroy(&arc_user_evicts_lock);
5155 	cv_destroy(&arc_user_evicts_cv);
5156 
5157 	refcount_destroy(&arc_anon->arcs_size);
5158 	refcount_destroy(&arc_mru->arcs_size);
5159 	refcount_destroy(&arc_mru_ghost->arcs_size);
5160 	refcount_destroy(&arc_mfu->arcs_size);
5161 	refcount_destroy(&arc_mfu_ghost->arcs_size);
5162 	refcount_destroy(&arc_l2c_only->arcs_size);
5163 
5164 	multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
5165 	multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
5166 	multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
5167 	multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
5168 	multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
5169 	multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
5170 	multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
5171 	multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
5172 
5173 	buf_fini();
5174 
5175 	ASSERT0(arc_loaned_bytes);
5176 }
5177 
5178 /*
5179  * Level 2 ARC
5180  *
5181  * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
5182  * It uses dedicated storage devices to hold cached data, which are populated
5183  * using large infrequent writes.  The main role of this cache is to boost
5184  * the performance of random read workloads.  The intended L2ARC devices
5185  * include short-stroked disks, solid state disks, and other media with
5186  * substantially faster read latency than disk.
5187  *
5188  *                 +-----------------------+
5189  *                 |         ARC           |
5190  *                 +-----------------------+
5191  *                    |         ^     ^
5192  *                    |         |     |
5193  *      l2arc_feed_thread()    arc_read()
5194  *                    |         |     |
5195  *                    |  l2arc read   |
5196  *                    V         |     |
5197  *               +---------------+    |
5198  *               |     L2ARC     |    |
5199  *               +---------------+    |
5200  *                   |    ^           |
5201  *          l2arc_write() |           |
5202  *                   |    |           |
5203  *                   V    |           |
5204  *                 +-------+      +-------+
5205  *                 | vdev  |      | vdev  |
5206  *                 | cache |      | cache |
5207  *                 +-------+      +-------+
5208  *                 +=========+     .-----.
5209  *                 :  L2ARC  :    |-_____-|
5210  *                 : devices :    | Disks |
5211  *                 +=========+    `-_____-'
5212  *
5213  * Read requests are satisfied from the following sources, in order:
5214  *
5215  *	1) ARC
5216  *	2) vdev cache of L2ARC devices
5217  *	3) L2ARC devices
5218  *	4) vdev cache of disks
5219  *	5) disks
5220  *
5221  * Some L2ARC device types exhibit extremely slow write performance.
5222  * To accommodate for this there are some significant differences between
5223  * the L2ARC and traditional cache design:
5224  *
5225  * 1. There is no eviction path from the ARC to the L2ARC.  Evictions from
5226  * the ARC behave as usual, freeing buffers and placing headers on ghost
5227  * lists.  The ARC does not send buffers to the L2ARC during eviction as
5228  * this would add inflated write latencies for all ARC memory pressure.
5229  *
5230  * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
5231  * It does this by periodically scanning buffers from the eviction-end of
5232  * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
5233  * not already there. It scans until a headroom of buffers is satisfied,
5234  * which itself is a buffer for ARC eviction. If a compressible buffer is
5235  * found during scanning and selected for writing to an L2ARC device, we
5236  * temporarily boost scanning headroom during the next scan cycle to make
5237  * sure we adapt to compression effects (which might significantly reduce
5238  * the data volume we write to L2ARC). The thread that does this is
5239  * l2arc_feed_thread(), illustrated below; example sizes are included to
5240  * provide a better sense of ratio than this diagram:
5241  *
5242  *	       head -->                        tail
5243  *	        +---------------------+----------+
5244  *	ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->.   # already on L2ARC
5245  *	        +---------------------+----------+   |   o L2ARC eligible
5246  *	ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->|   : ARC buffer
5247  *	        +---------------------+----------+   |
5248  *	             15.9 Gbytes      ^ 32 Mbytes    |
5249  *	                           headroom          |
5250  *	                                      l2arc_feed_thread()
5251  *	                                             |
5252  *	                 l2arc write hand <--[oooo]--'
5253  *	                         |           8 Mbyte
5254  *	                         |          write max
5255  *	                         V
5256  *		  +==============================+
5257  *	L2ARC dev |####|#|###|###|    |####| ... |
5258  *	          +==============================+
5259  *	                     32 Gbytes
5260  *
5261  * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
5262  * evicted, then the L2ARC has cached a buffer much sooner than it probably
5263  * needed to, potentially wasting L2ARC device bandwidth and storage.  It is
5264  * safe to say that this is an uncommon case, since buffers at the end of
5265  * the ARC lists have moved there due to inactivity.
5266  *
5267  * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
5268  * then the L2ARC simply misses copying some buffers.  This serves as a
5269  * pressure valve to prevent heavy read workloads from both stalling the ARC
5270  * with waits and clogging the L2ARC with writes.  This also helps prevent
5271  * the potential for the L2ARC to churn if it attempts to cache content too
5272  * quickly, such as during backups of the entire pool.
5273  *
5274  * 5. After system boot and before the ARC has filled main memory, there are
5275  * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
5276  * lists can remain mostly static.  Instead of searching from tail of these
5277  * lists as pictured, the l2arc_feed_thread() will search from the list heads
5278  * for eligible buffers, greatly increasing its chance of finding them.
5279  *
5280  * The L2ARC device write speed is also boosted during this time so that
5281  * the L2ARC warms up faster.  Since there have been no ARC evictions yet,
5282  * there are no L2ARC reads, and no fear of degrading read performance
5283  * through increased writes.
5284  *
5285  * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
5286  * the vdev queue can aggregate them into larger and fewer writes.  Each
5287  * device is written to in a rotor fashion, sweeping writes through
5288  * available space then repeating.
5289  *
5290  * 7. The L2ARC does not store dirty content.  It never needs to flush
5291  * write buffers back to disk based storage.
5292  *
5293  * 8. If an ARC buffer is written (and dirtied) which also exists in the
5294  * L2ARC, the now stale L2ARC buffer is immediately dropped.
5295  *
5296  * The performance of the L2ARC can be tweaked by a number of tunables, which
5297  * may be necessary for different workloads:
5298  *
5299  *	l2arc_write_max		max write bytes per interval
5300  *	l2arc_write_boost	extra write bytes during device warmup
5301  *	l2arc_noprefetch	skip caching prefetched buffers
5302  *	l2arc_headroom		number of max device writes to precache
5303  *	l2arc_headroom_boost	when we find compressed buffers during ARC
5304  *				scanning, we multiply headroom by this
5305  *				percentage factor for the next scan cycle,
5306  *				since more compressed buffers are likely to
5307  *				be present
5308  *	l2arc_feed_secs		seconds between L2ARC writing
5309  *
5310  * Tunables may be removed or added as future performance improvements are
5311  * integrated, and also may become zpool properties.
5312  *
5313  * There are three key functions that control how the L2ARC warms up:
5314  *
5315  *	l2arc_write_eligible()	check if a buffer is eligible to cache
5316  *	l2arc_write_size()	calculate how much to write
5317  *	l2arc_write_interval()	calculate sleep delay between writes
5318  *
5319  * These three functions determine what to write, how much, and how quickly
5320  * to send writes.
5321  */
5322 
5323 static boolean_t
5324 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
5325 {
5326 	/*
5327 	 * A buffer is *not* eligible for the L2ARC if it:
5328 	 * 1. belongs to a different spa.
5329 	 * 2. is already cached on the L2ARC.
5330 	 * 3. has an I/O in progress (it may be an incomplete read).
5331 	 * 4. is flagged not eligible (zfs property).
5332 	 */
5333 	if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
5334 	    HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
5335 		return (B_FALSE);
5336 
5337 	return (B_TRUE);
5338 }
5339 
5340 static uint64_t
5341 l2arc_write_size(void)
5342 {
5343 	uint64_t size;
5344 
5345 	/*
5346 	 * Make sure our globals have meaningful values in case the user
5347 	 * altered them.
5348 	 */
5349 	size = l2arc_write_max;
5350 	if (size == 0) {
5351 		cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
5352 		    "be greater than zero, resetting it to the default (%d)",
5353 		    L2ARC_WRITE_SIZE);
5354 		size = l2arc_write_max = L2ARC_WRITE_SIZE;
5355 	}
5356 
5357 	if (arc_warm == B_FALSE)
5358 		size += l2arc_write_boost;
5359 
5360 	return (size);
5361 
5362 }
5363 
5364 static clock_t
5365 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
5366 {
5367 	clock_t interval, next, now;
5368 
5369 	/*
5370 	 * If the ARC lists are busy, increase our write rate; if the
5371 	 * lists are stale, idle back.  This is achieved by checking
5372 	 * how much we previously wrote - if it was more than half of
5373 	 * what we wanted, schedule the next write much sooner.
5374 	 */
5375 	if (l2arc_feed_again && wrote > (wanted / 2))
5376 		interval = (hz * l2arc_feed_min_ms) / 1000;
5377 	else
5378 		interval = hz * l2arc_feed_secs;
5379 
5380 	now = ddi_get_lbolt();
5381 	next = MAX(now, MIN(now + interval, began + interval));
5382 
5383 	return (next);
5384 }
5385 
5386 /*
5387  * Cycle through L2ARC devices.  This is how L2ARC load balances.
5388  * If a device is returned, this also returns holding the spa config lock.
5389  */
5390 static l2arc_dev_t *
5391 l2arc_dev_get_next(void)
5392 {
5393 	l2arc_dev_t *first, *next = NULL;
5394 
5395 	/*
5396 	 * Lock out the removal of spas (spa_namespace_lock), then removal
5397 	 * of cache devices (l2arc_dev_mtx).  Once a device has been selected,
5398 	 * both locks will be dropped and a spa config lock held instead.
5399 	 */
5400 	mutex_enter(&spa_namespace_lock);
5401 	mutex_enter(&l2arc_dev_mtx);
5402 
5403 	/* if there are no vdevs, there is nothing to do */
5404 	if (l2arc_ndev == 0)
5405 		goto out;
5406 
5407 	first = NULL;
5408 	next = l2arc_dev_last;
5409 	do {
5410 		/* loop around the list looking for a non-faulted vdev */
5411 		if (next == NULL) {
5412 			next = list_head(l2arc_dev_list);
5413 		} else {
5414 			next = list_next(l2arc_dev_list, next);
5415 			if (next == NULL)
5416 				next = list_head(l2arc_dev_list);
5417 		}
5418 
5419 		/* if we have come back to the start, bail out */
5420 		if (first == NULL)
5421 			first = next;
5422 		else if (next == first)
5423 			break;
5424 
5425 	} while (vdev_is_dead(next->l2ad_vdev));
5426 
5427 	/* if we were unable to find any usable vdevs, return NULL */
5428 	if (vdev_is_dead(next->l2ad_vdev))
5429 		next = NULL;
5430 
5431 	l2arc_dev_last = next;
5432 
5433 out:
5434 	mutex_exit(&l2arc_dev_mtx);
5435 
5436 	/*
5437 	 * Grab the config lock to prevent the 'next' device from being
5438 	 * removed while we are writing to it.
5439 	 */
5440 	if (next != NULL)
5441 		spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
5442 	mutex_exit(&spa_namespace_lock);
5443 
5444 	return (next);
5445 }
5446 
5447 /*
5448  * Free buffers that were tagged for destruction.
5449  */
5450 static void
5451 l2arc_do_free_on_write()
5452 {
5453 	list_t *buflist;
5454 	l2arc_data_free_t *df, *df_prev;
5455 
5456 	mutex_enter(&l2arc_free_on_write_mtx);
5457 	buflist = l2arc_free_on_write;
5458 
5459 	for (df = list_tail(buflist); df; df = df_prev) {
5460 		df_prev = list_prev(buflist, df);
5461 		ASSERT(df->l2df_data != NULL);
5462 		ASSERT(df->l2df_func != NULL);
5463 		df->l2df_func(df->l2df_data, df->l2df_size);
5464 		list_remove(buflist, df);
5465 		kmem_free(df, sizeof (l2arc_data_free_t));
5466 	}
5467 
5468 	mutex_exit(&l2arc_free_on_write_mtx);
5469 }
5470 
5471 /*
5472  * A write to a cache device has completed.  Update all headers to allow
5473  * reads from these buffers to begin.
5474  */
5475 static void
5476 l2arc_write_done(zio_t *zio)
5477 {
5478 	l2arc_write_callback_t *cb;
5479 	l2arc_dev_t *dev;
5480 	list_t *buflist;
5481 	arc_buf_hdr_t *head, *hdr, *hdr_prev;
5482 	kmutex_t *hash_lock;
5483 	int64_t bytes_dropped = 0;
5484 
5485 	cb = zio->io_private;
5486 	ASSERT(cb != NULL);
5487 	dev = cb->l2wcb_dev;
5488 	ASSERT(dev != NULL);
5489 	head = cb->l2wcb_head;
5490 	ASSERT(head != NULL);
5491 	buflist = &dev->l2ad_buflist;
5492 	ASSERT(buflist != NULL);
5493 	DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
5494 	    l2arc_write_callback_t *, cb);
5495 
5496 	if (zio->io_error != 0)
5497 		ARCSTAT_BUMP(arcstat_l2_writes_error);
5498 
5499 	/*
5500 	 * All writes completed, or an error was hit.
5501 	 */
5502 top:
5503 	mutex_enter(&dev->l2ad_mtx);
5504 	for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
5505 		hdr_prev = list_prev(buflist, hdr);
5506 
5507 		hash_lock = HDR_LOCK(hdr);
5508 
5509 		/*
5510 		 * We cannot use mutex_enter or else we can deadlock
5511 		 * with l2arc_write_buffers (due to swapping the order
5512 		 * the hash lock and l2ad_mtx are taken).
5513 		 */
5514 		if (!mutex_tryenter(hash_lock)) {
5515 			/*
5516 			 * Missed the hash lock. We must retry so we
5517 			 * don't leave the ARC_FLAG_L2_WRITING bit set.
5518 			 */
5519 			ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
5520 
5521 			/*
5522 			 * We don't want to rescan the headers we've
5523 			 * already marked as having been written out, so
5524 			 * we reinsert the head node so we can pick up
5525 			 * where we left off.
5526 			 */
5527 			list_remove(buflist, head);
5528 			list_insert_after(buflist, hdr, head);
5529 
5530 			mutex_exit(&dev->l2ad_mtx);
5531 
5532 			/*
5533 			 * We wait for the hash lock to become available
5534 			 * to try and prevent busy waiting, and increase
5535 			 * the chance we'll be able to acquire the lock
5536 			 * the next time around.
5537 			 */
5538 			mutex_enter(hash_lock);
5539 			mutex_exit(hash_lock);
5540 			goto top;
5541 		}
5542 
5543 		/*
5544 		 * We could not have been moved into the arc_l2c_only
5545 		 * state while in-flight due to our ARC_FLAG_L2_WRITING
5546 		 * bit being set. Let's just ensure that's being enforced.
5547 		 */
5548 		ASSERT(HDR_HAS_L1HDR(hdr));
5549 
5550 		/*
5551 		 * We may have allocated a buffer for L2ARC compression,
5552 		 * we must release it to avoid leaking this data.
5553 		 */
5554 		l2arc_release_cdata_buf(hdr);
5555 
5556 		if (zio->io_error != 0) {
5557 			/*
5558 			 * Error - drop L2ARC entry.
5559 			 */
5560 			list_remove(buflist, hdr);
5561 			hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
5562 
5563 			ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize);
5564 			ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
5565 
5566 			bytes_dropped += hdr->b_l2hdr.b_asize;
5567 			(void) refcount_remove_many(&dev->l2ad_alloc,
5568 			    hdr->b_l2hdr.b_asize, hdr);
5569 		}
5570 
5571 		/*
5572 		 * Allow ARC to begin reads and ghost list evictions to
5573 		 * this L2ARC entry.
5574 		 */
5575 		hdr->b_flags &= ~ARC_FLAG_L2_WRITING;
5576 
5577 		mutex_exit(hash_lock);
5578 	}
5579 
5580 	atomic_inc_64(&l2arc_writes_done);
5581 	list_remove(buflist, head);
5582 	ASSERT(!HDR_HAS_L1HDR(head));
5583 	kmem_cache_free(hdr_l2only_cache, head);
5584 	mutex_exit(&dev->l2ad_mtx);
5585 
5586 	vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
5587 
5588 	l2arc_do_free_on_write();
5589 
5590 	kmem_free(cb, sizeof (l2arc_write_callback_t));
5591 }
5592 
5593 /*
5594  * A read to a cache device completed.  Validate buffer contents before
5595  * handing over to the regular ARC routines.
5596  */
5597 static void
5598 l2arc_read_done(zio_t *zio)
5599 {
5600 	l2arc_read_callback_t *cb;
5601 	arc_buf_hdr_t *hdr;
5602 	arc_buf_t *buf;
5603 	kmutex_t *hash_lock;
5604 	int equal;
5605 
5606 	ASSERT(zio->io_vd != NULL);
5607 	ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
5608 
5609 	spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
5610 
5611 	cb = zio->io_private;
5612 	ASSERT(cb != NULL);
5613 	buf = cb->l2rcb_buf;
5614 	ASSERT(buf != NULL);
5615 
5616 	hash_lock = HDR_LOCK(buf->b_hdr);
5617 	mutex_enter(hash_lock);
5618 	hdr = buf->b_hdr;
5619 	ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5620 
5621 	/*
5622 	 * If the buffer was compressed, decompress it first.
5623 	 */
5624 	if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
5625 		l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
5626 	ASSERT(zio->io_data != NULL);
5627 
5628 	/*
5629 	 * Check this survived the L2ARC journey.
5630 	 */
5631 	equal = arc_cksum_equal(buf);
5632 	if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
5633 		mutex_exit(hash_lock);
5634 		zio->io_private = buf;
5635 		zio->io_bp_copy = cb->l2rcb_bp;	/* XXX fix in L2ARC 2.0	*/
5636 		zio->io_bp = &zio->io_bp_copy;	/* XXX fix in L2ARC 2.0	*/
5637 		arc_read_done(zio);
5638 	} else {
5639 		mutex_exit(hash_lock);
5640 		/*
5641 		 * Buffer didn't survive caching.  Increment stats and
5642 		 * reissue to the original storage device.
5643 		 */
5644 		if (zio->io_error != 0) {
5645 			ARCSTAT_BUMP(arcstat_l2_io_error);
5646 		} else {
5647 			zio->io_error = SET_ERROR(EIO);
5648 		}
5649 		if (!equal)
5650 			ARCSTAT_BUMP(arcstat_l2_cksum_bad);
5651 
5652 		/*
5653 		 * If there's no waiter, issue an async i/o to the primary
5654 		 * storage now.  If there *is* a waiter, the caller must
5655 		 * issue the i/o in a context where it's OK to block.
5656 		 */
5657 		if (zio->io_waiter == NULL) {
5658 			zio_t *pio = zio_unique_parent(zio);
5659 
5660 			ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
5661 
5662 			zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
5663 			    buf->b_data, zio->io_size, arc_read_done, buf,
5664 			    zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
5665 		}
5666 	}
5667 
5668 	kmem_free(cb, sizeof (l2arc_read_callback_t));
5669 }
5670 
5671 /*
5672  * This is the list priority from which the L2ARC will search for pages to
5673  * cache.  This is used within loops (0..3) to cycle through lists in the
5674  * desired order.  This order can have a significant effect on cache
5675  * performance.
5676  *
5677  * Currently the metadata lists are hit first, MFU then MRU, followed by
5678  * the data lists.  This function returns a locked list, and also returns
5679  * the lock pointer.
5680  */
5681 static multilist_sublist_t *
5682 l2arc_sublist_lock(int list_num)
5683 {
5684 	multilist_t *ml = NULL;
5685 	unsigned int idx;
5686 
5687 	ASSERT(list_num >= 0 && list_num <= 3);
5688 
5689 	switch (list_num) {
5690 	case 0:
5691 		ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
5692 		break;
5693 	case 1:
5694 		ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
5695 		break;
5696 	case 2:
5697 		ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
5698 		break;
5699 	case 3:
5700 		ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
5701 		break;
5702 	}
5703 
5704 	/*
5705 	 * Return a randomly-selected sublist. This is acceptable
5706 	 * because the caller feeds only a little bit of data for each
5707 	 * call (8MB). Subsequent calls will result in different
5708 	 * sublists being selected.
5709 	 */
5710 	idx = multilist_get_random_index(ml);
5711 	return (multilist_sublist_lock(ml, idx));
5712 }
5713 
5714 /*
5715  * Evict buffers from the device write hand to the distance specified in
5716  * bytes.  This distance may span populated buffers, it may span nothing.
5717  * This is clearing a region on the L2ARC device ready for writing.
5718  * If the 'all' boolean is set, every buffer is evicted.
5719  */
5720 static void
5721 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
5722 {
5723 	list_t *buflist;
5724 	arc_buf_hdr_t *hdr, *hdr_prev;
5725 	kmutex_t *hash_lock;
5726 	uint64_t taddr;
5727 
5728 	buflist = &dev->l2ad_buflist;
5729 
5730 	if (!all && dev->l2ad_first) {
5731 		/*
5732 		 * This is the first sweep through the device.  There is
5733 		 * nothing to evict.
5734 		 */
5735 		return;
5736 	}
5737 
5738 	if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
5739 		/*
5740 		 * When nearing the end of the device, evict to the end
5741 		 * before the device write hand jumps to the start.
5742 		 */
5743 		taddr = dev->l2ad_end;
5744 	} else {
5745 		taddr = dev->l2ad_hand + distance;
5746 	}
5747 	DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
5748 	    uint64_t, taddr, boolean_t, all);
5749 
5750 top:
5751 	mutex_enter(&dev->l2ad_mtx);
5752 	for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
5753 		hdr_prev = list_prev(buflist, hdr);
5754 
5755 		hash_lock = HDR_LOCK(hdr);
5756 
5757 		/*
5758 		 * We cannot use mutex_enter or else we can deadlock
5759 		 * with l2arc_write_buffers (due to swapping the order
5760 		 * the hash lock and l2ad_mtx are taken).
5761 		 */
5762 		if (!mutex_tryenter(hash_lock)) {
5763 			/*
5764 			 * Missed the hash lock.  Retry.
5765 			 */
5766 			ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
5767 			mutex_exit(&dev->l2ad_mtx);
5768 			mutex_enter(hash_lock);
5769 			mutex_exit(hash_lock);
5770 			goto top;
5771 		}
5772 
5773 		if (HDR_L2_WRITE_HEAD(hdr)) {
5774 			/*
5775 			 * We hit a write head node.  Leave it for
5776 			 * l2arc_write_done().
5777 			 */
5778 			list_remove(buflist, hdr);
5779 			mutex_exit(hash_lock);
5780 			continue;
5781 		}
5782 
5783 		if (!all && HDR_HAS_L2HDR(hdr) &&
5784 		    (hdr->b_l2hdr.b_daddr > taddr ||
5785 		    hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
5786 			/*
5787 			 * We've evicted to the target address,
5788 			 * or the end of the device.
5789 			 */
5790 			mutex_exit(hash_lock);
5791 			break;
5792 		}
5793 
5794 		ASSERT(HDR_HAS_L2HDR(hdr));
5795 		if (!HDR_HAS_L1HDR(hdr)) {
5796 			ASSERT(!HDR_L2_READING(hdr));
5797 			/*
5798 			 * This doesn't exist in the ARC.  Destroy.
5799 			 * arc_hdr_destroy() will call list_remove()
5800 			 * and decrement arcstat_l2_size.
5801 			 */
5802 			arc_change_state(arc_anon, hdr, hash_lock);
5803 			arc_hdr_destroy(hdr);
5804 		} else {
5805 			ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
5806 			ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
5807 			/*
5808 			 * Invalidate issued or about to be issued
5809 			 * reads, since we may be about to write
5810 			 * over this location.
5811 			 */
5812 			if (HDR_L2_READING(hdr)) {
5813 				ARCSTAT_BUMP(arcstat_l2_evict_reading);
5814 				hdr->b_flags |= ARC_FLAG_L2_EVICTED;
5815 			}
5816 
5817 			/* Ensure this header has finished being written */
5818 			ASSERT(!HDR_L2_WRITING(hdr));
5819 			ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
5820 
5821 			arc_hdr_l2hdr_destroy(hdr);
5822 		}
5823 		mutex_exit(hash_lock);
5824 	}
5825 	mutex_exit(&dev->l2ad_mtx);
5826 }
5827 
5828 /*
5829  * Find and write ARC buffers to the L2ARC device.
5830  *
5831  * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
5832  * for reading until they have completed writing.
5833  * The headroom_boost is an in-out parameter used to maintain headroom boost
5834  * state between calls to this function.
5835  *
5836  * Returns the number of bytes actually written (which may be smaller than
5837  * the delta by which the device hand has changed due to alignment).
5838  */
5839 static uint64_t
5840 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
5841     boolean_t *headroom_boost)
5842 {
5843 	arc_buf_hdr_t *hdr, *hdr_prev, *head;
5844 	uint64_t write_asize, write_psize, write_sz, headroom,
5845 	    buf_compress_minsz;
5846 	void *buf_data;
5847 	boolean_t full;
5848 	l2arc_write_callback_t *cb;
5849 	zio_t *pio, *wzio;
5850 	uint64_t guid = spa_load_guid(spa);
5851 	const boolean_t do_headroom_boost = *headroom_boost;
5852 
5853 	ASSERT(dev->l2ad_vdev != NULL);
5854 
5855 	/* Lower the flag now, we might want to raise it again later. */
5856 	*headroom_boost = B_FALSE;
5857 
5858 	pio = NULL;
5859 	write_sz = write_asize = write_psize = 0;
5860 	full = B_FALSE;
5861 	head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
5862 	head->b_flags |= ARC_FLAG_L2_WRITE_HEAD;
5863 	head->b_flags |= ARC_FLAG_HAS_L2HDR;
5864 
5865 	/*
5866 	 * We will want to try to compress buffers that are at least 2x the
5867 	 * device sector size.
5868 	 */
5869 	buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
5870 
5871 	/*
5872 	 * Copy buffers for L2ARC writing.
5873 	 */
5874 	for (int try = 0; try <= 3; try++) {
5875 		multilist_sublist_t *mls = l2arc_sublist_lock(try);
5876 		uint64_t passed_sz = 0;
5877 
5878 		/*
5879 		 * L2ARC fast warmup.
5880 		 *
5881 		 * Until the ARC is warm and starts to evict, read from the
5882 		 * head of the ARC lists rather than the tail.
5883 		 */
5884 		if (arc_warm == B_FALSE)
5885 			hdr = multilist_sublist_head(mls);
5886 		else
5887 			hdr = multilist_sublist_tail(mls);
5888 
5889 		headroom = target_sz * l2arc_headroom;
5890 		if (do_headroom_boost)
5891 			headroom = (headroom * l2arc_headroom_boost) / 100;
5892 
5893 		for (; hdr; hdr = hdr_prev) {
5894 			kmutex_t *hash_lock;
5895 			uint64_t buf_sz;
5896 
5897 			if (arc_warm == B_FALSE)
5898 				hdr_prev = multilist_sublist_next(mls, hdr);
5899 			else
5900 				hdr_prev = multilist_sublist_prev(mls, hdr);
5901 
5902 			hash_lock = HDR_LOCK(hdr);
5903 			if (!mutex_tryenter(hash_lock)) {
5904 				/*
5905 				 * Skip this buffer rather than waiting.
5906 				 */
5907 				continue;
5908 			}
5909 
5910 			passed_sz += hdr->b_size;
5911 			if (passed_sz > headroom) {
5912 				/*
5913 				 * Searched too far.
5914 				 */
5915 				mutex_exit(hash_lock);
5916 				break;
5917 			}
5918 
5919 			if (!l2arc_write_eligible(guid, hdr)) {
5920 				mutex_exit(hash_lock);
5921 				continue;
5922 			}
5923 
5924 			if ((write_sz + hdr->b_size) > target_sz) {
5925 				full = B_TRUE;
5926 				mutex_exit(hash_lock);
5927 				break;
5928 			}
5929 
5930 			if (pio == NULL) {
5931 				/*
5932 				 * Insert a dummy header on the buflist so
5933 				 * l2arc_write_done() can find where the
5934 				 * write buffers begin without searching.
5935 				 */
5936 				mutex_enter(&dev->l2ad_mtx);
5937 				list_insert_head(&dev->l2ad_buflist, head);
5938 				mutex_exit(&dev->l2ad_mtx);
5939 
5940 				cb = kmem_alloc(
5941 				    sizeof (l2arc_write_callback_t), KM_SLEEP);
5942 				cb->l2wcb_dev = dev;
5943 				cb->l2wcb_head = head;
5944 				pio = zio_root(spa, l2arc_write_done, cb,
5945 				    ZIO_FLAG_CANFAIL);
5946 			}
5947 
5948 			/*
5949 			 * Create and add a new L2ARC header.
5950 			 */
5951 			hdr->b_l2hdr.b_dev = dev;
5952 			hdr->b_flags |= ARC_FLAG_L2_WRITING;
5953 			/*
5954 			 * Temporarily stash the data buffer in b_tmp_cdata.
5955 			 * The subsequent write step will pick it up from
5956 			 * there. This is because can't access b_l1hdr.b_buf
5957 			 * without holding the hash_lock, which we in turn
5958 			 * can't access without holding the ARC list locks
5959 			 * (which we want to avoid during compression/writing).
5960 			 */
5961 			HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
5962 			hdr->b_l2hdr.b_asize = hdr->b_size;
5963 			hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data;
5964 
5965 			/*
5966 			 * Explicitly set the b_daddr field to a known
5967 			 * value which means "invalid address". This
5968 			 * enables us to differentiate which stage of
5969 			 * l2arc_write_buffers() the particular header
5970 			 * is in (e.g. this loop, or the one below).
5971 			 * ARC_FLAG_L2_WRITING is not enough to make
5972 			 * this distinction, and we need to know in
5973 			 * order to do proper l2arc vdev accounting in
5974 			 * arc_release() and arc_hdr_destroy().
5975 			 *
5976 			 * Note, we can't use a new flag to distinguish
5977 			 * the two stages because we don't hold the
5978 			 * header's hash_lock below, in the second stage
5979 			 * of this function. Thus, we can't simply
5980 			 * change the b_flags field to denote that the
5981 			 * IO has been sent. We can change the b_daddr
5982 			 * field of the L2 portion, though, since we'll
5983 			 * be holding the l2ad_mtx; which is why we're
5984 			 * using it to denote the header's state change.
5985 			 */
5986 			hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET;
5987 
5988 			buf_sz = hdr->b_size;
5989 			hdr->b_flags |= ARC_FLAG_HAS_L2HDR;
5990 
5991 			mutex_enter(&dev->l2ad_mtx);
5992 			list_insert_head(&dev->l2ad_buflist, hdr);
5993 			mutex_exit(&dev->l2ad_mtx);
5994 
5995 			/*
5996 			 * Compute and store the buffer cksum before
5997 			 * writing.  On debug the cksum is verified first.
5998 			 */
5999 			arc_cksum_verify(hdr->b_l1hdr.b_buf);
6000 			arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE);
6001 
6002 			mutex_exit(hash_lock);
6003 
6004 			write_sz += buf_sz;
6005 		}
6006 
6007 		multilist_sublist_unlock(mls);
6008 
6009 		if (full == B_TRUE)
6010 			break;
6011 	}
6012 
6013 	/* No buffers selected for writing? */
6014 	if (pio == NULL) {
6015 		ASSERT0(write_sz);
6016 		ASSERT(!HDR_HAS_L1HDR(head));
6017 		kmem_cache_free(hdr_l2only_cache, head);
6018 		return (0);
6019 	}
6020 
6021 	mutex_enter(&dev->l2ad_mtx);
6022 
6023 	/*
6024 	 * Now start writing the buffers. We're starting at the write head
6025 	 * and work backwards, retracing the course of the buffer selector
6026 	 * loop above.
6027 	 */
6028 	for (hdr = list_prev(&dev->l2ad_buflist, head); hdr;
6029 	    hdr = list_prev(&dev->l2ad_buflist, hdr)) {
6030 		uint64_t buf_sz;
6031 
6032 		/*
6033 		 * We rely on the L1 portion of the header below, so
6034 		 * it's invalid for this header to have been evicted out
6035 		 * of the ghost cache, prior to being written out. The
6036 		 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6037 		 */
6038 		ASSERT(HDR_HAS_L1HDR(hdr));
6039 
6040 		/*
6041 		 * We shouldn't need to lock the buffer here, since we flagged
6042 		 * it as ARC_FLAG_L2_WRITING in the previous step, but we must
6043 		 * take care to only access its L2 cache parameters. In
6044 		 * particular, hdr->l1hdr.b_buf may be invalid by now due to
6045 		 * ARC eviction.
6046 		 */
6047 		hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
6048 
6049 		if ((HDR_L2COMPRESS(hdr)) &&
6050 		    hdr->b_l2hdr.b_asize >= buf_compress_minsz) {
6051 			if (l2arc_compress_buf(hdr)) {
6052 				/*
6053 				 * If compression succeeded, enable headroom
6054 				 * boost on the next scan cycle.
6055 				 */
6056 				*headroom_boost = B_TRUE;
6057 			}
6058 		}
6059 
6060 		/*
6061 		 * Pick up the buffer data we had previously stashed away
6062 		 * (and now potentially also compressed).
6063 		 */
6064 		buf_data = hdr->b_l1hdr.b_tmp_cdata;
6065 		buf_sz = hdr->b_l2hdr.b_asize;
6066 
6067 		/*
6068 		 * We need to do this regardless if buf_sz is zero or
6069 		 * not, otherwise, when this l2hdr is evicted we'll
6070 		 * remove a reference that was never added.
6071 		 */
6072 		(void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr);
6073 
6074 		/* Compression may have squashed the buffer to zero length. */
6075 		if (buf_sz != 0) {
6076 			uint64_t buf_p_sz;
6077 
6078 			wzio = zio_write_phys(pio, dev->l2ad_vdev,
6079 			    dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
6080 			    NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
6081 			    ZIO_FLAG_CANFAIL, B_FALSE);
6082 
6083 			DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
6084 			    zio_t *, wzio);
6085 			(void) zio_nowait(wzio);
6086 
6087 			write_asize += buf_sz;
6088 
6089 			/*
6090 			 * Keep the clock hand suitably device-aligned.
6091 			 */
6092 			buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
6093 			write_psize += buf_p_sz;
6094 			dev->l2ad_hand += buf_p_sz;
6095 		}
6096 	}
6097 
6098 	mutex_exit(&dev->l2ad_mtx);
6099 
6100 	ASSERT3U(write_asize, <=, target_sz);
6101 	ARCSTAT_BUMP(arcstat_l2_writes_sent);
6102 	ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
6103 	ARCSTAT_INCR(arcstat_l2_size, write_sz);
6104 	ARCSTAT_INCR(arcstat_l2_asize, write_asize);
6105 	vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0);
6106 
6107 	/*
6108 	 * Bump device hand to the device start if it is approaching the end.
6109 	 * l2arc_evict() will already have evicted ahead for this case.
6110 	 */
6111 	if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
6112 		dev->l2ad_hand = dev->l2ad_start;
6113 		dev->l2ad_first = B_FALSE;
6114 	}
6115 
6116 	dev->l2ad_writing = B_TRUE;
6117 	(void) zio_wait(pio);
6118 	dev->l2ad_writing = B_FALSE;
6119 
6120 	return (write_asize);
6121 }
6122 
6123 /*
6124  * Compresses an L2ARC buffer.
6125  * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its
6126  * size in l2hdr->b_asize. This routine tries to compress the data and
6127  * depending on the compression result there are three possible outcomes:
6128  * *) The buffer was incompressible. The original l2hdr contents were left
6129  *    untouched and are ready for writing to an L2 device.
6130  * *) The buffer was all-zeros, so there is no need to write it to an L2
6131  *    device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
6132  *    set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
6133  * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
6134  *    data buffer which holds the compressed data to be written, and b_asize
6135  *    tells us how much data there is. b_compress is set to the appropriate
6136  *    compression algorithm. Once writing is done, invoke
6137  *    l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
6138  *
6139  * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
6140  * buffer was incompressible).
6141  */
6142 static boolean_t
6143 l2arc_compress_buf(arc_buf_hdr_t *hdr)
6144 {
6145 	void *cdata;
6146 	size_t csize, len, rounded;
6147 	ASSERT(HDR_HAS_L2HDR(hdr));
6148 	l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
6149 
6150 	ASSERT(HDR_HAS_L1HDR(hdr));
6151 	ASSERT(HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF);
6152 	ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6153 
6154 	len = l2hdr->b_asize;
6155 	cdata = zio_data_buf_alloc(len);
6156 	ASSERT3P(cdata, !=, NULL);
6157 	csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata,
6158 	    cdata, l2hdr->b_asize);
6159 
6160 	rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE);
6161 	if (rounded > csize) {
6162 		bzero((char *)cdata + csize, rounded - csize);
6163 		csize = rounded;
6164 	}
6165 
6166 	if (csize == 0) {
6167 		/* zero block, indicate that there's nothing to write */
6168 		zio_data_buf_free(cdata, len);
6169 		HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_EMPTY);
6170 		l2hdr->b_asize = 0;
6171 		hdr->b_l1hdr.b_tmp_cdata = NULL;
6172 		ARCSTAT_BUMP(arcstat_l2_compress_zeros);
6173 		return (B_TRUE);
6174 	} else if (csize > 0 && csize < len) {
6175 		/*
6176 		 * Compression succeeded, we'll keep the cdata around for
6177 		 * writing and release it afterwards.
6178 		 */
6179 		HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_LZ4);
6180 		l2hdr->b_asize = csize;
6181 		hdr->b_l1hdr.b_tmp_cdata = cdata;
6182 		ARCSTAT_BUMP(arcstat_l2_compress_successes);
6183 		return (B_TRUE);
6184 	} else {
6185 		/*
6186 		 * Compression failed, release the compressed buffer.
6187 		 * l2hdr will be left unmodified.
6188 		 */
6189 		zio_data_buf_free(cdata, len);
6190 		ARCSTAT_BUMP(arcstat_l2_compress_failures);
6191 		return (B_FALSE);
6192 	}
6193 }
6194 
6195 /*
6196  * Decompresses a zio read back from an l2arc device. On success, the
6197  * underlying zio's io_data buffer is overwritten by the uncompressed
6198  * version. On decompression error (corrupt compressed stream), the
6199  * zio->io_error value is set to signal an I/O error.
6200  *
6201  * Please note that the compressed data stream is not checksummed, so
6202  * if the underlying device is experiencing data corruption, we may feed
6203  * corrupt data to the decompressor, so the decompressor needs to be
6204  * able to handle this situation (LZ4 does).
6205  */
6206 static void
6207 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
6208 {
6209 	ASSERT(L2ARC_IS_VALID_COMPRESS(c));
6210 
6211 	if (zio->io_error != 0) {
6212 		/*
6213 		 * An io error has occured, just restore the original io
6214 		 * size in preparation for a main pool read.
6215 		 */
6216 		zio->io_orig_size = zio->io_size = hdr->b_size;
6217 		return;
6218 	}
6219 
6220 	if (c == ZIO_COMPRESS_EMPTY) {
6221 		/*
6222 		 * An empty buffer results in a null zio, which means we
6223 		 * need to fill its io_data after we're done restoring the
6224 		 * buffer's contents.
6225 		 */
6226 		ASSERT(hdr->b_l1hdr.b_buf != NULL);
6227 		bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size);
6228 		zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data;
6229 	} else {
6230 		ASSERT(zio->io_data != NULL);
6231 		/*
6232 		 * We copy the compressed data from the start of the arc buffer
6233 		 * (the zio_read will have pulled in only what we need, the
6234 		 * rest is garbage which we will overwrite at decompression)
6235 		 * and then decompress back to the ARC data buffer. This way we
6236 		 * can minimize copying by simply decompressing back over the
6237 		 * original compressed data (rather than decompressing to an
6238 		 * aux buffer and then copying back the uncompressed buffer,
6239 		 * which is likely to be much larger).
6240 		 */
6241 		uint64_t csize;
6242 		void *cdata;
6243 
6244 		csize = zio->io_size;
6245 		cdata = zio_data_buf_alloc(csize);
6246 		bcopy(zio->io_data, cdata, csize);
6247 		if (zio_decompress_data(c, cdata, zio->io_data, csize,
6248 		    hdr->b_size) != 0)
6249 			zio->io_error = EIO;
6250 		zio_data_buf_free(cdata, csize);
6251 	}
6252 
6253 	/* Restore the expected uncompressed IO size. */
6254 	zio->io_orig_size = zio->io_size = hdr->b_size;
6255 }
6256 
6257 /*
6258  * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
6259  * This buffer serves as a temporary holder of compressed data while
6260  * the buffer entry is being written to an l2arc device. Once that is
6261  * done, we can dispose of it.
6262  */
6263 static void
6264 l2arc_release_cdata_buf(arc_buf_hdr_t *hdr)
6265 {
6266 	enum zio_compress comp = HDR_GET_COMPRESS(hdr);
6267 
6268 	ASSERT(HDR_HAS_L1HDR(hdr));
6269 	ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp));
6270 
6271 	if (comp == ZIO_COMPRESS_OFF) {
6272 		/*
6273 		 * In this case, b_tmp_cdata points to the same buffer
6274 		 * as the arc_buf_t's b_data field. We don't want to
6275 		 * free it, since the arc_buf_t will handle that.
6276 		 */
6277 		hdr->b_l1hdr.b_tmp_cdata = NULL;
6278 	} else if (comp == ZIO_COMPRESS_EMPTY) {
6279 		/*
6280 		 * In this case, b_tmp_cdata was compressed to an empty
6281 		 * buffer, thus there's nothing to free and b_tmp_cdata
6282 		 * should have been set to NULL in l2arc_write_buffers().
6283 		 */
6284 		ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
6285 	} else {
6286 		/*
6287 		 * If the data was compressed, then we've allocated a
6288 		 * temporary buffer for it, so now we need to release it.
6289 		 */
6290 		ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6291 		zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata,
6292 		    hdr->b_size);
6293 		hdr->b_l1hdr.b_tmp_cdata = NULL;
6294 	}
6295 
6296 }
6297 
6298 /*
6299  * This thread feeds the L2ARC at regular intervals.  This is the beating
6300  * heart of the L2ARC.
6301  */
6302 static void
6303 l2arc_feed_thread(void)
6304 {
6305 	callb_cpr_t cpr;
6306 	l2arc_dev_t *dev;
6307 	spa_t *spa;
6308 	uint64_t size, wrote;
6309 	clock_t begin, next = ddi_get_lbolt();
6310 	boolean_t headroom_boost = B_FALSE;
6311 
6312 	CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
6313 
6314 	mutex_enter(&l2arc_feed_thr_lock);
6315 
6316 	while (l2arc_thread_exit == 0) {
6317 		CALLB_CPR_SAFE_BEGIN(&cpr);
6318 		(void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
6319 		    next);
6320 		CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
6321 		next = ddi_get_lbolt() + hz;
6322 
6323 		/*
6324 		 * Quick check for L2ARC devices.
6325 		 */
6326 		mutex_enter(&l2arc_dev_mtx);
6327 		if (l2arc_ndev == 0) {
6328 			mutex_exit(&l2arc_dev_mtx);
6329 			continue;
6330 		}
6331 		mutex_exit(&l2arc_dev_mtx);
6332 		begin = ddi_get_lbolt();
6333 
6334 		/*
6335 		 * This selects the next l2arc device to write to, and in
6336 		 * doing so the next spa to feed from: dev->l2ad_spa.   This
6337 		 * will return NULL if there are now no l2arc devices or if
6338 		 * they are all faulted.
6339 		 *
6340 		 * If a device is returned, its spa's config lock is also
6341 		 * held to prevent device removal.  l2arc_dev_get_next()
6342 		 * will grab and release l2arc_dev_mtx.
6343 		 */
6344 		if ((dev = l2arc_dev_get_next()) == NULL)
6345 			continue;
6346 
6347 		spa = dev->l2ad_spa;
6348 		ASSERT(spa != NULL);
6349 
6350 		/*
6351 		 * If the pool is read-only then force the feed thread to
6352 		 * sleep a little longer.
6353 		 */
6354 		if (!spa_writeable(spa)) {
6355 			next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
6356 			spa_config_exit(spa, SCL_L2ARC, dev);
6357 			continue;
6358 		}
6359 
6360 		/*
6361 		 * Avoid contributing to memory pressure.
6362 		 */
6363 		if (arc_reclaim_needed()) {
6364 			ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
6365 			spa_config_exit(spa, SCL_L2ARC, dev);
6366 			continue;
6367 		}
6368 
6369 		ARCSTAT_BUMP(arcstat_l2_feeds);
6370 
6371 		size = l2arc_write_size();
6372 
6373 		/*
6374 		 * Evict L2ARC buffers that will be overwritten.
6375 		 */
6376 		l2arc_evict(dev, size, B_FALSE);
6377 
6378 		/*
6379 		 * Write ARC buffers.
6380 		 */
6381 		wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
6382 
6383 		/*
6384 		 * Calculate interval between writes.
6385 		 */
6386 		next = l2arc_write_interval(begin, size, wrote);
6387 		spa_config_exit(spa, SCL_L2ARC, dev);
6388 	}
6389 
6390 	l2arc_thread_exit = 0;
6391 	cv_broadcast(&l2arc_feed_thr_cv);
6392 	CALLB_CPR_EXIT(&cpr);		/* drops l2arc_feed_thr_lock */
6393 	thread_exit();
6394 }
6395 
6396 boolean_t
6397 l2arc_vdev_present(vdev_t *vd)
6398 {
6399 	l2arc_dev_t *dev;
6400 
6401 	mutex_enter(&l2arc_dev_mtx);
6402 	for (dev = list_head(l2arc_dev_list); dev != NULL;
6403 	    dev = list_next(l2arc_dev_list, dev)) {
6404 		if (dev->l2ad_vdev == vd)
6405 			break;
6406 	}
6407 	mutex_exit(&l2arc_dev_mtx);
6408 
6409 	return (dev != NULL);
6410 }
6411 
6412 /*
6413  * Add a vdev for use by the L2ARC.  By this point the spa has already
6414  * validated the vdev and opened it.
6415  */
6416 void
6417 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
6418 {
6419 	l2arc_dev_t *adddev;
6420 
6421 	ASSERT(!l2arc_vdev_present(vd));
6422 
6423 	/*
6424 	 * Create a new l2arc device entry.
6425 	 */
6426 	adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
6427 	adddev->l2ad_spa = spa;
6428 	adddev->l2ad_vdev = vd;
6429 	adddev->l2ad_start = VDEV_LABEL_START_SIZE;
6430 	adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
6431 	adddev->l2ad_hand = adddev->l2ad_start;
6432 	adddev->l2ad_first = B_TRUE;
6433 	adddev->l2ad_writing = B_FALSE;
6434 
6435 	mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
6436 	/*
6437 	 * This is a list of all ARC buffers that are still valid on the
6438 	 * device.
6439 	 */
6440 	list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
6441 	    offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
6442 
6443 	vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
6444 	refcount_create(&adddev->l2ad_alloc);
6445 
6446 	/*
6447 	 * Add device to global list
6448 	 */
6449 	mutex_enter(&l2arc_dev_mtx);
6450 	list_insert_head(l2arc_dev_list, adddev);
6451 	atomic_inc_64(&l2arc_ndev);
6452 	mutex_exit(&l2arc_dev_mtx);
6453 }
6454 
6455 /*
6456  * Remove a vdev from the L2ARC.
6457  */
6458 void
6459 l2arc_remove_vdev(vdev_t *vd)
6460 {
6461 	l2arc_dev_t *dev, *nextdev, *remdev = NULL;
6462 
6463 	/*
6464 	 * Find the device by vdev
6465 	 */
6466 	mutex_enter(&l2arc_dev_mtx);
6467 	for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
6468 		nextdev = list_next(l2arc_dev_list, dev);
6469 		if (vd == dev->l2ad_vdev) {
6470 			remdev = dev;
6471 			break;
6472 		}
6473 	}
6474 	ASSERT(remdev != NULL);
6475 
6476 	/*
6477 	 * Remove device from global list
6478 	 */
6479 	list_remove(l2arc_dev_list, remdev);
6480 	l2arc_dev_last = NULL;		/* may have been invalidated */
6481 	atomic_dec_64(&l2arc_ndev);
6482 	mutex_exit(&l2arc_dev_mtx);
6483 
6484 	/*
6485 	 * Clear all buflists and ARC references.  L2ARC device flush.
6486 	 */
6487 	l2arc_evict(remdev, 0, B_TRUE);
6488 	list_destroy(&remdev->l2ad_buflist);
6489 	mutex_destroy(&remdev->l2ad_mtx);
6490 	refcount_destroy(&remdev->l2ad_alloc);
6491 	kmem_free(remdev, sizeof (l2arc_dev_t));
6492 }
6493 
6494 void
6495 l2arc_init(void)
6496 {
6497 	l2arc_thread_exit = 0;
6498 	l2arc_ndev = 0;
6499 	l2arc_writes_sent = 0;
6500 	l2arc_writes_done = 0;
6501 
6502 	mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
6503 	cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
6504 	mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
6505 	mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
6506 
6507 	l2arc_dev_list = &L2ARC_dev_list;
6508 	l2arc_free_on_write = &L2ARC_free_on_write;
6509 	list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
6510 	    offsetof(l2arc_dev_t, l2ad_node));
6511 	list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
6512 	    offsetof(l2arc_data_free_t, l2df_list_node));
6513 }
6514 
6515 void
6516 l2arc_fini(void)
6517 {
6518 	/*
6519 	 * This is called from dmu_fini(), which is called from spa_fini();
6520 	 * Because of this, we can assume that all l2arc devices have
6521 	 * already been removed when the pools themselves were removed.
6522 	 */
6523 
6524 	l2arc_do_free_on_write();
6525 
6526 	mutex_destroy(&l2arc_feed_thr_lock);
6527 	cv_destroy(&l2arc_feed_thr_cv);
6528 	mutex_destroy(&l2arc_dev_mtx);
6529 	mutex_destroy(&l2arc_free_on_write_mtx);
6530 
6531 	list_destroy(l2arc_dev_list);
6532 	list_destroy(l2arc_free_on_write);
6533 }
6534 
6535 void
6536 l2arc_start(void)
6537 {
6538 	if (!(spa_mode_global & FWRITE))
6539 		return;
6540 
6541 	(void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
6542 	    TS_RUN, minclsyspri);
6543 }
6544 
6545 void
6546 l2arc_stop(void)
6547 {
6548 	if (!(spa_mode_global & FWRITE))
6549 		return;
6550 
6551 	mutex_enter(&l2arc_feed_thr_lock);
6552 	cv_signal(&l2arc_feed_thr_cv);	/* kick thread out of startup */
6553 	l2arc_thread_exit = 1;
6554 	while (l2arc_thread_exit != 0)
6555 		cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
6556 	mutex_exit(&l2arc_feed_thr_lock);
6557 }
6558