xref: /illumos-gate/usr/src/uts/common/fs/zfs/arc.c (revision 7b38fab6)
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) 2018, Joyent, Inc.
24  * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
25  * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26  * Copyright 2017 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  * Note that the majority of the performance stats are manipulated
103  * with atomic operations.
104  *
105  * The L2ARC uses the l2ad_mtx on each vdev for the following:
106  *
107  *	- L2ARC buflist creation
108  *	- L2ARC buflist eviction
109  *	- L2ARC write completion, which walks L2ARC buflists
110  *	- ARC header destruction, as it removes from L2ARC buflists
111  *	- ARC header release, as it removes from L2ARC buflists
112  */
113 
114 /*
115  * ARC operation:
116  *
117  * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
118  * This structure can point either to a block that is still in the cache or to
119  * one that is only accessible in an L2 ARC device, or it can provide
120  * information about a block that was recently evicted. If a block is
121  * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
122  * information to retrieve it from the L2ARC device. This information is
123  * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
124  * that is in this state cannot access the data directly.
125  *
126  * Blocks that are actively being referenced or have not been evicted
127  * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
128  * the arc_buf_hdr_t that will point to the data block in memory. A block can
129  * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
130  * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
131  * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
132  *
133  * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
134  * ability to store the physical data (b_pabd) associated with the DVA of the
135  * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
136  * it will match its on-disk compression characteristics. This behavior can be
137  * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
138  * compressed ARC functionality is disabled, the b_pabd will point to an
139  * uncompressed version of the on-disk data.
140  *
141  * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
142  * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
143  * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
144  * consumer. The ARC will provide references to this data and will keep it
145  * cached until it is no longer in use. The ARC caches only the L1ARC's physical
146  * data block and will evict any arc_buf_t that is no longer referenced. The
147  * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
148  * "overhead_size" kstat.
149  *
150  * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
151  * compressed form. The typical case is that consumers will want uncompressed
152  * data, and when that happens a new data buffer is allocated where the data is
153  * decompressed for them to use. Currently the only consumer who wants
154  * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
155  * exists on disk. When this happens, the arc_buf_t's data buffer is shared
156  * with the arc_buf_hdr_t.
157  *
158  * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
159  * first one is owned by a compressed send consumer (and therefore references
160  * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
161  * used by any other consumer (and has its own uncompressed copy of the data
162  * buffer).
163  *
164  *   arc_buf_hdr_t
165  *   +-----------+
166  *   | fields    |
167  *   | common to |
168  *   | L1- and   |
169  *   | L2ARC     |
170  *   +-----------+
171  *   | l2arc_buf_hdr_t
172  *   |           |
173  *   +-----------+
174  *   | l1arc_buf_hdr_t
175  *   |           |              arc_buf_t
176  *   | b_buf     +------------>+-----------+      arc_buf_t
177  *   | b_pabd    +-+           |b_next     +---->+-----------+
178  *   +-----------+ |           |-----------|     |b_next     +-->NULL
179  *                 |           |b_comp = T |     +-----------+
180  *                 |           |b_data     +-+   |b_comp = F |
181  *                 |           +-----------+ |   |b_data     +-+
182  *                 +->+------+               |   +-----------+ |
183  *        compressed  |      |               |                 |
184  *           data     |      |<--------------+                 | uncompressed
185  *                    +------+          compressed,            |     data
186  *                                        shared               +-->+------+
187  *                                         data                    |      |
188  *                                                                 |      |
189  *                                                                 +------+
190  *
191  * When a consumer reads a block, the ARC must first look to see if the
192  * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
193  * arc_buf_t and either copies uncompressed data into a new data buffer from an
194  * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
195  * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
196  * hdr is compressed and the desired compression characteristics of the
197  * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
198  * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
199  * the last buffer in the hdr's b_buf list, however a shared compressed buf can
200  * be anywhere in the hdr's list.
201  *
202  * The diagram below shows an example of an uncompressed ARC hdr that is
203  * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
204  * the last element in the buf list):
205  *
206  *                arc_buf_hdr_t
207  *                +-----------+
208  *                |           |
209  *                |           |
210  *                |           |
211  *                +-----------+
212  * l2arc_buf_hdr_t|           |
213  *                |           |
214  *                +-----------+
215  * l1arc_buf_hdr_t|           |
216  *                |           |                 arc_buf_t    (shared)
217  *                |    b_buf  +------------>+---------+      arc_buf_t
218  *                |           |             |b_next   +---->+---------+
219  *                |  b_pabd   +-+           |---------|     |b_next   +-->NULL
220  *                +-----------+ |           |         |     +---------+
221  *                              |           |b_data   +-+   |         |
222  *                              |           +---------+ |   |b_data   +-+
223  *                              +->+------+             |   +---------+ |
224  *                                 |      |             |               |
225  *                   uncompressed  |      |             |               |
226  *                        data     +------+             |               |
227  *                                    ^                 +->+------+     |
228  *                                    |       uncompressed |      |     |
229  *                                    |           data     |      |     |
230  *                                    |                    +------+     |
231  *                                    +---------------------------------+
232  *
233  * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
234  * since the physical block is about to be rewritten. The new data contents
235  * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
236  * it may compress the data before writing it to disk. The ARC will be called
237  * with the transformed data and will bcopy the transformed on-disk block into
238  * a newly allocated b_pabd. Writes are always done into buffers which have
239  * either been loaned (and hence are new and don't have other readers) or
240  * buffers which have been released (and hence have their own hdr, if there
241  * were originally other readers of the buf's original hdr). This ensures that
242  * the ARC only needs to update a single buf and its hdr after a write occurs.
243  *
244  * When the L2ARC is in use, it will also take advantage of the b_pabd. The
245  * L2ARC will always write the contents of b_pabd to the L2ARC. This means
246  * that when compressed ARC is enabled that the L2ARC blocks are identical
247  * to the on-disk block in the main data pool. This provides a significant
248  * advantage since the ARC can leverage the bp's checksum when reading from the
249  * L2ARC to determine if the contents are valid. However, if the compressed
250  * ARC is disabled, then the L2ARC's block must be transformed to look
251  * like the physical block in the main data pool before comparing the
252  * checksum and determining its validity.
253  */
254 
255 #include <sys/spa.h>
256 #include <sys/zio.h>
257 #include <sys/spa_impl.h>
258 #include <sys/zio_compress.h>
259 #include <sys/zio_checksum.h>
260 #include <sys/zfs_context.h>
261 #include <sys/arc.h>
262 #include <sys/refcount.h>
263 #include <sys/vdev.h>
264 #include <sys/vdev_impl.h>
265 #include <sys/dsl_pool.h>
266 #include <sys/zio_checksum.h>
267 #include <sys/multilist.h>
268 #include <sys/abd.h>
269 #ifdef _KERNEL
270 #include <sys/vmsystm.h>
271 #include <vm/anon.h>
272 #include <sys/fs/swapnode.h>
273 #include <sys/dnlc.h>
274 #endif
275 #include <sys/callb.h>
276 #include <sys/kstat.h>
277 #include <sys/zthr.h>
278 #include <zfs_fletcher.h>
279 #include <sys/aggsum.h>
280 #include <sys/cityhash.h>
281 
282 #ifndef _KERNEL
283 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
284 boolean_t arc_watch = B_FALSE;
285 int arc_procfd;
286 #endif
287 
288 /*
289  * This thread's job is to keep enough free memory in the system, by
290  * calling arc_kmem_reap_now() plus arc_shrink(), which improves
291  * arc_available_memory().
292  */
293 static zthr_t		*arc_reap_zthr;
294 
295 /*
296  * This thread's job is to keep arc_size under arc_c, by calling
297  * arc_adjust(), which improves arc_is_overflowing().
298  */
299 static zthr_t		*arc_adjust_zthr;
300 
301 static kmutex_t		arc_adjust_lock;
302 static kcondvar_t	arc_adjust_waiters_cv;
303 static boolean_t	arc_adjust_needed = B_FALSE;
304 
305 uint_t arc_reduce_dnlc_percent = 3;
306 
307 /*
308  * The number of headers to evict in arc_evict_state_impl() before
309  * dropping the sublist lock and evicting from another sublist. A lower
310  * value means we're more likely to evict the "correct" header (i.e. the
311  * oldest header in the arc state), but comes with higher overhead
312  * (i.e. more invocations of arc_evict_state_impl()).
313  */
314 int zfs_arc_evict_batch_limit = 10;
315 
316 /* number of seconds before growing cache again */
317 int arc_grow_retry = 60;
318 
319 /*
320  * Minimum time between calls to arc_kmem_reap_soon().  Note that this will
321  * be converted to ticks, so with the default hz=100, a setting of 15 ms
322  * will actually wait 2 ticks, or 20ms.
323  */
324 int arc_kmem_cache_reap_retry_ms = 1000;
325 
326 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
327 int zfs_arc_overflow_shift = 8;
328 
329 /* shift of arc_c for calculating both min and max arc_p */
330 int arc_p_min_shift = 4;
331 
332 /* log2(fraction of arc to reclaim) */
333 int arc_shrink_shift = 7;
334 
335 /*
336  * log2(fraction of ARC which must be free to allow growing).
337  * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
338  * when reading a new block into the ARC, we will evict an equal-sized block
339  * from the ARC.
340  *
341  * This must be less than arc_shrink_shift, so that when we shrink the ARC,
342  * we will still not allow it to grow.
343  */
344 int			arc_no_grow_shift = 5;
345 
346 
347 /*
348  * minimum lifespan of a prefetch block in clock ticks
349  * (initialized in arc_init())
350  */
351 static int		arc_min_prefetch_lifespan;
352 
353 /*
354  * If this percent of memory is free, don't throttle.
355  */
356 int arc_lotsfree_percent = 10;
357 
358 static boolean_t arc_initialized;
359 
360 /*
361  * The arc has filled available memory and has now warmed up.
362  */
363 static boolean_t arc_warm;
364 
365 /*
366  * log2 fraction of the zio arena to keep free.
367  */
368 int arc_zio_arena_free_shift = 2;
369 
370 /*
371  * These tunables are for performance analysis.
372  */
373 uint64_t zfs_arc_max;
374 uint64_t zfs_arc_min;
375 uint64_t zfs_arc_meta_limit = 0;
376 uint64_t zfs_arc_meta_min = 0;
377 int zfs_arc_grow_retry = 0;
378 int zfs_arc_shrink_shift = 0;
379 int zfs_arc_p_min_shift = 0;
380 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
381 
382 /*
383  * ARC dirty data constraints for arc_tempreserve_space() throttle
384  */
385 uint_t zfs_arc_dirty_limit_percent = 50;	/* total dirty data limit */
386 uint_t zfs_arc_anon_limit_percent = 25;		/* anon block dirty limit */
387 uint_t zfs_arc_pool_dirty_percent = 20;		/* each pool's anon allowance */
388 
389 boolean_t zfs_compressed_arc_enabled = B_TRUE;
390 
391 /*
392  * Note that buffers can be in one of 6 states:
393  *	ARC_anon	- anonymous (discussed below)
394  *	ARC_mru		- recently used, currently cached
395  *	ARC_mru_ghost	- recentely used, no longer in cache
396  *	ARC_mfu		- frequently used, currently cached
397  *	ARC_mfu_ghost	- frequently used, no longer in cache
398  *	ARC_l2c_only	- exists in L2ARC but not other states
399  * When there are no active references to the buffer, they are
400  * are linked onto a list in one of these arc states.  These are
401  * the only buffers that can be evicted or deleted.  Within each
402  * state there are multiple lists, one for meta-data and one for
403  * non-meta-data.  Meta-data (indirect blocks, blocks of dnodes,
404  * etc.) is tracked separately so that it can be managed more
405  * explicitly: favored over data, limited explicitly.
406  *
407  * Anonymous buffers are buffers that are not associated with
408  * a DVA.  These are buffers that hold dirty block copies
409  * before they are written to stable storage.  By definition,
410  * they are "ref'd" and are considered part of arc_mru
411  * that cannot be freed.  Generally, they will aquire a DVA
412  * as they are written and migrate onto the arc_mru list.
413  *
414  * The ARC_l2c_only state is for buffers that are in the second
415  * level ARC but no longer in any of the ARC_m* lists.  The second
416  * level ARC itself may also contain buffers that are in any of
417  * the ARC_m* states - meaning that a buffer can exist in two
418  * places.  The reason for the ARC_l2c_only state is to keep the
419  * buffer header in the hash table, so that reads that hit the
420  * second level ARC benefit from these fast lookups.
421  */
422 
423 typedef struct arc_state {
424 	/*
425 	 * list of evictable buffers
426 	 */
427 	multilist_t *arcs_list[ARC_BUFC_NUMTYPES];
428 	/*
429 	 * total amount of evictable data in this state
430 	 */
431 	zfs_refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
432 	/*
433 	 * total amount of data in this state; this includes: evictable,
434 	 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
435 	 */
436 	zfs_refcount_t arcs_size;
437 } arc_state_t;
438 
439 /* The 6 states: */
440 static arc_state_t ARC_anon;
441 static arc_state_t ARC_mru;
442 static arc_state_t ARC_mru_ghost;
443 static arc_state_t ARC_mfu;
444 static arc_state_t ARC_mfu_ghost;
445 static arc_state_t ARC_l2c_only;
446 
447 typedef struct arc_stats {
448 	kstat_named_t arcstat_hits;
449 	kstat_named_t arcstat_misses;
450 	kstat_named_t arcstat_demand_data_hits;
451 	kstat_named_t arcstat_demand_data_misses;
452 	kstat_named_t arcstat_demand_metadata_hits;
453 	kstat_named_t arcstat_demand_metadata_misses;
454 	kstat_named_t arcstat_prefetch_data_hits;
455 	kstat_named_t arcstat_prefetch_data_misses;
456 	kstat_named_t arcstat_prefetch_metadata_hits;
457 	kstat_named_t arcstat_prefetch_metadata_misses;
458 	kstat_named_t arcstat_mru_hits;
459 	kstat_named_t arcstat_mru_ghost_hits;
460 	kstat_named_t arcstat_mfu_hits;
461 	kstat_named_t arcstat_mfu_ghost_hits;
462 	kstat_named_t arcstat_deleted;
463 	/*
464 	 * Number of buffers that could not be evicted because the hash lock
465 	 * was held by another thread.  The lock may not necessarily be held
466 	 * by something using the same buffer, since hash locks are shared
467 	 * by multiple buffers.
468 	 */
469 	kstat_named_t arcstat_mutex_miss;
470 	/*
471 	 * Number of buffers skipped when updating the access state due to the
472 	 * header having already been released after acquiring the hash lock.
473 	 */
474 	kstat_named_t arcstat_access_skip;
475 	/*
476 	 * Number of buffers skipped because they have I/O in progress, are
477 	 * indirect prefetch buffers that have not lived long enough, or are
478 	 * not from the spa we're trying to evict from.
479 	 */
480 	kstat_named_t arcstat_evict_skip;
481 	/*
482 	 * Number of times arc_evict_state() was unable to evict enough
483 	 * buffers to reach it's target amount.
484 	 */
485 	kstat_named_t arcstat_evict_not_enough;
486 	kstat_named_t arcstat_evict_l2_cached;
487 	kstat_named_t arcstat_evict_l2_eligible;
488 	kstat_named_t arcstat_evict_l2_ineligible;
489 	kstat_named_t arcstat_evict_l2_skip;
490 	kstat_named_t arcstat_hash_elements;
491 	kstat_named_t arcstat_hash_elements_max;
492 	kstat_named_t arcstat_hash_collisions;
493 	kstat_named_t arcstat_hash_chains;
494 	kstat_named_t arcstat_hash_chain_max;
495 	kstat_named_t arcstat_p;
496 	kstat_named_t arcstat_c;
497 	kstat_named_t arcstat_c_min;
498 	kstat_named_t arcstat_c_max;
499 	/* Not updated directly; only synced in arc_kstat_update. */
500 	kstat_named_t arcstat_size;
501 	/*
502 	 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
503 	 * Note that the compressed bytes may match the uncompressed bytes
504 	 * if the block is either not compressed or compressed arc is disabled.
505 	 */
506 	kstat_named_t arcstat_compressed_size;
507 	/*
508 	 * Uncompressed size of the data stored in b_pabd. If compressed
509 	 * arc is disabled then this value will be identical to the stat
510 	 * above.
511 	 */
512 	kstat_named_t arcstat_uncompressed_size;
513 	/*
514 	 * Number of bytes stored in all the arc_buf_t's. This is classified
515 	 * as "overhead" since this data is typically short-lived and will
516 	 * be evicted from the arc when it becomes unreferenced unless the
517 	 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
518 	 * values have been set (see comment in dbuf.c for more information).
519 	 */
520 	kstat_named_t arcstat_overhead_size;
521 	/*
522 	 * Number of bytes consumed by internal ARC structures necessary
523 	 * for tracking purposes; these structures are not actually
524 	 * backed by ARC buffers. This includes arc_buf_hdr_t structures
525 	 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
526 	 * caches), and arc_buf_t structures (allocated via arc_buf_t
527 	 * cache).
528 	 * Not updated directly; only synced in arc_kstat_update.
529 	 */
530 	kstat_named_t arcstat_hdr_size;
531 	/*
532 	 * Number of bytes consumed by ARC buffers of type equal to
533 	 * ARC_BUFC_DATA. This is generally consumed by buffers backing
534 	 * on disk user data (e.g. plain file contents).
535 	 * Not updated directly; only synced in arc_kstat_update.
536 	 */
537 	kstat_named_t arcstat_data_size;
538 	/*
539 	 * Number of bytes consumed by ARC buffers of type equal to
540 	 * ARC_BUFC_METADATA. This is generally consumed by buffers
541 	 * backing on disk data that is used for internal ZFS
542 	 * structures (e.g. ZAP, dnode, indirect blocks, etc).
543 	 * Not updated directly; only synced in arc_kstat_update.
544 	 */
545 	kstat_named_t arcstat_metadata_size;
546 	/*
547 	 * Number of bytes consumed by various buffers and structures
548 	 * not actually backed with ARC buffers. This includes bonus
549 	 * buffers (allocated directly via zio_buf_* functions),
550 	 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
551 	 * cache), and dnode_t structures (allocated via dnode_t cache).
552 	 * Not updated directly; only synced in arc_kstat_update.
553 	 */
554 	kstat_named_t arcstat_other_size;
555 	/*
556 	 * Total number of bytes consumed by ARC buffers residing in the
557 	 * arc_anon state. This includes *all* buffers in the arc_anon
558 	 * state; e.g. data, metadata, evictable, and unevictable buffers
559 	 * are all included in this value.
560 	 * Not updated directly; only synced in arc_kstat_update.
561 	 */
562 	kstat_named_t arcstat_anon_size;
563 	/*
564 	 * Number of bytes consumed by ARC buffers that meet the
565 	 * following criteria: backing buffers of type ARC_BUFC_DATA,
566 	 * residing in the arc_anon state, and are eligible for eviction
567 	 * (e.g. have no outstanding holds on the buffer).
568 	 * Not updated directly; only synced in arc_kstat_update.
569 	 */
570 	kstat_named_t arcstat_anon_evictable_data;
571 	/*
572 	 * Number of bytes consumed by ARC buffers that meet the
573 	 * following criteria: backing buffers of type ARC_BUFC_METADATA,
574 	 * residing in the arc_anon state, and are eligible for eviction
575 	 * (e.g. have no outstanding holds on the buffer).
576 	 * Not updated directly; only synced in arc_kstat_update.
577 	 */
578 	kstat_named_t arcstat_anon_evictable_metadata;
579 	/*
580 	 * Total number of bytes consumed by ARC buffers residing in the
581 	 * arc_mru state. This includes *all* buffers in the arc_mru
582 	 * state; e.g. data, metadata, evictable, and unevictable buffers
583 	 * are all included in this value.
584 	 * Not updated directly; only synced in arc_kstat_update.
585 	 */
586 	kstat_named_t arcstat_mru_size;
587 	/*
588 	 * Number of bytes consumed by ARC buffers that meet the
589 	 * following criteria: backing buffers of type ARC_BUFC_DATA,
590 	 * residing in the arc_mru state, and are eligible for eviction
591 	 * (e.g. have no outstanding holds on the buffer).
592 	 * Not updated directly; only synced in arc_kstat_update.
593 	 */
594 	kstat_named_t arcstat_mru_evictable_data;
595 	/*
596 	 * Number of bytes consumed by ARC buffers that meet the
597 	 * following criteria: backing buffers of type ARC_BUFC_METADATA,
598 	 * residing in the arc_mru state, and are eligible for eviction
599 	 * (e.g. have no outstanding holds on the buffer).
600 	 * Not updated directly; only synced in arc_kstat_update.
601 	 */
602 	kstat_named_t arcstat_mru_evictable_metadata;
603 	/*
604 	 * Total number of bytes that *would have been* consumed by ARC
605 	 * buffers in the arc_mru_ghost state. The key thing to note
606 	 * here, is the fact that this size doesn't actually indicate
607 	 * RAM consumption. The ghost lists only consist of headers and
608 	 * don't actually have ARC buffers linked off of these headers.
609 	 * Thus, *if* the headers had associated ARC buffers, these
610 	 * buffers *would have* consumed this number of bytes.
611 	 * Not updated directly; only synced in arc_kstat_update.
612 	 */
613 	kstat_named_t arcstat_mru_ghost_size;
614 	/*
615 	 * Number of bytes that *would have been* consumed by ARC
616 	 * buffers that are eligible for eviction, of type
617 	 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
618 	 * Not updated directly; only synced in arc_kstat_update.
619 	 */
620 	kstat_named_t arcstat_mru_ghost_evictable_data;
621 	/*
622 	 * Number of bytes that *would have been* consumed by ARC
623 	 * buffers that are eligible for eviction, of type
624 	 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
625 	 * Not updated directly; only synced in arc_kstat_update.
626 	 */
627 	kstat_named_t arcstat_mru_ghost_evictable_metadata;
628 	/*
629 	 * Total number of bytes consumed by ARC buffers residing in the
630 	 * arc_mfu state. This includes *all* buffers in the arc_mfu
631 	 * state; e.g. data, metadata, evictable, and unevictable buffers
632 	 * are all included in this value.
633 	 * Not updated directly; only synced in arc_kstat_update.
634 	 */
635 	kstat_named_t arcstat_mfu_size;
636 	/*
637 	 * Number of bytes consumed by ARC buffers that are eligible for
638 	 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
639 	 * state.
640 	 * Not updated directly; only synced in arc_kstat_update.
641 	 */
642 	kstat_named_t arcstat_mfu_evictable_data;
643 	/*
644 	 * Number of bytes consumed by ARC buffers that are eligible for
645 	 * eviction, of type ARC_BUFC_METADATA, and reside in the
646 	 * arc_mfu state.
647 	 * Not updated directly; only synced in arc_kstat_update.
648 	 */
649 	kstat_named_t arcstat_mfu_evictable_metadata;
650 	/*
651 	 * Total number of bytes that *would have been* consumed by ARC
652 	 * buffers in the arc_mfu_ghost state. See the comment above
653 	 * arcstat_mru_ghost_size for more details.
654 	 * Not updated directly; only synced in arc_kstat_update.
655 	 */
656 	kstat_named_t arcstat_mfu_ghost_size;
657 	/*
658 	 * Number of bytes that *would have been* consumed by ARC
659 	 * buffers that are eligible for eviction, of type
660 	 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
661 	 * Not updated directly; only synced in arc_kstat_update.
662 	 */
663 	kstat_named_t arcstat_mfu_ghost_evictable_data;
664 	/*
665 	 * Number of bytes that *would have been* consumed by ARC
666 	 * buffers that are eligible for eviction, of type
667 	 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
668 	 * Not updated directly; only synced in arc_kstat_update.
669 	 */
670 	kstat_named_t arcstat_mfu_ghost_evictable_metadata;
671 	kstat_named_t arcstat_l2_hits;
672 	kstat_named_t arcstat_l2_misses;
673 	kstat_named_t arcstat_l2_feeds;
674 	kstat_named_t arcstat_l2_rw_clash;
675 	kstat_named_t arcstat_l2_read_bytes;
676 	kstat_named_t arcstat_l2_write_bytes;
677 	kstat_named_t arcstat_l2_writes_sent;
678 	kstat_named_t arcstat_l2_writes_done;
679 	kstat_named_t arcstat_l2_writes_error;
680 	kstat_named_t arcstat_l2_writes_lock_retry;
681 	kstat_named_t arcstat_l2_evict_lock_retry;
682 	kstat_named_t arcstat_l2_evict_reading;
683 	kstat_named_t arcstat_l2_evict_l1cached;
684 	kstat_named_t arcstat_l2_free_on_write;
685 	kstat_named_t arcstat_l2_abort_lowmem;
686 	kstat_named_t arcstat_l2_cksum_bad;
687 	kstat_named_t arcstat_l2_io_error;
688 	kstat_named_t arcstat_l2_lsize;
689 	kstat_named_t arcstat_l2_psize;
690 	/* Not updated directly; only synced in arc_kstat_update. */
691 	kstat_named_t arcstat_l2_hdr_size;
692 	kstat_named_t arcstat_memory_throttle_count;
693 	/* Not updated directly; only synced in arc_kstat_update. */
694 	kstat_named_t arcstat_meta_used;
695 	kstat_named_t arcstat_meta_limit;
696 	kstat_named_t arcstat_meta_max;
697 	kstat_named_t arcstat_meta_min;
698 	kstat_named_t arcstat_sync_wait_for_async;
699 	kstat_named_t arcstat_demand_hit_predictive_prefetch;
700 } arc_stats_t;
701 
702 static arc_stats_t arc_stats = {
703 	{ "hits",			KSTAT_DATA_UINT64 },
704 	{ "misses",			KSTAT_DATA_UINT64 },
705 	{ "demand_data_hits",		KSTAT_DATA_UINT64 },
706 	{ "demand_data_misses",		KSTAT_DATA_UINT64 },
707 	{ "demand_metadata_hits",	KSTAT_DATA_UINT64 },
708 	{ "demand_metadata_misses",	KSTAT_DATA_UINT64 },
709 	{ "prefetch_data_hits",		KSTAT_DATA_UINT64 },
710 	{ "prefetch_data_misses",	KSTAT_DATA_UINT64 },
711 	{ "prefetch_metadata_hits",	KSTAT_DATA_UINT64 },
712 	{ "prefetch_metadata_misses",	KSTAT_DATA_UINT64 },
713 	{ "mru_hits",			KSTAT_DATA_UINT64 },
714 	{ "mru_ghost_hits",		KSTAT_DATA_UINT64 },
715 	{ "mfu_hits",			KSTAT_DATA_UINT64 },
716 	{ "mfu_ghost_hits",		KSTAT_DATA_UINT64 },
717 	{ "deleted",			KSTAT_DATA_UINT64 },
718 	{ "mutex_miss",			KSTAT_DATA_UINT64 },
719 	{ "access_skip",		KSTAT_DATA_UINT64 },
720 	{ "evict_skip",			KSTAT_DATA_UINT64 },
721 	{ "evict_not_enough",		KSTAT_DATA_UINT64 },
722 	{ "evict_l2_cached",		KSTAT_DATA_UINT64 },
723 	{ "evict_l2_eligible",		KSTAT_DATA_UINT64 },
724 	{ "evict_l2_ineligible",	KSTAT_DATA_UINT64 },
725 	{ "evict_l2_skip",		KSTAT_DATA_UINT64 },
726 	{ "hash_elements",		KSTAT_DATA_UINT64 },
727 	{ "hash_elements_max",		KSTAT_DATA_UINT64 },
728 	{ "hash_collisions",		KSTAT_DATA_UINT64 },
729 	{ "hash_chains",		KSTAT_DATA_UINT64 },
730 	{ "hash_chain_max",		KSTAT_DATA_UINT64 },
731 	{ "p",				KSTAT_DATA_UINT64 },
732 	{ "c",				KSTAT_DATA_UINT64 },
733 	{ "c_min",			KSTAT_DATA_UINT64 },
734 	{ "c_max",			KSTAT_DATA_UINT64 },
735 	{ "size",			KSTAT_DATA_UINT64 },
736 	{ "compressed_size",		KSTAT_DATA_UINT64 },
737 	{ "uncompressed_size",		KSTAT_DATA_UINT64 },
738 	{ "overhead_size",		KSTAT_DATA_UINT64 },
739 	{ "hdr_size",			KSTAT_DATA_UINT64 },
740 	{ "data_size",			KSTAT_DATA_UINT64 },
741 	{ "metadata_size",		KSTAT_DATA_UINT64 },
742 	{ "other_size",			KSTAT_DATA_UINT64 },
743 	{ "anon_size",			KSTAT_DATA_UINT64 },
744 	{ "anon_evictable_data",	KSTAT_DATA_UINT64 },
745 	{ "anon_evictable_metadata",	KSTAT_DATA_UINT64 },
746 	{ "mru_size",			KSTAT_DATA_UINT64 },
747 	{ "mru_evictable_data",		KSTAT_DATA_UINT64 },
748 	{ "mru_evictable_metadata",	KSTAT_DATA_UINT64 },
749 	{ "mru_ghost_size",		KSTAT_DATA_UINT64 },
750 	{ "mru_ghost_evictable_data",	KSTAT_DATA_UINT64 },
751 	{ "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
752 	{ "mfu_size",			KSTAT_DATA_UINT64 },
753 	{ "mfu_evictable_data",		KSTAT_DATA_UINT64 },
754 	{ "mfu_evictable_metadata",	KSTAT_DATA_UINT64 },
755 	{ "mfu_ghost_size",		KSTAT_DATA_UINT64 },
756 	{ "mfu_ghost_evictable_data",	KSTAT_DATA_UINT64 },
757 	{ "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
758 	{ "l2_hits",			KSTAT_DATA_UINT64 },
759 	{ "l2_misses",			KSTAT_DATA_UINT64 },
760 	{ "l2_feeds",			KSTAT_DATA_UINT64 },
761 	{ "l2_rw_clash",		KSTAT_DATA_UINT64 },
762 	{ "l2_read_bytes",		KSTAT_DATA_UINT64 },
763 	{ "l2_write_bytes",		KSTAT_DATA_UINT64 },
764 	{ "l2_writes_sent",		KSTAT_DATA_UINT64 },
765 	{ "l2_writes_done",		KSTAT_DATA_UINT64 },
766 	{ "l2_writes_error",		KSTAT_DATA_UINT64 },
767 	{ "l2_writes_lock_retry",	KSTAT_DATA_UINT64 },
768 	{ "l2_evict_lock_retry",	KSTAT_DATA_UINT64 },
769 	{ "l2_evict_reading",		KSTAT_DATA_UINT64 },
770 	{ "l2_evict_l1cached",		KSTAT_DATA_UINT64 },
771 	{ "l2_free_on_write",		KSTAT_DATA_UINT64 },
772 	{ "l2_abort_lowmem",		KSTAT_DATA_UINT64 },
773 	{ "l2_cksum_bad",		KSTAT_DATA_UINT64 },
774 	{ "l2_io_error",		KSTAT_DATA_UINT64 },
775 	{ "l2_size",			KSTAT_DATA_UINT64 },
776 	{ "l2_asize",			KSTAT_DATA_UINT64 },
777 	{ "l2_hdr_size",		KSTAT_DATA_UINT64 },
778 	{ "memory_throttle_count",	KSTAT_DATA_UINT64 },
779 	{ "arc_meta_used",		KSTAT_DATA_UINT64 },
780 	{ "arc_meta_limit",		KSTAT_DATA_UINT64 },
781 	{ "arc_meta_max",		KSTAT_DATA_UINT64 },
782 	{ "arc_meta_min",		KSTAT_DATA_UINT64 },
783 	{ "sync_wait_for_async",	KSTAT_DATA_UINT64 },
784 	{ "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
785 };
786 
787 #define	ARCSTAT(stat)	(arc_stats.stat.value.ui64)
788 
789 #define	ARCSTAT_INCR(stat, val) \
790 	atomic_add_64(&arc_stats.stat.value.ui64, (val))
791 
792 #define	ARCSTAT_BUMP(stat)	ARCSTAT_INCR(stat, 1)
793 #define	ARCSTAT_BUMPDOWN(stat)	ARCSTAT_INCR(stat, -1)
794 
795 #define	ARCSTAT_MAX(stat, val) {					\
796 	uint64_t m;							\
797 	while ((val) > (m = arc_stats.stat.value.ui64) &&		\
798 	    (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val))))	\
799 		continue;						\
800 }
801 
802 #define	ARCSTAT_MAXSTAT(stat) \
803 	ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
804 
805 /*
806  * We define a macro to allow ARC hits/misses to be easily broken down by
807  * two separate conditions, giving a total of four different subtypes for
808  * each of hits and misses (so eight statistics total).
809  */
810 #define	ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
811 	if (cond1) {							\
812 		if (cond2) {						\
813 			ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
814 		} else {						\
815 			ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
816 		}							\
817 	} else {							\
818 		if (cond2) {						\
819 			ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
820 		} else {						\
821 			ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
822 		}							\
823 	}
824 
825 kstat_t			*arc_ksp;
826 static arc_state_t	*arc_anon;
827 static arc_state_t	*arc_mru;
828 static arc_state_t	*arc_mru_ghost;
829 static arc_state_t	*arc_mfu;
830 static arc_state_t	*arc_mfu_ghost;
831 static arc_state_t	*arc_l2c_only;
832 
833 /*
834  * There are several ARC variables that are critical to export as kstats --
835  * but we don't want to have to grovel around in the kstat whenever we wish to
836  * manipulate them.  For these variables, we therefore define them to be in
837  * terms of the statistic variable.  This assures that we are not introducing
838  * the possibility of inconsistency by having shadow copies of the variables,
839  * while still allowing the code to be readable.
840  */
841 #define	arc_p		ARCSTAT(arcstat_p)	/* target size of MRU */
842 #define	arc_c		ARCSTAT(arcstat_c)	/* target size of cache */
843 #define	arc_c_min	ARCSTAT(arcstat_c_min)	/* min target cache size */
844 #define	arc_c_max	ARCSTAT(arcstat_c_max)	/* max target cache size */
845 #define	arc_meta_limit	ARCSTAT(arcstat_meta_limit) /* max size for metadata */
846 #define	arc_meta_min	ARCSTAT(arcstat_meta_min) /* min size for metadata */
847 #define	arc_meta_max	ARCSTAT(arcstat_meta_max) /* max size of metadata */
848 
849 /* compressed size of entire arc */
850 #define	arc_compressed_size	ARCSTAT(arcstat_compressed_size)
851 /* uncompressed size of entire arc */
852 #define	arc_uncompressed_size	ARCSTAT(arcstat_uncompressed_size)
853 /* number of bytes in the arc from arc_buf_t's */
854 #define	arc_overhead_size	ARCSTAT(arcstat_overhead_size)
855 
856 /*
857  * There are also some ARC variables that we want to export, but that are
858  * updated so often that having the canonical representation be the statistic
859  * variable causes a performance bottleneck. We want to use aggsum_t's for these
860  * instead, but still be able to export the kstat in the same way as before.
861  * The solution is to always use the aggsum version, except in the kstat update
862  * callback.
863  */
864 aggsum_t arc_size;
865 aggsum_t arc_meta_used;
866 aggsum_t astat_data_size;
867 aggsum_t astat_metadata_size;
868 aggsum_t astat_hdr_size;
869 aggsum_t astat_other_size;
870 aggsum_t astat_l2_hdr_size;
871 
872 static int		arc_no_grow;	/* Don't try to grow cache size */
873 static hrtime_t		arc_growtime;
874 static uint64_t		arc_tempreserve;
875 static uint64_t		arc_loaned_bytes;
876 
877 typedef struct arc_callback arc_callback_t;
878 
879 struct arc_callback {
880 	void			*acb_private;
881 	arc_done_func_t		*acb_done;
882 	arc_buf_t		*acb_buf;
883 	boolean_t		acb_compressed;
884 	zio_t			*acb_zio_dummy;
885 	arc_callback_t		*acb_next;
886 };
887 
888 typedef struct arc_write_callback arc_write_callback_t;
889 
890 struct arc_write_callback {
891 	void		*awcb_private;
892 	arc_done_func_t	*awcb_ready;
893 	arc_done_func_t	*awcb_children_ready;
894 	arc_done_func_t	*awcb_physdone;
895 	arc_done_func_t	*awcb_done;
896 	arc_buf_t	*awcb_buf;
897 };
898 
899 /*
900  * ARC buffers are separated into multiple structs as a memory saving measure:
901  *   - Common fields struct, always defined, and embedded within it:
902  *       - L2-only fields, always allocated but undefined when not in L2ARC
903  *       - L1-only fields, only allocated when in L1ARC
904  *
905  *           Buffer in L1                     Buffer only in L2
906  *    +------------------------+          +------------------------+
907  *    | arc_buf_hdr_t          |          | arc_buf_hdr_t          |
908  *    |                        |          |                        |
909  *    |                        |          |                        |
910  *    |                        |          |                        |
911  *    +------------------------+          +------------------------+
912  *    | l2arc_buf_hdr_t        |          | l2arc_buf_hdr_t        |
913  *    | (undefined if L1-only) |          |                        |
914  *    +------------------------+          +------------------------+
915  *    | l1arc_buf_hdr_t        |
916  *    |                        |
917  *    |                        |
918  *    |                        |
919  *    |                        |
920  *    +------------------------+
921  *
922  * Because it's possible for the L2ARC to become extremely large, we can wind
923  * up eating a lot of memory in L2ARC buffer headers, so the size of a header
924  * is minimized by only allocating the fields necessary for an L1-cached buffer
925  * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
926  * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
927  * words in pointers. arc_hdr_realloc() is used to switch a header between
928  * these two allocation states.
929  */
930 typedef struct l1arc_buf_hdr {
931 	kmutex_t		b_freeze_lock;
932 	zio_cksum_t		*b_freeze_cksum;
933 #ifdef ZFS_DEBUG
934 	/*
935 	 * Used for debugging with kmem_flags - by allocating and freeing
936 	 * b_thawed when the buffer is thawed, we get a record of the stack
937 	 * trace that thawed it.
938 	 */
939 	void			*b_thawed;
940 #endif
941 
942 	arc_buf_t		*b_buf;
943 	uint32_t		b_bufcnt;
944 	/* for waiting on writes to complete */
945 	kcondvar_t		b_cv;
946 	uint8_t			b_byteswap;
947 
948 	/* protected by arc state mutex */
949 	arc_state_t		*b_state;
950 	multilist_node_t	b_arc_node;
951 
952 	/* updated atomically */
953 	clock_t			b_arc_access;
954 
955 	/* self protecting */
956 	zfs_refcount_t		b_refcnt;
957 
958 	arc_callback_t		*b_acb;
959 	abd_t			*b_pabd;
960 } l1arc_buf_hdr_t;
961 
962 typedef struct l2arc_dev l2arc_dev_t;
963 
964 typedef struct l2arc_buf_hdr {
965 	/* protected by arc_buf_hdr mutex */
966 	l2arc_dev_t		*b_dev;		/* L2ARC device */
967 	uint64_t		b_daddr;	/* disk address, offset byte */
968 
969 	list_node_t		b_l2node;
970 } l2arc_buf_hdr_t;
971 
972 struct arc_buf_hdr {
973 	/* protected by hash lock */
974 	dva_t			b_dva;
975 	uint64_t		b_birth;
976 
977 	arc_buf_contents_t	b_type;
978 	arc_buf_hdr_t		*b_hash_next;
979 	arc_flags_t		b_flags;
980 
981 	/*
982 	 * This field stores the size of the data buffer after
983 	 * compression, and is set in the arc's zio completion handlers.
984 	 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
985 	 *
986 	 * While the block pointers can store up to 32MB in their psize
987 	 * field, we can only store up to 32MB minus 512B. This is due
988 	 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
989 	 * a field of zeros represents 512B in the bp). We can't use a
990 	 * bias of 1 since we need to reserve a psize of zero, here, to
991 	 * represent holes and embedded blocks.
992 	 *
993 	 * This isn't a problem in practice, since the maximum size of a
994 	 * buffer is limited to 16MB, so we never need to store 32MB in
995 	 * this field. Even in the upstream illumos code base, the
996 	 * maximum size of a buffer is limited to 16MB.
997 	 */
998 	uint16_t		b_psize;
999 
1000 	/*
1001 	 * This field stores the size of the data buffer before
1002 	 * compression, and cannot change once set. It is in units
1003 	 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
1004 	 */
1005 	uint16_t		b_lsize;	/* immutable */
1006 	uint64_t		b_spa;		/* immutable */
1007 
1008 	/* L2ARC fields. Undefined when not in L2ARC. */
1009 	l2arc_buf_hdr_t		b_l2hdr;
1010 	/* L1ARC fields. Undefined when in l2arc_only state */
1011 	l1arc_buf_hdr_t		b_l1hdr;
1012 };
1013 
1014 #define	GHOST_STATE(state)	\
1015 	((state) == arc_mru_ghost || (state) == arc_mfu_ghost ||	\
1016 	(state) == arc_l2c_only)
1017 
1018 #define	HDR_IN_HASH_TABLE(hdr)	((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1019 #define	HDR_IO_IN_PROGRESS(hdr)	((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1020 #define	HDR_IO_ERROR(hdr)	((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1021 #define	HDR_PREFETCH(hdr)	((hdr)->b_flags & ARC_FLAG_PREFETCH)
1022 #define	HDR_COMPRESSION_ENABLED(hdr)	\
1023 	((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1024 
1025 #define	HDR_L2CACHE(hdr)	((hdr)->b_flags & ARC_FLAG_L2CACHE)
1026 #define	HDR_L2_READING(hdr)	\
1027 	(((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) &&	\
1028 	((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1029 #define	HDR_L2_WRITING(hdr)	((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1030 #define	HDR_L2_EVICTED(hdr)	((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1031 #define	HDR_L2_WRITE_HEAD(hdr)	((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1032 #define	HDR_SHARED_DATA(hdr)	((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1033 
1034 #define	HDR_ISTYPE_METADATA(hdr)	\
1035 	((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1036 #define	HDR_ISTYPE_DATA(hdr)	(!HDR_ISTYPE_METADATA(hdr))
1037 
1038 #define	HDR_HAS_L1HDR(hdr)	((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1039 #define	HDR_HAS_L2HDR(hdr)	((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1040 
1041 /* For storing compression mode in b_flags */
1042 #define	HDR_COMPRESS_OFFSET	(highbit64(ARC_FLAG_COMPRESS_0) - 1)
1043 
1044 #define	HDR_GET_COMPRESS(hdr)	((enum zio_compress)BF32_GET((hdr)->b_flags, \
1045 	HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1046 #define	HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1047 	HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1048 
1049 #define	ARC_BUF_LAST(buf)	((buf)->b_next == NULL)
1050 #define	ARC_BUF_SHARED(buf)	((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1051 #define	ARC_BUF_COMPRESSED(buf)	((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1052 
1053 /*
1054  * Other sizes
1055  */
1056 
1057 #define	HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1058 #define	HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1059 
1060 /*
1061  * Hash table routines
1062  */
1063 
1064 #define	HT_LOCK_PAD	64
1065 
1066 struct ht_lock {
1067 	kmutex_t	ht_lock;
1068 #ifdef _KERNEL
1069 	unsigned char	pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1070 #endif
1071 };
1072 
1073 #define	BUF_LOCKS 256
1074 typedef struct buf_hash_table {
1075 	uint64_t ht_mask;
1076 	arc_buf_hdr_t **ht_table;
1077 	struct ht_lock ht_locks[BUF_LOCKS];
1078 } buf_hash_table_t;
1079 
1080 static buf_hash_table_t buf_hash_table;
1081 
1082 #define	BUF_HASH_INDEX(spa, dva, birth) \
1083 	(buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1084 #define	BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1085 #define	BUF_HASH_LOCK(idx)	(&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1086 #define	HDR_LOCK(hdr) \
1087 	(BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1088 
1089 uint64_t zfs_crc64_table[256];
1090 
1091 /*
1092  * Level 2 ARC
1093  */
1094 
1095 #define	L2ARC_WRITE_SIZE	(8 * 1024 * 1024)	/* initial write max */
1096 #define	L2ARC_HEADROOM		2			/* num of writes */
1097 /*
1098  * If we discover during ARC scan any buffers to be compressed, we boost
1099  * our headroom for the next scanning cycle by this percentage multiple.
1100  */
1101 #define	L2ARC_HEADROOM_BOOST	200
1102 #define	L2ARC_FEED_SECS		1		/* caching interval secs */
1103 #define	L2ARC_FEED_MIN_MS	200		/* min caching interval ms */
1104 
1105 #define	l2arc_writes_sent	ARCSTAT(arcstat_l2_writes_sent)
1106 #define	l2arc_writes_done	ARCSTAT(arcstat_l2_writes_done)
1107 
1108 /* L2ARC Performance Tunables */
1109 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE;	/* default max write size */
1110 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE;	/* extra write during warmup */
1111 uint64_t l2arc_headroom = L2ARC_HEADROOM;	/* number of dev writes */
1112 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1113 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS;	/* interval seconds */
1114 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS;	/* min interval milliseconds */
1115 boolean_t l2arc_noprefetch = B_TRUE;		/* don't cache prefetch bufs */
1116 boolean_t l2arc_feed_again = B_TRUE;		/* turbo warmup */
1117 boolean_t l2arc_norw = B_TRUE;			/* no reads during writes */
1118 
1119 /*
1120  * L2ARC Internals
1121  */
1122 struct l2arc_dev {
1123 	vdev_t			*l2ad_vdev;	/* vdev */
1124 	spa_t			*l2ad_spa;	/* spa */
1125 	uint64_t		l2ad_hand;	/* next write location */
1126 	uint64_t		l2ad_start;	/* first addr on device */
1127 	uint64_t		l2ad_end;	/* last addr on device */
1128 	boolean_t		l2ad_first;	/* first sweep through */
1129 	boolean_t		l2ad_writing;	/* currently writing */
1130 	kmutex_t		l2ad_mtx;	/* lock for buffer list */
1131 	list_t			l2ad_buflist;	/* buffer list */
1132 	list_node_t		l2ad_node;	/* device list node */
1133 	zfs_refcount_t		l2ad_alloc;	/* allocated bytes */
1134 };
1135 
1136 static list_t L2ARC_dev_list;			/* device list */
1137 static list_t *l2arc_dev_list;			/* device list pointer */
1138 static kmutex_t l2arc_dev_mtx;			/* device list mutex */
1139 static l2arc_dev_t *l2arc_dev_last;		/* last device used */
1140 static list_t L2ARC_free_on_write;		/* free after write buf list */
1141 static list_t *l2arc_free_on_write;		/* free after write list ptr */
1142 static kmutex_t l2arc_free_on_write_mtx;	/* mutex for list */
1143 static uint64_t l2arc_ndev;			/* number of devices */
1144 
1145 typedef struct l2arc_read_callback {
1146 	arc_buf_hdr_t		*l2rcb_hdr;		/* read header */
1147 	blkptr_t		l2rcb_bp;		/* original blkptr */
1148 	zbookmark_phys_t	l2rcb_zb;		/* original bookmark */
1149 	int			l2rcb_flags;		/* original flags */
1150 	abd_t			*l2rcb_abd;		/* temporary buffer */
1151 } l2arc_read_callback_t;
1152 
1153 typedef struct l2arc_write_callback {
1154 	l2arc_dev_t	*l2wcb_dev;		/* device info */
1155 	arc_buf_hdr_t	*l2wcb_head;		/* head of write buflist */
1156 } l2arc_write_callback_t;
1157 
1158 typedef struct l2arc_data_free {
1159 	/* protected by l2arc_free_on_write_mtx */
1160 	abd_t		*l2df_abd;
1161 	size_t		l2df_size;
1162 	arc_buf_contents_t l2df_type;
1163 	list_node_t	l2df_list_node;
1164 } l2arc_data_free_t;
1165 
1166 static kmutex_t l2arc_feed_thr_lock;
1167 static kcondvar_t l2arc_feed_thr_cv;
1168 static uint8_t l2arc_thread_exit;
1169 
1170 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1171 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1172 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1173 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1174 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1175 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1176 static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1177 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *);
1178 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1179 static boolean_t arc_is_overflowing();
1180 static void arc_buf_watch(arc_buf_t *);
1181 
1182 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1183 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1184 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1185 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1186 
1187 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1188 static void l2arc_read_done(zio_t *);
1189 
1190 
1191 /*
1192  * We use Cityhash for this. It's fast, and has good hash properties without
1193  * requiring any large static buffers.
1194  */
1195 static uint64_t
1196 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1197 {
1198 	return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
1199 }
1200 
1201 #define	HDR_EMPTY(hdr)						\
1202 	((hdr)->b_dva.dva_word[0] == 0 &&			\
1203 	(hdr)->b_dva.dva_word[1] == 0)
1204 
1205 #define	HDR_EQUAL(spa, dva, birth, hdr)				\
1206 	((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) &&	\
1207 	((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) &&	\
1208 	((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1209 
1210 static void
1211 buf_discard_identity(arc_buf_hdr_t *hdr)
1212 {
1213 	hdr->b_dva.dva_word[0] = 0;
1214 	hdr->b_dva.dva_word[1] = 0;
1215 	hdr->b_birth = 0;
1216 }
1217 
1218 static arc_buf_hdr_t *
1219 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1220 {
1221 	const dva_t *dva = BP_IDENTITY(bp);
1222 	uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1223 	uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1224 	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1225 	arc_buf_hdr_t *hdr;
1226 
1227 	mutex_enter(hash_lock);
1228 	for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1229 	    hdr = hdr->b_hash_next) {
1230 		if (HDR_EQUAL(spa, dva, birth, hdr)) {
1231 			*lockp = hash_lock;
1232 			return (hdr);
1233 		}
1234 	}
1235 	mutex_exit(hash_lock);
1236 	*lockp = NULL;
1237 	return (NULL);
1238 }
1239 
1240 /*
1241  * Insert an entry into the hash table.  If there is already an element
1242  * equal to elem in the hash table, then the already existing element
1243  * will be returned and the new element will not be inserted.
1244  * Otherwise returns NULL.
1245  * If lockp == NULL, the caller is assumed to already hold the hash lock.
1246  */
1247 static arc_buf_hdr_t *
1248 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1249 {
1250 	uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1251 	kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1252 	arc_buf_hdr_t *fhdr;
1253 	uint32_t i;
1254 
1255 	ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1256 	ASSERT(hdr->b_birth != 0);
1257 	ASSERT(!HDR_IN_HASH_TABLE(hdr));
1258 
1259 	if (lockp != NULL) {
1260 		*lockp = hash_lock;
1261 		mutex_enter(hash_lock);
1262 	} else {
1263 		ASSERT(MUTEX_HELD(hash_lock));
1264 	}
1265 
1266 	for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1267 	    fhdr = fhdr->b_hash_next, i++) {
1268 		if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1269 			return (fhdr);
1270 	}
1271 
1272 	hdr->b_hash_next = buf_hash_table.ht_table[idx];
1273 	buf_hash_table.ht_table[idx] = hdr;
1274 	arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1275 
1276 	/* collect some hash table performance data */
1277 	if (i > 0) {
1278 		ARCSTAT_BUMP(arcstat_hash_collisions);
1279 		if (i == 1)
1280 			ARCSTAT_BUMP(arcstat_hash_chains);
1281 
1282 		ARCSTAT_MAX(arcstat_hash_chain_max, i);
1283 	}
1284 
1285 	ARCSTAT_BUMP(arcstat_hash_elements);
1286 	ARCSTAT_MAXSTAT(arcstat_hash_elements);
1287 
1288 	return (NULL);
1289 }
1290 
1291 static void
1292 buf_hash_remove(arc_buf_hdr_t *hdr)
1293 {
1294 	arc_buf_hdr_t *fhdr, **hdrp;
1295 	uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1296 
1297 	ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1298 	ASSERT(HDR_IN_HASH_TABLE(hdr));
1299 
1300 	hdrp = &buf_hash_table.ht_table[idx];
1301 	while ((fhdr = *hdrp) != hdr) {
1302 		ASSERT3P(fhdr, !=, NULL);
1303 		hdrp = &fhdr->b_hash_next;
1304 	}
1305 	*hdrp = hdr->b_hash_next;
1306 	hdr->b_hash_next = NULL;
1307 	arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1308 
1309 	/* collect some hash table performance data */
1310 	ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1311 
1312 	if (buf_hash_table.ht_table[idx] &&
1313 	    buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1314 		ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1315 }
1316 
1317 /*
1318  * Global data structures and functions for the buf kmem cache.
1319  */
1320 static kmem_cache_t *hdr_full_cache;
1321 static kmem_cache_t *hdr_l2only_cache;
1322 static kmem_cache_t *buf_cache;
1323 
1324 static void
1325 buf_fini(void)
1326 {
1327 	int i;
1328 
1329 	kmem_free(buf_hash_table.ht_table,
1330 	    (buf_hash_table.ht_mask + 1) * sizeof (void *));
1331 	for (i = 0; i < BUF_LOCKS; i++)
1332 		mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1333 	kmem_cache_destroy(hdr_full_cache);
1334 	kmem_cache_destroy(hdr_l2only_cache);
1335 	kmem_cache_destroy(buf_cache);
1336 }
1337 
1338 /*
1339  * Constructor callback - called when the cache is empty
1340  * and a new buf is requested.
1341  */
1342 /* ARGSUSED */
1343 static int
1344 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1345 {
1346 	arc_buf_hdr_t *hdr = vbuf;
1347 
1348 	bzero(hdr, HDR_FULL_SIZE);
1349 	cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1350 	zfs_refcount_create(&hdr->b_l1hdr.b_refcnt);
1351 	mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1352 	multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1353 	arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1354 
1355 	return (0);
1356 }
1357 
1358 /* ARGSUSED */
1359 static int
1360 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1361 {
1362 	arc_buf_hdr_t *hdr = vbuf;
1363 
1364 	bzero(hdr, HDR_L2ONLY_SIZE);
1365 	arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1366 
1367 	return (0);
1368 }
1369 
1370 /* ARGSUSED */
1371 static int
1372 buf_cons(void *vbuf, void *unused, int kmflag)
1373 {
1374 	arc_buf_t *buf = vbuf;
1375 
1376 	bzero(buf, sizeof (arc_buf_t));
1377 	mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1378 	arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1379 
1380 	return (0);
1381 }
1382 
1383 /*
1384  * Destructor callback - called when a cached buf is
1385  * no longer required.
1386  */
1387 /* ARGSUSED */
1388 static void
1389 hdr_full_dest(void *vbuf, void *unused)
1390 {
1391 	arc_buf_hdr_t *hdr = vbuf;
1392 
1393 	ASSERT(HDR_EMPTY(hdr));
1394 	cv_destroy(&hdr->b_l1hdr.b_cv);
1395 	zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1396 	mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1397 	ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1398 	arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1399 }
1400 
1401 /* ARGSUSED */
1402 static void
1403 hdr_l2only_dest(void *vbuf, void *unused)
1404 {
1405 	arc_buf_hdr_t *hdr = vbuf;
1406 
1407 	ASSERT(HDR_EMPTY(hdr));
1408 	arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1409 }
1410 
1411 /* ARGSUSED */
1412 static void
1413 buf_dest(void *vbuf, void *unused)
1414 {
1415 	arc_buf_t *buf = vbuf;
1416 
1417 	mutex_destroy(&buf->b_evict_lock);
1418 	arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1419 }
1420 
1421 /*
1422  * Reclaim callback -- invoked when memory is low.
1423  */
1424 /* ARGSUSED */
1425 static void
1426 hdr_recl(void *unused)
1427 {
1428 	dprintf("hdr_recl called\n");
1429 	/*
1430 	 * umem calls the reclaim func when we destroy the buf cache,
1431 	 * which is after we do arc_fini().
1432 	 */
1433 	if (arc_initialized)
1434 		zthr_wakeup(arc_reap_zthr);
1435 }
1436 
1437 static void
1438 buf_init(void)
1439 {
1440 	uint64_t *ct;
1441 	uint64_t hsize = 1ULL << 12;
1442 	int i, j;
1443 
1444 	/*
1445 	 * The hash table is big enough to fill all of physical memory
1446 	 * with an average block size of zfs_arc_average_blocksize (default 8K).
1447 	 * By default, the table will take up
1448 	 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1449 	 */
1450 	while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
1451 		hsize <<= 1;
1452 retry:
1453 	buf_hash_table.ht_mask = hsize - 1;
1454 	buf_hash_table.ht_table =
1455 	    kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1456 	if (buf_hash_table.ht_table == NULL) {
1457 		ASSERT(hsize > (1ULL << 8));
1458 		hsize >>= 1;
1459 		goto retry;
1460 	}
1461 
1462 	hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1463 	    0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1464 	hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1465 	    HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1466 	    NULL, NULL, 0);
1467 	buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1468 	    0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1469 
1470 	for (i = 0; i < 256; i++)
1471 		for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1472 			*ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1473 
1474 	for (i = 0; i < BUF_LOCKS; i++) {
1475 		mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1476 		    NULL, MUTEX_DEFAULT, NULL);
1477 	}
1478 }
1479 
1480 /*
1481  * This is the size that the buf occupies in memory. If the buf is compressed,
1482  * it will correspond to the compressed size. You should use this method of
1483  * getting the buf size unless you explicitly need the logical size.
1484  */
1485 int32_t
1486 arc_buf_size(arc_buf_t *buf)
1487 {
1488 	return (ARC_BUF_COMPRESSED(buf) ?
1489 	    HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1490 }
1491 
1492 int32_t
1493 arc_buf_lsize(arc_buf_t *buf)
1494 {
1495 	return (HDR_GET_LSIZE(buf->b_hdr));
1496 }
1497 
1498 enum zio_compress
1499 arc_get_compression(arc_buf_t *buf)
1500 {
1501 	return (ARC_BUF_COMPRESSED(buf) ?
1502 	    HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1503 }
1504 
1505 #define	ARC_MINTIME	(hz>>4) /* 62 ms */
1506 
1507 static inline boolean_t
1508 arc_buf_is_shared(arc_buf_t *buf)
1509 {
1510 	boolean_t shared = (buf->b_data != NULL &&
1511 	    buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1512 	    abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1513 	    buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1514 	IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1515 	IMPLY(shared, ARC_BUF_SHARED(buf));
1516 	IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1517 
1518 	/*
1519 	 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1520 	 * already being shared" requirement prevents us from doing that.
1521 	 */
1522 
1523 	return (shared);
1524 }
1525 
1526 /*
1527  * Free the checksum associated with this header. If there is no checksum, this
1528  * is a no-op.
1529  */
1530 static inline void
1531 arc_cksum_free(arc_buf_hdr_t *hdr)
1532 {
1533 	ASSERT(HDR_HAS_L1HDR(hdr));
1534 	mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1535 	if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1536 		kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1537 		hdr->b_l1hdr.b_freeze_cksum = NULL;
1538 	}
1539 	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1540 }
1541 
1542 /*
1543  * Return true iff at least one of the bufs on hdr is not compressed.
1544  */
1545 static boolean_t
1546 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1547 {
1548 	for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1549 		if (!ARC_BUF_COMPRESSED(b)) {
1550 			return (B_TRUE);
1551 		}
1552 	}
1553 	return (B_FALSE);
1554 }
1555 
1556 /*
1557  * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1558  * matches the checksum that is stored in the hdr. If there is no checksum,
1559  * or if the buf is compressed, this is a no-op.
1560  */
1561 static void
1562 arc_cksum_verify(arc_buf_t *buf)
1563 {
1564 	arc_buf_hdr_t *hdr = buf->b_hdr;
1565 	zio_cksum_t zc;
1566 
1567 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1568 		return;
1569 
1570 	if (ARC_BUF_COMPRESSED(buf)) {
1571 		ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1572 		    arc_hdr_has_uncompressed_buf(hdr));
1573 		return;
1574 	}
1575 
1576 	ASSERT(HDR_HAS_L1HDR(hdr));
1577 
1578 	mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1579 	if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1580 		mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1581 		return;
1582 	}
1583 
1584 	fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1585 	if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1586 		panic("buffer modified while frozen!");
1587 	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1588 }
1589 
1590 static boolean_t
1591 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1592 {
1593 	enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1594 	boolean_t valid_cksum;
1595 
1596 	ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1597 	VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1598 
1599 	/*
1600 	 * We rely on the blkptr's checksum to determine if the block
1601 	 * is valid or not. When compressed arc is enabled, the l2arc
1602 	 * writes the block to the l2arc just as it appears in the pool.
1603 	 * This allows us to use the blkptr's checksum to validate the
1604 	 * data that we just read off of the l2arc without having to store
1605 	 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1606 	 * arc is disabled, then the data written to the l2arc is always
1607 	 * uncompressed and won't match the block as it exists in the main
1608 	 * pool. When this is the case, we must first compress it if it is
1609 	 * compressed on the main pool before we can validate the checksum.
1610 	 */
1611 	if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1612 		ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1613 		uint64_t lsize = HDR_GET_LSIZE(hdr);
1614 		uint64_t csize;
1615 
1616 		abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
1617 		csize = zio_compress_data(compress, zio->io_abd,
1618 		    abd_to_buf(cdata), lsize);
1619 
1620 		ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1621 		if (csize < HDR_GET_PSIZE(hdr)) {
1622 			/*
1623 			 * Compressed blocks are always a multiple of the
1624 			 * smallest ashift in the pool. Ideally, we would
1625 			 * like to round up the csize to the next
1626 			 * spa_min_ashift but that value may have changed
1627 			 * since the block was last written. Instead,
1628 			 * we rely on the fact that the hdr's psize
1629 			 * was set to the psize of the block when it was
1630 			 * last written. We set the csize to that value
1631 			 * and zero out any part that should not contain
1632 			 * data.
1633 			 */
1634 			abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
1635 			csize = HDR_GET_PSIZE(hdr);
1636 		}
1637 		zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
1638 	}
1639 
1640 	/*
1641 	 * Block pointers always store the checksum for the logical data.
1642 	 * If the block pointer has the gang bit set, then the checksum
1643 	 * it represents is for the reconstituted data and not for an
1644 	 * individual gang member. The zio pipeline, however, must be able to
1645 	 * determine the checksum of each of the gang constituents so it
1646 	 * treats the checksum comparison differently than what we need
1647 	 * for l2arc blocks. This prevents us from using the
1648 	 * zio_checksum_error() interface directly. Instead we must call the
1649 	 * zio_checksum_error_impl() so that we can ensure the checksum is
1650 	 * generated using the correct checksum algorithm and accounts for the
1651 	 * logical I/O size and not just a gang fragment.
1652 	 */
1653 	valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1654 	    BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1655 	    zio->io_offset, NULL) == 0);
1656 	zio_pop_transforms(zio);
1657 	return (valid_cksum);
1658 }
1659 
1660 /*
1661  * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1662  * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1663  * isn't modified later on. If buf is compressed or there is already a checksum
1664  * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1665  */
1666 static void
1667 arc_cksum_compute(arc_buf_t *buf)
1668 {
1669 	arc_buf_hdr_t *hdr = buf->b_hdr;
1670 
1671 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1672 		return;
1673 
1674 	ASSERT(HDR_HAS_L1HDR(hdr));
1675 
1676 	mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1677 	if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1678 		ASSERT(arc_hdr_has_uncompressed_buf(hdr));
1679 		mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1680 		return;
1681 	} else if (ARC_BUF_COMPRESSED(buf)) {
1682 		mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1683 		return;
1684 	}
1685 
1686 	ASSERT(!ARC_BUF_COMPRESSED(buf));
1687 	hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1688 	    KM_SLEEP);
1689 	fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1690 	    hdr->b_l1hdr.b_freeze_cksum);
1691 	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1692 	arc_buf_watch(buf);
1693 }
1694 
1695 #ifndef _KERNEL
1696 typedef struct procctl {
1697 	long cmd;
1698 	prwatch_t prwatch;
1699 } procctl_t;
1700 #endif
1701 
1702 /* ARGSUSED */
1703 static void
1704 arc_buf_unwatch(arc_buf_t *buf)
1705 {
1706 #ifndef _KERNEL
1707 	if (arc_watch) {
1708 		int result;
1709 		procctl_t ctl;
1710 		ctl.cmd = PCWATCH;
1711 		ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1712 		ctl.prwatch.pr_size = 0;
1713 		ctl.prwatch.pr_wflags = 0;
1714 		result = write(arc_procfd, &ctl, sizeof (ctl));
1715 		ASSERT3U(result, ==, sizeof (ctl));
1716 	}
1717 #endif
1718 }
1719 
1720 /* ARGSUSED */
1721 static void
1722 arc_buf_watch(arc_buf_t *buf)
1723 {
1724 #ifndef _KERNEL
1725 	if (arc_watch) {
1726 		int result;
1727 		procctl_t ctl;
1728 		ctl.cmd = PCWATCH;
1729 		ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1730 		ctl.prwatch.pr_size = arc_buf_size(buf);
1731 		ctl.prwatch.pr_wflags = WA_WRITE;
1732 		result = write(arc_procfd, &ctl, sizeof (ctl));
1733 		ASSERT3U(result, ==, sizeof (ctl));
1734 	}
1735 #endif
1736 }
1737 
1738 static arc_buf_contents_t
1739 arc_buf_type(arc_buf_hdr_t *hdr)
1740 {
1741 	arc_buf_contents_t type;
1742 	if (HDR_ISTYPE_METADATA(hdr)) {
1743 		type = ARC_BUFC_METADATA;
1744 	} else {
1745 		type = ARC_BUFC_DATA;
1746 	}
1747 	VERIFY3U(hdr->b_type, ==, type);
1748 	return (type);
1749 }
1750 
1751 boolean_t
1752 arc_is_metadata(arc_buf_t *buf)
1753 {
1754 	return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1755 }
1756 
1757 static uint32_t
1758 arc_bufc_to_flags(arc_buf_contents_t type)
1759 {
1760 	switch (type) {
1761 	case ARC_BUFC_DATA:
1762 		/* metadata field is 0 if buffer contains normal data */
1763 		return (0);
1764 	case ARC_BUFC_METADATA:
1765 		return (ARC_FLAG_BUFC_METADATA);
1766 	default:
1767 		break;
1768 	}
1769 	panic("undefined ARC buffer type!");
1770 	return ((uint32_t)-1);
1771 }
1772 
1773 void
1774 arc_buf_thaw(arc_buf_t *buf)
1775 {
1776 	arc_buf_hdr_t *hdr = buf->b_hdr;
1777 
1778 	ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1779 	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1780 
1781 	arc_cksum_verify(buf);
1782 
1783 	/*
1784 	 * Compressed buffers do not manipulate the b_freeze_cksum or
1785 	 * allocate b_thawed.
1786 	 */
1787 	if (ARC_BUF_COMPRESSED(buf)) {
1788 		ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1789 		    arc_hdr_has_uncompressed_buf(hdr));
1790 		return;
1791 	}
1792 
1793 	ASSERT(HDR_HAS_L1HDR(hdr));
1794 	arc_cksum_free(hdr);
1795 
1796 	mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1797 #ifdef ZFS_DEBUG
1798 	if (zfs_flags & ZFS_DEBUG_MODIFY) {
1799 		if (hdr->b_l1hdr.b_thawed != NULL)
1800 			kmem_free(hdr->b_l1hdr.b_thawed, 1);
1801 		hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1802 	}
1803 #endif
1804 
1805 	mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1806 
1807 	arc_buf_unwatch(buf);
1808 }
1809 
1810 void
1811 arc_buf_freeze(arc_buf_t *buf)
1812 {
1813 	arc_buf_hdr_t *hdr = buf->b_hdr;
1814 	kmutex_t *hash_lock;
1815 
1816 	if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1817 		return;
1818 
1819 	if (ARC_BUF_COMPRESSED(buf)) {
1820 		ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1821 		    arc_hdr_has_uncompressed_buf(hdr));
1822 		return;
1823 	}
1824 
1825 	hash_lock = HDR_LOCK(hdr);
1826 	mutex_enter(hash_lock);
1827 
1828 	ASSERT(HDR_HAS_L1HDR(hdr));
1829 	ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
1830 	    hdr->b_l1hdr.b_state == arc_anon);
1831 	arc_cksum_compute(buf);
1832 	mutex_exit(hash_lock);
1833 }
1834 
1835 /*
1836  * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1837  * the following functions should be used to ensure that the flags are
1838  * updated in a thread-safe way. When manipulating the flags either
1839  * the hash_lock must be held or the hdr must be undiscoverable. This
1840  * ensures that we're not racing with any other threads when updating
1841  * the flags.
1842  */
1843 static inline void
1844 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1845 {
1846 	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1847 	hdr->b_flags |= flags;
1848 }
1849 
1850 static inline void
1851 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1852 {
1853 	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1854 	hdr->b_flags &= ~flags;
1855 }
1856 
1857 /*
1858  * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1859  * done in a special way since we have to clear and set bits
1860  * at the same time. Consumers that wish to set the compression bits
1861  * must use this function to ensure that the flags are updated in
1862  * thread-safe manner.
1863  */
1864 static void
1865 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1866 {
1867 	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1868 
1869 	/*
1870 	 * Holes and embedded blocks will always have a psize = 0 so
1871 	 * we ignore the compression of the blkptr and set the
1872 	 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
1873 	 * Holes and embedded blocks remain anonymous so we don't
1874 	 * want to uncompress them. Mark them as uncompressed.
1875 	 */
1876 	if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1877 		arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1878 		HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
1879 		ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1880 		ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1881 	} else {
1882 		arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1883 		HDR_SET_COMPRESS(hdr, cmp);
1884 		ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1885 		ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1886 	}
1887 }
1888 
1889 /*
1890  * Looks for another buf on the same hdr which has the data decompressed, copies
1891  * from it, and returns true. If no such buf exists, returns false.
1892  */
1893 static boolean_t
1894 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1895 {
1896 	arc_buf_hdr_t *hdr = buf->b_hdr;
1897 	boolean_t copied = B_FALSE;
1898 
1899 	ASSERT(HDR_HAS_L1HDR(hdr));
1900 	ASSERT3P(buf->b_data, !=, NULL);
1901 	ASSERT(!ARC_BUF_COMPRESSED(buf));
1902 
1903 	for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1904 	    from = from->b_next) {
1905 		/* can't use our own data buffer */
1906 		if (from == buf) {
1907 			continue;
1908 		}
1909 
1910 		if (!ARC_BUF_COMPRESSED(from)) {
1911 			bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
1912 			copied = B_TRUE;
1913 			break;
1914 		}
1915 	}
1916 
1917 	/*
1918 	 * There were no decompressed bufs, so there should not be a
1919 	 * checksum on the hdr either.
1920 	 */
1921 	EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1922 
1923 	return (copied);
1924 }
1925 
1926 /*
1927  * Given a buf that has a data buffer attached to it, this function will
1928  * efficiently fill the buf with data of the specified compression setting from
1929  * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1930  * are already sharing a data buf, no copy is performed.
1931  *
1932  * If the buf is marked as compressed but uncompressed data was requested, this
1933  * will allocate a new data buffer for the buf, remove that flag, and fill the
1934  * buf with uncompressed data. You can't request a compressed buf on a hdr with
1935  * uncompressed data, and (since we haven't added support for it yet) if you
1936  * want compressed data your buf must already be marked as compressed and have
1937  * the correct-sized data buffer.
1938  */
1939 static int
1940 arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
1941 {
1942 	arc_buf_hdr_t *hdr = buf->b_hdr;
1943 	boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
1944 	dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
1945 
1946 	ASSERT3P(buf->b_data, !=, NULL);
1947 	IMPLY(compressed, hdr_compressed);
1948 	IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
1949 
1950 	if (hdr_compressed == compressed) {
1951 		if (!arc_buf_is_shared(buf)) {
1952 			abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
1953 			    arc_buf_size(buf));
1954 		}
1955 	} else {
1956 		ASSERT(hdr_compressed);
1957 		ASSERT(!compressed);
1958 		ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
1959 
1960 		/*
1961 		 * If the buf is sharing its data with the hdr, unlink it and
1962 		 * allocate a new data buffer for the buf.
1963 		 */
1964 		if (arc_buf_is_shared(buf)) {
1965 			ASSERT(ARC_BUF_COMPRESSED(buf));
1966 
1967 			/* We need to give the buf it's own b_data */
1968 			buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
1969 			buf->b_data =
1970 			    arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
1971 			arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
1972 
1973 			/* Previously overhead was 0; just add new overhead */
1974 			ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
1975 		} else if (ARC_BUF_COMPRESSED(buf)) {
1976 			/* We need to reallocate the buf's b_data */
1977 			arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
1978 			    buf);
1979 			buf->b_data =
1980 			    arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
1981 
1982 			/* We increased the size of b_data; update overhead */
1983 			ARCSTAT_INCR(arcstat_overhead_size,
1984 			    HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
1985 		}
1986 
1987 		/*
1988 		 * Regardless of the buf's previous compression settings, it
1989 		 * should not be compressed at the end of this function.
1990 		 */
1991 		buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1992 
1993 		/*
1994 		 * Try copying the data from another buf which already has a
1995 		 * decompressed version. If that's not possible, it's time to
1996 		 * bite the bullet and decompress the data from the hdr.
1997 		 */
1998 		if (arc_buf_try_copy_decompressed_data(buf)) {
1999 			/* Skip byteswapping and checksumming (already done) */
2000 			ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2001 			return (0);
2002 		} else {
2003 			int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2004 			    hdr->b_l1hdr.b_pabd, buf->b_data,
2005 			    HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2006 
2007 			/*
2008 			 * Absent hardware errors or software bugs, this should
2009 			 * be impossible, but log it anyway so we can debug it.
2010 			 */
2011 			if (error != 0) {
2012 				zfs_dbgmsg(
2013 				    "hdr %p, compress %d, psize %d, lsize %d",
2014 				    hdr, HDR_GET_COMPRESS(hdr),
2015 				    HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2016 				return (SET_ERROR(EIO));
2017 			}
2018 		}
2019 	}
2020 
2021 	/* Byteswap the buf's data if necessary */
2022 	if (bswap != DMU_BSWAP_NUMFUNCS) {
2023 		ASSERT(!HDR_SHARED_DATA(hdr));
2024 		ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2025 		dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2026 	}
2027 
2028 	/* Compute the hdr's checksum if necessary */
2029 	arc_cksum_compute(buf);
2030 
2031 	return (0);
2032 }
2033 
2034 int
2035 arc_decompress(arc_buf_t *buf)
2036 {
2037 	return (arc_buf_fill(buf, B_FALSE));
2038 }
2039 
2040 /*
2041  * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
2042  */
2043 static uint64_t
2044 arc_hdr_size(arc_buf_hdr_t *hdr)
2045 {
2046 	uint64_t size;
2047 
2048 	if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2049 	    HDR_GET_PSIZE(hdr) > 0) {
2050 		size = HDR_GET_PSIZE(hdr);
2051 	} else {
2052 		ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2053 		size = HDR_GET_LSIZE(hdr);
2054 	}
2055 	return (size);
2056 }
2057 
2058 /*
2059  * Increment the amount of evictable space in the arc_state_t's refcount.
2060  * We account for the space used by the hdr and the arc buf individually
2061  * so that we can add and remove them from the refcount individually.
2062  */
2063 static void
2064 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2065 {
2066 	arc_buf_contents_t type = arc_buf_type(hdr);
2067 
2068 	ASSERT(HDR_HAS_L1HDR(hdr));
2069 
2070 	if (GHOST_STATE(state)) {
2071 		ASSERT0(hdr->b_l1hdr.b_bufcnt);
2072 		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2073 		ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2074 		(void) zfs_refcount_add_many(&state->arcs_esize[type],
2075 		    HDR_GET_LSIZE(hdr), hdr);
2076 		return;
2077 	}
2078 
2079 	ASSERT(!GHOST_STATE(state));
2080 	if (hdr->b_l1hdr.b_pabd != NULL) {
2081 		(void) zfs_refcount_add_many(&state->arcs_esize[type],
2082 		    arc_hdr_size(hdr), hdr);
2083 	}
2084 	for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2085 	    buf = buf->b_next) {
2086 		if (arc_buf_is_shared(buf))
2087 			continue;
2088 		(void) zfs_refcount_add_many(&state->arcs_esize[type],
2089 		    arc_buf_size(buf), buf);
2090 	}
2091 }
2092 
2093 /*
2094  * Decrement the amount of evictable space in the arc_state_t's refcount.
2095  * We account for the space used by the hdr and the arc buf individually
2096  * so that we can add and remove them from the refcount individually.
2097  */
2098 static void
2099 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2100 {
2101 	arc_buf_contents_t type = arc_buf_type(hdr);
2102 
2103 	ASSERT(HDR_HAS_L1HDR(hdr));
2104 
2105 	if (GHOST_STATE(state)) {
2106 		ASSERT0(hdr->b_l1hdr.b_bufcnt);
2107 		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2108 		ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2109 		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
2110 		    HDR_GET_LSIZE(hdr), hdr);
2111 		return;
2112 	}
2113 
2114 	ASSERT(!GHOST_STATE(state));
2115 	if (hdr->b_l1hdr.b_pabd != NULL) {
2116 		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
2117 		    arc_hdr_size(hdr), hdr);
2118 	}
2119 	for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2120 	    buf = buf->b_next) {
2121 		if (arc_buf_is_shared(buf))
2122 			continue;
2123 		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
2124 		    arc_buf_size(buf), buf);
2125 	}
2126 }
2127 
2128 /*
2129  * Add a reference to this hdr indicating that someone is actively
2130  * referencing that memory. When the refcount transitions from 0 to 1,
2131  * we remove it from the respective arc_state_t list to indicate that
2132  * it is not evictable.
2133  */
2134 static void
2135 add_reference(arc_buf_hdr_t *hdr, void *tag)
2136 {
2137 	ASSERT(HDR_HAS_L1HDR(hdr));
2138 	if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2139 		ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2140 		ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2141 		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2142 	}
2143 
2144 	arc_state_t *state = hdr->b_l1hdr.b_state;
2145 
2146 	if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2147 	    (state != arc_anon)) {
2148 		/* We don't use the L2-only state list. */
2149 		if (state != arc_l2c_only) {
2150 			multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2151 			    hdr);
2152 			arc_evictable_space_decrement(hdr, state);
2153 		}
2154 		/* remove the prefetch flag if we get a reference */
2155 		arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2156 	}
2157 }
2158 
2159 /*
2160  * Remove a reference from this hdr. When the reference transitions from
2161  * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2162  * list making it eligible for eviction.
2163  */
2164 static int
2165 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2166 {
2167 	int cnt;
2168 	arc_state_t *state = hdr->b_l1hdr.b_state;
2169 
2170 	ASSERT(HDR_HAS_L1HDR(hdr));
2171 	ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2172 	ASSERT(!GHOST_STATE(state));
2173 
2174 	/*
2175 	 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2176 	 * check to prevent usage of the arc_l2c_only list.
2177 	 */
2178 	if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2179 	    (state != arc_anon)) {
2180 		multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2181 		ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2182 		arc_evictable_space_increment(hdr, state);
2183 	}
2184 	return (cnt);
2185 }
2186 
2187 /*
2188  * Move the supplied buffer to the indicated state. The hash lock
2189  * for the buffer must be held by the caller.
2190  */
2191 static void
2192 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2193     kmutex_t *hash_lock)
2194 {
2195 	arc_state_t *old_state;
2196 	int64_t refcnt;
2197 	uint32_t bufcnt;
2198 	boolean_t update_old, update_new;
2199 	arc_buf_contents_t buftype = arc_buf_type(hdr);
2200 
2201 	/*
2202 	 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2203 	 * in arc_read() when bringing a buffer out of the L2ARC.  However, the
2204 	 * L1 hdr doesn't always exist when we change state to arc_anon before
2205 	 * destroying a header, in which case reallocating to add the L1 hdr is
2206 	 * pointless.
2207 	 */
2208 	if (HDR_HAS_L1HDR(hdr)) {
2209 		old_state = hdr->b_l1hdr.b_state;
2210 		refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt);
2211 		bufcnt = hdr->b_l1hdr.b_bufcnt;
2212 		update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL);
2213 	} else {
2214 		old_state = arc_l2c_only;
2215 		refcnt = 0;
2216 		bufcnt = 0;
2217 		update_old = B_FALSE;
2218 	}
2219 	update_new = update_old;
2220 
2221 	ASSERT(MUTEX_HELD(hash_lock));
2222 	ASSERT3P(new_state, !=, old_state);
2223 	ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2224 	ASSERT(old_state != arc_anon || bufcnt <= 1);
2225 
2226 	/*
2227 	 * If this buffer is evictable, transfer it from the
2228 	 * old state list to the new state list.
2229 	 */
2230 	if (refcnt == 0) {
2231 		if (old_state != arc_anon && old_state != arc_l2c_only) {
2232 			ASSERT(HDR_HAS_L1HDR(hdr));
2233 			multilist_remove(old_state->arcs_list[buftype], hdr);
2234 
2235 			if (GHOST_STATE(old_state)) {
2236 				ASSERT0(bufcnt);
2237 				ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2238 				update_old = B_TRUE;
2239 			}
2240 			arc_evictable_space_decrement(hdr, old_state);
2241 		}
2242 		if (new_state != arc_anon && new_state != arc_l2c_only) {
2243 
2244 			/*
2245 			 * An L1 header always exists here, since if we're
2246 			 * moving to some L1-cached state (i.e. not l2c_only or
2247 			 * anonymous), we realloc the header to add an L1hdr
2248 			 * beforehand.
2249 			 */
2250 			ASSERT(HDR_HAS_L1HDR(hdr));
2251 			multilist_insert(new_state->arcs_list[buftype], hdr);
2252 
2253 			if (GHOST_STATE(new_state)) {
2254 				ASSERT0(bufcnt);
2255 				ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2256 				update_new = B_TRUE;
2257 			}
2258 			arc_evictable_space_increment(hdr, new_state);
2259 		}
2260 	}
2261 
2262 	ASSERT(!HDR_EMPTY(hdr));
2263 	if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2264 		buf_hash_remove(hdr);
2265 
2266 	/* adjust state sizes (ignore arc_l2c_only) */
2267 
2268 	if (update_new && new_state != arc_l2c_only) {
2269 		ASSERT(HDR_HAS_L1HDR(hdr));
2270 		if (GHOST_STATE(new_state)) {
2271 			ASSERT0(bufcnt);
2272 
2273 			/*
2274 			 * When moving a header to a ghost state, we first
2275 			 * remove all arc buffers. Thus, we'll have a
2276 			 * bufcnt of zero, and no arc buffer to use for
2277 			 * the reference. As a result, we use the arc
2278 			 * header pointer for the reference.
2279 			 */
2280 			(void) zfs_refcount_add_many(&new_state->arcs_size,
2281 			    HDR_GET_LSIZE(hdr), hdr);
2282 			ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2283 		} else {
2284 			uint32_t buffers = 0;
2285 
2286 			/*
2287 			 * Each individual buffer holds a unique reference,
2288 			 * thus we must remove each of these references one
2289 			 * at a time.
2290 			 */
2291 			for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2292 			    buf = buf->b_next) {
2293 				ASSERT3U(bufcnt, !=, 0);
2294 				buffers++;
2295 
2296 				/*
2297 				 * When the arc_buf_t is sharing the data
2298 				 * block with the hdr, the owner of the
2299 				 * reference belongs to the hdr. Only
2300 				 * add to the refcount if the arc_buf_t is
2301 				 * not shared.
2302 				 */
2303 				if (arc_buf_is_shared(buf))
2304 					continue;
2305 
2306 				(void) zfs_refcount_add_many(
2307 				    &new_state->arcs_size,
2308 				    arc_buf_size(buf), buf);
2309 			}
2310 			ASSERT3U(bufcnt, ==, buffers);
2311 
2312 			if (hdr->b_l1hdr.b_pabd != NULL) {
2313 				(void) zfs_refcount_add_many(
2314 				    &new_state->arcs_size,
2315 				    arc_hdr_size(hdr), hdr);
2316 			} else {
2317 				ASSERT(GHOST_STATE(old_state));
2318 			}
2319 		}
2320 	}
2321 
2322 	if (update_old && old_state != arc_l2c_only) {
2323 		ASSERT(HDR_HAS_L1HDR(hdr));
2324 		if (GHOST_STATE(old_state)) {
2325 			ASSERT0(bufcnt);
2326 			ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2327 
2328 			/*
2329 			 * When moving a header off of a ghost state,
2330 			 * the header will not contain any arc buffers.
2331 			 * We use the arc header pointer for the reference
2332 			 * which is exactly what we did when we put the
2333 			 * header on the ghost state.
2334 			 */
2335 
2336 			(void) zfs_refcount_remove_many(&old_state->arcs_size,
2337 			    HDR_GET_LSIZE(hdr), hdr);
2338 		} else {
2339 			uint32_t buffers = 0;
2340 
2341 			/*
2342 			 * Each individual buffer holds a unique reference,
2343 			 * thus we must remove each of these references one
2344 			 * at a time.
2345 			 */
2346 			for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2347 			    buf = buf->b_next) {
2348 				ASSERT3U(bufcnt, !=, 0);
2349 				buffers++;
2350 
2351 				/*
2352 				 * When the arc_buf_t is sharing the data
2353 				 * block with the hdr, the owner of the
2354 				 * reference belongs to the hdr. Only
2355 				 * add to the refcount if the arc_buf_t is
2356 				 * not shared.
2357 				 */
2358 				if (arc_buf_is_shared(buf))
2359 					continue;
2360 
2361 				(void) zfs_refcount_remove_many(
2362 				    &old_state->arcs_size, arc_buf_size(buf),
2363 				    buf);
2364 			}
2365 			ASSERT3U(bufcnt, ==, buffers);
2366 			ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2367 			(void) zfs_refcount_remove_many(
2368 			    &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2369 		}
2370 	}
2371 
2372 	if (HDR_HAS_L1HDR(hdr))
2373 		hdr->b_l1hdr.b_state = new_state;
2374 
2375 	/*
2376 	 * L2 headers should never be on the L2 state list since they don't
2377 	 * have L1 headers allocated.
2378 	 */
2379 	ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2380 	    multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2381 }
2382 
2383 void
2384 arc_space_consume(uint64_t space, arc_space_type_t type)
2385 {
2386 	ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2387 
2388 	switch (type) {
2389 	case ARC_SPACE_DATA:
2390 		aggsum_add(&astat_data_size, space);
2391 		break;
2392 	case ARC_SPACE_META:
2393 		aggsum_add(&astat_metadata_size, space);
2394 		break;
2395 	case ARC_SPACE_OTHER:
2396 		aggsum_add(&astat_other_size, space);
2397 		break;
2398 	case ARC_SPACE_HDRS:
2399 		aggsum_add(&astat_hdr_size, space);
2400 		break;
2401 	case ARC_SPACE_L2HDRS:
2402 		aggsum_add(&astat_l2_hdr_size, space);
2403 		break;
2404 	}
2405 
2406 	if (type != ARC_SPACE_DATA)
2407 		aggsum_add(&arc_meta_used, space);
2408 
2409 	aggsum_add(&arc_size, space);
2410 }
2411 
2412 void
2413 arc_space_return(uint64_t space, arc_space_type_t type)
2414 {
2415 	ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2416 
2417 	switch (type) {
2418 	case ARC_SPACE_DATA:
2419 		aggsum_add(&astat_data_size, -space);
2420 		break;
2421 	case ARC_SPACE_META:
2422 		aggsum_add(&astat_metadata_size, -space);
2423 		break;
2424 	case ARC_SPACE_OTHER:
2425 		aggsum_add(&astat_other_size, -space);
2426 		break;
2427 	case ARC_SPACE_HDRS:
2428 		aggsum_add(&astat_hdr_size, -space);
2429 		break;
2430 	case ARC_SPACE_L2HDRS:
2431 		aggsum_add(&astat_l2_hdr_size, -space);
2432 		break;
2433 	}
2434 
2435 	if (type != ARC_SPACE_DATA) {
2436 		ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2437 		/*
2438 		 * We use the upper bound here rather than the precise value
2439 		 * because the arc_meta_max value doesn't need to be
2440 		 * precise. It's only consumed by humans via arcstats.
2441 		 */
2442 		if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2443 			arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2444 		aggsum_add(&arc_meta_used, -space);
2445 	}
2446 
2447 	ASSERT(aggsum_compare(&arc_size, space) >= 0);
2448 	aggsum_add(&arc_size, -space);
2449 }
2450 
2451 /*
2452  * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2453  * with the hdr's b_pabd.
2454  */
2455 static boolean_t
2456 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2457 {
2458 	/*
2459 	 * The criteria for sharing a hdr's data are:
2460 	 * 1. the hdr's compression matches the buf's compression
2461 	 * 2. the hdr doesn't need to be byteswapped
2462 	 * 3. the hdr isn't already being shared
2463 	 * 4. the buf is either compressed or it is the last buf in the hdr list
2464 	 *
2465 	 * Criterion #4 maintains the invariant that shared uncompressed
2466 	 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2467 	 * might ask, "if a compressed buf is allocated first, won't that be the
2468 	 * last thing in the list?", but in that case it's impossible to create
2469 	 * a shared uncompressed buf anyway (because the hdr must be compressed
2470 	 * to have the compressed buf). You might also think that #3 is
2471 	 * sufficient to make this guarantee, however it's possible
2472 	 * (specifically in the rare L2ARC write race mentioned in
2473 	 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2474 	 * is sharable, but wasn't at the time of its allocation. Rather than
2475 	 * allow a new shared uncompressed buf to be created and then shuffle
2476 	 * the list around to make it the last element, this simply disallows
2477 	 * sharing if the new buf isn't the first to be added.
2478 	 */
2479 	ASSERT3P(buf->b_hdr, ==, hdr);
2480 	boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF;
2481 	boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2482 	return (buf_compressed == hdr_compressed &&
2483 	    hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2484 	    !HDR_SHARED_DATA(hdr) &&
2485 	    (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2486 }
2487 
2488 /*
2489  * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2490  * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2491  * copy was made successfully, or an error code otherwise.
2492  */
2493 static int
2494 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2495     boolean_t fill, arc_buf_t **ret)
2496 {
2497 	arc_buf_t *buf;
2498 
2499 	ASSERT(HDR_HAS_L1HDR(hdr));
2500 	ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2501 	VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2502 	    hdr->b_type == ARC_BUFC_METADATA);
2503 	ASSERT3P(ret, !=, NULL);
2504 	ASSERT3P(*ret, ==, NULL);
2505 
2506 	buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2507 	buf->b_hdr = hdr;
2508 	buf->b_data = NULL;
2509 	buf->b_next = hdr->b_l1hdr.b_buf;
2510 	buf->b_flags = 0;
2511 
2512 	add_reference(hdr, tag);
2513 
2514 	/*
2515 	 * We're about to change the hdr's b_flags. We must either
2516 	 * hold the hash_lock or be undiscoverable.
2517 	 */
2518 	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2519 
2520 	/*
2521 	 * Only honor requests for compressed bufs if the hdr is actually
2522 	 * compressed.
2523 	 */
2524 	if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
2525 		buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2526 
2527 	/*
2528 	 * If the hdr's data can be shared then we share the data buffer and
2529 	 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2530 	 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2531 	 * buffer to store the buf's data.
2532 	 *
2533 	 * There are two additional restrictions here because we're sharing
2534 	 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2535 	 * actively involved in an L2ARC write, because if this buf is used by
2536 	 * an arc_write() then the hdr's data buffer will be released when the
2537 	 * write completes, even though the L2ARC write might still be using it.
2538 	 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2539 	 * need to be ABD-aware.
2540 	 */
2541 	boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2542 	    abd_is_linear(hdr->b_l1hdr.b_pabd);
2543 
2544 	/* Set up b_data and sharing */
2545 	if (can_share) {
2546 		buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2547 		buf->b_flags |= ARC_BUF_FLAG_SHARED;
2548 		arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2549 	} else {
2550 		buf->b_data =
2551 		    arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2552 		ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2553 	}
2554 	VERIFY3P(buf->b_data, !=, NULL);
2555 
2556 	hdr->b_l1hdr.b_buf = buf;
2557 	hdr->b_l1hdr.b_bufcnt += 1;
2558 
2559 	/*
2560 	 * If the user wants the data from the hdr, we need to either copy or
2561 	 * decompress the data.
2562 	 */
2563 	if (fill) {
2564 		return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
2565 	}
2566 
2567 	return (0);
2568 }
2569 
2570 static char *arc_onloan_tag = "onloan";
2571 
2572 static inline void
2573 arc_loaned_bytes_update(int64_t delta)
2574 {
2575 	atomic_add_64(&arc_loaned_bytes, delta);
2576 
2577 	/* assert that it did not wrap around */
2578 	ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2579 }
2580 
2581 /*
2582  * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2583  * flight data by arc_tempreserve_space() until they are "returned". Loaned
2584  * buffers must be returned to the arc before they can be used by the DMU or
2585  * freed.
2586  */
2587 arc_buf_t *
2588 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2589 {
2590 	arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2591 	    is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2592 
2593 	arc_loaned_bytes_update(arc_buf_size(buf));
2594 
2595 	return (buf);
2596 }
2597 
2598 arc_buf_t *
2599 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2600     enum zio_compress compression_type)
2601 {
2602 	arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2603 	    psize, lsize, compression_type);
2604 
2605 	arc_loaned_bytes_update(arc_buf_size(buf));
2606 
2607 	return (buf);
2608 }
2609 
2610 
2611 /*
2612  * Return a loaned arc buffer to the arc.
2613  */
2614 void
2615 arc_return_buf(arc_buf_t *buf, void *tag)
2616 {
2617 	arc_buf_hdr_t *hdr = buf->b_hdr;
2618 
2619 	ASSERT3P(buf->b_data, !=, NULL);
2620 	ASSERT(HDR_HAS_L1HDR(hdr));
2621 	(void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2622 	(void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2623 
2624 	arc_loaned_bytes_update(-arc_buf_size(buf));
2625 }
2626 
2627 /* Detach an arc_buf from a dbuf (tag) */
2628 void
2629 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2630 {
2631 	arc_buf_hdr_t *hdr = buf->b_hdr;
2632 
2633 	ASSERT3P(buf->b_data, !=, NULL);
2634 	ASSERT(HDR_HAS_L1HDR(hdr));
2635 	(void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2636 	(void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2637 
2638 	arc_loaned_bytes_update(arc_buf_size(buf));
2639 }
2640 
2641 static void
2642 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2643 {
2644 	l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2645 
2646 	df->l2df_abd = abd;
2647 	df->l2df_size = size;
2648 	df->l2df_type = type;
2649 	mutex_enter(&l2arc_free_on_write_mtx);
2650 	list_insert_head(l2arc_free_on_write, df);
2651 	mutex_exit(&l2arc_free_on_write_mtx);
2652 }
2653 
2654 static void
2655 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
2656 {
2657 	arc_state_t *state = hdr->b_l1hdr.b_state;
2658 	arc_buf_contents_t type = arc_buf_type(hdr);
2659 	uint64_t size = arc_hdr_size(hdr);
2660 
2661 	/* protected by hash lock, if in the hash table */
2662 	if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2663 		ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2664 		ASSERT(state != arc_anon && state != arc_l2c_only);
2665 
2666 		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
2667 		    size, hdr);
2668 	}
2669 	(void) zfs_refcount_remove_many(&state->arcs_size, size, hdr);
2670 	if (type == ARC_BUFC_METADATA) {
2671 		arc_space_return(size, ARC_SPACE_META);
2672 	} else {
2673 		ASSERT(type == ARC_BUFC_DATA);
2674 		arc_space_return(size, ARC_SPACE_DATA);
2675 	}
2676 
2677 	l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2678 }
2679 
2680 /*
2681  * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2682  * data buffer, we transfer the refcount ownership to the hdr and update
2683  * the appropriate kstats.
2684  */
2685 static void
2686 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2687 {
2688 	arc_state_t *state = hdr->b_l1hdr.b_state;
2689 
2690 	ASSERT(arc_can_share(hdr, buf));
2691 	ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2692 	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2693 
2694 	/*
2695 	 * Start sharing the data buffer. We transfer the
2696 	 * refcount ownership to the hdr since it always owns
2697 	 * the refcount whenever an arc_buf_t is shared.
2698 	 */
2699 	zfs_refcount_transfer_ownership(&state->arcs_size, buf, hdr);
2700 	hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2701 	abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2702 	    HDR_ISTYPE_METADATA(hdr));
2703 	arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2704 	buf->b_flags |= ARC_BUF_FLAG_SHARED;
2705 
2706 	/*
2707 	 * Since we've transferred ownership to the hdr we need
2708 	 * to increment its compressed and uncompressed kstats and
2709 	 * decrement the overhead size.
2710 	 */
2711 	ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2712 	ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2713 	ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
2714 }
2715 
2716 static void
2717 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2718 {
2719 	arc_state_t *state = hdr->b_l1hdr.b_state;
2720 
2721 	ASSERT(arc_buf_is_shared(buf));
2722 	ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2723 	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2724 
2725 	/*
2726 	 * We are no longer sharing this buffer so we need
2727 	 * to transfer its ownership to the rightful owner.
2728 	 */
2729 	zfs_refcount_transfer_ownership(&state->arcs_size, hdr, buf);
2730 	arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2731 	abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
2732 	abd_put(hdr->b_l1hdr.b_pabd);
2733 	hdr->b_l1hdr.b_pabd = NULL;
2734 	buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2735 
2736 	/*
2737 	 * Since the buffer is no longer shared between
2738 	 * the arc buf and the hdr, count it as overhead.
2739 	 */
2740 	ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2741 	ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2742 	ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2743 }
2744 
2745 /*
2746  * Remove an arc_buf_t from the hdr's buf list and return the last
2747  * arc_buf_t on the list. If no buffers remain on the list then return
2748  * NULL.
2749  */
2750 static arc_buf_t *
2751 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2752 {
2753 	ASSERT(HDR_HAS_L1HDR(hdr));
2754 	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2755 
2756 	arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
2757 	arc_buf_t *lastbuf = NULL;
2758 
2759 	/*
2760 	 * Remove the buf from the hdr list and locate the last
2761 	 * remaining buffer on the list.
2762 	 */
2763 	while (*bufp != NULL) {
2764 		if (*bufp == buf)
2765 			*bufp = buf->b_next;
2766 
2767 		/*
2768 		 * If we've removed a buffer in the middle of
2769 		 * the list then update the lastbuf and update
2770 		 * bufp.
2771 		 */
2772 		if (*bufp != NULL) {
2773 			lastbuf = *bufp;
2774 			bufp = &(*bufp)->b_next;
2775 		}
2776 	}
2777 	buf->b_next = NULL;
2778 	ASSERT3P(lastbuf, !=, buf);
2779 	IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
2780 	IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
2781 	IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
2782 
2783 	return (lastbuf);
2784 }
2785 
2786 /*
2787  * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
2788  * list and free it.
2789  */
2790 static void
2791 arc_buf_destroy_impl(arc_buf_t *buf)
2792 {
2793 	arc_buf_hdr_t *hdr = buf->b_hdr;
2794 
2795 	/*
2796 	 * Free up the data associated with the buf but only if we're not
2797 	 * sharing this with the hdr. If we are sharing it with the hdr, the
2798 	 * hdr is responsible for doing the free.
2799 	 */
2800 	if (buf->b_data != NULL) {
2801 		/*
2802 		 * We're about to change the hdr's b_flags. We must either
2803 		 * hold the hash_lock or be undiscoverable.
2804 		 */
2805 		ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2806 
2807 		arc_cksum_verify(buf);
2808 		arc_buf_unwatch(buf);
2809 
2810 		if (arc_buf_is_shared(buf)) {
2811 			arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2812 		} else {
2813 			uint64_t size = arc_buf_size(buf);
2814 			arc_free_data_buf(hdr, buf->b_data, size, buf);
2815 			ARCSTAT_INCR(arcstat_overhead_size, -size);
2816 		}
2817 		buf->b_data = NULL;
2818 
2819 		ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
2820 		hdr->b_l1hdr.b_bufcnt -= 1;
2821 	}
2822 
2823 	arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
2824 
2825 	if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
2826 		/*
2827 		 * If the current arc_buf_t is sharing its data buffer with the
2828 		 * hdr, then reassign the hdr's b_pabd to share it with the new
2829 		 * buffer at the end of the list. The shared buffer is always
2830 		 * the last one on the hdr's buffer list.
2831 		 *
2832 		 * There is an equivalent case for compressed bufs, but since
2833 		 * they aren't guaranteed to be the last buf in the list and
2834 		 * that is an exceedingly rare case, we just allow that space be
2835 		 * wasted temporarily.
2836 		 */
2837 		if (lastbuf != NULL) {
2838 			/* Only one buf can be shared at once */
2839 			VERIFY(!arc_buf_is_shared(lastbuf));
2840 			/* hdr is uncompressed so can't have compressed buf */
2841 			VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
2842 
2843 			ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2844 			arc_hdr_free_pabd(hdr);
2845 
2846 			/*
2847 			 * We must setup a new shared block between the
2848 			 * last buffer and the hdr. The data would have
2849 			 * been allocated by the arc buf so we need to transfer
2850 			 * ownership to the hdr since it's now being shared.
2851 			 */
2852 			arc_share_buf(hdr, lastbuf);
2853 		}
2854 	} else if (HDR_SHARED_DATA(hdr)) {
2855 		/*
2856 		 * Uncompressed shared buffers are always at the end
2857 		 * of the list. Compressed buffers don't have the
2858 		 * same requirements. This makes it hard to
2859 		 * simply assert that the lastbuf is shared so
2860 		 * we rely on the hdr's compression flags to determine
2861 		 * if we have a compressed, shared buffer.
2862 		 */
2863 		ASSERT3P(lastbuf, !=, NULL);
2864 		ASSERT(arc_buf_is_shared(lastbuf) ||
2865 		    HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
2866 	}
2867 
2868 	/*
2869 	 * Free the checksum if we're removing the last uncompressed buf from
2870 	 * this hdr.
2871 	 */
2872 	if (!arc_hdr_has_uncompressed_buf(hdr)) {
2873 		arc_cksum_free(hdr);
2874 	}
2875 
2876 	/* clean up the buf */
2877 	buf->b_hdr = NULL;
2878 	kmem_cache_free(buf_cache, buf);
2879 }
2880 
2881 static void
2882 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr)
2883 {
2884 	ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2885 	ASSERT(HDR_HAS_L1HDR(hdr));
2886 	ASSERT(!HDR_SHARED_DATA(hdr));
2887 
2888 	ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2889 	hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
2890 	hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
2891 	ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2892 
2893 	ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2894 	ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2895 }
2896 
2897 static void
2898 arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
2899 {
2900 	ASSERT(HDR_HAS_L1HDR(hdr));
2901 	ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2902 
2903 	/*
2904 	 * If the hdr is currently being written to the l2arc then
2905 	 * we defer freeing the data by adding it to the l2arc_free_on_write
2906 	 * list. The l2arc will free the data once it's finished
2907 	 * writing it to the l2arc device.
2908 	 */
2909 	if (HDR_L2_WRITING(hdr)) {
2910 		arc_hdr_free_on_write(hdr);
2911 		ARCSTAT_BUMP(arcstat_l2_free_on_write);
2912 	} else {
2913 		arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
2914 		    arc_hdr_size(hdr), hdr);
2915 	}
2916 	hdr->b_l1hdr.b_pabd = NULL;
2917 	hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
2918 
2919 	ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2920 	ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2921 }
2922 
2923 static arc_buf_hdr_t *
2924 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
2925     enum zio_compress compression_type, arc_buf_contents_t type)
2926 {
2927 	arc_buf_hdr_t *hdr;
2928 
2929 	VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
2930 
2931 	hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
2932 	ASSERT(HDR_EMPTY(hdr));
2933 	ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
2934 	ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
2935 	HDR_SET_PSIZE(hdr, psize);
2936 	HDR_SET_LSIZE(hdr, lsize);
2937 	hdr->b_spa = spa;
2938 	hdr->b_type = type;
2939 	hdr->b_flags = 0;
2940 	arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
2941 	arc_hdr_set_compress(hdr, compression_type);
2942 
2943 	hdr->b_l1hdr.b_state = arc_anon;
2944 	hdr->b_l1hdr.b_arc_access = 0;
2945 	hdr->b_l1hdr.b_bufcnt = 0;
2946 	hdr->b_l1hdr.b_buf = NULL;
2947 
2948 	/*
2949 	 * Allocate the hdr's buffer. This will contain either
2950 	 * the compressed or uncompressed data depending on the block
2951 	 * it references and compressed arc enablement.
2952 	 */
2953 	arc_hdr_alloc_pabd(hdr);
2954 	ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2955 
2956 	return (hdr);
2957 }
2958 
2959 /*
2960  * Transition between the two allocation states for the arc_buf_hdr struct.
2961  * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
2962  * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
2963  * version is used when a cache buffer is only in the L2ARC in order to reduce
2964  * memory usage.
2965  */
2966 static arc_buf_hdr_t *
2967 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
2968 {
2969 	ASSERT(HDR_HAS_L2HDR(hdr));
2970 
2971 	arc_buf_hdr_t *nhdr;
2972 	l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2973 
2974 	ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
2975 	    (old == hdr_l2only_cache && new == hdr_full_cache));
2976 
2977 	nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
2978 
2979 	ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
2980 	buf_hash_remove(hdr);
2981 
2982 	bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
2983 
2984 	if (new == hdr_full_cache) {
2985 		arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
2986 		/*
2987 		 * arc_access and arc_change_state need to be aware that a
2988 		 * header has just come out of L2ARC, so we set its state to
2989 		 * l2c_only even though it's about to change.
2990 		 */
2991 		nhdr->b_l1hdr.b_state = arc_l2c_only;
2992 
2993 		/* Verify previous threads set to NULL before freeing */
2994 		ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
2995 	} else {
2996 		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2997 		ASSERT0(hdr->b_l1hdr.b_bufcnt);
2998 		ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
2999 
3000 		/*
3001 		 * If we've reached here, We must have been called from
3002 		 * arc_evict_hdr(), as such we should have already been
3003 		 * removed from any ghost list we were previously on
3004 		 * (which protects us from racing with arc_evict_state),
3005 		 * thus no locking is needed during this check.
3006 		 */
3007 		ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3008 
3009 		/*
3010 		 * A buffer must not be moved into the arc_l2c_only
3011 		 * state if it's not finished being written out to the
3012 		 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3013 		 * might try to be accessed, even though it was removed.
3014 		 */
3015 		VERIFY(!HDR_L2_WRITING(hdr));
3016 		VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3017 
3018 #ifdef ZFS_DEBUG
3019 		if (hdr->b_l1hdr.b_thawed != NULL) {
3020 			kmem_free(hdr->b_l1hdr.b_thawed, 1);
3021 			hdr->b_l1hdr.b_thawed = NULL;
3022 		}
3023 #endif
3024 
3025 		arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3026 	}
3027 	/*
3028 	 * The header has been reallocated so we need to re-insert it into any
3029 	 * lists it was on.
3030 	 */
3031 	(void) buf_hash_insert(nhdr, NULL);
3032 
3033 	ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3034 
3035 	mutex_enter(&dev->l2ad_mtx);
3036 
3037 	/*
3038 	 * We must place the realloc'ed header back into the list at
3039 	 * the same spot. Otherwise, if it's placed earlier in the list,
3040 	 * l2arc_write_buffers() could find it during the function's
3041 	 * write phase, and try to write it out to the l2arc.
3042 	 */
3043 	list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3044 	list_remove(&dev->l2ad_buflist, hdr);
3045 
3046 	mutex_exit(&dev->l2ad_mtx);
3047 
3048 	/*
3049 	 * Since we're using the pointer address as the tag when
3050 	 * incrementing and decrementing the l2ad_alloc refcount, we
3051 	 * must remove the old pointer (that we're about to destroy) and
3052 	 * add the new pointer to the refcount. Otherwise we'd remove
3053 	 * the wrong pointer address when calling arc_hdr_destroy() later.
3054 	 */
3055 
3056 	(void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
3057 	    hdr);
3058 	(void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr),
3059 	    nhdr);
3060 
3061 	buf_discard_identity(hdr);
3062 	kmem_cache_free(old, hdr);
3063 
3064 	return (nhdr);
3065 }
3066 
3067 /*
3068  * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3069  * The buf is returned thawed since we expect the consumer to modify it.
3070  */
3071 arc_buf_t *
3072 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3073 {
3074 	arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3075 	    ZIO_COMPRESS_OFF, type);
3076 	ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3077 
3078 	arc_buf_t *buf = NULL;
3079 	VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf));
3080 	arc_buf_thaw(buf);
3081 
3082 	return (buf);
3083 }
3084 
3085 /*
3086  * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3087  * for bufs containing metadata.
3088  */
3089 arc_buf_t *
3090 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3091     enum zio_compress compression_type)
3092 {
3093 	ASSERT3U(lsize, >, 0);
3094 	ASSERT3U(lsize, >=, psize);
3095 	ASSERT(compression_type > ZIO_COMPRESS_OFF);
3096 	ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3097 
3098 	arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3099 	    compression_type, ARC_BUFC_DATA);
3100 	ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3101 
3102 	arc_buf_t *buf = NULL;
3103 	VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3104 	arc_buf_thaw(buf);
3105 	ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3106 
3107 	if (!arc_buf_is_shared(buf)) {
3108 		/*
3109 		 * To ensure that the hdr has the correct data in it if we call
3110 		 * arc_decompress() on this buf before it's been written to
3111 		 * disk, it's easiest if we just set up sharing between the
3112 		 * buf and the hdr.
3113 		 */
3114 		ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3115 		arc_hdr_free_pabd(hdr);
3116 		arc_share_buf(hdr, buf);
3117 	}
3118 
3119 	return (buf);
3120 }
3121 
3122 static void
3123 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3124 {
3125 	l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3126 	l2arc_dev_t *dev = l2hdr->b_dev;
3127 	uint64_t psize = arc_hdr_size(hdr);
3128 
3129 	ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3130 	ASSERT(HDR_HAS_L2HDR(hdr));
3131 
3132 	list_remove(&dev->l2ad_buflist, hdr);
3133 
3134 	ARCSTAT_INCR(arcstat_l2_psize, -psize);
3135 	ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3136 
3137 	vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3138 
3139 	(void) zfs_refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3140 	arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3141 }
3142 
3143 static void
3144 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3145 {
3146 	if (HDR_HAS_L1HDR(hdr)) {
3147 		ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3148 		    hdr->b_l1hdr.b_bufcnt > 0);
3149 		ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3150 		ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3151 	}
3152 	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3153 	ASSERT(!HDR_IN_HASH_TABLE(hdr));
3154 
3155 	if (!HDR_EMPTY(hdr))
3156 		buf_discard_identity(hdr);
3157 
3158 	if (HDR_HAS_L2HDR(hdr)) {
3159 		l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3160 		boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3161 
3162 		if (!buflist_held)
3163 			mutex_enter(&dev->l2ad_mtx);
3164 
3165 		/*
3166 		 * Even though we checked this conditional above, we
3167 		 * need to check this again now that we have the
3168 		 * l2ad_mtx. This is because we could be racing with
3169 		 * another thread calling l2arc_evict() which might have
3170 		 * destroyed this header's L2 portion as we were waiting
3171 		 * to acquire the l2ad_mtx. If that happens, we don't
3172 		 * want to re-destroy the header's L2 portion.
3173 		 */
3174 		if (HDR_HAS_L2HDR(hdr))
3175 			arc_hdr_l2hdr_destroy(hdr);
3176 
3177 		if (!buflist_held)
3178 			mutex_exit(&dev->l2ad_mtx);
3179 	}
3180 
3181 	if (HDR_HAS_L1HDR(hdr)) {
3182 		arc_cksum_free(hdr);
3183 
3184 		while (hdr->b_l1hdr.b_buf != NULL)
3185 			arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3186 
3187 #ifdef ZFS_DEBUG
3188 		if (hdr->b_l1hdr.b_thawed != NULL) {
3189 			kmem_free(hdr->b_l1hdr.b_thawed, 1);
3190 			hdr->b_l1hdr.b_thawed = NULL;
3191 		}
3192 #endif
3193 
3194 		if (hdr->b_l1hdr.b_pabd != NULL) {
3195 			arc_hdr_free_pabd(hdr);
3196 		}
3197 	}
3198 
3199 	ASSERT3P(hdr->b_hash_next, ==, NULL);
3200 	if (HDR_HAS_L1HDR(hdr)) {
3201 		ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3202 		ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3203 		kmem_cache_free(hdr_full_cache, hdr);
3204 	} else {
3205 		kmem_cache_free(hdr_l2only_cache, hdr);
3206 	}
3207 }
3208 
3209 void
3210 arc_buf_destroy(arc_buf_t *buf, void* tag)
3211 {
3212 	arc_buf_hdr_t *hdr = buf->b_hdr;
3213 	kmutex_t *hash_lock = HDR_LOCK(hdr);
3214 
3215 	if (hdr->b_l1hdr.b_state == arc_anon) {
3216 		ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3217 		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3218 		VERIFY0(remove_reference(hdr, NULL, tag));
3219 		arc_hdr_destroy(hdr);
3220 		return;
3221 	}
3222 
3223 	mutex_enter(hash_lock);
3224 	ASSERT3P(hdr, ==, buf->b_hdr);
3225 	ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3226 	ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3227 	ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3228 	ASSERT3P(buf->b_data, !=, NULL);
3229 
3230 	(void) remove_reference(hdr, hash_lock, tag);
3231 	arc_buf_destroy_impl(buf);
3232 	mutex_exit(hash_lock);
3233 }
3234 
3235 /*
3236  * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3237  * state of the header is dependent on it's state prior to entering this
3238  * function. The following transitions are possible:
3239  *
3240  *    - arc_mru -> arc_mru_ghost
3241  *    - arc_mfu -> arc_mfu_ghost
3242  *    - arc_mru_ghost -> arc_l2c_only
3243  *    - arc_mru_ghost -> deleted
3244  *    - arc_mfu_ghost -> arc_l2c_only
3245  *    - arc_mfu_ghost -> deleted
3246  */
3247 static int64_t
3248 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3249 {
3250 	arc_state_t *evicted_state, *state;
3251 	int64_t bytes_evicted = 0;
3252 
3253 	ASSERT(MUTEX_HELD(hash_lock));
3254 	ASSERT(HDR_HAS_L1HDR(hdr));
3255 
3256 	state = hdr->b_l1hdr.b_state;
3257 	if (GHOST_STATE(state)) {
3258 		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3259 		ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3260 
3261 		/*
3262 		 * l2arc_write_buffers() relies on a header's L1 portion
3263 		 * (i.e. its b_pabd field) during it's write phase.
3264 		 * Thus, we cannot push a header onto the arc_l2c_only
3265 		 * state (removing it's L1 piece) until the header is
3266 		 * done being written to the l2arc.
3267 		 */
3268 		if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3269 			ARCSTAT_BUMP(arcstat_evict_l2_skip);
3270 			return (bytes_evicted);
3271 		}
3272 
3273 		ARCSTAT_BUMP(arcstat_deleted);
3274 		bytes_evicted += HDR_GET_LSIZE(hdr);
3275 
3276 		DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3277 
3278 		ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3279 		if (HDR_HAS_L2HDR(hdr)) {
3280 			/*
3281 			 * This buffer is cached on the 2nd Level ARC;
3282 			 * don't destroy the header.
3283 			 */
3284 			arc_change_state(arc_l2c_only, hdr, hash_lock);
3285 			/*
3286 			 * dropping from L1+L2 cached to L2-only,
3287 			 * realloc to remove the L1 header.
3288 			 */
3289 			hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3290 			    hdr_l2only_cache);
3291 		} else {
3292 			arc_change_state(arc_anon, hdr, hash_lock);
3293 			arc_hdr_destroy(hdr);
3294 		}
3295 		return (bytes_evicted);
3296 	}
3297 
3298 	ASSERT(state == arc_mru || state == arc_mfu);
3299 	evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3300 
3301 	/* prefetch buffers have a minimum lifespan */
3302 	if (HDR_IO_IN_PROGRESS(hdr) ||
3303 	    ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3304 	    ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3305 	    arc_min_prefetch_lifespan)) {
3306 		ARCSTAT_BUMP(arcstat_evict_skip);
3307 		return (bytes_evicted);
3308 	}
3309 
3310 	ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3311 	while (hdr->b_l1hdr.b_buf) {
3312 		arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3313 		if (!mutex_tryenter(&buf->b_evict_lock)) {
3314 			ARCSTAT_BUMP(arcstat_mutex_miss);
3315 			break;
3316 		}
3317 		if (buf->b_data != NULL)
3318 			bytes_evicted += HDR_GET_LSIZE(hdr);
3319 		mutex_exit(&buf->b_evict_lock);
3320 		arc_buf_destroy_impl(buf);
3321 	}
3322 
3323 	if (HDR_HAS_L2HDR(hdr)) {
3324 		ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3325 	} else {
3326 		if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3327 			ARCSTAT_INCR(arcstat_evict_l2_eligible,
3328 			    HDR_GET_LSIZE(hdr));
3329 		} else {
3330 			ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3331 			    HDR_GET_LSIZE(hdr));
3332 		}
3333 	}
3334 
3335 	if (hdr->b_l1hdr.b_bufcnt == 0) {
3336 		arc_cksum_free(hdr);
3337 
3338 		bytes_evicted += arc_hdr_size(hdr);
3339 
3340 		/*
3341 		 * If this hdr is being evicted and has a compressed
3342 		 * buffer then we discard it here before we change states.
3343 		 * This ensures that the accounting is updated correctly
3344 		 * in arc_free_data_impl().
3345 		 */
3346 		arc_hdr_free_pabd(hdr);
3347 
3348 		arc_change_state(evicted_state, hdr, hash_lock);
3349 		ASSERT(HDR_IN_HASH_TABLE(hdr));
3350 		arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3351 		DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3352 	}
3353 
3354 	return (bytes_evicted);
3355 }
3356 
3357 static uint64_t
3358 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3359     uint64_t spa, int64_t bytes)
3360 {
3361 	multilist_sublist_t *mls;
3362 	uint64_t bytes_evicted = 0;
3363 	arc_buf_hdr_t *hdr;
3364 	kmutex_t *hash_lock;
3365 	int evict_count = 0;
3366 
3367 	ASSERT3P(marker, !=, NULL);
3368 	IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3369 
3370 	mls = multilist_sublist_lock(ml, idx);
3371 
3372 	for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3373 	    hdr = multilist_sublist_prev(mls, marker)) {
3374 		if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3375 		    (evict_count >= zfs_arc_evict_batch_limit))
3376 			break;
3377 
3378 		/*
3379 		 * To keep our iteration location, move the marker
3380 		 * forward. Since we're not holding hdr's hash lock, we
3381 		 * must be very careful and not remove 'hdr' from the
3382 		 * sublist. Otherwise, other consumers might mistake the
3383 		 * 'hdr' as not being on a sublist when they call the
3384 		 * multilist_link_active() function (they all rely on
3385 		 * the hash lock protecting concurrent insertions and
3386 		 * removals). multilist_sublist_move_forward() was
3387 		 * specifically implemented to ensure this is the case
3388 		 * (only 'marker' will be removed and re-inserted).
3389 		 */
3390 		multilist_sublist_move_forward(mls, marker);
3391 
3392 		/*
3393 		 * The only case where the b_spa field should ever be
3394 		 * zero, is the marker headers inserted by
3395 		 * arc_evict_state(). It's possible for multiple threads
3396 		 * to be calling arc_evict_state() concurrently (e.g.
3397 		 * dsl_pool_close() and zio_inject_fault()), so we must
3398 		 * skip any markers we see from these other threads.
3399 		 */
3400 		if (hdr->b_spa == 0)
3401 			continue;
3402 
3403 		/* we're only interested in evicting buffers of a certain spa */
3404 		if (spa != 0 && hdr->b_spa != spa) {
3405 			ARCSTAT_BUMP(arcstat_evict_skip);
3406 			continue;
3407 		}
3408 
3409 		hash_lock = HDR_LOCK(hdr);
3410 
3411 		/*
3412 		 * We aren't calling this function from any code path
3413 		 * that would already be holding a hash lock, so we're
3414 		 * asserting on this assumption to be defensive in case
3415 		 * this ever changes. Without this check, it would be
3416 		 * possible to incorrectly increment arcstat_mutex_miss
3417 		 * below (e.g. if the code changed such that we called
3418 		 * this function with a hash lock held).
3419 		 */
3420 		ASSERT(!MUTEX_HELD(hash_lock));
3421 
3422 		if (mutex_tryenter(hash_lock)) {
3423 			uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3424 			mutex_exit(hash_lock);
3425 
3426 			bytes_evicted += evicted;
3427 
3428 			/*
3429 			 * If evicted is zero, arc_evict_hdr() must have
3430 			 * decided to skip this header, don't increment
3431 			 * evict_count in this case.
3432 			 */
3433 			if (evicted != 0)
3434 				evict_count++;
3435 
3436 			/*
3437 			 * If arc_size isn't overflowing, signal any
3438 			 * threads that might happen to be waiting.
3439 			 *
3440 			 * For each header evicted, we wake up a single
3441 			 * thread. If we used cv_broadcast, we could
3442 			 * wake up "too many" threads causing arc_size
3443 			 * to significantly overflow arc_c; since
3444 			 * arc_get_data_impl() doesn't check for overflow
3445 			 * when it's woken up (it doesn't because it's
3446 			 * possible for the ARC to be overflowing while
3447 			 * full of un-evictable buffers, and the
3448 			 * function should proceed in this case).
3449 			 *
3450 			 * If threads are left sleeping, due to not
3451 			 * using cv_broadcast here, they will be woken
3452 			 * up via cv_broadcast in arc_adjust_cb() just
3453 			 * before arc_adjust_zthr sleeps.
3454 			 */
3455 			mutex_enter(&arc_adjust_lock);
3456 			if (!arc_is_overflowing())
3457 				cv_signal(&arc_adjust_waiters_cv);
3458 			mutex_exit(&arc_adjust_lock);
3459 		} else {
3460 			ARCSTAT_BUMP(arcstat_mutex_miss);
3461 		}
3462 	}
3463 
3464 	multilist_sublist_unlock(mls);
3465 
3466 	return (bytes_evicted);
3467 }
3468 
3469 /*
3470  * Evict buffers from the given arc state, until we've removed the
3471  * specified number of bytes. Move the removed buffers to the
3472  * appropriate evict state.
3473  *
3474  * This function makes a "best effort". It skips over any buffers
3475  * it can't get a hash_lock on, and so, may not catch all candidates.
3476  * It may also return without evicting as much space as requested.
3477  *
3478  * If bytes is specified using the special value ARC_EVICT_ALL, this
3479  * will evict all available (i.e. unlocked and evictable) buffers from
3480  * the given arc state; which is used by arc_flush().
3481  */
3482 static uint64_t
3483 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3484     arc_buf_contents_t type)
3485 {
3486 	uint64_t total_evicted = 0;
3487 	multilist_t *ml = state->arcs_list[type];
3488 	int num_sublists;
3489 	arc_buf_hdr_t **markers;
3490 
3491 	IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3492 
3493 	num_sublists = multilist_get_num_sublists(ml);
3494 
3495 	/*
3496 	 * If we've tried to evict from each sublist, made some
3497 	 * progress, but still have not hit the target number of bytes
3498 	 * to evict, we want to keep trying. The markers allow us to
3499 	 * pick up where we left off for each individual sublist, rather
3500 	 * than starting from the tail each time.
3501 	 */
3502 	markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3503 	for (int i = 0; i < num_sublists; i++) {
3504 		markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3505 
3506 		/*
3507 		 * A b_spa of 0 is used to indicate that this header is
3508 		 * a marker. This fact is used in arc_adjust_type() and
3509 		 * arc_evict_state_impl().
3510 		 */
3511 		markers[i]->b_spa = 0;
3512 
3513 		multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3514 		multilist_sublist_insert_tail(mls, markers[i]);
3515 		multilist_sublist_unlock(mls);
3516 	}
3517 
3518 	/*
3519 	 * While we haven't hit our target number of bytes to evict, or
3520 	 * we're evicting all available buffers.
3521 	 */
3522 	while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3523 		/*
3524 		 * Start eviction using a randomly selected sublist,
3525 		 * this is to try and evenly balance eviction across all
3526 		 * sublists. Always starting at the same sublist
3527 		 * (e.g. index 0) would cause evictions to favor certain
3528 		 * sublists over others.
3529 		 */
3530 		int sublist_idx = multilist_get_random_index(ml);
3531 		uint64_t scan_evicted = 0;
3532 
3533 		for (int i = 0; i < num_sublists; i++) {
3534 			uint64_t bytes_remaining;
3535 			uint64_t bytes_evicted;
3536 
3537 			if (bytes == ARC_EVICT_ALL)
3538 				bytes_remaining = ARC_EVICT_ALL;
3539 			else if (total_evicted < bytes)
3540 				bytes_remaining = bytes - total_evicted;
3541 			else
3542 				break;
3543 
3544 			bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3545 			    markers[sublist_idx], spa, bytes_remaining);
3546 
3547 			scan_evicted += bytes_evicted;
3548 			total_evicted += bytes_evicted;
3549 
3550 			/* we've reached the end, wrap to the beginning */
3551 			if (++sublist_idx >= num_sublists)
3552 				sublist_idx = 0;
3553 		}
3554 
3555 		/*
3556 		 * If we didn't evict anything during this scan, we have
3557 		 * no reason to believe we'll evict more during another
3558 		 * scan, so break the loop.
3559 		 */
3560 		if (scan_evicted == 0) {
3561 			/* This isn't possible, let's make that obvious */
3562 			ASSERT3S(bytes, !=, 0);
3563 
3564 			/*
3565 			 * When bytes is ARC_EVICT_ALL, the only way to
3566 			 * break the loop is when scan_evicted is zero.
3567 			 * In that case, we actually have evicted enough,
3568 			 * so we don't want to increment the kstat.
3569 			 */
3570 			if (bytes != ARC_EVICT_ALL) {
3571 				ASSERT3S(total_evicted, <, bytes);
3572 				ARCSTAT_BUMP(arcstat_evict_not_enough);
3573 			}
3574 
3575 			break;
3576 		}
3577 	}
3578 
3579 	for (int i = 0; i < num_sublists; i++) {
3580 		multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3581 		multilist_sublist_remove(mls, markers[i]);
3582 		multilist_sublist_unlock(mls);
3583 
3584 		kmem_cache_free(hdr_full_cache, markers[i]);
3585 	}
3586 	kmem_free(markers, sizeof (*markers) * num_sublists);
3587 
3588 	return (total_evicted);
3589 }
3590 
3591 /*
3592  * Flush all "evictable" data of the given type from the arc state
3593  * specified. This will not evict any "active" buffers (i.e. referenced).
3594  *
3595  * When 'retry' is set to B_FALSE, the function will make a single pass
3596  * over the state and evict any buffers that it can. Since it doesn't
3597  * continually retry the eviction, it might end up leaving some buffers
3598  * in the ARC due to lock misses.
3599  *
3600  * When 'retry' is set to B_TRUE, the function will continually retry the
3601  * eviction until *all* evictable buffers have been removed from the
3602  * state. As a result, if concurrent insertions into the state are
3603  * allowed (e.g. if the ARC isn't shutting down), this function might
3604  * wind up in an infinite loop, continually trying to evict buffers.
3605  */
3606 static uint64_t
3607 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
3608     boolean_t retry)
3609 {
3610 	uint64_t evicted = 0;
3611 
3612 	while (zfs_refcount_count(&state->arcs_esize[type]) != 0) {
3613 		evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
3614 
3615 		if (!retry)
3616 			break;
3617 	}
3618 
3619 	return (evicted);
3620 }
3621 
3622 /*
3623  * Evict the specified number of bytes from the state specified,
3624  * restricting eviction to the spa and type given. This function
3625  * prevents us from trying to evict more from a state's list than
3626  * is "evictable", and to skip evicting altogether when passed a
3627  * negative value for "bytes". In contrast, arc_evict_state() will
3628  * evict everything it can, when passed a negative value for "bytes".
3629  */
3630 static uint64_t
3631 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3632     arc_buf_contents_t type)
3633 {
3634 	int64_t delta;
3635 
3636 	if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) {
3637 		delta = MIN(zfs_refcount_count(&state->arcs_esize[type]),
3638 		    bytes);
3639 		return (arc_evict_state(state, spa, delta, type));
3640 	}
3641 
3642 	return (0);
3643 }
3644 
3645 /*
3646  * Evict metadata buffers from the cache, such that arc_meta_used is
3647  * capped by the arc_meta_limit tunable.
3648  */
3649 static uint64_t
3650 arc_adjust_meta(uint64_t meta_used)
3651 {
3652 	uint64_t total_evicted = 0;
3653 	int64_t target;
3654 
3655 	/*
3656 	 * If we're over the meta limit, we want to evict enough
3657 	 * metadata to get back under the meta limit. We don't want to
3658 	 * evict so much that we drop the MRU below arc_p, though. If
3659 	 * we're over the meta limit more than we're over arc_p, we
3660 	 * evict some from the MRU here, and some from the MFU below.
3661 	 */
3662 	target = MIN((int64_t)(meta_used - arc_meta_limit),
3663 	    (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
3664 	    zfs_refcount_count(&arc_mru->arcs_size) - arc_p));
3665 
3666 	total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3667 
3668 	/*
3669 	 * Similar to the above, we want to evict enough bytes to get us
3670 	 * below the meta limit, but not so much as to drop us below the
3671 	 * space allotted to the MFU (which is defined as arc_c - arc_p).
3672 	 */
3673 	target = MIN((int64_t)(meta_used - arc_meta_limit),
3674 	    (int64_t)(zfs_refcount_count(&arc_mfu->arcs_size) -
3675 	    (arc_c - arc_p)));
3676 
3677 	total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3678 
3679 	return (total_evicted);
3680 }
3681 
3682 /*
3683  * Return the type of the oldest buffer in the given arc state
3684  *
3685  * This function will select a random sublist of type ARC_BUFC_DATA and
3686  * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3687  * is compared, and the type which contains the "older" buffer will be
3688  * returned.
3689  */
3690 static arc_buf_contents_t
3691 arc_adjust_type(arc_state_t *state)
3692 {
3693 	multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
3694 	multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
3695 	int data_idx = multilist_get_random_index(data_ml);
3696 	int meta_idx = multilist_get_random_index(meta_ml);
3697 	multilist_sublist_t *data_mls;
3698 	multilist_sublist_t *meta_mls;
3699 	arc_buf_contents_t type;
3700 	arc_buf_hdr_t *data_hdr;
3701 	arc_buf_hdr_t *meta_hdr;
3702 
3703 	/*
3704 	 * We keep the sublist lock until we're finished, to prevent
3705 	 * the headers from being destroyed via arc_evict_state().
3706 	 */
3707 	data_mls = multilist_sublist_lock(data_ml, data_idx);
3708 	meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
3709 
3710 	/*
3711 	 * These two loops are to ensure we skip any markers that
3712 	 * might be at the tail of the lists due to arc_evict_state().
3713 	 */
3714 
3715 	for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
3716 	    data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
3717 		if (data_hdr->b_spa != 0)
3718 			break;
3719 	}
3720 
3721 	for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
3722 	    meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
3723 		if (meta_hdr->b_spa != 0)
3724 			break;
3725 	}
3726 
3727 	if (data_hdr == NULL && meta_hdr == NULL) {
3728 		type = ARC_BUFC_DATA;
3729 	} else if (data_hdr == NULL) {
3730 		ASSERT3P(meta_hdr, !=, NULL);
3731 		type = ARC_BUFC_METADATA;
3732 	} else if (meta_hdr == NULL) {
3733 		ASSERT3P(data_hdr, !=, NULL);
3734 		type = ARC_BUFC_DATA;
3735 	} else {
3736 		ASSERT3P(data_hdr, !=, NULL);
3737 		ASSERT3P(meta_hdr, !=, NULL);
3738 
3739 		/* The headers can't be on the sublist without an L1 header */
3740 		ASSERT(HDR_HAS_L1HDR(data_hdr));
3741 		ASSERT(HDR_HAS_L1HDR(meta_hdr));
3742 
3743 		if (data_hdr->b_l1hdr.b_arc_access <
3744 		    meta_hdr->b_l1hdr.b_arc_access) {
3745 			type = ARC_BUFC_DATA;
3746 		} else {
3747 			type = ARC_BUFC_METADATA;
3748 		}
3749 	}
3750 
3751 	multilist_sublist_unlock(meta_mls);
3752 	multilist_sublist_unlock(data_mls);
3753 
3754 	return (type);
3755 }
3756 
3757 /*
3758  * Evict buffers from the cache, such that arc_size is capped by arc_c.
3759  */
3760 static uint64_t
3761 arc_adjust(void)
3762 {
3763 	uint64_t total_evicted = 0;
3764 	uint64_t bytes;
3765 	int64_t target;
3766 	uint64_t asize = aggsum_value(&arc_size);
3767 	uint64_t ameta = aggsum_value(&arc_meta_used);
3768 
3769 	/*
3770 	 * If we're over arc_meta_limit, we want to correct that before
3771 	 * potentially evicting data buffers below.
3772 	 */
3773 	total_evicted += arc_adjust_meta(ameta);
3774 
3775 	/*
3776 	 * Adjust MRU size
3777 	 *
3778 	 * If we're over the target cache size, we want to evict enough
3779 	 * from the list to get back to our target size. We don't want
3780 	 * to evict too much from the MRU, such that it drops below
3781 	 * arc_p. So, if we're over our target cache size more than
3782 	 * the MRU is over arc_p, we'll evict enough to get back to
3783 	 * arc_p here, and then evict more from the MFU below.
3784 	 */
3785 	target = MIN((int64_t)(asize - arc_c),
3786 	    (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
3787 	    zfs_refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
3788 
3789 	/*
3790 	 * If we're below arc_meta_min, always prefer to evict data.
3791 	 * Otherwise, try to satisfy the requested number of bytes to
3792 	 * evict from the type which contains older buffers; in an
3793 	 * effort to keep newer buffers in the cache regardless of their
3794 	 * type. If we cannot satisfy the number of bytes from this
3795 	 * type, spill over into the next type.
3796 	 */
3797 	if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
3798 	    ameta > arc_meta_min) {
3799 		bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3800 		total_evicted += bytes;
3801 
3802 		/*
3803 		 * If we couldn't evict our target number of bytes from
3804 		 * metadata, we try to get the rest from data.
3805 		 */
3806 		target -= bytes;
3807 
3808 		total_evicted +=
3809 		    arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3810 	} else {
3811 		bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3812 		total_evicted += bytes;
3813 
3814 		/*
3815 		 * If we couldn't evict our target number of bytes from
3816 		 * data, we try to get the rest from metadata.
3817 		 */
3818 		target -= bytes;
3819 
3820 		total_evicted +=
3821 		    arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3822 	}
3823 
3824 	/*
3825 	 * Adjust MFU size
3826 	 *
3827 	 * Now that we've tried to evict enough from the MRU to get its
3828 	 * size back to arc_p, if we're still above the target cache
3829 	 * size, we evict the rest from the MFU.
3830 	 */
3831 	target = asize - arc_c;
3832 
3833 	if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
3834 	    ameta > arc_meta_min) {
3835 		bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3836 		total_evicted += bytes;
3837 
3838 		/*
3839 		 * If we couldn't evict our target number of bytes from
3840 		 * metadata, we try to get the rest from data.
3841 		 */
3842 		target -= bytes;
3843 
3844 		total_evicted +=
3845 		    arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3846 	} else {
3847 		bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3848 		total_evicted += bytes;
3849 
3850 		/*
3851 		 * If we couldn't evict our target number of bytes from
3852 		 * data, we try to get the rest from data.
3853 		 */
3854 		target -= bytes;
3855 
3856 		total_evicted +=
3857 		    arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3858 	}
3859 
3860 	/*
3861 	 * Adjust ghost lists
3862 	 *
3863 	 * In addition to the above, the ARC also defines target values
3864 	 * for the ghost lists. The sum of the mru list and mru ghost
3865 	 * list should never exceed the target size of the cache, and
3866 	 * the sum of the mru list, mfu list, mru ghost list, and mfu
3867 	 * ghost list should never exceed twice the target size of the
3868 	 * cache. The following logic enforces these limits on the ghost
3869 	 * caches, and evicts from them as needed.
3870 	 */
3871 	target = zfs_refcount_count(&arc_mru->arcs_size) +
3872 	    zfs_refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3873 
3874 	bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3875 	total_evicted += bytes;
3876 
3877 	target -= bytes;
3878 
3879 	total_evicted +=
3880 	    arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3881 
3882 	/*
3883 	 * We assume the sum of the mru list and mfu list is less than
3884 	 * or equal to arc_c (we enforced this above), which means we
3885 	 * can use the simpler of the two equations below:
3886 	 *
3887 	 *	mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3888 	 *		    mru ghost + mfu ghost <= arc_c
3889 	 */
3890 	target = zfs_refcount_count(&arc_mru_ghost->arcs_size) +
3891 	    zfs_refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3892 
3893 	bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3894 	total_evicted += bytes;
3895 
3896 	target -= bytes;
3897 
3898 	total_evicted +=
3899 	    arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3900 
3901 	return (total_evicted);
3902 }
3903 
3904 void
3905 arc_flush(spa_t *spa, boolean_t retry)
3906 {
3907 	uint64_t guid = 0;
3908 
3909 	/*
3910 	 * If retry is B_TRUE, a spa must not be specified since we have
3911 	 * no good way to determine if all of a spa's buffers have been
3912 	 * evicted from an arc state.
3913 	 */
3914 	ASSERT(!retry || spa == 0);
3915 
3916 	if (spa != NULL)
3917 		guid = spa_load_guid(spa);
3918 
3919 	(void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3920 	(void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3921 
3922 	(void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3923 	(void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3924 
3925 	(void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3926 	(void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3927 
3928 	(void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3929 	(void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3930 }
3931 
3932 static void
3933 arc_reduce_target_size(int64_t to_free)
3934 {
3935 	uint64_t asize = aggsum_value(&arc_size);
3936 	if (arc_c > arc_c_min) {
3937 
3938 		if (arc_c > arc_c_min + to_free)
3939 			atomic_add_64(&arc_c, -to_free);
3940 		else
3941 			arc_c = arc_c_min;
3942 
3943 		atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3944 		if (asize < arc_c)
3945 			arc_c = MAX(asize, arc_c_min);
3946 		if (arc_p > arc_c)
3947 			arc_p = (arc_c >> 1);
3948 		ASSERT(arc_c >= arc_c_min);
3949 		ASSERT((int64_t)arc_p >= 0);
3950 	}
3951 
3952 	if (asize > arc_c) {
3953 		/* See comment in arc_adjust_cb_check() on why lock+flag */
3954 		mutex_enter(&arc_adjust_lock);
3955 		arc_adjust_needed = B_TRUE;
3956 		mutex_exit(&arc_adjust_lock);
3957 		zthr_wakeup(arc_adjust_zthr);
3958 	}
3959 }
3960 
3961 typedef enum free_memory_reason_t {
3962 	FMR_UNKNOWN,
3963 	FMR_NEEDFREE,
3964 	FMR_LOTSFREE,
3965 	FMR_SWAPFS_MINFREE,
3966 	FMR_PAGES_PP_MAXIMUM,
3967 	FMR_HEAP_ARENA,
3968 	FMR_ZIO_ARENA,
3969 } free_memory_reason_t;
3970 
3971 int64_t last_free_memory;
3972 free_memory_reason_t last_free_reason;
3973 
3974 /*
3975  * Additional reserve of pages for pp_reserve.
3976  */
3977 int64_t arc_pages_pp_reserve = 64;
3978 
3979 /*
3980  * Additional reserve of pages for swapfs.
3981  */
3982 int64_t arc_swapfs_reserve = 64;
3983 
3984 /*
3985  * Return the amount of memory that can be consumed before reclaim will be
3986  * needed.  Positive if there is sufficient free memory, negative indicates
3987  * the amount of memory that needs to be freed up.
3988  */
3989 static int64_t
3990 arc_available_memory(void)
3991 {
3992 	int64_t lowest = INT64_MAX;
3993 	int64_t n;
3994 	free_memory_reason_t r = FMR_UNKNOWN;
3995 
3996 #ifdef _KERNEL
3997 	if (needfree > 0) {
3998 		n = PAGESIZE * (-needfree);
3999 		if (n < lowest) {
4000 			lowest = n;
4001 			r = FMR_NEEDFREE;
4002 		}
4003 	}
4004 
4005 	/*
4006 	 * check that we're out of range of the pageout scanner.  It starts to
4007 	 * schedule paging if freemem is less than lotsfree and needfree.
4008 	 * lotsfree is the high-water mark for pageout, and needfree is the
4009 	 * number of needed free pages.  We add extra pages here to make sure
4010 	 * the scanner doesn't start up while we're freeing memory.
4011 	 */
4012 	n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4013 	if (n < lowest) {
4014 		lowest = n;
4015 		r = FMR_LOTSFREE;
4016 	}
4017 
4018 	/*
4019 	 * check to make sure that swapfs has enough space so that anon
4020 	 * reservations can still succeed. anon_resvmem() checks that the
4021 	 * availrmem is greater than swapfs_minfree, and the number of reserved
4022 	 * swap pages.  We also add a bit of extra here just to prevent
4023 	 * circumstances from getting really dire.
4024 	 */
4025 	n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4026 	    desfree - arc_swapfs_reserve);
4027 	if (n < lowest) {
4028 		lowest = n;
4029 		r = FMR_SWAPFS_MINFREE;
4030 	}
4031 
4032 
4033 	/*
4034 	 * Check that we have enough availrmem that memory locking (e.g., via
4035 	 * mlock(3C) or memcntl(2)) can still succeed.  (pages_pp_maximum
4036 	 * stores the number of pages that cannot be locked; when availrmem
4037 	 * drops below pages_pp_maximum, page locking mechanisms such as
4038 	 * page_pp_lock() will fail.)
4039 	 */
4040 	n = PAGESIZE * (availrmem - pages_pp_maximum -
4041 	    arc_pages_pp_reserve);
4042 	if (n < lowest) {
4043 		lowest = n;
4044 		r = FMR_PAGES_PP_MAXIMUM;
4045 	}
4046 
4047 #if defined(__i386)
4048 	/*
4049 	 * If we're on an i386 platform, it's possible that we'll exhaust the
4050 	 * kernel heap space before we ever run out of available physical
4051 	 * memory.  Most checks of the size of the heap_area compare against
4052 	 * tune.t_minarmem, which is the minimum available real memory that we
4053 	 * can have in the system.  However, this is generally fixed at 25 pages
4054 	 * which is so low that it's useless.  In this comparison, we seek to
4055 	 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4056 	 * heap is allocated.  (Or, in the calculation, if less than 1/4th is
4057 	 * free)
4058 	 */
4059 	n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
4060 	    (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
4061 	if (n < lowest) {
4062 		lowest = n;
4063 		r = FMR_HEAP_ARENA;
4064 	}
4065 #endif
4066 
4067 	/*
4068 	 * If zio data pages are being allocated out of a separate heap segment,
4069 	 * then enforce that the size of available vmem for this arena remains
4070 	 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4071 	 *
4072 	 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4073 	 * memory (in the zio_arena) free, which can avoid memory
4074 	 * fragmentation issues.
4075 	 */
4076 	if (zio_arena != NULL) {
4077 		n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4078 		    (vmem_size(zio_arena, VMEM_ALLOC) >>
4079 		    arc_zio_arena_free_shift);
4080 		if (n < lowest) {
4081 			lowest = n;
4082 			r = FMR_ZIO_ARENA;
4083 		}
4084 	}
4085 #else
4086 	/* Every 100 calls, free a small amount */
4087 	if (spa_get_random(100) == 0)
4088 		lowest = -1024;
4089 #endif
4090 
4091 	last_free_memory = lowest;
4092 	last_free_reason = r;
4093 
4094 	return (lowest);
4095 }
4096 
4097 
4098 /*
4099  * Determine if the system is under memory pressure and is asking
4100  * to reclaim memory. A return value of B_TRUE indicates that the system
4101  * is under memory pressure and that the arc should adjust accordingly.
4102  */
4103 static boolean_t
4104 arc_reclaim_needed(void)
4105 {
4106 	return (arc_available_memory() < 0);
4107 }
4108 
4109 static void
4110 arc_kmem_reap_soon(void)
4111 {
4112 	size_t			i;
4113 	kmem_cache_t		*prev_cache = NULL;
4114 	kmem_cache_t		*prev_data_cache = NULL;
4115 	extern kmem_cache_t	*zio_buf_cache[];
4116 	extern kmem_cache_t	*zio_data_buf_cache[];
4117 	extern kmem_cache_t	*range_seg_cache;
4118 	extern kmem_cache_t	*abd_chunk_cache;
4119 
4120 #ifdef _KERNEL
4121 	if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) {
4122 		/*
4123 		 * We are exceeding our meta-data cache limit.
4124 		 * Purge some DNLC entries to release holds on meta-data.
4125 		 */
4126 		dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4127 	}
4128 #if defined(__i386)
4129 	/*
4130 	 * Reclaim unused memory from all kmem caches.
4131 	 */
4132 	kmem_reap();
4133 #endif
4134 #endif
4135 
4136 	for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4137 		if (zio_buf_cache[i] != prev_cache) {
4138 			prev_cache = zio_buf_cache[i];
4139 			kmem_cache_reap_soon(zio_buf_cache[i]);
4140 		}
4141 		if (zio_data_buf_cache[i] != prev_data_cache) {
4142 			prev_data_cache = zio_data_buf_cache[i];
4143 			kmem_cache_reap_soon(zio_data_buf_cache[i]);
4144 		}
4145 	}
4146 	kmem_cache_reap_soon(abd_chunk_cache);
4147 	kmem_cache_reap_soon(buf_cache);
4148 	kmem_cache_reap_soon(hdr_full_cache);
4149 	kmem_cache_reap_soon(hdr_l2only_cache);
4150 	kmem_cache_reap_soon(range_seg_cache);
4151 
4152 	if (zio_arena != NULL) {
4153 		/*
4154 		 * Ask the vmem arena to reclaim unused memory from its
4155 		 * quantum caches.
4156 		 */
4157 		vmem_qcache_reap(zio_arena);
4158 	}
4159 }
4160 
4161 /* ARGSUSED */
4162 static boolean_t
4163 arc_adjust_cb_check(void *arg, zthr_t *zthr)
4164 {
4165 	/*
4166 	 * This is necessary in order for the mdb ::arc dcmd to
4167 	 * show up to date information. Since the ::arc command
4168 	 * does not call the kstat's update function, without
4169 	 * this call, the command may show stale stats for the
4170 	 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4171 	 * with this change, the data might be up to 1 second
4172 	 * out of date(the arc_adjust_zthr has a maximum sleep
4173 	 * time of 1 second); but that should suffice.  The
4174 	 * arc_state_t structures can be queried directly if more
4175 	 * accurate information is needed.
4176 	 */
4177 	if (arc_ksp != NULL)
4178 		arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4179 
4180 	/*
4181 	 * We have to rely on arc_get_data_impl() to tell us when to adjust,
4182 	 * rather than checking if we are overflowing here, so that we are
4183 	 * sure to not leave arc_get_data_impl() waiting on
4184 	 * arc_adjust_waiters_cv.  If we have become "not overflowing" since
4185 	 * arc_get_data_impl() checked, we need to wake it up.  We could
4186 	 * broadcast the CV here, but arc_get_data_impl() may have not yet
4187 	 * gone to sleep.  We would need to use a mutex to ensure that this
4188 	 * function doesn't broadcast until arc_get_data_impl() has gone to
4189 	 * sleep (e.g. the arc_adjust_lock).  However, the lock ordering of
4190 	 * such a lock would necessarily be incorrect with respect to the
4191 	 * zthr_lock, which is held before this function is called, and is
4192 	 * held by arc_get_data_impl() when it calls zthr_wakeup().
4193 	 */
4194 	return (arc_adjust_needed);
4195 }
4196 
4197 /*
4198  * Keep arc_size under arc_c by running arc_adjust which evicts data
4199  * from the ARC.
4200  */
4201 /* ARGSUSED */
4202 static int
4203 arc_adjust_cb(void *arg, zthr_t *zthr)
4204 {
4205 	uint64_t evicted = 0;
4206 
4207 	/* Evict from cache */
4208 	evicted = arc_adjust();
4209 
4210 	/*
4211 	 * If evicted is zero, we couldn't evict anything
4212 	 * via arc_adjust(). This could be due to hash lock
4213 	 * collisions, but more likely due to the majority of
4214 	 * arc buffers being unevictable. Therefore, even if
4215 	 * arc_size is above arc_c, another pass is unlikely to
4216 	 * be helpful and could potentially cause us to enter an
4217 	 * infinite loop.  Additionally, zthr_iscancelled() is
4218 	 * checked here so that if the arc is shutting down, the
4219 	 * broadcast will wake any remaining arc adjust waiters.
4220 	 */
4221 	mutex_enter(&arc_adjust_lock);
4222 	arc_adjust_needed = !zthr_iscancelled(arc_adjust_zthr) &&
4223 	    evicted > 0 && aggsum_compare(&arc_size, arc_c) > 0;
4224 	if (!arc_adjust_needed) {
4225 		/*
4226 		 * We're either no longer overflowing, or we
4227 		 * can't evict anything more, so we should wake
4228 		 * up any waiters.
4229 		 */
4230 		cv_broadcast(&arc_adjust_waiters_cv);
4231 	}
4232 	mutex_exit(&arc_adjust_lock);
4233 
4234 	return (0);
4235 }
4236 
4237 /* ARGSUSED */
4238 static boolean_t
4239 arc_reap_cb_check(void *arg, zthr_t *zthr)
4240 {
4241 	int64_t free_memory = arc_available_memory();
4242 
4243 	/*
4244 	 * If a kmem reap is already active, don't schedule more.  We must
4245 	 * check for this because kmem_cache_reap_soon() won't actually
4246 	 * block on the cache being reaped (this is to prevent callers from
4247 	 * becoming implicitly blocked by a system-wide kmem reap -- which,
4248 	 * on a system with many, many full magazines, can take minutes).
4249 	 */
4250 	if (!kmem_cache_reap_active() &&
4251 	    free_memory < 0) {
4252 		arc_no_grow = B_TRUE;
4253 		arc_warm = B_TRUE;
4254 		/*
4255 		 * Wait at least zfs_grow_retry (default 60) seconds
4256 		 * before considering growing.
4257 		 */
4258 		arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4259 		return (B_TRUE);
4260 	} else if (free_memory < arc_c >> arc_no_grow_shift) {
4261 		arc_no_grow = B_TRUE;
4262 	} else if (gethrtime() >= arc_growtime) {
4263 		arc_no_grow = B_FALSE;
4264 	}
4265 
4266 	return (B_FALSE);
4267 }
4268 
4269 /*
4270  * Keep enough free memory in the system by reaping the ARC's kmem
4271  * caches.  To cause more slabs to be reapable, we may reduce the
4272  * target size of the cache (arc_c), causing the arc_adjust_cb()
4273  * to free more buffers.
4274  */
4275 /* ARGSUSED */
4276 static int
4277 arc_reap_cb(void *arg, zthr_t *zthr)
4278 {
4279 	int64_t free_memory;
4280 
4281 	/*
4282 	 * Kick off asynchronous kmem_reap()'s of all our caches.
4283 	 */
4284 	arc_kmem_reap_soon();
4285 
4286 	/*
4287 	 * Wait at least arc_kmem_cache_reap_retry_ms between
4288 	 * arc_kmem_reap_soon() calls. Without this check it is possible to
4289 	 * end up in a situation where we spend lots of time reaping
4290 	 * caches, while we're near arc_c_min.  Waiting here also gives the
4291 	 * subsequent free memory check a chance of finding that the
4292 	 * asynchronous reap has already freed enough memory, and we don't
4293 	 * need to call arc_reduce_target_size().
4294 	 */
4295 	delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000);
4296 
4297 	/*
4298 	 * Reduce the target size as needed to maintain the amount of free
4299 	 * memory in the system at a fraction of the arc_size (1/128th by
4300 	 * default).  If oversubscribed (free_memory < 0) then reduce the
4301 	 * target arc_size by the deficit amount plus the fractional
4302 	 * amount.  If free memory is positive but less then the fractional
4303 	 * amount, reduce by what is needed to hit the fractional amount.
4304 	 */
4305 	free_memory = arc_available_memory();
4306 
4307 	int64_t to_free =
4308 	    (arc_c >> arc_shrink_shift) - free_memory;
4309 	if (to_free > 0) {
4310 #ifdef _KERNEL
4311 		to_free = MAX(to_free, ptob(needfree));
4312 #endif
4313 		arc_reduce_target_size(to_free);
4314 	}
4315 
4316 	return (0);
4317 }
4318 
4319 /*
4320  * Adapt arc info given the number of bytes we are trying to add and
4321  * the state that we are comming from.  This function is only called
4322  * when we are adding new content to the cache.
4323  */
4324 static void
4325 arc_adapt(int bytes, arc_state_t *state)
4326 {
4327 	int mult;
4328 	uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4329 	int64_t mrug_size = zfs_refcount_count(&arc_mru_ghost->arcs_size);
4330 	int64_t mfug_size = zfs_refcount_count(&arc_mfu_ghost->arcs_size);
4331 
4332 	if (state == arc_l2c_only)
4333 		return;
4334 
4335 	ASSERT(bytes > 0);
4336 	/*
4337 	 * Adapt the target size of the MRU list:
4338 	 *	- if we just hit in the MRU ghost list, then increase
4339 	 *	  the target size of the MRU list.
4340 	 *	- if we just hit in the MFU ghost list, then increase
4341 	 *	  the target size of the MFU list by decreasing the
4342 	 *	  target size of the MRU list.
4343 	 */
4344 	if (state == arc_mru_ghost) {
4345 		mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4346 		mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4347 
4348 		arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4349 	} else if (state == arc_mfu_ghost) {
4350 		uint64_t delta;
4351 
4352 		mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4353 		mult = MIN(mult, 10);
4354 
4355 		delta = MIN(bytes * mult, arc_p);
4356 		arc_p = MAX(arc_p_min, arc_p - delta);
4357 	}
4358 	ASSERT((int64_t)arc_p >= 0);
4359 
4360 	/*
4361 	 * Wake reap thread if we do not have any available memory
4362 	 */
4363 	if (arc_reclaim_needed()) {
4364 		zthr_wakeup(arc_reap_zthr);
4365 		return;
4366 	}
4367 
4368 
4369 	if (arc_no_grow)
4370 		return;
4371 
4372 	if (arc_c >= arc_c_max)
4373 		return;
4374 
4375 	/*
4376 	 * If we're within (2 * maxblocksize) bytes of the target
4377 	 * cache size, increment the target cache size
4378 	 */
4379 	if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >
4380 	    0) {
4381 		atomic_add_64(&arc_c, (int64_t)bytes);
4382 		if (arc_c > arc_c_max)
4383 			arc_c = arc_c_max;
4384 		else if (state == arc_anon)
4385 			atomic_add_64(&arc_p, (int64_t)bytes);
4386 		if (arc_p > arc_c)
4387 			arc_p = arc_c;
4388 	}
4389 	ASSERT((int64_t)arc_p >= 0);
4390 }
4391 
4392 /*
4393  * Check if arc_size has grown past our upper threshold, determined by
4394  * zfs_arc_overflow_shift.
4395  */
4396 static boolean_t
4397 arc_is_overflowing(void)
4398 {
4399 	/* Always allow at least one block of overflow */
4400 	uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4401 	    arc_c >> zfs_arc_overflow_shift);
4402 
4403 	/*
4404 	 * We just compare the lower bound here for performance reasons. Our
4405 	 * primary goals are to make sure that the arc never grows without
4406 	 * bound, and that it can reach its maximum size. This check
4407 	 * accomplishes both goals. The maximum amount we could run over by is
4408 	 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
4409 	 * in the ARC. In practice, that's in the tens of MB, which is low
4410 	 * enough to be safe.
4411 	 */
4412 	return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
4413 }
4414 
4415 static abd_t *
4416 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4417 {
4418 	arc_buf_contents_t type = arc_buf_type(hdr);
4419 
4420 	arc_get_data_impl(hdr, size, tag);
4421 	if (type == ARC_BUFC_METADATA) {
4422 		return (abd_alloc(size, B_TRUE));
4423 	} else {
4424 		ASSERT(type == ARC_BUFC_DATA);
4425 		return (abd_alloc(size, B_FALSE));
4426 	}
4427 }
4428 
4429 static void *
4430 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4431 {
4432 	arc_buf_contents_t type = arc_buf_type(hdr);
4433 
4434 	arc_get_data_impl(hdr, size, tag);
4435 	if (type == ARC_BUFC_METADATA) {
4436 		return (zio_buf_alloc(size));
4437 	} else {
4438 		ASSERT(type == ARC_BUFC_DATA);
4439 		return (zio_data_buf_alloc(size));
4440 	}
4441 }
4442 
4443 /*
4444  * Allocate a block and return it to the caller. If we are hitting the
4445  * hard limit for the cache size, we must sleep, waiting for the eviction
4446  * thread to catch up. If we're past the target size but below the hard
4447  * limit, we'll only signal the reclaim thread and continue on.
4448  */
4449 static void
4450 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4451 {
4452 	arc_state_t *state = hdr->b_l1hdr.b_state;
4453 	arc_buf_contents_t type = arc_buf_type(hdr);
4454 
4455 	arc_adapt(size, state);
4456 
4457 	/*
4458 	 * If arc_size is currently overflowing, and has grown past our
4459 	 * upper limit, we must be adding data faster than the evict
4460 	 * thread can evict. Thus, to ensure we don't compound the
4461 	 * problem by adding more data and forcing arc_size to grow even
4462 	 * further past it's target size, we halt and wait for the
4463 	 * eviction thread to catch up.
4464 	 *
4465 	 * It's also possible that the reclaim thread is unable to evict
4466 	 * enough buffers to get arc_size below the overflow limit (e.g.
4467 	 * due to buffers being un-evictable, or hash lock collisions).
4468 	 * In this case, we want to proceed regardless if we're
4469 	 * overflowing; thus we don't use a while loop here.
4470 	 */
4471 	if (arc_is_overflowing()) {
4472 		mutex_enter(&arc_adjust_lock);
4473 
4474 		/*
4475 		 * Now that we've acquired the lock, we may no longer be
4476 		 * over the overflow limit, lets check.
4477 		 *
4478 		 * We're ignoring the case of spurious wake ups. If that
4479 		 * were to happen, it'd let this thread consume an ARC
4480 		 * buffer before it should have (i.e. before we're under
4481 		 * the overflow limit and were signalled by the reclaim
4482 		 * thread). As long as that is a rare occurrence, it
4483 		 * shouldn't cause any harm.
4484 		 */
4485 		if (arc_is_overflowing()) {
4486 			arc_adjust_needed = B_TRUE;
4487 			zthr_wakeup(arc_adjust_zthr);
4488 			(void) cv_wait(&arc_adjust_waiters_cv,
4489 			    &arc_adjust_lock);
4490 		}
4491 		mutex_exit(&arc_adjust_lock);
4492 	}
4493 
4494 	VERIFY3U(hdr->b_type, ==, type);
4495 	if (type == ARC_BUFC_METADATA) {
4496 		arc_space_consume(size, ARC_SPACE_META);
4497 	} else {
4498 		arc_space_consume(size, ARC_SPACE_DATA);
4499 	}
4500 
4501 	/*
4502 	 * Update the state size.  Note that ghost states have a
4503 	 * "ghost size" and so don't need to be updated.
4504 	 */
4505 	if (!GHOST_STATE(state)) {
4506 
4507 		(void) zfs_refcount_add_many(&state->arcs_size, size, tag);
4508 
4509 		/*
4510 		 * If this is reached via arc_read, the link is
4511 		 * protected by the hash lock. If reached via
4512 		 * arc_buf_alloc, the header should not be accessed by
4513 		 * any other thread. And, if reached via arc_read_done,
4514 		 * the hash lock will protect it if it's found in the
4515 		 * hash table; otherwise no other thread should be
4516 		 * trying to [add|remove]_reference it.
4517 		 */
4518 		if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4519 			ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4520 			(void) zfs_refcount_add_many(&state->arcs_esize[type],
4521 			    size, tag);
4522 		}
4523 
4524 		/*
4525 		 * If we are growing the cache, and we are adding anonymous
4526 		 * data, and we have outgrown arc_p, update arc_p
4527 		 */
4528 		if (aggsum_compare(&arc_size, arc_c) < 0 &&
4529 		    hdr->b_l1hdr.b_state == arc_anon &&
4530 		    (zfs_refcount_count(&arc_anon->arcs_size) +
4531 		    zfs_refcount_count(&arc_mru->arcs_size) > arc_p))
4532 			arc_p = MIN(arc_c, arc_p + size);
4533 	}
4534 }
4535 
4536 static void
4537 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
4538 {
4539 	arc_free_data_impl(hdr, size, tag);
4540 	abd_free(abd);
4541 }
4542 
4543 static void
4544 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
4545 {
4546 	arc_buf_contents_t type = arc_buf_type(hdr);
4547 
4548 	arc_free_data_impl(hdr, size, tag);
4549 	if (type == ARC_BUFC_METADATA) {
4550 		zio_buf_free(buf, size);
4551 	} else {
4552 		ASSERT(type == ARC_BUFC_DATA);
4553 		zio_data_buf_free(buf, size);
4554 	}
4555 }
4556 
4557 /*
4558  * Free the arc data buffer.
4559  */
4560 static void
4561 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4562 {
4563 	arc_state_t *state = hdr->b_l1hdr.b_state;
4564 	arc_buf_contents_t type = arc_buf_type(hdr);
4565 
4566 	/* protected by hash lock, if in the hash table */
4567 	if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4568 		ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4569 		ASSERT(state != arc_anon && state != arc_l2c_only);
4570 
4571 		(void) zfs_refcount_remove_many(&state->arcs_esize[type],
4572 		    size, tag);
4573 	}
4574 	(void) zfs_refcount_remove_many(&state->arcs_size, size, tag);
4575 
4576 	VERIFY3U(hdr->b_type, ==, type);
4577 	if (type == ARC_BUFC_METADATA) {
4578 		arc_space_return(size, ARC_SPACE_META);
4579 	} else {
4580 		ASSERT(type == ARC_BUFC_DATA);
4581 		arc_space_return(size, ARC_SPACE_DATA);
4582 	}
4583 }
4584 
4585 /*
4586  * This routine is called whenever a buffer is accessed.
4587  * NOTE: the hash lock is dropped in this function.
4588  */
4589 static void
4590 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4591 {
4592 	clock_t now;
4593 
4594 	ASSERT(MUTEX_HELD(hash_lock));
4595 	ASSERT(HDR_HAS_L1HDR(hdr));
4596 
4597 	if (hdr->b_l1hdr.b_state == arc_anon) {
4598 		/*
4599 		 * This buffer is not in the cache, and does not
4600 		 * appear in our "ghost" list.  Add the new buffer
4601 		 * to the MRU state.
4602 		 */
4603 
4604 		ASSERT0(hdr->b_l1hdr.b_arc_access);
4605 		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4606 		DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4607 		arc_change_state(arc_mru, hdr, hash_lock);
4608 
4609 	} else if (hdr->b_l1hdr.b_state == arc_mru) {
4610 		now = ddi_get_lbolt();
4611 
4612 		/*
4613 		 * If this buffer is here because of a prefetch, then either:
4614 		 * - clear the flag if this is a "referencing" read
4615 		 *   (any subsequent access will bump this into the MFU state).
4616 		 * or
4617 		 * - move the buffer to the head of the list if this is
4618 		 *   another prefetch (to make it less likely to be evicted).
4619 		 */
4620 		if (HDR_PREFETCH(hdr)) {
4621 			if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4622 				/* link protected by hash lock */
4623 				ASSERT(multilist_link_active(
4624 				    &hdr->b_l1hdr.b_arc_node));
4625 			} else {
4626 				arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4627 				ARCSTAT_BUMP(arcstat_mru_hits);
4628 			}
4629 			hdr->b_l1hdr.b_arc_access = now;
4630 			return;
4631 		}
4632 
4633 		/*
4634 		 * This buffer has been "accessed" only once so far,
4635 		 * but it is still in the cache. Move it to the MFU
4636 		 * state.
4637 		 */
4638 		if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
4639 			/*
4640 			 * More than 125ms have passed since we
4641 			 * instantiated this buffer.  Move it to the
4642 			 * most frequently used state.
4643 			 */
4644 			hdr->b_l1hdr.b_arc_access = now;
4645 			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4646 			arc_change_state(arc_mfu, hdr, hash_lock);
4647 		}
4648 		ARCSTAT_BUMP(arcstat_mru_hits);
4649 	} else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
4650 		arc_state_t	*new_state;
4651 		/*
4652 		 * This buffer has been "accessed" recently, but
4653 		 * was evicted from the cache.  Move it to the
4654 		 * MFU state.
4655 		 */
4656 
4657 		if (HDR_PREFETCH(hdr)) {
4658 			new_state = arc_mru;
4659 			if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
4660 				arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4661 			DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4662 		} else {
4663 			new_state = arc_mfu;
4664 			DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4665 		}
4666 
4667 		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4668 		arc_change_state(new_state, hdr, hash_lock);
4669 
4670 		ARCSTAT_BUMP(arcstat_mru_ghost_hits);
4671 	} else if (hdr->b_l1hdr.b_state == arc_mfu) {
4672 		/*
4673 		 * This buffer has been accessed more than once and is
4674 		 * still in the cache.  Keep it in the MFU state.
4675 		 *
4676 		 * NOTE: an add_reference() that occurred when we did
4677 		 * the arc_read() will have kicked this off the list.
4678 		 * If it was a prefetch, we will explicitly move it to
4679 		 * the head of the list now.
4680 		 */
4681 		if ((HDR_PREFETCH(hdr)) != 0) {
4682 			ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4683 			/* link protected by hash_lock */
4684 			ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4685 		}
4686 		ARCSTAT_BUMP(arcstat_mfu_hits);
4687 		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4688 	} else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
4689 		arc_state_t	*new_state = arc_mfu;
4690 		/*
4691 		 * This buffer has been accessed more than once but has
4692 		 * been evicted from the cache.  Move it back to the
4693 		 * MFU state.
4694 		 */
4695 
4696 		if (HDR_PREFETCH(hdr)) {
4697 			/*
4698 			 * This is a prefetch access...
4699 			 * move this block back to the MRU state.
4700 			 */
4701 			ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
4702 			new_state = arc_mru;
4703 		}
4704 
4705 		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4706 		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4707 		arc_change_state(new_state, hdr, hash_lock);
4708 
4709 		ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
4710 	} else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
4711 		/*
4712 		 * This buffer is on the 2nd Level ARC.
4713 		 */
4714 
4715 		hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4716 		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4717 		arc_change_state(arc_mfu, hdr, hash_lock);
4718 	} else {
4719 		ASSERT(!"invalid arc state");
4720 	}
4721 }
4722 
4723 /*
4724  * This routine is called by dbuf_hold() to update the arc_access() state
4725  * which otherwise would be skipped for entries in the dbuf cache.
4726  */
4727 void
4728 arc_buf_access(arc_buf_t *buf)
4729 {
4730 	mutex_enter(&buf->b_evict_lock);
4731 	arc_buf_hdr_t *hdr = buf->b_hdr;
4732 
4733 	/*
4734 	 * Avoid taking the hash_lock when possible as an optimization.
4735 	 * The header must be checked again under the hash_lock in order
4736 	 * to handle the case where it is concurrently being released.
4737 	 */
4738 	if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
4739 		mutex_exit(&buf->b_evict_lock);
4740 		return;
4741 	}
4742 
4743 	kmutex_t *hash_lock = HDR_LOCK(hdr);
4744 	mutex_enter(hash_lock);
4745 
4746 	if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
4747 		mutex_exit(hash_lock);
4748 		mutex_exit(&buf->b_evict_lock);
4749 		ARCSTAT_BUMP(arcstat_access_skip);
4750 		return;
4751 	}
4752 
4753 	mutex_exit(&buf->b_evict_lock);
4754 
4755 	ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4756 	    hdr->b_l1hdr.b_state == arc_mfu);
4757 
4758 	DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4759 	arc_access(hdr, hash_lock);
4760 	mutex_exit(hash_lock);
4761 
4762 	ARCSTAT_BUMP(arcstat_hits);
4763 	ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4764 	    demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
4765 }
4766 
4767 /* a generic arc_done_func_t which you can use */
4768 /* ARGSUSED */
4769 void
4770 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
4771 {
4772 	if (zio == NULL || zio->io_error == 0)
4773 		bcopy(buf->b_data, arg, arc_buf_size(buf));
4774 	arc_buf_destroy(buf, arg);
4775 }
4776 
4777 /* a generic arc_done_func_t */
4778 void
4779 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
4780 {
4781 	arc_buf_t **bufp = arg;
4782 	if (buf == NULL) {
4783 		ASSERT(zio == NULL || zio->io_error != 0);
4784 		*bufp = NULL;
4785 	} else {
4786 		ASSERT(zio == NULL || zio->io_error == 0);
4787 		*bufp = buf;
4788 		ASSERT(buf->b_data != NULL);
4789 	}
4790 }
4791 
4792 static void
4793 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
4794 {
4795 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
4796 		ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
4797 		ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
4798 	} else {
4799 		if (HDR_COMPRESSION_ENABLED(hdr)) {
4800 			ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
4801 			    BP_GET_COMPRESS(bp));
4802 		}
4803 		ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
4804 		ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
4805 	}
4806 }
4807 
4808 static void
4809 arc_read_done(zio_t *zio)
4810 {
4811 	arc_buf_hdr_t	*hdr = zio->io_private;
4812 	kmutex_t	*hash_lock = NULL;
4813 	arc_callback_t	*callback_list;
4814 	arc_callback_t	*acb;
4815 	boolean_t	freeable = B_FALSE;
4816 	boolean_t	no_zio_error = (zio->io_error == 0);
4817 
4818 	/*
4819 	 * The hdr was inserted into hash-table and removed from lists
4820 	 * prior to starting I/O.  We should find this header, since
4821 	 * it's in the hash table, and it should be legit since it's
4822 	 * not possible to evict it during the I/O.  The only possible
4823 	 * reason for it not to be found is if we were freed during the
4824 	 * read.
4825 	 */
4826 	if (HDR_IN_HASH_TABLE(hdr)) {
4827 		ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
4828 		ASSERT3U(hdr->b_dva.dva_word[0], ==,
4829 		    BP_IDENTITY(zio->io_bp)->dva_word[0]);
4830 		ASSERT3U(hdr->b_dva.dva_word[1], ==,
4831 		    BP_IDENTITY(zio->io_bp)->dva_word[1]);
4832 
4833 		arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
4834 		    &hash_lock);
4835 
4836 		ASSERT((found == hdr &&
4837 		    DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
4838 		    (found == hdr && HDR_L2_READING(hdr)));
4839 		ASSERT3P(hash_lock, !=, NULL);
4840 	}
4841 
4842 	if (no_zio_error) {
4843 		/* byteswap if necessary */
4844 		if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
4845 			if (BP_GET_LEVEL(zio->io_bp) > 0) {
4846 				hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
4847 			} else {
4848 				hdr->b_l1hdr.b_byteswap =
4849 				    DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
4850 			}
4851 		} else {
4852 			hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
4853 		}
4854 	}
4855 
4856 	arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
4857 	if (l2arc_noprefetch && HDR_PREFETCH(hdr))
4858 		arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
4859 
4860 	callback_list = hdr->b_l1hdr.b_acb;
4861 	ASSERT3P(callback_list, !=, NULL);
4862 
4863 	if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) {
4864 		/*
4865 		 * Only call arc_access on anonymous buffers.  This is because
4866 		 * if we've issued an I/O for an evicted buffer, we've already
4867 		 * called arc_access (to prevent any simultaneous readers from
4868 		 * getting confused).
4869 		 */
4870 		arc_access(hdr, hash_lock);
4871 	}
4872 
4873 	/*
4874 	 * If a read request has a callback (i.e. acb_done is not NULL), then we
4875 	 * make a buf containing the data according to the parameters which were
4876 	 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
4877 	 * aren't needlessly decompressing the data multiple times.
4878 	 */
4879 	int callback_cnt = 0;
4880 	for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
4881 		if (!acb->acb_done)
4882 			continue;
4883 
4884 		/* This is a demand read since prefetches don't use callbacks */
4885 		callback_cnt++;
4886 
4887 		if (no_zio_error) {
4888 			int error = arc_buf_alloc_impl(hdr, acb->acb_private,
4889 			    acb->acb_compressed, zio->io_error == 0,
4890 			    &acb->acb_buf);
4891 			if (error != 0) {
4892 				/*
4893 				 * Decompression failed.  Set io_error
4894 				 * so that when we call acb_done (below),
4895 				 * we will indicate that the read failed.
4896 				 * Note that in the unusual case where one
4897 				 * callback is compressed and another
4898 				 * uncompressed, we will mark all of them
4899 				 * as failed, even though the uncompressed
4900 				 * one can't actually fail.  In this case,
4901 				 * the hdr will not be anonymous, because
4902 				 * if there are multiple callbacks, it's
4903 				 * because multiple threads found the same
4904 				 * arc buf in the hash table.
4905 				 */
4906 				zio->io_error = error;
4907 			}
4908 		}
4909 	}
4910 	/*
4911 	 * If there are multiple callbacks, we must have the hash lock,
4912 	 * because the only way for multiple threads to find this hdr is
4913 	 * in the hash table.  This ensures that if there are multiple
4914 	 * callbacks, the hdr is not anonymous.  If it were anonymous,
4915 	 * we couldn't use arc_buf_destroy() in the error case below.
4916 	 */
4917 	ASSERT(callback_cnt < 2 || hash_lock != NULL);
4918 
4919 	hdr->b_l1hdr.b_acb = NULL;
4920 	arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
4921 	if (callback_cnt == 0) {
4922 		ASSERT(HDR_PREFETCH(hdr));
4923 		ASSERT0(hdr->b_l1hdr.b_bufcnt);
4924 		ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
4925 	}
4926 
4927 	ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
4928 	    callback_list != NULL);
4929 
4930 	if (no_zio_error) {
4931 		arc_hdr_verify(hdr, zio->io_bp);
4932 	} else {
4933 		arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
4934 		if (hdr->b_l1hdr.b_state != arc_anon)
4935 			arc_change_state(arc_anon, hdr, hash_lock);
4936 		if (HDR_IN_HASH_TABLE(hdr))
4937 			buf_hash_remove(hdr);
4938 		freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4939 	}
4940 
4941 	/*
4942 	 * Broadcast before we drop the hash_lock to avoid the possibility
4943 	 * that the hdr (and hence the cv) might be freed before we get to
4944 	 * the cv_broadcast().
4945 	 */
4946 	cv_broadcast(&hdr->b_l1hdr.b_cv);
4947 
4948 	if (hash_lock != NULL) {
4949 		mutex_exit(hash_lock);
4950 	} else {
4951 		/*
4952 		 * This block was freed while we waited for the read to
4953 		 * complete.  It has been removed from the hash table and
4954 		 * moved to the anonymous state (so that it won't show up
4955 		 * in the cache).
4956 		 */
4957 		ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
4958 		freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4959 	}
4960 
4961 	/* execute each callback and free its structure */
4962 	while ((acb = callback_list) != NULL) {
4963 		if (acb->acb_done != NULL) {
4964 			if (zio->io_error != 0 && acb->acb_buf != NULL) {
4965 				/*
4966 				 * If arc_buf_alloc_impl() fails during
4967 				 * decompression, the buf will still be
4968 				 * allocated, and needs to be freed here.
4969 				 */
4970 				arc_buf_destroy(acb->acb_buf, acb->acb_private);
4971 				acb->acb_buf = NULL;
4972 			}
4973 			acb->acb_done(zio, acb->acb_buf, acb->acb_private);
4974 		}
4975 
4976 		if (acb->acb_zio_dummy != NULL) {
4977 			acb->acb_zio_dummy->io_error = zio->io_error;
4978 			zio_nowait(acb->acb_zio_dummy);
4979 		}
4980 
4981 		callback_list = acb->acb_next;
4982 		kmem_free(acb, sizeof (arc_callback_t));
4983 	}
4984 
4985 	if (freeable)
4986 		arc_hdr_destroy(hdr);
4987 }
4988 
4989 /*
4990  * "Read" the block at the specified DVA (in bp) via the
4991  * cache.  If the block is found in the cache, invoke the provided
4992  * callback immediately and return.  Note that the `zio' parameter
4993  * in the callback will be NULL in this case, since no IO was
4994  * required.  If the block is not in the cache pass the read request
4995  * on to the spa with a substitute callback function, so that the
4996  * requested block will be added to the cache.
4997  *
4998  * If a read request arrives for a block that has a read in-progress,
4999  * either wait for the in-progress read to complete (and return the
5000  * results); or, if this is a read with a "done" func, add a record
5001  * to the read to invoke the "done" func when the read completes,
5002  * and return; or just return.
5003  *
5004  * arc_read_done() will invoke all the requested "done" functions
5005  * for readers of this block.
5006  */
5007 int
5008 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
5009     void *private, zio_priority_t priority, int zio_flags,
5010     arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5011 {
5012 	arc_buf_hdr_t *hdr = NULL;
5013 	kmutex_t *hash_lock = NULL;
5014 	zio_t *rzio;
5015 	uint64_t guid = spa_load_guid(spa);
5016 	boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0;
5017 
5018 	ASSERT(!BP_IS_EMBEDDED(bp) ||
5019 	    BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5020 
5021 top:
5022 	if (!BP_IS_EMBEDDED(bp)) {
5023 		/*
5024 		 * Embedded BP's have no DVA and require no I/O to "read".
5025 		 * Create an anonymous arc buf to back it.
5026 		 */
5027 		hdr = buf_hash_find(guid, bp, &hash_lock);
5028 	}
5029 
5030 	if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) {
5031 		arc_buf_t *buf = NULL;
5032 		*arc_flags |= ARC_FLAG_CACHED;
5033 
5034 		if (HDR_IO_IN_PROGRESS(hdr)) {
5035 
5036 			if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5037 			    priority == ZIO_PRIORITY_SYNC_READ) {
5038 				/*
5039 				 * This sync read must wait for an
5040 				 * in-progress async read (e.g. a predictive
5041 				 * prefetch).  Async reads are queued
5042 				 * separately at the vdev_queue layer, so
5043 				 * this is a form of priority inversion.
5044 				 * Ideally, we would "inherit" the demand
5045 				 * i/o's priority by moving the i/o from
5046 				 * the async queue to the synchronous queue,
5047 				 * but there is currently no mechanism to do
5048 				 * so.  Track this so that we can evaluate
5049 				 * the magnitude of this potential performance
5050 				 * problem.
5051 				 *
5052 				 * Note that if the prefetch i/o is already
5053 				 * active (has been issued to the device),
5054 				 * the prefetch improved performance, because
5055 				 * we issued it sooner than we would have
5056 				 * without the prefetch.
5057 				 */
5058 				DTRACE_PROBE1(arc__sync__wait__for__async,
5059 				    arc_buf_hdr_t *, hdr);
5060 				ARCSTAT_BUMP(arcstat_sync_wait_for_async);
5061 			}
5062 			if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5063 				arc_hdr_clear_flags(hdr,
5064 				    ARC_FLAG_PREDICTIVE_PREFETCH);
5065 			}
5066 
5067 			if (*arc_flags & ARC_FLAG_WAIT) {
5068 				cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5069 				mutex_exit(hash_lock);
5070 				goto top;
5071 			}
5072 			ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5073 
5074 			if (done) {
5075 				arc_callback_t *acb = NULL;
5076 
5077 				acb = kmem_zalloc(sizeof (arc_callback_t),
5078 				    KM_SLEEP);
5079 				acb->acb_done = done;
5080 				acb->acb_private = private;
5081 				acb->acb_compressed = compressed_read;
5082 				if (pio != NULL)
5083 					acb->acb_zio_dummy = zio_null(pio,
5084 					    spa, NULL, NULL, NULL, zio_flags);
5085 
5086 				ASSERT3P(acb->acb_done, !=, NULL);
5087 				acb->acb_next = hdr->b_l1hdr.b_acb;
5088 				hdr->b_l1hdr.b_acb = acb;
5089 				mutex_exit(hash_lock);
5090 				return (0);
5091 			}
5092 			mutex_exit(hash_lock);
5093 			return (0);
5094 		}
5095 
5096 		ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5097 		    hdr->b_l1hdr.b_state == arc_mfu);
5098 
5099 		if (done) {
5100 			if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5101 				/*
5102 				 * This is a demand read which does not have to
5103 				 * wait for i/o because we did a predictive
5104 				 * prefetch i/o for it, which has completed.
5105 				 */
5106 				DTRACE_PROBE1(
5107 				    arc__demand__hit__predictive__prefetch,
5108 				    arc_buf_hdr_t *, hdr);
5109 				ARCSTAT_BUMP(
5110 				    arcstat_demand_hit_predictive_prefetch);
5111 				arc_hdr_clear_flags(hdr,
5112 				    ARC_FLAG_PREDICTIVE_PREFETCH);
5113 			}
5114 			ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5115 
5116 			/* Get a buf with the desired data in it. */
5117 			VERIFY0(arc_buf_alloc_impl(hdr, private,
5118 			    compressed_read, B_TRUE, &buf));
5119 		} else if (*arc_flags & ARC_FLAG_PREFETCH &&
5120 		    zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5121 			arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5122 		}
5123 		DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5124 		arc_access(hdr, hash_lock);
5125 		if (*arc_flags & ARC_FLAG_L2CACHE)
5126 			arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5127 		mutex_exit(hash_lock);
5128 		ARCSTAT_BUMP(arcstat_hits);
5129 		ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5130 		    demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5131 		    data, metadata, hits);
5132 
5133 		if (done)
5134 			done(NULL, buf, private);
5135 	} else {
5136 		uint64_t lsize = BP_GET_LSIZE(bp);
5137 		uint64_t psize = BP_GET_PSIZE(bp);
5138 		arc_callback_t *acb;
5139 		vdev_t *vd = NULL;
5140 		uint64_t addr = 0;
5141 		boolean_t devw = B_FALSE;
5142 		uint64_t size;
5143 
5144 		if (hdr == NULL) {
5145 			/* this block is not in the cache */
5146 			arc_buf_hdr_t *exists = NULL;
5147 			arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5148 			hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5149 			    BP_GET_COMPRESS(bp), type);
5150 
5151 			if (!BP_IS_EMBEDDED(bp)) {
5152 				hdr->b_dva = *BP_IDENTITY(bp);
5153 				hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5154 				exists = buf_hash_insert(hdr, &hash_lock);
5155 			}
5156 			if (exists != NULL) {
5157 				/* somebody beat us to the hash insert */
5158 				mutex_exit(hash_lock);
5159 				buf_discard_identity(hdr);
5160 				arc_hdr_destroy(hdr);
5161 				goto top; /* restart the IO request */
5162 			}
5163 		} else {
5164 			/*
5165 			 * This block is in the ghost cache. If it was L2-only
5166 			 * (and thus didn't have an L1 hdr), we realloc the
5167 			 * header to add an L1 hdr.
5168 			 */
5169 			if (!HDR_HAS_L1HDR(hdr)) {
5170 				hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5171 				    hdr_full_cache);
5172 			}
5173 			ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5174 			ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5175 			ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5176 			ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5177 			ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5178 			ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5179 
5180 			/*
5181 			 * This is a delicate dance that we play here.
5182 			 * This hdr is in the ghost list so we access it
5183 			 * to move it out of the ghost list before we
5184 			 * initiate the read. If it's a prefetch then
5185 			 * it won't have a callback so we'll remove the
5186 			 * reference that arc_buf_alloc_impl() created. We
5187 			 * do this after we've called arc_access() to
5188 			 * avoid hitting an assert in remove_reference().
5189 			 */
5190 			arc_access(hdr, hash_lock);
5191 			arc_hdr_alloc_pabd(hdr);
5192 		}
5193 		ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5194 		size = arc_hdr_size(hdr);
5195 
5196 		/*
5197 		 * If compression is enabled on the hdr, then will do
5198 		 * RAW I/O and will store the compressed data in the hdr's
5199 		 * data block. Otherwise, the hdr's data block will contain
5200 		 * the uncompressed data.
5201 		 */
5202 		if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5203 			zio_flags |= ZIO_FLAG_RAW;
5204 		}
5205 
5206 		if (*arc_flags & ARC_FLAG_PREFETCH)
5207 			arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5208 		if (*arc_flags & ARC_FLAG_L2CACHE)
5209 			arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5210 		if (BP_GET_LEVEL(bp) > 0)
5211 			arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5212 		if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5213 			arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5214 		ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5215 
5216 		acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5217 		acb->acb_done = done;
5218 		acb->acb_private = private;
5219 		acb->acb_compressed = compressed_read;
5220 
5221 		ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5222 		hdr->b_l1hdr.b_acb = acb;
5223 		arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5224 
5225 		if (HDR_HAS_L2HDR(hdr) &&
5226 		    (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5227 			devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5228 			addr = hdr->b_l2hdr.b_daddr;
5229 			/*
5230 			 * Lock out L2ARC device removal.
5231 			 */
5232 			if (vdev_is_dead(vd) ||
5233 			    !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5234 				vd = NULL;
5235 		}
5236 
5237 		if (priority == ZIO_PRIORITY_ASYNC_READ)
5238 			arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5239 		else
5240 			arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5241 
5242 		if (hash_lock != NULL)
5243 			mutex_exit(hash_lock);
5244 
5245 		/*
5246 		 * At this point, we have a level 1 cache miss.  Try again in
5247 		 * L2ARC if possible.
5248 		 */
5249 		ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5250 
5251 		DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5252 		    uint64_t, lsize, zbookmark_phys_t *, zb);
5253 		ARCSTAT_BUMP(arcstat_misses);
5254 		ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5255 		    demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5256 		    data, metadata, misses);
5257 
5258 		if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5259 			/*
5260 			 * Read from the L2ARC if the following are true:
5261 			 * 1. The L2ARC vdev was previously cached.
5262 			 * 2. This buffer still has L2ARC metadata.
5263 			 * 3. This buffer isn't currently writing to the L2ARC.
5264 			 * 4. The L2ARC entry wasn't evicted, which may
5265 			 *    also have invalidated the vdev.
5266 			 * 5. This isn't prefetch and l2arc_noprefetch is set.
5267 			 */
5268 			if (HDR_HAS_L2HDR(hdr) &&
5269 			    !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5270 			    !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5271 				l2arc_read_callback_t *cb;
5272 				abd_t *abd;
5273 				uint64_t asize;
5274 
5275 				DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5276 				ARCSTAT_BUMP(arcstat_l2_hits);
5277 
5278 				cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5279 				    KM_SLEEP);
5280 				cb->l2rcb_hdr = hdr;
5281 				cb->l2rcb_bp = *bp;
5282 				cb->l2rcb_zb = *zb;
5283 				cb->l2rcb_flags = zio_flags;
5284 
5285 				asize = vdev_psize_to_asize(vd, size);
5286 				if (asize != size) {
5287 					abd = abd_alloc_for_io(asize,
5288 					    HDR_ISTYPE_METADATA(hdr));
5289 					cb->l2rcb_abd = abd;
5290 				} else {
5291 					abd = hdr->b_l1hdr.b_pabd;
5292 				}
5293 
5294 				ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5295 				    addr + asize <= vd->vdev_psize -
5296 				    VDEV_LABEL_END_SIZE);
5297 
5298 				/*
5299 				 * l2arc read.  The SCL_L2ARC lock will be
5300 				 * released by l2arc_read_done().
5301 				 * Issue a null zio if the underlying buffer
5302 				 * was squashed to zero size by compression.
5303 				 */
5304 				ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
5305 				    ZIO_COMPRESS_EMPTY);
5306 				rzio = zio_read_phys(pio, vd, addr,
5307 				    asize, abd,
5308 				    ZIO_CHECKSUM_OFF,
5309 				    l2arc_read_done, cb, priority,
5310 				    zio_flags | ZIO_FLAG_DONT_CACHE |
5311 				    ZIO_FLAG_CANFAIL |
5312 				    ZIO_FLAG_DONT_PROPAGATE |
5313 				    ZIO_FLAG_DONT_RETRY, B_FALSE);
5314 				DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5315 				    zio_t *, rzio);
5316 				ARCSTAT_INCR(arcstat_l2_read_bytes, size);
5317 
5318 				if (*arc_flags & ARC_FLAG_NOWAIT) {
5319 					zio_nowait(rzio);
5320 					return (0);
5321 				}
5322 
5323 				ASSERT(*arc_flags & ARC_FLAG_WAIT);
5324 				if (zio_wait(rzio) == 0)
5325 					return (0);
5326 
5327 				/* l2arc read error; goto zio_read() */
5328 			} else {
5329 				DTRACE_PROBE1(l2arc__miss,
5330 				    arc_buf_hdr_t *, hdr);
5331 				ARCSTAT_BUMP(arcstat_l2_misses);
5332 				if (HDR_L2_WRITING(hdr))
5333 					ARCSTAT_BUMP(arcstat_l2_rw_clash);
5334 				spa_config_exit(spa, SCL_L2ARC, vd);
5335 			}
5336 		} else {
5337 			if (vd != NULL)
5338 				spa_config_exit(spa, SCL_L2ARC, vd);
5339 			if (l2arc_ndev != 0) {
5340 				DTRACE_PROBE1(l2arc__miss,
5341 				    arc_buf_hdr_t *, hdr);
5342 				ARCSTAT_BUMP(arcstat_l2_misses);
5343 			}
5344 		}
5345 
5346 		rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size,
5347 		    arc_read_done, hdr, priority, zio_flags, zb);
5348 
5349 		if (*arc_flags & ARC_FLAG_WAIT)
5350 			return (zio_wait(rzio));
5351 
5352 		ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5353 		zio_nowait(rzio);
5354 	}
5355 	return (0);
5356 }
5357 
5358 /*
5359  * Notify the arc that a block was freed, and thus will never be used again.
5360  */
5361 void
5362 arc_freed(spa_t *spa, const blkptr_t *bp)
5363 {
5364 	arc_buf_hdr_t *hdr;
5365 	kmutex_t *hash_lock;
5366 	uint64_t guid = spa_load_guid(spa);
5367 
5368 	ASSERT(!BP_IS_EMBEDDED(bp));
5369 
5370 	hdr = buf_hash_find(guid, bp, &hash_lock);
5371 	if (hdr == NULL)
5372 		return;
5373 
5374 	/*
5375 	 * We might be trying to free a block that is still doing I/O
5376 	 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5377 	 * dmu_sync-ed block). If this block is being prefetched, then it
5378 	 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5379 	 * until the I/O completes. A block may also have a reference if it is
5380 	 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5381 	 * have written the new block to its final resting place on disk but
5382 	 * without the dedup flag set. This would have left the hdr in the MRU
5383 	 * state and discoverable. When the txg finally syncs it detects that
5384 	 * the block was overridden in open context and issues an override I/O.
5385 	 * Since this is a dedup block, the override I/O will determine if the
5386 	 * block is already in the DDT. If so, then it will replace the io_bp
5387 	 * with the bp from the DDT and allow the I/O to finish. When the I/O
5388 	 * reaches the done callback, dbuf_write_override_done, it will
5389 	 * check to see if the io_bp and io_bp_override are identical.
5390 	 * If they are not, then it indicates that the bp was replaced with
5391 	 * the bp in the DDT and the override bp is freed. This allows
5392 	 * us to arrive here with a reference on a block that is being
5393 	 * freed. So if we have an I/O in progress, or a reference to
5394 	 * this hdr, then we don't destroy the hdr.
5395 	 */
5396 	if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
5397 	    zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
5398 		arc_change_state(arc_anon, hdr, hash_lock);
5399 		arc_hdr_destroy(hdr);
5400 		mutex_exit(hash_lock);
5401 	} else {
5402 		mutex_exit(hash_lock);
5403 	}
5404 
5405 }
5406 
5407 /*
5408  * Release this buffer from the cache, making it an anonymous buffer.  This
5409  * must be done after a read and prior to modifying the buffer contents.
5410  * If the buffer has more than one reference, we must make
5411  * a new hdr for the buffer.
5412  */
5413 void
5414 arc_release(arc_buf_t *buf, void *tag)
5415 {
5416 	arc_buf_hdr_t *hdr = buf->b_hdr;
5417 
5418 	/*
5419 	 * It would be nice to assert that if it's DMU metadata (level >
5420 	 * 0 || it's the dnode file), then it must be syncing context.
5421 	 * But we don't know that information at this level.
5422 	 */
5423 
5424 	mutex_enter(&buf->b_evict_lock);
5425 
5426 	ASSERT(HDR_HAS_L1HDR(hdr));
5427 
5428 	/*
5429 	 * We don't grab the hash lock prior to this check, because if
5430 	 * the buffer's header is in the arc_anon state, it won't be
5431 	 * linked into the hash table.
5432 	 */
5433 	if (hdr->b_l1hdr.b_state == arc_anon) {
5434 		mutex_exit(&buf->b_evict_lock);
5435 		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5436 		ASSERT(!HDR_IN_HASH_TABLE(hdr));
5437 		ASSERT(!HDR_HAS_L2HDR(hdr));
5438 		ASSERT(HDR_EMPTY(hdr));
5439 
5440 		ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5441 		ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
5442 		ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
5443 
5444 		hdr->b_l1hdr.b_arc_access = 0;
5445 
5446 		/*
5447 		 * If the buf is being overridden then it may already
5448 		 * have a hdr that is not empty.
5449 		 */
5450 		buf_discard_identity(hdr);
5451 		arc_buf_thaw(buf);
5452 
5453 		return;
5454 	}
5455 
5456 	kmutex_t *hash_lock = HDR_LOCK(hdr);
5457 	mutex_enter(hash_lock);
5458 
5459 	/*
5460 	 * This assignment is only valid as long as the hash_lock is
5461 	 * held, we must be careful not to reference state or the
5462 	 * b_state field after dropping the lock.
5463 	 */
5464 	arc_state_t *state = hdr->b_l1hdr.b_state;
5465 	ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5466 	ASSERT3P(state, !=, arc_anon);
5467 
5468 	/* this buffer is not on any list */
5469 	ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
5470 
5471 	if (HDR_HAS_L2HDR(hdr)) {
5472 		mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5473 
5474 		/*
5475 		 * We have to recheck this conditional again now that
5476 		 * we're holding the l2ad_mtx to prevent a race with
5477 		 * another thread which might be concurrently calling
5478 		 * l2arc_evict(). In that case, l2arc_evict() might have
5479 		 * destroyed the header's L2 portion as we were waiting
5480 		 * to acquire the l2ad_mtx.
5481 		 */
5482 		if (HDR_HAS_L2HDR(hdr))
5483 			arc_hdr_l2hdr_destroy(hdr);
5484 
5485 		mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5486 	}
5487 
5488 	/*
5489 	 * Do we have more than one buf?
5490 	 */
5491 	if (hdr->b_l1hdr.b_bufcnt > 1) {
5492 		arc_buf_hdr_t *nhdr;
5493 		uint64_t spa = hdr->b_spa;
5494 		uint64_t psize = HDR_GET_PSIZE(hdr);
5495 		uint64_t lsize = HDR_GET_LSIZE(hdr);
5496 		enum zio_compress compress = HDR_GET_COMPRESS(hdr);
5497 		arc_buf_contents_t type = arc_buf_type(hdr);
5498 		VERIFY3U(hdr->b_type, ==, type);
5499 
5500 		ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
5501 		(void) remove_reference(hdr, hash_lock, tag);
5502 
5503 		if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
5504 			ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5505 			ASSERT(ARC_BUF_LAST(buf));
5506 		}
5507 
5508 		/*
5509 		 * Pull the data off of this hdr and attach it to
5510 		 * a new anonymous hdr. Also find the last buffer
5511 		 * in the hdr's buffer list.
5512 		 */
5513 		arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
5514 		ASSERT3P(lastbuf, !=, NULL);
5515 
5516 		/*
5517 		 * If the current arc_buf_t and the hdr are sharing their data
5518 		 * buffer, then we must stop sharing that block.
5519 		 */
5520 		if (arc_buf_is_shared(buf)) {
5521 			VERIFY(!arc_buf_is_shared(lastbuf));
5522 
5523 			/*
5524 			 * First, sever the block sharing relationship between
5525 			 * buf and the arc_buf_hdr_t.
5526 			 */
5527 			arc_unshare_buf(hdr, buf);
5528 
5529 			/*
5530 			 * Now we need to recreate the hdr's b_pabd. Since we
5531 			 * have lastbuf handy, we try to share with it, but if
5532 			 * we can't then we allocate a new b_pabd and copy the
5533 			 * data from buf into it.
5534 			 */
5535 			if (arc_can_share(hdr, lastbuf)) {
5536 				arc_share_buf(hdr, lastbuf);
5537 			} else {
5538 				arc_hdr_alloc_pabd(hdr);
5539 				abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
5540 				    buf->b_data, psize);
5541 			}
5542 			VERIFY3P(lastbuf->b_data, !=, NULL);
5543 		} else if (HDR_SHARED_DATA(hdr)) {
5544 			/*
5545 			 * Uncompressed shared buffers are always at the end
5546 			 * of the list. Compressed buffers don't have the
5547 			 * same requirements. This makes it hard to
5548 			 * simply assert that the lastbuf is shared so
5549 			 * we rely on the hdr's compression flags to determine
5550 			 * if we have a compressed, shared buffer.
5551 			 */
5552 			ASSERT(arc_buf_is_shared(lastbuf) ||
5553 			    HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
5554 			ASSERT(!ARC_BUF_SHARED(buf));
5555 		}
5556 		ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5557 		ASSERT3P(state, !=, arc_l2c_only);
5558 
5559 		(void) zfs_refcount_remove_many(&state->arcs_size,
5560 		    arc_buf_size(buf), buf);
5561 
5562 		if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5563 			ASSERT3P(state, !=, arc_l2c_only);
5564 			(void) zfs_refcount_remove_many(
5565 			    &state->arcs_esize[type],
5566 			    arc_buf_size(buf), buf);
5567 		}
5568 
5569 		hdr->b_l1hdr.b_bufcnt -= 1;
5570 		arc_cksum_verify(buf);
5571 		arc_buf_unwatch(buf);
5572 
5573 		mutex_exit(hash_lock);
5574 
5575 		/*
5576 		 * Allocate a new hdr. The new hdr will contain a b_pabd
5577 		 * buffer which will be freed in arc_write().
5578 		 */
5579 		nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
5580 		ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
5581 		ASSERT0(nhdr->b_l1hdr.b_bufcnt);
5582 		ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt));
5583 		VERIFY3U(nhdr->b_type, ==, type);
5584 		ASSERT(!HDR_SHARED_DATA(nhdr));
5585 
5586 		nhdr->b_l1hdr.b_buf = buf;
5587 		nhdr->b_l1hdr.b_bufcnt = 1;
5588 		(void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
5589 		buf->b_hdr = nhdr;
5590 
5591 		mutex_exit(&buf->b_evict_lock);
5592 		(void) zfs_refcount_add_many(&arc_anon->arcs_size,
5593 		    arc_buf_size(buf), buf);
5594 	} else {
5595 		mutex_exit(&buf->b_evict_lock);
5596 		ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
5597 		/* protected by hash lock, or hdr is on arc_anon */
5598 		ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5599 		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5600 		arc_change_state(arc_anon, hdr, hash_lock);
5601 		hdr->b_l1hdr.b_arc_access = 0;
5602 		mutex_exit(hash_lock);
5603 
5604 		buf_discard_identity(hdr);
5605 		arc_buf_thaw(buf);
5606 	}
5607 }
5608 
5609 int
5610 arc_released(arc_buf_t *buf)
5611 {
5612 	int released;
5613 
5614 	mutex_enter(&buf->b_evict_lock);
5615 	released = (buf->b_data != NULL &&
5616 	    buf->b_hdr->b_l1hdr.b_state == arc_anon);
5617 	mutex_exit(&buf->b_evict_lock);
5618 	return (released);
5619 }
5620 
5621 #ifdef ZFS_DEBUG
5622 int
5623 arc_referenced(arc_buf_t *buf)
5624 {
5625 	int referenced;
5626 
5627 	mutex_enter(&buf->b_evict_lock);
5628 	referenced = (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
5629 	mutex_exit(&buf->b_evict_lock);
5630 	return (referenced);
5631 }
5632 #endif
5633 
5634 static void
5635 arc_write_ready(zio_t *zio)
5636 {
5637 	arc_write_callback_t *callback = zio->io_private;
5638 	arc_buf_t *buf = callback->awcb_buf;
5639 	arc_buf_hdr_t *hdr = buf->b_hdr;
5640 	uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
5641 
5642 	ASSERT(HDR_HAS_L1HDR(hdr));
5643 	ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
5644 	ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
5645 
5646 	/*
5647 	 * If we're reexecuting this zio because the pool suspended, then
5648 	 * cleanup any state that was previously set the first time the
5649 	 * callback was invoked.
5650 	 */
5651 	if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
5652 		arc_cksum_free(hdr);
5653 		arc_buf_unwatch(buf);
5654 		if (hdr->b_l1hdr.b_pabd != NULL) {
5655 			if (arc_buf_is_shared(buf)) {
5656 				arc_unshare_buf(hdr, buf);
5657 			} else {
5658 				arc_hdr_free_pabd(hdr);
5659 			}
5660 		}
5661 	}
5662 	ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5663 	ASSERT(!HDR_SHARED_DATA(hdr));
5664 	ASSERT(!arc_buf_is_shared(buf));
5665 
5666 	callback->awcb_ready(zio, buf, callback->awcb_private);
5667 
5668 	if (HDR_IO_IN_PROGRESS(hdr))
5669 		ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
5670 
5671 	arc_cksum_compute(buf);
5672 	arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5673 
5674 	enum zio_compress compress;
5675 	if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5676 		compress = ZIO_COMPRESS_OFF;
5677 	} else {
5678 		ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
5679 		compress = BP_GET_COMPRESS(zio->io_bp);
5680 	}
5681 	HDR_SET_PSIZE(hdr, psize);
5682 	arc_hdr_set_compress(hdr, compress);
5683 
5684 
5685 	/*
5686 	 * Fill the hdr with data. If the hdr is compressed, the data we want
5687 	 * is available from the zio, otherwise we can take it from the buf.
5688 	 *
5689 	 * We might be able to share the buf's data with the hdr here. However,
5690 	 * doing so would cause the ARC to be full of linear ABDs if we write a
5691 	 * lot of shareable data. As a compromise, we check whether scattered
5692 	 * ABDs are allowed, and assume that if they are then the user wants
5693 	 * the ARC to be primarily filled with them regardless of the data being
5694 	 * written. Therefore, if they're allowed then we allocate one and copy
5695 	 * the data into it; otherwise, we share the data directly if we can.
5696 	 */
5697 	if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
5698 		arc_hdr_alloc_pabd(hdr);
5699 
5700 		/*
5701 		 * Ideally, we would always copy the io_abd into b_pabd, but the
5702 		 * user may have disabled compressed ARC, thus we must check the
5703 		 * hdr's compression setting rather than the io_bp's.
5704 		 */
5705 		if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5706 			ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=,
5707 			    ZIO_COMPRESS_OFF);
5708 			ASSERT3U(psize, >, 0);
5709 
5710 			abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
5711 		} else {
5712 			ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
5713 
5714 			abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
5715 			    arc_buf_size(buf));
5716 		}
5717 	} else {
5718 		ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
5719 		ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
5720 		ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5721 
5722 		arc_share_buf(hdr, buf);
5723 	}
5724 
5725 	arc_hdr_verify(hdr, zio->io_bp);
5726 }
5727 
5728 static void
5729 arc_write_children_ready(zio_t *zio)
5730 {
5731 	arc_write_callback_t *callback = zio->io_private;
5732 	arc_buf_t *buf = callback->awcb_buf;
5733 
5734 	callback->awcb_children_ready(zio, buf, callback->awcb_private);
5735 }
5736 
5737 /*
5738  * The SPA calls this callback for each physical write that happens on behalf
5739  * of a logical write.  See the comment in dbuf_write_physdone() for details.
5740  */
5741 static void
5742 arc_write_physdone(zio_t *zio)
5743 {
5744 	arc_write_callback_t *cb = zio->io_private;
5745 	if (cb->awcb_physdone != NULL)
5746 		cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
5747 }
5748 
5749 static void
5750 arc_write_done(zio_t *zio)
5751 {
5752 	arc_write_callback_t *callback = zio->io_private;
5753 	arc_buf_t *buf = callback->awcb_buf;
5754 	arc_buf_hdr_t *hdr = buf->b_hdr;
5755 
5756 	ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5757 
5758 	if (zio->io_error == 0) {
5759 		arc_hdr_verify(hdr, zio->io_bp);
5760 
5761 		if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5762 			buf_discard_identity(hdr);
5763 		} else {
5764 			hdr->b_dva = *BP_IDENTITY(zio->io_bp);
5765 			hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
5766 		}
5767 	} else {
5768 		ASSERT(HDR_EMPTY(hdr));
5769 	}
5770 
5771 	/*
5772 	 * If the block to be written was all-zero or compressed enough to be
5773 	 * embedded in the BP, no write was performed so there will be no
5774 	 * dva/birth/checksum.  The buffer must therefore remain anonymous
5775 	 * (and uncached).
5776 	 */
5777 	if (!HDR_EMPTY(hdr)) {
5778 		arc_buf_hdr_t *exists;
5779 		kmutex_t *hash_lock;
5780 
5781 		ASSERT3U(zio->io_error, ==, 0);
5782 
5783 		arc_cksum_verify(buf);
5784 
5785 		exists = buf_hash_insert(hdr, &hash_lock);
5786 		if (exists != NULL) {
5787 			/*
5788 			 * This can only happen if we overwrite for
5789 			 * sync-to-convergence, because we remove
5790 			 * buffers from the hash table when we arc_free().
5791 			 */
5792 			if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
5793 				if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5794 					panic("bad overwrite, hdr=%p exists=%p",
5795 					    (void *)hdr, (void *)exists);
5796 				ASSERT(zfs_refcount_is_zero(
5797 				    &exists->b_l1hdr.b_refcnt));
5798 				arc_change_state(arc_anon, exists, hash_lock);
5799 				mutex_exit(hash_lock);
5800 				arc_hdr_destroy(exists);
5801 				exists = buf_hash_insert(hdr, &hash_lock);
5802 				ASSERT3P(exists, ==, NULL);
5803 			} else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
5804 				/* nopwrite */
5805 				ASSERT(zio->io_prop.zp_nopwrite);
5806 				if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5807 					panic("bad nopwrite, hdr=%p exists=%p",
5808 					    (void *)hdr, (void *)exists);
5809 			} else {
5810 				/* Dedup */
5811 				ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
5812 				ASSERT(hdr->b_l1hdr.b_state == arc_anon);
5813 				ASSERT(BP_GET_DEDUP(zio->io_bp));
5814 				ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
5815 			}
5816 		}
5817 		arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5818 		/* if it's not anon, we are doing a scrub */
5819 		if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
5820 			arc_access(hdr, hash_lock);
5821 		mutex_exit(hash_lock);
5822 	} else {
5823 		arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5824 	}
5825 
5826 	ASSERT(!zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5827 	callback->awcb_done(zio, buf, callback->awcb_private);
5828 
5829 	abd_put(zio->io_abd);
5830 	kmem_free(callback, sizeof (arc_write_callback_t));
5831 }
5832 
5833 zio_t *
5834 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
5835     boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready,
5836     arc_done_func_t *children_ready, arc_done_func_t *physdone,
5837     arc_done_func_t *done, void *private, zio_priority_t priority,
5838     int zio_flags, const zbookmark_phys_t *zb)
5839 {
5840 	arc_buf_hdr_t *hdr = buf->b_hdr;
5841 	arc_write_callback_t *callback;
5842 	zio_t *zio;
5843 	zio_prop_t localprop = *zp;
5844 
5845 	ASSERT3P(ready, !=, NULL);
5846 	ASSERT3P(done, !=, NULL);
5847 	ASSERT(!HDR_IO_ERROR(hdr));
5848 	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5849 	ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5850 	ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
5851 	if (l2arc)
5852 		arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5853 	if (ARC_BUF_COMPRESSED(buf)) {
5854 		/*
5855 		 * We're writing a pre-compressed buffer.  Make the
5856 		 * compression algorithm requested by the zio_prop_t match
5857 		 * the pre-compressed buffer's compression algorithm.
5858 		 */
5859 		localprop.zp_compress = HDR_GET_COMPRESS(hdr);
5860 
5861 		ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
5862 		zio_flags |= ZIO_FLAG_RAW;
5863 	}
5864 	callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
5865 	callback->awcb_ready = ready;
5866 	callback->awcb_children_ready = children_ready;
5867 	callback->awcb_physdone = physdone;
5868 	callback->awcb_done = done;
5869 	callback->awcb_private = private;
5870 	callback->awcb_buf = buf;
5871 
5872 	/*
5873 	 * The hdr's b_pabd is now stale, free it now. A new data block
5874 	 * will be allocated when the zio pipeline calls arc_write_ready().
5875 	 */
5876 	if (hdr->b_l1hdr.b_pabd != NULL) {
5877 		/*
5878 		 * If the buf is currently sharing the data block with
5879 		 * the hdr then we need to break that relationship here.
5880 		 * The hdr will remain with a NULL data pointer and the
5881 		 * buf will take sole ownership of the block.
5882 		 */
5883 		if (arc_buf_is_shared(buf)) {
5884 			arc_unshare_buf(hdr, buf);
5885 		} else {
5886 			arc_hdr_free_pabd(hdr);
5887 		}
5888 		VERIFY3P(buf->b_data, !=, NULL);
5889 		arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
5890 	}
5891 	ASSERT(!arc_buf_is_shared(buf));
5892 	ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5893 
5894 	zio = zio_write(pio, spa, txg, bp,
5895 	    abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
5896 	    HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
5897 	    (children_ready != NULL) ? arc_write_children_ready : NULL,
5898 	    arc_write_physdone, arc_write_done, callback,
5899 	    priority, zio_flags, zb);
5900 
5901 	return (zio);
5902 }
5903 
5904 static int
5905 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
5906 {
5907 #ifdef _KERNEL
5908 	uint64_t available_memory = ptob(freemem);
5909 
5910 #if defined(__i386)
5911 	available_memory =
5912 	    MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
5913 #endif
5914 
5915 	if (freemem > physmem * arc_lotsfree_percent / 100)
5916 		return (0);
5917 
5918 	if (txg > spa->spa_lowmem_last_txg) {
5919 		spa->spa_lowmem_last_txg = txg;
5920 		spa->spa_lowmem_page_load = 0;
5921 	}
5922 	/*
5923 	 * If we are in pageout, we know that memory is already tight,
5924 	 * the arc is already going to be evicting, so we just want to
5925 	 * continue to let page writes occur as quickly as possible.
5926 	 */
5927 	if (curproc == proc_pageout) {
5928 		if (spa->spa_lowmem_page_load >
5929 		    MAX(ptob(minfree), available_memory) / 4)
5930 			return (SET_ERROR(ERESTART));
5931 		/* Note: reserve is inflated, so we deflate */
5932 		atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8);
5933 		return (0);
5934 	} else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) {
5935 		/* memory is low, delay before restarting */
5936 		ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
5937 		return (SET_ERROR(EAGAIN));
5938 	}
5939 	spa->spa_lowmem_page_load = 0;
5940 #endif /* _KERNEL */
5941 	return (0);
5942 }
5943 
5944 void
5945 arc_tempreserve_clear(uint64_t reserve)
5946 {
5947 	atomic_add_64(&arc_tempreserve, -reserve);
5948 	ASSERT((int64_t)arc_tempreserve >= 0);
5949 }
5950 
5951 int
5952 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
5953 {
5954 	int error;
5955 	uint64_t anon_size;
5956 
5957 	if (reserve > arc_c/4 && !arc_no_grow)
5958 		arc_c = MIN(arc_c_max, reserve * 4);
5959 	if (reserve > arc_c)
5960 		return (SET_ERROR(ENOMEM));
5961 
5962 	/*
5963 	 * Don't count loaned bufs as in flight dirty data to prevent long
5964 	 * network delays from blocking transactions that are ready to be
5965 	 * assigned to a txg.
5966 	 */
5967 
5968 	/* assert that it has not wrapped around */
5969 	ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
5970 
5971 	anon_size = MAX((int64_t)(zfs_refcount_count(&arc_anon->arcs_size) -
5972 	    arc_loaned_bytes), 0);
5973 
5974 	/*
5975 	 * Writes will, almost always, require additional memory allocations
5976 	 * in order to compress/encrypt/etc the data.  We therefore need to
5977 	 * make sure that there is sufficient available memory for this.
5978 	 */
5979 	error = arc_memory_throttle(spa, reserve, txg);
5980 	if (error != 0)
5981 		return (error);
5982 
5983 	/*
5984 	 * Throttle writes when the amount of dirty data in the cache
5985 	 * gets too large.  We try to keep the cache less than half full
5986 	 * of dirty blocks so that our sync times don't grow too large.
5987 	 *
5988 	 * In the case of one pool being built on another pool, we want
5989 	 * to make sure we don't end up throttling the lower (backing)
5990 	 * pool when the upper pool is the majority contributor to dirty
5991 	 * data. To insure we make forward progress during throttling, we
5992 	 * also check the current pool's net dirty data and only throttle
5993 	 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
5994 	 * data in the cache.
5995 	 *
5996 	 * Note: if two requests come in concurrently, we might let them
5997 	 * both succeed, when one of them should fail.  Not a huge deal.
5998 	 */
5999 	uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
6000 	uint64_t spa_dirty_anon = spa_dirty_data(spa);
6001 
6002 	if (total_dirty > arc_c * zfs_arc_dirty_limit_percent / 100 &&
6003 	    anon_size > arc_c * zfs_arc_anon_limit_percent / 100 &&
6004 	    spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
6005 		uint64_t meta_esize =
6006 		    zfs_refcount_count(
6007 		    &arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6008 		uint64_t data_esize =
6009 		    zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6010 		dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6011 		    "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6012 		    arc_tempreserve >> 10, meta_esize >> 10,
6013 		    data_esize >> 10, reserve >> 10, arc_c >> 10);
6014 		return (SET_ERROR(ERESTART));
6015 	}
6016 	atomic_add_64(&arc_tempreserve, reserve);
6017 	return (0);
6018 }
6019 
6020 static void
6021 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6022     kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6023 {
6024 	size->value.ui64 = zfs_refcount_count(&state->arcs_size);
6025 	evict_data->value.ui64 =
6026 	    zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6027 	evict_metadata->value.ui64 =
6028 	    zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6029 }
6030 
6031 static int
6032 arc_kstat_update(kstat_t *ksp, int rw)
6033 {
6034 	arc_stats_t *as = ksp->ks_data;
6035 
6036 	if (rw == KSTAT_WRITE) {
6037 		return (EACCES);
6038 	} else {
6039 		arc_kstat_update_state(arc_anon,
6040 		    &as->arcstat_anon_size,
6041 		    &as->arcstat_anon_evictable_data,
6042 		    &as->arcstat_anon_evictable_metadata);
6043 		arc_kstat_update_state(arc_mru,
6044 		    &as->arcstat_mru_size,
6045 		    &as->arcstat_mru_evictable_data,
6046 		    &as->arcstat_mru_evictable_metadata);
6047 		arc_kstat_update_state(arc_mru_ghost,
6048 		    &as->arcstat_mru_ghost_size,
6049 		    &as->arcstat_mru_ghost_evictable_data,
6050 		    &as->arcstat_mru_ghost_evictable_metadata);
6051 		arc_kstat_update_state(arc_mfu,
6052 		    &as->arcstat_mfu_size,
6053 		    &as->arcstat_mfu_evictable_data,
6054 		    &as->arcstat_mfu_evictable_metadata);
6055 		arc_kstat_update_state(arc_mfu_ghost,
6056 		    &as->arcstat_mfu_ghost_size,
6057 		    &as->arcstat_mfu_ghost_evictable_data,
6058 		    &as->arcstat_mfu_ghost_evictable_metadata);
6059 
6060 		ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
6061 		ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
6062 		ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
6063 		ARCSTAT(arcstat_metadata_size) =
6064 		    aggsum_value(&astat_metadata_size);
6065 		ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
6066 		ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size);
6067 		ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
6068 	}
6069 
6070 	return (0);
6071 }
6072 
6073 /*
6074  * This function *must* return indices evenly distributed between all
6075  * sublists of the multilist. This is needed due to how the ARC eviction
6076  * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6077  * distributed between all sublists and uses this assumption when
6078  * deciding which sublist to evict from and how much to evict from it.
6079  */
6080 unsigned int
6081 arc_state_multilist_index_func(multilist_t *ml, void *obj)
6082 {
6083 	arc_buf_hdr_t *hdr = obj;
6084 
6085 	/*
6086 	 * We rely on b_dva to generate evenly distributed index
6087 	 * numbers using buf_hash below. So, as an added precaution,
6088 	 * let's make sure we never add empty buffers to the arc lists.
6089 	 */
6090 	ASSERT(!HDR_EMPTY(hdr));
6091 
6092 	/*
6093 	 * The assumption here, is the hash value for a given
6094 	 * arc_buf_hdr_t will remain constant throughout it's lifetime
6095 	 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
6096 	 * Thus, we don't need to store the header's sublist index
6097 	 * on insertion, as this index can be recalculated on removal.
6098 	 *
6099 	 * Also, the low order bits of the hash value are thought to be
6100 	 * distributed evenly. Otherwise, in the case that the multilist
6101 	 * has a power of two number of sublists, each sublists' usage
6102 	 * would not be evenly distributed.
6103 	 */
6104 	return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
6105 	    multilist_get_num_sublists(ml));
6106 }
6107 
6108 static void
6109 arc_state_init(void)
6110 {
6111 	arc_anon = &ARC_anon;
6112 	arc_mru = &ARC_mru;
6113 	arc_mru_ghost = &ARC_mru_ghost;
6114 	arc_mfu = &ARC_mfu;
6115 	arc_mfu_ghost = &ARC_mfu_ghost;
6116 	arc_l2c_only = &ARC_l2c_only;
6117 
6118 	arc_mru->arcs_list[ARC_BUFC_METADATA] =
6119 	    multilist_create(sizeof (arc_buf_hdr_t),
6120 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6121 	    arc_state_multilist_index_func);
6122 	arc_mru->arcs_list[ARC_BUFC_DATA] =
6123 	    multilist_create(sizeof (arc_buf_hdr_t),
6124 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6125 	    arc_state_multilist_index_func);
6126 	arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
6127 	    multilist_create(sizeof (arc_buf_hdr_t),
6128 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6129 	    arc_state_multilist_index_func);
6130 	arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
6131 	    multilist_create(sizeof (arc_buf_hdr_t),
6132 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6133 	    arc_state_multilist_index_func);
6134 	arc_mfu->arcs_list[ARC_BUFC_METADATA] =
6135 	    multilist_create(sizeof (arc_buf_hdr_t),
6136 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6137 	    arc_state_multilist_index_func);
6138 	arc_mfu->arcs_list[ARC_BUFC_DATA] =
6139 	    multilist_create(sizeof (arc_buf_hdr_t),
6140 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6141 	    arc_state_multilist_index_func);
6142 	arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
6143 	    multilist_create(sizeof (arc_buf_hdr_t),
6144 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6145 	    arc_state_multilist_index_func);
6146 	arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
6147 	    multilist_create(sizeof (arc_buf_hdr_t),
6148 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6149 	    arc_state_multilist_index_func);
6150 	arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
6151 	    multilist_create(sizeof (arc_buf_hdr_t),
6152 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6153 	    arc_state_multilist_index_func);
6154 	arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
6155 	    multilist_create(sizeof (arc_buf_hdr_t),
6156 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6157 	    arc_state_multilist_index_func);
6158 
6159 	zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6160 	zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6161 	zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6162 	zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6163 	zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6164 	zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6165 	zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6166 	zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6167 	zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6168 	zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6169 	zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6170 	zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6171 
6172 	zfs_refcount_create(&arc_anon->arcs_size);
6173 	zfs_refcount_create(&arc_mru->arcs_size);
6174 	zfs_refcount_create(&arc_mru_ghost->arcs_size);
6175 	zfs_refcount_create(&arc_mfu->arcs_size);
6176 	zfs_refcount_create(&arc_mfu_ghost->arcs_size);
6177 	zfs_refcount_create(&arc_l2c_only->arcs_size);
6178 
6179 	aggsum_init(&arc_meta_used, 0);
6180 	aggsum_init(&arc_size, 0);
6181 	aggsum_init(&astat_data_size, 0);
6182 	aggsum_init(&astat_metadata_size, 0);
6183 	aggsum_init(&astat_hdr_size, 0);
6184 	aggsum_init(&astat_other_size, 0);
6185 	aggsum_init(&astat_l2_hdr_size, 0);
6186 }
6187 
6188 static void
6189 arc_state_fini(void)
6190 {
6191 	zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6192 	zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6193 	zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6194 	zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6195 	zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6196 	zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6197 	zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6198 	zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6199 	zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6200 	zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6201 	zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6202 	zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6203 
6204 	zfs_refcount_destroy(&arc_anon->arcs_size);
6205 	zfs_refcount_destroy(&arc_mru->arcs_size);
6206 	zfs_refcount_destroy(&arc_mru_ghost->arcs_size);
6207 	zfs_refcount_destroy(&arc_mfu->arcs_size);
6208 	zfs_refcount_destroy(&arc_mfu_ghost->arcs_size);
6209 	zfs_refcount_destroy(&arc_l2c_only->arcs_size);
6210 
6211 	multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
6212 	multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
6213 	multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
6214 	multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
6215 	multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
6216 	multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
6217 	multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
6218 	multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
6219 
6220 	aggsum_fini(&arc_meta_used);
6221 	aggsum_fini(&arc_size);
6222 	aggsum_fini(&astat_data_size);
6223 	aggsum_fini(&astat_metadata_size);
6224 	aggsum_fini(&astat_hdr_size);
6225 	aggsum_fini(&astat_other_size);
6226 	aggsum_fini(&astat_l2_hdr_size);
6227 }
6228 
6229 uint64_t
6230 arc_max_bytes(void)
6231 {
6232 	return (arc_c_max);
6233 }
6234 
6235 void
6236 arc_init(void)
6237 {
6238 	/*
6239 	 * allmem is "all memory that we could possibly use".
6240 	 */
6241 #ifdef _KERNEL
6242 	uint64_t allmem = ptob(physmem - swapfs_minfree);
6243 #else
6244 	uint64_t allmem = (physmem * PAGESIZE) / 2;
6245 #endif
6246 	mutex_init(&arc_adjust_lock, NULL, MUTEX_DEFAULT, NULL);
6247 	cv_init(&arc_adjust_waiters_cv, NULL, CV_DEFAULT, NULL);
6248 
6249 	/* Convert seconds to clock ticks */
6250 	arc_min_prefetch_lifespan = 1 * hz;
6251 
6252 	/* set min cache to 1/32 of all memory, or 64MB, whichever is more */
6253 	arc_c_min = MAX(allmem / 32, 64 << 20);
6254 	/* set max to 3/4 of all memory, or all but 1GB, whichever is more */
6255 	if (allmem >= 1 << 30)
6256 		arc_c_max = allmem - (1 << 30);
6257 	else
6258 		arc_c_max = arc_c_min;
6259 	arc_c_max = MAX(allmem * 3 / 4, arc_c_max);
6260 
6261 	/*
6262 	 * In userland, there's only the memory pressure that we artificially
6263 	 * create (see arc_available_memory()).  Don't let arc_c get too
6264 	 * small, because it can cause transactions to be larger than
6265 	 * arc_c, causing arc_tempreserve_space() to fail.
6266 	 */
6267 #ifndef _KERNEL
6268 	arc_c_min = arc_c_max / 2;
6269 #endif
6270 
6271 	/*
6272 	 * Allow the tunables to override our calculations if they are
6273 	 * reasonable (ie. over 64MB)
6274 	 */
6275 	if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) {
6276 		arc_c_max = zfs_arc_max;
6277 		arc_c_min = MIN(arc_c_min, arc_c_max);
6278 	}
6279 	if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max)
6280 		arc_c_min = zfs_arc_min;
6281 
6282 	arc_c = arc_c_max;
6283 	arc_p = (arc_c >> 1);
6284 
6285 	/* limit meta-data to 1/4 of the arc capacity */
6286 	arc_meta_limit = arc_c_max / 4;
6287 
6288 #ifdef _KERNEL
6289 	/*
6290 	 * Metadata is stored in the kernel's heap.  Don't let us
6291 	 * use more than half the heap for the ARC.
6292 	 */
6293 	arc_meta_limit = MIN(arc_meta_limit,
6294 	    vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
6295 #endif
6296 
6297 	/* Allow the tunable to override if it is reasonable */
6298 	if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
6299 		arc_meta_limit = zfs_arc_meta_limit;
6300 
6301 	if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
6302 		arc_c_min = arc_meta_limit / 2;
6303 
6304 	if (zfs_arc_meta_min > 0) {
6305 		arc_meta_min = zfs_arc_meta_min;
6306 	} else {
6307 		arc_meta_min = arc_c_min / 2;
6308 	}
6309 
6310 	if (zfs_arc_grow_retry > 0)
6311 		arc_grow_retry = zfs_arc_grow_retry;
6312 
6313 	if (zfs_arc_shrink_shift > 0)
6314 		arc_shrink_shift = zfs_arc_shrink_shift;
6315 
6316 	/*
6317 	 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6318 	 */
6319 	if (arc_no_grow_shift >= arc_shrink_shift)
6320 		arc_no_grow_shift = arc_shrink_shift - 1;
6321 
6322 	if (zfs_arc_p_min_shift > 0)
6323 		arc_p_min_shift = zfs_arc_p_min_shift;
6324 
6325 	/* if kmem_flags are set, lets try to use less memory */
6326 	if (kmem_debugging())
6327 		arc_c = arc_c / 2;
6328 	if (arc_c < arc_c_min)
6329 		arc_c = arc_c_min;
6330 
6331 	arc_state_init();
6332 
6333 	/*
6334 	 * The arc must be "uninitialized", so that hdr_recl() (which is
6335 	 * registered by buf_init()) will not access arc_reap_zthr before
6336 	 * it is created.
6337 	 */
6338 	ASSERT(!arc_initialized);
6339 	buf_init();
6340 
6341 	arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
6342 	    sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
6343 
6344 	if (arc_ksp != NULL) {
6345 		arc_ksp->ks_data = &arc_stats;
6346 		arc_ksp->ks_update = arc_kstat_update;
6347 		kstat_install(arc_ksp);
6348 	}
6349 
6350 	arc_adjust_zthr = zthr_create(arc_adjust_cb_check,
6351 	    arc_adjust_cb, NULL);
6352 	arc_reap_zthr = zthr_create_timer(arc_reap_cb_check,
6353 	    arc_reap_cb, NULL, SEC2NSEC(1));
6354 
6355 	arc_initialized = B_TRUE;
6356 	arc_warm = B_FALSE;
6357 
6358 	/*
6359 	 * Calculate maximum amount of dirty data per pool.
6360 	 *
6361 	 * If it has been set by /etc/system, take that.
6362 	 * Otherwise, use a percentage of physical memory defined by
6363 	 * zfs_dirty_data_max_percent (default 10%) with a cap at
6364 	 * zfs_dirty_data_max_max (default 4GB).
6365 	 */
6366 	if (zfs_dirty_data_max == 0) {
6367 		zfs_dirty_data_max = physmem * PAGESIZE *
6368 		    zfs_dirty_data_max_percent / 100;
6369 		zfs_dirty_data_max = MIN(zfs_dirty_data_max,
6370 		    zfs_dirty_data_max_max);
6371 	}
6372 }
6373 
6374 void
6375 arc_fini(void)
6376 {
6377 	/* Use B_TRUE to ensure *all* buffers are evicted */
6378 	arc_flush(NULL, B_TRUE);
6379 
6380 	arc_initialized = B_FALSE;
6381 
6382 	if (arc_ksp != NULL) {
6383 		kstat_delete(arc_ksp);
6384 		arc_ksp = NULL;
6385 	}
6386 
6387 	(void) zthr_cancel(arc_adjust_zthr);
6388 	zthr_destroy(arc_adjust_zthr);
6389 
6390 	(void) zthr_cancel(arc_reap_zthr);
6391 	zthr_destroy(arc_reap_zthr);
6392 
6393 	mutex_destroy(&arc_adjust_lock);
6394 	cv_destroy(&arc_adjust_waiters_cv);
6395 
6396 	/*
6397 	 * buf_fini() must proceed arc_state_fini() because buf_fin() may
6398 	 * trigger the release of kmem magazines, which can callback to
6399 	 * arc_space_return() which accesses aggsums freed in act_state_fini().
6400 	 */
6401 	buf_fini();
6402 	arc_state_fini();
6403 
6404 	ASSERT0(arc_loaned_bytes);
6405 }
6406 
6407 /*
6408  * Level 2 ARC
6409  *
6410  * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6411  * It uses dedicated storage devices to hold cached data, which are populated
6412  * using large infrequent writes.  The main role of this cache is to boost
6413  * the performance of random read workloads.  The intended L2ARC devices
6414  * include short-stroked disks, solid state disks, and other media with
6415  * substantially faster read latency than disk.
6416  *
6417  *                 +-----------------------+
6418  *                 |         ARC           |
6419  *                 +-----------------------+
6420  *                    |         ^     ^
6421  *                    |         |     |
6422  *      l2arc_feed_thread()    arc_read()
6423  *                    |         |     |
6424  *                    |  l2arc read   |
6425  *                    V         |     |
6426  *               +---------------+    |
6427  *               |     L2ARC     |    |
6428  *               +---------------+    |
6429  *                   |    ^           |
6430  *          l2arc_write() |           |
6431  *                   |    |           |
6432  *                   V    |           |
6433  *                 +-------+      +-------+
6434  *                 | vdev  |      | vdev  |
6435  *                 | cache |      | cache |
6436  *                 +-------+      +-------+
6437  *                 +=========+     .-----.
6438  *                 :  L2ARC  :    |-_____-|
6439  *                 : devices :    | Disks |
6440  *                 +=========+    `-_____-'
6441  *
6442  * Read requests are satisfied from the following sources, in order:
6443  *
6444  *	1) ARC
6445  *	2) vdev cache of L2ARC devices
6446  *	3) L2ARC devices
6447  *	4) vdev cache of disks
6448  *	5) disks
6449  *
6450  * Some L2ARC device types exhibit extremely slow write performance.
6451  * To accommodate for this there are some significant differences between
6452  * the L2ARC and traditional cache design:
6453  *
6454  * 1. There is no eviction path from the ARC to the L2ARC.  Evictions from
6455  * the ARC behave as usual, freeing buffers and placing headers on ghost
6456  * lists.  The ARC does not send buffers to the L2ARC during eviction as
6457  * this would add inflated write latencies for all ARC memory pressure.
6458  *
6459  * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6460  * It does this by periodically scanning buffers from the eviction-end of
6461  * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6462  * not already there. It scans until a headroom of buffers is satisfied,
6463  * which itself is a buffer for ARC eviction. If a compressible buffer is
6464  * found during scanning and selected for writing to an L2ARC device, we
6465  * temporarily boost scanning headroom during the next scan cycle to make
6466  * sure we adapt to compression effects (which might significantly reduce
6467  * the data volume we write to L2ARC). The thread that does this is
6468  * l2arc_feed_thread(), illustrated below; example sizes are included to
6469  * provide a better sense of ratio than this diagram:
6470  *
6471  *	       head -->                        tail
6472  *	        +---------------------+----------+
6473  *	ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->.   # already on L2ARC
6474  *	        +---------------------+----------+   |   o L2ARC eligible
6475  *	ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->|   : ARC buffer
6476  *	        +---------------------+----------+   |
6477  *	             15.9 Gbytes      ^ 32 Mbytes    |
6478  *	                           headroom          |
6479  *	                                      l2arc_feed_thread()
6480  *	                                             |
6481  *	                 l2arc write hand <--[oooo]--'
6482  *	                         |           8 Mbyte
6483  *	                         |          write max
6484  *	                         V
6485  *		  +==============================+
6486  *	L2ARC dev |####|#|###|###|    |####| ... |
6487  *	          +==============================+
6488  *	                     32 Gbytes
6489  *
6490  * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6491  * evicted, then the L2ARC has cached a buffer much sooner than it probably
6492  * needed to, potentially wasting L2ARC device bandwidth and storage.  It is
6493  * safe to say that this is an uncommon case, since buffers at the end of
6494  * the ARC lists have moved there due to inactivity.
6495  *
6496  * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6497  * then the L2ARC simply misses copying some buffers.  This serves as a
6498  * pressure valve to prevent heavy read workloads from both stalling the ARC
6499  * with waits and clogging the L2ARC with writes.  This also helps prevent
6500  * the potential for the L2ARC to churn if it attempts to cache content too
6501  * quickly, such as during backups of the entire pool.
6502  *
6503  * 5. After system boot and before the ARC has filled main memory, there are
6504  * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6505  * lists can remain mostly static.  Instead of searching from tail of these
6506  * lists as pictured, the l2arc_feed_thread() will search from the list heads
6507  * for eligible buffers, greatly increasing its chance of finding them.
6508  *
6509  * The L2ARC device write speed is also boosted during this time so that
6510  * the L2ARC warms up faster.  Since there have been no ARC evictions yet,
6511  * there are no L2ARC reads, and no fear of degrading read performance
6512  * through increased writes.
6513  *
6514  * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6515  * the vdev queue can aggregate them into larger and fewer writes.  Each
6516  * device is written to in a rotor fashion, sweeping writes through
6517  * available space then repeating.
6518  *
6519  * 7. The L2ARC does not store dirty content.  It never needs to flush
6520  * write buffers back to disk based storage.
6521  *
6522  * 8. If an ARC buffer is written (and dirtied) which also exists in the
6523  * L2ARC, the now stale L2ARC buffer is immediately dropped.
6524  *
6525  * The performance of the L2ARC can be tweaked by a number of tunables, which
6526  * may be necessary for different workloads:
6527  *
6528  *	l2arc_write_max		max write bytes per interval
6529  *	l2arc_write_boost	extra write bytes during device warmup
6530  *	l2arc_noprefetch	skip caching prefetched buffers
6531  *	l2arc_headroom		number of max device writes to precache
6532  *	l2arc_headroom_boost	when we find compressed buffers during ARC
6533  *				scanning, we multiply headroom by this
6534  *				percentage factor for the next scan cycle,
6535  *				since more compressed buffers are likely to
6536  *				be present
6537  *	l2arc_feed_secs		seconds between L2ARC writing
6538  *
6539  * Tunables may be removed or added as future performance improvements are
6540  * integrated, and also may become zpool properties.
6541  *
6542  * There are three key functions that control how the L2ARC warms up:
6543  *
6544  *	l2arc_write_eligible()	check if a buffer is eligible to cache
6545  *	l2arc_write_size()	calculate how much to write
6546  *	l2arc_write_interval()	calculate sleep delay between writes
6547  *
6548  * These three functions determine what to write, how much, and how quickly
6549  * to send writes.
6550  */
6551 
6552 static boolean_t
6553 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
6554 {
6555 	/*
6556 	 * A buffer is *not* eligible for the L2ARC if it:
6557 	 * 1. belongs to a different spa.
6558 	 * 2. is already cached on the L2ARC.
6559 	 * 3. has an I/O in progress (it may be an incomplete read).
6560 	 * 4. is flagged not eligible (zfs property).
6561 	 */
6562 	if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
6563 	    HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
6564 		return (B_FALSE);
6565 
6566 	return (B_TRUE);
6567 }
6568 
6569 static uint64_t
6570 l2arc_write_size(void)
6571 {
6572 	uint64_t size;
6573 
6574 	/*
6575 	 * Make sure our globals have meaningful values in case the user
6576 	 * altered them.
6577 	 */
6578 	size = l2arc_write_max;
6579 	if (size == 0) {
6580 		cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
6581 		    "be greater than zero, resetting it to the default (%d)",
6582 		    L2ARC_WRITE_SIZE);
6583 		size = l2arc_write_max = L2ARC_WRITE_SIZE;
6584 	}
6585 
6586 	if (arc_warm == B_FALSE)
6587 		size += l2arc_write_boost;
6588 
6589 	return (size);
6590 
6591 }
6592 
6593 static clock_t
6594 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
6595 {
6596 	clock_t interval, next, now;
6597 
6598 	/*
6599 	 * If the ARC lists are busy, increase our write rate; if the
6600 	 * lists are stale, idle back.  This is achieved by checking
6601 	 * how much we previously wrote - if it was more than half of
6602 	 * what we wanted, schedule the next write much sooner.
6603 	 */
6604 	if (l2arc_feed_again && wrote > (wanted / 2))
6605 		interval = (hz * l2arc_feed_min_ms) / 1000;
6606 	else
6607 		interval = hz * l2arc_feed_secs;
6608 
6609 	now = ddi_get_lbolt();
6610 	next = MAX(now, MIN(now + interval, began + interval));
6611 
6612 	return (next);
6613 }
6614 
6615 /*
6616  * Cycle through L2ARC devices.  This is how L2ARC load balances.
6617  * If a device is returned, this also returns holding the spa config lock.
6618  */
6619 static l2arc_dev_t *
6620 l2arc_dev_get_next(void)
6621 {
6622 	l2arc_dev_t *first, *next = NULL;
6623 
6624 	/*
6625 	 * Lock out the removal of spas (spa_namespace_lock), then removal
6626 	 * of cache devices (l2arc_dev_mtx).  Once a device has been selected,
6627 	 * both locks will be dropped and a spa config lock held instead.
6628 	 */
6629 	mutex_enter(&spa_namespace_lock);
6630 	mutex_enter(&l2arc_dev_mtx);
6631 
6632 	/* if there are no vdevs, there is nothing to do */
6633 	if (l2arc_ndev == 0)
6634 		goto out;
6635 
6636 	first = NULL;
6637 	next = l2arc_dev_last;
6638 	do {
6639 		/* loop around the list looking for a non-faulted vdev */
6640 		if (next == NULL) {
6641 			next = list_head(l2arc_dev_list);
6642 		} else {
6643 			next = list_next(l2arc_dev_list, next);
6644 			if (next == NULL)
6645 				next = list_head(l2arc_dev_list);
6646 		}
6647 
6648 		/* if we have come back to the start, bail out */
6649 		if (first == NULL)
6650 			first = next;
6651 		else if (next == first)
6652 			break;
6653 
6654 	} while (vdev_is_dead(next->l2ad_vdev));
6655 
6656 	/* if we were unable to find any usable vdevs, return NULL */
6657 	if (vdev_is_dead(next->l2ad_vdev))
6658 		next = NULL;
6659 
6660 	l2arc_dev_last = next;
6661 
6662 out:
6663 	mutex_exit(&l2arc_dev_mtx);
6664 
6665 	/*
6666 	 * Grab the config lock to prevent the 'next' device from being
6667 	 * removed while we are writing to it.
6668 	 */
6669 	if (next != NULL)
6670 		spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
6671 	mutex_exit(&spa_namespace_lock);
6672 
6673 	return (next);
6674 }
6675 
6676 /*
6677  * Free buffers that were tagged for destruction.
6678  */
6679 static void
6680 l2arc_do_free_on_write()
6681 {
6682 	list_t *buflist;
6683 	l2arc_data_free_t *df, *df_prev;
6684 
6685 	mutex_enter(&l2arc_free_on_write_mtx);
6686 	buflist = l2arc_free_on_write;
6687 
6688 	for (df = list_tail(buflist); df; df = df_prev) {
6689 		df_prev = list_prev(buflist, df);
6690 		ASSERT3P(df->l2df_abd, !=, NULL);
6691 		abd_free(df->l2df_abd);
6692 		list_remove(buflist, df);
6693 		kmem_free(df, sizeof (l2arc_data_free_t));
6694 	}
6695 
6696 	mutex_exit(&l2arc_free_on_write_mtx);
6697 }
6698 
6699 /*
6700  * A write to a cache device has completed.  Update all headers to allow
6701  * reads from these buffers to begin.
6702  */
6703 static void
6704 l2arc_write_done(zio_t *zio)
6705 {
6706 	l2arc_write_callback_t *cb;
6707 	l2arc_dev_t *dev;
6708 	list_t *buflist;
6709 	arc_buf_hdr_t *head, *hdr, *hdr_prev;
6710 	kmutex_t *hash_lock;
6711 	int64_t bytes_dropped = 0;
6712 
6713 	cb = zio->io_private;
6714 	ASSERT3P(cb, !=, NULL);
6715 	dev = cb->l2wcb_dev;
6716 	ASSERT3P(dev, !=, NULL);
6717 	head = cb->l2wcb_head;
6718 	ASSERT3P(head, !=, NULL);
6719 	buflist = &dev->l2ad_buflist;
6720 	ASSERT3P(buflist, !=, NULL);
6721 	DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
6722 	    l2arc_write_callback_t *, cb);
6723 
6724 	if (zio->io_error != 0)
6725 		ARCSTAT_BUMP(arcstat_l2_writes_error);
6726 
6727 	/*
6728 	 * All writes completed, or an error was hit.
6729 	 */
6730 top:
6731 	mutex_enter(&dev->l2ad_mtx);
6732 	for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
6733 		hdr_prev = list_prev(buflist, hdr);
6734 
6735 		hash_lock = HDR_LOCK(hdr);
6736 
6737 		/*
6738 		 * We cannot use mutex_enter or else we can deadlock
6739 		 * with l2arc_write_buffers (due to swapping the order
6740 		 * the hash lock and l2ad_mtx are taken).
6741 		 */
6742 		if (!mutex_tryenter(hash_lock)) {
6743 			/*
6744 			 * Missed the hash lock. We must retry so we
6745 			 * don't leave the ARC_FLAG_L2_WRITING bit set.
6746 			 */
6747 			ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
6748 
6749 			/*
6750 			 * We don't want to rescan the headers we've
6751 			 * already marked as having been written out, so
6752 			 * we reinsert the head node so we can pick up
6753 			 * where we left off.
6754 			 */
6755 			list_remove(buflist, head);
6756 			list_insert_after(buflist, hdr, head);
6757 
6758 			mutex_exit(&dev->l2ad_mtx);
6759 
6760 			/*
6761 			 * We wait for the hash lock to become available
6762 			 * to try and prevent busy waiting, and increase
6763 			 * the chance we'll be able to acquire the lock
6764 			 * the next time around.
6765 			 */
6766 			mutex_enter(hash_lock);
6767 			mutex_exit(hash_lock);
6768 			goto top;
6769 		}
6770 
6771 		/*
6772 		 * We could not have been moved into the arc_l2c_only
6773 		 * state while in-flight due to our ARC_FLAG_L2_WRITING
6774 		 * bit being set. Let's just ensure that's being enforced.
6775 		 */
6776 		ASSERT(HDR_HAS_L1HDR(hdr));
6777 
6778 		if (zio->io_error != 0) {
6779 			/*
6780 			 * Error - drop L2ARC entry.
6781 			 */
6782 			list_remove(buflist, hdr);
6783 			arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
6784 
6785 			ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr));
6786 			ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
6787 
6788 			bytes_dropped += arc_hdr_size(hdr);
6789 			(void) zfs_refcount_remove_many(&dev->l2ad_alloc,
6790 			    arc_hdr_size(hdr), hdr);
6791 		}
6792 
6793 		/*
6794 		 * Allow ARC to begin reads and ghost list evictions to
6795 		 * this L2ARC entry.
6796 		 */
6797 		arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
6798 
6799 		mutex_exit(hash_lock);
6800 	}
6801 
6802 	atomic_inc_64(&l2arc_writes_done);
6803 	list_remove(buflist, head);
6804 	ASSERT(!HDR_HAS_L1HDR(head));
6805 	kmem_cache_free(hdr_l2only_cache, head);
6806 	mutex_exit(&dev->l2ad_mtx);
6807 
6808 	vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
6809 
6810 	l2arc_do_free_on_write();
6811 
6812 	kmem_free(cb, sizeof (l2arc_write_callback_t));
6813 }
6814 
6815 /*
6816  * A read to a cache device completed.  Validate buffer contents before
6817  * handing over to the regular ARC routines.
6818  */
6819 static void
6820 l2arc_read_done(zio_t *zio)
6821 {
6822 	l2arc_read_callback_t *cb;
6823 	arc_buf_hdr_t *hdr;
6824 	kmutex_t *hash_lock;
6825 	boolean_t valid_cksum;
6826 
6827 	ASSERT3P(zio->io_vd, !=, NULL);
6828 	ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
6829 
6830 	spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
6831 
6832 	cb = zio->io_private;
6833 	ASSERT3P(cb, !=, NULL);
6834 	hdr = cb->l2rcb_hdr;
6835 	ASSERT3P(hdr, !=, NULL);
6836 
6837 	hash_lock = HDR_LOCK(hdr);
6838 	mutex_enter(hash_lock);
6839 	ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6840 
6841 	/*
6842 	 * If the data was read into a temporary buffer,
6843 	 * move it and free the buffer.
6844 	 */
6845 	if (cb->l2rcb_abd != NULL) {
6846 		ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
6847 		if (zio->io_error == 0) {
6848 			abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd,
6849 			    arc_hdr_size(hdr));
6850 		}
6851 
6852 		/*
6853 		 * The following must be done regardless of whether
6854 		 * there was an error:
6855 		 * - free the temporary buffer
6856 		 * - point zio to the real ARC buffer
6857 		 * - set zio size accordingly
6858 		 * These are required because zio is either re-used for
6859 		 * an I/O of the block in the case of the error
6860 		 * or the zio is passed to arc_read_done() and it
6861 		 * needs real data.
6862 		 */
6863 		abd_free(cb->l2rcb_abd);
6864 		zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
6865 		zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
6866 	}
6867 
6868 	ASSERT3P(zio->io_abd, !=, NULL);
6869 
6870 	/*
6871 	 * Check this survived the L2ARC journey.
6872 	 */
6873 	ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd);
6874 	zio->io_bp_copy = cb->l2rcb_bp;	/* XXX fix in L2ARC 2.0	*/
6875 	zio->io_bp = &zio->io_bp_copy;	/* XXX fix in L2ARC 2.0	*/
6876 
6877 	valid_cksum = arc_cksum_is_equal(hdr, zio);
6878 	if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
6879 		mutex_exit(hash_lock);
6880 		zio->io_private = hdr;
6881 		arc_read_done(zio);
6882 	} else {
6883 		mutex_exit(hash_lock);
6884 		/*
6885 		 * Buffer didn't survive caching.  Increment stats and
6886 		 * reissue to the original storage device.
6887 		 */
6888 		if (zio->io_error != 0) {
6889 			ARCSTAT_BUMP(arcstat_l2_io_error);
6890 		} else {
6891 			zio->io_error = SET_ERROR(EIO);
6892 		}
6893 		if (!valid_cksum)
6894 			ARCSTAT_BUMP(arcstat_l2_cksum_bad);
6895 
6896 		/*
6897 		 * If there's no waiter, issue an async i/o to the primary
6898 		 * storage now.  If there *is* a waiter, the caller must
6899 		 * issue the i/o in a context where it's OK to block.
6900 		 */
6901 		if (zio->io_waiter == NULL) {
6902 			zio_t *pio = zio_unique_parent(zio);
6903 
6904 			ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
6905 
6906 			zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
6907 			    hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done,
6908 			    hdr, zio->io_priority, cb->l2rcb_flags,
6909 			    &cb->l2rcb_zb));
6910 		}
6911 	}
6912 
6913 	kmem_free(cb, sizeof (l2arc_read_callback_t));
6914 }
6915 
6916 /*
6917  * This is the list priority from which the L2ARC will search for pages to
6918  * cache.  This is used within loops (0..3) to cycle through lists in the
6919  * desired order.  This order can have a significant effect on cache
6920  * performance.
6921  *
6922  * Currently the metadata lists are hit first, MFU then MRU, followed by
6923  * the data lists.  This function returns a locked list, and also returns
6924  * the lock pointer.
6925  */
6926 static multilist_sublist_t *
6927 l2arc_sublist_lock(int list_num)
6928 {
6929 	multilist_t *ml = NULL;
6930 	unsigned int idx;
6931 
6932 	ASSERT(list_num >= 0 && list_num <= 3);
6933 
6934 	switch (list_num) {
6935 	case 0:
6936 		ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
6937 		break;
6938 	case 1:
6939 		ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
6940 		break;
6941 	case 2:
6942 		ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
6943 		break;
6944 	case 3:
6945 		ml = arc_mru->arcs_list[ARC_BUFC_DATA];
6946 		break;
6947 	}
6948 
6949 	/*
6950 	 * Return a randomly-selected sublist. This is acceptable
6951 	 * because the caller feeds only a little bit of data for each
6952 	 * call (8MB). Subsequent calls will result in different
6953 	 * sublists being selected.
6954 	 */
6955 	idx = multilist_get_random_index(ml);
6956 	return (multilist_sublist_lock(ml, idx));
6957 }
6958 
6959 /*
6960  * Evict buffers from the device write hand to the distance specified in
6961  * bytes.  This distance may span populated buffers, it may span nothing.
6962  * This is clearing a region on the L2ARC device ready for writing.
6963  * If the 'all' boolean is set, every buffer is evicted.
6964  */
6965 static void
6966 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
6967 {
6968 	list_t *buflist;
6969 	arc_buf_hdr_t *hdr, *hdr_prev;
6970 	kmutex_t *hash_lock;
6971 	uint64_t taddr;
6972 
6973 	buflist = &dev->l2ad_buflist;
6974 
6975 	if (!all && dev->l2ad_first) {
6976 		/*
6977 		 * This is the first sweep through the device.  There is
6978 		 * nothing to evict.
6979 		 */
6980 		return;
6981 	}
6982 
6983 	if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
6984 		/*
6985 		 * When nearing the end of the device, evict to the end
6986 		 * before the device write hand jumps to the start.
6987 		 */
6988 		taddr = dev->l2ad_end;
6989 	} else {
6990 		taddr = dev->l2ad_hand + distance;
6991 	}
6992 	DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
6993 	    uint64_t, taddr, boolean_t, all);
6994 
6995 top:
6996 	mutex_enter(&dev->l2ad_mtx);
6997 	for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
6998 		hdr_prev = list_prev(buflist, hdr);
6999 
7000 		hash_lock = HDR_LOCK(hdr);
7001 
7002 		/*
7003 		 * We cannot use mutex_enter or else we can deadlock
7004 		 * with l2arc_write_buffers (due to swapping the order
7005 		 * the hash lock and l2ad_mtx are taken).
7006 		 */
7007 		if (!mutex_tryenter(hash_lock)) {
7008 			/*
7009 			 * Missed the hash lock.  Retry.
7010 			 */
7011 			ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
7012 			mutex_exit(&dev->l2ad_mtx);
7013 			mutex_enter(hash_lock);
7014 			mutex_exit(hash_lock);
7015 			goto top;
7016 		}
7017 
7018 		/*
7019 		 * A header can't be on this list if it doesn't have L2 header.
7020 		 */
7021 		ASSERT(HDR_HAS_L2HDR(hdr));
7022 
7023 		/* Ensure this header has finished being written. */
7024 		ASSERT(!HDR_L2_WRITING(hdr));
7025 		ASSERT(!HDR_L2_WRITE_HEAD(hdr));
7026 
7027 		if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
7028 		    hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
7029 			/*
7030 			 * We've evicted to the target address,
7031 			 * or the end of the device.
7032 			 */
7033 			mutex_exit(hash_lock);
7034 			break;
7035 		}
7036 
7037 		if (!HDR_HAS_L1HDR(hdr)) {
7038 			ASSERT(!HDR_L2_READING(hdr));
7039 			/*
7040 			 * This doesn't exist in the ARC.  Destroy.
7041 			 * arc_hdr_destroy() will call list_remove()
7042 			 * and decrement arcstat_l2_lsize.
7043 			 */
7044 			arc_change_state(arc_anon, hdr, hash_lock);
7045 			arc_hdr_destroy(hdr);
7046 		} else {
7047 			ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
7048 			ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
7049 			/*
7050 			 * Invalidate issued or about to be issued
7051 			 * reads, since we may be about to write
7052 			 * over this location.
7053 			 */
7054 			if (HDR_L2_READING(hdr)) {
7055 				ARCSTAT_BUMP(arcstat_l2_evict_reading);
7056 				arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
7057 			}
7058 
7059 			arc_hdr_l2hdr_destroy(hdr);
7060 		}
7061 		mutex_exit(hash_lock);
7062 	}
7063 	mutex_exit(&dev->l2ad_mtx);
7064 }
7065 
7066 /*
7067  * Find and write ARC buffers to the L2ARC device.
7068  *
7069  * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7070  * for reading until they have completed writing.
7071  * The headroom_boost is an in-out parameter used to maintain headroom boost
7072  * state between calls to this function.
7073  *
7074  * Returns the number of bytes actually written (which may be smaller than
7075  * the delta by which the device hand has changed due to alignment).
7076  */
7077 static uint64_t
7078 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
7079 {
7080 	arc_buf_hdr_t *hdr, *hdr_prev, *head;
7081 	uint64_t write_asize, write_psize, write_lsize, headroom;
7082 	boolean_t full;
7083 	l2arc_write_callback_t *cb;
7084 	zio_t *pio, *wzio;
7085 	uint64_t guid = spa_load_guid(spa);
7086 
7087 	ASSERT3P(dev->l2ad_vdev, !=, NULL);
7088 
7089 	pio = NULL;
7090 	write_lsize = write_asize = write_psize = 0;
7091 	full = B_FALSE;
7092 	head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
7093 	arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
7094 
7095 	/*
7096 	 * Copy buffers for L2ARC writing.
7097 	 */
7098 	for (int try = 0; try <= 3; try++) {
7099 		multilist_sublist_t *mls = l2arc_sublist_lock(try);
7100 		uint64_t passed_sz = 0;
7101 
7102 		/*
7103 		 * L2ARC fast warmup.
7104 		 *
7105 		 * Until the ARC is warm and starts to evict, read from the
7106 		 * head of the ARC lists rather than the tail.
7107 		 */
7108 		if (arc_warm == B_FALSE)
7109 			hdr = multilist_sublist_head(mls);
7110 		else
7111 			hdr = multilist_sublist_tail(mls);
7112 
7113 		headroom = target_sz * l2arc_headroom;
7114 		if (zfs_compressed_arc_enabled)
7115 			headroom = (headroom * l2arc_headroom_boost) / 100;
7116 
7117 		for (; hdr; hdr = hdr_prev) {
7118 			kmutex_t *hash_lock;
7119 
7120 			if (arc_warm == B_FALSE)
7121 				hdr_prev = multilist_sublist_next(mls, hdr);
7122 			else
7123 				hdr_prev = multilist_sublist_prev(mls, hdr);
7124 
7125 			hash_lock = HDR_LOCK(hdr);
7126 			if (!mutex_tryenter(hash_lock)) {
7127 				/*
7128 				 * Skip this buffer rather than waiting.
7129 				 */
7130 				continue;
7131 			}
7132 
7133 			passed_sz += HDR_GET_LSIZE(hdr);
7134 			if (passed_sz > headroom) {
7135 				/*
7136 				 * Searched too far.
7137 				 */
7138 				mutex_exit(hash_lock);
7139 				break;
7140 			}
7141 
7142 			if (!l2arc_write_eligible(guid, hdr)) {
7143 				mutex_exit(hash_lock);
7144 				continue;
7145 			}
7146 
7147 			/*
7148 			 * We rely on the L1 portion of the header below, so
7149 			 * it's invalid for this header to have been evicted out
7150 			 * of the ghost cache, prior to being written out. The
7151 			 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
7152 			 */
7153 			ASSERT(HDR_HAS_L1HDR(hdr));
7154 
7155 			ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
7156 			ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
7157 			ASSERT3U(arc_hdr_size(hdr), >, 0);
7158 			uint64_t psize = arc_hdr_size(hdr);
7159 			uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
7160 			    psize);
7161 
7162 			if ((write_asize + asize) > target_sz) {
7163 				full = B_TRUE;
7164 				mutex_exit(hash_lock);
7165 				break;
7166 			}
7167 
7168 			if (pio == NULL) {
7169 				/*
7170 				 * Insert a dummy header on the buflist so
7171 				 * l2arc_write_done() can find where the
7172 				 * write buffers begin without searching.
7173 				 */
7174 				mutex_enter(&dev->l2ad_mtx);
7175 				list_insert_head(&dev->l2ad_buflist, head);
7176 				mutex_exit(&dev->l2ad_mtx);
7177 
7178 				cb = kmem_alloc(
7179 				    sizeof (l2arc_write_callback_t), KM_SLEEP);
7180 				cb->l2wcb_dev = dev;
7181 				cb->l2wcb_head = head;
7182 				pio = zio_root(spa, l2arc_write_done, cb,
7183 				    ZIO_FLAG_CANFAIL);
7184 			}
7185 
7186 			hdr->b_l2hdr.b_dev = dev;
7187 			hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
7188 			arc_hdr_set_flags(hdr,
7189 			    ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
7190 
7191 			mutex_enter(&dev->l2ad_mtx);
7192 			list_insert_head(&dev->l2ad_buflist, hdr);
7193 			mutex_exit(&dev->l2ad_mtx);
7194 
7195 			(void) zfs_refcount_add_many(&dev->l2ad_alloc, psize,
7196 			    hdr);
7197 
7198 			/*
7199 			 * Normally the L2ARC can use the hdr's data, but if
7200 			 * we're sharing data between the hdr and one of its
7201 			 * bufs, L2ARC needs its own copy of the data so that
7202 			 * the ZIO below can't race with the buf consumer.
7203 			 * Another case where we need to create a copy of the
7204 			 * data is when the buffer size is not device-aligned
7205 			 * and we need to pad the block to make it such.
7206 			 * That also keeps the clock hand suitably aligned.
7207 			 *
7208 			 * To ensure that the copy will be available for the
7209 			 * lifetime of the ZIO and be cleaned up afterwards, we
7210 			 * add it to the l2arc_free_on_write queue.
7211 			 */
7212 			abd_t *to_write;
7213 			if (!HDR_SHARED_DATA(hdr) && psize == asize) {
7214 				to_write = hdr->b_l1hdr.b_pabd;
7215 			} else {
7216 				to_write = abd_alloc_for_io(asize,
7217 				    HDR_ISTYPE_METADATA(hdr));
7218 				abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
7219 				if (asize != psize) {
7220 					abd_zero_off(to_write, psize,
7221 					    asize - psize);
7222 				}
7223 				l2arc_free_abd_on_write(to_write, asize,
7224 				    arc_buf_type(hdr));
7225 			}
7226 			wzio = zio_write_phys(pio, dev->l2ad_vdev,
7227 			    hdr->b_l2hdr.b_daddr, asize, to_write,
7228 			    ZIO_CHECKSUM_OFF, NULL, hdr,
7229 			    ZIO_PRIORITY_ASYNC_WRITE,
7230 			    ZIO_FLAG_CANFAIL, B_FALSE);
7231 
7232 			write_lsize += HDR_GET_LSIZE(hdr);
7233 			DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
7234 			    zio_t *, wzio);
7235 
7236 			write_psize += psize;
7237 			write_asize += asize;
7238 			dev->l2ad_hand += asize;
7239 
7240 			mutex_exit(hash_lock);
7241 
7242 			(void) zio_nowait(wzio);
7243 		}
7244 
7245 		multilist_sublist_unlock(mls);
7246 
7247 		if (full == B_TRUE)
7248 			break;
7249 	}
7250 
7251 	/* No buffers selected for writing? */
7252 	if (pio == NULL) {
7253 		ASSERT0(write_lsize);
7254 		ASSERT(!HDR_HAS_L1HDR(head));
7255 		kmem_cache_free(hdr_l2only_cache, head);
7256 		return (0);
7257 	}
7258 
7259 	ASSERT3U(write_asize, <=, target_sz);
7260 	ARCSTAT_BUMP(arcstat_l2_writes_sent);
7261 	ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
7262 	ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
7263 	ARCSTAT_INCR(arcstat_l2_psize, write_psize);
7264 	vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
7265 
7266 	/*
7267 	 * Bump device hand to the device start if it is approaching the end.
7268 	 * l2arc_evict() will already have evicted ahead for this case.
7269 	 */
7270 	if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
7271 		dev->l2ad_hand = dev->l2ad_start;
7272 		dev->l2ad_first = B_FALSE;
7273 	}
7274 
7275 	dev->l2ad_writing = B_TRUE;
7276 	(void) zio_wait(pio);
7277 	dev->l2ad_writing = B_FALSE;
7278 
7279 	return (write_asize);
7280 }
7281 
7282 /*
7283  * This thread feeds the L2ARC at regular intervals.  This is the beating
7284  * heart of the L2ARC.
7285  */
7286 /* ARGSUSED */
7287 static void
7288 l2arc_feed_thread(void *unused)
7289 {
7290 	callb_cpr_t cpr;
7291 	l2arc_dev_t *dev;
7292 	spa_t *spa;
7293 	uint64_t size, wrote;
7294 	clock_t begin, next = ddi_get_lbolt();
7295 
7296 	CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
7297 
7298 	mutex_enter(&l2arc_feed_thr_lock);
7299 
7300 	while (l2arc_thread_exit == 0) {
7301 		CALLB_CPR_SAFE_BEGIN(&cpr);
7302 		(void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
7303 		    next);
7304 		CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
7305 		next = ddi_get_lbolt() + hz;
7306 
7307 		/*
7308 		 * Quick check for L2ARC devices.
7309 		 */
7310 		mutex_enter(&l2arc_dev_mtx);
7311 		if (l2arc_ndev == 0) {
7312 			mutex_exit(&l2arc_dev_mtx);
7313 			continue;
7314 		}
7315 		mutex_exit(&l2arc_dev_mtx);
7316 		begin = ddi_get_lbolt();
7317 
7318 		/*
7319 		 * This selects the next l2arc device to write to, and in
7320 		 * doing so the next spa to feed from: dev->l2ad_spa.   This
7321 		 * will return NULL if there are now no l2arc devices or if
7322 		 * they are all faulted.
7323 		 *
7324 		 * If a device is returned, its spa's config lock is also
7325 		 * held to prevent device removal.  l2arc_dev_get_next()
7326 		 * will grab and release l2arc_dev_mtx.
7327 		 */
7328 		if ((dev = l2arc_dev_get_next()) == NULL)
7329 			continue;
7330 
7331 		spa = dev->l2ad_spa;
7332 		ASSERT3P(spa, !=, NULL);
7333 
7334 		/*
7335 		 * If the pool is read-only then force the feed thread to
7336 		 * sleep a little longer.
7337 		 */
7338 		if (!spa_writeable(spa)) {
7339 			next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
7340 			spa_config_exit(spa, SCL_L2ARC, dev);
7341 			continue;
7342 		}
7343 
7344 		/*
7345 		 * Avoid contributing to memory pressure.
7346 		 */
7347 		if (arc_reclaim_needed()) {
7348 			ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
7349 			spa_config_exit(spa, SCL_L2ARC, dev);
7350 			continue;
7351 		}
7352 
7353 		ARCSTAT_BUMP(arcstat_l2_feeds);
7354 
7355 		size = l2arc_write_size();
7356 
7357 		/*
7358 		 * Evict L2ARC buffers that will be overwritten.
7359 		 */
7360 		l2arc_evict(dev, size, B_FALSE);
7361 
7362 		/*
7363 		 * Write ARC buffers.
7364 		 */
7365 		wrote = l2arc_write_buffers(spa, dev, size);
7366 
7367 		/*
7368 		 * Calculate interval between writes.
7369 		 */
7370 		next = l2arc_write_interval(begin, size, wrote);
7371 		spa_config_exit(spa, SCL_L2ARC, dev);
7372 	}
7373 
7374 	l2arc_thread_exit = 0;
7375 	cv_broadcast(&l2arc_feed_thr_cv);
7376 	CALLB_CPR_EXIT(&cpr);		/* drops l2arc_feed_thr_lock */
7377 	thread_exit();
7378 }
7379 
7380 boolean_t
7381 l2arc_vdev_present(vdev_t *vd)
7382 {
7383 	l2arc_dev_t *dev;
7384 
7385 	mutex_enter(&l2arc_dev_mtx);
7386 	for (dev = list_head(l2arc_dev_list); dev != NULL;
7387 	    dev = list_next(l2arc_dev_list, dev)) {
7388 		if (dev->l2ad_vdev == vd)
7389 			break;
7390 	}
7391 	mutex_exit(&l2arc_dev_mtx);
7392 
7393 	return (dev != NULL);
7394 }
7395 
7396 /*
7397  * Add a vdev for use by the L2ARC.  By this point the spa has already
7398  * validated the vdev and opened it.
7399  */
7400 void
7401 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
7402 {
7403 	l2arc_dev_t *adddev;
7404 
7405 	ASSERT(!l2arc_vdev_present(vd));
7406 
7407 	/*
7408 	 * Create a new l2arc device entry.
7409 	 */
7410 	adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
7411 	adddev->l2ad_spa = spa;
7412 	adddev->l2ad_vdev = vd;
7413 	adddev->l2ad_start = VDEV_LABEL_START_SIZE;
7414 	adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
7415 	adddev->l2ad_hand = adddev->l2ad_start;
7416 	adddev->l2ad_first = B_TRUE;
7417 	adddev->l2ad_writing = B_FALSE;
7418 
7419 	mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
7420 	/*
7421 	 * This is a list of all ARC buffers that are still valid on the
7422 	 * device.
7423 	 */
7424 	list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
7425 	    offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
7426 
7427 	vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
7428 	zfs_refcount_create(&adddev->l2ad_alloc);
7429 
7430 	/*
7431 	 * Add device to global list
7432 	 */
7433 	mutex_enter(&l2arc_dev_mtx);
7434 	list_insert_head(l2arc_dev_list, adddev);
7435 	atomic_inc_64(&l2arc_ndev);
7436 	mutex_exit(&l2arc_dev_mtx);
7437 }
7438 
7439 /*
7440  * Remove a vdev from the L2ARC.
7441  */
7442 void
7443 l2arc_remove_vdev(vdev_t *vd)
7444 {
7445 	l2arc_dev_t *dev, *nextdev, *remdev = NULL;
7446 
7447 	/*
7448 	 * Find the device by vdev
7449 	 */
7450 	mutex_enter(&l2arc_dev_mtx);
7451 	for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
7452 		nextdev = list_next(l2arc_dev_list, dev);
7453 		if (vd == dev->l2ad_vdev) {
7454 			remdev = dev;
7455 			break;
7456 		}
7457 	}
7458 	ASSERT3P(remdev, !=, NULL);
7459 
7460 	/*
7461 	 * Remove device from global list
7462 	 */
7463 	list_remove(l2arc_dev_list, remdev);
7464 	l2arc_dev_last = NULL;		/* may have been invalidated */
7465 	atomic_dec_64(&l2arc_ndev);
7466 	mutex_exit(&l2arc_dev_mtx);
7467 
7468 	/*
7469 	 * Clear all buflists and ARC references.  L2ARC device flush.
7470 	 */
7471 	l2arc_evict(remdev, 0, B_TRUE);
7472 	list_destroy(&remdev->l2ad_buflist);
7473 	mutex_destroy(&remdev->l2ad_mtx);
7474 	zfs_refcount_destroy(&remdev->l2ad_alloc);
7475 	kmem_free(remdev, sizeof (l2arc_dev_t));
7476 }
7477 
7478 void
7479 l2arc_init(void)
7480 {
7481 	l2arc_thread_exit = 0;
7482 	l2arc_ndev = 0;
7483 	l2arc_writes_sent = 0;
7484 	l2arc_writes_done = 0;
7485 
7486 	mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
7487 	cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
7488 	mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
7489 	mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
7490 
7491 	l2arc_dev_list = &L2ARC_dev_list;
7492 	l2arc_free_on_write = &L2ARC_free_on_write;
7493 	list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
7494 	    offsetof(l2arc_dev_t, l2ad_node));
7495 	list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
7496 	    offsetof(l2arc_data_free_t, l2df_list_node));
7497 }
7498 
7499 void
7500 l2arc_fini(void)
7501 {
7502 	/*
7503 	 * This is called from dmu_fini(), which is called from spa_fini();
7504 	 * Because of this, we can assume that all l2arc devices have
7505 	 * already been removed when the pools themselves were removed.
7506 	 */
7507 
7508 	l2arc_do_free_on_write();
7509 
7510 	mutex_destroy(&l2arc_feed_thr_lock);
7511 	cv_destroy(&l2arc_feed_thr_cv);
7512 	mutex_destroy(&l2arc_dev_mtx);
7513 	mutex_destroy(&l2arc_free_on_write_mtx);
7514 
7515 	list_destroy(l2arc_dev_list);
7516 	list_destroy(l2arc_free_on_write);
7517 }
7518 
7519 void
7520 l2arc_start(void)
7521 {
7522 	if (!(spa_mode_global & FWRITE))
7523 		return;
7524 
7525 	(void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
7526 	    TS_RUN, minclsyspri);
7527 }
7528 
7529 void
7530 l2arc_stop(void)
7531 {
7532 	if (!(spa_mode_global & FWRITE))
7533 		return;
7534 
7535 	mutex_enter(&l2arc_feed_thr_lock);
7536 	cv_signal(&l2arc_feed_thr_cv);	/* kick thread out of startup */
7537 	l2arc_thread_exit = 1;
7538 	while (l2arc_thread_exit != 0)
7539 		cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
7540 	mutex_exit(&l2arc_feed_thr_lock);
7541 }
7542