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.
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.
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]
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.
30 * DVA-based Adjustable Replacement Cache
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:
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.
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.
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
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.
70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71 * by N. Megiddo & D. Modha, FAST 2003
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
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).
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.
93 * buf_hash_remove() expects the appropriate hash mutex to be
94 * already held before it is invoked.
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.
102 * It as also possible to register a callback which is run when the
103 * arc_meta_limit is reached and no buffers can be safely evicted. In
104 * this case the arc user should drop a reference on some arc buffers so
105 * they can be reclaimed and the arc_meta_limit honored. For example,
106 * when using the ZPL each dentry holds a references on a znode. These
107 * dentries must be pruned before the arc buffer holding the znode can
110 * Note that the majority of the performance stats are manipulated
111 * with atomic operations.
113 * The L2ARC uses the l2ad_mtx on each vdev for the following:
115 * - L2ARC buflist creation
116 * - L2ARC buflist eviction
117 * - L2ARC write completion, which walks L2ARC buflists
118 * - ARC header destruction, as it removes from L2ARC buflists
119 * - ARC header release, as it removes from L2ARC buflists
125 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
126 * This structure can point either to a block that is still in the cache or to
127 * one that is only accessible in an L2 ARC device, or it can provide
128 * information about a block that was recently evicted. If a block is
129 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
130 * information to retrieve it from the L2ARC device. This information is
131 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
132 * that is in this state cannot access the data directly.
134 * Blocks that are actively being referenced or have not been evicted
135 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
136 * the arc_buf_hdr_t that will point to the data block in memory. A block can
137 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
138 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
139 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
141 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
142 * ability to store the physical data (b_pabd) associated with the DVA of the
143 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
144 * it will match its on-disk compression characteristics. This behavior can be
145 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
146 * compressed ARC functionality is disabled, the b_pabd will point to an
147 * uncompressed version of the on-disk data.
149 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
150 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
151 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
152 * consumer. The ARC will provide references to this data and will keep it
153 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
154 * data block and will evict any arc_buf_t that is no longer referenced. The
155 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
156 * "overhead_size" kstat.
158 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
159 * compressed form. The typical case is that consumers will want uncompressed
160 * data, and when that happens a new data buffer is allocated where the data is
161 * decompressed for them to use. Currently the only consumer who wants
162 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
163 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
164 * with the arc_buf_hdr_t.
166 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
167 * first one is owned by a compressed send consumer (and therefore references
168 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
169 * used by any other consumer (and has its own uncompressed copy of the data
184 * | b_buf +------------>+-----------+ arc_buf_t
185 * | b_pabd +-+ |b_next +---->+-----------+
186 * +-----------+ | |-----------| |b_next +-->NULL
187 * | |b_comp = T | +-----------+
188 * | |b_data +-+ |b_comp = F |
189 * | +-----------+ | |b_data +-+
190 * +->+------+ | +-----------+ |
192 * data | |<--------------+ | uncompressed
193 * +------+ compressed, | data
194 * shared +-->+------+
199 * When a consumer reads a block, the ARC must first look to see if the
200 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
201 * arc_buf_t and either copies uncompressed data into a new data buffer from an
202 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
203 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
204 * hdr is compressed and the desired compression characteristics of the
205 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
206 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
207 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
208 * be anywhere in the hdr's list.
210 * The diagram below shows an example of an uncompressed ARC hdr that is
211 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
212 * the last element in the buf list):
224 * | | arc_buf_t (shared)
225 * | b_buf +------------>+---------+ arc_buf_t
226 * | | |b_next +---->+---------+
227 * | b_pabd +-+ |---------| |b_next +-->NULL
228 * +-----------+ | | | +---------+
230 * | +---------+ | |b_data +-+
231 * +->+------+ | +---------+ |
233 * uncompressed | | | |
236 * | uncompressed | | |
239 * +---------------------------------+
241 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
242 * since the physical block is about to be rewritten. The new data contents
243 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
244 * it may compress the data before writing it to disk. The ARC will be called
245 * with the transformed data and will bcopy the transformed on-disk block into
246 * a newly allocated b_pabd. Writes are always done into buffers which have
247 * either been loaned (and hence are new and don't have other readers) or
248 * buffers which have been released (and hence have their own hdr, if there
249 * were originally other readers of the buf's original hdr). This ensures that
250 * the ARC only needs to update a single buf and its hdr after a write occurs.
252 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
253 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
254 * that when compressed ARC is enabled that the L2ARC blocks are identical
255 * to the on-disk block in the main data pool. This provides a significant
256 * advantage since the ARC can leverage the bp's checksum when reading from the
257 * L2ARC to determine if the contents are valid. However, if the compressed
258 * ARC is disabled, then the L2ARC's block must be transformed to look
259 * like the physical block in the main data pool before comparing the
260 * checksum and determining its validity.
265 #include <sys/spa_impl.h>
266 #include <sys/zio_compress.h>
267 #include <sys/zio_checksum.h>
268 #include <sys/zfs_context.h>
270 #include <sys/refcount.h>
271 #include <sys/vdev.h>
272 #include <sys/vdev_impl.h>
273 #include <sys/dsl_pool.h>
274 #include <sys/zio_checksum.h>
275 #include <sys/multilist.h>
278 #include <sys/dnlc.h>
279 #include <sys/racct.h>
281 #include <sys/callb.h>
282 #include <sys/kstat.h>
283 #include <sys/trim_map.h>
284 #include <zfs_fletcher.h>
286 #include <sys/aggsum.h>
287 #include <sys/cityhash.h>
289 #include <machine/vmparam.h>
293 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
294 boolean_t arc_watch = B_FALSE;
299 static kmutex_t arc_reclaim_lock;
300 static kcondvar_t arc_reclaim_thread_cv;
301 static boolean_t arc_reclaim_thread_exit;
302 static kcondvar_t arc_reclaim_waiters_cv;
304 static kmutex_t arc_dnlc_evicts_lock;
305 static kcondvar_t arc_dnlc_evicts_cv;
306 static boolean_t arc_dnlc_evicts_thread_exit;
308 uint_t arc_reduce_dnlc_percent = 3;
311 * The number of headers to evict in arc_evict_state_impl() before
312 * dropping the sublist lock and evicting from another sublist. A lower
313 * value means we're more likely to evict the "correct" header (i.e. the
314 * oldest header in the arc state), but comes with higher overhead
315 * (i.e. more invocations of arc_evict_state_impl()).
317 int zfs_arc_evict_batch_limit = 10;
319 /* number of seconds before growing cache again */
320 static int arc_grow_retry = 60;
322 /* number of milliseconds before attempting a kmem-cache-reap */
323 static int arc_kmem_cache_reap_retry_ms = 0;
325 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
326 int zfs_arc_overflow_shift = 8;
328 /* shift of arc_c for calculating both min and max arc_p */
329 static int arc_p_min_shift = 4;
331 /* log2(fraction of arc to reclaim) */
332 static int arc_shrink_shift = 7;
335 * log2(fraction of ARC which must be free to allow growing).
336 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
337 * when reading a new block into the ARC, we will evict an equal-sized block
340 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
341 * we will still not allow it to grow.
343 int arc_no_grow_shift = 5;
347 * minimum lifespan of a prefetch block in clock ticks
348 * (initialized in arc_init())
350 static int zfs_arc_min_prefetch_ms = 1;
351 static int zfs_arc_min_prescient_prefetch_ms = 6;
354 * If this percent of memory is free, don't throttle.
356 int arc_lotsfree_percent = 10;
359 extern boolean_t zfs_prefetch_disable;
362 * The arc has filled available memory and has now warmed up.
364 static boolean_t arc_warm;
367 * log2 fraction of the zio arena to keep free.
369 int arc_zio_arena_free_shift = 2;
372 * These tunables are for performance analysis.
374 uint64_t zfs_arc_max;
375 uint64_t zfs_arc_min;
376 uint64_t zfs_arc_meta_limit = 0;
377 uint64_t zfs_arc_meta_min = 0;
378 uint64_t zfs_arc_dnode_limit = 0;
379 uint64_t zfs_arc_dnode_reduce_percent = 10;
380 int zfs_arc_grow_retry = 0;
381 int zfs_arc_shrink_shift = 0;
382 int zfs_arc_no_grow_shift = 0;
383 int zfs_arc_p_min_shift = 0;
384 uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
385 u_int zfs_arc_free_target = 0;
387 /* Absolute min for arc min / max is 16MB. */
388 static uint64_t arc_abs_min = 16 << 20;
391 * ARC dirty data constraints for arc_tempreserve_space() throttle
393 uint_t zfs_arc_dirty_limit_percent = 50; /* total dirty data limit */
394 uint_t zfs_arc_anon_limit_percent = 25; /* anon block dirty limit */
395 uint_t zfs_arc_pool_dirty_percent = 20; /* each pool's anon allowance */
397 boolean_t zfs_compressed_arc_enabled = B_TRUE;
399 static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
400 static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
401 static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS);
402 static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS);
403 static int sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS);
405 #if defined(__FreeBSD__) && defined(_KERNEL)
407 arc_free_target_init(void *unused __unused)
410 zfs_arc_free_target = vm_cnt.v_free_target;
412 SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
413 arc_free_target_init, NULL);
415 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
416 TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
417 TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
418 TUNABLE_INT("vfs.zfs.arc_grow_retry", &zfs_arc_grow_retry);
419 TUNABLE_INT("vfs.zfs.arc_no_grow_shift", &zfs_arc_no_grow_shift);
420 SYSCTL_DECL(_vfs_zfs);
421 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN,
422 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size");
423 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN,
424 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size");
425 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_no_grow_shift, CTLTYPE_U32 | CTLFLAG_RWTUN,
426 0, sizeof(uint32_t), sysctl_vfs_zfs_arc_no_grow_shift, "U",
427 "log2(fraction of ARC which must be free to allow growing)");
428 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
429 &zfs_arc_average_blocksize, 0,
430 "ARC average blocksize");
431 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
432 &arc_shrink_shift, 0,
433 "log2(fraction of arc to reclaim)");
434 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_grow_retry, CTLFLAG_RW,
436 "Wait in seconds before considering growing ARC");
437 SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN,
438 &zfs_compressed_arc_enabled, 0,
439 "Enable compressed ARC");
440 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_kmem_cache_reap_retry_ms, CTLFLAG_RWTUN,
441 &arc_kmem_cache_reap_retry_ms, 0,
442 "Interval between ARC kmem_cache reapings");
445 * We don't have a tunable for arc_free_target due to the dependency on
446 * pagedaemon initialisation.
448 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
449 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
450 sysctl_vfs_zfs_arc_free_target, "IU",
451 "Desired number of free pages below which ARC triggers reclaim");
454 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
459 val = zfs_arc_free_target;
460 err = sysctl_handle_int(oidp, &val, 0, req);
461 if (err != 0 || req->newptr == NULL)
466 if (val > vm_cnt.v_page_count)
469 zfs_arc_free_target = val;
475 * Must be declared here, before the definition of corresponding kstat
476 * macro which uses the same names will confuse the compiler.
478 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
479 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
480 sysctl_vfs_zfs_arc_meta_limit, "QU",
481 "ARC metadata limit");
485 * Note that buffers can be in one of 6 states:
486 * ARC_anon - anonymous (discussed below)
487 * ARC_mru - recently used, currently cached
488 * ARC_mru_ghost - recentely used, no longer in cache
489 * ARC_mfu - frequently used, currently cached
490 * ARC_mfu_ghost - frequently used, no longer in cache
491 * ARC_l2c_only - exists in L2ARC but not other states
492 * When there are no active references to the buffer, they are
493 * are linked onto a list in one of these arc states. These are
494 * the only buffers that can be evicted or deleted. Within each
495 * state there are multiple lists, one for meta-data and one for
496 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
497 * etc.) is tracked separately so that it can be managed more
498 * explicitly: favored over data, limited explicitly.
500 * Anonymous buffers are buffers that are not associated with
501 * a DVA. These are buffers that hold dirty block copies
502 * before they are written to stable storage. By definition,
503 * they are "ref'd" and are considered part of arc_mru
504 * that cannot be freed. Generally, they will aquire a DVA
505 * as they are written and migrate onto the arc_mru list.
507 * The ARC_l2c_only state is for buffers that are in the second
508 * level ARC but no longer in any of the ARC_m* lists. The second
509 * level ARC itself may also contain buffers that are in any of
510 * the ARC_m* states - meaning that a buffer can exist in two
511 * places. The reason for the ARC_l2c_only state is to keep the
512 * buffer header in the hash table, so that reads that hit the
513 * second level ARC benefit from these fast lookups.
516 typedef struct arc_state {
518 * list of evictable buffers
520 multilist_t *arcs_list[ARC_BUFC_NUMTYPES];
522 * total amount of evictable data in this state
524 refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
526 * total amount of data in this state; this includes: evictable,
527 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
529 refcount_t arcs_size;
531 * supports the "dbufs" kstat
533 arc_state_type_t arcs_state;
537 * Percentage that can be consumed by dnodes of ARC meta buffers.
539 int zfs_arc_meta_prune = 10000;
540 unsigned long zfs_arc_dnode_limit_percent = 10;
541 int zfs_arc_meta_strategy = ARC_STRATEGY_META_BALANCED;
542 int zfs_arc_meta_adjust_restarts = 4096;
545 static arc_state_t ARC_anon;
546 static arc_state_t ARC_mru;
547 static arc_state_t ARC_mru_ghost;
548 static arc_state_t ARC_mfu;
549 static arc_state_t ARC_mfu_ghost;
550 static arc_state_t ARC_l2c_only;
552 typedef struct arc_stats {
553 kstat_named_t arcstat_hits;
554 kstat_named_t arcstat_misses;
555 kstat_named_t arcstat_demand_data_hits;
556 kstat_named_t arcstat_demand_data_misses;
557 kstat_named_t arcstat_demand_metadata_hits;
558 kstat_named_t arcstat_demand_metadata_misses;
559 kstat_named_t arcstat_prefetch_data_hits;
560 kstat_named_t arcstat_prefetch_data_misses;
561 kstat_named_t arcstat_prefetch_metadata_hits;
562 kstat_named_t arcstat_prefetch_metadata_misses;
563 kstat_named_t arcstat_mru_hits;
564 kstat_named_t arcstat_mru_ghost_hits;
565 kstat_named_t arcstat_mfu_hits;
566 kstat_named_t arcstat_mfu_ghost_hits;
567 kstat_named_t arcstat_allocated;
568 kstat_named_t arcstat_deleted;
570 * Number of buffers that could not be evicted because the hash lock
571 * was held by another thread. The lock may not necessarily be held
572 * by something using the same buffer, since hash locks are shared
573 * by multiple buffers.
575 kstat_named_t arcstat_mutex_miss;
577 * Number of buffers skipped when updating the access state due to the
578 * header having already been released after acquiring the hash lock.
580 kstat_named_t arcstat_access_skip;
582 * Number of buffers skipped because they have I/O in progress, are
583 * indirect prefetch buffers that have not lived long enough, or are
584 * not from the spa we're trying to evict from.
586 kstat_named_t arcstat_evict_skip;
588 * Number of times arc_evict_state() was unable to evict enough
589 * buffers to reach it's target amount.
591 kstat_named_t arcstat_evict_not_enough;
592 kstat_named_t arcstat_evict_l2_cached;
593 kstat_named_t arcstat_evict_l2_eligible;
594 kstat_named_t arcstat_evict_l2_ineligible;
595 kstat_named_t arcstat_evict_l2_skip;
596 kstat_named_t arcstat_hash_elements;
597 kstat_named_t arcstat_hash_elements_max;
598 kstat_named_t arcstat_hash_collisions;
599 kstat_named_t arcstat_hash_chains;
600 kstat_named_t arcstat_hash_chain_max;
601 kstat_named_t arcstat_p;
602 kstat_named_t arcstat_c;
603 kstat_named_t arcstat_c_min;
604 kstat_named_t arcstat_c_max;
605 /* Not updated directly; only synced in arc_kstat_update. */
606 kstat_named_t arcstat_size;
608 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
609 * Note that the compressed bytes may match the uncompressed bytes
610 * if the block is either not compressed or compressed arc is disabled.
612 kstat_named_t arcstat_compressed_size;
614 * Uncompressed size of the data stored in b_pabd. If compressed
615 * arc is disabled then this value will be identical to the stat
618 kstat_named_t arcstat_uncompressed_size;
620 * Number of bytes stored in all the arc_buf_t's. This is classified
621 * as "overhead" since this data is typically short-lived and will
622 * be evicted from the arc when it becomes unreferenced unless the
623 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
624 * values have been set (see comment in dbuf.c for more information).
626 kstat_named_t arcstat_overhead_size;
628 * Number of bytes consumed by internal ARC structures necessary
629 * for tracking purposes; these structures are not actually
630 * backed by ARC buffers. This includes arc_buf_hdr_t structures
631 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
632 * caches), and arc_buf_t structures (allocated via arc_buf_t
634 * Not updated directly; only synced in arc_kstat_update.
636 kstat_named_t arcstat_hdr_size;
638 * Number of bytes consumed by ARC buffers of type equal to
639 * ARC_BUFC_DATA. This is generally consumed by buffers backing
640 * on disk user data (e.g. plain file contents).
641 * Not updated directly; only synced in arc_kstat_update.
643 kstat_named_t arcstat_data_size;
645 * Number of bytes consumed by ARC buffers of type equal to
646 * ARC_BUFC_METADATA. This is generally consumed by buffers
647 * backing on disk data that is used for internal ZFS
648 * structures (e.g. ZAP, dnode, indirect blocks, etc).
649 * Not updated directly; only synced in arc_kstat_update.
651 kstat_named_t arcstat_metadata_size;
653 * Number of bytes consumed by dmu_buf_impl_t objects.
655 kstat_named_t arcstat_dbuf_size;
657 * Number of bytes consumed by dnode_t objects.
659 kstat_named_t arcstat_dnode_size;
661 * Number of bytes consumed by bonus buffers.
663 kstat_named_t arcstat_bonus_size;
665 * Total number of bytes consumed by ARC buffers residing in the
666 * arc_anon state. This includes *all* buffers in the arc_anon
667 * state; e.g. data, metadata, evictable, and unevictable buffers
668 * are all included in this value.
669 * Not updated directly; only synced in arc_kstat_update.
671 kstat_named_t arcstat_anon_size;
673 * Number of bytes consumed by ARC buffers that meet the
674 * following criteria: backing buffers of type ARC_BUFC_DATA,
675 * residing in the arc_anon state, and are eligible for eviction
676 * (e.g. have no outstanding holds on the buffer).
677 * Not updated directly; only synced in arc_kstat_update.
679 kstat_named_t arcstat_anon_evictable_data;
681 * Number of bytes consumed by ARC buffers that meet the
682 * following criteria: backing buffers of type ARC_BUFC_METADATA,
683 * residing in the arc_anon state, and are eligible for eviction
684 * (e.g. have no outstanding holds on the buffer).
685 * Not updated directly; only synced in arc_kstat_update.
687 kstat_named_t arcstat_anon_evictable_metadata;
689 * Total number of bytes consumed by ARC buffers residing in the
690 * arc_mru state. This includes *all* buffers in the arc_mru
691 * state; e.g. data, metadata, evictable, and unevictable buffers
692 * are all included in this value.
693 * Not updated directly; only synced in arc_kstat_update.
695 kstat_named_t arcstat_mru_size;
697 * Number of bytes consumed by ARC buffers that meet the
698 * following criteria: backing buffers of type ARC_BUFC_DATA,
699 * residing in the arc_mru state, and are eligible for eviction
700 * (e.g. have no outstanding holds on the buffer).
701 * Not updated directly; only synced in arc_kstat_update.
703 kstat_named_t arcstat_mru_evictable_data;
705 * Number of bytes consumed by ARC buffers that meet the
706 * following criteria: backing buffers of type ARC_BUFC_METADATA,
707 * residing in the arc_mru state, and are eligible for eviction
708 * (e.g. have no outstanding holds on the buffer).
709 * Not updated directly; only synced in arc_kstat_update.
711 kstat_named_t arcstat_mru_evictable_metadata;
713 * Total number of bytes that *would have been* consumed by ARC
714 * buffers in the arc_mru_ghost state. The key thing to note
715 * here, is the fact that this size doesn't actually indicate
716 * RAM consumption. The ghost lists only consist of headers and
717 * don't actually have ARC buffers linked off of these headers.
718 * Thus, *if* the headers had associated ARC buffers, these
719 * buffers *would have* consumed this number of bytes.
720 * Not updated directly; only synced in arc_kstat_update.
722 kstat_named_t arcstat_mru_ghost_size;
724 * Number of bytes that *would have been* consumed by ARC
725 * buffers that are eligible for eviction, of type
726 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
727 * Not updated directly; only synced in arc_kstat_update.
729 kstat_named_t arcstat_mru_ghost_evictable_data;
731 * Number of bytes that *would have been* consumed by ARC
732 * buffers that are eligible for eviction, of type
733 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
734 * Not updated directly; only synced in arc_kstat_update.
736 kstat_named_t arcstat_mru_ghost_evictable_metadata;
738 * Total number of bytes consumed by ARC buffers residing in the
739 * arc_mfu state. This includes *all* buffers in the arc_mfu
740 * state; e.g. data, metadata, evictable, and unevictable buffers
741 * are all included in this value.
742 * Not updated directly; only synced in arc_kstat_update.
744 kstat_named_t arcstat_mfu_size;
746 * Number of bytes consumed by ARC buffers that are eligible for
747 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
749 * Not updated directly; only synced in arc_kstat_update.
751 kstat_named_t arcstat_mfu_evictable_data;
753 * Number of bytes consumed by ARC buffers that are eligible for
754 * eviction, of type ARC_BUFC_METADATA, and reside in the
756 * Not updated directly; only synced in arc_kstat_update.
758 kstat_named_t arcstat_mfu_evictable_metadata;
760 * Total number of bytes that *would have been* consumed by ARC
761 * buffers in the arc_mfu_ghost state. See the comment above
762 * arcstat_mru_ghost_size for more details.
763 * Not updated directly; only synced in arc_kstat_update.
765 kstat_named_t arcstat_mfu_ghost_size;
767 * Number of bytes that *would have been* consumed by ARC
768 * buffers that are eligible for eviction, of type
769 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
770 * Not updated directly; only synced in arc_kstat_update.
772 kstat_named_t arcstat_mfu_ghost_evictable_data;
774 * Number of bytes that *would have been* consumed by ARC
775 * buffers that are eligible for eviction, of type
776 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
777 * Not updated directly; only synced in arc_kstat_update.
779 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
780 kstat_named_t arcstat_l2_hits;
781 kstat_named_t arcstat_l2_misses;
782 kstat_named_t arcstat_l2_feeds;
783 kstat_named_t arcstat_l2_rw_clash;
784 kstat_named_t arcstat_l2_read_bytes;
785 kstat_named_t arcstat_l2_write_bytes;
786 kstat_named_t arcstat_l2_writes_sent;
787 kstat_named_t arcstat_l2_writes_done;
788 kstat_named_t arcstat_l2_writes_error;
789 kstat_named_t arcstat_l2_writes_lock_retry;
790 kstat_named_t arcstat_l2_evict_lock_retry;
791 kstat_named_t arcstat_l2_evict_reading;
792 kstat_named_t arcstat_l2_evict_l1cached;
793 kstat_named_t arcstat_l2_free_on_write;
794 kstat_named_t arcstat_l2_abort_lowmem;
795 kstat_named_t arcstat_l2_cksum_bad;
796 kstat_named_t arcstat_l2_io_error;
797 kstat_named_t arcstat_l2_lsize;
798 kstat_named_t arcstat_l2_psize;
799 /* Not updated directly; only synced in arc_kstat_update. */
800 kstat_named_t arcstat_l2_hdr_size;
801 kstat_named_t arcstat_l2_write_trylock_fail;
802 kstat_named_t arcstat_l2_write_passed_headroom;
803 kstat_named_t arcstat_l2_write_spa_mismatch;
804 kstat_named_t arcstat_l2_write_in_l2;
805 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
806 kstat_named_t arcstat_l2_write_not_cacheable;
807 kstat_named_t arcstat_l2_write_full;
808 kstat_named_t arcstat_l2_write_buffer_iter;
809 kstat_named_t arcstat_l2_write_pios;
810 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
811 kstat_named_t arcstat_l2_write_buffer_list_iter;
812 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
813 kstat_named_t arcstat_memory_throttle_count;
814 kstat_named_t arcstat_memory_direct_count;
815 kstat_named_t arcstat_memory_indirect_count;
816 kstat_named_t arcstat_memory_all_bytes;
817 kstat_named_t arcstat_memory_free_bytes;
818 kstat_named_t arcstat_memory_available_bytes;
819 kstat_named_t arcstat_no_grow;
820 kstat_named_t arcstat_tempreserve;
821 kstat_named_t arcstat_loaned_bytes;
822 kstat_named_t arcstat_prune;
823 /* Not updated directly; only synced in arc_kstat_update. */
824 kstat_named_t arcstat_meta_used;
825 kstat_named_t arcstat_meta_limit;
826 kstat_named_t arcstat_dnode_limit;
827 kstat_named_t arcstat_meta_max;
828 kstat_named_t arcstat_meta_min;
829 kstat_named_t arcstat_async_upgrade_sync;
830 kstat_named_t arcstat_demand_hit_predictive_prefetch;
831 kstat_named_t arcstat_demand_hit_prescient_prefetch;
834 static arc_stats_t arc_stats = {
835 { "hits", KSTAT_DATA_UINT64 },
836 { "misses", KSTAT_DATA_UINT64 },
837 { "demand_data_hits", KSTAT_DATA_UINT64 },
838 { "demand_data_misses", KSTAT_DATA_UINT64 },
839 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
840 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
841 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
842 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
843 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
844 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
845 { "mru_hits", KSTAT_DATA_UINT64 },
846 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
847 { "mfu_hits", KSTAT_DATA_UINT64 },
848 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
849 { "allocated", KSTAT_DATA_UINT64 },
850 { "deleted", KSTAT_DATA_UINT64 },
851 { "mutex_miss", KSTAT_DATA_UINT64 },
852 { "access_skip", KSTAT_DATA_UINT64 },
853 { "evict_skip", KSTAT_DATA_UINT64 },
854 { "evict_not_enough", KSTAT_DATA_UINT64 },
855 { "evict_l2_cached", KSTAT_DATA_UINT64 },
856 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
857 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
858 { "evict_l2_skip", KSTAT_DATA_UINT64 },
859 { "hash_elements", KSTAT_DATA_UINT64 },
860 { "hash_elements_max", KSTAT_DATA_UINT64 },
861 { "hash_collisions", KSTAT_DATA_UINT64 },
862 { "hash_chains", KSTAT_DATA_UINT64 },
863 { "hash_chain_max", KSTAT_DATA_UINT64 },
864 { "p", KSTAT_DATA_UINT64 },
865 { "c", KSTAT_DATA_UINT64 },
866 { "c_min", KSTAT_DATA_UINT64 },
867 { "c_max", KSTAT_DATA_UINT64 },
868 { "size", KSTAT_DATA_UINT64 },
869 { "compressed_size", KSTAT_DATA_UINT64 },
870 { "uncompressed_size", KSTAT_DATA_UINT64 },
871 { "overhead_size", KSTAT_DATA_UINT64 },
872 { "hdr_size", KSTAT_DATA_UINT64 },
873 { "data_size", KSTAT_DATA_UINT64 },
874 { "metadata_size", KSTAT_DATA_UINT64 },
875 { "dbuf_size", KSTAT_DATA_UINT64 },
876 { "dnode_size", KSTAT_DATA_UINT64 },
877 { "bonus_size", KSTAT_DATA_UINT64 },
878 { "anon_size", KSTAT_DATA_UINT64 },
879 { "anon_evictable_data", KSTAT_DATA_UINT64 },
880 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
881 { "mru_size", KSTAT_DATA_UINT64 },
882 { "mru_evictable_data", KSTAT_DATA_UINT64 },
883 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
884 { "mru_ghost_size", KSTAT_DATA_UINT64 },
885 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
886 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
887 { "mfu_size", KSTAT_DATA_UINT64 },
888 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
889 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
890 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
891 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
892 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
893 { "l2_hits", KSTAT_DATA_UINT64 },
894 { "l2_misses", KSTAT_DATA_UINT64 },
895 { "l2_feeds", KSTAT_DATA_UINT64 },
896 { "l2_rw_clash", KSTAT_DATA_UINT64 },
897 { "l2_read_bytes", KSTAT_DATA_UINT64 },
898 { "l2_write_bytes", KSTAT_DATA_UINT64 },
899 { "l2_writes_sent", KSTAT_DATA_UINT64 },
900 { "l2_writes_done", KSTAT_DATA_UINT64 },
901 { "l2_writes_error", KSTAT_DATA_UINT64 },
902 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
903 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
904 { "l2_evict_reading", KSTAT_DATA_UINT64 },
905 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
906 { "l2_free_on_write", KSTAT_DATA_UINT64 },
907 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
908 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
909 { "l2_io_error", KSTAT_DATA_UINT64 },
910 { "l2_size", KSTAT_DATA_UINT64 },
911 { "l2_asize", KSTAT_DATA_UINT64 },
912 { "l2_hdr_size", KSTAT_DATA_UINT64 },
913 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
914 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
915 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
916 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
917 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
918 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
919 { "l2_write_full", KSTAT_DATA_UINT64 },
920 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
921 { "l2_write_pios", KSTAT_DATA_UINT64 },
922 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
923 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
924 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
925 { "memory_throttle_count", KSTAT_DATA_UINT64 },
926 { "memory_direct_count", KSTAT_DATA_UINT64 },
927 { "memory_indirect_count", KSTAT_DATA_UINT64 },
928 { "memory_all_bytes", KSTAT_DATA_UINT64 },
929 { "memory_free_bytes", KSTAT_DATA_UINT64 },
930 { "memory_available_bytes", KSTAT_DATA_UINT64 },
931 { "arc_no_grow", KSTAT_DATA_UINT64 },
932 { "arc_tempreserve", KSTAT_DATA_UINT64 },
933 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
934 { "arc_prune", KSTAT_DATA_UINT64 },
935 { "arc_meta_used", KSTAT_DATA_UINT64 },
936 { "arc_meta_limit", KSTAT_DATA_UINT64 },
937 { "arc_dnode_limit", KSTAT_DATA_UINT64 },
938 { "arc_meta_max", KSTAT_DATA_UINT64 },
939 { "arc_meta_min", KSTAT_DATA_UINT64 },
940 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
941 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
942 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
945 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
947 #define ARCSTAT_INCR(stat, val) \
948 atomic_add_64(&arc_stats.stat.value.ui64, (val))
950 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
951 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
953 #define ARCSTAT_MAX(stat, val) { \
955 while ((val) > (m = arc_stats.stat.value.ui64) && \
956 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
960 #define ARCSTAT_MAXSTAT(stat) \
961 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
964 * We define a macro to allow ARC hits/misses to be easily broken down by
965 * two separate conditions, giving a total of four different subtypes for
966 * each of hits and misses (so eight statistics total).
968 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
971 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
973 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
977 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
979 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
984 static arc_state_t *arc_anon;
985 static arc_state_t *arc_mru;
986 static arc_state_t *arc_mru_ghost;
987 static arc_state_t *arc_mfu;
988 static arc_state_t *arc_mfu_ghost;
989 static arc_state_t *arc_l2c_only;
992 * There are several ARC variables that are critical to export as kstats --
993 * but we don't want to have to grovel around in the kstat whenever we wish to
994 * manipulate them. For these variables, we therefore define them to be in
995 * terms of the statistic variable. This assures that we are not introducing
996 * the possibility of inconsistency by having shadow copies of the variables,
997 * while still allowing the code to be readable.
999 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
1000 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
1001 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
1002 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
1003 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
1004 #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
1005 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
1006 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
1007 #define arc_dbuf_size ARCSTAT(arcstat_dbuf_size) /* dbuf metadata */
1008 #define arc_dnode_size ARCSTAT(arcstat_dnode_size) /* dnode metadata */
1009 #define arc_bonus_size ARCSTAT(arcstat_bonus_size) /* bonus buffer metadata */
1011 /* compressed size of entire arc */
1012 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
1013 /* uncompressed size of entire arc */
1014 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
1015 /* number of bytes in the arc from arc_buf_t's */
1016 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
1019 * There are also some ARC variables that we want to export, but that are
1020 * updated so often that having the canonical representation be the statistic
1021 * variable causes a performance bottleneck. We want to use aggsum_t's for these
1022 * instead, but still be able to export the kstat in the same way as before.
1023 * The solution is to always use the aggsum version, except in the kstat update
1027 aggsum_t arc_meta_used;
1028 aggsum_t astat_data_size;
1029 aggsum_t astat_metadata_size;
1030 aggsum_t astat_hdr_size;
1031 aggsum_t astat_bonus_size;
1032 aggsum_t astat_dnode_size;
1033 aggsum_t astat_dbuf_size;
1034 aggsum_t astat_l2_hdr_size;
1036 static list_t arc_prune_list;
1037 static kmutex_t arc_prune_mtx;
1038 static taskq_t *arc_prune_taskq;
1040 static int arc_no_grow; /* Don't try to grow cache size */
1041 static uint64_t arc_tempreserve;
1042 static uint64_t arc_loaned_bytes;
1044 typedef struct arc_callback arc_callback_t;
1046 struct arc_callback {
1048 arc_read_done_func_t *acb_done;
1050 boolean_t acb_compressed;
1051 zio_t *acb_zio_dummy;
1052 zio_t *acb_zio_head;
1053 arc_callback_t *acb_next;
1056 typedef struct arc_write_callback arc_write_callback_t;
1058 struct arc_write_callback {
1060 arc_write_done_func_t *awcb_ready;
1061 arc_write_done_func_t *awcb_children_ready;
1062 arc_write_done_func_t *awcb_physdone;
1063 arc_write_done_func_t *awcb_done;
1064 arc_buf_t *awcb_buf;
1068 * ARC buffers are separated into multiple structs as a memory saving measure:
1069 * - Common fields struct, always defined, and embedded within it:
1070 * - L2-only fields, always allocated but undefined when not in L2ARC
1071 * - L1-only fields, only allocated when in L1ARC
1073 * Buffer in L1 Buffer only in L2
1074 * +------------------------+ +------------------------+
1075 * | arc_buf_hdr_t | | arc_buf_hdr_t |
1079 * +------------------------+ +------------------------+
1080 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
1081 * | (undefined if L1-only) | | |
1082 * +------------------------+ +------------------------+
1083 * | l1arc_buf_hdr_t |
1088 * +------------------------+
1090 * Because it's possible for the L2ARC to become extremely large, we can wind
1091 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
1092 * is minimized by only allocating the fields necessary for an L1-cached buffer
1093 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
1094 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
1095 * words in pointers. arc_hdr_realloc() is used to switch a header between
1096 * these two allocation states.
1098 typedef struct l1arc_buf_hdr {
1099 kmutex_t b_freeze_lock;
1100 zio_cksum_t *b_freeze_cksum;
1103 * Used for debugging with kmem_flags - by allocating and freeing
1104 * b_thawed when the buffer is thawed, we get a record of the stack
1105 * trace that thawed it.
1112 /* for waiting on writes to complete */
1116 /* protected by arc state mutex */
1117 arc_state_t *b_state;
1118 multilist_node_t b_arc_node;
1120 /* updated atomically */
1121 clock_t b_arc_access;
1122 uint32_t b_mru_hits;
1123 uint32_t b_mru_ghost_hits;
1124 uint32_t b_mfu_hits;
1125 uint32_t b_mfu_ghost_hits;
1128 /* self protecting */
1129 refcount_t b_refcnt;
1131 arc_callback_t *b_acb;
1135 typedef struct l2arc_dev l2arc_dev_t;
1137 typedef struct l2arc_buf_hdr {
1138 /* protected by arc_buf_hdr mutex */
1139 l2arc_dev_t *b_dev; /* L2ARC device */
1140 uint64_t b_daddr; /* disk address, offset byte */
1143 list_node_t b_l2node;
1146 struct arc_buf_hdr {
1147 /* protected by hash lock */
1151 arc_buf_contents_t b_type;
1152 arc_buf_hdr_t *b_hash_next;
1153 arc_flags_t b_flags;
1156 * This field stores the size of the data buffer after
1157 * compression, and is set in the arc's zio completion handlers.
1158 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
1160 * While the block pointers can store up to 32MB in their psize
1161 * field, we can only store up to 32MB minus 512B. This is due
1162 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
1163 * a field of zeros represents 512B in the bp). We can't use a
1164 * bias of 1 since we need to reserve a psize of zero, here, to
1165 * represent holes and embedded blocks.
1167 * This isn't a problem in practice, since the maximum size of a
1168 * buffer is limited to 16MB, so we never need to store 32MB in
1169 * this field. Even in the upstream illumos code base, the
1170 * maximum size of a buffer is limited to 16MB.
1175 * This field stores the size of the data buffer before
1176 * compression, and cannot change once set. It is in units
1177 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
1179 uint16_t b_lsize; /* immutable */
1180 uint64_t b_spa; /* immutable */
1182 /* L2ARC fields. Undefined when not in L2ARC. */
1183 l2arc_buf_hdr_t b_l2hdr;
1184 /* L1ARC fields. Undefined when in l2arc_only state */
1185 l1arc_buf_hdr_t b_l1hdr;
1188 #if defined(__FreeBSD__) && defined(_KERNEL)
1190 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
1195 val = arc_meta_limit;
1196 err = sysctl_handle_64(oidp, &val, 0, req);
1197 if (err != 0 || req->newptr == NULL)
1200 if (val <= 0 || val > arc_c_max)
1203 arc_meta_limit = val;
1208 sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS)
1213 val = arc_no_grow_shift;
1214 err = sysctl_handle_32(oidp, &val, 0, req);
1215 if (err != 0 || req->newptr == NULL)
1218 if (val >= arc_shrink_shift)
1221 arc_no_grow_shift = val;
1226 sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
1232 err = sysctl_handle_64(oidp, &val, 0, req);
1233 if (err != 0 || req->newptr == NULL)
1236 if (zfs_arc_max == 0) {
1237 /* Loader tunable so blindly set */
1242 if (val < arc_abs_min || val > kmem_size())
1244 if (val < arc_c_min)
1246 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
1252 arc_p = (arc_c >> 1);
1254 if (zfs_arc_meta_limit == 0) {
1255 /* limit meta-data to 1/4 of the arc capacity */
1256 arc_meta_limit = arc_c_max / 4;
1259 /* if kmem_flags are set, lets try to use less memory */
1260 if (kmem_debugging())
1263 zfs_arc_max = arc_c;
1269 sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
1275 err = sysctl_handle_64(oidp, &val, 0, req);
1276 if (err != 0 || req->newptr == NULL)
1279 if (zfs_arc_min == 0) {
1280 /* Loader tunable so blindly set */
1285 if (val < arc_abs_min || val > arc_c_max)
1290 if (zfs_arc_meta_min == 0)
1291 arc_meta_min = arc_c_min / 2;
1293 if (arc_c < arc_c_min)
1296 zfs_arc_min = arc_c_min;
1302 #define GHOST_STATE(state) \
1303 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
1304 (state) == arc_l2c_only)
1306 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1307 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1308 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1309 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1310 #define HDR_PRESCIENT_PREFETCH(hdr) \
1311 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
1312 #define HDR_COMPRESSION_ENABLED(hdr) \
1313 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1315 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1316 #define HDR_L2_READING(hdr) \
1317 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1318 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1319 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1320 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1321 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1322 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1324 #define HDR_ISTYPE_METADATA(hdr) \
1325 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1326 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1328 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1329 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1331 /* For storing compression mode in b_flags */
1332 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1334 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1335 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1336 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1337 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1339 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1340 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1341 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1347 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1348 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1351 * Hash table routines
1354 #define HT_LOCK_PAD CACHE_LINE_SIZE
1359 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1363 #define BUF_LOCKS 256
1364 typedef struct buf_hash_table {
1366 arc_buf_hdr_t **ht_table;
1367 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1370 static buf_hash_table_t buf_hash_table;
1372 #define BUF_HASH_INDEX(spa, dva, birth) \
1373 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1374 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1375 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1376 #define HDR_LOCK(hdr) \
1377 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1379 uint64_t zfs_crc64_table[256];
1385 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1386 #define L2ARC_HEADROOM 2 /* num of writes */
1388 * If we discover during ARC scan any buffers to be compressed, we boost
1389 * our headroom for the next scanning cycle by this percentage multiple.
1391 #define L2ARC_HEADROOM_BOOST 200
1392 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1393 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1395 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1396 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1398 /* L2ARC Performance Tunables */
1399 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1400 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1401 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1402 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1403 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1404 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1405 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1406 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1407 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1409 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1410 &l2arc_write_max, 0, "max write size");
1411 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1412 &l2arc_write_boost, 0, "extra write during warmup");
1413 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1414 &l2arc_headroom, 0, "number of dev writes");
1415 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1416 &l2arc_feed_secs, 0, "interval seconds");
1417 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1418 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1420 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1421 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1422 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1423 &l2arc_feed_again, 0, "turbo warmup");
1424 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1425 &l2arc_norw, 0, "no reads during writes");
1427 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1428 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1429 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD,
1430 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1431 "size of anonymous state");
1432 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD,
1433 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1434 "size of anonymous state");
1436 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1437 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1438 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD,
1439 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1440 "size of metadata in mru state");
1441 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD,
1442 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1443 "size of data in mru state");
1445 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1446 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1447 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD,
1448 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1449 "size of metadata in mru ghost state");
1450 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD,
1451 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1452 "size of data in mru ghost state");
1454 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1455 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1456 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD,
1457 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1458 "size of metadata in mfu state");
1459 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD,
1460 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1461 "size of data in mfu state");
1463 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1464 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1465 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD,
1466 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1467 "size of metadata in mfu ghost state");
1468 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD,
1469 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1470 "size of data in mfu ghost state");
1472 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1473 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1475 SYSCTL_UINT(_vfs_zfs, OID_AUTO, arc_min_prefetch_ms, CTLFLAG_RW,
1476 &zfs_arc_min_prefetch_ms, 0, "Min life of prefetch block in ms");
1477 SYSCTL_UINT(_vfs_zfs, OID_AUTO, arc_min_prescient_prefetch_ms, CTLFLAG_RW,
1478 &zfs_arc_min_prescient_prefetch_ms, 0, "Min life of prescient prefetched block in ms");
1484 vdev_t *l2ad_vdev; /* vdev */
1485 spa_t *l2ad_spa; /* spa */
1486 uint64_t l2ad_hand; /* next write location */
1487 uint64_t l2ad_start; /* first addr on device */
1488 uint64_t l2ad_end; /* last addr on device */
1489 boolean_t l2ad_first; /* first sweep through */
1490 boolean_t l2ad_writing; /* currently writing */
1491 kmutex_t l2ad_mtx; /* lock for buffer list */
1492 list_t l2ad_buflist; /* buffer list */
1493 list_node_t l2ad_node; /* device list node */
1494 refcount_t l2ad_alloc; /* allocated bytes */
1497 static list_t L2ARC_dev_list; /* device list */
1498 static list_t *l2arc_dev_list; /* device list pointer */
1499 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1500 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1501 static list_t L2ARC_free_on_write; /* free after write buf list */
1502 static list_t *l2arc_free_on_write; /* free after write list ptr */
1503 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1504 static uint64_t l2arc_ndev; /* number of devices */
1506 typedef struct l2arc_read_callback {
1507 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1508 blkptr_t l2rcb_bp; /* original blkptr */
1509 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1510 int l2rcb_flags; /* original flags */
1511 abd_t *l2rcb_abd; /* temporary buffer */
1512 } l2arc_read_callback_t;
1514 typedef struct l2arc_write_callback {
1515 l2arc_dev_t *l2wcb_dev; /* device info */
1516 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1517 } l2arc_write_callback_t;
1519 typedef struct l2arc_data_free {
1520 /* protected by l2arc_free_on_write_mtx */
1523 arc_buf_contents_t l2df_type;
1524 list_node_t l2df_list_node;
1525 } l2arc_data_free_t;
1527 static kmutex_t l2arc_feed_thr_lock;
1528 static kcondvar_t l2arc_feed_thr_cv;
1529 static uint8_t l2arc_thread_exit;
1531 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1532 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1533 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1534 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1535 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1536 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1537 static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1538 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *);
1539 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1540 static boolean_t arc_is_overflowing();
1541 static void arc_buf_watch(arc_buf_t *);
1542 static void arc_prune_async(int64_t);
1544 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1545 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1546 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1547 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1549 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1550 static void l2arc_read_done(zio_t *);
1553 l2arc_trim(const arc_buf_hdr_t *hdr)
1555 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1557 ASSERT(HDR_HAS_L2HDR(hdr));
1558 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1560 if (HDR_GET_PSIZE(hdr) != 0) {
1561 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1562 HDR_GET_PSIZE(hdr), 0);
1567 * We use Cityhash for this. It's fast, and has good hash properties without
1568 * requiring any large static buffers.
1571 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1573 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
1576 #define HDR_EMPTY(hdr) \
1577 ((hdr)->b_dva.dva_word[0] == 0 && \
1578 (hdr)->b_dva.dva_word[1] == 0)
1580 #define HDR_EQUAL(spa, dva, birth, hdr) \
1581 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1582 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1583 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1586 buf_discard_identity(arc_buf_hdr_t *hdr)
1588 hdr->b_dva.dva_word[0] = 0;
1589 hdr->b_dva.dva_word[1] = 0;
1593 static arc_buf_hdr_t *
1594 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1596 const dva_t *dva = BP_IDENTITY(bp);
1597 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1598 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1599 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1602 mutex_enter(hash_lock);
1603 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1604 hdr = hdr->b_hash_next) {
1605 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1610 mutex_exit(hash_lock);
1616 * Insert an entry into the hash table. If there is already an element
1617 * equal to elem in the hash table, then the already existing element
1618 * will be returned and the new element will not be inserted.
1619 * Otherwise returns NULL.
1620 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1622 static arc_buf_hdr_t *
1623 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1625 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1626 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1627 arc_buf_hdr_t *fhdr;
1630 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1631 ASSERT(hdr->b_birth != 0);
1632 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1634 if (lockp != NULL) {
1636 mutex_enter(hash_lock);
1638 ASSERT(MUTEX_HELD(hash_lock));
1641 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1642 fhdr = fhdr->b_hash_next, i++) {
1643 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1647 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1648 buf_hash_table.ht_table[idx] = hdr;
1649 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1651 /* collect some hash table performance data */
1653 ARCSTAT_BUMP(arcstat_hash_collisions);
1655 ARCSTAT_BUMP(arcstat_hash_chains);
1657 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1660 ARCSTAT_BUMP(arcstat_hash_elements);
1661 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1667 buf_hash_remove(arc_buf_hdr_t *hdr)
1669 arc_buf_hdr_t *fhdr, **hdrp;
1670 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1672 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1673 ASSERT(HDR_IN_HASH_TABLE(hdr));
1675 hdrp = &buf_hash_table.ht_table[idx];
1676 while ((fhdr = *hdrp) != hdr) {
1677 ASSERT3P(fhdr, !=, NULL);
1678 hdrp = &fhdr->b_hash_next;
1680 *hdrp = hdr->b_hash_next;
1681 hdr->b_hash_next = NULL;
1682 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1684 /* collect some hash table performance data */
1685 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1687 if (buf_hash_table.ht_table[idx] &&
1688 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1689 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1693 * Global data structures and functions for the buf kmem cache.
1695 static kmem_cache_t *hdr_full_cache;
1696 static kmem_cache_t *hdr_l2only_cache;
1697 static kmem_cache_t *buf_cache;
1704 kmem_free(buf_hash_table.ht_table,
1705 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1706 for (i = 0; i < BUF_LOCKS; i++)
1707 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1708 kmem_cache_destroy(hdr_full_cache);
1709 kmem_cache_destroy(hdr_l2only_cache);
1710 kmem_cache_destroy(buf_cache);
1714 * Constructor callback - called when the cache is empty
1715 * and a new buf is requested.
1719 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1721 arc_buf_hdr_t *hdr = vbuf;
1723 bzero(hdr, HDR_FULL_SIZE);
1724 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1725 refcount_create(&hdr->b_l1hdr.b_refcnt);
1726 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1727 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1728 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1735 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1737 arc_buf_hdr_t *hdr = vbuf;
1739 bzero(hdr, HDR_L2ONLY_SIZE);
1740 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1747 buf_cons(void *vbuf, void *unused, int kmflag)
1749 arc_buf_t *buf = vbuf;
1751 bzero(buf, sizeof (arc_buf_t));
1752 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1753 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1759 * Destructor callback - called when a cached buf is
1760 * no longer required.
1764 hdr_full_dest(void *vbuf, void *unused)
1766 arc_buf_hdr_t *hdr = vbuf;
1768 ASSERT(HDR_EMPTY(hdr));
1769 cv_destroy(&hdr->b_l1hdr.b_cv);
1770 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1771 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1772 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1773 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1778 hdr_l2only_dest(void *vbuf, void *unused)
1780 arc_buf_hdr_t *hdr = vbuf;
1782 ASSERT(HDR_EMPTY(hdr));
1783 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1788 buf_dest(void *vbuf, void *unused)
1790 arc_buf_t *buf = vbuf;
1792 mutex_destroy(&buf->b_evict_lock);
1793 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1797 * Reclaim callback -- invoked when memory is low.
1801 hdr_recl(void *unused)
1803 dprintf("hdr_recl called\n");
1805 * umem calls the reclaim func when we destroy the buf cache,
1806 * which is after we do arc_fini().
1809 cv_signal(&arc_reclaim_thread_cv);
1816 uint64_t hsize = 1ULL << 12;
1820 * The hash table is big enough to fill all of physical memory
1821 * with an average block size of zfs_arc_average_blocksize (default 8K).
1822 * By default, the table will take up
1823 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1825 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1828 buf_hash_table.ht_mask = hsize - 1;
1829 buf_hash_table.ht_table =
1830 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1831 if (buf_hash_table.ht_table == NULL) {
1832 ASSERT(hsize > (1ULL << 8));
1837 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1838 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1839 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1840 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1842 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1843 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1845 for (i = 0; i < 256; i++)
1846 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1847 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1849 for (i = 0; i < BUF_LOCKS; i++) {
1850 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1851 NULL, MUTEX_DEFAULT, NULL);
1856 * This is the size that the buf occupies in memory. If the buf is compressed,
1857 * it will correspond to the compressed size. You should use this method of
1858 * getting the buf size unless you explicitly need the logical size.
1861 arc_buf_size(arc_buf_t *buf)
1863 return (ARC_BUF_COMPRESSED(buf) ?
1864 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1868 arc_buf_lsize(arc_buf_t *buf)
1870 return (HDR_GET_LSIZE(buf->b_hdr));
1874 arc_get_compression(arc_buf_t *buf)
1876 return (ARC_BUF_COMPRESSED(buf) ?
1877 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1880 #define ARC_MINTIME (hz>>4) /* 62 ms */
1882 static inline boolean_t
1883 arc_buf_is_shared(arc_buf_t *buf)
1885 boolean_t shared = (buf->b_data != NULL &&
1886 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1887 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1888 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1889 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1890 IMPLY(shared, ARC_BUF_SHARED(buf));
1891 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1894 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1895 * already being shared" requirement prevents us from doing that.
1902 * Free the checksum associated with this header. If there is no checksum, this
1906 arc_cksum_free(arc_buf_hdr_t *hdr)
1908 ASSERT(HDR_HAS_L1HDR(hdr));
1909 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1910 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1911 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1912 hdr->b_l1hdr.b_freeze_cksum = NULL;
1914 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1918 * Return true iff at least one of the bufs on hdr is not compressed.
1921 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1923 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1924 if (!ARC_BUF_COMPRESSED(b)) {
1932 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1933 * matches the checksum that is stored in the hdr. If there is no checksum,
1934 * or if the buf is compressed, this is a no-op.
1937 arc_cksum_verify(arc_buf_t *buf)
1939 arc_buf_hdr_t *hdr = buf->b_hdr;
1942 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1945 if (ARC_BUF_COMPRESSED(buf)) {
1946 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1947 arc_hdr_has_uncompressed_buf(hdr));
1951 ASSERT(HDR_HAS_L1HDR(hdr));
1953 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1954 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1955 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1959 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1960 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1961 panic("buffer modified while frozen!");
1962 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1966 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1968 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1969 boolean_t valid_cksum;
1971 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1972 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1975 * We rely on the blkptr's checksum to determine if the block
1976 * is valid or not. When compressed arc is enabled, the l2arc
1977 * writes the block to the l2arc just as it appears in the pool.
1978 * This allows us to use the blkptr's checksum to validate the
1979 * data that we just read off of the l2arc without having to store
1980 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1981 * arc is disabled, then the data written to the l2arc is always
1982 * uncompressed and won't match the block as it exists in the main
1983 * pool. When this is the case, we must first compress it if it is
1984 * compressed on the main pool before we can validate the checksum.
1986 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1987 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1988 uint64_t lsize = HDR_GET_LSIZE(hdr);
1991 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
1992 csize = zio_compress_data(compress, zio->io_abd,
1993 abd_to_buf(cdata), lsize);
1995 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1996 if (csize < HDR_GET_PSIZE(hdr)) {
1998 * Compressed blocks are always a multiple of the
1999 * smallest ashift in the pool. Ideally, we would
2000 * like to round up the csize to the next
2001 * spa_min_ashift but that value may have changed
2002 * since the block was last written. Instead,
2003 * we rely on the fact that the hdr's psize
2004 * was set to the psize of the block when it was
2005 * last written. We set the csize to that value
2006 * and zero out any part that should not contain
2009 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
2010 csize = HDR_GET_PSIZE(hdr);
2012 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
2016 * Block pointers always store the checksum for the logical data.
2017 * If the block pointer has the gang bit set, then the checksum
2018 * it represents is for the reconstituted data and not for an
2019 * individual gang member. The zio pipeline, however, must be able to
2020 * determine the checksum of each of the gang constituents so it
2021 * treats the checksum comparison differently than what we need
2022 * for l2arc blocks. This prevents us from using the
2023 * zio_checksum_error() interface directly. Instead we must call the
2024 * zio_checksum_error_impl() so that we can ensure the checksum is
2025 * generated using the correct checksum algorithm and accounts for the
2026 * logical I/O size and not just a gang fragment.
2028 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
2029 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
2030 zio->io_offset, NULL) == 0);
2031 zio_pop_transforms(zio);
2032 return (valid_cksum);
2036 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
2037 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
2038 * isn't modified later on. If buf is compressed or there is already a checksum
2039 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
2042 arc_cksum_compute(arc_buf_t *buf)
2044 arc_buf_hdr_t *hdr = buf->b_hdr;
2046 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2049 ASSERT(HDR_HAS_L1HDR(hdr));
2051 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
2052 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
2053 ASSERT(arc_hdr_has_uncompressed_buf(hdr));
2054 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2056 } else if (ARC_BUF_COMPRESSED(buf)) {
2057 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2061 ASSERT(!ARC_BUF_COMPRESSED(buf));
2062 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
2064 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
2065 hdr->b_l1hdr.b_freeze_cksum);
2066 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2074 typedef struct procctl {
2082 arc_buf_unwatch(arc_buf_t *buf)
2089 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2090 ctl.prwatch.pr_size = 0;
2091 ctl.prwatch.pr_wflags = 0;
2092 result = write(arc_procfd, &ctl, sizeof (ctl));
2093 ASSERT3U(result, ==, sizeof (ctl));
2100 arc_buf_watch(arc_buf_t *buf)
2107 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2108 ctl.prwatch.pr_size = arc_buf_size(buf);
2109 ctl.prwatch.pr_wflags = WA_WRITE;
2110 result = write(arc_procfd, &ctl, sizeof (ctl));
2111 ASSERT3U(result, ==, sizeof (ctl));
2115 #endif /* illumos */
2117 static arc_buf_contents_t
2118 arc_buf_type(arc_buf_hdr_t *hdr)
2120 arc_buf_contents_t type;
2121 if (HDR_ISTYPE_METADATA(hdr)) {
2122 type = ARC_BUFC_METADATA;
2124 type = ARC_BUFC_DATA;
2126 VERIFY3U(hdr->b_type, ==, type);
2131 arc_is_metadata(arc_buf_t *buf)
2133 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
2137 arc_bufc_to_flags(arc_buf_contents_t type)
2141 /* metadata field is 0 if buffer contains normal data */
2143 case ARC_BUFC_METADATA:
2144 return (ARC_FLAG_BUFC_METADATA);
2148 panic("undefined ARC buffer type!");
2149 return ((uint32_t)-1);
2153 arc_buf_thaw(arc_buf_t *buf)
2155 arc_buf_hdr_t *hdr = buf->b_hdr;
2157 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2158 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2160 arc_cksum_verify(buf);
2163 * Compressed buffers do not manipulate the b_freeze_cksum or
2164 * allocate b_thawed.
2166 if (ARC_BUF_COMPRESSED(buf)) {
2167 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2168 arc_hdr_has_uncompressed_buf(hdr));
2172 ASSERT(HDR_HAS_L1HDR(hdr));
2173 arc_cksum_free(hdr);
2175 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
2177 if (zfs_flags & ZFS_DEBUG_MODIFY) {
2178 if (hdr->b_l1hdr.b_thawed != NULL)
2179 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2180 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
2184 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2187 arc_buf_unwatch(buf);
2192 arc_buf_freeze(arc_buf_t *buf)
2194 arc_buf_hdr_t *hdr = buf->b_hdr;
2195 kmutex_t *hash_lock;
2197 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2200 if (ARC_BUF_COMPRESSED(buf)) {
2201 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2202 arc_hdr_has_uncompressed_buf(hdr));
2206 hash_lock = HDR_LOCK(hdr);
2207 mutex_enter(hash_lock);
2209 ASSERT(HDR_HAS_L1HDR(hdr));
2210 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
2211 hdr->b_l1hdr.b_state == arc_anon);
2212 arc_cksum_compute(buf);
2213 mutex_exit(hash_lock);
2217 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
2218 * the following functions should be used to ensure that the flags are
2219 * updated in a thread-safe way. When manipulating the flags either
2220 * the hash_lock must be held or the hdr must be undiscoverable. This
2221 * ensures that we're not racing with any other threads when updating
2225 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2227 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2228 hdr->b_flags |= flags;
2232 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2234 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2235 hdr->b_flags &= ~flags;
2239 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
2240 * done in a special way since we have to clear and set bits
2241 * at the same time. Consumers that wish to set the compression bits
2242 * must use this function to ensure that the flags are updated in
2243 * thread-safe manner.
2246 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
2248 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2251 * Holes and embedded blocks will always have a psize = 0 so
2252 * we ignore the compression of the blkptr and set the
2253 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
2254 * Holes and embedded blocks remain anonymous so we don't
2255 * want to uncompress them. Mark them as uncompressed.
2257 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
2258 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2259 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
2260 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
2261 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2263 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2264 HDR_SET_COMPRESS(hdr, cmp);
2265 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
2266 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
2271 * Looks for another buf on the same hdr which has the data decompressed, copies
2272 * from it, and returns true. If no such buf exists, returns false.
2275 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
2277 arc_buf_hdr_t *hdr = buf->b_hdr;
2278 boolean_t copied = B_FALSE;
2280 ASSERT(HDR_HAS_L1HDR(hdr));
2281 ASSERT3P(buf->b_data, !=, NULL);
2282 ASSERT(!ARC_BUF_COMPRESSED(buf));
2284 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
2285 from = from->b_next) {
2286 /* can't use our own data buffer */
2291 if (!ARC_BUF_COMPRESSED(from)) {
2292 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
2299 * There were no decompressed bufs, so there should not be a
2300 * checksum on the hdr either.
2302 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
2308 * Given a buf that has a data buffer attached to it, this function will
2309 * efficiently fill the buf with data of the specified compression setting from
2310 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2311 * are already sharing a data buf, no copy is performed.
2313 * If the buf is marked as compressed but uncompressed data was requested, this
2314 * will allocate a new data buffer for the buf, remove that flag, and fill the
2315 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2316 * uncompressed data, and (since we haven't added support for it yet) if you
2317 * want compressed data your buf must already be marked as compressed and have
2318 * the correct-sized data buffer.
2321 arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
2323 arc_buf_hdr_t *hdr = buf->b_hdr;
2324 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
2325 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2327 ASSERT3P(buf->b_data, !=, NULL);
2328 IMPLY(compressed, hdr_compressed);
2329 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2331 if (hdr_compressed == compressed) {
2332 if (!arc_buf_is_shared(buf)) {
2333 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2337 ASSERT(hdr_compressed);
2338 ASSERT(!compressed);
2339 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2342 * If the buf is sharing its data with the hdr, unlink it and
2343 * allocate a new data buffer for the buf.
2345 if (arc_buf_is_shared(buf)) {
2346 ASSERT(ARC_BUF_COMPRESSED(buf));
2348 /* We need to give the buf it's own b_data */
2349 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2351 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2352 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2354 /* Previously overhead was 0; just add new overhead */
2355 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2356 } else if (ARC_BUF_COMPRESSED(buf)) {
2357 /* We need to reallocate the buf's b_data */
2358 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2361 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2363 /* We increased the size of b_data; update overhead */
2364 ARCSTAT_INCR(arcstat_overhead_size,
2365 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2369 * Regardless of the buf's previous compression settings, it
2370 * should not be compressed at the end of this function.
2372 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2375 * Try copying the data from another buf which already has a
2376 * decompressed version. If that's not possible, it's time to
2377 * bite the bullet and decompress the data from the hdr.
2379 if (arc_buf_try_copy_decompressed_data(buf)) {
2380 /* Skip byteswapping and checksumming (already done) */
2381 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2384 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2385 hdr->b_l1hdr.b_pabd, buf->b_data,
2386 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2389 * Absent hardware errors or software bugs, this should
2390 * be impossible, but log it anyway so we can debug it.
2394 "hdr %p, compress %d, psize %d, lsize %d",
2395 hdr, HDR_GET_COMPRESS(hdr),
2396 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2397 return (SET_ERROR(EIO));
2402 /* Byteswap the buf's data if necessary */
2403 if (bswap != DMU_BSWAP_NUMFUNCS) {
2404 ASSERT(!HDR_SHARED_DATA(hdr));
2405 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2406 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2409 /* Compute the hdr's checksum if necessary */
2410 arc_cksum_compute(buf);
2416 arc_decompress(arc_buf_t *buf)
2418 return (arc_buf_fill(buf, B_FALSE));
2422 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
2425 arc_hdr_size(arc_buf_hdr_t *hdr)
2429 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2430 HDR_GET_PSIZE(hdr) > 0) {
2431 size = HDR_GET_PSIZE(hdr);
2433 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2434 size = HDR_GET_LSIZE(hdr);
2440 * Increment the amount of evictable space in the arc_state_t's refcount.
2441 * We account for the space used by the hdr and the arc buf individually
2442 * so that we can add and remove them from the refcount individually.
2445 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2447 arc_buf_contents_t type = arc_buf_type(hdr);
2449 ASSERT(HDR_HAS_L1HDR(hdr));
2451 if (GHOST_STATE(state)) {
2452 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2453 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2454 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2455 (void) refcount_add_many(&state->arcs_esize[type],
2456 HDR_GET_LSIZE(hdr), hdr);
2460 ASSERT(!GHOST_STATE(state));
2461 if (hdr->b_l1hdr.b_pabd != NULL) {
2462 (void) refcount_add_many(&state->arcs_esize[type],
2463 arc_hdr_size(hdr), hdr);
2465 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2466 buf = buf->b_next) {
2467 if (arc_buf_is_shared(buf))
2469 (void) refcount_add_many(&state->arcs_esize[type],
2470 arc_buf_size(buf), buf);
2475 * Decrement the amount of evictable space in the arc_state_t's refcount.
2476 * We account for the space used by the hdr and the arc buf individually
2477 * so that we can add and remove them from the refcount individually.
2480 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2482 arc_buf_contents_t type = arc_buf_type(hdr);
2484 ASSERT(HDR_HAS_L1HDR(hdr));
2486 if (GHOST_STATE(state)) {
2487 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2488 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2489 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2490 (void) refcount_remove_many(&state->arcs_esize[type],
2491 HDR_GET_LSIZE(hdr), hdr);
2495 ASSERT(!GHOST_STATE(state));
2496 if (hdr->b_l1hdr.b_pabd != NULL) {
2497 (void) refcount_remove_many(&state->arcs_esize[type],
2498 arc_hdr_size(hdr), hdr);
2500 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2501 buf = buf->b_next) {
2502 if (arc_buf_is_shared(buf))
2504 (void) refcount_remove_many(&state->arcs_esize[type],
2505 arc_buf_size(buf), buf);
2510 * Add a reference to this hdr indicating that someone is actively
2511 * referencing that memory. When the refcount transitions from 0 to 1,
2512 * we remove it from the respective arc_state_t list to indicate that
2513 * it is not evictable.
2516 add_reference(arc_buf_hdr_t *hdr, void *tag)
2518 ASSERT(HDR_HAS_L1HDR(hdr));
2519 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2520 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2521 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2522 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2525 arc_state_t *state = hdr->b_l1hdr.b_state;
2527 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2528 (state != arc_anon)) {
2529 /* We don't use the L2-only state list. */
2530 if (state != arc_l2c_only) {
2531 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2533 arc_evictable_space_decrement(hdr, state);
2535 /* remove the prefetch flag if we get a reference */
2536 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2541 * Remove a reference from this hdr. When the reference transitions from
2542 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2543 * list making it eligible for eviction.
2546 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2549 arc_state_t *state = hdr->b_l1hdr.b_state;
2551 ASSERT(HDR_HAS_L1HDR(hdr));
2552 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2553 ASSERT(!GHOST_STATE(state));
2556 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2557 * check to prevent usage of the arc_l2c_only list.
2559 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2560 (state != arc_anon)) {
2561 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2562 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2563 arc_evictable_space_increment(hdr, state);
2569 * Returns detailed information about a specific arc buffer. When the
2570 * state_index argument is set the function will calculate the arc header
2571 * list position for its arc state. Since this requires a linear traversal
2572 * callers are strongly encourage not to do this. However, it can be helpful
2573 * for targeted analysis so the functionality is provided.
2576 arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
2578 arc_buf_hdr_t *hdr = ab->b_hdr;
2579 l1arc_buf_hdr_t *l1hdr = NULL;
2580 l2arc_buf_hdr_t *l2hdr = NULL;
2581 arc_state_t *state = NULL;
2583 memset(abi, 0, sizeof (arc_buf_info_t));
2588 abi->abi_flags = hdr->b_flags;
2590 if (HDR_HAS_L1HDR(hdr)) {
2591 l1hdr = &hdr->b_l1hdr;
2592 state = l1hdr->b_state;
2594 if (HDR_HAS_L2HDR(hdr))
2595 l2hdr = &hdr->b_l2hdr;
2598 abi->abi_bufcnt = l1hdr->b_bufcnt;
2599 abi->abi_access = l1hdr->b_arc_access;
2600 abi->abi_mru_hits = l1hdr->b_mru_hits;
2601 abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits;
2602 abi->abi_mfu_hits = l1hdr->b_mfu_hits;
2603 abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits;
2604 abi->abi_holds = refcount_count(&l1hdr->b_refcnt);
2608 abi->abi_l2arc_dattr = l2hdr->b_daddr;
2609 abi->abi_l2arc_hits = l2hdr->b_hits;
2612 abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
2613 abi->abi_state_contents = arc_buf_type(hdr);
2614 abi->abi_size = arc_hdr_size(hdr);
2618 * Move the supplied buffer to the indicated state. The hash lock
2619 * for the buffer must be held by the caller.
2622 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2623 kmutex_t *hash_lock)
2625 arc_state_t *old_state;
2628 boolean_t update_old, update_new;
2629 arc_buf_contents_t buftype = arc_buf_type(hdr);
2632 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2633 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2634 * L1 hdr doesn't always exist when we change state to arc_anon before
2635 * destroying a header, in which case reallocating to add the L1 hdr is
2638 if (HDR_HAS_L1HDR(hdr)) {
2639 old_state = hdr->b_l1hdr.b_state;
2640 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2641 bufcnt = hdr->b_l1hdr.b_bufcnt;
2642 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL);
2644 old_state = arc_l2c_only;
2647 update_old = B_FALSE;
2649 update_new = update_old;
2651 ASSERT(MUTEX_HELD(hash_lock));
2652 ASSERT3P(new_state, !=, old_state);
2653 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2654 ASSERT(old_state != arc_anon || bufcnt <= 1);
2657 * If this buffer is evictable, transfer it from the
2658 * old state list to the new state list.
2661 if (old_state != arc_anon && old_state != arc_l2c_only) {
2662 ASSERT(HDR_HAS_L1HDR(hdr));
2663 multilist_remove(old_state->arcs_list[buftype], hdr);
2665 if (GHOST_STATE(old_state)) {
2667 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2668 update_old = B_TRUE;
2670 arc_evictable_space_decrement(hdr, old_state);
2672 if (new_state != arc_anon && new_state != arc_l2c_only) {
2675 * An L1 header always exists here, since if we're
2676 * moving to some L1-cached state (i.e. not l2c_only or
2677 * anonymous), we realloc the header to add an L1hdr
2680 ASSERT(HDR_HAS_L1HDR(hdr));
2681 multilist_insert(new_state->arcs_list[buftype], hdr);
2683 if (GHOST_STATE(new_state)) {
2685 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2686 update_new = B_TRUE;
2688 arc_evictable_space_increment(hdr, new_state);
2692 ASSERT(!HDR_EMPTY(hdr));
2693 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2694 buf_hash_remove(hdr);
2696 /* adjust state sizes (ignore arc_l2c_only) */
2698 if (update_new && new_state != arc_l2c_only) {
2699 ASSERT(HDR_HAS_L1HDR(hdr));
2700 if (GHOST_STATE(new_state)) {
2704 * When moving a header to a ghost state, we first
2705 * remove all arc buffers. Thus, we'll have a
2706 * bufcnt of zero, and no arc buffer to use for
2707 * the reference. As a result, we use the arc
2708 * header pointer for the reference.
2710 (void) refcount_add_many(&new_state->arcs_size,
2711 HDR_GET_LSIZE(hdr), hdr);
2712 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2714 uint32_t buffers = 0;
2717 * Each individual buffer holds a unique reference,
2718 * thus we must remove each of these references one
2721 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2722 buf = buf->b_next) {
2723 ASSERT3U(bufcnt, !=, 0);
2727 * When the arc_buf_t is sharing the data
2728 * block with the hdr, the owner of the
2729 * reference belongs to the hdr. Only
2730 * add to the refcount if the arc_buf_t is
2733 if (arc_buf_is_shared(buf))
2736 (void) refcount_add_many(&new_state->arcs_size,
2737 arc_buf_size(buf), buf);
2739 ASSERT3U(bufcnt, ==, buffers);
2741 if (hdr->b_l1hdr.b_pabd != NULL) {
2742 (void) refcount_add_many(&new_state->arcs_size,
2743 arc_hdr_size(hdr), hdr);
2745 ASSERT(GHOST_STATE(old_state));
2750 if (update_old && old_state != arc_l2c_only) {
2751 ASSERT(HDR_HAS_L1HDR(hdr));
2752 if (GHOST_STATE(old_state)) {
2754 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2757 * When moving a header off of a ghost state,
2758 * the header will not contain any arc buffers.
2759 * We use the arc header pointer for the reference
2760 * which is exactly what we did when we put the
2761 * header on the ghost state.
2764 (void) refcount_remove_many(&old_state->arcs_size,
2765 HDR_GET_LSIZE(hdr), hdr);
2767 uint32_t buffers = 0;
2770 * Each individual buffer holds a unique reference,
2771 * thus we must remove each of these references one
2774 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2775 buf = buf->b_next) {
2776 ASSERT3U(bufcnt, !=, 0);
2780 * When the arc_buf_t is sharing the data
2781 * block with the hdr, the owner of the
2782 * reference belongs to the hdr. Only
2783 * add to the refcount if the arc_buf_t is
2786 if (arc_buf_is_shared(buf))
2789 (void) refcount_remove_many(
2790 &old_state->arcs_size, arc_buf_size(buf),
2793 ASSERT3U(bufcnt, ==, buffers);
2794 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2795 (void) refcount_remove_many(
2796 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2800 if (HDR_HAS_L1HDR(hdr))
2801 hdr->b_l1hdr.b_state = new_state;
2804 * L2 headers should never be on the L2 state list since they don't
2805 * have L1 headers allocated.
2807 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2808 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2812 arc_space_consume(uint64_t space, arc_space_type_t type)
2814 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2817 case ARC_SPACE_DATA:
2818 aggsum_add(&astat_data_size, space);
2820 case ARC_SPACE_META:
2821 aggsum_add(&astat_metadata_size, space);
2823 case ARC_SPACE_BONUS:
2824 aggsum_add(&astat_bonus_size, space);
2826 case ARC_SPACE_DNODE:
2827 aggsum_add(&astat_dnode_size, space);
2829 case ARC_SPACE_DBUF:
2830 aggsum_add(&astat_dbuf_size, space);
2832 case ARC_SPACE_HDRS:
2833 aggsum_add(&astat_hdr_size, space);
2835 case ARC_SPACE_L2HDRS:
2836 aggsum_add(&astat_l2_hdr_size, space);
2840 if (type != ARC_SPACE_DATA)
2841 aggsum_add(&arc_meta_used, space);
2843 aggsum_add(&arc_size, space);
2847 arc_space_return(uint64_t space, arc_space_type_t type)
2849 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2852 case ARC_SPACE_DATA:
2853 aggsum_add(&astat_data_size, -space);
2855 case ARC_SPACE_META:
2856 aggsum_add(&astat_metadata_size, -space);
2858 case ARC_SPACE_BONUS:
2859 aggsum_add(&astat_bonus_size, -space);
2861 case ARC_SPACE_DNODE:
2862 aggsum_add(&astat_dnode_size, -space);
2864 case ARC_SPACE_DBUF:
2865 aggsum_add(&astat_dbuf_size, -space);
2867 case ARC_SPACE_HDRS:
2868 aggsum_add(&astat_hdr_size, -space);
2870 case ARC_SPACE_L2HDRS:
2871 aggsum_add(&astat_l2_hdr_size, -space);
2875 if (type != ARC_SPACE_DATA) {
2876 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2878 * We use the upper bound here rather than the precise value
2879 * because the arc_meta_max value doesn't need to be
2880 * precise. It's only consumed by humans via arcstats.
2882 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2883 arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2884 aggsum_add(&arc_meta_used, -space);
2887 ASSERT(aggsum_compare(&arc_size, space) >= 0);
2888 aggsum_add(&arc_size, -space);
2892 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2893 * with the hdr's b_pabd.
2896 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2899 * The criteria for sharing a hdr's data are:
2900 * 1. the hdr's compression matches the buf's compression
2901 * 2. the hdr doesn't need to be byteswapped
2902 * 3. the hdr isn't already being shared
2903 * 4. the buf is either compressed or it is the last buf in the hdr list
2905 * Criterion #4 maintains the invariant that shared uncompressed
2906 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2907 * might ask, "if a compressed buf is allocated first, won't that be the
2908 * last thing in the list?", but in that case it's impossible to create
2909 * a shared uncompressed buf anyway (because the hdr must be compressed
2910 * to have the compressed buf). You might also think that #3 is
2911 * sufficient to make this guarantee, however it's possible
2912 * (specifically in the rare L2ARC write race mentioned in
2913 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2914 * is sharable, but wasn't at the time of its allocation. Rather than
2915 * allow a new shared uncompressed buf to be created and then shuffle
2916 * the list around to make it the last element, this simply disallows
2917 * sharing if the new buf isn't the first to be added.
2919 ASSERT3P(buf->b_hdr, ==, hdr);
2920 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF;
2921 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2922 return (buf_compressed == hdr_compressed &&
2923 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2924 !HDR_SHARED_DATA(hdr) &&
2925 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2929 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2930 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2931 * copy was made successfully, or an error code otherwise.
2934 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2935 boolean_t fill, arc_buf_t **ret)
2939 ASSERT(HDR_HAS_L1HDR(hdr));
2940 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2941 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2942 hdr->b_type == ARC_BUFC_METADATA);
2943 ASSERT3P(ret, !=, NULL);
2944 ASSERT3P(*ret, ==, NULL);
2946 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2949 buf->b_next = hdr->b_l1hdr.b_buf;
2952 add_reference(hdr, tag);
2955 * We're about to change the hdr's b_flags. We must either
2956 * hold the hash_lock or be undiscoverable.
2958 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2961 * Only honor requests for compressed bufs if the hdr is actually
2964 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
2965 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2968 * If the hdr's data can be shared then we share the data buffer and
2969 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2970 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2971 * buffer to store the buf's data.
2973 * There are two additional restrictions here because we're sharing
2974 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2975 * actively involved in an L2ARC write, because if this buf is used by
2976 * an arc_write() then the hdr's data buffer will be released when the
2977 * write completes, even though the L2ARC write might still be using it.
2978 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2979 * need to be ABD-aware.
2981 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2982 abd_is_linear(hdr->b_l1hdr.b_pabd);
2984 /* Set up b_data and sharing */
2986 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2987 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2988 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2991 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2992 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2994 VERIFY3P(buf->b_data, !=, NULL);
2996 hdr->b_l1hdr.b_buf = buf;
2997 hdr->b_l1hdr.b_bufcnt += 1;
3000 * If the user wants the data from the hdr, we need to either copy or
3001 * decompress the data.
3004 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
3010 static char *arc_onloan_tag = "onloan";
3013 arc_loaned_bytes_update(int64_t delta)
3015 atomic_add_64(&arc_loaned_bytes, delta);
3017 /* assert that it did not wrap around */
3018 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
3022 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
3023 * flight data by arc_tempreserve_space() until they are "returned". Loaned
3024 * buffers must be returned to the arc before they can be used by the DMU or
3028 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
3030 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
3031 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
3033 arc_loaned_bytes_update(arc_buf_size(buf));
3039 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
3040 enum zio_compress compression_type)
3042 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
3043 psize, lsize, compression_type);
3045 arc_loaned_bytes_update(arc_buf_size(buf));
3052 * Return a loaned arc buffer to the arc.
3055 arc_return_buf(arc_buf_t *buf, void *tag)
3057 arc_buf_hdr_t *hdr = buf->b_hdr;
3059 ASSERT3P(buf->b_data, !=, NULL);
3060 ASSERT(HDR_HAS_L1HDR(hdr));
3061 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
3062 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
3064 arc_loaned_bytes_update(-arc_buf_size(buf));
3067 /* Detach an arc_buf from a dbuf (tag) */
3069 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
3071 arc_buf_hdr_t *hdr = buf->b_hdr;
3073 ASSERT3P(buf->b_data, !=, NULL);
3074 ASSERT(HDR_HAS_L1HDR(hdr));
3075 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
3076 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
3078 arc_loaned_bytes_update(arc_buf_size(buf));
3082 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
3084 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
3087 df->l2df_size = size;
3088 df->l2df_type = type;
3089 mutex_enter(&l2arc_free_on_write_mtx);
3090 list_insert_head(l2arc_free_on_write, df);
3091 mutex_exit(&l2arc_free_on_write_mtx);
3095 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
3097 arc_state_t *state = hdr->b_l1hdr.b_state;
3098 arc_buf_contents_t type = arc_buf_type(hdr);
3099 uint64_t size = arc_hdr_size(hdr);
3101 /* protected by hash lock, if in the hash table */
3102 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3103 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3104 ASSERT(state != arc_anon && state != arc_l2c_only);
3106 (void) refcount_remove_many(&state->arcs_esize[type],
3109 (void) refcount_remove_many(&state->arcs_size, size, hdr);
3110 if (type == ARC_BUFC_METADATA) {
3111 arc_space_return(size, ARC_SPACE_META);
3113 ASSERT(type == ARC_BUFC_DATA);
3114 arc_space_return(size, ARC_SPACE_DATA);
3117 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
3121 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
3122 * data buffer, we transfer the refcount ownership to the hdr and update
3123 * the appropriate kstats.
3126 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3128 arc_state_t *state = hdr->b_l1hdr.b_state;
3130 ASSERT(arc_can_share(hdr, buf));
3131 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3132 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3135 * Start sharing the data buffer. We transfer the
3136 * refcount ownership to the hdr since it always owns
3137 * the refcount whenever an arc_buf_t is shared.
3139 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
3140 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
3141 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
3142 HDR_ISTYPE_METADATA(hdr));
3143 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
3144 buf->b_flags |= ARC_BUF_FLAG_SHARED;
3147 * Since we've transferred ownership to the hdr we need
3148 * to increment its compressed and uncompressed kstats and
3149 * decrement the overhead size.
3151 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3152 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3153 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
3157 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3159 arc_state_t *state = hdr->b_l1hdr.b_state;
3161 ASSERT(arc_buf_is_shared(buf));
3162 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3163 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3166 * We are no longer sharing this buffer so we need
3167 * to transfer its ownership to the rightful owner.
3169 refcount_transfer_ownership(&state->arcs_size, hdr, buf);
3170 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3171 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3172 abd_put(hdr->b_l1hdr.b_pabd);
3173 hdr->b_l1hdr.b_pabd = NULL;
3174 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3177 * Since the buffer is no longer shared between
3178 * the arc buf and the hdr, count it as overhead.
3180 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3181 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3182 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3186 * Remove an arc_buf_t from the hdr's buf list and return the last
3187 * arc_buf_t on the list. If no buffers remain on the list then return
3191 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3193 ASSERT(HDR_HAS_L1HDR(hdr));
3194 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3196 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3197 arc_buf_t *lastbuf = NULL;
3200 * Remove the buf from the hdr list and locate the last
3201 * remaining buffer on the list.
3203 while (*bufp != NULL) {
3205 *bufp = buf->b_next;
3208 * If we've removed a buffer in the middle of
3209 * the list then update the lastbuf and update
3212 if (*bufp != NULL) {
3214 bufp = &(*bufp)->b_next;
3218 ASSERT3P(lastbuf, !=, buf);
3219 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3220 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3221 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3227 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3231 arc_buf_destroy_impl(arc_buf_t *buf)
3233 arc_buf_hdr_t *hdr = buf->b_hdr;
3236 * Free up the data associated with the buf but only if we're not
3237 * sharing this with the hdr. If we are sharing it with the hdr, the
3238 * hdr is responsible for doing the free.
3240 if (buf->b_data != NULL) {
3242 * We're about to change the hdr's b_flags. We must either
3243 * hold the hash_lock or be undiscoverable.
3245 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3247 arc_cksum_verify(buf);
3249 arc_buf_unwatch(buf);
3252 if (arc_buf_is_shared(buf)) {
3253 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3255 uint64_t size = arc_buf_size(buf);
3256 arc_free_data_buf(hdr, buf->b_data, size, buf);
3257 ARCSTAT_INCR(arcstat_overhead_size, -size);
3261 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3262 hdr->b_l1hdr.b_bufcnt -= 1;
3265 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3267 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3269 * If the current arc_buf_t is sharing its data buffer with the
3270 * hdr, then reassign the hdr's b_pabd to share it with the new
3271 * buffer at the end of the list. The shared buffer is always
3272 * the last one on the hdr's buffer list.
3274 * There is an equivalent case for compressed bufs, but since
3275 * they aren't guaranteed to be the last buf in the list and
3276 * that is an exceedingly rare case, we just allow that space be
3277 * wasted temporarily.
3279 if (lastbuf != NULL) {
3280 /* Only one buf can be shared at once */
3281 VERIFY(!arc_buf_is_shared(lastbuf));
3282 /* hdr is uncompressed so can't have compressed buf */
3283 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3285 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3286 arc_hdr_free_pabd(hdr);
3289 * We must setup a new shared block between the
3290 * last buffer and the hdr. The data would have
3291 * been allocated by the arc buf so we need to transfer
3292 * ownership to the hdr since it's now being shared.
3294 arc_share_buf(hdr, lastbuf);
3296 } else if (HDR_SHARED_DATA(hdr)) {
3298 * Uncompressed shared buffers are always at the end
3299 * of the list. Compressed buffers don't have the
3300 * same requirements. This makes it hard to
3301 * simply assert that the lastbuf is shared so
3302 * we rely on the hdr's compression flags to determine
3303 * if we have a compressed, shared buffer.
3305 ASSERT3P(lastbuf, !=, NULL);
3306 ASSERT(arc_buf_is_shared(lastbuf) ||
3307 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
3311 * Free the checksum if we're removing the last uncompressed buf from
3314 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3315 arc_cksum_free(hdr);
3318 /* clean up the buf */
3320 kmem_cache_free(buf_cache, buf);
3324 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr)
3326 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3327 ASSERT(HDR_HAS_L1HDR(hdr));
3328 ASSERT(!HDR_SHARED_DATA(hdr));
3330 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3331 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
3332 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3333 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3335 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3336 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3340 arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
3342 ASSERT(HDR_HAS_L1HDR(hdr));
3343 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3346 * If the hdr is currently being written to the l2arc then
3347 * we defer freeing the data by adding it to the l2arc_free_on_write
3348 * list. The l2arc will free the data once it's finished
3349 * writing it to the l2arc device.
3351 if (HDR_L2_WRITING(hdr)) {
3352 arc_hdr_free_on_write(hdr);
3353 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3355 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
3356 arc_hdr_size(hdr), hdr);
3358 hdr->b_l1hdr.b_pabd = NULL;
3359 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3361 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3362 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3365 static arc_buf_hdr_t *
3366 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3367 enum zio_compress compression_type, arc_buf_contents_t type)
3371 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3373 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3374 ASSERT(HDR_EMPTY(hdr));
3375 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3376 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
3377 HDR_SET_PSIZE(hdr, psize);
3378 HDR_SET_LSIZE(hdr, lsize);
3382 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3383 arc_hdr_set_compress(hdr, compression_type);
3385 hdr->b_l1hdr.b_state = arc_anon;
3386 hdr->b_l1hdr.b_arc_access = 0;
3387 hdr->b_l1hdr.b_bufcnt = 0;
3388 hdr->b_l1hdr.b_buf = NULL;
3391 * Allocate the hdr's buffer. This will contain either
3392 * the compressed or uncompressed data depending on the block
3393 * it references and compressed arc enablement.
3395 arc_hdr_alloc_pabd(hdr);
3396 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3402 * Transition between the two allocation states for the arc_buf_hdr struct.
3403 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3404 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3405 * version is used when a cache buffer is only in the L2ARC in order to reduce
3408 static arc_buf_hdr_t *
3409 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3411 ASSERT(HDR_HAS_L2HDR(hdr));
3413 arc_buf_hdr_t *nhdr;
3414 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3416 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3417 (old == hdr_l2only_cache && new == hdr_full_cache));
3419 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3421 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3422 buf_hash_remove(hdr);
3424 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3426 if (new == hdr_full_cache) {
3427 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3429 * arc_access and arc_change_state need to be aware that a
3430 * header has just come out of L2ARC, so we set its state to
3431 * l2c_only even though it's about to change.
3433 nhdr->b_l1hdr.b_state = arc_l2c_only;
3435 /* Verify previous threads set to NULL before freeing */
3436 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3438 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3439 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3440 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3443 * If we've reached here, We must have been called from
3444 * arc_evict_hdr(), as such we should have already been
3445 * removed from any ghost list we were previously on
3446 * (which protects us from racing with arc_evict_state),
3447 * thus no locking is needed during this check.
3449 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3452 * A buffer must not be moved into the arc_l2c_only
3453 * state if it's not finished being written out to the
3454 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3455 * might try to be accessed, even though it was removed.
3457 VERIFY(!HDR_L2_WRITING(hdr));
3458 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3461 if (hdr->b_l1hdr.b_thawed != NULL) {
3462 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3463 hdr->b_l1hdr.b_thawed = NULL;
3467 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3470 * The header has been reallocated so we need to re-insert it into any
3473 (void) buf_hash_insert(nhdr, NULL);
3475 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3477 mutex_enter(&dev->l2ad_mtx);
3480 * We must place the realloc'ed header back into the list at
3481 * the same spot. Otherwise, if it's placed earlier in the list,
3482 * l2arc_write_buffers() could find it during the function's
3483 * write phase, and try to write it out to the l2arc.
3485 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3486 list_remove(&dev->l2ad_buflist, hdr);
3488 mutex_exit(&dev->l2ad_mtx);
3491 * Since we're using the pointer address as the tag when
3492 * incrementing and decrementing the l2ad_alloc refcount, we
3493 * must remove the old pointer (that we're about to destroy) and
3494 * add the new pointer to the refcount. Otherwise we'd remove
3495 * the wrong pointer address when calling arc_hdr_destroy() later.
3498 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
3499 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
3501 buf_discard_identity(hdr);
3502 kmem_cache_free(old, hdr);
3508 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3509 * The buf is returned thawed since we expect the consumer to modify it.
3512 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3514 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3515 ZIO_COMPRESS_OFF, type);
3516 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3518 arc_buf_t *buf = NULL;
3519 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf));
3526 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3527 * for bufs containing metadata.
3530 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3531 enum zio_compress compression_type)
3533 ASSERT3U(lsize, >, 0);
3534 ASSERT3U(lsize, >=, psize);
3535 ASSERT(compression_type > ZIO_COMPRESS_OFF);
3536 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3538 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3539 compression_type, ARC_BUFC_DATA);
3540 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3542 arc_buf_t *buf = NULL;
3543 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3545 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3547 if (!arc_buf_is_shared(buf)) {
3549 * To ensure that the hdr has the correct data in it if we call
3550 * arc_decompress() on this buf before it's been written to
3551 * disk, it's easiest if we just set up sharing between the
3554 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3555 arc_hdr_free_pabd(hdr);
3556 arc_share_buf(hdr, buf);
3563 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3565 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3566 l2arc_dev_t *dev = l2hdr->b_dev;
3567 uint64_t psize = arc_hdr_size(hdr);
3569 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3570 ASSERT(HDR_HAS_L2HDR(hdr));
3572 list_remove(&dev->l2ad_buflist, hdr);
3574 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3575 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3577 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3579 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3580 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3584 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3586 if (HDR_HAS_L1HDR(hdr)) {
3587 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3588 hdr->b_l1hdr.b_bufcnt > 0);
3589 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3590 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3592 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3593 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3595 if (!HDR_EMPTY(hdr))
3596 buf_discard_identity(hdr);
3598 if (HDR_HAS_L2HDR(hdr)) {
3599 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3600 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3603 mutex_enter(&dev->l2ad_mtx);
3606 * Even though we checked this conditional above, we
3607 * need to check this again now that we have the
3608 * l2ad_mtx. This is because we could be racing with
3609 * another thread calling l2arc_evict() which might have
3610 * destroyed this header's L2 portion as we were waiting
3611 * to acquire the l2ad_mtx. If that happens, we don't
3612 * want to re-destroy the header's L2 portion.
3614 if (HDR_HAS_L2HDR(hdr)) {
3616 arc_hdr_l2hdr_destroy(hdr);
3620 mutex_exit(&dev->l2ad_mtx);
3623 if (HDR_HAS_L1HDR(hdr)) {
3624 arc_cksum_free(hdr);
3626 while (hdr->b_l1hdr.b_buf != NULL)
3627 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3630 if (hdr->b_l1hdr.b_thawed != NULL) {
3631 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3632 hdr->b_l1hdr.b_thawed = NULL;
3636 if (hdr->b_l1hdr.b_pabd != NULL) {
3637 arc_hdr_free_pabd(hdr);
3641 ASSERT3P(hdr->b_hash_next, ==, NULL);
3642 if (HDR_HAS_L1HDR(hdr)) {
3643 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3644 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3645 kmem_cache_free(hdr_full_cache, hdr);
3647 kmem_cache_free(hdr_l2only_cache, hdr);
3652 arc_buf_destroy(arc_buf_t *buf, void* tag)
3654 arc_buf_hdr_t *hdr = buf->b_hdr;
3655 kmutex_t *hash_lock = HDR_LOCK(hdr);
3657 if (hdr->b_l1hdr.b_state == arc_anon) {
3658 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3659 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3660 VERIFY0(remove_reference(hdr, NULL, tag));
3661 arc_hdr_destroy(hdr);
3665 mutex_enter(hash_lock);
3666 ASSERT3P(hdr, ==, buf->b_hdr);
3667 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3668 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3669 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3670 ASSERT3P(buf->b_data, !=, NULL);
3672 (void) remove_reference(hdr, hash_lock, tag);
3673 arc_buf_destroy_impl(buf);
3674 mutex_exit(hash_lock);
3678 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3679 * state of the header is dependent on its state prior to entering this
3680 * function. The following transitions are possible:
3682 * - arc_mru -> arc_mru_ghost
3683 * - arc_mfu -> arc_mfu_ghost
3684 * - arc_mru_ghost -> arc_l2c_only
3685 * - arc_mru_ghost -> deleted
3686 * - arc_mfu_ghost -> arc_l2c_only
3687 * - arc_mfu_ghost -> deleted
3690 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3692 arc_state_t *evicted_state, *state;
3693 int64_t bytes_evicted = 0;
3694 int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3695 zfs_arc_min_prescient_prefetch_ms : zfs_arc_min_prefetch_ms;
3697 ASSERT(MUTEX_HELD(hash_lock));
3698 ASSERT(HDR_HAS_L1HDR(hdr));
3700 state = hdr->b_l1hdr.b_state;
3701 if (GHOST_STATE(state)) {
3702 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3703 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3706 * l2arc_write_buffers() relies on a header's L1 portion
3707 * (i.e. its b_pabd field) during it's write phase.
3708 * Thus, we cannot push a header onto the arc_l2c_only
3709 * state (removing it's L1 piece) until the header is
3710 * done being written to the l2arc.
3712 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3713 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3714 return (bytes_evicted);
3717 ARCSTAT_BUMP(arcstat_deleted);
3718 bytes_evicted += HDR_GET_LSIZE(hdr);
3720 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3722 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3723 if (HDR_HAS_L2HDR(hdr)) {
3725 * This buffer is cached on the 2nd Level ARC;
3726 * don't destroy the header.
3728 arc_change_state(arc_l2c_only, hdr, hash_lock);
3730 * dropping from L1+L2 cached to L2-only,
3731 * realloc to remove the L1 header.
3733 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3736 arc_change_state(arc_anon, hdr, hash_lock);
3737 arc_hdr_destroy(hdr);
3739 return (bytes_evicted);
3742 ASSERT(state == arc_mru || state == arc_mfu);
3743 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3745 /* prefetch buffers have a minimum lifespan */
3746 if (HDR_IO_IN_PROGRESS(hdr) ||
3747 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3748 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < min_lifetime * hz)) {
3749 ARCSTAT_BUMP(arcstat_evict_skip);
3750 return (bytes_evicted);
3753 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3754 while (hdr->b_l1hdr.b_buf) {
3755 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3756 if (!mutex_tryenter(&buf->b_evict_lock)) {
3757 ARCSTAT_BUMP(arcstat_mutex_miss);
3760 if (buf->b_data != NULL)
3761 bytes_evicted += HDR_GET_LSIZE(hdr);
3762 mutex_exit(&buf->b_evict_lock);
3763 arc_buf_destroy_impl(buf);
3766 if (HDR_HAS_L2HDR(hdr)) {
3767 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3769 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3770 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3771 HDR_GET_LSIZE(hdr));
3773 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3774 HDR_GET_LSIZE(hdr));
3778 if (hdr->b_l1hdr.b_bufcnt == 0) {
3779 arc_cksum_free(hdr);
3781 bytes_evicted += arc_hdr_size(hdr);
3784 * If this hdr is being evicted and has a compressed
3785 * buffer then we discard it here before we change states.
3786 * This ensures that the accounting is updated correctly
3787 * in arc_free_data_impl().
3789 arc_hdr_free_pabd(hdr);
3791 arc_change_state(evicted_state, hdr, hash_lock);
3792 ASSERT(HDR_IN_HASH_TABLE(hdr));
3793 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3794 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3797 return (bytes_evicted);
3801 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3802 uint64_t spa, int64_t bytes)
3804 multilist_sublist_t *mls;
3805 uint64_t bytes_evicted = 0;
3807 kmutex_t *hash_lock;
3808 int evict_count = 0;
3810 ASSERT3P(marker, !=, NULL);
3811 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3813 mls = multilist_sublist_lock(ml, idx);
3815 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3816 hdr = multilist_sublist_prev(mls, marker)) {
3817 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3818 (evict_count >= zfs_arc_evict_batch_limit))
3822 * To keep our iteration location, move the marker
3823 * forward. Since we're not holding hdr's hash lock, we
3824 * must be very careful and not remove 'hdr' from the
3825 * sublist. Otherwise, other consumers might mistake the
3826 * 'hdr' as not being on a sublist when they call the
3827 * multilist_link_active() function (they all rely on
3828 * the hash lock protecting concurrent insertions and
3829 * removals). multilist_sublist_move_forward() was
3830 * specifically implemented to ensure this is the case
3831 * (only 'marker' will be removed and re-inserted).
3833 multilist_sublist_move_forward(mls, marker);
3836 * The only case where the b_spa field should ever be
3837 * zero, is the marker headers inserted by
3838 * arc_evict_state(). It's possible for multiple threads
3839 * to be calling arc_evict_state() concurrently (e.g.
3840 * dsl_pool_close() and zio_inject_fault()), so we must
3841 * skip any markers we see from these other threads.
3843 if (hdr->b_spa == 0)
3846 /* we're only interested in evicting buffers of a certain spa */
3847 if (spa != 0 && hdr->b_spa != spa) {
3848 ARCSTAT_BUMP(arcstat_evict_skip);
3852 hash_lock = HDR_LOCK(hdr);
3855 * We aren't calling this function from any code path
3856 * that would already be holding a hash lock, so we're
3857 * asserting on this assumption to be defensive in case
3858 * this ever changes. Without this check, it would be
3859 * possible to incorrectly increment arcstat_mutex_miss
3860 * below (e.g. if the code changed such that we called
3861 * this function with a hash lock held).
3863 ASSERT(!MUTEX_HELD(hash_lock));
3865 if (mutex_tryenter(hash_lock)) {
3866 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3867 mutex_exit(hash_lock);
3869 bytes_evicted += evicted;
3872 * If evicted is zero, arc_evict_hdr() must have
3873 * decided to skip this header, don't increment
3874 * evict_count in this case.
3880 * If arc_size isn't overflowing, signal any
3881 * threads that might happen to be waiting.
3883 * For each header evicted, we wake up a single
3884 * thread. If we used cv_broadcast, we could
3885 * wake up "too many" threads causing arc_size
3886 * to significantly overflow arc_c; since
3887 * arc_get_data_impl() doesn't check for overflow
3888 * when it's woken up (it doesn't because it's
3889 * possible for the ARC to be overflowing while
3890 * full of un-evictable buffers, and the
3891 * function should proceed in this case).
3893 * If threads are left sleeping, due to not
3894 * using cv_broadcast, they will be woken up
3895 * just before arc_reclaim_thread() sleeps.
3897 mutex_enter(&arc_reclaim_lock);
3898 if (!arc_is_overflowing())
3899 cv_signal(&arc_reclaim_waiters_cv);
3900 mutex_exit(&arc_reclaim_lock);
3902 ARCSTAT_BUMP(arcstat_mutex_miss);
3906 multilist_sublist_unlock(mls);
3908 return (bytes_evicted);
3912 * Evict buffers from the given arc state, until we've removed the
3913 * specified number of bytes. Move the removed buffers to the
3914 * appropriate evict state.
3916 * This function makes a "best effort". It skips over any buffers
3917 * it can't get a hash_lock on, and so, may not catch all candidates.
3918 * It may also return without evicting as much space as requested.
3920 * If bytes is specified using the special value ARC_EVICT_ALL, this
3921 * will evict all available (i.e. unlocked and evictable) buffers from
3922 * the given arc state; which is used by arc_flush().
3925 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3926 arc_buf_contents_t type)
3928 uint64_t total_evicted = 0;
3929 multilist_t *ml = state->arcs_list[type];
3931 arc_buf_hdr_t **markers;
3933 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3935 num_sublists = multilist_get_num_sublists(ml);
3938 * If we've tried to evict from each sublist, made some
3939 * progress, but still have not hit the target number of bytes
3940 * to evict, we want to keep trying. The markers allow us to
3941 * pick up where we left off for each individual sublist, rather
3942 * than starting from the tail each time.
3944 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3945 for (int i = 0; i < num_sublists; i++) {
3946 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3949 * A b_spa of 0 is used to indicate that this header is
3950 * a marker. This fact is used in arc_adjust_type() and
3951 * arc_evict_state_impl().
3953 markers[i]->b_spa = 0;
3955 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3956 multilist_sublist_insert_tail(mls, markers[i]);
3957 multilist_sublist_unlock(mls);
3961 * While we haven't hit our target number of bytes to evict, or
3962 * we're evicting all available buffers.
3964 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3965 int sublist_idx = multilist_get_random_index(ml);
3966 uint64_t scan_evicted = 0;
3969 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
3970 * Request that 10% of the LRUs be scanned by the superblock
3973 if (type == ARC_BUFC_DATA && aggsum_compare(&astat_dnode_size,
3974 arc_dnode_limit) > 0) {
3975 arc_prune_async((aggsum_upper_bound(&astat_dnode_size) -
3976 arc_dnode_limit) / sizeof (dnode_t) /
3977 zfs_arc_dnode_reduce_percent);
3981 * Start eviction using a randomly selected sublist,
3982 * this is to try and evenly balance eviction across all
3983 * sublists. Always starting at the same sublist
3984 * (e.g. index 0) would cause evictions to favor certain
3985 * sublists over others.
3987 for (int i = 0; i < num_sublists; i++) {
3988 uint64_t bytes_remaining;
3989 uint64_t bytes_evicted;
3991 if (bytes == ARC_EVICT_ALL)
3992 bytes_remaining = ARC_EVICT_ALL;
3993 else if (total_evicted < bytes)
3994 bytes_remaining = bytes - total_evicted;
3998 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3999 markers[sublist_idx], spa, bytes_remaining);
4001 scan_evicted += bytes_evicted;
4002 total_evicted += bytes_evicted;
4004 /* we've reached the end, wrap to the beginning */
4005 if (++sublist_idx >= num_sublists)
4010 * If we didn't evict anything during this scan, we have
4011 * no reason to believe we'll evict more during another
4012 * scan, so break the loop.
4014 if (scan_evicted == 0) {
4015 /* This isn't possible, let's make that obvious */
4016 ASSERT3S(bytes, !=, 0);
4019 * When bytes is ARC_EVICT_ALL, the only way to
4020 * break the loop is when scan_evicted is zero.
4021 * In that case, we actually have evicted enough,
4022 * so we don't want to increment the kstat.
4024 if (bytes != ARC_EVICT_ALL) {
4025 ASSERT3S(total_evicted, <, bytes);
4026 ARCSTAT_BUMP(arcstat_evict_not_enough);
4033 for (int i = 0; i < num_sublists; i++) {
4034 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4035 multilist_sublist_remove(mls, markers[i]);
4036 multilist_sublist_unlock(mls);
4038 kmem_cache_free(hdr_full_cache, markers[i]);
4040 kmem_free(markers, sizeof (*markers) * num_sublists);
4042 return (total_evicted);
4046 * Flush all "evictable" data of the given type from the arc state
4047 * specified. This will not evict any "active" buffers (i.e. referenced).
4049 * When 'retry' is set to B_FALSE, the function will make a single pass
4050 * over the state and evict any buffers that it can. Since it doesn't
4051 * continually retry the eviction, it might end up leaving some buffers
4052 * in the ARC due to lock misses.
4054 * When 'retry' is set to B_TRUE, the function will continually retry the
4055 * eviction until *all* evictable buffers have been removed from the
4056 * state. As a result, if concurrent insertions into the state are
4057 * allowed (e.g. if the ARC isn't shutting down), this function might
4058 * wind up in an infinite loop, continually trying to evict buffers.
4061 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4064 uint64_t evicted = 0;
4066 while (refcount_count(&state->arcs_esize[type]) != 0) {
4067 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
4077 * Helper function for arc_prune_async() it is responsible for safely
4078 * handling the execution of a registered arc_prune_func_t.
4081 arc_prune_task(void *ptr)
4083 arc_prune_t *ap = (arc_prune_t *)ptr;
4084 arc_prune_func_t *func = ap->p_pfunc;
4087 func(ap->p_adjust, ap->p_private);
4089 refcount_remove(&ap->p_refcnt, func);
4093 * Notify registered consumers they must drop holds on a portion of the ARC
4094 * buffered they reference. This provides a mechanism to ensure the ARC can
4095 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
4096 * is analogous to dnlc_reduce_cache() but more generic.
4098 * This operation is performed asynchronously so it may be safely called
4099 * in the context of the arc_reclaim_thread(). A reference is taken here
4100 * for each registered arc_prune_t and the arc_prune_task() is responsible
4101 * for releasing it once the registered arc_prune_func_t has completed.
4104 arc_prune_async(int64_t adjust)
4108 mutex_enter(&arc_prune_mtx);
4109 for (ap = list_head(&arc_prune_list); ap != NULL;
4110 ap = list_next(&arc_prune_list, ap)) {
4112 if (refcount_count(&ap->p_refcnt) >= 2)
4115 refcount_add(&ap->p_refcnt, ap->p_pfunc);
4116 ap->p_adjust = adjust;
4117 if (taskq_dispatch(arc_prune_taskq, arc_prune_task,
4118 ap, TQ_SLEEP) == TASKQID_INVALID) {
4119 refcount_remove(&ap->p_refcnt, ap->p_pfunc);
4122 ARCSTAT_BUMP(arcstat_prune);
4124 mutex_exit(&arc_prune_mtx);
4128 * Evict the specified number of bytes from the state specified,
4129 * restricting eviction to the spa and type given. This function
4130 * prevents us from trying to evict more from a state's list than
4131 * is "evictable", and to skip evicting altogether when passed a
4132 * negative value for "bytes". In contrast, arc_evict_state() will
4133 * evict everything it can, when passed a negative value for "bytes".
4136 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
4137 arc_buf_contents_t type)
4141 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
4142 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
4143 return (arc_evict_state(state, spa, delta, type));
4150 * The goal of this function is to evict enough meta data buffers from the
4151 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
4152 * more complicated than it appears because it is common for data buffers
4153 * to have holds on meta data buffers. In addition, dnode meta data buffers
4154 * will be held by the dnodes in the block preventing them from being freed.
4155 * This means we can't simply traverse the ARC and expect to always find
4156 * enough unheld meta data buffer to release.
4158 * Therefore, this function has been updated to make alternating passes
4159 * over the ARC releasing data buffers and then newly unheld meta data
4160 * buffers. This ensures forward progress is maintained and meta_used
4161 * will decrease. Normally this is sufficient, but if required the ARC
4162 * will call the registered prune callbacks causing dentry and inodes to
4163 * be dropped from the VFS cache. This will make dnode meta data buffers
4164 * available for reclaim.
4167 arc_adjust_meta_balanced(uint64_t meta_used)
4169 int64_t delta, prune = 0, adjustmnt;
4170 uint64_t total_evicted = 0;
4171 arc_buf_contents_t type = ARC_BUFC_DATA;
4172 int restarts = MAX(zfs_arc_meta_adjust_restarts, 0);
4176 * This slightly differs than the way we evict from the mru in
4177 * arc_adjust because we don't have a "target" value (i.e. no
4178 * "meta" arc_p). As a result, I think we can completely
4179 * cannibalize the metadata in the MRU before we evict the
4180 * metadata from the MFU. I think we probably need to implement a
4181 * "metadata arc_p" value to do this properly.
4183 adjustmnt = meta_used - arc_meta_limit;
4185 if (adjustmnt > 0 && refcount_count(&arc_mru->arcs_esize[type]) > 0) {
4186 delta = MIN(refcount_count(&arc_mru->arcs_esize[type]),
4188 total_evicted += arc_adjust_impl(arc_mru, 0, delta, type);
4193 * We can't afford to recalculate adjustmnt here. If we do,
4194 * new metadata buffers can sneak into the MRU or ANON lists,
4195 * thus penalize the MFU metadata. Although the fudge factor is
4196 * small, it has been empirically shown to be significant for
4197 * certain workloads (e.g. creating many empty directories). As
4198 * such, we use the original calculation for adjustmnt, and
4199 * simply decrement the amount of data evicted from the MRU.
4202 if (adjustmnt > 0 && refcount_count(&arc_mfu->arcs_esize[type]) > 0) {
4203 delta = MIN(refcount_count(&arc_mfu->arcs_esize[type]),
4205 total_evicted += arc_adjust_impl(arc_mfu, 0, delta, type);
4208 adjustmnt = meta_used - arc_meta_limit;
4210 if (adjustmnt > 0 &&
4211 refcount_count(&arc_mru_ghost->arcs_esize[type]) > 0) {
4212 delta = MIN(adjustmnt,
4213 refcount_count(&arc_mru_ghost->arcs_esize[type]));
4214 total_evicted += arc_adjust_impl(arc_mru_ghost, 0, delta, type);
4218 if (adjustmnt > 0 &&
4219 refcount_count(&arc_mfu_ghost->arcs_esize[type]) > 0) {
4220 delta = MIN(adjustmnt,
4221 refcount_count(&arc_mfu_ghost->arcs_esize[type]));
4222 total_evicted += arc_adjust_impl(arc_mfu_ghost, 0, delta, type);
4226 * If after attempting to make the requested adjustment to the ARC
4227 * the meta limit is still being exceeded then request that the
4228 * higher layers drop some cached objects which have holds on ARC
4229 * meta buffers. Requests to the upper layers will be made with
4230 * increasingly large scan sizes until the ARC is below the limit.
4232 if (meta_used > arc_meta_limit) {
4233 if (type == ARC_BUFC_DATA) {
4234 type = ARC_BUFC_METADATA;
4236 type = ARC_BUFC_DATA;
4238 if (zfs_arc_meta_prune) {
4239 prune += zfs_arc_meta_prune;
4240 arc_prune_async(prune);
4249 return (total_evicted);
4253 * Evict metadata buffers from the cache, such that arc_meta_used is
4254 * capped by the arc_meta_limit tunable.
4257 arc_adjust_meta_only(uint64_t meta_used)
4259 uint64_t total_evicted = 0;
4263 * If we're over the meta limit, we want to evict enough
4264 * metadata to get back under the meta limit. We don't want to
4265 * evict so much that we drop the MRU below arc_p, though. If
4266 * we're over the meta limit more than we're over arc_p, we
4267 * evict some from the MRU here, and some from the MFU below.
4269 target = MIN((int64_t)(meta_used - arc_meta_limit),
4270 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4271 refcount_count(&arc_mru->arcs_size) - arc_p));
4273 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4276 * Similar to the above, we want to evict enough bytes to get us
4277 * below the meta limit, but not so much as to drop us below the
4278 * space allotted to the MFU (which is defined as arc_c - arc_p).
4280 target = MIN((int64_t)(meta_used - arc_meta_limit),
4281 (int64_t)(refcount_count(&arc_mfu->arcs_size) -
4284 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4286 return (total_evicted);
4290 arc_adjust_meta(uint64_t meta_used)
4292 if (zfs_arc_meta_strategy == ARC_STRATEGY_META_ONLY)
4293 return (arc_adjust_meta_only(meta_used));
4295 return (arc_adjust_meta_balanced(meta_used));
4299 * Return the type of the oldest buffer in the given arc state
4301 * This function will select a random sublist of type ARC_BUFC_DATA and
4302 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4303 * is compared, and the type which contains the "older" buffer will be
4306 static arc_buf_contents_t
4307 arc_adjust_type(arc_state_t *state)
4309 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
4310 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
4311 int data_idx = multilist_get_random_index(data_ml);
4312 int meta_idx = multilist_get_random_index(meta_ml);
4313 multilist_sublist_t *data_mls;
4314 multilist_sublist_t *meta_mls;
4315 arc_buf_contents_t type;
4316 arc_buf_hdr_t *data_hdr;
4317 arc_buf_hdr_t *meta_hdr;
4320 * We keep the sublist lock until we're finished, to prevent
4321 * the headers from being destroyed via arc_evict_state().
4323 data_mls = multilist_sublist_lock(data_ml, data_idx);
4324 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4327 * These two loops are to ensure we skip any markers that
4328 * might be at the tail of the lists due to arc_evict_state().
4331 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4332 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4333 if (data_hdr->b_spa != 0)
4337 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4338 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4339 if (meta_hdr->b_spa != 0)
4343 if (data_hdr == NULL && meta_hdr == NULL) {
4344 type = ARC_BUFC_DATA;
4345 } else if (data_hdr == NULL) {
4346 ASSERT3P(meta_hdr, !=, NULL);
4347 type = ARC_BUFC_METADATA;
4348 } else if (meta_hdr == NULL) {
4349 ASSERT3P(data_hdr, !=, NULL);
4350 type = ARC_BUFC_DATA;
4352 ASSERT3P(data_hdr, !=, NULL);
4353 ASSERT3P(meta_hdr, !=, NULL);
4355 /* The headers can't be on the sublist without an L1 header */
4356 ASSERT(HDR_HAS_L1HDR(data_hdr));
4357 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4359 if (data_hdr->b_l1hdr.b_arc_access <
4360 meta_hdr->b_l1hdr.b_arc_access) {
4361 type = ARC_BUFC_DATA;
4363 type = ARC_BUFC_METADATA;
4367 multilist_sublist_unlock(meta_mls);
4368 multilist_sublist_unlock(data_mls);
4374 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4379 uint64_t total_evicted = 0;
4382 uint64_t asize = aggsum_value(&arc_size);
4383 uint64_t ameta = aggsum_value(&arc_meta_used);
4386 * If we're over arc_meta_limit, we want to correct that before
4387 * potentially evicting data buffers below.
4389 total_evicted += arc_adjust_meta(ameta);
4394 * If we're over the target cache size, we want to evict enough
4395 * from the list to get back to our target size. We don't want
4396 * to evict too much from the MRU, such that it drops below
4397 * arc_p. So, if we're over our target cache size more than
4398 * the MRU is over arc_p, we'll evict enough to get back to
4399 * arc_p here, and then evict more from the MFU below.
4401 target = MIN((int64_t)(asize - arc_c),
4402 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4403 refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4406 * If we're below arc_meta_min, always prefer to evict data.
4407 * Otherwise, try to satisfy the requested number of bytes to
4408 * evict from the type which contains older buffers; in an
4409 * effort to keep newer buffers in the cache regardless of their
4410 * type. If we cannot satisfy the number of bytes from this
4411 * type, spill over into the next type.
4413 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4414 ameta > arc_meta_min) {
4415 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4416 total_evicted += bytes;
4419 * If we couldn't evict our target number of bytes from
4420 * metadata, we try to get the rest from data.
4425 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4427 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4428 total_evicted += bytes;
4431 * If we couldn't evict our target number of bytes from
4432 * data, we try to get the rest from metadata.
4437 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4443 * Now that we've tried to evict enough from the MRU to get its
4444 * size back to arc_p, if we're still above the target cache
4445 * size, we evict the rest from the MFU.
4447 target = asize - arc_c;
4449 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4450 ameta > arc_meta_min) {
4451 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4452 total_evicted += bytes;
4455 * If we couldn't evict our target number of bytes from
4456 * metadata, we try to get the rest from data.
4461 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4463 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4464 total_evicted += bytes;
4467 * If we couldn't evict our target number of bytes from
4468 * data, we try to get the rest from data.
4473 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4477 * Adjust ghost lists
4479 * In addition to the above, the ARC also defines target values
4480 * for the ghost lists. The sum of the mru list and mru ghost
4481 * list should never exceed the target size of the cache, and
4482 * the sum of the mru list, mfu list, mru ghost list, and mfu
4483 * ghost list should never exceed twice the target size of the
4484 * cache. The following logic enforces these limits on the ghost
4485 * caches, and evicts from them as needed.
4487 target = refcount_count(&arc_mru->arcs_size) +
4488 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4490 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4491 total_evicted += bytes;
4496 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4499 * We assume the sum of the mru list and mfu list is less than
4500 * or equal to arc_c (we enforced this above), which means we
4501 * can use the simpler of the two equations below:
4503 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4504 * mru ghost + mfu ghost <= arc_c
4506 target = refcount_count(&arc_mru_ghost->arcs_size) +
4507 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4509 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4510 total_evicted += bytes;
4515 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4517 return (total_evicted);
4521 arc_flush(spa_t *spa, boolean_t retry)
4526 * If retry is B_TRUE, a spa must not be specified since we have
4527 * no good way to determine if all of a spa's buffers have been
4528 * evicted from an arc state.
4530 ASSERT(!retry || spa == 0);
4533 guid = spa_load_guid(spa);
4535 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4536 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4538 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4539 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4541 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4542 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4544 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4545 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4549 arc_shrink(int64_t to_free)
4551 uint64_t asize = aggsum_value(&arc_size);
4552 if (arc_c > arc_c_min) {
4553 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
4554 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
4555 if (arc_c > arc_c_min + to_free)
4556 atomic_add_64(&arc_c, -to_free);
4560 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4562 arc_c = MAX(asize, arc_c_min);
4564 arc_p = (arc_c >> 1);
4566 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
4569 ASSERT(arc_c >= arc_c_min);
4570 ASSERT((int64_t)arc_p >= 0);
4573 if (asize > arc_c) {
4574 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, asize,
4576 (void) arc_adjust();
4580 typedef enum free_memory_reason_t {
4585 FMR_PAGES_PP_MAXIMUM,
4588 } free_memory_reason_t;
4590 int64_t last_free_memory;
4591 free_memory_reason_t last_free_reason;
4594 * Additional reserve of pages for pp_reserve.
4596 int64_t arc_pages_pp_reserve = 64;
4599 * Additional reserve of pages for swapfs.
4601 int64_t arc_swapfs_reserve = 64;
4604 * Return the amount of memory that can be consumed before reclaim will be
4605 * needed. Positive if there is sufficient free memory, negative indicates
4606 * the amount of memory that needs to be freed up.
4609 arc_available_memory(void)
4611 int64_t lowest = INT64_MAX;
4613 free_memory_reason_t r = FMR_UNKNOWN;
4618 * Cooperate with pagedaemon when it's time for it to scan
4619 * and reclaim some pages.
4621 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
4629 n = PAGESIZE * (-needfree);
4637 * check that we're out of range of the pageout scanner. It starts to
4638 * schedule paging if freemem is less than lotsfree and needfree.
4639 * lotsfree is the high-water mark for pageout, and needfree is the
4640 * number of needed free pages. We add extra pages here to make sure
4641 * the scanner doesn't start up while we're freeing memory.
4643 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4650 * check to make sure that swapfs has enough space so that anon
4651 * reservations can still succeed. anon_resvmem() checks that the
4652 * availrmem is greater than swapfs_minfree, and the number of reserved
4653 * swap pages. We also add a bit of extra here just to prevent
4654 * circumstances from getting really dire.
4656 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4657 desfree - arc_swapfs_reserve);
4660 r = FMR_SWAPFS_MINFREE;
4665 * Check that we have enough availrmem that memory locking (e.g., via
4666 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4667 * stores the number of pages that cannot be locked; when availrmem
4668 * drops below pages_pp_maximum, page locking mechanisms such as
4669 * page_pp_lock() will fail.)
4671 n = PAGESIZE * (availrmem - pages_pp_maximum -
4672 arc_pages_pp_reserve);
4675 r = FMR_PAGES_PP_MAXIMUM;
4678 #endif /* __FreeBSD__ */
4679 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
4681 * If we're on an i386 platform, it's possible that we'll exhaust the
4682 * kernel heap space before we ever run out of available physical
4683 * memory. Most checks of the size of the heap_area compare against
4684 * tune.t_minarmem, which is the minimum available real memory that we
4685 * can have in the system. However, this is generally fixed at 25 pages
4686 * which is so low that it's useless. In this comparison, we seek to
4687 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4688 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4691 n = uma_avail() - (long)(uma_limit() / 4);
4699 * If zio data pages are being allocated out of a separate heap segment,
4700 * then enforce that the size of available vmem for this arena remains
4701 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4703 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4704 * memory (in the zio_arena) free, which can avoid memory
4705 * fragmentation issues.
4707 if (zio_arena != NULL) {
4708 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4709 (vmem_size(zio_arena, VMEM_ALLOC) >>
4710 arc_zio_arena_free_shift);
4718 /* Every 100 calls, free a small amount */
4719 if (spa_get_random(100) == 0)
4721 #endif /* _KERNEL */
4723 last_free_memory = lowest;
4724 last_free_reason = r;
4725 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
4731 * Determine if the system is under memory pressure and is asking
4732 * to reclaim memory. A return value of B_TRUE indicates that the system
4733 * is under memory pressure and that the arc should adjust accordingly.
4736 arc_reclaim_needed(void)
4738 return (arc_available_memory() < 0);
4741 extern kmem_cache_t *zio_buf_cache[];
4742 extern kmem_cache_t *zio_data_buf_cache[];
4743 extern kmem_cache_t *range_seg_cache;
4744 extern kmem_cache_t *abd_chunk_cache;
4746 static __noinline void
4747 arc_kmem_reap_now(void)
4750 kmem_cache_t *prev_cache = NULL;
4751 kmem_cache_t *prev_data_cache = NULL;
4753 DTRACE_PROBE(arc__kmem_reap_start);
4755 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) {
4757 * We are exceeding our meta-data cache limit.
4758 * Purge some DNLC entries to release holds on meta-data.
4760 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4764 * Reclaim unused memory from all kmem caches.
4771 * If a kmem reap is already active, don't schedule more. We must
4772 * check for this because kmem_cache_reap_soon() won't actually
4773 * block on the cache being reaped (this is to prevent callers from
4774 * becoming implicitly blocked by a system-wide kmem reap -- which,
4775 * on a system with many, many full magazines, can take minutes).
4777 if (kmem_cache_reap_active())
4780 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4781 if (zio_buf_cache[i] != prev_cache) {
4782 prev_cache = zio_buf_cache[i];
4783 kmem_cache_reap_soon(zio_buf_cache[i]);
4785 if (zio_data_buf_cache[i] != prev_data_cache) {
4786 prev_data_cache = zio_data_buf_cache[i];
4787 kmem_cache_reap_soon(zio_data_buf_cache[i]);
4790 kmem_cache_reap_soon(abd_chunk_cache);
4791 kmem_cache_reap_soon(buf_cache);
4792 kmem_cache_reap_soon(hdr_full_cache);
4793 kmem_cache_reap_soon(hdr_l2only_cache);
4794 kmem_cache_reap_soon(range_seg_cache);
4797 if (zio_arena != NULL) {
4799 * Ask the vmem arena to reclaim unused memory from its
4802 vmem_qcache_reap(zio_arena);
4805 DTRACE_PROBE(arc__kmem_reap_end);
4809 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4810 * enough data and signal them to proceed. When this happens, the threads in
4811 * arc_get_data_impl() are sleeping while holding the hash lock for their
4812 * particular arc header. Thus, we must be careful to never sleep on a
4813 * hash lock in this thread. This is to prevent the following deadlock:
4815 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4816 * waiting for the reclaim thread to signal it.
4818 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4819 * fails, and goes to sleep forever.
4821 * This possible deadlock is avoided by always acquiring a hash lock
4822 * using mutex_tryenter() from arc_reclaim_thread().
4826 arc_reclaim_thread(void *unused __unused)
4828 hrtime_t growtime = 0;
4829 hrtime_t kmem_reap_time = 0;
4832 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4834 mutex_enter(&arc_reclaim_lock);
4835 while (!arc_reclaim_thread_exit) {
4836 uint64_t evicted = 0;
4839 * This is necessary in order for the mdb ::arc dcmd to
4840 * show up to date information. Since the ::arc command
4841 * does not call the kstat's update function, without
4842 * this call, the command may show stale stats for the
4843 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4844 * with this change, the data might be up to 1 second
4845 * out of date; but that should suffice. The arc_state_t
4846 * structures can be queried directly if more accurate
4847 * information is needed.
4849 if (arc_ksp != NULL)
4850 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4852 mutex_exit(&arc_reclaim_lock);
4855 * We call arc_adjust() before (possibly) calling
4856 * arc_kmem_reap_now(), so that we can wake up
4857 * arc_get_data_impl() sooner.
4859 evicted = arc_adjust();
4861 int64_t free_memory = arc_available_memory();
4862 if (free_memory < 0) {
4863 hrtime_t curtime = gethrtime();
4864 arc_no_grow = B_TRUE;
4868 * Wait at least zfs_grow_retry (default 60) seconds
4869 * before considering growing.
4871 growtime = curtime + SEC2NSEC(arc_grow_retry);
4874 * Wait at least arc_kmem_cache_reap_retry_ms
4875 * between arc_kmem_reap_now() calls. Without
4876 * this check it is possible to end up in a
4877 * situation where we spend lots of time
4878 * reaping caches, while we're near arc_c_min.
4880 if (curtime >= kmem_reap_time) {
4881 arc_kmem_reap_now();
4882 kmem_reap_time = gethrtime() +
4883 MSEC2NSEC(arc_kmem_cache_reap_retry_ms);
4887 * If we are still low on memory, shrink the ARC
4888 * so that we have arc_shrink_min free space.
4890 free_memory = arc_available_memory();
4893 (arc_c >> arc_shrink_shift) - free_memory;
4897 to_free = MAX(to_free, ptob(needfree));
4900 arc_shrink(to_free);
4902 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4903 arc_no_grow = B_TRUE;
4904 } else if (gethrtime() >= growtime) {
4905 arc_no_grow = B_FALSE;
4908 mutex_enter(&arc_reclaim_lock);
4911 * If evicted is zero, we couldn't evict anything via
4912 * arc_adjust(). This could be due to hash lock
4913 * collisions, but more likely due to the majority of
4914 * arc buffers being unevictable. Therefore, even if
4915 * arc_size is above arc_c, another pass is unlikely to
4916 * be helpful and could potentially cause us to enter an
4919 if (aggsum_compare(&arc_size, arc_c) <= 0|| evicted == 0) {
4921 * We're either no longer overflowing, or we
4922 * can't evict anything more, so we should wake
4923 * up any threads before we go to sleep.
4925 cv_broadcast(&arc_reclaim_waiters_cv);
4928 * Block until signaled, or after one second (we
4929 * might need to perform arc_kmem_reap_now()
4930 * even if we aren't being signalled)
4932 CALLB_CPR_SAFE_BEGIN(&cpr);
4933 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4934 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4935 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4939 arc_reclaim_thread_exit = B_FALSE;
4940 cv_broadcast(&arc_reclaim_thread_cv);
4941 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
4945 static u_int arc_dnlc_evicts_arg;
4946 extern struct vfsops zfs_vfsops;
4949 arc_dnlc_evicts_thread(void *dummy __unused)
4954 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG);
4956 mutex_enter(&arc_dnlc_evicts_lock);
4957 while (!arc_dnlc_evicts_thread_exit) {
4958 CALLB_CPR_SAFE_BEGIN(&cpr);
4959 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
4960 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock);
4961 if (arc_dnlc_evicts_arg != 0) {
4962 percent = arc_dnlc_evicts_arg;
4963 mutex_exit(&arc_dnlc_evicts_lock);
4965 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops);
4967 mutex_enter(&arc_dnlc_evicts_lock);
4969 * Clear our token only after vnlru_free()
4970 * pass is done, to avoid false queueing of
4973 arc_dnlc_evicts_arg = 0;
4976 arc_dnlc_evicts_thread_exit = FALSE;
4977 cv_broadcast(&arc_dnlc_evicts_cv);
4978 CALLB_CPR_EXIT(&cpr);
4983 dnlc_reduce_cache(void *arg)
4987 percent = (u_int)(uintptr_t)arg;
4988 mutex_enter(&arc_dnlc_evicts_lock);
4989 if (arc_dnlc_evicts_arg == 0) {
4990 arc_dnlc_evicts_arg = percent;
4991 cv_broadcast(&arc_dnlc_evicts_cv);
4993 mutex_exit(&arc_dnlc_evicts_lock);
4997 * Adapt arc info given the number of bytes we are trying to add and
4998 * the state that we are comming from. This function is only called
4999 * when we are adding new content to the cache.
5002 arc_adapt(int bytes, arc_state_t *state)
5005 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
5006 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
5007 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
5009 if (state == arc_l2c_only)
5014 * Adapt the target size of the MRU list:
5015 * - if we just hit in the MRU ghost list, then increase
5016 * the target size of the MRU list.
5017 * - if we just hit in the MFU ghost list, then increase
5018 * the target size of the MFU list by decreasing the
5019 * target size of the MRU list.
5021 if (state == arc_mru_ghost) {
5022 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
5023 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
5025 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
5026 } else if (state == arc_mfu_ghost) {
5029 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
5030 mult = MIN(mult, 10);
5032 delta = MIN(bytes * mult, arc_p);
5033 arc_p = MAX(arc_p_min, arc_p - delta);
5035 ASSERT((int64_t)arc_p >= 0);
5037 if (arc_reclaim_needed()) {
5038 cv_signal(&arc_reclaim_thread_cv);
5045 if (arc_c >= arc_c_max)
5049 * If we're within (2 * maxblocksize) bytes of the target
5050 * cache size, increment the target cache size
5052 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >
5054 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
5055 atomic_add_64(&arc_c, (int64_t)bytes);
5056 if (arc_c > arc_c_max)
5058 else if (state == arc_anon)
5059 atomic_add_64(&arc_p, (int64_t)bytes);
5063 ASSERT((int64_t)arc_p >= 0);
5067 * Check if arc_size has grown past our upper threshold, determined by
5068 * zfs_arc_overflow_shift.
5071 arc_is_overflowing(void)
5073 /* Always allow at least one block of overflow */
5074 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
5075 arc_c >> zfs_arc_overflow_shift);
5078 * We just compare the lower bound here for performance reasons. Our
5079 * primary goals are to make sure that the arc never grows without
5080 * bound, and that it can reach its maximum size. This check
5081 * accomplishes both goals. The maximum amount we could run over by is
5082 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
5083 * in the ARC. In practice, that's in the tens of MB, which is low
5084 * enough to be safe.
5086 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
5090 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5092 arc_buf_contents_t type = arc_buf_type(hdr);
5094 arc_get_data_impl(hdr, size, tag);
5095 if (type == ARC_BUFC_METADATA) {
5096 return (abd_alloc(size, B_TRUE));
5098 ASSERT(type == ARC_BUFC_DATA);
5099 return (abd_alloc(size, B_FALSE));
5104 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5106 arc_buf_contents_t type = arc_buf_type(hdr);
5108 arc_get_data_impl(hdr, size, tag);
5109 if (type == ARC_BUFC_METADATA) {
5110 return (zio_buf_alloc(size));
5112 ASSERT(type == ARC_BUFC_DATA);
5113 return (zio_data_buf_alloc(size));
5118 * Allocate a block and return it to the caller. If we are hitting the
5119 * hard limit for the cache size, we must sleep, waiting for the eviction
5120 * thread to catch up. If we're past the target size but below the hard
5121 * limit, we'll only signal the reclaim thread and continue on.
5124 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5126 arc_state_t *state = hdr->b_l1hdr.b_state;
5127 arc_buf_contents_t type = arc_buf_type(hdr);
5129 arc_adapt(size, state);
5132 * If arc_size is currently overflowing, and has grown past our
5133 * upper limit, we must be adding data faster than the evict
5134 * thread can evict. Thus, to ensure we don't compound the
5135 * problem by adding more data and forcing arc_size to grow even
5136 * further past it's target size, we halt and wait for the
5137 * eviction thread to catch up.
5139 * It's also possible that the reclaim thread is unable to evict
5140 * enough buffers to get arc_size below the overflow limit (e.g.
5141 * due to buffers being un-evictable, or hash lock collisions).
5142 * In this case, we want to proceed regardless if we're
5143 * overflowing; thus we don't use a while loop here.
5145 if (arc_is_overflowing()) {
5146 mutex_enter(&arc_reclaim_lock);
5149 * Now that we've acquired the lock, we may no longer be
5150 * over the overflow limit, lets check.
5152 * We're ignoring the case of spurious wake ups. If that
5153 * were to happen, it'd let this thread consume an ARC
5154 * buffer before it should have (i.e. before we're under
5155 * the overflow limit and were signalled by the reclaim
5156 * thread). As long as that is a rare occurrence, it
5157 * shouldn't cause any harm.
5159 if (arc_is_overflowing()) {
5160 cv_signal(&arc_reclaim_thread_cv);
5161 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
5164 mutex_exit(&arc_reclaim_lock);
5167 VERIFY3U(hdr->b_type, ==, type);
5168 if (type == ARC_BUFC_METADATA) {
5169 arc_space_consume(size, ARC_SPACE_META);
5171 arc_space_consume(size, ARC_SPACE_DATA);
5175 * Update the state size. Note that ghost states have a
5176 * "ghost size" and so don't need to be updated.
5178 if (!GHOST_STATE(state)) {
5180 (void) refcount_add_many(&state->arcs_size, size, tag);
5183 * If this is reached via arc_read, the link is
5184 * protected by the hash lock. If reached via
5185 * arc_buf_alloc, the header should not be accessed by
5186 * any other thread. And, if reached via arc_read_done,
5187 * the hash lock will protect it if it's found in the
5188 * hash table; otherwise no other thread should be
5189 * trying to [add|remove]_reference it.
5191 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5192 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5193 (void) refcount_add_many(&state->arcs_esize[type],
5198 * If we are growing the cache, and we are adding anonymous
5199 * data, and we have outgrown arc_p, update arc_p
5201 if (aggsum_compare(&arc_size, arc_c) < 0 &&
5202 hdr->b_l1hdr.b_state == arc_anon &&
5203 (refcount_count(&arc_anon->arcs_size) +
5204 refcount_count(&arc_mru->arcs_size) > arc_p))
5205 arc_p = MIN(arc_c, arc_p + size);
5207 ARCSTAT_BUMP(arcstat_allocated);
5211 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
5213 arc_free_data_impl(hdr, size, tag);
5218 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
5220 arc_buf_contents_t type = arc_buf_type(hdr);
5222 arc_free_data_impl(hdr, size, tag);
5223 if (type == ARC_BUFC_METADATA) {
5224 zio_buf_free(buf, size);
5226 ASSERT(type == ARC_BUFC_DATA);
5227 zio_data_buf_free(buf, size);
5232 * Free the arc data buffer.
5235 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5237 arc_state_t *state = hdr->b_l1hdr.b_state;
5238 arc_buf_contents_t type = arc_buf_type(hdr);
5240 /* protected by hash lock, if in the hash table */
5241 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5242 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5243 ASSERT(state != arc_anon && state != arc_l2c_only);
5245 (void) refcount_remove_many(&state->arcs_esize[type],
5248 (void) refcount_remove_many(&state->arcs_size, size, tag);
5250 VERIFY3U(hdr->b_type, ==, type);
5251 if (type == ARC_BUFC_METADATA) {
5252 arc_space_return(size, ARC_SPACE_META);
5254 ASSERT(type == ARC_BUFC_DATA);
5255 arc_space_return(size, ARC_SPACE_DATA);
5260 * This routine is called whenever a buffer is accessed.
5261 * NOTE: the hash lock is dropped in this function.
5264 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
5268 ASSERT(MUTEX_HELD(hash_lock));
5269 ASSERT(HDR_HAS_L1HDR(hdr));
5271 if (hdr->b_l1hdr.b_state == arc_anon) {
5273 * This buffer is not in the cache, and does not
5274 * appear in our "ghost" list. Add the new buffer
5278 ASSERT0(hdr->b_l1hdr.b_arc_access);
5279 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5280 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5281 arc_change_state(arc_mru, hdr, hash_lock);
5283 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5284 now = ddi_get_lbolt();
5287 * If this buffer is here because of a prefetch, then either:
5288 * - clear the flag if this is a "referencing" read
5289 * (any subsequent access will bump this into the MFU state).
5291 * - move the buffer to the head of the list if this is
5292 * another prefetch (to make it less likely to be evicted).
5294 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5295 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5296 /* link protected by hash lock */
5297 ASSERT(multilist_link_active(
5298 &hdr->b_l1hdr.b_arc_node));
5300 arc_hdr_clear_flags(hdr,
5302 ARC_FLAG_PRESCIENT_PREFETCH);
5303 ARCSTAT_BUMP(arcstat_mru_hits);
5305 hdr->b_l1hdr.b_arc_access = now;
5310 * This buffer has been "accessed" only once so far,
5311 * but it is still in the cache. Move it to the MFU
5314 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
5316 * More than 125ms have passed since we
5317 * instantiated this buffer. Move it to the
5318 * most frequently used state.
5320 hdr->b_l1hdr.b_arc_access = now;
5321 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5322 arc_change_state(arc_mfu, hdr, hash_lock);
5324 atomic_inc_32(&hdr->b_l1hdr.b_mru_hits);
5325 ARCSTAT_BUMP(arcstat_mru_hits);
5326 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5327 arc_state_t *new_state;
5329 * This buffer has been "accessed" recently, but
5330 * was evicted from the cache. Move it to the
5334 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5335 new_state = arc_mru;
5336 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
5337 arc_hdr_clear_flags(hdr,
5339 ARC_FLAG_PRESCIENT_PREFETCH);
5341 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5343 new_state = arc_mfu;
5344 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5347 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5348 arc_change_state(new_state, hdr, hash_lock);
5350 atomic_inc_32(&hdr->b_l1hdr.b_mru_ghost_hits);
5351 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5352 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5354 * This buffer has been accessed more than once and is
5355 * still in the cache. Keep it in the MFU state.
5357 * NOTE: an add_reference() that occurred when we did
5358 * the arc_read() will have kicked this off the list.
5359 * If it was a prefetch, we will explicitly move it to
5360 * the head of the list now.
5363 atomic_inc_32(&hdr->b_l1hdr.b_mfu_hits);
5364 ARCSTAT_BUMP(arcstat_mfu_hits);
5365 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5366 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5367 arc_state_t *new_state = arc_mfu;
5369 * This buffer has been accessed more than once but has
5370 * been evicted from the cache. Move it back to the
5374 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5376 * This is a prefetch access...
5377 * move this block back to the MRU state.
5379 new_state = arc_mru;
5382 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5383 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5384 arc_change_state(new_state, hdr, hash_lock);
5386 atomic_inc_32(&hdr->b_l1hdr.b_mfu_ghost_hits);
5387 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5388 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5390 * This buffer is on the 2nd Level ARC.
5393 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5394 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5395 arc_change_state(arc_mfu, hdr, hash_lock);
5397 ASSERT(!"invalid arc state");
5402 * This routine is called by dbuf_hold() to update the arc_access() state
5403 * which otherwise would be skipped for entries in the dbuf cache.
5406 arc_buf_access(arc_buf_t *buf)
5408 mutex_enter(&buf->b_evict_lock);
5409 arc_buf_hdr_t *hdr = buf->b_hdr;
5412 * Avoid taking the hash_lock when possible as an optimization.
5413 * The header must be checked again under the hash_lock in order
5414 * to handle the case where it is concurrently being released.
5416 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5417 mutex_exit(&buf->b_evict_lock);
5418 ARCSTAT_BUMP(arcstat_access_skip);
5422 kmutex_t *hash_lock = HDR_LOCK(hdr);
5423 mutex_enter(hash_lock);
5425 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5426 mutex_exit(hash_lock);
5427 mutex_exit(&buf->b_evict_lock);
5428 ARCSTAT_BUMP(arcstat_access_skip);
5432 mutex_exit(&buf->b_evict_lock);
5434 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5435 hdr->b_l1hdr.b_state == arc_mfu);
5437 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5438 arc_access(hdr, hash_lock);
5439 mutex_exit(hash_lock);
5441 ARCSTAT_BUMP(arcstat_hits);
5442 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5443 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5446 /* a generic arc_read_done_func_t which you can use */
5449 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5450 arc_buf_t *buf, void *arg)
5455 bcopy(buf->b_data, arg, arc_buf_size(buf));
5456 arc_buf_destroy(buf, arg);
5459 /* a generic arc_read_done_func_t */
5462 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5463 arc_buf_t *buf, void *arg)
5465 arc_buf_t **bufp = arg;
5467 ASSERT(zio == NULL || zio->io_error != 0);
5470 ASSERT(zio == NULL || zio->io_error == 0);
5472 ASSERT(buf->b_data != NULL);
5477 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5479 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5480 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5481 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
5483 if (HDR_COMPRESSION_ENABLED(hdr)) {
5484 ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
5485 BP_GET_COMPRESS(bp));
5487 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5488 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5493 arc_read_done(zio_t *zio)
5495 arc_buf_hdr_t *hdr = zio->io_private;
5496 kmutex_t *hash_lock = NULL;
5497 arc_callback_t *callback_list;
5498 arc_callback_t *acb;
5499 boolean_t freeable = B_FALSE;
5500 boolean_t no_zio_error = (zio->io_error == 0);
5503 * The hdr was inserted into hash-table and removed from lists
5504 * prior to starting I/O. We should find this header, since
5505 * it's in the hash table, and it should be legit since it's
5506 * not possible to evict it during the I/O. The only possible
5507 * reason for it not to be found is if we were freed during the
5510 if (HDR_IN_HASH_TABLE(hdr)) {
5511 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5512 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5513 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5514 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5515 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5517 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
5520 ASSERT((found == hdr &&
5521 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5522 (found == hdr && HDR_L2_READING(hdr)));
5523 ASSERT3P(hash_lock, !=, NULL);
5527 /* byteswap if necessary */
5528 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5529 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5530 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5532 hdr->b_l1hdr.b_byteswap =
5533 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5536 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5540 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5541 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5542 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5544 callback_list = hdr->b_l1hdr.b_acb;
5545 ASSERT3P(callback_list, !=, NULL);
5547 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) {
5549 * Only call arc_access on anonymous buffers. This is because
5550 * if we've issued an I/O for an evicted buffer, we've already
5551 * called arc_access (to prevent any simultaneous readers from
5552 * getting confused).
5554 arc_access(hdr, hash_lock);
5558 * If a read request has a callback (i.e. acb_done is not NULL), then we
5559 * make a buf containing the data according to the parameters which were
5560 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5561 * aren't needlessly decompressing the data multiple times.
5563 int callback_cnt = 0;
5564 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5571 int error = arc_buf_alloc_impl(hdr, acb->acb_private,
5572 acb->acb_compressed, zio->io_error == 0,
5576 * Decompression failed. Set io_error
5577 * so that when we call acb_done (below),
5578 * we will indicate that the read failed.
5579 * Note that in the unusual case where one
5580 * callback is compressed and another
5581 * uncompressed, we will mark all of them
5582 * as failed, even though the uncompressed
5583 * one can't actually fail. In this case,
5584 * the hdr will not be anonymous, because
5585 * if there are multiple callbacks, it's
5586 * because multiple threads found the same
5587 * arc buf in the hash table.
5589 zio->io_error = error;
5594 * If there are multiple callbacks, we must have the hash lock,
5595 * because the only way for multiple threads to find this hdr is
5596 * in the hash table. This ensures that if there are multiple
5597 * callbacks, the hdr is not anonymous. If it were anonymous,
5598 * we couldn't use arc_buf_destroy() in the error case below.
5600 ASSERT(callback_cnt < 2 || hash_lock != NULL);
5602 hdr->b_l1hdr.b_acb = NULL;
5603 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5604 if (callback_cnt == 0) {
5605 ASSERT(HDR_PREFETCH(hdr));
5606 ASSERT0(hdr->b_l1hdr.b_bufcnt);
5607 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5610 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5611 callback_list != NULL);
5614 arc_hdr_verify(hdr, zio->io_bp);
5616 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5617 if (hdr->b_l1hdr.b_state != arc_anon)
5618 arc_change_state(arc_anon, hdr, hash_lock);
5619 if (HDR_IN_HASH_TABLE(hdr))
5620 buf_hash_remove(hdr);
5621 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5625 * Broadcast before we drop the hash_lock to avoid the possibility
5626 * that the hdr (and hence the cv) might be freed before we get to
5627 * the cv_broadcast().
5629 cv_broadcast(&hdr->b_l1hdr.b_cv);
5631 if (hash_lock != NULL) {
5632 mutex_exit(hash_lock);
5635 * This block was freed while we waited for the read to
5636 * complete. It has been removed from the hash table and
5637 * moved to the anonymous state (so that it won't show up
5640 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5641 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5644 /* execute each callback and free its structure */
5645 while ((acb = callback_list) != NULL) {
5646 if (acb->acb_done != NULL) {
5647 if (zio->io_error != 0 && acb->acb_buf != NULL) {
5649 * If arc_buf_alloc_impl() fails during
5650 * decompression, the buf will still be
5651 * allocated, and needs to be freed here.
5653 arc_buf_destroy(acb->acb_buf, acb->acb_private);
5654 acb->acb_buf = NULL;
5656 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5657 acb->acb_buf, acb->acb_private);
5660 if (acb->acb_zio_dummy != NULL) {
5661 acb->acb_zio_dummy->io_error = zio->io_error;
5662 zio_nowait(acb->acb_zio_dummy);
5665 callback_list = acb->acb_next;
5666 kmem_free(acb, sizeof (arc_callback_t));
5670 arc_hdr_destroy(hdr);
5674 * "Read" the block at the specified DVA (in bp) via the
5675 * cache. If the block is found in the cache, invoke the provided
5676 * callback immediately and return. Note that the `zio' parameter
5677 * in the callback will be NULL in this case, since no IO was
5678 * required. If the block is not in the cache pass the read request
5679 * on to the spa with a substitute callback function, so that the
5680 * requested block will be added to the cache.
5682 * If a read request arrives for a block that has a read in-progress,
5683 * either wait for the in-progress read to complete (and return the
5684 * results); or, if this is a read with a "done" func, add a record
5685 * to the read to invoke the "done" func when the read completes,
5686 * and return; or just return.
5688 * arc_read_done() will invoke all the requested "done" functions
5689 * for readers of this block.
5692 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_read_done_func_t *done,
5693 void *private, zio_priority_t priority, int zio_flags,
5694 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5696 arc_buf_hdr_t *hdr = NULL;
5697 kmutex_t *hash_lock = NULL;
5699 uint64_t guid = spa_load_guid(spa);
5700 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0;
5703 ASSERT(!BP_IS_EMBEDDED(bp) ||
5704 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5707 if (!BP_IS_EMBEDDED(bp)) {
5709 * Embedded BP's have no DVA and require no I/O to "read".
5710 * Create an anonymous arc buf to back it.
5712 hdr = buf_hash_find(guid, bp, &hash_lock);
5715 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) {
5716 arc_buf_t *buf = NULL;
5717 *arc_flags |= ARC_FLAG_CACHED;
5719 if (HDR_IO_IN_PROGRESS(hdr)) {
5720 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
5722 ASSERT3P(head_zio, !=, NULL);
5723 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5724 priority == ZIO_PRIORITY_SYNC_READ) {
5726 * This is a sync read that needs to wait for
5727 * an in-flight async read. Request that the
5728 * zio have its priority upgraded.
5730 zio_change_priority(head_zio, priority);
5731 DTRACE_PROBE1(arc__async__upgrade__sync,
5732 arc_buf_hdr_t *, hdr);
5733 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
5735 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5736 arc_hdr_clear_flags(hdr,
5737 ARC_FLAG_PREDICTIVE_PREFETCH);
5740 if (*arc_flags & ARC_FLAG_WAIT) {
5741 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5742 mutex_exit(hash_lock);
5745 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5748 arc_callback_t *acb = NULL;
5750 acb = kmem_zalloc(sizeof (arc_callback_t),
5752 acb->acb_done = done;
5753 acb->acb_private = private;
5754 acb->acb_compressed = compressed_read;
5756 acb->acb_zio_dummy = zio_null(pio,
5757 spa, NULL, NULL, NULL, zio_flags);
5759 ASSERT3P(acb->acb_done, !=, NULL);
5760 acb->acb_zio_head = head_zio;
5761 acb->acb_next = hdr->b_l1hdr.b_acb;
5762 hdr->b_l1hdr.b_acb = acb;
5763 mutex_exit(hash_lock);
5766 mutex_exit(hash_lock);
5770 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5771 hdr->b_l1hdr.b_state == arc_mfu);
5774 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5776 * This is a demand read which does not have to
5777 * wait for i/o because we did a predictive
5778 * prefetch i/o for it, which has completed.
5781 arc__demand__hit__predictive__prefetch,
5782 arc_buf_hdr_t *, hdr);
5784 arcstat_demand_hit_predictive_prefetch);
5785 arc_hdr_clear_flags(hdr,
5786 ARC_FLAG_PREDICTIVE_PREFETCH);
5789 if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
5791 arcstat_demand_hit_prescient_prefetch);
5792 arc_hdr_clear_flags(hdr,
5793 ARC_FLAG_PRESCIENT_PREFETCH);
5796 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5797 /* Get a buf with the desired data in it. */
5798 rc = arc_buf_alloc_impl(hdr, private,
5799 compressed_read, B_TRUE, &buf);
5801 arc_buf_destroy(buf, private);
5804 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
5805 rc == 0 || rc != ENOENT);
5806 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
5807 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5808 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5810 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5811 arc_access(hdr, hash_lock);
5812 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5813 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5814 if (*arc_flags & ARC_FLAG_L2CACHE)
5815 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5816 mutex_exit(hash_lock);
5817 ARCSTAT_BUMP(arcstat_hits);
5818 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5819 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5820 data, metadata, hits);
5823 done(NULL, zb, bp, buf, private);
5825 uint64_t lsize = BP_GET_LSIZE(bp);
5826 uint64_t psize = BP_GET_PSIZE(bp);
5827 arc_callback_t *acb;
5830 boolean_t devw = B_FALSE;
5834 /* this block is not in the cache */
5835 arc_buf_hdr_t *exists = NULL;
5836 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5837 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5838 BP_GET_COMPRESS(bp), type);
5840 if (!BP_IS_EMBEDDED(bp)) {
5841 hdr->b_dva = *BP_IDENTITY(bp);
5842 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5843 exists = buf_hash_insert(hdr, &hash_lock);
5845 if (exists != NULL) {
5846 /* somebody beat us to the hash insert */
5847 mutex_exit(hash_lock);
5848 buf_discard_identity(hdr);
5849 arc_hdr_destroy(hdr);
5850 goto top; /* restart the IO request */
5854 * This block is in the ghost cache. If it was L2-only
5855 * (and thus didn't have an L1 hdr), we realloc the
5856 * header to add an L1 hdr.
5858 if (!HDR_HAS_L1HDR(hdr)) {
5859 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5862 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5863 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5864 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5865 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5866 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5867 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5870 * This is a delicate dance that we play here.
5871 * This hdr is in the ghost list so we access it
5872 * to move it out of the ghost list before we
5873 * initiate the read. If it's a prefetch then
5874 * it won't have a callback so we'll remove the
5875 * reference that arc_buf_alloc_impl() created. We
5876 * do this after we've called arc_access() to
5877 * avoid hitting an assert in remove_reference().
5879 arc_access(hdr, hash_lock);
5880 arc_hdr_alloc_pabd(hdr);
5882 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5883 size = arc_hdr_size(hdr);
5886 * If compression is enabled on the hdr, then will do
5887 * RAW I/O and will store the compressed data in the hdr's
5888 * data block. Otherwise, the hdr's data block will contain
5889 * the uncompressed data.
5891 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5892 zio_flags |= ZIO_FLAG_RAW;
5895 if (*arc_flags & ARC_FLAG_PREFETCH)
5896 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5897 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5898 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5900 if (*arc_flags & ARC_FLAG_L2CACHE)
5901 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5902 if (BP_GET_LEVEL(bp) > 0)
5903 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5904 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5905 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5906 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5908 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5909 acb->acb_done = done;
5910 acb->acb_private = private;
5911 acb->acb_compressed = compressed_read;
5913 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5914 hdr->b_l1hdr.b_acb = acb;
5915 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5917 if (HDR_HAS_L2HDR(hdr) &&
5918 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5919 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5920 addr = hdr->b_l2hdr.b_daddr;
5922 * Lock out L2ARC device removal.
5924 if (vdev_is_dead(vd) ||
5925 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5930 * We count both async reads and scrub IOs as asynchronous so
5931 * that both can be upgraded in the event of a cache hit while
5932 * the read IO is still in-flight.
5934 if (priority == ZIO_PRIORITY_ASYNC_READ ||
5935 priority == ZIO_PRIORITY_SCRUB)
5936 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5938 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5941 * At this point, we have a level 1 cache miss. Try again in
5942 * L2ARC if possible.
5944 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5946 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5947 uint64_t, lsize, zbookmark_phys_t *, zb);
5948 ARCSTAT_BUMP(arcstat_misses);
5949 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5950 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5951 data, metadata, misses);
5956 racct_add_force(curproc, RACCT_READBPS, size);
5957 racct_add_force(curproc, RACCT_READIOPS, 1);
5958 PROC_UNLOCK(curproc);
5961 curthread->td_ru.ru_inblock++;
5964 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5966 * Read from the L2ARC if the following are true:
5967 * 1. The L2ARC vdev was previously cached.
5968 * 2. This buffer still has L2ARC metadata.
5969 * 3. This buffer isn't currently writing to the L2ARC.
5970 * 4. The L2ARC entry wasn't evicted, which may
5971 * also have invalidated the vdev.
5972 * 5. This isn't prefetch and l2arc_noprefetch is set.
5974 if (HDR_HAS_L2HDR(hdr) &&
5975 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5976 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5977 l2arc_read_callback_t *cb;
5981 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5982 ARCSTAT_BUMP(arcstat_l2_hits);
5983 atomic_inc_32(&hdr->b_l2hdr.b_hits);
5985 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5987 cb->l2rcb_hdr = hdr;
5990 cb->l2rcb_flags = zio_flags;
5992 asize = vdev_psize_to_asize(vd, size);
5993 if (asize != size) {
5994 abd = abd_alloc_for_io(asize,
5995 HDR_ISTYPE_METADATA(hdr));
5996 cb->l2rcb_abd = abd;
5998 abd = hdr->b_l1hdr.b_pabd;
6001 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
6002 addr + asize <= vd->vdev_psize -
6003 VDEV_LABEL_END_SIZE);
6006 * l2arc read. The SCL_L2ARC lock will be
6007 * released by l2arc_read_done().
6008 * Issue a null zio if the underlying buffer
6009 * was squashed to zero size by compression.
6011 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
6012 ZIO_COMPRESS_EMPTY);
6013 rzio = zio_read_phys(pio, vd, addr,
6016 l2arc_read_done, cb, priority,
6017 zio_flags | ZIO_FLAG_DONT_CACHE |
6019 ZIO_FLAG_DONT_PROPAGATE |
6020 ZIO_FLAG_DONT_RETRY, B_FALSE);
6021 acb->acb_zio_head = rzio;
6023 if (hash_lock != NULL)
6024 mutex_exit(hash_lock);
6026 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
6028 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
6030 if (*arc_flags & ARC_FLAG_NOWAIT) {
6035 ASSERT(*arc_flags & ARC_FLAG_WAIT);
6036 if (zio_wait(rzio) == 0)
6039 /* l2arc read error; goto zio_read() */
6040 if (hash_lock != NULL)
6041 mutex_enter(hash_lock);
6043 DTRACE_PROBE1(l2arc__miss,
6044 arc_buf_hdr_t *, hdr);
6045 ARCSTAT_BUMP(arcstat_l2_misses);
6046 if (HDR_L2_WRITING(hdr))
6047 ARCSTAT_BUMP(arcstat_l2_rw_clash);
6048 spa_config_exit(spa, SCL_L2ARC, vd);
6052 spa_config_exit(spa, SCL_L2ARC, vd);
6053 if (l2arc_ndev != 0) {
6054 DTRACE_PROBE1(l2arc__miss,
6055 arc_buf_hdr_t *, hdr);
6056 ARCSTAT_BUMP(arcstat_l2_misses);
6060 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size,
6061 arc_read_done, hdr, priority, zio_flags, zb);
6062 acb->acb_zio_head = rzio;
6064 if (hash_lock != NULL)
6065 mutex_exit(hash_lock);
6067 if (*arc_flags & ARC_FLAG_WAIT)
6068 return (zio_wait(rzio));
6070 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6077 arc_add_prune_callback(arc_prune_func_t *func, void *private)
6081 p = kmem_alloc(sizeof (*p), KM_SLEEP);
6083 p->p_private = private;
6084 list_link_init(&p->p_node);
6085 refcount_create(&p->p_refcnt);
6087 mutex_enter(&arc_prune_mtx);
6088 refcount_add(&p->p_refcnt, &arc_prune_list);
6089 list_insert_head(&arc_prune_list, p);
6090 mutex_exit(&arc_prune_mtx);
6096 arc_remove_prune_callback(arc_prune_t *p)
6098 boolean_t wait = B_FALSE;
6099 mutex_enter(&arc_prune_mtx);
6100 list_remove(&arc_prune_list, p);
6101 if (refcount_remove(&p->p_refcnt, &arc_prune_list) > 0)
6103 mutex_exit(&arc_prune_mtx);
6105 /* wait for arc_prune_task to finish */
6107 taskq_wait(arc_prune_taskq);
6108 ASSERT0(refcount_count(&p->p_refcnt));
6109 refcount_destroy(&p->p_refcnt);
6110 kmem_free(p, sizeof (*p));
6114 * Notify the arc that a block was freed, and thus will never be used again.
6117 arc_freed(spa_t *spa, const blkptr_t *bp)
6120 kmutex_t *hash_lock;
6121 uint64_t guid = spa_load_guid(spa);
6123 ASSERT(!BP_IS_EMBEDDED(bp));
6125 hdr = buf_hash_find(guid, bp, &hash_lock);
6130 * We might be trying to free a block that is still doing I/O
6131 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
6132 * dmu_sync-ed block). If this block is being prefetched, then it
6133 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
6134 * until the I/O completes. A block may also have a reference if it is
6135 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6136 * have written the new block to its final resting place on disk but
6137 * without the dedup flag set. This would have left the hdr in the MRU
6138 * state and discoverable. When the txg finally syncs it detects that
6139 * the block was overridden in open context and issues an override I/O.
6140 * Since this is a dedup block, the override I/O will determine if the
6141 * block is already in the DDT. If so, then it will replace the io_bp
6142 * with the bp from the DDT and allow the I/O to finish. When the I/O
6143 * reaches the done callback, dbuf_write_override_done, it will
6144 * check to see if the io_bp and io_bp_override are identical.
6145 * If they are not, then it indicates that the bp was replaced with
6146 * the bp in the DDT and the override bp is freed. This allows
6147 * us to arrive here with a reference on a block that is being
6148 * freed. So if we have an I/O in progress, or a reference to
6149 * this hdr, then we don't destroy the hdr.
6151 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
6152 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
6153 arc_change_state(arc_anon, hdr, hash_lock);
6154 arc_hdr_destroy(hdr);
6155 mutex_exit(hash_lock);
6157 mutex_exit(hash_lock);
6163 * Release this buffer from the cache, making it an anonymous buffer. This
6164 * must be done after a read and prior to modifying the buffer contents.
6165 * If the buffer has more than one reference, we must make
6166 * a new hdr for the buffer.
6169 arc_release(arc_buf_t *buf, void *tag)
6171 arc_buf_hdr_t *hdr = buf->b_hdr;
6174 * It would be nice to assert that if it's DMU metadata (level >
6175 * 0 || it's the dnode file), then it must be syncing context.
6176 * But we don't know that information at this level.
6179 mutex_enter(&buf->b_evict_lock);
6181 ASSERT(HDR_HAS_L1HDR(hdr));
6184 * We don't grab the hash lock prior to this check, because if
6185 * the buffer's header is in the arc_anon state, it won't be
6186 * linked into the hash table.
6188 if (hdr->b_l1hdr.b_state == arc_anon) {
6189 mutex_exit(&buf->b_evict_lock);
6190 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6191 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6192 ASSERT(!HDR_HAS_L2HDR(hdr));
6193 ASSERT(HDR_EMPTY(hdr));
6194 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6195 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6196 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
6198 hdr->b_l1hdr.b_arc_access = 0;
6201 * If the buf is being overridden then it may already
6202 * have a hdr that is not empty.
6204 buf_discard_identity(hdr);
6210 kmutex_t *hash_lock = HDR_LOCK(hdr);
6211 mutex_enter(hash_lock);
6214 * This assignment is only valid as long as the hash_lock is
6215 * held, we must be careful not to reference state or the
6216 * b_state field after dropping the lock.
6218 arc_state_t *state = hdr->b_l1hdr.b_state;
6219 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6220 ASSERT3P(state, !=, arc_anon);
6222 /* this buffer is not on any list */
6223 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6225 if (HDR_HAS_L2HDR(hdr)) {
6226 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6229 * We have to recheck this conditional again now that
6230 * we're holding the l2ad_mtx to prevent a race with
6231 * another thread which might be concurrently calling
6232 * l2arc_evict(). In that case, l2arc_evict() might have
6233 * destroyed the header's L2 portion as we were waiting
6234 * to acquire the l2ad_mtx.
6236 if (HDR_HAS_L2HDR(hdr)) {
6238 arc_hdr_l2hdr_destroy(hdr);
6241 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6245 * Do we have more than one buf?
6247 if (hdr->b_l1hdr.b_bufcnt > 1) {
6248 arc_buf_hdr_t *nhdr;
6249 uint64_t spa = hdr->b_spa;
6250 uint64_t psize = HDR_GET_PSIZE(hdr);
6251 uint64_t lsize = HDR_GET_LSIZE(hdr);
6252 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
6253 arc_buf_contents_t type = arc_buf_type(hdr);
6254 VERIFY3U(hdr->b_type, ==, type);
6256 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6257 (void) remove_reference(hdr, hash_lock, tag);
6259 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
6260 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6261 ASSERT(ARC_BUF_LAST(buf));
6265 * Pull the data off of this hdr and attach it to
6266 * a new anonymous hdr. Also find the last buffer
6267 * in the hdr's buffer list.
6269 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6270 ASSERT3P(lastbuf, !=, NULL);
6273 * If the current arc_buf_t and the hdr are sharing their data
6274 * buffer, then we must stop sharing that block.
6276 if (arc_buf_is_shared(buf)) {
6277 VERIFY(!arc_buf_is_shared(lastbuf));
6280 * First, sever the block sharing relationship between
6281 * buf and the arc_buf_hdr_t.
6283 arc_unshare_buf(hdr, buf);
6286 * Now we need to recreate the hdr's b_pabd. Since we
6287 * have lastbuf handy, we try to share with it, but if
6288 * we can't then we allocate a new b_pabd and copy the
6289 * data from buf into it.
6291 if (arc_can_share(hdr, lastbuf)) {
6292 arc_share_buf(hdr, lastbuf);
6294 arc_hdr_alloc_pabd(hdr);
6295 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6296 buf->b_data, psize);
6298 VERIFY3P(lastbuf->b_data, !=, NULL);
6299 } else if (HDR_SHARED_DATA(hdr)) {
6301 * Uncompressed shared buffers are always at the end
6302 * of the list. Compressed buffers don't have the
6303 * same requirements. This makes it hard to
6304 * simply assert that the lastbuf is shared so
6305 * we rely on the hdr's compression flags to determine
6306 * if we have a compressed, shared buffer.
6308 ASSERT(arc_buf_is_shared(lastbuf) ||
6309 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
6310 ASSERT(!ARC_BUF_SHARED(buf));
6312 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
6313 ASSERT3P(state, !=, arc_l2c_only);
6315 (void) refcount_remove_many(&state->arcs_size,
6316 arc_buf_size(buf), buf);
6318 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6319 ASSERT3P(state, !=, arc_l2c_only);
6320 (void) refcount_remove_many(&state->arcs_esize[type],
6321 arc_buf_size(buf), buf);
6324 hdr->b_l1hdr.b_bufcnt -= 1;
6325 arc_cksum_verify(buf);
6327 arc_buf_unwatch(buf);
6330 mutex_exit(hash_lock);
6333 * Allocate a new hdr. The new hdr will contain a b_pabd
6334 * buffer which will be freed in arc_write().
6336 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
6337 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6338 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6339 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
6340 VERIFY3U(nhdr->b_type, ==, type);
6341 ASSERT(!HDR_SHARED_DATA(nhdr));
6343 nhdr->b_l1hdr.b_buf = buf;
6344 nhdr->b_l1hdr.b_bufcnt = 1;
6345 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6348 mutex_exit(&buf->b_evict_lock);
6349 (void) refcount_add_many(&arc_anon->arcs_size,
6350 arc_buf_size(buf), buf);
6352 mutex_exit(&buf->b_evict_lock);
6353 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6354 /* protected by hash lock, or hdr is on arc_anon */
6355 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6356 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6357 arc_change_state(arc_anon, hdr, hash_lock);
6358 hdr->b_l1hdr.b_arc_access = 0;
6359 mutex_exit(hash_lock);
6361 buf_discard_identity(hdr);
6367 arc_released(arc_buf_t *buf)
6371 mutex_enter(&buf->b_evict_lock);
6372 released = (buf->b_data != NULL &&
6373 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6374 mutex_exit(&buf->b_evict_lock);
6380 arc_referenced(arc_buf_t *buf)
6384 mutex_enter(&buf->b_evict_lock);
6385 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6386 mutex_exit(&buf->b_evict_lock);
6387 return (referenced);
6392 arc_write_ready(zio_t *zio)
6394 arc_write_callback_t *callback = zio->io_private;
6395 arc_buf_t *buf = callback->awcb_buf;
6396 arc_buf_hdr_t *hdr = buf->b_hdr;
6397 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
6399 ASSERT(HDR_HAS_L1HDR(hdr));
6400 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6401 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6404 * If we're reexecuting this zio because the pool suspended, then
6405 * cleanup any state that was previously set the first time the
6406 * callback was invoked.
6408 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6409 arc_cksum_free(hdr);
6411 arc_buf_unwatch(buf);
6413 if (hdr->b_l1hdr.b_pabd != NULL) {
6414 if (arc_buf_is_shared(buf)) {
6415 arc_unshare_buf(hdr, buf);
6417 arc_hdr_free_pabd(hdr);
6421 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6422 ASSERT(!HDR_SHARED_DATA(hdr));
6423 ASSERT(!arc_buf_is_shared(buf));
6425 callback->awcb_ready(zio, buf, callback->awcb_private);
6427 if (HDR_IO_IN_PROGRESS(hdr))
6428 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6430 arc_cksum_compute(buf);
6431 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6433 enum zio_compress compress;
6434 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6435 compress = ZIO_COMPRESS_OFF;
6437 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
6438 compress = BP_GET_COMPRESS(zio->io_bp);
6440 HDR_SET_PSIZE(hdr, psize);
6441 arc_hdr_set_compress(hdr, compress);
6445 * Fill the hdr with data. If the hdr is compressed, the data we want
6446 * is available from the zio, otherwise we can take it from the buf.
6448 * We might be able to share the buf's data with the hdr here. However,
6449 * doing so would cause the ARC to be full of linear ABDs if we write a
6450 * lot of shareable data. As a compromise, we check whether scattered
6451 * ABDs are allowed, and assume that if they are then the user wants
6452 * the ARC to be primarily filled with them regardless of the data being
6453 * written. Therefore, if they're allowed then we allocate one and copy
6454 * the data into it; otherwise, we share the data directly if we can.
6456 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
6457 arc_hdr_alloc_pabd(hdr);
6460 * Ideally, we would always copy the io_abd into b_pabd, but the
6461 * user may have disabled compressed ARC, thus we must check the
6462 * hdr's compression setting rather than the io_bp's.
6464 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
6465 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=,
6467 ASSERT3U(psize, >, 0);
6469 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6471 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6473 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6477 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6478 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6479 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6481 arc_share_buf(hdr, buf);
6484 arc_hdr_verify(hdr, zio->io_bp);
6488 arc_write_children_ready(zio_t *zio)
6490 arc_write_callback_t *callback = zio->io_private;
6491 arc_buf_t *buf = callback->awcb_buf;
6493 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6497 * The SPA calls this callback for each physical write that happens on behalf
6498 * of a logical write. See the comment in dbuf_write_physdone() for details.
6501 arc_write_physdone(zio_t *zio)
6503 arc_write_callback_t *cb = zio->io_private;
6504 if (cb->awcb_physdone != NULL)
6505 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
6509 arc_write_done(zio_t *zio)
6511 arc_write_callback_t *callback = zio->io_private;
6512 arc_buf_t *buf = callback->awcb_buf;
6513 arc_buf_hdr_t *hdr = buf->b_hdr;
6515 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6517 if (zio->io_error == 0) {
6518 arc_hdr_verify(hdr, zio->io_bp);
6520 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6521 buf_discard_identity(hdr);
6523 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6524 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6527 ASSERT(HDR_EMPTY(hdr));
6531 * If the block to be written was all-zero or compressed enough to be
6532 * embedded in the BP, no write was performed so there will be no
6533 * dva/birth/checksum. The buffer must therefore remain anonymous
6536 if (!HDR_EMPTY(hdr)) {
6537 arc_buf_hdr_t *exists;
6538 kmutex_t *hash_lock;
6540 ASSERT3U(zio->io_error, ==, 0);
6542 arc_cksum_verify(buf);
6544 exists = buf_hash_insert(hdr, &hash_lock);
6545 if (exists != NULL) {
6547 * This can only happen if we overwrite for
6548 * sync-to-convergence, because we remove
6549 * buffers from the hash table when we arc_free().
6551 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6552 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6553 panic("bad overwrite, hdr=%p exists=%p",
6554 (void *)hdr, (void *)exists);
6555 ASSERT(refcount_is_zero(
6556 &exists->b_l1hdr.b_refcnt));
6557 arc_change_state(arc_anon, exists, hash_lock);
6558 mutex_exit(hash_lock);
6559 arc_hdr_destroy(exists);
6560 exists = buf_hash_insert(hdr, &hash_lock);
6561 ASSERT3P(exists, ==, NULL);
6562 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6564 ASSERT(zio->io_prop.zp_nopwrite);
6565 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6566 panic("bad nopwrite, hdr=%p exists=%p",
6567 (void *)hdr, (void *)exists);
6570 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6571 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6572 ASSERT(BP_GET_DEDUP(zio->io_bp));
6573 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6576 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6577 /* if it's not anon, we are doing a scrub */
6578 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6579 arc_access(hdr, hash_lock);
6580 mutex_exit(hash_lock);
6582 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6585 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6586 callback->awcb_done(zio, buf, callback->awcb_private);
6588 abd_put(zio->io_abd);
6589 kmem_free(callback, sizeof (arc_write_callback_t));
6593 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
6594 boolean_t l2arc, const zio_prop_t *zp, arc_write_done_func_t *ready,
6595 arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
6596 arc_write_done_func_t *done, void *private, zio_priority_t priority,
6597 int zio_flags, const zbookmark_phys_t *zb)
6599 arc_buf_hdr_t *hdr = buf->b_hdr;
6600 arc_write_callback_t *callback;
6602 zio_prop_t localprop = *zp;
6604 ASSERT3P(ready, !=, NULL);
6605 ASSERT3P(done, !=, NULL);
6606 ASSERT(!HDR_IO_ERROR(hdr));
6607 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6608 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6609 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6611 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6612 if (ARC_BUF_COMPRESSED(buf)) {
6614 * We're writing a pre-compressed buffer. Make the
6615 * compression algorithm requested by the zio_prop_t match
6616 * the pre-compressed buffer's compression algorithm.
6618 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6620 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6621 zio_flags |= ZIO_FLAG_RAW;
6623 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6624 callback->awcb_ready = ready;
6625 callback->awcb_children_ready = children_ready;
6626 callback->awcb_physdone = physdone;
6627 callback->awcb_done = done;
6628 callback->awcb_private = private;
6629 callback->awcb_buf = buf;
6632 * The hdr's b_pabd is now stale, free it now. A new data block
6633 * will be allocated when the zio pipeline calls arc_write_ready().
6635 if (hdr->b_l1hdr.b_pabd != NULL) {
6637 * If the buf is currently sharing the data block with
6638 * the hdr then we need to break that relationship here.
6639 * The hdr will remain with a NULL data pointer and the
6640 * buf will take sole ownership of the block.
6642 if (arc_buf_is_shared(buf)) {
6643 arc_unshare_buf(hdr, buf);
6645 arc_hdr_free_pabd(hdr);
6647 VERIFY3P(buf->b_data, !=, NULL);
6648 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6650 ASSERT(!arc_buf_is_shared(buf));
6651 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6653 zio = zio_write(pio, spa, txg, bp,
6654 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6655 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6656 (children_ready != NULL) ? arc_write_children_ready : NULL,
6657 arc_write_physdone, arc_write_done, callback,
6658 priority, zio_flags, zb);
6664 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
6667 uint64_t available_memory = ptob(freemem);
6669 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
6670 available_memory = MIN(available_memory, uma_avail());
6673 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
6676 if (txg > spa->spa_lowmem_last_txg) {
6677 spa->spa_lowmem_last_txg = txg;
6678 spa->spa_lowmem_page_load = 0;
6681 * If we are in pageout, we know that memory is already tight,
6682 * the arc is already going to be evicting, so we just want to
6683 * continue to let page writes occur as quickly as possible.
6685 if (curproc == pageproc) {
6686 if (spa->spa_lowmem_page_load >
6687 MAX(ptob(minfree), available_memory) / 4)
6688 return (SET_ERROR(ERESTART));
6689 /* Note: reserve is inflated, so we deflate */
6690 atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8);
6692 } else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) {
6693 /* memory is low, delay before restarting */
6694 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
6695 return (SET_ERROR(EAGAIN));
6697 spa->spa_lowmem_page_load = 0;
6698 #endif /* _KERNEL */
6703 arc_tempreserve_clear(uint64_t reserve)
6705 atomic_add_64(&arc_tempreserve, -reserve);
6706 ASSERT((int64_t)arc_tempreserve >= 0);
6710 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
6715 if (reserve > arc_c/4 && !arc_no_grow) {
6716 arc_c = MIN(arc_c_max, reserve * 4);
6717 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
6719 if (reserve > arc_c)
6720 return (SET_ERROR(ENOMEM));
6723 * Don't count loaned bufs as in flight dirty data to prevent long
6724 * network delays from blocking transactions that are ready to be
6725 * assigned to a txg.
6728 /* assert that it has not wrapped around */
6729 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6731 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
6732 arc_loaned_bytes), 0);
6735 * Writes will, almost always, require additional memory allocations
6736 * in order to compress/encrypt/etc the data. We therefore need to
6737 * make sure that there is sufficient available memory for this.
6739 error = arc_memory_throttle(spa, reserve, txg);
6744 * Throttle writes when the amount of dirty data in the cache
6745 * gets too large. We try to keep the cache less than half full
6746 * of dirty blocks so that our sync times don't grow too large.
6748 * In the case of one pool being built on another pool, we want
6749 * to make sure we don't end up throttling the lower (backing)
6750 * pool when the upper pool is the majority contributor to dirty
6751 * data. To insure we make forward progress during throttling, we
6752 * also check the current pool's net dirty data and only throttle
6753 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
6754 * data in the cache.
6756 * Note: if two requests come in concurrently, we might let them
6757 * both succeed, when one of them should fail. Not a huge deal.
6759 uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
6760 uint64_t spa_dirty_anon = spa_dirty_data(spa);
6762 if (total_dirty > arc_c * zfs_arc_dirty_limit_percent / 100 &&
6763 anon_size > arc_c * zfs_arc_anon_limit_percent / 100 &&
6764 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
6765 uint64_t meta_esize =
6766 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6767 uint64_t data_esize =
6768 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6769 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6770 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6771 arc_tempreserve >> 10, meta_esize >> 10,
6772 data_esize >> 10, reserve >> 10, arc_c >> 10);
6773 return (SET_ERROR(ERESTART));
6775 atomic_add_64(&arc_tempreserve, reserve);
6780 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6781 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6783 size->value.ui64 = refcount_count(&state->arcs_size);
6784 evict_data->value.ui64 =
6785 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6786 evict_metadata->value.ui64 =
6787 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6791 arc_kstat_update(kstat_t *ksp, int rw)
6793 arc_stats_t *as = ksp->ks_data;
6795 if (rw == KSTAT_WRITE) {
6798 arc_kstat_update_state(arc_anon,
6799 &as->arcstat_anon_size,
6800 &as->arcstat_anon_evictable_data,
6801 &as->arcstat_anon_evictable_metadata);
6802 arc_kstat_update_state(arc_mru,
6803 &as->arcstat_mru_size,
6804 &as->arcstat_mru_evictable_data,
6805 &as->arcstat_mru_evictable_metadata);
6806 arc_kstat_update_state(arc_mru_ghost,
6807 &as->arcstat_mru_ghost_size,
6808 &as->arcstat_mru_ghost_evictable_data,
6809 &as->arcstat_mru_ghost_evictable_metadata);
6810 arc_kstat_update_state(arc_mfu,
6811 &as->arcstat_mfu_size,
6812 &as->arcstat_mfu_evictable_data,
6813 &as->arcstat_mfu_evictable_metadata);
6814 arc_kstat_update_state(arc_mfu_ghost,
6815 &as->arcstat_mfu_ghost_size,
6816 &as->arcstat_mfu_ghost_evictable_data,
6817 &as->arcstat_mfu_ghost_evictable_metadata);
6819 ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
6820 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
6821 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
6822 ARCSTAT(arcstat_metadata_size) =
6823 aggsum_value(&astat_metadata_size);
6824 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
6825 ARCSTAT(arcstat_bonus_size) = aggsum_value(&astat_bonus_size);
6826 ARCSTAT(arcstat_dnode_size) = aggsum_value(&astat_dnode_size);
6827 ARCSTAT(arcstat_dbuf_size) = aggsum_value(&astat_dbuf_size);
6828 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
6835 * This function *must* return indices evenly distributed between all
6836 * sublists of the multilist. This is needed due to how the ARC eviction
6837 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6838 * distributed between all sublists and uses this assumption when
6839 * deciding which sublist to evict from and how much to evict from it.
6842 arc_state_multilist_index_func(multilist_t *ml, void *obj)
6844 arc_buf_hdr_t *hdr = obj;
6847 * We rely on b_dva to generate evenly distributed index
6848 * numbers using buf_hash below. So, as an added precaution,
6849 * let's make sure we never add empty buffers to the arc lists.
6851 ASSERT(!HDR_EMPTY(hdr));
6854 * The assumption here, is the hash value for a given
6855 * arc_buf_hdr_t will remain constant throughout it's lifetime
6856 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
6857 * Thus, we don't need to store the header's sublist index
6858 * on insertion, as this index can be recalculated on removal.
6860 * Also, the low order bits of the hash value are thought to be
6861 * distributed evenly. Otherwise, in the case that the multilist
6862 * has a power of two number of sublists, each sublists' usage
6863 * would not be evenly distributed.
6865 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
6866 multilist_get_num_sublists(ml));
6870 static eventhandler_tag arc_event_lowmem = NULL;
6873 arc_lowmem(void *arg __unused, int howto __unused)
6876 mutex_enter(&arc_reclaim_lock);
6877 DTRACE_PROBE1(arc__needfree, int64_t, ((int64_t)freemem - zfs_arc_free_target) * PAGESIZE);
6878 cv_signal(&arc_reclaim_thread_cv);
6881 * It is unsafe to block here in arbitrary threads, because we can come
6882 * here from ARC itself and may hold ARC locks and thus risk a deadlock
6883 * with ARC reclaim thread.
6885 if (curproc == pageproc)
6886 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
6887 mutex_exit(&arc_reclaim_lock);
6892 arc_state_init(void)
6894 arc_anon = &ARC_anon;
6896 arc_mru_ghost = &ARC_mru_ghost;
6898 arc_mfu_ghost = &ARC_mfu_ghost;
6899 arc_l2c_only = &ARC_l2c_only;
6901 arc_mru->arcs_list[ARC_BUFC_METADATA] =
6902 multilist_create(sizeof (arc_buf_hdr_t),
6903 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6904 arc_state_multilist_index_func);
6905 arc_mru->arcs_list[ARC_BUFC_DATA] =
6906 multilist_create(sizeof (arc_buf_hdr_t),
6907 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6908 arc_state_multilist_index_func);
6909 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
6910 multilist_create(sizeof (arc_buf_hdr_t),
6911 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6912 arc_state_multilist_index_func);
6913 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
6914 multilist_create(sizeof (arc_buf_hdr_t),
6915 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6916 arc_state_multilist_index_func);
6917 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
6918 multilist_create(sizeof (arc_buf_hdr_t),
6919 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6920 arc_state_multilist_index_func);
6921 arc_mfu->arcs_list[ARC_BUFC_DATA] =
6922 multilist_create(sizeof (arc_buf_hdr_t),
6923 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6924 arc_state_multilist_index_func);
6925 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
6926 multilist_create(sizeof (arc_buf_hdr_t),
6927 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6928 arc_state_multilist_index_func);
6929 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
6930 multilist_create(sizeof (arc_buf_hdr_t),
6931 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6932 arc_state_multilist_index_func);
6933 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
6934 multilist_create(sizeof (arc_buf_hdr_t),
6935 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6936 arc_state_multilist_index_func);
6937 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
6938 multilist_create(sizeof (arc_buf_hdr_t),
6939 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6940 arc_state_multilist_index_func);
6942 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6943 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6944 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6945 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6946 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6947 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6948 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6949 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6950 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6951 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6952 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6953 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6955 refcount_create(&arc_anon->arcs_size);
6956 refcount_create(&arc_mru->arcs_size);
6957 refcount_create(&arc_mru_ghost->arcs_size);
6958 refcount_create(&arc_mfu->arcs_size);
6959 refcount_create(&arc_mfu_ghost->arcs_size);
6960 refcount_create(&arc_l2c_only->arcs_size);
6962 aggsum_init(&arc_meta_used, 0);
6963 aggsum_init(&arc_size, 0);
6964 aggsum_init(&astat_data_size, 0);
6965 aggsum_init(&astat_metadata_size, 0);
6966 aggsum_init(&astat_hdr_size, 0);
6967 aggsum_init(&astat_bonus_size, 0);
6968 aggsum_init(&astat_dnode_size, 0);
6969 aggsum_init(&astat_dbuf_size, 0);
6970 aggsum_init(&astat_l2_hdr_size, 0);
6974 arc_state_fini(void)
6976 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6977 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6978 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6979 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6980 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6981 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6982 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6983 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6984 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6985 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6986 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6987 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6989 refcount_destroy(&arc_anon->arcs_size);
6990 refcount_destroy(&arc_mru->arcs_size);
6991 refcount_destroy(&arc_mru_ghost->arcs_size);
6992 refcount_destroy(&arc_mfu->arcs_size);
6993 refcount_destroy(&arc_mfu_ghost->arcs_size);
6994 refcount_destroy(&arc_l2c_only->arcs_size);
6996 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
6997 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
6998 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
6999 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
7000 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
7001 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
7002 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
7003 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7015 int i, prefetch_tunable_set = 0;
7018 * allmem is "all memory that we could possibly use".
7022 uint64_t allmem = ptob(physmem - swapfs_minfree);
7024 uint64_t allmem = (physmem * PAGESIZE) / 2;
7027 uint64_t allmem = kmem_size();
7031 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
7032 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
7033 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
7035 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
7036 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL);
7038 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
7039 arc_c_min = MAX(allmem / 32, arc_abs_min);
7040 /* set max to 5/8 of all memory, or all but 1GB, whichever is more */
7041 if (allmem >= 1 << 30)
7042 arc_c_max = allmem - (1 << 30);
7044 arc_c_max = arc_c_min;
7045 arc_c_max = MAX(allmem * 5 / 8, arc_c_max);
7048 * In userland, there's only the memory pressure that we artificially
7049 * create (see arc_available_memory()). Don't let arc_c get too
7050 * small, because it can cause transactions to be larger than
7051 * arc_c, causing arc_tempreserve_space() to fail.
7054 arc_c_min = arc_c_max / 2;
7059 * Allow the tunables to override our calculations if they are
7062 if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) {
7063 arc_c_max = zfs_arc_max;
7064 arc_c_min = MIN(arc_c_min, arc_c_max);
7066 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
7067 arc_c_min = zfs_arc_min;
7071 arc_p = (arc_c >> 1);
7073 /* limit meta-data to 1/4 of the arc capacity */
7074 arc_meta_limit = arc_c_max / 4;
7078 * Metadata is stored in the kernel's heap. Don't let us
7079 * use more than half the heap for the ARC.
7082 arc_meta_limit = MIN(arc_meta_limit, uma_limit() / 2);
7083 arc_dnode_limit = arc_meta_limit / 10;
7085 arc_meta_limit = MIN(arc_meta_limit,
7086 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
7090 /* Allow the tunable to override if it is reasonable */
7091 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
7092 arc_meta_limit = zfs_arc_meta_limit;
7094 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
7095 arc_c_min = arc_meta_limit / 2;
7097 if (zfs_arc_meta_min > 0) {
7098 arc_meta_min = zfs_arc_meta_min;
7100 arc_meta_min = arc_c_min / 2;
7103 /* Valid range: <arc_meta_min> - <arc_c_max> */
7104 if ((zfs_arc_dnode_limit) && (zfs_arc_dnode_limit != arc_dnode_limit) &&
7105 (zfs_arc_dnode_limit >= zfs_arc_meta_min) &&
7106 (zfs_arc_dnode_limit <= arc_c_max))
7107 arc_dnode_limit = zfs_arc_dnode_limit;
7109 if (zfs_arc_grow_retry > 0)
7110 arc_grow_retry = zfs_arc_grow_retry;
7112 if (zfs_arc_shrink_shift > 0)
7113 arc_shrink_shift = zfs_arc_shrink_shift;
7115 if (zfs_arc_no_grow_shift > 0)
7116 arc_no_grow_shift = zfs_arc_no_grow_shift;
7118 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
7120 if (arc_no_grow_shift >= arc_shrink_shift)
7121 arc_no_grow_shift = arc_shrink_shift - 1;
7123 if (zfs_arc_p_min_shift > 0)
7124 arc_p_min_shift = zfs_arc_p_min_shift;
7126 /* if kmem_flags are set, lets try to use less memory */
7127 if (kmem_debugging())
7129 if (arc_c < arc_c_min)
7132 zfs_arc_min = arc_c_min;
7133 zfs_arc_max = arc_c_max;
7138 list_create(&arc_prune_list, sizeof (arc_prune_t),
7139 offsetof(arc_prune_t, p_node));
7140 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
7142 arc_prune_taskq = taskq_create("arc_prune", max_ncpus, minclsyspri,
7143 max_ncpus, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
7145 arc_reclaim_thread_exit = B_FALSE;
7146 arc_dnlc_evicts_thread_exit = FALSE;
7148 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
7149 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
7151 if (arc_ksp != NULL) {
7152 arc_ksp->ks_data = &arc_stats;
7153 arc_ksp->ks_update = arc_kstat_update;
7154 kstat_install(arc_ksp);
7157 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
7158 TS_RUN, minclsyspri);
7161 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
7162 EVENTHANDLER_PRI_FIRST);
7165 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0,
7166 TS_RUN, minclsyspri);
7172 * Calculate maximum amount of dirty data per pool.
7174 * If it has been set by /etc/system, take that.
7175 * Otherwise, use a percentage of physical memory defined by
7176 * zfs_dirty_data_max_percent (default 10%) with a cap at
7177 * zfs_dirty_data_max_max (default 4GB).
7179 if (zfs_dirty_data_max == 0) {
7180 zfs_dirty_data_max = ptob(physmem) *
7181 zfs_dirty_data_max_percent / 100;
7182 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
7183 zfs_dirty_data_max_max);
7187 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
7188 prefetch_tunable_set = 1;
7191 if (prefetch_tunable_set == 0) {
7192 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
7194 printf(" add \"vfs.zfs.prefetch_disable=0\" "
7195 "to /boot/loader.conf.\n");
7196 zfs_prefetch_disable = 1;
7199 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
7200 prefetch_tunable_set == 0) {
7201 printf("ZFS NOTICE: Prefetch is disabled by default if less "
7202 "than 4GB of RAM is present;\n"
7203 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
7204 "to /boot/loader.conf.\n");
7205 zfs_prefetch_disable = 1;
7208 /* Warn about ZFS memory and address space requirements. */
7209 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
7210 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
7211 "expect unstable behavior.\n");
7213 if (allmem < 512 * (1 << 20)) {
7214 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
7215 "expect unstable behavior.\n");
7216 printf(" Consider tuning vm.kmem_size and "
7217 "vm.kmem_size_max\n");
7218 printf(" in /boot/loader.conf.\n");
7229 if (arc_event_lowmem != NULL)
7230 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
7233 mutex_enter(&arc_reclaim_lock);
7234 arc_reclaim_thread_exit = B_TRUE;
7236 * The reclaim thread will set arc_reclaim_thread_exit back to
7237 * B_FALSE when it is finished exiting; we're waiting for that.
7239 while (arc_reclaim_thread_exit) {
7240 cv_signal(&arc_reclaim_thread_cv);
7241 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
7243 mutex_exit(&arc_reclaim_lock);
7245 /* Use B_TRUE to ensure *all* buffers are evicted */
7246 arc_flush(NULL, B_TRUE);
7248 mutex_enter(&arc_dnlc_evicts_lock);
7249 arc_dnlc_evicts_thread_exit = TRUE;
7251 * The user evicts thread will set arc_user_evicts_thread_exit
7252 * to FALSE when it is finished exiting; we're waiting for that.
7254 while (arc_dnlc_evicts_thread_exit) {
7255 cv_signal(&arc_dnlc_evicts_cv);
7256 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
7258 mutex_exit(&arc_dnlc_evicts_lock);
7262 if (arc_ksp != NULL) {
7263 kstat_delete(arc_ksp);
7267 taskq_wait(arc_prune_taskq);
7268 taskq_destroy(arc_prune_taskq);
7270 mutex_enter(&arc_prune_mtx);
7271 while ((p = list_head(&arc_prune_list)) != NULL) {
7272 list_remove(&arc_prune_list, p);
7273 refcount_remove(&p->p_refcnt, &arc_prune_list);
7274 refcount_destroy(&p->p_refcnt);
7275 kmem_free(p, sizeof (*p));
7277 mutex_exit(&arc_prune_mtx);
7279 list_destroy(&arc_prune_list);
7280 mutex_destroy(&arc_prune_mtx);
7281 mutex_destroy(&arc_reclaim_lock);
7282 cv_destroy(&arc_reclaim_thread_cv);
7283 cv_destroy(&arc_reclaim_waiters_cv);
7285 mutex_destroy(&arc_dnlc_evicts_lock);
7286 cv_destroy(&arc_dnlc_evicts_cv);
7291 ASSERT0(arc_loaned_bytes);
7297 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7298 * It uses dedicated storage devices to hold cached data, which are populated
7299 * using large infrequent writes. The main role of this cache is to boost
7300 * the performance of random read workloads. The intended L2ARC devices
7301 * include short-stroked disks, solid state disks, and other media with
7302 * substantially faster read latency than disk.
7304 * +-----------------------+
7306 * +-----------------------+
7309 * l2arc_feed_thread() arc_read()
7313 * +---------------+ |
7315 * +---------------+ |
7320 * +-------+ +-------+
7322 * | cache | | cache |
7323 * +-------+ +-------+
7324 * +=========+ .-----.
7325 * : L2ARC : |-_____-|
7326 * : devices : | Disks |
7327 * +=========+ `-_____-'
7329 * Read requests are satisfied from the following sources, in order:
7332 * 2) vdev cache of L2ARC devices
7334 * 4) vdev cache of disks
7337 * Some L2ARC device types exhibit extremely slow write performance.
7338 * To accommodate for this there are some significant differences between
7339 * the L2ARC and traditional cache design:
7341 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
7342 * the ARC behave as usual, freeing buffers and placing headers on ghost
7343 * lists. The ARC does not send buffers to the L2ARC during eviction as
7344 * this would add inflated write latencies for all ARC memory pressure.
7346 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7347 * It does this by periodically scanning buffers from the eviction-end of
7348 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7349 * not already there. It scans until a headroom of buffers is satisfied,
7350 * which itself is a buffer for ARC eviction. If a compressible buffer is
7351 * found during scanning and selected for writing to an L2ARC device, we
7352 * temporarily boost scanning headroom during the next scan cycle to make
7353 * sure we adapt to compression effects (which might significantly reduce
7354 * the data volume we write to L2ARC). The thread that does this is
7355 * l2arc_feed_thread(), illustrated below; example sizes are included to
7356 * provide a better sense of ratio than this diagram:
7359 * +---------------------+----------+
7360 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
7361 * +---------------------+----------+ | o L2ARC eligible
7362 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
7363 * +---------------------+----------+ |
7364 * 15.9 Gbytes ^ 32 Mbytes |
7366 * l2arc_feed_thread()
7368 * l2arc write hand <--[oooo]--'
7372 * +==============================+
7373 * L2ARC dev |####|#|###|###| |####| ... |
7374 * +==============================+
7377 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
7378 * evicted, then the L2ARC has cached a buffer much sooner than it probably
7379 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
7380 * safe to say that this is an uncommon case, since buffers at the end of
7381 * the ARC lists have moved there due to inactivity.
7383 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
7384 * then the L2ARC simply misses copying some buffers. This serves as a
7385 * pressure valve to prevent heavy read workloads from both stalling the ARC
7386 * with waits and clogging the L2ARC with writes. This also helps prevent
7387 * the potential for the L2ARC to churn if it attempts to cache content too
7388 * quickly, such as during backups of the entire pool.
7390 * 5. After system boot and before the ARC has filled main memory, there are
7391 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
7392 * lists can remain mostly static. Instead of searching from tail of these
7393 * lists as pictured, the l2arc_feed_thread() will search from the list heads
7394 * for eligible buffers, greatly increasing its chance of finding them.
7396 * The L2ARC device write speed is also boosted during this time so that
7397 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
7398 * there are no L2ARC reads, and no fear of degrading read performance
7399 * through increased writes.
7401 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
7402 * the vdev queue can aggregate them into larger and fewer writes. Each
7403 * device is written to in a rotor fashion, sweeping writes through
7404 * available space then repeating.
7406 * 7. The L2ARC does not store dirty content. It never needs to flush
7407 * write buffers back to disk based storage.
7409 * 8. If an ARC buffer is written (and dirtied) which also exists in the
7410 * L2ARC, the now stale L2ARC buffer is immediately dropped.
7412 * The performance of the L2ARC can be tweaked by a number of tunables, which
7413 * may be necessary for different workloads:
7415 * l2arc_write_max max write bytes per interval
7416 * l2arc_write_boost extra write bytes during device warmup
7417 * l2arc_noprefetch skip caching prefetched buffers
7418 * l2arc_headroom number of max device writes to precache
7419 * l2arc_headroom_boost when we find compressed buffers during ARC
7420 * scanning, we multiply headroom by this
7421 * percentage factor for the next scan cycle,
7422 * since more compressed buffers are likely to
7424 * l2arc_feed_secs seconds between L2ARC writing
7426 * Tunables may be removed or added as future performance improvements are
7427 * integrated, and also may become zpool properties.
7429 * There are three key functions that control how the L2ARC warms up:
7431 * l2arc_write_eligible() check if a buffer is eligible to cache
7432 * l2arc_write_size() calculate how much to write
7433 * l2arc_write_interval() calculate sleep delay between writes
7435 * These three functions determine what to write, how much, and how quickly
7440 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
7443 * A buffer is *not* eligible for the L2ARC if it:
7444 * 1. belongs to a different spa.
7445 * 2. is already cached on the L2ARC.
7446 * 3. has an I/O in progress (it may be an incomplete read).
7447 * 4. is flagged not eligible (zfs property).
7449 if (hdr->b_spa != spa_guid) {
7450 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
7453 if (HDR_HAS_L2HDR(hdr)) {
7454 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
7457 if (HDR_IO_IN_PROGRESS(hdr)) {
7458 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
7461 if (!HDR_L2CACHE(hdr)) {
7462 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
7470 l2arc_write_size(void)
7475 * Make sure our globals have meaningful values in case the user
7478 size = l2arc_write_max;
7480 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
7481 "be greater than zero, resetting it to the default (%d)",
7483 size = l2arc_write_max = L2ARC_WRITE_SIZE;
7486 if (arc_warm == B_FALSE)
7487 size += l2arc_write_boost;
7494 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
7496 clock_t interval, next, now;
7499 * If the ARC lists are busy, increase our write rate; if the
7500 * lists are stale, idle back. This is achieved by checking
7501 * how much we previously wrote - if it was more than half of
7502 * what we wanted, schedule the next write much sooner.
7504 if (l2arc_feed_again && wrote > (wanted / 2))
7505 interval = (hz * l2arc_feed_min_ms) / 1000;
7507 interval = hz * l2arc_feed_secs;
7509 now = ddi_get_lbolt();
7510 next = MAX(now, MIN(now + interval, began + interval));
7516 * Cycle through L2ARC devices. This is how L2ARC load balances.
7517 * If a device is returned, this also returns holding the spa config lock.
7519 static l2arc_dev_t *
7520 l2arc_dev_get_next(void)
7522 l2arc_dev_t *first, *next = NULL;
7525 * Lock out the removal of spas (spa_namespace_lock), then removal
7526 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
7527 * both locks will be dropped and a spa config lock held instead.
7529 mutex_enter(&spa_namespace_lock);
7530 mutex_enter(&l2arc_dev_mtx);
7532 /* if there are no vdevs, there is nothing to do */
7533 if (l2arc_ndev == 0)
7537 next = l2arc_dev_last;
7539 /* loop around the list looking for a non-faulted vdev */
7541 next = list_head(l2arc_dev_list);
7543 next = list_next(l2arc_dev_list, next);
7545 next = list_head(l2arc_dev_list);
7548 /* if we have come back to the start, bail out */
7551 else if (next == first)
7554 } while (vdev_is_dead(next->l2ad_vdev));
7556 /* if we were unable to find any usable vdevs, return NULL */
7557 if (vdev_is_dead(next->l2ad_vdev))
7560 l2arc_dev_last = next;
7563 mutex_exit(&l2arc_dev_mtx);
7566 * Grab the config lock to prevent the 'next' device from being
7567 * removed while we are writing to it.
7570 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
7571 mutex_exit(&spa_namespace_lock);
7577 * Free buffers that were tagged for destruction.
7580 l2arc_do_free_on_write()
7583 l2arc_data_free_t *df, *df_prev;
7585 mutex_enter(&l2arc_free_on_write_mtx);
7586 buflist = l2arc_free_on_write;
7588 for (df = list_tail(buflist); df; df = df_prev) {
7589 df_prev = list_prev(buflist, df);
7590 ASSERT3P(df->l2df_abd, !=, NULL);
7591 abd_free(df->l2df_abd);
7592 list_remove(buflist, df);
7593 kmem_free(df, sizeof (l2arc_data_free_t));
7596 mutex_exit(&l2arc_free_on_write_mtx);
7600 * A write to a cache device has completed. Update all headers to allow
7601 * reads from these buffers to begin.
7604 l2arc_write_done(zio_t *zio)
7606 l2arc_write_callback_t *cb;
7609 arc_buf_hdr_t *head, *hdr, *hdr_prev;
7610 kmutex_t *hash_lock;
7611 int64_t bytes_dropped = 0;
7613 cb = zio->io_private;
7614 ASSERT3P(cb, !=, NULL);
7615 dev = cb->l2wcb_dev;
7616 ASSERT3P(dev, !=, NULL);
7617 head = cb->l2wcb_head;
7618 ASSERT3P(head, !=, NULL);
7619 buflist = &dev->l2ad_buflist;
7620 ASSERT3P(buflist, !=, NULL);
7621 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
7622 l2arc_write_callback_t *, cb);
7624 if (zio->io_error != 0)
7625 ARCSTAT_BUMP(arcstat_l2_writes_error);
7628 * All writes completed, or an error was hit.
7631 mutex_enter(&dev->l2ad_mtx);
7632 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
7633 hdr_prev = list_prev(buflist, hdr);
7635 hash_lock = HDR_LOCK(hdr);
7638 * We cannot use mutex_enter or else we can deadlock
7639 * with l2arc_write_buffers (due to swapping the order
7640 * the hash lock and l2ad_mtx are taken).
7642 if (!mutex_tryenter(hash_lock)) {
7644 * Missed the hash lock. We must retry so we
7645 * don't leave the ARC_FLAG_L2_WRITING bit set.
7647 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
7650 * We don't want to rescan the headers we've
7651 * already marked as having been written out, so
7652 * we reinsert the head node so we can pick up
7653 * where we left off.
7655 list_remove(buflist, head);
7656 list_insert_after(buflist, hdr, head);
7658 mutex_exit(&dev->l2ad_mtx);
7661 * We wait for the hash lock to become available
7662 * to try and prevent busy waiting, and increase
7663 * the chance we'll be able to acquire the lock
7664 * the next time around.
7666 mutex_enter(hash_lock);
7667 mutex_exit(hash_lock);
7672 * We could not have been moved into the arc_l2c_only
7673 * state while in-flight due to our ARC_FLAG_L2_WRITING
7674 * bit being set. Let's just ensure that's being enforced.
7676 ASSERT(HDR_HAS_L1HDR(hdr));
7678 if (zio->io_error != 0) {
7680 * Error - drop L2ARC entry.
7682 list_remove(buflist, hdr);
7684 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
7686 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr));
7687 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
7689 bytes_dropped += arc_hdr_size(hdr);
7690 (void) refcount_remove_many(&dev->l2ad_alloc,
7691 arc_hdr_size(hdr), hdr);
7695 * Allow ARC to begin reads and ghost list evictions to
7698 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
7700 mutex_exit(hash_lock);
7703 atomic_inc_64(&l2arc_writes_done);
7704 list_remove(buflist, head);
7705 ASSERT(!HDR_HAS_L1HDR(head));
7706 kmem_cache_free(hdr_l2only_cache, head);
7707 mutex_exit(&dev->l2ad_mtx);
7709 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
7711 l2arc_do_free_on_write();
7713 kmem_free(cb, sizeof (l2arc_write_callback_t));
7717 * A read to a cache device completed. Validate buffer contents before
7718 * handing over to the regular ARC routines.
7721 l2arc_read_done(zio_t *zio)
7723 l2arc_read_callback_t *cb;
7725 kmutex_t *hash_lock;
7726 boolean_t valid_cksum;
7728 ASSERT3P(zio->io_vd, !=, NULL);
7729 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
7731 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
7733 cb = zio->io_private;
7734 ASSERT3P(cb, !=, NULL);
7735 hdr = cb->l2rcb_hdr;
7736 ASSERT3P(hdr, !=, NULL);
7738 hash_lock = HDR_LOCK(hdr);
7739 mutex_enter(hash_lock);
7740 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
7743 * If the data was read into a temporary buffer,
7744 * move it and free the buffer.
7746 if (cb->l2rcb_abd != NULL) {
7747 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
7748 if (zio->io_error == 0) {
7749 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd,
7754 * The following must be done regardless of whether
7755 * there was an error:
7756 * - free the temporary buffer
7757 * - point zio to the real ARC buffer
7758 * - set zio size accordingly
7759 * These are required because zio is either re-used for
7760 * an I/O of the block in the case of the error
7761 * or the zio is passed to arc_read_done() and it
7764 abd_free(cb->l2rcb_abd);
7765 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
7766 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
7769 ASSERT3P(zio->io_abd, !=, NULL);
7772 * Check this survived the L2ARC journey.
7774 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd);
7775 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
7776 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
7778 valid_cksum = arc_cksum_is_equal(hdr, zio);
7779 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
7780 mutex_exit(hash_lock);
7781 zio->io_private = hdr;
7784 mutex_exit(hash_lock);
7786 * Buffer didn't survive caching. Increment stats and
7787 * reissue to the original storage device.
7789 if (zio->io_error != 0) {
7790 ARCSTAT_BUMP(arcstat_l2_io_error);
7792 zio->io_error = SET_ERROR(EIO);
7795 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
7798 * If there's no waiter, issue an async i/o to the primary
7799 * storage now. If there *is* a waiter, the caller must
7800 * issue the i/o in a context where it's OK to block.
7802 if (zio->io_waiter == NULL) {
7803 zio_t *pio = zio_unique_parent(zio);
7805 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
7807 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
7808 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done,
7809 hdr, zio->io_priority, cb->l2rcb_flags,
7814 kmem_free(cb, sizeof (l2arc_read_callback_t));
7818 * This is the list priority from which the L2ARC will search for pages to
7819 * cache. This is used within loops (0..3) to cycle through lists in the
7820 * desired order. This order can have a significant effect on cache
7823 * Currently the metadata lists are hit first, MFU then MRU, followed by
7824 * the data lists. This function returns a locked list, and also returns
7827 static multilist_sublist_t *
7828 l2arc_sublist_lock(int list_num)
7830 multilist_t *ml = NULL;
7833 ASSERT(list_num >= 0 && list_num <= 3);
7837 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
7840 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
7843 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
7846 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
7851 * Return a randomly-selected sublist. This is acceptable
7852 * because the caller feeds only a little bit of data for each
7853 * call (8MB). Subsequent calls will result in different
7854 * sublists being selected.
7856 idx = multilist_get_random_index(ml);
7857 return (multilist_sublist_lock(ml, idx));
7861 * Evict buffers from the device write hand to the distance specified in
7862 * bytes. This distance may span populated buffers, it may span nothing.
7863 * This is clearing a region on the L2ARC device ready for writing.
7864 * If the 'all' boolean is set, every buffer is evicted.
7867 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
7870 arc_buf_hdr_t *hdr, *hdr_prev;
7871 kmutex_t *hash_lock;
7874 buflist = &dev->l2ad_buflist;
7876 if (!all && dev->l2ad_first) {
7878 * This is the first sweep through the device. There is
7884 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
7886 * When nearing the end of the device, evict to the end
7887 * before the device write hand jumps to the start.
7889 taddr = dev->l2ad_end;
7891 taddr = dev->l2ad_hand + distance;
7893 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
7894 uint64_t, taddr, boolean_t, all);
7897 mutex_enter(&dev->l2ad_mtx);
7898 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
7899 hdr_prev = list_prev(buflist, hdr);
7901 hash_lock = HDR_LOCK(hdr);
7904 * We cannot use mutex_enter or else we can deadlock
7905 * with l2arc_write_buffers (due to swapping the order
7906 * the hash lock and l2ad_mtx are taken).
7908 if (!mutex_tryenter(hash_lock)) {
7910 * Missed the hash lock. Retry.
7912 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
7913 mutex_exit(&dev->l2ad_mtx);
7914 mutex_enter(hash_lock);
7915 mutex_exit(hash_lock);
7920 * A header can't be on this list if it doesn't have L2 header.
7922 ASSERT(HDR_HAS_L2HDR(hdr));
7924 /* Ensure this header has finished being written. */
7925 ASSERT(!HDR_L2_WRITING(hdr));
7926 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
7928 if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
7929 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
7931 * We've evicted to the target address,
7932 * or the end of the device.
7934 mutex_exit(hash_lock);
7938 if (!HDR_HAS_L1HDR(hdr)) {
7939 ASSERT(!HDR_L2_READING(hdr));
7941 * This doesn't exist in the ARC. Destroy.
7942 * arc_hdr_destroy() will call list_remove()
7943 * and decrement arcstat_l2_lsize.
7945 arc_change_state(arc_anon, hdr, hash_lock);
7946 arc_hdr_destroy(hdr);
7948 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
7949 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
7951 * Invalidate issued or about to be issued
7952 * reads, since we may be about to write
7953 * over this location.
7955 if (HDR_L2_READING(hdr)) {
7956 ARCSTAT_BUMP(arcstat_l2_evict_reading);
7957 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
7960 arc_hdr_l2hdr_destroy(hdr);
7962 mutex_exit(hash_lock);
7964 mutex_exit(&dev->l2ad_mtx);
7968 * Find and write ARC buffers to the L2ARC device.
7970 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7971 * for reading until they have completed writing.
7972 * The headroom_boost is an in-out parameter used to maintain headroom boost
7973 * state between calls to this function.
7975 * Returns the number of bytes actually written (which may be smaller than
7976 * the delta by which the device hand has changed due to alignment).
7979 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
7981 arc_buf_hdr_t *hdr, *hdr_prev, *head;
7982 uint64_t write_asize, write_psize, write_lsize, headroom;
7984 l2arc_write_callback_t *cb;
7986 uint64_t guid = spa_load_guid(spa);
7989 ASSERT3P(dev->l2ad_vdev, !=, NULL);
7992 write_lsize = write_asize = write_psize = 0;
7994 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
7995 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
7997 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
7999 * Copy buffers for L2ARC writing.
8001 for (try = 0; try <= 3; try++) {
8002 multilist_sublist_t *mls = l2arc_sublist_lock(try);
8003 uint64_t passed_sz = 0;
8005 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
8008 * L2ARC fast warmup.
8010 * Until the ARC is warm and starts to evict, read from the
8011 * head of the ARC lists rather than the tail.
8013 if (arc_warm == B_FALSE)
8014 hdr = multilist_sublist_head(mls);
8016 hdr = multilist_sublist_tail(mls);
8018 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
8020 headroom = target_sz * l2arc_headroom;
8021 if (zfs_compressed_arc_enabled)
8022 headroom = (headroom * l2arc_headroom_boost) / 100;
8024 for (; hdr; hdr = hdr_prev) {
8025 kmutex_t *hash_lock;
8027 if (arc_warm == B_FALSE)
8028 hdr_prev = multilist_sublist_next(mls, hdr);
8030 hdr_prev = multilist_sublist_prev(mls, hdr);
8031 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned,
8032 HDR_GET_LSIZE(hdr));
8034 hash_lock = HDR_LOCK(hdr);
8035 if (!mutex_tryenter(hash_lock)) {
8036 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
8038 * Skip this buffer rather than waiting.
8043 passed_sz += HDR_GET_LSIZE(hdr);
8044 if (passed_sz > headroom) {
8048 mutex_exit(hash_lock);
8049 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
8053 if (!l2arc_write_eligible(guid, hdr)) {
8054 mutex_exit(hash_lock);
8059 * We rely on the L1 portion of the header below, so
8060 * it's invalid for this header to have been evicted out
8061 * of the ghost cache, prior to being written out. The
8062 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8064 ASSERT(HDR_HAS_L1HDR(hdr));
8066 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8067 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8068 ASSERT3U(arc_hdr_size(hdr), >, 0);
8069 uint64_t psize = arc_hdr_size(hdr);
8070 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
8073 if ((write_asize + asize) > target_sz) {
8075 mutex_exit(hash_lock);
8076 ARCSTAT_BUMP(arcstat_l2_write_full);
8082 * Insert a dummy header on the buflist so
8083 * l2arc_write_done() can find where the
8084 * write buffers begin without searching.
8086 mutex_enter(&dev->l2ad_mtx);
8087 list_insert_head(&dev->l2ad_buflist, head);
8088 mutex_exit(&dev->l2ad_mtx);
8091 sizeof (l2arc_write_callback_t), KM_SLEEP);
8092 cb->l2wcb_dev = dev;
8093 cb->l2wcb_head = head;
8094 pio = zio_root(spa, l2arc_write_done, cb,
8096 ARCSTAT_BUMP(arcstat_l2_write_pios);
8099 hdr->b_l2hdr.b_dev = dev;
8100 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
8101 arc_hdr_set_flags(hdr,
8102 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
8104 mutex_enter(&dev->l2ad_mtx);
8105 list_insert_head(&dev->l2ad_buflist, hdr);
8106 mutex_exit(&dev->l2ad_mtx);
8108 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr);
8111 * Normally the L2ARC can use the hdr's data, but if
8112 * we're sharing data between the hdr and one of its
8113 * bufs, L2ARC needs its own copy of the data so that
8114 * the ZIO below can't race with the buf consumer.
8115 * Another case where we need to create a copy of the
8116 * data is when the buffer size is not device-aligned
8117 * and we need to pad the block to make it such.
8118 * That also keeps the clock hand suitably aligned.
8120 * To ensure that the copy will be available for the
8121 * lifetime of the ZIO and be cleaned up afterwards, we
8122 * add it to the l2arc_free_on_write queue.
8125 if (!HDR_SHARED_DATA(hdr) && psize == asize) {
8126 to_write = hdr->b_l1hdr.b_pabd;
8128 to_write = abd_alloc_for_io(asize,
8129 HDR_ISTYPE_METADATA(hdr));
8130 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
8131 if (asize != psize) {
8132 abd_zero_off(to_write, psize,
8135 l2arc_free_abd_on_write(to_write, asize,
8138 wzio = zio_write_phys(pio, dev->l2ad_vdev,
8139 hdr->b_l2hdr.b_daddr, asize, to_write,
8140 ZIO_CHECKSUM_OFF, NULL, hdr,
8141 ZIO_PRIORITY_ASYNC_WRITE,
8142 ZIO_FLAG_CANFAIL, B_FALSE);
8144 write_lsize += HDR_GET_LSIZE(hdr);
8145 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
8148 write_psize += psize;
8149 write_asize += asize;
8150 dev->l2ad_hand += asize;
8152 mutex_exit(hash_lock);
8154 (void) zio_nowait(wzio);
8157 multilist_sublist_unlock(mls);
8163 /* No buffers selected for writing? */
8165 ASSERT0(write_lsize);
8166 ASSERT(!HDR_HAS_L1HDR(head));
8167 kmem_cache_free(hdr_l2only_cache, head);
8171 ASSERT3U(write_psize, <=, target_sz);
8172 ARCSTAT_BUMP(arcstat_l2_writes_sent);
8173 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
8174 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
8175 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
8176 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
8179 * Bump device hand to the device start if it is approaching the end.
8180 * l2arc_evict() will already have evicted ahead for this case.
8182 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
8183 dev->l2ad_hand = dev->l2ad_start;
8184 dev->l2ad_first = B_FALSE;
8187 dev->l2ad_writing = B_TRUE;
8188 (void) zio_wait(pio);
8189 dev->l2ad_writing = B_FALSE;
8191 return (write_asize);
8195 * This thread feeds the L2ARC at regular intervals. This is the beating
8196 * heart of the L2ARC.
8200 l2arc_feed_thread(void *unused __unused)
8205 uint64_t size, wrote;
8206 clock_t begin, next = ddi_get_lbolt();
8208 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
8210 mutex_enter(&l2arc_feed_thr_lock);
8212 while (l2arc_thread_exit == 0) {
8213 CALLB_CPR_SAFE_BEGIN(&cpr);
8214 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
8215 next - ddi_get_lbolt());
8216 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
8217 next = ddi_get_lbolt() + hz;
8220 * Quick check for L2ARC devices.
8222 mutex_enter(&l2arc_dev_mtx);
8223 if (l2arc_ndev == 0) {
8224 mutex_exit(&l2arc_dev_mtx);
8227 mutex_exit(&l2arc_dev_mtx);
8228 begin = ddi_get_lbolt();
8231 * This selects the next l2arc device to write to, and in
8232 * doing so the next spa to feed from: dev->l2ad_spa. This
8233 * will return NULL if there are now no l2arc devices or if
8234 * they are all faulted.
8236 * If a device is returned, its spa's config lock is also
8237 * held to prevent device removal. l2arc_dev_get_next()
8238 * will grab and release l2arc_dev_mtx.
8240 if ((dev = l2arc_dev_get_next()) == NULL)
8243 spa = dev->l2ad_spa;
8244 ASSERT3P(spa, !=, NULL);
8247 * If the pool is read-only then force the feed thread to
8248 * sleep a little longer.
8250 if (!spa_writeable(spa)) {
8251 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
8252 spa_config_exit(spa, SCL_L2ARC, dev);
8257 * Avoid contributing to memory pressure.
8259 if (arc_reclaim_needed()) {
8260 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
8261 spa_config_exit(spa, SCL_L2ARC, dev);
8265 ARCSTAT_BUMP(arcstat_l2_feeds);
8267 size = l2arc_write_size();
8270 * Evict L2ARC buffers that will be overwritten.
8272 l2arc_evict(dev, size, B_FALSE);
8275 * Write ARC buffers.
8277 wrote = l2arc_write_buffers(spa, dev, size);
8280 * Calculate interval between writes.
8282 next = l2arc_write_interval(begin, size, wrote);
8283 spa_config_exit(spa, SCL_L2ARC, dev);
8286 l2arc_thread_exit = 0;
8287 cv_broadcast(&l2arc_feed_thr_cv);
8288 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
8293 l2arc_vdev_present(vdev_t *vd)
8297 mutex_enter(&l2arc_dev_mtx);
8298 for (dev = list_head(l2arc_dev_list); dev != NULL;
8299 dev = list_next(l2arc_dev_list, dev)) {
8300 if (dev->l2ad_vdev == vd)
8303 mutex_exit(&l2arc_dev_mtx);
8305 return (dev != NULL);
8309 * Add a vdev for use by the L2ARC. By this point the spa has already
8310 * validated the vdev and opened it.
8313 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
8315 l2arc_dev_t *adddev;
8317 ASSERT(!l2arc_vdev_present(vd));
8319 vdev_ashift_optimize(vd);
8322 * Create a new l2arc device entry.
8324 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
8325 adddev->l2ad_spa = spa;
8326 adddev->l2ad_vdev = vd;
8327 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
8328 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
8329 adddev->l2ad_hand = adddev->l2ad_start;
8330 adddev->l2ad_first = B_TRUE;
8331 adddev->l2ad_writing = B_FALSE;
8333 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
8335 * This is a list of all ARC buffers that are still valid on the
8338 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
8339 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
8341 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
8342 refcount_create(&adddev->l2ad_alloc);
8345 * Add device to global list
8347 mutex_enter(&l2arc_dev_mtx);
8348 list_insert_head(l2arc_dev_list, adddev);
8349 atomic_inc_64(&l2arc_ndev);
8350 mutex_exit(&l2arc_dev_mtx);
8354 * Remove a vdev from the L2ARC.
8357 l2arc_remove_vdev(vdev_t *vd)
8359 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
8362 * Find the device by vdev
8364 mutex_enter(&l2arc_dev_mtx);
8365 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
8366 nextdev = list_next(l2arc_dev_list, dev);
8367 if (vd == dev->l2ad_vdev) {
8372 ASSERT3P(remdev, !=, NULL);
8375 * Remove device from global list
8377 list_remove(l2arc_dev_list, remdev);
8378 l2arc_dev_last = NULL; /* may have been invalidated */
8379 atomic_dec_64(&l2arc_ndev);
8380 mutex_exit(&l2arc_dev_mtx);
8383 * Clear all buflists and ARC references. L2ARC device flush.
8385 l2arc_evict(remdev, 0, B_TRUE);
8386 list_destroy(&remdev->l2ad_buflist);
8387 mutex_destroy(&remdev->l2ad_mtx);
8388 refcount_destroy(&remdev->l2ad_alloc);
8389 kmem_free(remdev, sizeof (l2arc_dev_t));
8395 l2arc_thread_exit = 0;
8397 l2arc_writes_sent = 0;
8398 l2arc_writes_done = 0;
8400 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
8401 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
8402 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
8403 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
8405 l2arc_dev_list = &L2ARC_dev_list;
8406 l2arc_free_on_write = &L2ARC_free_on_write;
8407 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
8408 offsetof(l2arc_dev_t, l2ad_node));
8409 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
8410 offsetof(l2arc_data_free_t, l2df_list_node));
8417 * This is called from dmu_fini(), which is called from spa_fini();
8418 * Because of this, we can assume that all l2arc devices have
8419 * already been removed when the pools themselves were removed.
8422 l2arc_do_free_on_write();
8424 mutex_destroy(&l2arc_feed_thr_lock);
8425 cv_destroy(&l2arc_feed_thr_cv);
8426 mutex_destroy(&l2arc_dev_mtx);
8427 mutex_destroy(&l2arc_free_on_write_mtx);
8429 list_destroy(l2arc_dev_list);
8430 list_destroy(l2arc_free_on_write);
8436 if (!(spa_mode_global & FWRITE))
8439 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
8440 TS_RUN, minclsyspri);
8446 if (!(spa_mode_global & FWRITE))
8449 mutex_enter(&l2arc_feed_thr_lock);
8450 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
8451 l2arc_thread_exit = 1;
8452 while (l2arc_thread_exit != 0)
8453 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
8454 mutex_exit(&l2arc_feed_thr_lock);