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_ONLY;
542 int zfs_arc_meta_adjust_restarts = 4096;
544 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_meta_strategy, CTLFLAG_RWTUN,
545 &zfs_arc_meta_strategy, 0,
546 "ARC metadata reclamation strategy "
547 "(0 = metadata only, 1 = balance data and metadata)");
550 static arc_state_t ARC_anon;
551 static arc_state_t ARC_mru;
552 static arc_state_t ARC_mru_ghost;
553 static arc_state_t ARC_mfu;
554 static arc_state_t ARC_mfu_ghost;
555 static arc_state_t ARC_l2c_only;
557 typedef struct arc_stats {
558 kstat_named_t arcstat_hits;
559 kstat_named_t arcstat_misses;
560 kstat_named_t arcstat_demand_data_hits;
561 kstat_named_t arcstat_demand_data_misses;
562 kstat_named_t arcstat_demand_metadata_hits;
563 kstat_named_t arcstat_demand_metadata_misses;
564 kstat_named_t arcstat_prefetch_data_hits;
565 kstat_named_t arcstat_prefetch_data_misses;
566 kstat_named_t arcstat_prefetch_metadata_hits;
567 kstat_named_t arcstat_prefetch_metadata_misses;
568 kstat_named_t arcstat_mru_hits;
569 kstat_named_t arcstat_mru_ghost_hits;
570 kstat_named_t arcstat_mfu_hits;
571 kstat_named_t arcstat_mfu_ghost_hits;
572 kstat_named_t arcstat_allocated;
573 kstat_named_t arcstat_deleted;
575 * Number of buffers that could not be evicted because the hash lock
576 * was held by another thread. The lock may not necessarily be held
577 * by something using the same buffer, since hash locks are shared
578 * by multiple buffers.
580 kstat_named_t arcstat_mutex_miss;
582 * Number of buffers skipped when updating the access state due to the
583 * header having already been released after acquiring the hash lock.
585 kstat_named_t arcstat_access_skip;
587 * Number of buffers skipped because they have I/O in progress, are
588 * indirect prefetch buffers that have not lived long enough, or are
589 * not from the spa we're trying to evict from.
591 kstat_named_t arcstat_evict_skip;
593 * Number of times arc_evict_state() was unable to evict enough
594 * buffers to reach it's target amount.
596 kstat_named_t arcstat_evict_not_enough;
597 kstat_named_t arcstat_evict_l2_cached;
598 kstat_named_t arcstat_evict_l2_eligible;
599 kstat_named_t arcstat_evict_l2_ineligible;
600 kstat_named_t arcstat_evict_l2_skip;
601 kstat_named_t arcstat_hash_elements;
602 kstat_named_t arcstat_hash_elements_max;
603 kstat_named_t arcstat_hash_collisions;
604 kstat_named_t arcstat_hash_chains;
605 kstat_named_t arcstat_hash_chain_max;
606 kstat_named_t arcstat_p;
607 kstat_named_t arcstat_c;
608 kstat_named_t arcstat_c_min;
609 kstat_named_t arcstat_c_max;
610 /* Not updated directly; only synced in arc_kstat_update. */
611 kstat_named_t arcstat_size;
613 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
614 * Note that the compressed bytes may match the uncompressed bytes
615 * if the block is either not compressed or compressed arc is disabled.
617 kstat_named_t arcstat_compressed_size;
619 * Uncompressed size of the data stored in b_pabd. If compressed
620 * arc is disabled then this value will be identical to the stat
623 kstat_named_t arcstat_uncompressed_size;
625 * Number of bytes stored in all the arc_buf_t's. This is classified
626 * as "overhead" since this data is typically short-lived and will
627 * be evicted from the arc when it becomes unreferenced unless the
628 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
629 * values have been set (see comment in dbuf.c for more information).
631 kstat_named_t arcstat_overhead_size;
633 * Number of bytes consumed by internal ARC structures necessary
634 * for tracking purposes; these structures are not actually
635 * backed by ARC buffers. This includes arc_buf_hdr_t structures
636 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
637 * caches), and arc_buf_t structures (allocated via arc_buf_t
639 * Not updated directly; only synced in arc_kstat_update.
641 kstat_named_t arcstat_hdr_size;
643 * Number of bytes consumed by ARC buffers of type equal to
644 * ARC_BUFC_DATA. This is generally consumed by buffers backing
645 * on disk user data (e.g. plain file contents).
646 * Not updated directly; only synced in arc_kstat_update.
648 kstat_named_t arcstat_data_size;
650 * Number of bytes consumed by ARC buffers of type equal to
651 * ARC_BUFC_METADATA. This is generally consumed by buffers
652 * backing on disk data that is used for internal ZFS
653 * structures (e.g. ZAP, dnode, indirect blocks, etc).
654 * Not updated directly; only synced in arc_kstat_update.
656 kstat_named_t arcstat_metadata_size;
658 * Number of bytes consumed by dmu_buf_impl_t objects.
660 kstat_named_t arcstat_dbuf_size;
662 * Number of bytes consumed by dnode_t objects.
664 kstat_named_t arcstat_dnode_size;
666 * Number of bytes consumed by bonus buffers.
668 kstat_named_t arcstat_bonus_size;
669 #if defined(__FreeBSD__) && defined(COMPAT_FREEBSD11)
671 * Sum of the previous three counters, provided for compatibility.
673 kstat_named_t arcstat_other_size;
676 * Total number of bytes consumed by ARC buffers residing in the
677 * arc_anon state. This includes *all* buffers in the arc_anon
678 * state; e.g. data, metadata, evictable, and unevictable buffers
679 * are all included in this value.
680 * Not updated directly; only synced in arc_kstat_update.
682 kstat_named_t arcstat_anon_size;
684 * Number of bytes consumed by ARC buffers that meet the
685 * following criteria: backing buffers of type ARC_BUFC_DATA,
686 * residing in the arc_anon state, and are eligible for eviction
687 * (e.g. have no outstanding holds on the buffer).
688 * Not updated directly; only synced in arc_kstat_update.
690 kstat_named_t arcstat_anon_evictable_data;
692 * Number of bytes consumed by ARC buffers that meet the
693 * following criteria: backing buffers of type ARC_BUFC_METADATA,
694 * residing in the arc_anon state, and are eligible for eviction
695 * (e.g. have no outstanding holds on the buffer).
696 * Not updated directly; only synced in arc_kstat_update.
698 kstat_named_t arcstat_anon_evictable_metadata;
700 * Total number of bytes consumed by ARC buffers residing in the
701 * arc_mru state. This includes *all* buffers in the arc_mru
702 * state; e.g. data, metadata, evictable, and unevictable buffers
703 * are all included in this value.
704 * Not updated directly; only synced in arc_kstat_update.
706 kstat_named_t arcstat_mru_size;
708 * Number of bytes consumed by ARC buffers that meet the
709 * following criteria: backing buffers of type ARC_BUFC_DATA,
710 * residing in the arc_mru state, and are eligible for eviction
711 * (e.g. have no outstanding holds on the buffer).
712 * Not updated directly; only synced in arc_kstat_update.
714 kstat_named_t arcstat_mru_evictable_data;
716 * Number of bytes consumed by ARC buffers that meet the
717 * following criteria: backing buffers of type ARC_BUFC_METADATA,
718 * residing in the arc_mru state, and are eligible for eviction
719 * (e.g. have no outstanding holds on the buffer).
720 * Not updated directly; only synced in arc_kstat_update.
722 kstat_named_t arcstat_mru_evictable_metadata;
724 * Total number of bytes that *would have been* consumed by ARC
725 * buffers in the arc_mru_ghost state. The key thing to note
726 * here, is the fact that this size doesn't actually indicate
727 * RAM consumption. The ghost lists only consist of headers and
728 * don't actually have ARC buffers linked off of these headers.
729 * Thus, *if* the headers had associated ARC buffers, these
730 * buffers *would have* consumed this number of bytes.
731 * Not updated directly; only synced in arc_kstat_update.
733 kstat_named_t arcstat_mru_ghost_size;
735 * Number of bytes that *would have been* consumed by ARC
736 * buffers that are eligible for eviction, of type
737 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
738 * Not updated directly; only synced in arc_kstat_update.
740 kstat_named_t arcstat_mru_ghost_evictable_data;
742 * Number of bytes that *would have been* consumed by ARC
743 * buffers that are eligible for eviction, of type
744 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
745 * Not updated directly; only synced in arc_kstat_update.
747 kstat_named_t arcstat_mru_ghost_evictable_metadata;
749 * Total number of bytes consumed by ARC buffers residing in the
750 * arc_mfu state. This includes *all* buffers in the arc_mfu
751 * state; e.g. data, metadata, evictable, and unevictable buffers
752 * are all included in this value.
753 * Not updated directly; only synced in arc_kstat_update.
755 kstat_named_t arcstat_mfu_size;
757 * Number of bytes consumed by ARC buffers that are eligible for
758 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
760 * Not updated directly; only synced in arc_kstat_update.
762 kstat_named_t arcstat_mfu_evictable_data;
764 * Number of bytes consumed by ARC buffers that are eligible for
765 * eviction, of type ARC_BUFC_METADATA, and reside in the
767 * Not updated directly; only synced in arc_kstat_update.
769 kstat_named_t arcstat_mfu_evictable_metadata;
771 * Total number of bytes that *would have been* consumed by ARC
772 * buffers in the arc_mfu_ghost state. See the comment above
773 * arcstat_mru_ghost_size for more details.
774 * Not updated directly; only synced in arc_kstat_update.
776 kstat_named_t arcstat_mfu_ghost_size;
778 * Number of bytes that *would have been* consumed by ARC
779 * buffers that are eligible for eviction, of type
780 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
781 * Not updated directly; only synced in arc_kstat_update.
783 kstat_named_t arcstat_mfu_ghost_evictable_data;
785 * Number of bytes that *would have been* consumed by ARC
786 * buffers that are eligible for eviction, of type
787 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
788 * Not updated directly; only synced in arc_kstat_update.
790 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
791 kstat_named_t arcstat_l2_hits;
792 kstat_named_t arcstat_l2_misses;
793 kstat_named_t arcstat_l2_feeds;
794 kstat_named_t arcstat_l2_rw_clash;
795 kstat_named_t arcstat_l2_read_bytes;
796 kstat_named_t arcstat_l2_write_bytes;
797 kstat_named_t arcstat_l2_writes_sent;
798 kstat_named_t arcstat_l2_writes_done;
799 kstat_named_t arcstat_l2_writes_error;
800 kstat_named_t arcstat_l2_writes_lock_retry;
801 kstat_named_t arcstat_l2_evict_lock_retry;
802 kstat_named_t arcstat_l2_evict_reading;
803 kstat_named_t arcstat_l2_evict_l1cached;
804 kstat_named_t arcstat_l2_free_on_write;
805 kstat_named_t arcstat_l2_abort_lowmem;
806 kstat_named_t arcstat_l2_cksum_bad;
807 kstat_named_t arcstat_l2_io_error;
808 kstat_named_t arcstat_l2_lsize;
809 kstat_named_t arcstat_l2_psize;
810 /* Not updated directly; only synced in arc_kstat_update. */
811 kstat_named_t arcstat_l2_hdr_size;
812 kstat_named_t arcstat_l2_write_trylock_fail;
813 kstat_named_t arcstat_l2_write_passed_headroom;
814 kstat_named_t arcstat_l2_write_spa_mismatch;
815 kstat_named_t arcstat_l2_write_in_l2;
816 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
817 kstat_named_t arcstat_l2_write_not_cacheable;
818 kstat_named_t arcstat_l2_write_full;
819 kstat_named_t arcstat_l2_write_buffer_iter;
820 kstat_named_t arcstat_l2_write_pios;
821 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
822 kstat_named_t arcstat_l2_write_buffer_list_iter;
823 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
824 kstat_named_t arcstat_memory_throttle_count;
825 kstat_named_t arcstat_memory_direct_count;
826 kstat_named_t arcstat_memory_indirect_count;
827 kstat_named_t arcstat_memory_all_bytes;
828 kstat_named_t arcstat_memory_free_bytes;
829 kstat_named_t arcstat_memory_available_bytes;
830 kstat_named_t arcstat_no_grow;
831 kstat_named_t arcstat_tempreserve;
832 kstat_named_t arcstat_loaned_bytes;
833 kstat_named_t arcstat_prune;
834 /* Not updated directly; only synced in arc_kstat_update. */
835 kstat_named_t arcstat_meta_used;
836 kstat_named_t arcstat_meta_limit;
837 kstat_named_t arcstat_dnode_limit;
838 kstat_named_t arcstat_meta_max;
839 kstat_named_t arcstat_meta_min;
840 kstat_named_t arcstat_async_upgrade_sync;
841 kstat_named_t arcstat_demand_hit_predictive_prefetch;
842 kstat_named_t arcstat_demand_hit_prescient_prefetch;
845 static arc_stats_t arc_stats = {
846 { "hits", KSTAT_DATA_UINT64 },
847 { "misses", KSTAT_DATA_UINT64 },
848 { "demand_data_hits", KSTAT_DATA_UINT64 },
849 { "demand_data_misses", KSTAT_DATA_UINT64 },
850 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
851 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
852 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
853 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
854 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
855 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
856 { "mru_hits", KSTAT_DATA_UINT64 },
857 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
858 { "mfu_hits", KSTAT_DATA_UINT64 },
859 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
860 { "allocated", KSTAT_DATA_UINT64 },
861 { "deleted", KSTAT_DATA_UINT64 },
862 { "mutex_miss", KSTAT_DATA_UINT64 },
863 { "access_skip", KSTAT_DATA_UINT64 },
864 { "evict_skip", KSTAT_DATA_UINT64 },
865 { "evict_not_enough", KSTAT_DATA_UINT64 },
866 { "evict_l2_cached", KSTAT_DATA_UINT64 },
867 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
868 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
869 { "evict_l2_skip", KSTAT_DATA_UINT64 },
870 { "hash_elements", KSTAT_DATA_UINT64 },
871 { "hash_elements_max", KSTAT_DATA_UINT64 },
872 { "hash_collisions", KSTAT_DATA_UINT64 },
873 { "hash_chains", KSTAT_DATA_UINT64 },
874 { "hash_chain_max", KSTAT_DATA_UINT64 },
875 { "p", KSTAT_DATA_UINT64 },
876 { "c", KSTAT_DATA_UINT64 },
877 { "c_min", KSTAT_DATA_UINT64 },
878 { "c_max", KSTAT_DATA_UINT64 },
879 { "size", KSTAT_DATA_UINT64 },
880 { "compressed_size", KSTAT_DATA_UINT64 },
881 { "uncompressed_size", KSTAT_DATA_UINT64 },
882 { "overhead_size", KSTAT_DATA_UINT64 },
883 { "hdr_size", KSTAT_DATA_UINT64 },
884 { "data_size", KSTAT_DATA_UINT64 },
885 { "metadata_size", KSTAT_DATA_UINT64 },
886 { "dbuf_size", KSTAT_DATA_UINT64 },
887 { "dnode_size", KSTAT_DATA_UINT64 },
888 { "bonus_size", KSTAT_DATA_UINT64 },
889 #if defined(__FreeBSD__) && defined(COMPAT_FREEBSD11)
890 { "other_size", KSTAT_DATA_UINT64 },
892 { "anon_size", KSTAT_DATA_UINT64 },
893 { "anon_evictable_data", KSTAT_DATA_UINT64 },
894 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
895 { "mru_size", KSTAT_DATA_UINT64 },
896 { "mru_evictable_data", KSTAT_DATA_UINT64 },
897 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
898 { "mru_ghost_size", KSTAT_DATA_UINT64 },
899 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
900 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
901 { "mfu_size", KSTAT_DATA_UINT64 },
902 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
903 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
904 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
905 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
906 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
907 { "l2_hits", KSTAT_DATA_UINT64 },
908 { "l2_misses", KSTAT_DATA_UINT64 },
909 { "l2_feeds", KSTAT_DATA_UINT64 },
910 { "l2_rw_clash", KSTAT_DATA_UINT64 },
911 { "l2_read_bytes", KSTAT_DATA_UINT64 },
912 { "l2_write_bytes", KSTAT_DATA_UINT64 },
913 { "l2_writes_sent", KSTAT_DATA_UINT64 },
914 { "l2_writes_done", KSTAT_DATA_UINT64 },
915 { "l2_writes_error", KSTAT_DATA_UINT64 },
916 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
917 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
918 { "l2_evict_reading", KSTAT_DATA_UINT64 },
919 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
920 { "l2_free_on_write", KSTAT_DATA_UINT64 },
921 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
922 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
923 { "l2_io_error", KSTAT_DATA_UINT64 },
924 { "l2_size", KSTAT_DATA_UINT64 },
925 { "l2_asize", KSTAT_DATA_UINT64 },
926 { "l2_hdr_size", KSTAT_DATA_UINT64 },
927 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
928 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
929 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
930 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
931 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
932 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
933 { "l2_write_full", KSTAT_DATA_UINT64 },
934 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
935 { "l2_write_pios", KSTAT_DATA_UINT64 },
936 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
937 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
938 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
939 { "memory_throttle_count", KSTAT_DATA_UINT64 },
940 { "memory_direct_count", KSTAT_DATA_UINT64 },
941 { "memory_indirect_count", KSTAT_DATA_UINT64 },
942 { "memory_all_bytes", KSTAT_DATA_UINT64 },
943 { "memory_free_bytes", KSTAT_DATA_UINT64 },
944 { "memory_available_bytes", KSTAT_DATA_UINT64 },
945 { "arc_no_grow", KSTAT_DATA_UINT64 },
946 { "arc_tempreserve", KSTAT_DATA_UINT64 },
947 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
948 { "arc_prune", KSTAT_DATA_UINT64 },
949 { "arc_meta_used", KSTAT_DATA_UINT64 },
950 { "arc_meta_limit", KSTAT_DATA_UINT64 },
951 { "arc_dnode_limit", KSTAT_DATA_UINT64 },
952 { "arc_meta_max", KSTAT_DATA_UINT64 },
953 { "arc_meta_min", KSTAT_DATA_UINT64 },
954 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
955 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
956 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
959 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
961 #define ARCSTAT_INCR(stat, val) \
962 atomic_add_64(&arc_stats.stat.value.ui64, (val))
964 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
965 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
967 #define ARCSTAT_MAX(stat, val) { \
969 while ((val) > (m = arc_stats.stat.value.ui64) && \
970 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
974 #define ARCSTAT_MAXSTAT(stat) \
975 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
978 * We define a macro to allow ARC hits/misses to be easily broken down by
979 * two separate conditions, giving a total of four different subtypes for
980 * each of hits and misses (so eight statistics total).
982 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
985 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
987 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
991 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
993 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
998 static arc_state_t *arc_anon;
999 static arc_state_t *arc_mru;
1000 static arc_state_t *arc_mru_ghost;
1001 static arc_state_t *arc_mfu;
1002 static arc_state_t *arc_mfu_ghost;
1003 static arc_state_t *arc_l2c_only;
1006 * There are several ARC variables that are critical to export as kstats --
1007 * but we don't want to have to grovel around in the kstat whenever we wish to
1008 * manipulate them. For these variables, we therefore define them to be in
1009 * terms of the statistic variable. This assures that we are not introducing
1010 * the possibility of inconsistency by having shadow copies of the variables,
1011 * while still allowing the code to be readable.
1013 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
1014 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
1015 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
1016 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
1017 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
1018 #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
1019 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
1020 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
1021 #define arc_dbuf_size ARCSTAT(arcstat_dbuf_size) /* dbuf metadata */
1022 #define arc_dnode_size ARCSTAT(arcstat_dnode_size) /* dnode metadata */
1023 #define arc_bonus_size ARCSTAT(arcstat_bonus_size) /* bonus buffer metadata */
1025 /* compressed size of entire arc */
1026 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
1027 /* uncompressed size of entire arc */
1028 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
1029 /* number of bytes in the arc from arc_buf_t's */
1030 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
1033 * There are also some ARC variables that we want to export, but that are
1034 * updated so often that having the canonical representation be the statistic
1035 * variable causes a performance bottleneck. We want to use aggsum_t's for these
1036 * instead, but still be able to export the kstat in the same way as before.
1037 * The solution is to always use the aggsum version, except in the kstat update
1041 aggsum_t arc_meta_used;
1042 aggsum_t astat_data_size;
1043 aggsum_t astat_metadata_size;
1044 aggsum_t astat_hdr_size;
1045 aggsum_t astat_bonus_size;
1046 aggsum_t astat_dnode_size;
1047 aggsum_t astat_dbuf_size;
1048 aggsum_t astat_l2_hdr_size;
1050 static list_t arc_prune_list;
1051 static kmutex_t arc_prune_mtx;
1052 static taskq_t *arc_prune_taskq;
1054 static int arc_no_grow; /* Don't try to grow cache size */
1055 static uint64_t arc_tempreserve;
1056 static uint64_t arc_loaned_bytes;
1058 typedef struct arc_callback arc_callback_t;
1060 struct arc_callback {
1062 arc_read_done_func_t *acb_done;
1064 boolean_t acb_compressed;
1065 zio_t *acb_zio_dummy;
1066 zio_t *acb_zio_head;
1067 arc_callback_t *acb_next;
1070 typedef struct arc_write_callback arc_write_callback_t;
1072 struct arc_write_callback {
1074 arc_write_done_func_t *awcb_ready;
1075 arc_write_done_func_t *awcb_children_ready;
1076 arc_write_done_func_t *awcb_physdone;
1077 arc_write_done_func_t *awcb_done;
1078 arc_buf_t *awcb_buf;
1082 * ARC buffers are separated into multiple structs as a memory saving measure:
1083 * - Common fields struct, always defined, and embedded within it:
1084 * - L2-only fields, always allocated but undefined when not in L2ARC
1085 * - L1-only fields, only allocated when in L1ARC
1087 * Buffer in L1 Buffer only in L2
1088 * +------------------------+ +------------------------+
1089 * | arc_buf_hdr_t | | arc_buf_hdr_t |
1093 * +------------------------+ +------------------------+
1094 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
1095 * | (undefined if L1-only) | | |
1096 * +------------------------+ +------------------------+
1097 * | l1arc_buf_hdr_t |
1102 * +------------------------+
1104 * Because it's possible for the L2ARC to become extremely large, we can wind
1105 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
1106 * is minimized by only allocating the fields necessary for an L1-cached buffer
1107 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
1108 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
1109 * words in pointers. arc_hdr_realloc() is used to switch a header between
1110 * these two allocation states.
1112 typedef struct l1arc_buf_hdr {
1113 kmutex_t b_freeze_lock;
1114 zio_cksum_t *b_freeze_cksum;
1117 * Used for debugging with kmem_flags - by allocating and freeing
1118 * b_thawed when the buffer is thawed, we get a record of the stack
1119 * trace that thawed it.
1126 /* for waiting on writes to complete */
1130 /* protected by arc state mutex */
1131 arc_state_t *b_state;
1132 multilist_node_t b_arc_node;
1134 /* updated atomically */
1135 clock_t b_arc_access;
1136 uint32_t b_mru_hits;
1137 uint32_t b_mru_ghost_hits;
1138 uint32_t b_mfu_hits;
1139 uint32_t b_mfu_ghost_hits;
1142 /* self protecting */
1143 refcount_t b_refcnt;
1145 arc_callback_t *b_acb;
1149 typedef struct l2arc_dev l2arc_dev_t;
1151 typedef struct l2arc_buf_hdr {
1152 /* protected by arc_buf_hdr mutex */
1153 l2arc_dev_t *b_dev; /* L2ARC device */
1154 uint64_t b_daddr; /* disk address, offset byte */
1157 list_node_t b_l2node;
1160 struct arc_buf_hdr {
1161 /* protected by hash lock */
1165 arc_buf_contents_t b_type;
1166 arc_buf_hdr_t *b_hash_next;
1167 arc_flags_t b_flags;
1170 * This field stores the size of the data buffer after
1171 * compression, and is set in the arc's zio completion handlers.
1172 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
1174 * While the block pointers can store up to 32MB in their psize
1175 * field, we can only store up to 32MB minus 512B. This is due
1176 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
1177 * a field of zeros represents 512B in the bp). We can't use a
1178 * bias of 1 since we need to reserve a psize of zero, here, to
1179 * represent holes and embedded blocks.
1181 * This isn't a problem in practice, since the maximum size of a
1182 * buffer is limited to 16MB, so we never need to store 32MB in
1183 * this field. Even in the upstream illumos code base, the
1184 * maximum size of a buffer is limited to 16MB.
1189 * This field stores the size of the data buffer before
1190 * compression, and cannot change once set. It is in units
1191 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
1193 uint16_t b_lsize; /* immutable */
1194 uint64_t b_spa; /* immutable */
1196 /* L2ARC fields. Undefined when not in L2ARC. */
1197 l2arc_buf_hdr_t b_l2hdr;
1198 /* L1ARC fields. Undefined when in l2arc_only state */
1199 l1arc_buf_hdr_t b_l1hdr;
1202 #if defined(__FreeBSD__) && defined(_KERNEL)
1204 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
1209 val = arc_meta_limit;
1210 err = sysctl_handle_64(oidp, &val, 0, req);
1211 if (err != 0 || req->newptr == NULL)
1214 if (val <= 0 || val > arc_c_max)
1217 arc_meta_limit = val;
1222 sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS)
1227 val = arc_no_grow_shift;
1228 err = sysctl_handle_32(oidp, &val, 0, req);
1229 if (err != 0 || req->newptr == NULL)
1232 if (val >= arc_shrink_shift)
1235 arc_no_grow_shift = val;
1240 sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
1246 err = sysctl_handle_64(oidp, &val, 0, req);
1247 if (err != 0 || req->newptr == NULL)
1250 if (zfs_arc_max == 0) {
1251 /* Loader tunable so blindly set */
1256 if (val < arc_abs_min || val > kmem_size())
1258 if (val < arc_c_min)
1260 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
1266 arc_p = (arc_c >> 1);
1268 if (zfs_arc_meta_limit == 0) {
1269 /* limit meta-data to 1/4 of the arc capacity */
1270 arc_meta_limit = arc_c_max / 4;
1273 /* if kmem_flags are set, lets try to use less memory */
1274 if (kmem_debugging())
1277 zfs_arc_max = arc_c;
1283 sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
1289 err = sysctl_handle_64(oidp, &val, 0, req);
1290 if (err != 0 || req->newptr == NULL)
1293 if (zfs_arc_min == 0) {
1294 /* Loader tunable so blindly set */
1299 if (val < arc_abs_min || val > arc_c_max)
1304 if (zfs_arc_meta_min == 0)
1305 arc_meta_min = arc_c_min / 2;
1307 if (arc_c < arc_c_min)
1310 zfs_arc_min = arc_c_min;
1316 #define GHOST_STATE(state) \
1317 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
1318 (state) == arc_l2c_only)
1320 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1321 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1322 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1323 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1324 #define HDR_PRESCIENT_PREFETCH(hdr) \
1325 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
1326 #define HDR_COMPRESSION_ENABLED(hdr) \
1327 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1329 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1330 #define HDR_L2_READING(hdr) \
1331 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1332 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1333 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1334 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1335 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1336 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1338 #define HDR_ISTYPE_METADATA(hdr) \
1339 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1340 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1342 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1343 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1345 /* For storing compression mode in b_flags */
1346 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1348 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1349 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1350 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1351 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1353 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1354 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1355 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1361 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1362 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1365 * Hash table routines
1368 #define HT_LOCK_PAD CACHE_LINE_SIZE
1373 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1377 #define BUF_LOCKS 256
1378 typedef struct buf_hash_table {
1380 arc_buf_hdr_t **ht_table;
1381 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1384 static buf_hash_table_t buf_hash_table;
1386 #define BUF_HASH_INDEX(spa, dva, birth) \
1387 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1388 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1389 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1390 #define HDR_LOCK(hdr) \
1391 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1393 uint64_t zfs_crc64_table[256];
1399 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1400 #define L2ARC_HEADROOM 2 /* num of writes */
1402 * If we discover during ARC scan any buffers to be compressed, we boost
1403 * our headroom for the next scanning cycle by this percentage multiple.
1405 #define L2ARC_HEADROOM_BOOST 200
1406 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1407 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1409 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1410 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1412 /* L2ARC Performance Tunables */
1413 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1414 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1415 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1416 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1417 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1418 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1419 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1420 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1421 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1423 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1424 &l2arc_write_max, 0, "max write size");
1425 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1426 &l2arc_write_boost, 0, "extra write during warmup");
1427 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1428 &l2arc_headroom, 0, "number of dev writes");
1429 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1430 &l2arc_feed_secs, 0, "interval seconds");
1431 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1432 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1434 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1435 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1436 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1437 &l2arc_feed_again, 0, "turbo warmup");
1438 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1439 &l2arc_norw, 0, "no reads during writes");
1441 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1442 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1443 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD,
1444 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1445 "size of anonymous state");
1446 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD,
1447 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1448 "size of anonymous state");
1450 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1451 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1452 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD,
1453 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1454 "size of metadata in mru state");
1455 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD,
1456 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1457 "size of data in mru state");
1459 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1460 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1461 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD,
1462 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1463 "size of metadata in mru ghost state");
1464 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD,
1465 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1466 "size of data in mru ghost state");
1468 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1469 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1470 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD,
1471 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1472 "size of metadata in mfu state");
1473 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD,
1474 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1475 "size of data in mfu state");
1477 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1478 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1479 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD,
1480 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1481 "size of metadata in mfu ghost state");
1482 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD,
1483 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1484 "size of data in mfu ghost state");
1486 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1487 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1489 SYSCTL_UINT(_vfs_zfs, OID_AUTO, arc_min_prefetch_ms, CTLFLAG_RW,
1490 &zfs_arc_min_prefetch_ms, 0, "Min life of prefetch block in ms");
1491 SYSCTL_UINT(_vfs_zfs, OID_AUTO, arc_min_prescient_prefetch_ms, CTLFLAG_RW,
1492 &zfs_arc_min_prescient_prefetch_ms, 0, "Min life of prescient prefetched block in ms");
1498 vdev_t *l2ad_vdev; /* vdev */
1499 spa_t *l2ad_spa; /* spa */
1500 uint64_t l2ad_hand; /* next write location */
1501 uint64_t l2ad_start; /* first addr on device */
1502 uint64_t l2ad_end; /* last addr on device */
1503 boolean_t l2ad_first; /* first sweep through */
1504 boolean_t l2ad_writing; /* currently writing */
1505 kmutex_t l2ad_mtx; /* lock for buffer list */
1506 list_t l2ad_buflist; /* buffer list */
1507 list_node_t l2ad_node; /* device list node */
1508 refcount_t l2ad_alloc; /* allocated bytes */
1511 static list_t L2ARC_dev_list; /* device list */
1512 static list_t *l2arc_dev_list; /* device list pointer */
1513 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1514 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1515 static list_t L2ARC_free_on_write; /* free after write buf list */
1516 static list_t *l2arc_free_on_write; /* free after write list ptr */
1517 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1518 static uint64_t l2arc_ndev; /* number of devices */
1520 typedef struct l2arc_read_callback {
1521 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1522 blkptr_t l2rcb_bp; /* original blkptr */
1523 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1524 int l2rcb_flags; /* original flags */
1525 abd_t *l2rcb_abd; /* temporary buffer */
1526 } l2arc_read_callback_t;
1528 typedef struct l2arc_write_callback {
1529 l2arc_dev_t *l2wcb_dev; /* device info */
1530 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1531 } l2arc_write_callback_t;
1533 typedef struct l2arc_data_free {
1534 /* protected by l2arc_free_on_write_mtx */
1537 arc_buf_contents_t l2df_type;
1538 list_node_t l2df_list_node;
1539 } l2arc_data_free_t;
1541 static kmutex_t l2arc_feed_thr_lock;
1542 static kcondvar_t l2arc_feed_thr_cv;
1543 static uint8_t l2arc_thread_exit;
1545 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1546 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1547 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1548 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1549 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1550 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1551 static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1552 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *);
1553 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1554 static boolean_t arc_is_overflowing();
1555 static void arc_buf_watch(arc_buf_t *);
1556 static void arc_prune_async(int64_t);
1558 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1559 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1560 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1561 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1563 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1564 static void l2arc_read_done(zio_t *);
1567 l2arc_trim(const arc_buf_hdr_t *hdr)
1569 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1571 ASSERT(HDR_HAS_L2HDR(hdr));
1572 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1574 if (HDR_GET_PSIZE(hdr) != 0) {
1575 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1576 HDR_GET_PSIZE(hdr), 0);
1581 * We use Cityhash for this. It's fast, and has good hash properties without
1582 * requiring any large static buffers.
1585 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1587 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
1590 #define HDR_EMPTY(hdr) \
1591 ((hdr)->b_dva.dva_word[0] == 0 && \
1592 (hdr)->b_dva.dva_word[1] == 0)
1594 #define HDR_EQUAL(spa, dva, birth, hdr) \
1595 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1596 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1597 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1600 buf_discard_identity(arc_buf_hdr_t *hdr)
1602 hdr->b_dva.dva_word[0] = 0;
1603 hdr->b_dva.dva_word[1] = 0;
1607 static arc_buf_hdr_t *
1608 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1610 const dva_t *dva = BP_IDENTITY(bp);
1611 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1612 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1613 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1616 mutex_enter(hash_lock);
1617 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1618 hdr = hdr->b_hash_next) {
1619 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1624 mutex_exit(hash_lock);
1630 * Insert an entry into the hash table. If there is already an element
1631 * equal to elem in the hash table, then the already existing element
1632 * will be returned and the new element will not be inserted.
1633 * Otherwise returns NULL.
1634 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1636 static arc_buf_hdr_t *
1637 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1639 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1640 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1641 arc_buf_hdr_t *fhdr;
1644 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1645 ASSERT(hdr->b_birth != 0);
1646 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1648 if (lockp != NULL) {
1650 mutex_enter(hash_lock);
1652 ASSERT(MUTEX_HELD(hash_lock));
1655 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1656 fhdr = fhdr->b_hash_next, i++) {
1657 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1661 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1662 buf_hash_table.ht_table[idx] = hdr;
1663 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1665 /* collect some hash table performance data */
1667 ARCSTAT_BUMP(arcstat_hash_collisions);
1669 ARCSTAT_BUMP(arcstat_hash_chains);
1671 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1674 ARCSTAT_BUMP(arcstat_hash_elements);
1675 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1681 buf_hash_remove(arc_buf_hdr_t *hdr)
1683 arc_buf_hdr_t *fhdr, **hdrp;
1684 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1686 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1687 ASSERT(HDR_IN_HASH_TABLE(hdr));
1689 hdrp = &buf_hash_table.ht_table[idx];
1690 while ((fhdr = *hdrp) != hdr) {
1691 ASSERT3P(fhdr, !=, NULL);
1692 hdrp = &fhdr->b_hash_next;
1694 *hdrp = hdr->b_hash_next;
1695 hdr->b_hash_next = NULL;
1696 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1698 /* collect some hash table performance data */
1699 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1701 if (buf_hash_table.ht_table[idx] &&
1702 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1703 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1707 * Global data structures and functions for the buf kmem cache.
1709 static kmem_cache_t *hdr_full_cache;
1710 static kmem_cache_t *hdr_l2only_cache;
1711 static kmem_cache_t *buf_cache;
1718 kmem_free(buf_hash_table.ht_table,
1719 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1720 for (i = 0; i < BUF_LOCKS; i++)
1721 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1722 kmem_cache_destroy(hdr_full_cache);
1723 kmem_cache_destroy(hdr_l2only_cache);
1724 kmem_cache_destroy(buf_cache);
1728 * Constructor callback - called when the cache is empty
1729 * and a new buf is requested.
1733 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1735 arc_buf_hdr_t *hdr = vbuf;
1737 bzero(hdr, HDR_FULL_SIZE);
1738 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1739 refcount_create(&hdr->b_l1hdr.b_refcnt);
1740 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1741 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1742 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1749 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1751 arc_buf_hdr_t *hdr = vbuf;
1753 bzero(hdr, HDR_L2ONLY_SIZE);
1754 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1761 buf_cons(void *vbuf, void *unused, int kmflag)
1763 arc_buf_t *buf = vbuf;
1765 bzero(buf, sizeof (arc_buf_t));
1766 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1767 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1773 * Destructor callback - called when a cached buf is
1774 * no longer required.
1778 hdr_full_dest(void *vbuf, void *unused)
1780 arc_buf_hdr_t *hdr = vbuf;
1782 ASSERT(HDR_EMPTY(hdr));
1783 cv_destroy(&hdr->b_l1hdr.b_cv);
1784 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1785 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1786 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1787 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1792 hdr_l2only_dest(void *vbuf, void *unused)
1794 arc_buf_hdr_t *hdr = vbuf;
1796 ASSERT(HDR_EMPTY(hdr));
1797 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1802 buf_dest(void *vbuf, void *unused)
1804 arc_buf_t *buf = vbuf;
1806 mutex_destroy(&buf->b_evict_lock);
1807 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1811 * Reclaim callback -- invoked when memory is low.
1815 hdr_recl(void *unused)
1817 dprintf("hdr_recl called\n");
1819 * umem calls the reclaim func when we destroy the buf cache,
1820 * which is after we do arc_fini().
1823 cv_signal(&arc_reclaim_thread_cv);
1830 uint64_t hsize = 1ULL << 12;
1834 * The hash table is big enough to fill all of physical memory
1835 * with an average block size of zfs_arc_average_blocksize (default 8K).
1836 * By default, the table will take up
1837 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1839 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1842 buf_hash_table.ht_mask = hsize - 1;
1843 buf_hash_table.ht_table =
1844 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1845 if (buf_hash_table.ht_table == NULL) {
1846 ASSERT(hsize > (1ULL << 8));
1851 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1852 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1853 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1854 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1856 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1857 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1859 for (i = 0; i < 256; i++)
1860 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1861 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1863 for (i = 0; i < BUF_LOCKS; i++) {
1864 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1865 NULL, MUTEX_DEFAULT, NULL);
1870 * This is the size that the buf occupies in memory. If the buf is compressed,
1871 * it will correspond to the compressed size. You should use this method of
1872 * getting the buf size unless you explicitly need the logical size.
1875 arc_buf_size(arc_buf_t *buf)
1877 return (ARC_BUF_COMPRESSED(buf) ?
1878 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1882 arc_buf_lsize(arc_buf_t *buf)
1884 return (HDR_GET_LSIZE(buf->b_hdr));
1888 arc_get_compression(arc_buf_t *buf)
1890 return (ARC_BUF_COMPRESSED(buf) ?
1891 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1894 #define ARC_MINTIME (hz>>4) /* 62 ms */
1896 static inline boolean_t
1897 arc_buf_is_shared(arc_buf_t *buf)
1899 boolean_t shared = (buf->b_data != NULL &&
1900 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1901 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1902 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1903 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1904 IMPLY(shared, ARC_BUF_SHARED(buf));
1905 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1908 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1909 * already being shared" requirement prevents us from doing that.
1916 * Free the checksum associated with this header. If there is no checksum, this
1920 arc_cksum_free(arc_buf_hdr_t *hdr)
1922 ASSERT(HDR_HAS_L1HDR(hdr));
1923 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1924 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1925 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1926 hdr->b_l1hdr.b_freeze_cksum = NULL;
1928 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1932 * Return true iff at least one of the bufs on hdr is not compressed.
1935 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1937 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1938 if (!ARC_BUF_COMPRESSED(b)) {
1946 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1947 * matches the checksum that is stored in the hdr. If there is no checksum,
1948 * or if the buf is compressed, this is a no-op.
1951 arc_cksum_verify(arc_buf_t *buf)
1953 arc_buf_hdr_t *hdr = buf->b_hdr;
1956 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1959 if (ARC_BUF_COMPRESSED(buf)) {
1960 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1961 arc_hdr_has_uncompressed_buf(hdr));
1965 ASSERT(HDR_HAS_L1HDR(hdr));
1967 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1968 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1969 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1973 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1974 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1975 panic("buffer modified while frozen!");
1976 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1980 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1982 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1983 boolean_t valid_cksum;
1985 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1986 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1989 * We rely on the blkptr's checksum to determine if the block
1990 * is valid or not. When compressed arc is enabled, the l2arc
1991 * writes the block to the l2arc just as it appears in the pool.
1992 * This allows us to use the blkptr's checksum to validate the
1993 * data that we just read off of the l2arc without having to store
1994 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1995 * arc is disabled, then the data written to the l2arc is always
1996 * uncompressed and won't match the block as it exists in the main
1997 * pool. When this is the case, we must first compress it if it is
1998 * compressed on the main pool before we can validate the checksum.
2000 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
2001 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2002 uint64_t lsize = HDR_GET_LSIZE(hdr);
2005 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
2006 csize = zio_compress_data(compress, zio->io_abd,
2007 abd_to_buf(cdata), lsize);
2009 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
2010 if (csize < HDR_GET_PSIZE(hdr)) {
2012 * Compressed blocks are always a multiple of the
2013 * smallest ashift in the pool. Ideally, we would
2014 * like to round up the csize to the next
2015 * spa_min_ashift but that value may have changed
2016 * since the block was last written. Instead,
2017 * we rely on the fact that the hdr's psize
2018 * was set to the psize of the block when it was
2019 * last written. We set the csize to that value
2020 * and zero out any part that should not contain
2023 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
2024 csize = HDR_GET_PSIZE(hdr);
2026 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
2030 * Block pointers always store the checksum for the logical data.
2031 * If the block pointer has the gang bit set, then the checksum
2032 * it represents is for the reconstituted data and not for an
2033 * individual gang member. The zio pipeline, however, must be able to
2034 * determine the checksum of each of the gang constituents so it
2035 * treats the checksum comparison differently than what we need
2036 * for l2arc blocks. This prevents us from using the
2037 * zio_checksum_error() interface directly. Instead we must call the
2038 * zio_checksum_error_impl() so that we can ensure the checksum is
2039 * generated using the correct checksum algorithm and accounts for the
2040 * logical I/O size and not just a gang fragment.
2042 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
2043 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
2044 zio->io_offset, NULL) == 0);
2045 zio_pop_transforms(zio);
2046 return (valid_cksum);
2050 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
2051 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
2052 * isn't modified later on. If buf is compressed or there is already a checksum
2053 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
2056 arc_cksum_compute(arc_buf_t *buf)
2058 arc_buf_hdr_t *hdr = buf->b_hdr;
2060 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2063 ASSERT(HDR_HAS_L1HDR(hdr));
2065 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
2066 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
2067 ASSERT(arc_hdr_has_uncompressed_buf(hdr));
2068 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2070 } else if (ARC_BUF_COMPRESSED(buf)) {
2071 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2075 ASSERT(!ARC_BUF_COMPRESSED(buf));
2076 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
2078 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
2079 hdr->b_l1hdr.b_freeze_cksum);
2080 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2088 typedef struct procctl {
2096 arc_buf_unwatch(arc_buf_t *buf)
2103 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2104 ctl.prwatch.pr_size = 0;
2105 ctl.prwatch.pr_wflags = 0;
2106 result = write(arc_procfd, &ctl, sizeof (ctl));
2107 ASSERT3U(result, ==, sizeof (ctl));
2114 arc_buf_watch(arc_buf_t *buf)
2121 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2122 ctl.prwatch.pr_size = arc_buf_size(buf);
2123 ctl.prwatch.pr_wflags = WA_WRITE;
2124 result = write(arc_procfd, &ctl, sizeof (ctl));
2125 ASSERT3U(result, ==, sizeof (ctl));
2129 #endif /* illumos */
2131 static arc_buf_contents_t
2132 arc_buf_type(arc_buf_hdr_t *hdr)
2134 arc_buf_contents_t type;
2135 if (HDR_ISTYPE_METADATA(hdr)) {
2136 type = ARC_BUFC_METADATA;
2138 type = ARC_BUFC_DATA;
2140 VERIFY3U(hdr->b_type, ==, type);
2145 arc_is_metadata(arc_buf_t *buf)
2147 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
2151 arc_bufc_to_flags(arc_buf_contents_t type)
2155 /* metadata field is 0 if buffer contains normal data */
2157 case ARC_BUFC_METADATA:
2158 return (ARC_FLAG_BUFC_METADATA);
2162 panic("undefined ARC buffer type!");
2163 return ((uint32_t)-1);
2167 arc_buf_thaw(arc_buf_t *buf)
2169 arc_buf_hdr_t *hdr = buf->b_hdr;
2171 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2172 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2174 arc_cksum_verify(buf);
2177 * Compressed buffers do not manipulate the b_freeze_cksum or
2178 * allocate b_thawed.
2180 if (ARC_BUF_COMPRESSED(buf)) {
2181 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2182 arc_hdr_has_uncompressed_buf(hdr));
2186 ASSERT(HDR_HAS_L1HDR(hdr));
2187 arc_cksum_free(hdr);
2189 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
2191 if (zfs_flags & ZFS_DEBUG_MODIFY) {
2192 if (hdr->b_l1hdr.b_thawed != NULL)
2193 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2194 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
2198 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2201 arc_buf_unwatch(buf);
2206 arc_buf_freeze(arc_buf_t *buf)
2208 arc_buf_hdr_t *hdr = buf->b_hdr;
2209 kmutex_t *hash_lock;
2211 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2214 if (ARC_BUF_COMPRESSED(buf)) {
2215 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2216 arc_hdr_has_uncompressed_buf(hdr));
2220 hash_lock = HDR_LOCK(hdr);
2221 mutex_enter(hash_lock);
2223 ASSERT(HDR_HAS_L1HDR(hdr));
2224 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
2225 hdr->b_l1hdr.b_state == arc_anon);
2226 arc_cksum_compute(buf);
2227 mutex_exit(hash_lock);
2231 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
2232 * the following functions should be used to ensure that the flags are
2233 * updated in a thread-safe way. When manipulating the flags either
2234 * the hash_lock must be held or the hdr must be undiscoverable. This
2235 * ensures that we're not racing with any other threads when updating
2239 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2241 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2242 hdr->b_flags |= flags;
2246 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2248 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2249 hdr->b_flags &= ~flags;
2253 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
2254 * done in a special way since we have to clear and set bits
2255 * at the same time. Consumers that wish to set the compression bits
2256 * must use this function to ensure that the flags are updated in
2257 * thread-safe manner.
2260 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
2262 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2265 * Holes and embedded blocks will always have a psize = 0 so
2266 * we ignore the compression of the blkptr and set the
2267 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
2268 * Holes and embedded blocks remain anonymous so we don't
2269 * want to uncompress them. Mark them as uncompressed.
2271 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
2272 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2273 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
2274 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
2275 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2277 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2278 HDR_SET_COMPRESS(hdr, cmp);
2279 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
2280 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
2285 * Looks for another buf on the same hdr which has the data decompressed, copies
2286 * from it, and returns true. If no such buf exists, returns false.
2289 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
2291 arc_buf_hdr_t *hdr = buf->b_hdr;
2292 boolean_t copied = B_FALSE;
2294 ASSERT(HDR_HAS_L1HDR(hdr));
2295 ASSERT3P(buf->b_data, !=, NULL);
2296 ASSERT(!ARC_BUF_COMPRESSED(buf));
2298 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
2299 from = from->b_next) {
2300 /* can't use our own data buffer */
2305 if (!ARC_BUF_COMPRESSED(from)) {
2306 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
2313 * There were no decompressed bufs, so there should not be a
2314 * checksum on the hdr either.
2316 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
2322 * Given a buf that has a data buffer attached to it, this function will
2323 * efficiently fill the buf with data of the specified compression setting from
2324 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2325 * are already sharing a data buf, no copy is performed.
2327 * If the buf is marked as compressed but uncompressed data was requested, this
2328 * will allocate a new data buffer for the buf, remove that flag, and fill the
2329 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2330 * uncompressed data, and (since we haven't added support for it yet) if you
2331 * want compressed data your buf must already be marked as compressed and have
2332 * the correct-sized data buffer.
2335 arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
2337 arc_buf_hdr_t *hdr = buf->b_hdr;
2338 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
2339 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2341 ASSERT3P(buf->b_data, !=, NULL);
2342 IMPLY(compressed, hdr_compressed);
2343 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2345 if (hdr_compressed == compressed) {
2346 if (!arc_buf_is_shared(buf)) {
2347 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2351 ASSERT(hdr_compressed);
2352 ASSERT(!compressed);
2353 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2356 * If the buf is sharing its data with the hdr, unlink it and
2357 * allocate a new data buffer for the buf.
2359 if (arc_buf_is_shared(buf)) {
2360 ASSERT(ARC_BUF_COMPRESSED(buf));
2362 /* We need to give the buf it's own b_data */
2363 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2365 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2366 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2368 /* Previously overhead was 0; just add new overhead */
2369 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2370 } else if (ARC_BUF_COMPRESSED(buf)) {
2371 /* We need to reallocate the buf's b_data */
2372 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2375 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2377 /* We increased the size of b_data; update overhead */
2378 ARCSTAT_INCR(arcstat_overhead_size,
2379 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2383 * Regardless of the buf's previous compression settings, it
2384 * should not be compressed at the end of this function.
2386 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2389 * Try copying the data from another buf which already has a
2390 * decompressed version. If that's not possible, it's time to
2391 * bite the bullet and decompress the data from the hdr.
2393 if (arc_buf_try_copy_decompressed_data(buf)) {
2394 /* Skip byteswapping and checksumming (already done) */
2395 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2398 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2399 hdr->b_l1hdr.b_pabd, buf->b_data,
2400 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2403 * Absent hardware errors or software bugs, this should
2404 * be impossible, but log it anyway so we can debug it.
2408 "hdr %p, compress %d, psize %d, lsize %d",
2409 hdr, HDR_GET_COMPRESS(hdr),
2410 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2411 return (SET_ERROR(EIO));
2416 /* Byteswap the buf's data if necessary */
2417 if (bswap != DMU_BSWAP_NUMFUNCS) {
2418 ASSERT(!HDR_SHARED_DATA(hdr));
2419 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2420 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2423 /* Compute the hdr's checksum if necessary */
2424 arc_cksum_compute(buf);
2430 arc_decompress(arc_buf_t *buf)
2432 return (arc_buf_fill(buf, B_FALSE));
2436 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
2439 arc_hdr_size(arc_buf_hdr_t *hdr)
2443 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2444 HDR_GET_PSIZE(hdr) > 0) {
2445 size = HDR_GET_PSIZE(hdr);
2447 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2448 size = HDR_GET_LSIZE(hdr);
2454 * Increment the amount of evictable space in the arc_state_t's refcount.
2455 * We account for the space used by the hdr and the arc buf individually
2456 * so that we can add and remove them from the refcount individually.
2459 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2461 arc_buf_contents_t type = arc_buf_type(hdr);
2463 ASSERT(HDR_HAS_L1HDR(hdr));
2465 if (GHOST_STATE(state)) {
2466 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2467 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2468 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2469 (void) refcount_add_many(&state->arcs_esize[type],
2470 HDR_GET_LSIZE(hdr), hdr);
2474 ASSERT(!GHOST_STATE(state));
2475 if (hdr->b_l1hdr.b_pabd != NULL) {
2476 (void) refcount_add_many(&state->arcs_esize[type],
2477 arc_hdr_size(hdr), hdr);
2479 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2480 buf = buf->b_next) {
2481 if (arc_buf_is_shared(buf))
2483 (void) refcount_add_many(&state->arcs_esize[type],
2484 arc_buf_size(buf), buf);
2489 * Decrement the amount of evictable space in the arc_state_t's refcount.
2490 * We account for the space used by the hdr and the arc buf individually
2491 * so that we can add and remove them from the refcount individually.
2494 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2496 arc_buf_contents_t type = arc_buf_type(hdr);
2498 ASSERT(HDR_HAS_L1HDR(hdr));
2500 if (GHOST_STATE(state)) {
2501 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2502 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2503 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2504 (void) refcount_remove_many(&state->arcs_esize[type],
2505 HDR_GET_LSIZE(hdr), hdr);
2509 ASSERT(!GHOST_STATE(state));
2510 if (hdr->b_l1hdr.b_pabd != NULL) {
2511 (void) refcount_remove_many(&state->arcs_esize[type],
2512 arc_hdr_size(hdr), hdr);
2514 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2515 buf = buf->b_next) {
2516 if (arc_buf_is_shared(buf))
2518 (void) refcount_remove_many(&state->arcs_esize[type],
2519 arc_buf_size(buf), buf);
2524 * Add a reference to this hdr indicating that someone is actively
2525 * referencing that memory. When the refcount transitions from 0 to 1,
2526 * we remove it from the respective arc_state_t list to indicate that
2527 * it is not evictable.
2530 add_reference(arc_buf_hdr_t *hdr, void *tag)
2532 ASSERT(HDR_HAS_L1HDR(hdr));
2533 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2534 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2535 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2536 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2539 arc_state_t *state = hdr->b_l1hdr.b_state;
2541 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2542 (state != arc_anon)) {
2543 /* We don't use the L2-only state list. */
2544 if (state != arc_l2c_only) {
2545 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2547 arc_evictable_space_decrement(hdr, state);
2549 /* remove the prefetch flag if we get a reference */
2550 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2555 * Remove a reference from this hdr. When the reference transitions from
2556 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2557 * list making it eligible for eviction.
2560 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2563 arc_state_t *state = hdr->b_l1hdr.b_state;
2565 ASSERT(HDR_HAS_L1HDR(hdr));
2566 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2567 ASSERT(!GHOST_STATE(state));
2570 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2571 * check to prevent usage of the arc_l2c_only list.
2573 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2574 (state != arc_anon)) {
2575 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2576 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2577 arc_evictable_space_increment(hdr, state);
2583 * Returns detailed information about a specific arc buffer. When the
2584 * state_index argument is set the function will calculate the arc header
2585 * list position for its arc state. Since this requires a linear traversal
2586 * callers are strongly encourage not to do this. However, it can be helpful
2587 * for targeted analysis so the functionality is provided.
2590 arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
2592 arc_buf_hdr_t *hdr = ab->b_hdr;
2593 l1arc_buf_hdr_t *l1hdr = NULL;
2594 l2arc_buf_hdr_t *l2hdr = NULL;
2595 arc_state_t *state = NULL;
2597 memset(abi, 0, sizeof (arc_buf_info_t));
2602 abi->abi_flags = hdr->b_flags;
2604 if (HDR_HAS_L1HDR(hdr)) {
2605 l1hdr = &hdr->b_l1hdr;
2606 state = l1hdr->b_state;
2608 if (HDR_HAS_L2HDR(hdr))
2609 l2hdr = &hdr->b_l2hdr;
2612 abi->abi_bufcnt = l1hdr->b_bufcnt;
2613 abi->abi_access = l1hdr->b_arc_access;
2614 abi->abi_mru_hits = l1hdr->b_mru_hits;
2615 abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits;
2616 abi->abi_mfu_hits = l1hdr->b_mfu_hits;
2617 abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits;
2618 abi->abi_holds = refcount_count(&l1hdr->b_refcnt);
2622 abi->abi_l2arc_dattr = l2hdr->b_daddr;
2623 abi->abi_l2arc_hits = l2hdr->b_hits;
2626 abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
2627 abi->abi_state_contents = arc_buf_type(hdr);
2628 abi->abi_size = arc_hdr_size(hdr);
2632 * Move the supplied buffer to the indicated state. The hash lock
2633 * for the buffer must be held by the caller.
2636 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2637 kmutex_t *hash_lock)
2639 arc_state_t *old_state;
2642 boolean_t update_old, update_new;
2643 arc_buf_contents_t buftype = arc_buf_type(hdr);
2646 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2647 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2648 * L1 hdr doesn't always exist when we change state to arc_anon before
2649 * destroying a header, in which case reallocating to add the L1 hdr is
2652 if (HDR_HAS_L1HDR(hdr)) {
2653 old_state = hdr->b_l1hdr.b_state;
2654 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2655 bufcnt = hdr->b_l1hdr.b_bufcnt;
2656 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL);
2658 old_state = arc_l2c_only;
2661 update_old = B_FALSE;
2663 update_new = update_old;
2665 ASSERT(MUTEX_HELD(hash_lock));
2666 ASSERT3P(new_state, !=, old_state);
2667 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2668 ASSERT(old_state != arc_anon || bufcnt <= 1);
2671 * If this buffer is evictable, transfer it from the
2672 * old state list to the new state list.
2675 if (old_state != arc_anon && old_state != arc_l2c_only) {
2676 ASSERT(HDR_HAS_L1HDR(hdr));
2677 multilist_remove(old_state->arcs_list[buftype], hdr);
2679 if (GHOST_STATE(old_state)) {
2681 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2682 update_old = B_TRUE;
2684 arc_evictable_space_decrement(hdr, old_state);
2686 if (new_state != arc_anon && new_state != arc_l2c_only) {
2689 * An L1 header always exists here, since if we're
2690 * moving to some L1-cached state (i.e. not l2c_only or
2691 * anonymous), we realloc the header to add an L1hdr
2694 ASSERT(HDR_HAS_L1HDR(hdr));
2695 multilist_insert(new_state->arcs_list[buftype], hdr);
2697 if (GHOST_STATE(new_state)) {
2699 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2700 update_new = B_TRUE;
2702 arc_evictable_space_increment(hdr, new_state);
2706 ASSERT(!HDR_EMPTY(hdr));
2707 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2708 buf_hash_remove(hdr);
2710 /* adjust state sizes (ignore arc_l2c_only) */
2712 if (update_new && new_state != arc_l2c_only) {
2713 ASSERT(HDR_HAS_L1HDR(hdr));
2714 if (GHOST_STATE(new_state)) {
2718 * When moving a header to a ghost state, we first
2719 * remove all arc buffers. Thus, we'll have a
2720 * bufcnt of zero, and no arc buffer to use for
2721 * the reference. As a result, we use the arc
2722 * header pointer for the reference.
2724 (void) refcount_add_many(&new_state->arcs_size,
2725 HDR_GET_LSIZE(hdr), hdr);
2726 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2728 uint32_t buffers = 0;
2731 * Each individual buffer holds a unique reference,
2732 * thus we must remove each of these references one
2735 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2736 buf = buf->b_next) {
2737 ASSERT3U(bufcnt, !=, 0);
2741 * When the arc_buf_t is sharing the data
2742 * block with the hdr, the owner of the
2743 * reference belongs to the hdr. Only
2744 * add to the refcount if the arc_buf_t is
2747 if (arc_buf_is_shared(buf))
2750 (void) refcount_add_many(&new_state->arcs_size,
2751 arc_buf_size(buf), buf);
2753 ASSERT3U(bufcnt, ==, buffers);
2755 if (hdr->b_l1hdr.b_pabd != NULL) {
2756 (void) refcount_add_many(&new_state->arcs_size,
2757 arc_hdr_size(hdr), hdr);
2759 ASSERT(GHOST_STATE(old_state));
2764 if (update_old && old_state != arc_l2c_only) {
2765 ASSERT(HDR_HAS_L1HDR(hdr));
2766 if (GHOST_STATE(old_state)) {
2768 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2771 * When moving a header off of a ghost state,
2772 * the header will not contain any arc buffers.
2773 * We use the arc header pointer for the reference
2774 * which is exactly what we did when we put the
2775 * header on the ghost state.
2778 (void) refcount_remove_many(&old_state->arcs_size,
2779 HDR_GET_LSIZE(hdr), hdr);
2781 uint32_t buffers = 0;
2784 * Each individual buffer holds a unique reference,
2785 * thus we must remove each of these references one
2788 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2789 buf = buf->b_next) {
2790 ASSERT3U(bufcnt, !=, 0);
2794 * When the arc_buf_t is sharing the data
2795 * block with the hdr, the owner of the
2796 * reference belongs to the hdr. Only
2797 * add to the refcount if the arc_buf_t is
2800 if (arc_buf_is_shared(buf))
2803 (void) refcount_remove_many(
2804 &old_state->arcs_size, arc_buf_size(buf),
2807 ASSERT3U(bufcnt, ==, buffers);
2808 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2809 (void) refcount_remove_many(
2810 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2814 if (HDR_HAS_L1HDR(hdr))
2815 hdr->b_l1hdr.b_state = new_state;
2818 * L2 headers should never be on the L2 state list since they don't
2819 * have L1 headers allocated.
2821 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2822 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2826 arc_space_consume(uint64_t space, arc_space_type_t type)
2828 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2831 case ARC_SPACE_DATA:
2832 aggsum_add(&astat_data_size, space);
2834 case ARC_SPACE_META:
2835 aggsum_add(&astat_metadata_size, space);
2837 case ARC_SPACE_BONUS:
2838 aggsum_add(&astat_bonus_size, space);
2840 case ARC_SPACE_DNODE:
2841 aggsum_add(&astat_dnode_size, space);
2843 case ARC_SPACE_DBUF:
2844 aggsum_add(&astat_dbuf_size, space);
2846 case ARC_SPACE_HDRS:
2847 aggsum_add(&astat_hdr_size, space);
2849 case ARC_SPACE_L2HDRS:
2850 aggsum_add(&astat_l2_hdr_size, space);
2854 if (type != ARC_SPACE_DATA)
2855 aggsum_add(&arc_meta_used, space);
2857 aggsum_add(&arc_size, space);
2861 arc_space_return(uint64_t space, arc_space_type_t type)
2863 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2866 case ARC_SPACE_DATA:
2867 aggsum_add(&astat_data_size, -space);
2869 case ARC_SPACE_META:
2870 aggsum_add(&astat_metadata_size, -space);
2872 case ARC_SPACE_BONUS:
2873 aggsum_add(&astat_bonus_size, -space);
2875 case ARC_SPACE_DNODE:
2876 aggsum_add(&astat_dnode_size, -space);
2878 case ARC_SPACE_DBUF:
2879 aggsum_add(&astat_dbuf_size, -space);
2881 case ARC_SPACE_HDRS:
2882 aggsum_add(&astat_hdr_size, -space);
2884 case ARC_SPACE_L2HDRS:
2885 aggsum_add(&astat_l2_hdr_size, -space);
2889 if (type != ARC_SPACE_DATA) {
2890 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2892 * We use the upper bound here rather than the precise value
2893 * because the arc_meta_max value doesn't need to be
2894 * precise. It's only consumed by humans via arcstats.
2896 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2897 arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2898 aggsum_add(&arc_meta_used, -space);
2901 ASSERT(aggsum_compare(&arc_size, space) >= 0);
2902 aggsum_add(&arc_size, -space);
2906 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2907 * with the hdr's b_pabd.
2910 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2913 * The criteria for sharing a hdr's data are:
2914 * 1. the hdr's compression matches the buf's compression
2915 * 2. the hdr doesn't need to be byteswapped
2916 * 3. the hdr isn't already being shared
2917 * 4. the buf is either compressed or it is the last buf in the hdr list
2919 * Criterion #4 maintains the invariant that shared uncompressed
2920 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2921 * might ask, "if a compressed buf is allocated first, won't that be the
2922 * last thing in the list?", but in that case it's impossible to create
2923 * a shared uncompressed buf anyway (because the hdr must be compressed
2924 * to have the compressed buf). You might also think that #3 is
2925 * sufficient to make this guarantee, however it's possible
2926 * (specifically in the rare L2ARC write race mentioned in
2927 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2928 * is sharable, but wasn't at the time of its allocation. Rather than
2929 * allow a new shared uncompressed buf to be created and then shuffle
2930 * the list around to make it the last element, this simply disallows
2931 * sharing if the new buf isn't the first to be added.
2933 ASSERT3P(buf->b_hdr, ==, hdr);
2934 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF;
2935 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2936 return (buf_compressed == hdr_compressed &&
2937 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2938 !HDR_SHARED_DATA(hdr) &&
2939 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2943 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2944 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2945 * copy was made successfully, or an error code otherwise.
2948 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2949 boolean_t fill, arc_buf_t **ret)
2953 ASSERT(HDR_HAS_L1HDR(hdr));
2954 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2955 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2956 hdr->b_type == ARC_BUFC_METADATA);
2957 ASSERT3P(ret, !=, NULL);
2958 ASSERT3P(*ret, ==, NULL);
2960 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2963 buf->b_next = hdr->b_l1hdr.b_buf;
2966 add_reference(hdr, tag);
2969 * We're about to change the hdr's b_flags. We must either
2970 * hold the hash_lock or be undiscoverable.
2972 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2975 * Only honor requests for compressed bufs if the hdr is actually
2978 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
2979 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2982 * If the hdr's data can be shared then we share the data buffer and
2983 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2984 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2985 * buffer to store the buf's data.
2987 * There are two additional restrictions here because we're sharing
2988 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2989 * actively involved in an L2ARC write, because if this buf is used by
2990 * an arc_write() then the hdr's data buffer will be released when the
2991 * write completes, even though the L2ARC write might still be using it.
2992 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2993 * need to be ABD-aware.
2995 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2996 abd_is_linear(hdr->b_l1hdr.b_pabd);
2998 /* Set up b_data and sharing */
3000 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
3001 buf->b_flags |= ARC_BUF_FLAG_SHARED;
3002 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
3005 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
3006 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3008 VERIFY3P(buf->b_data, !=, NULL);
3010 hdr->b_l1hdr.b_buf = buf;
3011 hdr->b_l1hdr.b_bufcnt += 1;
3014 * If the user wants the data from the hdr, we need to either copy or
3015 * decompress the data.
3018 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
3024 static char *arc_onloan_tag = "onloan";
3027 arc_loaned_bytes_update(int64_t delta)
3029 atomic_add_64(&arc_loaned_bytes, delta);
3031 /* assert that it did not wrap around */
3032 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
3036 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
3037 * flight data by arc_tempreserve_space() until they are "returned". Loaned
3038 * buffers must be returned to the arc before they can be used by the DMU or
3042 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
3044 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
3045 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
3047 arc_loaned_bytes_update(arc_buf_size(buf));
3053 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
3054 enum zio_compress compression_type)
3056 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
3057 psize, lsize, compression_type);
3059 arc_loaned_bytes_update(arc_buf_size(buf));
3066 * Return a loaned arc buffer to the arc.
3069 arc_return_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, tag);
3076 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
3078 arc_loaned_bytes_update(-arc_buf_size(buf));
3081 /* Detach an arc_buf from a dbuf (tag) */
3083 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
3085 arc_buf_hdr_t *hdr = buf->b_hdr;
3087 ASSERT3P(buf->b_data, !=, NULL);
3088 ASSERT(HDR_HAS_L1HDR(hdr));
3089 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
3090 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
3092 arc_loaned_bytes_update(arc_buf_size(buf));
3096 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
3098 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
3101 df->l2df_size = size;
3102 df->l2df_type = type;
3103 mutex_enter(&l2arc_free_on_write_mtx);
3104 list_insert_head(l2arc_free_on_write, df);
3105 mutex_exit(&l2arc_free_on_write_mtx);
3109 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
3111 arc_state_t *state = hdr->b_l1hdr.b_state;
3112 arc_buf_contents_t type = arc_buf_type(hdr);
3113 uint64_t size = arc_hdr_size(hdr);
3115 /* protected by hash lock, if in the hash table */
3116 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3117 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3118 ASSERT(state != arc_anon && state != arc_l2c_only);
3120 (void) refcount_remove_many(&state->arcs_esize[type],
3123 (void) refcount_remove_many(&state->arcs_size, size, hdr);
3124 if (type == ARC_BUFC_METADATA) {
3125 arc_space_return(size, ARC_SPACE_META);
3127 ASSERT(type == ARC_BUFC_DATA);
3128 arc_space_return(size, ARC_SPACE_DATA);
3131 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
3135 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
3136 * data buffer, we transfer the refcount ownership to the hdr and update
3137 * the appropriate kstats.
3140 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3142 arc_state_t *state = hdr->b_l1hdr.b_state;
3144 ASSERT(arc_can_share(hdr, buf));
3145 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3146 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3149 * Start sharing the data buffer. We transfer the
3150 * refcount ownership to the hdr since it always owns
3151 * the refcount whenever an arc_buf_t is shared.
3153 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
3154 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
3155 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
3156 HDR_ISTYPE_METADATA(hdr));
3157 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
3158 buf->b_flags |= ARC_BUF_FLAG_SHARED;
3161 * Since we've transferred ownership to the hdr we need
3162 * to increment its compressed and uncompressed kstats and
3163 * decrement the overhead size.
3165 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3166 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3167 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
3171 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3173 arc_state_t *state = hdr->b_l1hdr.b_state;
3175 ASSERT(arc_buf_is_shared(buf));
3176 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3177 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3180 * We are no longer sharing this buffer so we need
3181 * to transfer its ownership to the rightful owner.
3183 refcount_transfer_ownership(&state->arcs_size, hdr, buf);
3184 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3185 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3186 abd_put(hdr->b_l1hdr.b_pabd);
3187 hdr->b_l1hdr.b_pabd = NULL;
3188 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3191 * Since the buffer is no longer shared between
3192 * the arc buf and the hdr, count it as overhead.
3194 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3195 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3196 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3200 * Remove an arc_buf_t from the hdr's buf list and return the last
3201 * arc_buf_t on the list. If no buffers remain on the list then return
3205 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3207 ASSERT(HDR_HAS_L1HDR(hdr));
3208 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3210 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3211 arc_buf_t *lastbuf = NULL;
3214 * Remove the buf from the hdr list and locate the last
3215 * remaining buffer on the list.
3217 while (*bufp != NULL) {
3219 *bufp = buf->b_next;
3222 * If we've removed a buffer in the middle of
3223 * the list then update the lastbuf and update
3226 if (*bufp != NULL) {
3228 bufp = &(*bufp)->b_next;
3232 ASSERT3P(lastbuf, !=, buf);
3233 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3234 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3235 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3241 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3245 arc_buf_destroy_impl(arc_buf_t *buf)
3247 arc_buf_hdr_t *hdr = buf->b_hdr;
3250 * Free up the data associated with the buf but only if we're not
3251 * sharing this with the hdr. If we are sharing it with the hdr, the
3252 * hdr is responsible for doing the free.
3254 if (buf->b_data != NULL) {
3256 * We're about to change the hdr's b_flags. We must either
3257 * hold the hash_lock or be undiscoverable.
3259 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3261 arc_cksum_verify(buf);
3263 arc_buf_unwatch(buf);
3266 if (arc_buf_is_shared(buf)) {
3267 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3269 uint64_t size = arc_buf_size(buf);
3270 arc_free_data_buf(hdr, buf->b_data, size, buf);
3271 ARCSTAT_INCR(arcstat_overhead_size, -size);
3275 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3276 hdr->b_l1hdr.b_bufcnt -= 1;
3279 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3281 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3283 * If the current arc_buf_t is sharing its data buffer with the
3284 * hdr, then reassign the hdr's b_pabd to share it with the new
3285 * buffer at the end of the list. The shared buffer is always
3286 * the last one on the hdr's buffer list.
3288 * There is an equivalent case for compressed bufs, but since
3289 * they aren't guaranteed to be the last buf in the list and
3290 * that is an exceedingly rare case, we just allow that space be
3291 * wasted temporarily.
3293 if (lastbuf != NULL) {
3294 /* Only one buf can be shared at once */
3295 VERIFY(!arc_buf_is_shared(lastbuf));
3296 /* hdr is uncompressed so can't have compressed buf */
3297 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3299 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3300 arc_hdr_free_pabd(hdr);
3303 * We must setup a new shared block between the
3304 * last buffer and the hdr. The data would have
3305 * been allocated by the arc buf so we need to transfer
3306 * ownership to the hdr since it's now being shared.
3308 arc_share_buf(hdr, lastbuf);
3310 } else if (HDR_SHARED_DATA(hdr)) {
3312 * Uncompressed shared buffers are always at the end
3313 * of the list. Compressed buffers don't have the
3314 * same requirements. This makes it hard to
3315 * simply assert that the lastbuf is shared so
3316 * we rely on the hdr's compression flags to determine
3317 * if we have a compressed, shared buffer.
3319 ASSERT3P(lastbuf, !=, NULL);
3320 ASSERT(arc_buf_is_shared(lastbuf) ||
3321 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
3325 * Free the checksum if we're removing the last uncompressed buf from
3328 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3329 arc_cksum_free(hdr);
3332 /* clean up the buf */
3334 kmem_cache_free(buf_cache, buf);
3338 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr)
3340 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3341 ASSERT(HDR_HAS_L1HDR(hdr));
3342 ASSERT(!HDR_SHARED_DATA(hdr));
3344 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3345 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
3346 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3347 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3349 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3350 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3354 arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
3356 ASSERT(HDR_HAS_L1HDR(hdr));
3357 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3360 * If the hdr is currently being written to the l2arc then
3361 * we defer freeing the data by adding it to the l2arc_free_on_write
3362 * list. The l2arc will free the data once it's finished
3363 * writing it to the l2arc device.
3365 if (HDR_L2_WRITING(hdr)) {
3366 arc_hdr_free_on_write(hdr);
3367 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3369 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
3370 arc_hdr_size(hdr), hdr);
3372 hdr->b_l1hdr.b_pabd = NULL;
3373 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3375 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3376 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3379 static arc_buf_hdr_t *
3380 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3381 enum zio_compress compression_type, arc_buf_contents_t type)
3385 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3387 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3388 ASSERT(HDR_EMPTY(hdr));
3389 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3390 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
3391 HDR_SET_PSIZE(hdr, psize);
3392 HDR_SET_LSIZE(hdr, lsize);
3396 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3397 arc_hdr_set_compress(hdr, compression_type);
3399 hdr->b_l1hdr.b_state = arc_anon;
3400 hdr->b_l1hdr.b_arc_access = 0;
3401 hdr->b_l1hdr.b_bufcnt = 0;
3402 hdr->b_l1hdr.b_buf = NULL;
3405 * Allocate the hdr's buffer. This will contain either
3406 * the compressed or uncompressed data depending on the block
3407 * it references and compressed arc enablement.
3409 arc_hdr_alloc_pabd(hdr);
3410 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3416 * Transition between the two allocation states for the arc_buf_hdr struct.
3417 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3418 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3419 * version is used when a cache buffer is only in the L2ARC in order to reduce
3422 static arc_buf_hdr_t *
3423 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3425 ASSERT(HDR_HAS_L2HDR(hdr));
3427 arc_buf_hdr_t *nhdr;
3428 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3430 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3431 (old == hdr_l2only_cache && new == hdr_full_cache));
3433 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3435 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3436 buf_hash_remove(hdr);
3438 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3440 if (new == hdr_full_cache) {
3441 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3443 * arc_access and arc_change_state need to be aware that a
3444 * header has just come out of L2ARC, so we set its state to
3445 * l2c_only even though it's about to change.
3447 nhdr->b_l1hdr.b_state = arc_l2c_only;
3449 /* Verify previous threads set to NULL before freeing */
3450 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3452 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3453 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3454 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3457 * If we've reached here, We must have been called from
3458 * arc_evict_hdr(), as such we should have already been
3459 * removed from any ghost list we were previously on
3460 * (which protects us from racing with arc_evict_state),
3461 * thus no locking is needed during this check.
3463 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3466 * A buffer must not be moved into the arc_l2c_only
3467 * state if it's not finished being written out to the
3468 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3469 * might try to be accessed, even though it was removed.
3471 VERIFY(!HDR_L2_WRITING(hdr));
3472 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3475 if (hdr->b_l1hdr.b_thawed != NULL) {
3476 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3477 hdr->b_l1hdr.b_thawed = NULL;
3481 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3484 * The header has been reallocated so we need to re-insert it into any
3487 (void) buf_hash_insert(nhdr, NULL);
3489 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3491 mutex_enter(&dev->l2ad_mtx);
3494 * We must place the realloc'ed header back into the list at
3495 * the same spot. Otherwise, if it's placed earlier in the list,
3496 * l2arc_write_buffers() could find it during the function's
3497 * write phase, and try to write it out to the l2arc.
3499 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3500 list_remove(&dev->l2ad_buflist, hdr);
3502 mutex_exit(&dev->l2ad_mtx);
3505 * Since we're using the pointer address as the tag when
3506 * incrementing and decrementing the l2ad_alloc refcount, we
3507 * must remove the old pointer (that we're about to destroy) and
3508 * add the new pointer to the refcount. Otherwise we'd remove
3509 * the wrong pointer address when calling arc_hdr_destroy() later.
3512 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
3513 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
3515 buf_discard_identity(hdr);
3516 kmem_cache_free(old, hdr);
3522 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3523 * The buf is returned thawed since we expect the consumer to modify it.
3526 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3528 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3529 ZIO_COMPRESS_OFF, type);
3530 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3532 arc_buf_t *buf = NULL;
3533 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf));
3540 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3541 * for bufs containing metadata.
3544 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3545 enum zio_compress compression_type)
3547 ASSERT3U(lsize, >, 0);
3548 ASSERT3U(lsize, >=, psize);
3549 ASSERT(compression_type > ZIO_COMPRESS_OFF);
3550 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3552 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3553 compression_type, ARC_BUFC_DATA);
3554 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3556 arc_buf_t *buf = NULL;
3557 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3559 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3561 if (!arc_buf_is_shared(buf)) {
3563 * To ensure that the hdr has the correct data in it if we call
3564 * arc_decompress() on this buf before it's been written to
3565 * disk, it's easiest if we just set up sharing between the
3568 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3569 arc_hdr_free_pabd(hdr);
3570 arc_share_buf(hdr, buf);
3577 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3579 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3580 l2arc_dev_t *dev = l2hdr->b_dev;
3581 uint64_t psize = arc_hdr_size(hdr);
3583 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3584 ASSERT(HDR_HAS_L2HDR(hdr));
3586 list_remove(&dev->l2ad_buflist, hdr);
3588 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3589 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3591 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3593 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3594 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3598 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3600 if (HDR_HAS_L1HDR(hdr)) {
3601 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3602 hdr->b_l1hdr.b_bufcnt > 0);
3603 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3604 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3606 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3607 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3609 if (!HDR_EMPTY(hdr))
3610 buf_discard_identity(hdr);
3612 if (HDR_HAS_L2HDR(hdr)) {
3613 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3614 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3617 mutex_enter(&dev->l2ad_mtx);
3620 * Even though we checked this conditional above, we
3621 * need to check this again now that we have the
3622 * l2ad_mtx. This is because we could be racing with
3623 * another thread calling l2arc_evict() which might have
3624 * destroyed this header's L2 portion as we were waiting
3625 * to acquire the l2ad_mtx. If that happens, we don't
3626 * want to re-destroy the header's L2 portion.
3628 if (HDR_HAS_L2HDR(hdr)) {
3630 arc_hdr_l2hdr_destroy(hdr);
3634 mutex_exit(&dev->l2ad_mtx);
3637 if (HDR_HAS_L1HDR(hdr)) {
3638 arc_cksum_free(hdr);
3640 while (hdr->b_l1hdr.b_buf != NULL)
3641 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3644 if (hdr->b_l1hdr.b_thawed != NULL) {
3645 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3646 hdr->b_l1hdr.b_thawed = NULL;
3650 if (hdr->b_l1hdr.b_pabd != NULL) {
3651 arc_hdr_free_pabd(hdr);
3655 ASSERT3P(hdr->b_hash_next, ==, NULL);
3656 if (HDR_HAS_L1HDR(hdr)) {
3657 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3658 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3659 kmem_cache_free(hdr_full_cache, hdr);
3661 kmem_cache_free(hdr_l2only_cache, hdr);
3666 arc_buf_destroy(arc_buf_t *buf, void* tag)
3668 arc_buf_hdr_t *hdr = buf->b_hdr;
3669 kmutex_t *hash_lock = HDR_LOCK(hdr);
3671 if (hdr->b_l1hdr.b_state == arc_anon) {
3672 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3673 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3674 VERIFY0(remove_reference(hdr, NULL, tag));
3675 arc_hdr_destroy(hdr);
3679 mutex_enter(hash_lock);
3680 ASSERT3P(hdr, ==, buf->b_hdr);
3681 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3682 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3683 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3684 ASSERT3P(buf->b_data, !=, NULL);
3686 (void) remove_reference(hdr, hash_lock, tag);
3687 arc_buf_destroy_impl(buf);
3688 mutex_exit(hash_lock);
3692 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3693 * state of the header is dependent on its state prior to entering this
3694 * function. The following transitions are possible:
3696 * - arc_mru -> arc_mru_ghost
3697 * - arc_mfu -> arc_mfu_ghost
3698 * - arc_mru_ghost -> arc_l2c_only
3699 * - arc_mru_ghost -> deleted
3700 * - arc_mfu_ghost -> arc_l2c_only
3701 * - arc_mfu_ghost -> deleted
3704 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3706 arc_state_t *evicted_state, *state;
3707 int64_t bytes_evicted = 0;
3708 int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3709 zfs_arc_min_prescient_prefetch_ms : zfs_arc_min_prefetch_ms;
3711 ASSERT(MUTEX_HELD(hash_lock));
3712 ASSERT(HDR_HAS_L1HDR(hdr));
3714 state = hdr->b_l1hdr.b_state;
3715 if (GHOST_STATE(state)) {
3716 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3717 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3720 * l2arc_write_buffers() relies on a header's L1 portion
3721 * (i.e. its b_pabd field) during it's write phase.
3722 * Thus, we cannot push a header onto the arc_l2c_only
3723 * state (removing it's L1 piece) until the header is
3724 * done being written to the l2arc.
3726 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3727 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3728 return (bytes_evicted);
3731 ARCSTAT_BUMP(arcstat_deleted);
3732 bytes_evicted += HDR_GET_LSIZE(hdr);
3734 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3736 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3737 if (HDR_HAS_L2HDR(hdr)) {
3739 * This buffer is cached on the 2nd Level ARC;
3740 * don't destroy the header.
3742 arc_change_state(arc_l2c_only, hdr, hash_lock);
3744 * dropping from L1+L2 cached to L2-only,
3745 * realloc to remove the L1 header.
3747 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3750 arc_change_state(arc_anon, hdr, hash_lock);
3751 arc_hdr_destroy(hdr);
3753 return (bytes_evicted);
3756 ASSERT(state == arc_mru || state == arc_mfu);
3757 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3759 /* prefetch buffers have a minimum lifespan */
3760 if (HDR_IO_IN_PROGRESS(hdr) ||
3761 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3762 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < min_lifetime * hz)) {
3763 ARCSTAT_BUMP(arcstat_evict_skip);
3764 return (bytes_evicted);
3767 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3768 while (hdr->b_l1hdr.b_buf) {
3769 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3770 if (!mutex_tryenter(&buf->b_evict_lock)) {
3771 ARCSTAT_BUMP(arcstat_mutex_miss);
3774 if (buf->b_data != NULL)
3775 bytes_evicted += HDR_GET_LSIZE(hdr);
3776 mutex_exit(&buf->b_evict_lock);
3777 arc_buf_destroy_impl(buf);
3780 if (HDR_HAS_L2HDR(hdr)) {
3781 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3783 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3784 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3785 HDR_GET_LSIZE(hdr));
3787 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3788 HDR_GET_LSIZE(hdr));
3792 if (hdr->b_l1hdr.b_bufcnt == 0) {
3793 arc_cksum_free(hdr);
3795 bytes_evicted += arc_hdr_size(hdr);
3798 * If this hdr is being evicted and has a compressed
3799 * buffer then we discard it here before we change states.
3800 * This ensures that the accounting is updated correctly
3801 * in arc_free_data_impl().
3803 arc_hdr_free_pabd(hdr);
3805 arc_change_state(evicted_state, hdr, hash_lock);
3806 ASSERT(HDR_IN_HASH_TABLE(hdr));
3807 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3808 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3811 return (bytes_evicted);
3815 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3816 uint64_t spa, int64_t bytes)
3818 multilist_sublist_t *mls;
3819 uint64_t bytes_evicted = 0;
3821 kmutex_t *hash_lock;
3822 int evict_count = 0;
3824 ASSERT3P(marker, !=, NULL);
3825 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3827 mls = multilist_sublist_lock(ml, idx);
3829 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3830 hdr = multilist_sublist_prev(mls, marker)) {
3831 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3832 (evict_count >= zfs_arc_evict_batch_limit))
3836 * To keep our iteration location, move the marker
3837 * forward. Since we're not holding hdr's hash lock, we
3838 * must be very careful and not remove 'hdr' from the
3839 * sublist. Otherwise, other consumers might mistake the
3840 * 'hdr' as not being on a sublist when they call the
3841 * multilist_link_active() function (they all rely on
3842 * the hash lock protecting concurrent insertions and
3843 * removals). multilist_sublist_move_forward() was
3844 * specifically implemented to ensure this is the case
3845 * (only 'marker' will be removed and re-inserted).
3847 multilist_sublist_move_forward(mls, marker);
3850 * The only case where the b_spa field should ever be
3851 * zero, is the marker headers inserted by
3852 * arc_evict_state(). It's possible for multiple threads
3853 * to be calling arc_evict_state() concurrently (e.g.
3854 * dsl_pool_close() and zio_inject_fault()), so we must
3855 * skip any markers we see from these other threads.
3857 if (hdr->b_spa == 0)
3860 /* we're only interested in evicting buffers of a certain spa */
3861 if (spa != 0 && hdr->b_spa != spa) {
3862 ARCSTAT_BUMP(arcstat_evict_skip);
3866 hash_lock = HDR_LOCK(hdr);
3869 * We aren't calling this function from any code path
3870 * that would already be holding a hash lock, so we're
3871 * asserting on this assumption to be defensive in case
3872 * this ever changes. Without this check, it would be
3873 * possible to incorrectly increment arcstat_mutex_miss
3874 * below (e.g. if the code changed such that we called
3875 * this function with a hash lock held).
3877 ASSERT(!MUTEX_HELD(hash_lock));
3879 if (mutex_tryenter(hash_lock)) {
3880 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3881 mutex_exit(hash_lock);
3883 bytes_evicted += evicted;
3886 * If evicted is zero, arc_evict_hdr() must have
3887 * decided to skip this header, don't increment
3888 * evict_count in this case.
3894 * If arc_size isn't overflowing, signal any
3895 * threads that might happen to be waiting.
3897 * For each header evicted, we wake up a single
3898 * thread. If we used cv_broadcast, we could
3899 * wake up "too many" threads causing arc_size
3900 * to significantly overflow arc_c; since
3901 * arc_get_data_impl() doesn't check for overflow
3902 * when it's woken up (it doesn't because it's
3903 * possible for the ARC to be overflowing while
3904 * full of un-evictable buffers, and the
3905 * function should proceed in this case).
3907 * If threads are left sleeping, due to not
3908 * using cv_broadcast, they will be woken up
3909 * just before arc_reclaim_thread() sleeps.
3911 mutex_enter(&arc_reclaim_lock);
3912 if (!arc_is_overflowing())
3913 cv_signal(&arc_reclaim_waiters_cv);
3914 mutex_exit(&arc_reclaim_lock);
3916 ARCSTAT_BUMP(arcstat_mutex_miss);
3920 multilist_sublist_unlock(mls);
3922 return (bytes_evicted);
3926 * Evict buffers from the given arc state, until we've removed the
3927 * specified number of bytes. Move the removed buffers to the
3928 * appropriate evict state.
3930 * This function makes a "best effort". It skips over any buffers
3931 * it can't get a hash_lock on, and so, may not catch all candidates.
3932 * It may also return without evicting as much space as requested.
3934 * If bytes is specified using the special value ARC_EVICT_ALL, this
3935 * will evict all available (i.e. unlocked and evictable) buffers from
3936 * the given arc state; which is used by arc_flush().
3939 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3940 arc_buf_contents_t type)
3942 uint64_t total_evicted = 0;
3943 multilist_t *ml = state->arcs_list[type];
3945 arc_buf_hdr_t **markers;
3947 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3949 num_sublists = multilist_get_num_sublists(ml);
3952 * If we've tried to evict from each sublist, made some
3953 * progress, but still have not hit the target number of bytes
3954 * to evict, we want to keep trying. The markers allow us to
3955 * pick up where we left off for each individual sublist, rather
3956 * than starting from the tail each time.
3958 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3959 for (int i = 0; i < num_sublists; i++) {
3960 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3963 * A b_spa of 0 is used to indicate that this header is
3964 * a marker. This fact is used in arc_adjust_type() and
3965 * arc_evict_state_impl().
3967 markers[i]->b_spa = 0;
3969 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3970 multilist_sublist_insert_tail(mls, markers[i]);
3971 multilist_sublist_unlock(mls);
3975 * While we haven't hit our target number of bytes to evict, or
3976 * we're evicting all available buffers.
3978 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3979 int sublist_idx = multilist_get_random_index(ml);
3980 uint64_t scan_evicted = 0;
3983 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
3984 * Request that 10% of the LRUs be scanned by the superblock
3987 if (type == ARC_BUFC_DATA && aggsum_compare(&astat_dnode_size,
3988 arc_dnode_limit) > 0) {
3989 arc_prune_async((aggsum_upper_bound(&astat_dnode_size) -
3990 arc_dnode_limit) / sizeof (dnode_t) /
3991 zfs_arc_dnode_reduce_percent);
3995 * Start eviction using a randomly selected sublist,
3996 * this is to try and evenly balance eviction across all
3997 * sublists. Always starting at the same sublist
3998 * (e.g. index 0) would cause evictions to favor certain
3999 * sublists over others.
4001 for (int i = 0; i < num_sublists; i++) {
4002 uint64_t bytes_remaining;
4003 uint64_t bytes_evicted;
4005 if (bytes == ARC_EVICT_ALL)
4006 bytes_remaining = ARC_EVICT_ALL;
4007 else if (total_evicted < bytes)
4008 bytes_remaining = bytes - total_evicted;
4012 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
4013 markers[sublist_idx], spa, bytes_remaining);
4015 scan_evicted += bytes_evicted;
4016 total_evicted += bytes_evicted;
4018 /* we've reached the end, wrap to the beginning */
4019 if (++sublist_idx >= num_sublists)
4024 * If we didn't evict anything during this scan, we have
4025 * no reason to believe we'll evict more during another
4026 * scan, so break the loop.
4028 if (scan_evicted == 0) {
4029 /* This isn't possible, let's make that obvious */
4030 ASSERT3S(bytes, !=, 0);
4033 * When bytes is ARC_EVICT_ALL, the only way to
4034 * break the loop is when scan_evicted is zero.
4035 * In that case, we actually have evicted enough,
4036 * so we don't want to increment the kstat.
4038 if (bytes != ARC_EVICT_ALL) {
4039 ASSERT3S(total_evicted, <, bytes);
4040 ARCSTAT_BUMP(arcstat_evict_not_enough);
4047 for (int i = 0; i < num_sublists; i++) {
4048 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4049 multilist_sublist_remove(mls, markers[i]);
4050 multilist_sublist_unlock(mls);
4052 kmem_cache_free(hdr_full_cache, markers[i]);
4054 kmem_free(markers, sizeof (*markers) * num_sublists);
4056 return (total_evicted);
4060 * Flush all "evictable" data of the given type from the arc state
4061 * specified. This will not evict any "active" buffers (i.e. referenced).
4063 * When 'retry' is set to B_FALSE, the function will make a single pass
4064 * over the state and evict any buffers that it can. Since it doesn't
4065 * continually retry the eviction, it might end up leaving some buffers
4066 * in the ARC due to lock misses.
4068 * When 'retry' is set to B_TRUE, the function will continually retry the
4069 * eviction until *all* evictable buffers have been removed from the
4070 * state. As a result, if concurrent insertions into the state are
4071 * allowed (e.g. if the ARC isn't shutting down), this function might
4072 * wind up in an infinite loop, continually trying to evict buffers.
4075 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4078 uint64_t evicted = 0;
4080 while (refcount_count(&state->arcs_esize[type]) != 0) {
4081 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
4091 * Helper function for arc_prune_async() it is responsible for safely
4092 * handling the execution of a registered arc_prune_func_t.
4095 arc_prune_task(void *ptr)
4097 arc_prune_t *ap = (arc_prune_t *)ptr;
4098 arc_prune_func_t *func = ap->p_pfunc;
4101 func(ap->p_adjust, ap->p_private);
4103 refcount_remove(&ap->p_refcnt, func);
4107 * Notify registered consumers they must drop holds on a portion of the ARC
4108 * buffered they reference. This provides a mechanism to ensure the ARC can
4109 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
4110 * is analogous to dnlc_reduce_cache() but more generic.
4112 * This operation is performed asynchronously so it may be safely called
4113 * in the context of the arc_reclaim_thread(). A reference is taken here
4114 * for each registered arc_prune_t and the arc_prune_task() is responsible
4115 * for releasing it once the registered arc_prune_func_t has completed.
4118 arc_prune_async(int64_t adjust)
4122 mutex_enter(&arc_prune_mtx);
4123 for (ap = list_head(&arc_prune_list); ap != NULL;
4124 ap = list_next(&arc_prune_list, ap)) {
4126 if (refcount_count(&ap->p_refcnt) >= 2)
4129 refcount_add(&ap->p_refcnt, ap->p_pfunc);
4130 ap->p_adjust = adjust;
4131 if (taskq_dispatch(arc_prune_taskq, arc_prune_task,
4132 ap, TQ_SLEEP) == TASKQID_INVALID) {
4133 refcount_remove(&ap->p_refcnt, ap->p_pfunc);
4136 ARCSTAT_BUMP(arcstat_prune);
4138 mutex_exit(&arc_prune_mtx);
4142 * Evict the specified number of bytes from the state specified,
4143 * restricting eviction to the spa and type given. This function
4144 * prevents us from trying to evict more from a state's list than
4145 * is "evictable", and to skip evicting altogether when passed a
4146 * negative value for "bytes". In contrast, arc_evict_state() will
4147 * evict everything it can, when passed a negative value for "bytes".
4150 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
4151 arc_buf_contents_t type)
4155 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
4156 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
4157 return (arc_evict_state(state, spa, delta, type));
4164 * The goal of this function is to evict enough meta data buffers from the
4165 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
4166 * more complicated than it appears because it is common for data buffers
4167 * to have holds on meta data buffers. In addition, dnode meta data buffers
4168 * will be held by the dnodes in the block preventing them from being freed.
4169 * This means we can't simply traverse the ARC and expect to always find
4170 * enough unheld meta data buffer to release.
4172 * Therefore, this function has been updated to make alternating passes
4173 * over the ARC releasing data buffers and then newly unheld meta data
4174 * buffers. This ensures forward progress is maintained and meta_used
4175 * will decrease. Normally this is sufficient, but if required the ARC
4176 * will call the registered prune callbacks causing dentry and inodes to
4177 * be dropped from the VFS cache. This will make dnode meta data buffers
4178 * available for reclaim.
4181 arc_adjust_meta_balanced(uint64_t meta_used)
4183 int64_t delta, prune = 0, adjustmnt;
4184 uint64_t total_evicted = 0;
4185 arc_buf_contents_t type = ARC_BUFC_DATA;
4186 int restarts = MAX(zfs_arc_meta_adjust_restarts, 0);
4190 * This slightly differs than the way we evict from the mru in
4191 * arc_adjust because we don't have a "target" value (i.e. no
4192 * "meta" arc_p). As a result, I think we can completely
4193 * cannibalize the metadata in the MRU before we evict the
4194 * metadata from the MFU. I think we probably need to implement a
4195 * "metadata arc_p" value to do this properly.
4197 adjustmnt = meta_used - arc_meta_limit;
4199 if (adjustmnt > 0 && refcount_count(&arc_mru->arcs_esize[type]) > 0) {
4200 delta = MIN(refcount_count(&arc_mru->arcs_esize[type]),
4202 total_evicted += arc_adjust_impl(arc_mru, 0, delta, type);
4207 * We can't afford to recalculate adjustmnt here. If we do,
4208 * new metadata buffers can sneak into the MRU or ANON lists,
4209 * thus penalize the MFU metadata. Although the fudge factor is
4210 * small, it has been empirically shown to be significant for
4211 * certain workloads (e.g. creating many empty directories). As
4212 * such, we use the original calculation for adjustmnt, and
4213 * simply decrement the amount of data evicted from the MRU.
4216 if (adjustmnt > 0 && refcount_count(&arc_mfu->arcs_esize[type]) > 0) {
4217 delta = MIN(refcount_count(&arc_mfu->arcs_esize[type]),
4219 total_evicted += arc_adjust_impl(arc_mfu, 0, delta, type);
4222 adjustmnt = meta_used - arc_meta_limit;
4224 if (adjustmnt > 0 &&
4225 refcount_count(&arc_mru_ghost->arcs_esize[type]) > 0) {
4226 delta = MIN(adjustmnt,
4227 refcount_count(&arc_mru_ghost->arcs_esize[type]));
4228 total_evicted += arc_adjust_impl(arc_mru_ghost, 0, delta, type);
4232 if (adjustmnt > 0 &&
4233 refcount_count(&arc_mfu_ghost->arcs_esize[type]) > 0) {
4234 delta = MIN(adjustmnt,
4235 refcount_count(&arc_mfu_ghost->arcs_esize[type]));
4236 total_evicted += arc_adjust_impl(arc_mfu_ghost, 0, delta, type);
4240 * If after attempting to make the requested adjustment to the ARC
4241 * the meta limit is still being exceeded then request that the
4242 * higher layers drop some cached objects which have holds on ARC
4243 * meta buffers. Requests to the upper layers will be made with
4244 * increasingly large scan sizes until the ARC is below the limit.
4246 if (meta_used > arc_meta_limit) {
4247 if (type == ARC_BUFC_DATA) {
4248 type = ARC_BUFC_METADATA;
4250 type = ARC_BUFC_DATA;
4252 if (zfs_arc_meta_prune) {
4253 prune += zfs_arc_meta_prune;
4254 arc_prune_async(prune);
4263 return (total_evicted);
4267 * Evict metadata buffers from the cache, such that arc_meta_used is
4268 * capped by the arc_meta_limit tunable.
4271 arc_adjust_meta_only(uint64_t meta_used)
4273 uint64_t total_evicted = 0;
4277 * If we're over the meta limit, we want to evict enough
4278 * metadata to get back under the meta limit. We don't want to
4279 * evict so much that we drop the MRU below arc_p, though. If
4280 * we're over the meta limit more than we're over arc_p, we
4281 * evict some from the MRU here, and some from the MFU below.
4283 target = MIN((int64_t)(meta_used - arc_meta_limit),
4284 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4285 refcount_count(&arc_mru->arcs_size) - arc_p));
4287 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4290 * Similar to the above, we want to evict enough bytes to get us
4291 * below the meta limit, but not so much as to drop us below the
4292 * space allotted to the MFU (which is defined as arc_c - arc_p).
4294 target = MIN((int64_t)(meta_used - arc_meta_limit),
4295 (int64_t)(refcount_count(&arc_mfu->arcs_size) -
4298 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4300 return (total_evicted);
4304 arc_adjust_meta(uint64_t meta_used)
4306 if (zfs_arc_meta_strategy == ARC_STRATEGY_META_ONLY)
4307 return (arc_adjust_meta_only(meta_used));
4309 return (arc_adjust_meta_balanced(meta_used));
4313 * Return the type of the oldest buffer in the given arc state
4315 * This function will select a random sublist of type ARC_BUFC_DATA and
4316 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4317 * is compared, and the type which contains the "older" buffer will be
4320 static arc_buf_contents_t
4321 arc_adjust_type(arc_state_t *state)
4323 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
4324 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
4325 int data_idx = multilist_get_random_index(data_ml);
4326 int meta_idx = multilist_get_random_index(meta_ml);
4327 multilist_sublist_t *data_mls;
4328 multilist_sublist_t *meta_mls;
4329 arc_buf_contents_t type;
4330 arc_buf_hdr_t *data_hdr;
4331 arc_buf_hdr_t *meta_hdr;
4334 * We keep the sublist lock until we're finished, to prevent
4335 * the headers from being destroyed via arc_evict_state().
4337 data_mls = multilist_sublist_lock(data_ml, data_idx);
4338 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4341 * These two loops are to ensure we skip any markers that
4342 * might be at the tail of the lists due to arc_evict_state().
4345 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4346 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4347 if (data_hdr->b_spa != 0)
4351 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4352 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4353 if (meta_hdr->b_spa != 0)
4357 if (data_hdr == NULL && meta_hdr == NULL) {
4358 type = ARC_BUFC_DATA;
4359 } else if (data_hdr == NULL) {
4360 ASSERT3P(meta_hdr, !=, NULL);
4361 type = ARC_BUFC_METADATA;
4362 } else if (meta_hdr == NULL) {
4363 ASSERT3P(data_hdr, !=, NULL);
4364 type = ARC_BUFC_DATA;
4366 ASSERT3P(data_hdr, !=, NULL);
4367 ASSERT3P(meta_hdr, !=, NULL);
4369 /* The headers can't be on the sublist without an L1 header */
4370 ASSERT(HDR_HAS_L1HDR(data_hdr));
4371 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4373 if (data_hdr->b_l1hdr.b_arc_access <
4374 meta_hdr->b_l1hdr.b_arc_access) {
4375 type = ARC_BUFC_DATA;
4377 type = ARC_BUFC_METADATA;
4381 multilist_sublist_unlock(meta_mls);
4382 multilist_sublist_unlock(data_mls);
4388 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4393 uint64_t total_evicted = 0;
4396 uint64_t asize = aggsum_value(&arc_size);
4397 uint64_t ameta = aggsum_value(&arc_meta_used);
4400 * If we're over arc_meta_limit, we want to correct that before
4401 * potentially evicting data buffers below.
4403 total_evicted += arc_adjust_meta(ameta);
4408 * If we're over the target cache size, we want to evict enough
4409 * from the list to get back to our target size. We don't want
4410 * to evict too much from the MRU, such that it drops below
4411 * arc_p. So, if we're over our target cache size more than
4412 * the MRU is over arc_p, we'll evict enough to get back to
4413 * arc_p here, and then evict more from the MFU below.
4415 target = MIN((int64_t)(asize - arc_c),
4416 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4417 refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4420 * If we're below arc_meta_min, always prefer to evict data.
4421 * Otherwise, try to satisfy the requested number of bytes to
4422 * evict from the type which contains older buffers; in an
4423 * effort to keep newer buffers in the cache regardless of their
4424 * type. If we cannot satisfy the number of bytes from this
4425 * type, spill over into the next type.
4427 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4428 ameta > arc_meta_min) {
4429 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4430 total_evicted += bytes;
4433 * If we couldn't evict our target number of bytes from
4434 * metadata, we try to get the rest from data.
4439 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4441 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4442 total_evicted += bytes;
4445 * If we couldn't evict our target number of bytes from
4446 * data, we try to get the rest from metadata.
4451 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4455 * Re-sum ARC stats after the first round of evictions.
4457 asize = aggsum_value(&arc_size);
4458 ameta = aggsum_value(&arc_meta_used);
4463 * Now that we've tried to evict enough from the MRU to get its
4464 * size back to arc_p, if we're still above the target cache
4465 * size, we evict the rest from the MFU.
4467 target = asize - arc_c;
4469 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4470 ameta > arc_meta_min) {
4471 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4472 total_evicted += bytes;
4475 * If we couldn't evict our target number of bytes from
4476 * metadata, we try to get the rest from data.
4481 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4483 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4484 total_evicted += bytes;
4487 * If we couldn't evict our target number of bytes from
4488 * data, we try to get the rest from data.
4493 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4497 * Adjust ghost lists
4499 * In addition to the above, the ARC also defines target values
4500 * for the ghost lists. The sum of the mru list and mru ghost
4501 * list should never exceed the target size of the cache, and
4502 * the sum of the mru list, mfu list, mru ghost list, and mfu
4503 * ghost list should never exceed twice the target size of the
4504 * cache. The following logic enforces these limits on the ghost
4505 * caches, and evicts from them as needed.
4507 target = refcount_count(&arc_mru->arcs_size) +
4508 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4510 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4511 total_evicted += bytes;
4516 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4519 * We assume the sum of the mru list and mfu list is less than
4520 * or equal to arc_c (we enforced this above), which means we
4521 * can use the simpler of the two equations below:
4523 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4524 * mru ghost + mfu ghost <= arc_c
4526 target = refcount_count(&arc_mru_ghost->arcs_size) +
4527 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4529 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4530 total_evicted += bytes;
4535 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4537 return (total_evicted);
4541 arc_flush(spa_t *spa, boolean_t retry)
4546 * If retry is B_TRUE, a spa must not be specified since we have
4547 * no good way to determine if all of a spa's buffers have been
4548 * evicted from an arc state.
4550 ASSERT(!retry || spa == 0);
4553 guid = spa_load_guid(spa);
4555 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4556 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4558 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4559 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4561 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4562 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4564 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4565 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4569 arc_shrink(int64_t to_free)
4571 uint64_t asize = aggsum_value(&arc_size);
4572 if (arc_c > arc_c_min) {
4573 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
4574 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
4575 if (arc_c > arc_c_min + to_free)
4576 atomic_add_64(&arc_c, -to_free);
4580 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4582 arc_c = MAX(asize, arc_c_min);
4584 arc_p = (arc_c >> 1);
4586 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
4589 ASSERT(arc_c >= arc_c_min);
4590 ASSERT((int64_t)arc_p >= 0);
4593 if (asize > arc_c) {
4594 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, asize,
4596 return (arc_adjust());
4601 typedef enum free_memory_reason_t {
4606 FMR_PAGES_PP_MAXIMUM,
4609 } free_memory_reason_t;
4611 int64_t last_free_memory;
4612 free_memory_reason_t last_free_reason;
4615 * Additional reserve of pages for pp_reserve.
4617 int64_t arc_pages_pp_reserve = 64;
4620 * Additional reserve of pages for swapfs.
4622 int64_t arc_swapfs_reserve = 64;
4625 * Return the amount of memory that can be consumed before reclaim will be
4626 * needed. Positive if there is sufficient free memory, negative indicates
4627 * the amount of memory that needs to be freed up.
4630 arc_available_memory(void)
4632 int64_t lowest = INT64_MAX;
4634 free_memory_reason_t r = FMR_UNKNOWN;
4639 * Cooperate with pagedaemon when it's time for it to scan
4640 * and reclaim some pages.
4642 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
4650 n = PAGESIZE * (-needfree);
4658 * check that we're out of range of the pageout scanner. It starts to
4659 * schedule paging if freemem is less than lotsfree and needfree.
4660 * lotsfree is the high-water mark for pageout, and needfree is the
4661 * number of needed free pages. We add extra pages here to make sure
4662 * the scanner doesn't start up while we're freeing memory.
4664 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4671 * check to make sure that swapfs has enough space so that anon
4672 * reservations can still succeed. anon_resvmem() checks that the
4673 * availrmem is greater than swapfs_minfree, and the number of reserved
4674 * swap pages. We also add a bit of extra here just to prevent
4675 * circumstances from getting really dire.
4677 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4678 desfree - arc_swapfs_reserve);
4681 r = FMR_SWAPFS_MINFREE;
4686 * Check that we have enough availrmem that memory locking (e.g., via
4687 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4688 * stores the number of pages that cannot be locked; when availrmem
4689 * drops below pages_pp_maximum, page locking mechanisms such as
4690 * page_pp_lock() will fail.)
4692 n = PAGESIZE * (availrmem - pages_pp_maximum -
4693 arc_pages_pp_reserve);
4696 r = FMR_PAGES_PP_MAXIMUM;
4699 #endif /* __FreeBSD__ */
4700 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
4702 * If we're on an i386 platform, it's possible that we'll exhaust the
4703 * kernel heap space before we ever run out of available physical
4704 * memory. Most checks of the size of the heap_area compare against
4705 * tune.t_minarmem, which is the minimum available real memory that we
4706 * can have in the system. However, this is generally fixed at 25 pages
4707 * which is so low that it's useless. In this comparison, we seek to
4708 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4709 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4712 n = uma_avail() - (long)(uma_limit() / 4);
4720 * If zio data pages are being allocated out of a separate heap segment,
4721 * then enforce that the size of available vmem for this arena remains
4722 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4724 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4725 * memory (in the zio_arena) free, which can avoid memory
4726 * fragmentation issues.
4728 if (zio_arena != NULL) {
4729 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4730 (vmem_size(zio_arena, VMEM_ALLOC) >>
4731 arc_zio_arena_free_shift);
4739 /* Every 100 calls, free a small amount */
4740 if (spa_get_random(100) == 0)
4742 #endif /* _KERNEL */
4744 last_free_memory = lowest;
4745 last_free_reason = r;
4746 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
4752 * Determine if the system is under memory pressure and is asking
4753 * to reclaim memory. A return value of B_TRUE indicates that the system
4754 * is under memory pressure and that the arc should adjust accordingly.
4757 arc_reclaim_needed(void)
4759 return (arc_available_memory() < 0);
4762 extern kmem_cache_t *zio_buf_cache[];
4763 extern kmem_cache_t *zio_data_buf_cache[];
4764 extern kmem_cache_t *range_seg_cache;
4765 extern kmem_cache_t *abd_chunk_cache;
4767 static __noinline void
4768 arc_kmem_reap_now(void)
4771 kmem_cache_t *prev_cache = NULL;
4772 kmem_cache_t *prev_data_cache = NULL;
4774 DTRACE_PROBE(arc__kmem_reap_start);
4776 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) {
4778 * We are exceeding our meta-data cache limit.
4779 * Purge some DNLC entries to release holds on meta-data.
4781 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4785 * Reclaim unused memory from all kmem caches.
4792 * If a kmem reap is already active, don't schedule more. We must
4793 * check for this because kmem_cache_reap_soon() won't actually
4794 * block on the cache being reaped (this is to prevent callers from
4795 * becoming implicitly blocked by a system-wide kmem reap -- which,
4796 * on a system with many, many full magazines, can take minutes).
4798 if (kmem_cache_reap_active())
4801 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4802 if (zio_buf_cache[i] != prev_cache) {
4803 prev_cache = zio_buf_cache[i];
4804 kmem_cache_reap_soon(zio_buf_cache[i]);
4806 if (zio_data_buf_cache[i] != prev_data_cache) {
4807 prev_data_cache = zio_data_buf_cache[i];
4808 kmem_cache_reap_soon(zio_data_buf_cache[i]);
4811 kmem_cache_reap_soon(abd_chunk_cache);
4812 kmem_cache_reap_soon(buf_cache);
4813 kmem_cache_reap_soon(hdr_full_cache);
4814 kmem_cache_reap_soon(hdr_l2only_cache);
4815 kmem_cache_reap_soon(range_seg_cache);
4818 if (zio_arena != NULL) {
4820 * Ask the vmem arena to reclaim unused memory from its
4823 vmem_qcache_reap(zio_arena);
4826 DTRACE_PROBE(arc__kmem_reap_end);
4830 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4831 * enough data and signal them to proceed. When this happens, the threads in
4832 * arc_get_data_impl() are sleeping while holding the hash lock for their
4833 * particular arc header. Thus, we must be careful to never sleep on a
4834 * hash lock in this thread. This is to prevent the following deadlock:
4836 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4837 * waiting for the reclaim thread to signal it.
4839 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4840 * fails, and goes to sleep forever.
4842 * This possible deadlock is avoided by always acquiring a hash lock
4843 * using mutex_tryenter() from arc_reclaim_thread().
4847 arc_reclaim_thread(void *unused __unused)
4849 hrtime_t growtime = 0;
4850 hrtime_t kmem_reap_time = 0;
4853 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4855 mutex_enter(&arc_reclaim_lock);
4856 while (!arc_reclaim_thread_exit) {
4857 uint64_t evicted = 0;
4860 * This is necessary in order for the mdb ::arc dcmd to
4861 * show up to date information. Since the ::arc command
4862 * does not call the kstat's update function, without
4863 * this call, the command may show stale stats for the
4864 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4865 * with this change, the data might be up to 1 second
4866 * out of date; but that should suffice. The arc_state_t
4867 * structures can be queried directly if more accurate
4868 * information is needed.
4870 if (arc_ksp != NULL)
4871 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4873 mutex_exit(&arc_reclaim_lock);
4876 * We call arc_adjust() before (possibly) calling
4877 * arc_kmem_reap_now(), so that we can wake up
4878 * arc_get_data_impl() sooner.
4880 evicted = arc_adjust();
4882 int64_t free_memory = arc_available_memory();
4883 if (free_memory < 0) {
4884 hrtime_t curtime = gethrtime();
4885 arc_no_grow = B_TRUE;
4889 * Wait at least zfs_grow_retry (default 60) seconds
4890 * before considering growing.
4892 growtime = curtime + SEC2NSEC(arc_grow_retry);
4895 * Wait at least arc_kmem_cache_reap_retry_ms
4896 * between arc_kmem_reap_now() calls. Without
4897 * this check it is possible to end up in a
4898 * situation where we spend lots of time
4899 * reaping caches, while we're near arc_c_min.
4901 if (curtime >= kmem_reap_time) {
4902 arc_kmem_reap_now();
4903 kmem_reap_time = gethrtime() +
4904 MSEC2NSEC(arc_kmem_cache_reap_retry_ms);
4908 * If we are still low on memory, shrink the ARC
4909 * so that we have arc_shrink_min free space.
4911 free_memory = arc_available_memory();
4914 (arc_c >> arc_shrink_shift) - free_memory;
4918 to_free = MAX(to_free, ptob(needfree));
4921 evicted += arc_shrink(to_free);
4923 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4924 arc_no_grow = B_TRUE;
4925 } else if (gethrtime() >= growtime) {
4926 arc_no_grow = B_FALSE;
4929 mutex_enter(&arc_reclaim_lock);
4932 * If evicted is zero, we couldn't evict anything via
4933 * arc_adjust(). This could be due to hash lock
4934 * collisions, but more likely due to the majority of
4935 * arc buffers being unevictable. Therefore, even if
4936 * arc_size is above arc_c, another pass is unlikely to
4937 * be helpful and could potentially cause us to enter an
4940 if (aggsum_compare(&arc_size, arc_c) <= 0|| evicted == 0) {
4942 * We're either no longer overflowing, or we
4943 * can't evict anything more, so we should wake
4944 * up any threads before we go to sleep.
4946 cv_broadcast(&arc_reclaim_waiters_cv);
4949 * Block until signaled, or after one second (we
4950 * might need to perform arc_kmem_reap_now()
4951 * even if we aren't being signalled)
4953 CALLB_CPR_SAFE_BEGIN(&cpr);
4954 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4955 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4956 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4960 arc_reclaim_thread_exit = B_FALSE;
4961 cv_broadcast(&arc_reclaim_thread_cv);
4962 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
4966 static u_int arc_dnlc_evicts_arg;
4967 extern struct vfsops zfs_vfsops;
4970 arc_dnlc_evicts_thread(void *dummy __unused)
4975 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG);
4977 mutex_enter(&arc_dnlc_evicts_lock);
4978 while (!arc_dnlc_evicts_thread_exit) {
4979 CALLB_CPR_SAFE_BEGIN(&cpr);
4980 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
4981 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock);
4982 if (arc_dnlc_evicts_arg != 0) {
4983 percent = arc_dnlc_evicts_arg;
4984 mutex_exit(&arc_dnlc_evicts_lock);
4986 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops);
4988 mutex_enter(&arc_dnlc_evicts_lock);
4990 * Clear our token only after vnlru_free()
4991 * pass is done, to avoid false queueing of
4994 arc_dnlc_evicts_arg = 0;
4997 arc_dnlc_evicts_thread_exit = FALSE;
4998 cv_broadcast(&arc_dnlc_evicts_cv);
4999 CALLB_CPR_EXIT(&cpr);
5004 dnlc_reduce_cache(void *arg)
5008 percent = (u_int)(uintptr_t)arg;
5009 mutex_enter(&arc_dnlc_evicts_lock);
5010 if (arc_dnlc_evicts_arg == 0) {
5011 arc_dnlc_evicts_arg = percent;
5012 cv_broadcast(&arc_dnlc_evicts_cv);
5014 mutex_exit(&arc_dnlc_evicts_lock);
5018 * Adapt arc info given the number of bytes we are trying to add and
5019 * the state that we are comming from. This function is only called
5020 * when we are adding new content to the cache.
5023 arc_adapt(int bytes, arc_state_t *state)
5026 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
5027 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
5028 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
5030 if (state == arc_l2c_only)
5035 * Adapt the target size of the MRU list:
5036 * - if we just hit in the MRU ghost list, then increase
5037 * the target size of the MRU list.
5038 * - if we just hit in the MFU ghost list, then increase
5039 * the target size of the MFU list by decreasing the
5040 * target size of the MRU list.
5042 if (state == arc_mru_ghost) {
5043 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
5044 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
5046 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
5047 } else if (state == arc_mfu_ghost) {
5050 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
5051 mult = MIN(mult, 10);
5053 delta = MIN(bytes * mult, arc_p);
5054 arc_p = MAX(arc_p_min, arc_p - delta);
5056 ASSERT((int64_t)arc_p >= 0);
5058 if (arc_reclaim_needed()) {
5059 cv_signal(&arc_reclaim_thread_cv);
5066 if (arc_c >= arc_c_max)
5070 * If we're within (2 * maxblocksize) bytes of the target
5071 * cache size, increment the target cache size
5073 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >
5075 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
5076 atomic_add_64(&arc_c, (int64_t)bytes);
5077 if (arc_c > arc_c_max)
5079 else if (state == arc_anon)
5080 atomic_add_64(&arc_p, (int64_t)bytes);
5084 ASSERT((int64_t)arc_p >= 0);
5088 * Check if arc_size has grown past our upper threshold, determined by
5089 * zfs_arc_overflow_shift.
5092 arc_is_overflowing(void)
5094 /* Always allow at least one block of overflow */
5095 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
5096 arc_c >> zfs_arc_overflow_shift);
5099 * We just compare the lower bound here for performance reasons. Our
5100 * primary goals are to make sure that the arc never grows without
5101 * bound, and that it can reach its maximum size. This check
5102 * accomplishes both goals. The maximum amount we could run over by is
5103 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
5104 * in the ARC. In practice, that's in the tens of MB, which is low
5105 * enough to be safe.
5107 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
5111 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5113 arc_buf_contents_t type = arc_buf_type(hdr);
5115 arc_get_data_impl(hdr, size, tag);
5116 if (type == ARC_BUFC_METADATA) {
5117 return (abd_alloc(size, B_TRUE));
5119 ASSERT(type == ARC_BUFC_DATA);
5120 return (abd_alloc(size, B_FALSE));
5125 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5127 arc_buf_contents_t type = arc_buf_type(hdr);
5129 arc_get_data_impl(hdr, size, tag);
5130 if (type == ARC_BUFC_METADATA) {
5131 return (zio_buf_alloc(size));
5133 ASSERT(type == ARC_BUFC_DATA);
5134 return (zio_data_buf_alloc(size));
5139 * Allocate a block and return it to the caller. If we are hitting the
5140 * hard limit for the cache size, we must sleep, waiting for the eviction
5141 * thread to catch up. If we're past the target size but below the hard
5142 * limit, we'll only signal the reclaim thread and continue on.
5145 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5147 arc_state_t *state = hdr->b_l1hdr.b_state;
5148 arc_buf_contents_t type = arc_buf_type(hdr);
5150 arc_adapt(size, state);
5153 * If arc_size is currently overflowing, and has grown past our
5154 * upper limit, we must be adding data faster than the evict
5155 * thread can evict. Thus, to ensure we don't compound the
5156 * problem by adding more data and forcing arc_size to grow even
5157 * further past it's target size, we halt and wait for the
5158 * eviction thread to catch up.
5160 * It's also possible that the reclaim thread is unable to evict
5161 * enough buffers to get arc_size below the overflow limit (e.g.
5162 * due to buffers being un-evictable, or hash lock collisions).
5163 * In this case, we want to proceed regardless if we're
5164 * overflowing; thus we don't use a while loop here.
5166 if (arc_is_overflowing()) {
5167 mutex_enter(&arc_reclaim_lock);
5170 * Now that we've acquired the lock, we may no longer be
5171 * over the overflow limit, lets check.
5173 * We're ignoring the case of spurious wake ups. If that
5174 * were to happen, it'd let this thread consume an ARC
5175 * buffer before it should have (i.e. before we're under
5176 * the overflow limit and were signalled by the reclaim
5177 * thread). As long as that is a rare occurrence, it
5178 * shouldn't cause any harm.
5180 if (arc_is_overflowing()) {
5181 cv_signal(&arc_reclaim_thread_cv);
5182 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
5185 mutex_exit(&arc_reclaim_lock);
5188 VERIFY3U(hdr->b_type, ==, type);
5189 if (type == ARC_BUFC_METADATA) {
5190 arc_space_consume(size, ARC_SPACE_META);
5192 arc_space_consume(size, ARC_SPACE_DATA);
5196 * Update the state size. Note that ghost states have a
5197 * "ghost size" and so don't need to be updated.
5199 if (!GHOST_STATE(state)) {
5201 (void) refcount_add_many(&state->arcs_size, size, tag);
5204 * If this is reached via arc_read, the link is
5205 * protected by the hash lock. If reached via
5206 * arc_buf_alloc, the header should not be accessed by
5207 * any other thread. And, if reached via arc_read_done,
5208 * the hash lock will protect it if it's found in the
5209 * hash table; otherwise no other thread should be
5210 * trying to [add|remove]_reference it.
5212 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5213 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5214 (void) refcount_add_many(&state->arcs_esize[type],
5219 * If we are growing the cache, and we are adding anonymous
5220 * data, and we have outgrown arc_p, update arc_p
5222 if (aggsum_compare(&arc_size, arc_c) < 0 &&
5223 hdr->b_l1hdr.b_state == arc_anon &&
5224 (refcount_count(&arc_anon->arcs_size) +
5225 refcount_count(&arc_mru->arcs_size) > arc_p))
5226 arc_p = MIN(arc_c, arc_p + size);
5228 ARCSTAT_BUMP(arcstat_allocated);
5232 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
5234 arc_free_data_impl(hdr, size, tag);
5239 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
5241 arc_buf_contents_t type = arc_buf_type(hdr);
5243 arc_free_data_impl(hdr, size, tag);
5244 if (type == ARC_BUFC_METADATA) {
5245 zio_buf_free(buf, size);
5247 ASSERT(type == ARC_BUFC_DATA);
5248 zio_data_buf_free(buf, size);
5253 * Free the arc data buffer.
5256 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5258 arc_state_t *state = hdr->b_l1hdr.b_state;
5259 arc_buf_contents_t type = arc_buf_type(hdr);
5261 /* protected by hash lock, if in the hash table */
5262 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5263 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5264 ASSERT(state != arc_anon && state != arc_l2c_only);
5266 (void) refcount_remove_many(&state->arcs_esize[type],
5269 (void) refcount_remove_many(&state->arcs_size, size, tag);
5271 VERIFY3U(hdr->b_type, ==, type);
5272 if (type == ARC_BUFC_METADATA) {
5273 arc_space_return(size, ARC_SPACE_META);
5275 ASSERT(type == ARC_BUFC_DATA);
5276 arc_space_return(size, ARC_SPACE_DATA);
5281 * This routine is called whenever a buffer is accessed.
5282 * NOTE: the hash lock is dropped in this function.
5285 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
5289 ASSERT(MUTEX_HELD(hash_lock));
5290 ASSERT(HDR_HAS_L1HDR(hdr));
5292 if (hdr->b_l1hdr.b_state == arc_anon) {
5294 * This buffer is not in the cache, and does not
5295 * appear in our "ghost" list. Add the new buffer
5299 ASSERT0(hdr->b_l1hdr.b_arc_access);
5300 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5301 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5302 arc_change_state(arc_mru, hdr, hash_lock);
5304 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5305 now = ddi_get_lbolt();
5308 * If this buffer is here because of a prefetch, then either:
5309 * - clear the flag if this is a "referencing" read
5310 * (any subsequent access will bump this into the MFU state).
5312 * - move the buffer to the head of the list if this is
5313 * another prefetch (to make it less likely to be evicted).
5315 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5316 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5317 /* link protected by hash lock */
5318 ASSERT(multilist_link_active(
5319 &hdr->b_l1hdr.b_arc_node));
5321 arc_hdr_clear_flags(hdr,
5323 ARC_FLAG_PRESCIENT_PREFETCH);
5324 ARCSTAT_BUMP(arcstat_mru_hits);
5326 hdr->b_l1hdr.b_arc_access = now;
5331 * This buffer has been "accessed" only once so far,
5332 * but it is still in the cache. Move it to the MFU
5335 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
5337 * More than 125ms have passed since we
5338 * instantiated this buffer. Move it to the
5339 * most frequently used state.
5341 hdr->b_l1hdr.b_arc_access = now;
5342 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5343 arc_change_state(arc_mfu, hdr, hash_lock);
5345 atomic_inc_32(&hdr->b_l1hdr.b_mru_hits);
5346 ARCSTAT_BUMP(arcstat_mru_hits);
5347 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5348 arc_state_t *new_state;
5350 * This buffer has been "accessed" recently, but
5351 * was evicted from the cache. Move it to the
5355 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5356 new_state = arc_mru;
5357 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
5358 arc_hdr_clear_flags(hdr,
5360 ARC_FLAG_PRESCIENT_PREFETCH);
5362 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5364 new_state = arc_mfu;
5365 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5368 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5369 arc_change_state(new_state, hdr, hash_lock);
5371 atomic_inc_32(&hdr->b_l1hdr.b_mru_ghost_hits);
5372 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5373 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5375 * This buffer has been accessed more than once and is
5376 * still in the cache. Keep it in the MFU state.
5378 * NOTE: an add_reference() that occurred when we did
5379 * the arc_read() will have kicked this off the list.
5380 * If it was a prefetch, we will explicitly move it to
5381 * the head of the list now.
5384 atomic_inc_32(&hdr->b_l1hdr.b_mfu_hits);
5385 ARCSTAT_BUMP(arcstat_mfu_hits);
5386 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5387 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5388 arc_state_t *new_state = arc_mfu;
5390 * This buffer has been accessed more than once but has
5391 * been evicted from the cache. Move it back to the
5395 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5397 * This is a prefetch access...
5398 * move this block back to the MRU state.
5400 new_state = arc_mru;
5403 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5404 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5405 arc_change_state(new_state, hdr, hash_lock);
5407 atomic_inc_32(&hdr->b_l1hdr.b_mfu_ghost_hits);
5408 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5409 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5411 * This buffer is on the 2nd Level ARC.
5414 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5415 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5416 arc_change_state(arc_mfu, hdr, hash_lock);
5418 ASSERT(!"invalid arc state");
5423 * This routine is called by dbuf_hold() to update the arc_access() state
5424 * which otherwise would be skipped for entries in the dbuf cache.
5427 arc_buf_access(arc_buf_t *buf)
5429 mutex_enter(&buf->b_evict_lock);
5430 arc_buf_hdr_t *hdr = buf->b_hdr;
5433 * Avoid taking the hash_lock when possible as an optimization.
5434 * The header must be checked again under the hash_lock in order
5435 * to handle the case where it is concurrently being released.
5437 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5438 mutex_exit(&buf->b_evict_lock);
5439 ARCSTAT_BUMP(arcstat_access_skip);
5443 kmutex_t *hash_lock = HDR_LOCK(hdr);
5444 mutex_enter(hash_lock);
5446 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5447 mutex_exit(hash_lock);
5448 mutex_exit(&buf->b_evict_lock);
5449 ARCSTAT_BUMP(arcstat_access_skip);
5453 mutex_exit(&buf->b_evict_lock);
5455 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5456 hdr->b_l1hdr.b_state == arc_mfu);
5458 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5459 arc_access(hdr, hash_lock);
5460 mutex_exit(hash_lock);
5462 ARCSTAT_BUMP(arcstat_hits);
5463 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5464 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5467 /* a generic arc_read_done_func_t which you can use */
5470 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5471 arc_buf_t *buf, void *arg)
5476 bcopy(buf->b_data, arg, arc_buf_size(buf));
5477 arc_buf_destroy(buf, arg);
5480 /* a generic arc_read_done_func_t */
5483 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5484 arc_buf_t *buf, void *arg)
5486 arc_buf_t **bufp = arg;
5488 ASSERT(zio == NULL || zio->io_error != 0);
5491 ASSERT(zio == NULL || zio->io_error == 0);
5493 ASSERT(buf->b_data != NULL);
5498 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5500 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5501 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5502 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
5504 if (HDR_COMPRESSION_ENABLED(hdr)) {
5505 ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
5506 BP_GET_COMPRESS(bp));
5508 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5509 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5514 arc_read_done(zio_t *zio)
5516 arc_buf_hdr_t *hdr = zio->io_private;
5517 kmutex_t *hash_lock = NULL;
5518 arc_callback_t *callback_list;
5519 arc_callback_t *acb;
5520 boolean_t freeable = B_FALSE;
5521 boolean_t no_zio_error = (zio->io_error == 0);
5524 * The hdr was inserted into hash-table and removed from lists
5525 * prior to starting I/O. We should find this header, since
5526 * it's in the hash table, and it should be legit since it's
5527 * not possible to evict it during the I/O. The only possible
5528 * reason for it not to be found is if we were freed during the
5531 if (HDR_IN_HASH_TABLE(hdr)) {
5532 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5533 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5534 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5535 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5536 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5538 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
5541 ASSERT((found == hdr &&
5542 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5543 (found == hdr && HDR_L2_READING(hdr)));
5544 ASSERT3P(hash_lock, !=, NULL);
5548 /* byteswap if necessary */
5549 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5550 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5551 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5553 hdr->b_l1hdr.b_byteswap =
5554 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5557 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5561 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5562 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5563 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5565 callback_list = hdr->b_l1hdr.b_acb;
5566 ASSERT3P(callback_list, !=, NULL);
5568 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) {
5570 * Only call arc_access on anonymous buffers. This is because
5571 * if we've issued an I/O for an evicted buffer, we've already
5572 * called arc_access (to prevent any simultaneous readers from
5573 * getting confused).
5575 arc_access(hdr, hash_lock);
5579 * If a read request has a callback (i.e. acb_done is not NULL), then we
5580 * make a buf containing the data according to the parameters which were
5581 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5582 * aren't needlessly decompressing the data multiple times.
5584 int callback_cnt = 0;
5585 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5592 int error = arc_buf_alloc_impl(hdr, acb->acb_private,
5593 acb->acb_compressed, zio->io_error == 0,
5597 * Decompression failed. Set io_error
5598 * so that when we call acb_done (below),
5599 * we will indicate that the read failed.
5600 * Note that in the unusual case where one
5601 * callback is compressed and another
5602 * uncompressed, we will mark all of them
5603 * as failed, even though the uncompressed
5604 * one can't actually fail. In this case,
5605 * the hdr will not be anonymous, because
5606 * if there are multiple callbacks, it's
5607 * because multiple threads found the same
5608 * arc buf in the hash table.
5610 zio->io_error = error;
5615 * If there are multiple callbacks, we must have the hash lock,
5616 * because the only way for multiple threads to find this hdr is
5617 * in the hash table. This ensures that if there are multiple
5618 * callbacks, the hdr is not anonymous. If it were anonymous,
5619 * we couldn't use arc_buf_destroy() in the error case below.
5621 ASSERT(callback_cnt < 2 || hash_lock != NULL);
5623 hdr->b_l1hdr.b_acb = NULL;
5624 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5625 if (callback_cnt == 0) {
5626 ASSERT(HDR_PREFETCH(hdr));
5627 ASSERT0(hdr->b_l1hdr.b_bufcnt);
5628 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5631 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5632 callback_list != NULL);
5635 arc_hdr_verify(hdr, zio->io_bp);
5637 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5638 if (hdr->b_l1hdr.b_state != arc_anon)
5639 arc_change_state(arc_anon, hdr, hash_lock);
5640 if (HDR_IN_HASH_TABLE(hdr))
5641 buf_hash_remove(hdr);
5642 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5646 * Broadcast before we drop the hash_lock to avoid the possibility
5647 * that the hdr (and hence the cv) might be freed before we get to
5648 * the cv_broadcast().
5650 cv_broadcast(&hdr->b_l1hdr.b_cv);
5652 if (hash_lock != NULL) {
5653 mutex_exit(hash_lock);
5656 * This block was freed while we waited for the read to
5657 * complete. It has been removed from the hash table and
5658 * moved to the anonymous state (so that it won't show up
5661 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5662 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5665 /* execute each callback and free its structure */
5666 while ((acb = callback_list) != NULL) {
5667 if (acb->acb_done != NULL) {
5668 if (zio->io_error != 0 && acb->acb_buf != NULL) {
5670 * If arc_buf_alloc_impl() fails during
5671 * decompression, the buf will still be
5672 * allocated, and needs to be freed here.
5674 arc_buf_destroy(acb->acb_buf, acb->acb_private);
5675 acb->acb_buf = NULL;
5677 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5678 acb->acb_buf, acb->acb_private);
5681 if (acb->acb_zio_dummy != NULL) {
5682 acb->acb_zio_dummy->io_error = zio->io_error;
5683 zio_nowait(acb->acb_zio_dummy);
5686 callback_list = acb->acb_next;
5687 kmem_free(acb, sizeof (arc_callback_t));
5691 arc_hdr_destroy(hdr);
5695 * "Read" the block at the specified DVA (in bp) via the
5696 * cache. If the block is found in the cache, invoke the provided
5697 * callback immediately and return. Note that the `zio' parameter
5698 * in the callback will be NULL in this case, since no IO was
5699 * required. If the block is not in the cache pass the read request
5700 * on to the spa with a substitute callback function, so that the
5701 * requested block will be added to the cache.
5703 * If a read request arrives for a block that has a read in-progress,
5704 * either wait for the in-progress read to complete (and return the
5705 * results); or, if this is a read with a "done" func, add a record
5706 * to the read to invoke the "done" func when the read completes,
5707 * and return; or just return.
5709 * arc_read_done() will invoke all the requested "done" functions
5710 * for readers of this block.
5713 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_read_done_func_t *done,
5714 void *private, zio_priority_t priority, int zio_flags,
5715 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5717 arc_buf_hdr_t *hdr = NULL;
5718 kmutex_t *hash_lock = NULL;
5720 uint64_t guid = spa_load_guid(spa);
5721 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0;
5724 ASSERT(!BP_IS_EMBEDDED(bp) ||
5725 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5728 if (!BP_IS_EMBEDDED(bp)) {
5730 * Embedded BP's have no DVA and require no I/O to "read".
5731 * Create an anonymous arc buf to back it.
5733 hdr = buf_hash_find(guid, bp, &hash_lock);
5736 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) {
5737 arc_buf_t *buf = NULL;
5738 *arc_flags |= ARC_FLAG_CACHED;
5740 if (HDR_IO_IN_PROGRESS(hdr)) {
5741 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
5743 ASSERT3P(head_zio, !=, NULL);
5744 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5745 priority == ZIO_PRIORITY_SYNC_READ) {
5747 * This is a sync read that needs to wait for
5748 * an in-flight async read. Request that the
5749 * zio have its priority upgraded.
5751 zio_change_priority(head_zio, priority);
5752 DTRACE_PROBE1(arc__async__upgrade__sync,
5753 arc_buf_hdr_t *, hdr);
5754 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
5756 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5757 arc_hdr_clear_flags(hdr,
5758 ARC_FLAG_PREDICTIVE_PREFETCH);
5761 if (*arc_flags & ARC_FLAG_WAIT) {
5762 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5763 mutex_exit(hash_lock);
5766 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5769 arc_callback_t *acb = NULL;
5771 acb = kmem_zalloc(sizeof (arc_callback_t),
5773 acb->acb_done = done;
5774 acb->acb_private = private;
5775 acb->acb_compressed = compressed_read;
5777 acb->acb_zio_dummy = zio_null(pio,
5778 spa, NULL, NULL, NULL, zio_flags);
5780 ASSERT3P(acb->acb_done, !=, NULL);
5781 acb->acb_zio_head = head_zio;
5782 acb->acb_next = hdr->b_l1hdr.b_acb;
5783 hdr->b_l1hdr.b_acb = acb;
5784 mutex_exit(hash_lock);
5787 mutex_exit(hash_lock);
5791 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5792 hdr->b_l1hdr.b_state == arc_mfu);
5795 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5797 * This is a demand read which does not have to
5798 * wait for i/o because we did a predictive
5799 * prefetch i/o for it, which has completed.
5802 arc__demand__hit__predictive__prefetch,
5803 arc_buf_hdr_t *, hdr);
5805 arcstat_demand_hit_predictive_prefetch);
5806 arc_hdr_clear_flags(hdr,
5807 ARC_FLAG_PREDICTIVE_PREFETCH);
5810 if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
5812 arcstat_demand_hit_prescient_prefetch);
5813 arc_hdr_clear_flags(hdr,
5814 ARC_FLAG_PRESCIENT_PREFETCH);
5817 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5818 /* Get a buf with the desired data in it. */
5819 rc = arc_buf_alloc_impl(hdr, private,
5820 compressed_read, B_TRUE, &buf);
5822 arc_buf_destroy(buf, private);
5825 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
5826 rc == 0 || rc != ENOENT);
5827 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
5828 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5829 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5831 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5832 arc_access(hdr, hash_lock);
5833 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5834 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5835 if (*arc_flags & ARC_FLAG_L2CACHE)
5836 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5837 mutex_exit(hash_lock);
5838 ARCSTAT_BUMP(arcstat_hits);
5839 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5840 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5841 data, metadata, hits);
5844 done(NULL, zb, bp, buf, private);
5846 uint64_t lsize = BP_GET_LSIZE(bp);
5847 uint64_t psize = BP_GET_PSIZE(bp);
5848 arc_callback_t *acb;
5851 boolean_t devw = B_FALSE;
5855 /* this block is not in the cache */
5856 arc_buf_hdr_t *exists = NULL;
5857 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5858 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5859 BP_GET_COMPRESS(bp), type);
5861 if (!BP_IS_EMBEDDED(bp)) {
5862 hdr->b_dva = *BP_IDENTITY(bp);
5863 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5864 exists = buf_hash_insert(hdr, &hash_lock);
5866 if (exists != NULL) {
5867 /* somebody beat us to the hash insert */
5868 mutex_exit(hash_lock);
5869 buf_discard_identity(hdr);
5870 arc_hdr_destroy(hdr);
5871 goto top; /* restart the IO request */
5875 * This block is in the ghost cache. If it was L2-only
5876 * (and thus didn't have an L1 hdr), we realloc the
5877 * header to add an L1 hdr.
5879 if (!HDR_HAS_L1HDR(hdr)) {
5880 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5883 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5884 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5885 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5886 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5887 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5888 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5891 * This is a delicate dance that we play here.
5892 * This hdr is in the ghost list so we access it
5893 * to move it out of the ghost list before we
5894 * initiate the read. If it's a prefetch then
5895 * it won't have a callback so we'll remove the
5896 * reference that arc_buf_alloc_impl() created. We
5897 * do this after we've called arc_access() to
5898 * avoid hitting an assert in remove_reference().
5900 arc_access(hdr, hash_lock);
5901 arc_hdr_alloc_pabd(hdr);
5903 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5904 size = arc_hdr_size(hdr);
5907 * If compression is enabled on the hdr, then will do
5908 * RAW I/O and will store the compressed data in the hdr's
5909 * data block. Otherwise, the hdr's data block will contain
5910 * the uncompressed data.
5912 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5913 zio_flags |= ZIO_FLAG_RAW;
5916 if (*arc_flags & ARC_FLAG_PREFETCH)
5917 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5918 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5919 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5921 if (*arc_flags & ARC_FLAG_L2CACHE)
5922 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5923 if (BP_GET_LEVEL(bp) > 0)
5924 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5925 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5926 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5927 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5929 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5930 acb->acb_done = done;
5931 acb->acb_private = private;
5932 acb->acb_compressed = compressed_read;
5934 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5935 hdr->b_l1hdr.b_acb = acb;
5936 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5938 if (HDR_HAS_L2HDR(hdr) &&
5939 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5940 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5941 addr = hdr->b_l2hdr.b_daddr;
5943 * Lock out L2ARC device removal.
5945 if (vdev_is_dead(vd) ||
5946 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5951 * We count both async reads and scrub IOs as asynchronous so
5952 * that both can be upgraded in the event of a cache hit while
5953 * the read IO is still in-flight.
5955 if (priority == ZIO_PRIORITY_ASYNC_READ ||
5956 priority == ZIO_PRIORITY_SCRUB)
5957 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5959 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5962 * At this point, we have a level 1 cache miss. Try again in
5963 * L2ARC if possible.
5965 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5967 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5968 uint64_t, lsize, zbookmark_phys_t *, zb);
5969 ARCSTAT_BUMP(arcstat_misses);
5970 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5971 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5972 data, metadata, misses);
5977 racct_add_force(curproc, RACCT_READBPS, size);
5978 racct_add_force(curproc, RACCT_READIOPS, 1);
5979 PROC_UNLOCK(curproc);
5982 curthread->td_ru.ru_inblock++;
5985 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5987 * Read from the L2ARC if the following are true:
5988 * 1. The L2ARC vdev was previously cached.
5989 * 2. This buffer still has L2ARC metadata.
5990 * 3. This buffer isn't currently writing to the L2ARC.
5991 * 4. The L2ARC entry wasn't evicted, which may
5992 * also have invalidated the vdev.
5993 * 5. This isn't prefetch and l2arc_noprefetch is set.
5995 if (HDR_HAS_L2HDR(hdr) &&
5996 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5997 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5998 l2arc_read_callback_t *cb;
6002 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
6003 ARCSTAT_BUMP(arcstat_l2_hits);
6004 atomic_inc_32(&hdr->b_l2hdr.b_hits);
6006 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
6008 cb->l2rcb_hdr = hdr;
6011 cb->l2rcb_flags = zio_flags;
6013 asize = vdev_psize_to_asize(vd, size);
6014 if (asize != size) {
6015 abd = abd_alloc_for_io(asize,
6016 HDR_ISTYPE_METADATA(hdr));
6017 cb->l2rcb_abd = abd;
6019 abd = hdr->b_l1hdr.b_pabd;
6022 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
6023 addr + asize <= vd->vdev_psize -
6024 VDEV_LABEL_END_SIZE);
6027 * l2arc read. The SCL_L2ARC lock will be
6028 * released by l2arc_read_done().
6029 * Issue a null zio if the underlying buffer
6030 * was squashed to zero size by compression.
6032 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
6033 ZIO_COMPRESS_EMPTY);
6034 rzio = zio_read_phys(pio, vd, addr,
6037 l2arc_read_done, cb, priority,
6038 zio_flags | ZIO_FLAG_DONT_CACHE |
6040 ZIO_FLAG_DONT_PROPAGATE |
6041 ZIO_FLAG_DONT_RETRY, B_FALSE);
6042 acb->acb_zio_head = rzio;
6044 if (hash_lock != NULL)
6045 mutex_exit(hash_lock);
6047 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
6049 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
6051 if (*arc_flags & ARC_FLAG_NOWAIT) {
6056 ASSERT(*arc_flags & ARC_FLAG_WAIT);
6057 if (zio_wait(rzio) == 0)
6060 /* l2arc read error; goto zio_read() */
6061 if (hash_lock != NULL)
6062 mutex_enter(hash_lock);
6064 DTRACE_PROBE1(l2arc__miss,
6065 arc_buf_hdr_t *, hdr);
6066 ARCSTAT_BUMP(arcstat_l2_misses);
6067 if (HDR_L2_WRITING(hdr))
6068 ARCSTAT_BUMP(arcstat_l2_rw_clash);
6069 spa_config_exit(spa, SCL_L2ARC, vd);
6073 spa_config_exit(spa, SCL_L2ARC, vd);
6074 if (l2arc_ndev != 0) {
6075 DTRACE_PROBE1(l2arc__miss,
6076 arc_buf_hdr_t *, hdr);
6077 ARCSTAT_BUMP(arcstat_l2_misses);
6081 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size,
6082 arc_read_done, hdr, priority, zio_flags, zb);
6083 acb->acb_zio_head = rzio;
6085 if (hash_lock != NULL)
6086 mutex_exit(hash_lock);
6088 if (*arc_flags & ARC_FLAG_WAIT)
6089 return (zio_wait(rzio));
6091 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6098 arc_add_prune_callback(arc_prune_func_t *func, void *private)
6102 p = kmem_alloc(sizeof (*p), KM_SLEEP);
6104 p->p_private = private;
6105 list_link_init(&p->p_node);
6106 refcount_create(&p->p_refcnt);
6108 mutex_enter(&arc_prune_mtx);
6109 refcount_add(&p->p_refcnt, &arc_prune_list);
6110 list_insert_head(&arc_prune_list, p);
6111 mutex_exit(&arc_prune_mtx);
6117 arc_remove_prune_callback(arc_prune_t *p)
6119 boolean_t wait = B_FALSE;
6120 mutex_enter(&arc_prune_mtx);
6121 list_remove(&arc_prune_list, p);
6122 if (refcount_remove(&p->p_refcnt, &arc_prune_list) > 0)
6124 mutex_exit(&arc_prune_mtx);
6126 /* wait for arc_prune_task to finish */
6128 taskq_wait(arc_prune_taskq);
6129 ASSERT0(refcount_count(&p->p_refcnt));
6130 refcount_destroy(&p->p_refcnt);
6131 kmem_free(p, sizeof (*p));
6135 * Notify the arc that a block was freed, and thus will never be used again.
6138 arc_freed(spa_t *spa, const blkptr_t *bp)
6141 kmutex_t *hash_lock;
6142 uint64_t guid = spa_load_guid(spa);
6144 ASSERT(!BP_IS_EMBEDDED(bp));
6146 hdr = buf_hash_find(guid, bp, &hash_lock);
6151 * We might be trying to free a block that is still doing I/O
6152 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
6153 * dmu_sync-ed block). If this block is being prefetched, then it
6154 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
6155 * until the I/O completes. A block may also have a reference if it is
6156 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6157 * have written the new block to its final resting place on disk but
6158 * without the dedup flag set. This would have left the hdr in the MRU
6159 * state and discoverable. When the txg finally syncs it detects that
6160 * the block was overridden in open context and issues an override I/O.
6161 * Since this is a dedup block, the override I/O will determine if the
6162 * block is already in the DDT. If so, then it will replace the io_bp
6163 * with the bp from the DDT and allow the I/O to finish. When the I/O
6164 * reaches the done callback, dbuf_write_override_done, it will
6165 * check to see if the io_bp and io_bp_override are identical.
6166 * If they are not, then it indicates that the bp was replaced with
6167 * the bp in the DDT and the override bp is freed. This allows
6168 * us to arrive here with a reference on a block that is being
6169 * freed. So if we have an I/O in progress, or a reference to
6170 * this hdr, then we don't destroy the hdr.
6172 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
6173 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
6174 arc_change_state(arc_anon, hdr, hash_lock);
6175 arc_hdr_destroy(hdr);
6176 mutex_exit(hash_lock);
6178 mutex_exit(hash_lock);
6184 * Release this buffer from the cache, making it an anonymous buffer. This
6185 * must be done after a read and prior to modifying the buffer contents.
6186 * If the buffer has more than one reference, we must make
6187 * a new hdr for the buffer.
6190 arc_release(arc_buf_t *buf, void *tag)
6192 arc_buf_hdr_t *hdr = buf->b_hdr;
6195 * It would be nice to assert that if it's DMU metadata (level >
6196 * 0 || it's the dnode file), then it must be syncing context.
6197 * But we don't know that information at this level.
6200 mutex_enter(&buf->b_evict_lock);
6202 ASSERT(HDR_HAS_L1HDR(hdr));
6205 * We don't grab the hash lock prior to this check, because if
6206 * the buffer's header is in the arc_anon state, it won't be
6207 * linked into the hash table.
6209 if (hdr->b_l1hdr.b_state == arc_anon) {
6210 mutex_exit(&buf->b_evict_lock);
6211 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6212 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6213 ASSERT(!HDR_HAS_L2HDR(hdr));
6214 ASSERT(HDR_EMPTY(hdr));
6215 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6216 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6217 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
6219 hdr->b_l1hdr.b_arc_access = 0;
6222 * If the buf is being overridden then it may already
6223 * have a hdr that is not empty.
6225 buf_discard_identity(hdr);
6231 kmutex_t *hash_lock = HDR_LOCK(hdr);
6232 mutex_enter(hash_lock);
6235 * This assignment is only valid as long as the hash_lock is
6236 * held, we must be careful not to reference state or the
6237 * b_state field after dropping the lock.
6239 arc_state_t *state = hdr->b_l1hdr.b_state;
6240 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6241 ASSERT3P(state, !=, arc_anon);
6243 /* this buffer is not on any list */
6244 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6246 if (HDR_HAS_L2HDR(hdr)) {
6247 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6250 * We have to recheck this conditional again now that
6251 * we're holding the l2ad_mtx to prevent a race with
6252 * another thread which might be concurrently calling
6253 * l2arc_evict(). In that case, l2arc_evict() might have
6254 * destroyed the header's L2 portion as we were waiting
6255 * to acquire the l2ad_mtx.
6257 if (HDR_HAS_L2HDR(hdr)) {
6259 arc_hdr_l2hdr_destroy(hdr);
6262 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6266 * Do we have more than one buf?
6268 if (hdr->b_l1hdr.b_bufcnt > 1) {
6269 arc_buf_hdr_t *nhdr;
6270 uint64_t spa = hdr->b_spa;
6271 uint64_t psize = HDR_GET_PSIZE(hdr);
6272 uint64_t lsize = HDR_GET_LSIZE(hdr);
6273 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
6274 arc_buf_contents_t type = arc_buf_type(hdr);
6275 VERIFY3U(hdr->b_type, ==, type);
6277 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6278 (void) remove_reference(hdr, hash_lock, tag);
6280 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
6281 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6282 ASSERT(ARC_BUF_LAST(buf));
6286 * Pull the data off of this hdr and attach it to
6287 * a new anonymous hdr. Also find the last buffer
6288 * in the hdr's buffer list.
6290 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6291 ASSERT3P(lastbuf, !=, NULL);
6294 * If the current arc_buf_t and the hdr are sharing their data
6295 * buffer, then we must stop sharing that block.
6297 if (arc_buf_is_shared(buf)) {
6298 VERIFY(!arc_buf_is_shared(lastbuf));
6301 * First, sever the block sharing relationship between
6302 * buf and the arc_buf_hdr_t.
6304 arc_unshare_buf(hdr, buf);
6307 * Now we need to recreate the hdr's b_pabd. Since we
6308 * have lastbuf handy, we try to share with it, but if
6309 * we can't then we allocate a new b_pabd and copy the
6310 * data from buf into it.
6312 if (arc_can_share(hdr, lastbuf)) {
6313 arc_share_buf(hdr, lastbuf);
6315 arc_hdr_alloc_pabd(hdr);
6316 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6317 buf->b_data, psize);
6319 VERIFY3P(lastbuf->b_data, !=, NULL);
6320 } else if (HDR_SHARED_DATA(hdr)) {
6322 * Uncompressed shared buffers are always at the end
6323 * of the list. Compressed buffers don't have the
6324 * same requirements. This makes it hard to
6325 * simply assert that the lastbuf is shared so
6326 * we rely on the hdr's compression flags to determine
6327 * if we have a compressed, shared buffer.
6329 ASSERT(arc_buf_is_shared(lastbuf) ||
6330 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
6331 ASSERT(!ARC_BUF_SHARED(buf));
6333 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
6334 ASSERT3P(state, !=, arc_l2c_only);
6336 (void) refcount_remove_many(&state->arcs_size,
6337 arc_buf_size(buf), buf);
6339 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6340 ASSERT3P(state, !=, arc_l2c_only);
6341 (void) refcount_remove_many(&state->arcs_esize[type],
6342 arc_buf_size(buf), buf);
6345 hdr->b_l1hdr.b_bufcnt -= 1;
6346 arc_cksum_verify(buf);
6348 arc_buf_unwatch(buf);
6351 mutex_exit(hash_lock);
6354 * Allocate a new hdr. The new hdr will contain a b_pabd
6355 * buffer which will be freed in arc_write().
6357 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
6358 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6359 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6360 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
6361 VERIFY3U(nhdr->b_type, ==, type);
6362 ASSERT(!HDR_SHARED_DATA(nhdr));
6364 nhdr->b_l1hdr.b_buf = buf;
6365 nhdr->b_l1hdr.b_bufcnt = 1;
6366 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6369 mutex_exit(&buf->b_evict_lock);
6370 (void) refcount_add_many(&arc_anon->arcs_size,
6371 arc_buf_size(buf), buf);
6373 mutex_exit(&buf->b_evict_lock);
6374 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6375 /* protected by hash lock, or hdr is on arc_anon */
6376 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6377 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6378 arc_change_state(arc_anon, hdr, hash_lock);
6379 hdr->b_l1hdr.b_arc_access = 0;
6380 mutex_exit(hash_lock);
6382 buf_discard_identity(hdr);
6388 arc_released(arc_buf_t *buf)
6392 mutex_enter(&buf->b_evict_lock);
6393 released = (buf->b_data != NULL &&
6394 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6395 mutex_exit(&buf->b_evict_lock);
6401 arc_referenced(arc_buf_t *buf)
6405 mutex_enter(&buf->b_evict_lock);
6406 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6407 mutex_exit(&buf->b_evict_lock);
6408 return (referenced);
6413 arc_write_ready(zio_t *zio)
6415 arc_write_callback_t *callback = zio->io_private;
6416 arc_buf_t *buf = callback->awcb_buf;
6417 arc_buf_hdr_t *hdr = buf->b_hdr;
6418 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
6420 ASSERT(HDR_HAS_L1HDR(hdr));
6421 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6422 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6425 * If we're reexecuting this zio because the pool suspended, then
6426 * cleanup any state that was previously set the first time the
6427 * callback was invoked.
6429 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6430 arc_cksum_free(hdr);
6432 arc_buf_unwatch(buf);
6434 if (hdr->b_l1hdr.b_pabd != NULL) {
6435 if (arc_buf_is_shared(buf)) {
6436 arc_unshare_buf(hdr, buf);
6438 arc_hdr_free_pabd(hdr);
6442 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6443 ASSERT(!HDR_SHARED_DATA(hdr));
6444 ASSERT(!arc_buf_is_shared(buf));
6446 callback->awcb_ready(zio, buf, callback->awcb_private);
6448 if (HDR_IO_IN_PROGRESS(hdr))
6449 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6451 arc_cksum_compute(buf);
6452 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6454 enum zio_compress compress;
6455 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6456 compress = ZIO_COMPRESS_OFF;
6458 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
6459 compress = BP_GET_COMPRESS(zio->io_bp);
6461 HDR_SET_PSIZE(hdr, psize);
6462 arc_hdr_set_compress(hdr, compress);
6466 * Fill the hdr with data. If the hdr is compressed, the data we want
6467 * is available from the zio, otherwise we can take it from the buf.
6469 * We might be able to share the buf's data with the hdr here. However,
6470 * doing so would cause the ARC to be full of linear ABDs if we write a
6471 * lot of shareable data. As a compromise, we check whether scattered
6472 * ABDs are allowed, and assume that if they are then the user wants
6473 * the ARC to be primarily filled with them regardless of the data being
6474 * written. Therefore, if they're allowed then we allocate one and copy
6475 * the data into it; otherwise, we share the data directly if we can.
6477 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
6478 arc_hdr_alloc_pabd(hdr);
6481 * Ideally, we would always copy the io_abd into b_pabd, but the
6482 * user may have disabled compressed ARC, thus we must check the
6483 * hdr's compression setting rather than the io_bp's.
6485 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
6486 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=,
6488 ASSERT3U(psize, >, 0);
6490 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6492 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6494 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6498 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6499 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6500 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6502 arc_share_buf(hdr, buf);
6505 arc_hdr_verify(hdr, zio->io_bp);
6509 arc_write_children_ready(zio_t *zio)
6511 arc_write_callback_t *callback = zio->io_private;
6512 arc_buf_t *buf = callback->awcb_buf;
6514 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6518 * The SPA calls this callback for each physical write that happens on behalf
6519 * of a logical write. See the comment in dbuf_write_physdone() for details.
6522 arc_write_physdone(zio_t *zio)
6524 arc_write_callback_t *cb = zio->io_private;
6525 if (cb->awcb_physdone != NULL)
6526 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
6530 arc_write_done(zio_t *zio)
6532 arc_write_callback_t *callback = zio->io_private;
6533 arc_buf_t *buf = callback->awcb_buf;
6534 arc_buf_hdr_t *hdr = buf->b_hdr;
6536 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6538 if (zio->io_error == 0) {
6539 arc_hdr_verify(hdr, zio->io_bp);
6541 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6542 buf_discard_identity(hdr);
6544 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6545 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6548 ASSERT(HDR_EMPTY(hdr));
6552 * If the block to be written was all-zero or compressed enough to be
6553 * embedded in the BP, no write was performed so there will be no
6554 * dva/birth/checksum. The buffer must therefore remain anonymous
6557 if (!HDR_EMPTY(hdr)) {
6558 arc_buf_hdr_t *exists;
6559 kmutex_t *hash_lock;
6561 ASSERT3U(zio->io_error, ==, 0);
6563 arc_cksum_verify(buf);
6565 exists = buf_hash_insert(hdr, &hash_lock);
6566 if (exists != NULL) {
6568 * This can only happen if we overwrite for
6569 * sync-to-convergence, because we remove
6570 * buffers from the hash table when we arc_free().
6572 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6573 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6574 panic("bad overwrite, hdr=%p exists=%p",
6575 (void *)hdr, (void *)exists);
6576 ASSERT(refcount_is_zero(
6577 &exists->b_l1hdr.b_refcnt));
6578 arc_change_state(arc_anon, exists, hash_lock);
6579 mutex_exit(hash_lock);
6580 arc_hdr_destroy(exists);
6581 exists = buf_hash_insert(hdr, &hash_lock);
6582 ASSERT3P(exists, ==, NULL);
6583 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6585 ASSERT(zio->io_prop.zp_nopwrite);
6586 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6587 panic("bad nopwrite, hdr=%p exists=%p",
6588 (void *)hdr, (void *)exists);
6591 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6592 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6593 ASSERT(BP_GET_DEDUP(zio->io_bp));
6594 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6597 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6598 /* if it's not anon, we are doing a scrub */
6599 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6600 arc_access(hdr, hash_lock);
6601 mutex_exit(hash_lock);
6603 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6606 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6607 callback->awcb_done(zio, buf, callback->awcb_private);
6609 abd_put(zio->io_abd);
6610 kmem_free(callback, sizeof (arc_write_callback_t));
6614 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
6615 boolean_t l2arc, const zio_prop_t *zp, arc_write_done_func_t *ready,
6616 arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
6617 arc_write_done_func_t *done, void *private, zio_priority_t priority,
6618 int zio_flags, const zbookmark_phys_t *zb)
6620 arc_buf_hdr_t *hdr = buf->b_hdr;
6621 arc_write_callback_t *callback;
6623 zio_prop_t localprop = *zp;
6625 ASSERT3P(ready, !=, NULL);
6626 ASSERT3P(done, !=, NULL);
6627 ASSERT(!HDR_IO_ERROR(hdr));
6628 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6629 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6630 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6632 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6633 if (ARC_BUF_COMPRESSED(buf)) {
6635 * We're writing a pre-compressed buffer. Make the
6636 * compression algorithm requested by the zio_prop_t match
6637 * the pre-compressed buffer's compression algorithm.
6639 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6641 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6642 zio_flags |= ZIO_FLAG_RAW;
6644 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6645 callback->awcb_ready = ready;
6646 callback->awcb_children_ready = children_ready;
6647 callback->awcb_physdone = physdone;
6648 callback->awcb_done = done;
6649 callback->awcb_private = private;
6650 callback->awcb_buf = buf;
6653 * The hdr's b_pabd is now stale, free it now. A new data block
6654 * will be allocated when the zio pipeline calls arc_write_ready().
6656 if (hdr->b_l1hdr.b_pabd != NULL) {
6658 * If the buf is currently sharing the data block with
6659 * the hdr then we need to break that relationship here.
6660 * The hdr will remain with a NULL data pointer and the
6661 * buf will take sole ownership of the block.
6663 if (arc_buf_is_shared(buf)) {
6664 arc_unshare_buf(hdr, buf);
6666 arc_hdr_free_pabd(hdr);
6668 VERIFY3P(buf->b_data, !=, NULL);
6669 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6671 ASSERT(!arc_buf_is_shared(buf));
6672 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6674 zio = zio_write(pio, spa, txg, bp,
6675 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6676 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6677 (children_ready != NULL) ? arc_write_children_ready : NULL,
6678 arc_write_physdone, arc_write_done, callback,
6679 priority, zio_flags, zb);
6685 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
6688 uint64_t available_memory = ptob(freemem);
6690 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
6691 available_memory = MIN(available_memory, uma_avail());
6694 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
6697 if (txg > spa->spa_lowmem_last_txg) {
6698 spa->spa_lowmem_last_txg = txg;
6699 spa->spa_lowmem_page_load = 0;
6702 * If we are in pageout, we know that memory is already tight,
6703 * the arc is already going to be evicting, so we just want to
6704 * continue to let page writes occur as quickly as possible.
6706 if (curproc == pageproc) {
6707 if (spa->spa_lowmem_page_load >
6708 MAX(ptob(minfree), available_memory) / 4)
6709 return (SET_ERROR(ERESTART));
6710 /* Note: reserve is inflated, so we deflate */
6711 atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8);
6713 } else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) {
6714 /* memory is low, delay before restarting */
6715 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
6716 return (SET_ERROR(EAGAIN));
6718 spa->spa_lowmem_page_load = 0;
6719 #endif /* _KERNEL */
6724 arc_tempreserve_clear(uint64_t reserve)
6726 atomic_add_64(&arc_tempreserve, -reserve);
6727 ASSERT((int64_t)arc_tempreserve >= 0);
6731 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
6736 if (reserve > arc_c/4 && !arc_no_grow) {
6737 arc_c = MIN(arc_c_max, reserve * 4);
6738 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
6740 if (reserve > arc_c)
6741 return (SET_ERROR(ENOMEM));
6744 * Don't count loaned bufs as in flight dirty data to prevent long
6745 * network delays from blocking transactions that are ready to be
6746 * assigned to a txg.
6749 /* assert that it has not wrapped around */
6750 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6752 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
6753 arc_loaned_bytes), 0);
6756 * Writes will, almost always, require additional memory allocations
6757 * in order to compress/encrypt/etc the data. We therefore need to
6758 * make sure that there is sufficient available memory for this.
6760 error = arc_memory_throttle(spa, reserve, txg);
6765 * Throttle writes when the amount of dirty data in the cache
6766 * gets too large. We try to keep the cache less than half full
6767 * of dirty blocks so that our sync times don't grow too large.
6769 * In the case of one pool being built on another pool, we want
6770 * to make sure we don't end up throttling the lower (backing)
6771 * pool when the upper pool is the majority contributor to dirty
6772 * data. To insure we make forward progress during throttling, we
6773 * also check the current pool's net dirty data and only throttle
6774 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
6775 * data in the cache.
6777 * Note: if two requests come in concurrently, we might let them
6778 * both succeed, when one of them should fail. Not a huge deal.
6780 uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
6781 uint64_t spa_dirty_anon = spa_dirty_data(spa);
6783 if (total_dirty > arc_c * zfs_arc_dirty_limit_percent / 100 &&
6784 anon_size > arc_c * zfs_arc_anon_limit_percent / 100 &&
6785 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
6786 uint64_t meta_esize =
6787 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6788 uint64_t data_esize =
6789 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6790 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6791 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6792 arc_tempreserve >> 10, meta_esize >> 10,
6793 data_esize >> 10, reserve >> 10, arc_c >> 10);
6794 return (SET_ERROR(ERESTART));
6796 atomic_add_64(&arc_tempreserve, reserve);
6801 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6802 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6804 size->value.ui64 = refcount_count(&state->arcs_size);
6805 evict_data->value.ui64 =
6806 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6807 evict_metadata->value.ui64 =
6808 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6812 arc_kstat_update(kstat_t *ksp, int rw)
6814 arc_stats_t *as = ksp->ks_data;
6816 if (rw == KSTAT_WRITE) {
6819 arc_kstat_update_state(arc_anon,
6820 &as->arcstat_anon_size,
6821 &as->arcstat_anon_evictable_data,
6822 &as->arcstat_anon_evictable_metadata);
6823 arc_kstat_update_state(arc_mru,
6824 &as->arcstat_mru_size,
6825 &as->arcstat_mru_evictable_data,
6826 &as->arcstat_mru_evictable_metadata);
6827 arc_kstat_update_state(arc_mru_ghost,
6828 &as->arcstat_mru_ghost_size,
6829 &as->arcstat_mru_ghost_evictable_data,
6830 &as->arcstat_mru_ghost_evictable_metadata);
6831 arc_kstat_update_state(arc_mfu,
6832 &as->arcstat_mfu_size,
6833 &as->arcstat_mfu_evictable_data,
6834 &as->arcstat_mfu_evictable_metadata);
6835 arc_kstat_update_state(arc_mfu_ghost,
6836 &as->arcstat_mfu_ghost_size,
6837 &as->arcstat_mfu_ghost_evictable_data,
6838 &as->arcstat_mfu_ghost_evictable_metadata);
6840 ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
6841 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
6842 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
6843 ARCSTAT(arcstat_metadata_size) =
6844 aggsum_value(&astat_metadata_size);
6845 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
6846 ARCSTAT(arcstat_bonus_size) = aggsum_value(&astat_bonus_size);
6847 ARCSTAT(arcstat_dnode_size) = aggsum_value(&astat_dnode_size);
6848 ARCSTAT(arcstat_dbuf_size) = aggsum_value(&astat_dbuf_size);
6849 #if defined(__FreeBSD__) && defined(COMPAT_FREEBSD11)
6850 ARCSTAT(arcstat_other_size) = aggsum_value(&astat_bonus_size) +
6851 aggsum_value(&astat_dnode_size) +
6852 aggsum_value(&astat_dbuf_size);
6854 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
6861 * This function *must* return indices evenly distributed between all
6862 * sublists of the multilist. This is needed due to how the ARC eviction
6863 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6864 * distributed between all sublists and uses this assumption when
6865 * deciding which sublist to evict from and how much to evict from it.
6868 arc_state_multilist_index_func(multilist_t *ml, void *obj)
6870 arc_buf_hdr_t *hdr = obj;
6873 * We rely on b_dva to generate evenly distributed index
6874 * numbers using buf_hash below. So, as an added precaution,
6875 * let's make sure we never add empty buffers to the arc lists.
6877 ASSERT(!HDR_EMPTY(hdr));
6880 * The assumption here, is the hash value for a given
6881 * arc_buf_hdr_t will remain constant throughout it's lifetime
6882 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
6883 * Thus, we don't need to store the header's sublist index
6884 * on insertion, as this index can be recalculated on removal.
6886 * Also, the low order bits of the hash value are thought to be
6887 * distributed evenly. Otherwise, in the case that the multilist
6888 * has a power of two number of sublists, each sublists' usage
6889 * would not be evenly distributed.
6891 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
6892 multilist_get_num_sublists(ml));
6896 static eventhandler_tag arc_event_lowmem = NULL;
6899 arc_lowmem(void *arg __unused, int howto __unused)
6902 mutex_enter(&arc_reclaim_lock);
6903 DTRACE_PROBE1(arc__needfree, int64_t, ((int64_t)freemem - zfs_arc_free_target) * PAGESIZE);
6904 cv_signal(&arc_reclaim_thread_cv);
6907 * It is unsafe to block here in arbitrary threads, because we can come
6908 * here from ARC itself and may hold ARC locks and thus risk a deadlock
6909 * with ARC reclaim thread.
6911 if (curproc == pageproc)
6912 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
6913 mutex_exit(&arc_reclaim_lock);
6918 arc_state_init(void)
6920 arc_anon = &ARC_anon;
6922 arc_mru_ghost = &ARC_mru_ghost;
6924 arc_mfu_ghost = &ARC_mfu_ghost;
6925 arc_l2c_only = &ARC_l2c_only;
6927 arc_mru->arcs_list[ARC_BUFC_METADATA] =
6928 multilist_create(sizeof (arc_buf_hdr_t),
6929 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6930 arc_state_multilist_index_func);
6931 arc_mru->arcs_list[ARC_BUFC_DATA] =
6932 multilist_create(sizeof (arc_buf_hdr_t),
6933 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6934 arc_state_multilist_index_func);
6935 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
6936 multilist_create(sizeof (arc_buf_hdr_t),
6937 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6938 arc_state_multilist_index_func);
6939 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
6940 multilist_create(sizeof (arc_buf_hdr_t),
6941 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6942 arc_state_multilist_index_func);
6943 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
6944 multilist_create(sizeof (arc_buf_hdr_t),
6945 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6946 arc_state_multilist_index_func);
6947 arc_mfu->arcs_list[ARC_BUFC_DATA] =
6948 multilist_create(sizeof (arc_buf_hdr_t),
6949 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6950 arc_state_multilist_index_func);
6951 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
6952 multilist_create(sizeof (arc_buf_hdr_t),
6953 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6954 arc_state_multilist_index_func);
6955 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
6956 multilist_create(sizeof (arc_buf_hdr_t),
6957 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6958 arc_state_multilist_index_func);
6959 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
6960 multilist_create(sizeof (arc_buf_hdr_t),
6961 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6962 arc_state_multilist_index_func);
6963 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
6964 multilist_create(sizeof (arc_buf_hdr_t),
6965 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6966 arc_state_multilist_index_func);
6968 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6969 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6970 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6971 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6972 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6973 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6974 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6975 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6976 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6977 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6978 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6979 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6981 refcount_create(&arc_anon->arcs_size);
6982 refcount_create(&arc_mru->arcs_size);
6983 refcount_create(&arc_mru_ghost->arcs_size);
6984 refcount_create(&arc_mfu->arcs_size);
6985 refcount_create(&arc_mfu_ghost->arcs_size);
6986 refcount_create(&arc_l2c_only->arcs_size);
6988 aggsum_init(&arc_meta_used, 0);
6989 aggsum_init(&arc_size, 0);
6990 aggsum_init(&astat_data_size, 0);
6991 aggsum_init(&astat_metadata_size, 0);
6992 aggsum_init(&astat_hdr_size, 0);
6993 aggsum_init(&astat_bonus_size, 0);
6994 aggsum_init(&astat_dnode_size, 0);
6995 aggsum_init(&astat_dbuf_size, 0);
6996 aggsum_init(&astat_l2_hdr_size, 0);
7000 arc_state_fini(void)
7002 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7003 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7004 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7005 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7006 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7007 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7008 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7009 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7010 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7011 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7012 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7013 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7015 refcount_destroy(&arc_anon->arcs_size);
7016 refcount_destroy(&arc_mru->arcs_size);
7017 refcount_destroy(&arc_mru_ghost->arcs_size);
7018 refcount_destroy(&arc_mfu->arcs_size);
7019 refcount_destroy(&arc_mfu_ghost->arcs_size);
7020 refcount_destroy(&arc_l2c_only->arcs_size);
7022 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
7023 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
7024 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
7025 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
7026 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
7027 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
7028 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
7029 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7041 int i, prefetch_tunable_set = 0;
7044 * allmem is "all memory that we could possibly use".
7048 uint64_t allmem = ptob(physmem - swapfs_minfree);
7050 uint64_t allmem = (physmem * PAGESIZE) / 2;
7053 uint64_t allmem = kmem_size();
7057 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
7058 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
7059 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
7061 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
7062 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL);
7064 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
7065 arc_c_min = MAX(allmem / 32, arc_abs_min);
7066 /* set max to 5/8 of all memory, or all but 1GB, whichever is more */
7067 if (allmem >= 1 << 30)
7068 arc_c_max = allmem - (1 << 30);
7070 arc_c_max = arc_c_min;
7071 arc_c_max = MAX(allmem * 5 / 8, arc_c_max);
7074 * In userland, there's only the memory pressure that we artificially
7075 * create (see arc_available_memory()). Don't let arc_c get too
7076 * small, because it can cause transactions to be larger than
7077 * arc_c, causing arc_tempreserve_space() to fail.
7080 arc_c_min = arc_c_max / 2;
7085 * Allow the tunables to override our calculations if they are
7088 if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) {
7089 arc_c_max = zfs_arc_max;
7090 arc_c_min = MIN(arc_c_min, arc_c_max);
7092 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
7093 arc_c_min = zfs_arc_min;
7097 arc_p = (arc_c >> 1);
7099 /* limit meta-data to 1/4 of the arc capacity */
7100 arc_meta_limit = arc_c_max / 4;
7104 * Metadata is stored in the kernel's heap. Don't let us
7105 * use more than half the heap for the ARC.
7108 arc_meta_limit = MIN(arc_meta_limit, uma_limit() / 2);
7109 arc_dnode_limit = arc_meta_limit / 10;
7111 arc_meta_limit = MIN(arc_meta_limit,
7112 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
7116 /* Allow the tunable to override if it is reasonable */
7117 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
7118 arc_meta_limit = zfs_arc_meta_limit;
7120 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
7121 arc_c_min = arc_meta_limit / 2;
7123 if (zfs_arc_meta_min > 0) {
7124 arc_meta_min = zfs_arc_meta_min;
7126 arc_meta_min = arc_c_min / 2;
7129 /* Valid range: <arc_meta_min> - <arc_c_max> */
7130 if ((zfs_arc_dnode_limit) && (zfs_arc_dnode_limit != arc_dnode_limit) &&
7131 (zfs_arc_dnode_limit >= zfs_arc_meta_min) &&
7132 (zfs_arc_dnode_limit <= arc_c_max))
7133 arc_dnode_limit = zfs_arc_dnode_limit;
7135 if (zfs_arc_grow_retry > 0)
7136 arc_grow_retry = zfs_arc_grow_retry;
7138 if (zfs_arc_shrink_shift > 0)
7139 arc_shrink_shift = zfs_arc_shrink_shift;
7141 if (zfs_arc_no_grow_shift > 0)
7142 arc_no_grow_shift = zfs_arc_no_grow_shift;
7144 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
7146 if (arc_no_grow_shift >= arc_shrink_shift)
7147 arc_no_grow_shift = arc_shrink_shift - 1;
7149 if (zfs_arc_p_min_shift > 0)
7150 arc_p_min_shift = zfs_arc_p_min_shift;
7152 /* if kmem_flags are set, lets try to use less memory */
7153 if (kmem_debugging())
7155 if (arc_c < arc_c_min)
7158 zfs_arc_min = arc_c_min;
7159 zfs_arc_max = arc_c_max;
7164 list_create(&arc_prune_list, sizeof (arc_prune_t),
7165 offsetof(arc_prune_t, p_node));
7166 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
7168 arc_prune_taskq = taskq_create("arc_prune", max_ncpus, minclsyspri,
7169 max_ncpus, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
7171 arc_reclaim_thread_exit = B_FALSE;
7172 arc_dnlc_evicts_thread_exit = FALSE;
7174 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
7175 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
7177 if (arc_ksp != NULL) {
7178 arc_ksp->ks_data = &arc_stats;
7179 arc_ksp->ks_update = arc_kstat_update;
7180 kstat_install(arc_ksp);
7183 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
7184 TS_RUN, minclsyspri);
7187 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
7188 EVENTHANDLER_PRI_FIRST);
7191 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0,
7192 TS_RUN, minclsyspri);
7198 * Calculate maximum amount of dirty data per pool.
7200 * If it has been set by /etc/system, take that.
7201 * Otherwise, use a percentage of physical memory defined by
7202 * zfs_dirty_data_max_percent (default 10%) with a cap at
7203 * zfs_dirty_data_max_max (default 4GB).
7205 if (zfs_dirty_data_max == 0) {
7206 zfs_dirty_data_max = ptob(physmem) *
7207 zfs_dirty_data_max_percent / 100;
7208 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
7209 zfs_dirty_data_max_max);
7213 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
7214 prefetch_tunable_set = 1;
7217 if (prefetch_tunable_set == 0) {
7218 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
7220 printf(" add \"vfs.zfs.prefetch_disable=0\" "
7221 "to /boot/loader.conf.\n");
7222 zfs_prefetch_disable = 1;
7225 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
7226 prefetch_tunable_set == 0) {
7227 printf("ZFS NOTICE: Prefetch is disabled by default if less "
7228 "than 4GB of RAM is present;\n"
7229 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
7230 "to /boot/loader.conf.\n");
7231 zfs_prefetch_disable = 1;
7234 /* Warn about ZFS memory and address space requirements. */
7235 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
7236 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
7237 "expect unstable behavior.\n");
7239 if (allmem < 512 * (1 << 20)) {
7240 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
7241 "expect unstable behavior.\n");
7242 printf(" Consider tuning vm.kmem_size and "
7243 "vm.kmem_size_max\n");
7244 printf(" in /boot/loader.conf.\n");
7255 if (arc_event_lowmem != NULL)
7256 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
7259 mutex_enter(&arc_reclaim_lock);
7260 arc_reclaim_thread_exit = B_TRUE;
7262 * The reclaim thread will set arc_reclaim_thread_exit back to
7263 * B_FALSE when it is finished exiting; we're waiting for that.
7265 while (arc_reclaim_thread_exit) {
7266 cv_signal(&arc_reclaim_thread_cv);
7267 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
7269 mutex_exit(&arc_reclaim_lock);
7271 /* Use B_TRUE to ensure *all* buffers are evicted */
7272 arc_flush(NULL, B_TRUE);
7274 mutex_enter(&arc_dnlc_evicts_lock);
7275 arc_dnlc_evicts_thread_exit = TRUE;
7277 * The user evicts thread will set arc_user_evicts_thread_exit
7278 * to FALSE when it is finished exiting; we're waiting for that.
7280 while (arc_dnlc_evicts_thread_exit) {
7281 cv_signal(&arc_dnlc_evicts_cv);
7282 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
7284 mutex_exit(&arc_dnlc_evicts_lock);
7288 if (arc_ksp != NULL) {
7289 kstat_delete(arc_ksp);
7293 taskq_wait(arc_prune_taskq);
7294 taskq_destroy(arc_prune_taskq);
7296 mutex_enter(&arc_prune_mtx);
7297 while ((p = list_head(&arc_prune_list)) != NULL) {
7298 list_remove(&arc_prune_list, p);
7299 refcount_remove(&p->p_refcnt, &arc_prune_list);
7300 refcount_destroy(&p->p_refcnt);
7301 kmem_free(p, sizeof (*p));
7303 mutex_exit(&arc_prune_mtx);
7305 list_destroy(&arc_prune_list);
7306 mutex_destroy(&arc_prune_mtx);
7307 mutex_destroy(&arc_reclaim_lock);
7308 cv_destroy(&arc_reclaim_thread_cv);
7309 cv_destroy(&arc_reclaim_waiters_cv);
7311 mutex_destroy(&arc_dnlc_evicts_lock);
7312 cv_destroy(&arc_dnlc_evicts_cv);
7317 ASSERT0(arc_loaned_bytes);
7323 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7324 * It uses dedicated storage devices to hold cached data, which are populated
7325 * using large infrequent writes. The main role of this cache is to boost
7326 * the performance of random read workloads. The intended L2ARC devices
7327 * include short-stroked disks, solid state disks, and other media with
7328 * substantially faster read latency than disk.
7330 * +-----------------------+
7332 * +-----------------------+
7335 * l2arc_feed_thread() arc_read()
7339 * +---------------+ |
7341 * +---------------+ |
7346 * +-------+ +-------+
7348 * | cache | | cache |
7349 * +-------+ +-------+
7350 * +=========+ .-----.
7351 * : L2ARC : |-_____-|
7352 * : devices : | Disks |
7353 * +=========+ `-_____-'
7355 * Read requests are satisfied from the following sources, in order:
7358 * 2) vdev cache of L2ARC devices
7360 * 4) vdev cache of disks
7363 * Some L2ARC device types exhibit extremely slow write performance.
7364 * To accommodate for this there are some significant differences between
7365 * the L2ARC and traditional cache design:
7367 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
7368 * the ARC behave as usual, freeing buffers and placing headers on ghost
7369 * lists. The ARC does not send buffers to the L2ARC during eviction as
7370 * this would add inflated write latencies for all ARC memory pressure.
7372 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7373 * It does this by periodically scanning buffers from the eviction-end of
7374 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7375 * not already there. It scans until a headroom of buffers is satisfied,
7376 * which itself is a buffer for ARC eviction. If a compressible buffer is
7377 * found during scanning and selected for writing to an L2ARC device, we
7378 * temporarily boost scanning headroom during the next scan cycle to make
7379 * sure we adapt to compression effects (which might significantly reduce
7380 * the data volume we write to L2ARC). The thread that does this is
7381 * l2arc_feed_thread(), illustrated below; example sizes are included to
7382 * provide a better sense of ratio than this diagram:
7385 * +---------------------+----------+
7386 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
7387 * +---------------------+----------+ | o L2ARC eligible
7388 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
7389 * +---------------------+----------+ |
7390 * 15.9 Gbytes ^ 32 Mbytes |
7392 * l2arc_feed_thread()
7394 * l2arc write hand <--[oooo]--'
7398 * +==============================+
7399 * L2ARC dev |####|#|###|###| |####| ... |
7400 * +==============================+
7403 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
7404 * evicted, then the L2ARC has cached a buffer much sooner than it probably
7405 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
7406 * safe to say that this is an uncommon case, since buffers at the end of
7407 * the ARC lists have moved there due to inactivity.
7409 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
7410 * then the L2ARC simply misses copying some buffers. This serves as a
7411 * pressure valve to prevent heavy read workloads from both stalling the ARC
7412 * with waits and clogging the L2ARC with writes. This also helps prevent
7413 * the potential for the L2ARC to churn if it attempts to cache content too
7414 * quickly, such as during backups of the entire pool.
7416 * 5. After system boot and before the ARC has filled main memory, there are
7417 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
7418 * lists can remain mostly static. Instead of searching from tail of these
7419 * lists as pictured, the l2arc_feed_thread() will search from the list heads
7420 * for eligible buffers, greatly increasing its chance of finding them.
7422 * The L2ARC device write speed is also boosted during this time so that
7423 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
7424 * there are no L2ARC reads, and no fear of degrading read performance
7425 * through increased writes.
7427 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
7428 * the vdev queue can aggregate them into larger and fewer writes. Each
7429 * device is written to in a rotor fashion, sweeping writes through
7430 * available space then repeating.
7432 * 7. The L2ARC does not store dirty content. It never needs to flush
7433 * write buffers back to disk based storage.
7435 * 8. If an ARC buffer is written (and dirtied) which also exists in the
7436 * L2ARC, the now stale L2ARC buffer is immediately dropped.
7438 * The performance of the L2ARC can be tweaked by a number of tunables, which
7439 * may be necessary for different workloads:
7441 * l2arc_write_max max write bytes per interval
7442 * l2arc_write_boost extra write bytes during device warmup
7443 * l2arc_noprefetch skip caching prefetched buffers
7444 * l2arc_headroom number of max device writes to precache
7445 * l2arc_headroom_boost when we find compressed buffers during ARC
7446 * scanning, we multiply headroom by this
7447 * percentage factor for the next scan cycle,
7448 * since more compressed buffers are likely to
7450 * l2arc_feed_secs seconds between L2ARC writing
7452 * Tunables may be removed or added as future performance improvements are
7453 * integrated, and also may become zpool properties.
7455 * There are three key functions that control how the L2ARC warms up:
7457 * l2arc_write_eligible() check if a buffer is eligible to cache
7458 * l2arc_write_size() calculate how much to write
7459 * l2arc_write_interval() calculate sleep delay between writes
7461 * These three functions determine what to write, how much, and how quickly
7466 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
7469 * A buffer is *not* eligible for the L2ARC if it:
7470 * 1. belongs to a different spa.
7471 * 2. is already cached on the L2ARC.
7472 * 3. has an I/O in progress (it may be an incomplete read).
7473 * 4. is flagged not eligible (zfs property).
7475 if (hdr->b_spa != spa_guid) {
7476 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
7479 if (HDR_HAS_L2HDR(hdr)) {
7480 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
7483 if (HDR_IO_IN_PROGRESS(hdr)) {
7484 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
7487 if (!HDR_L2CACHE(hdr)) {
7488 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
7496 l2arc_write_size(void)
7501 * Make sure our globals have meaningful values in case the user
7504 size = l2arc_write_max;
7506 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
7507 "be greater than zero, resetting it to the default (%d)",
7509 size = l2arc_write_max = L2ARC_WRITE_SIZE;
7512 if (arc_warm == B_FALSE)
7513 size += l2arc_write_boost;
7520 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
7522 clock_t interval, next, now;
7525 * If the ARC lists are busy, increase our write rate; if the
7526 * lists are stale, idle back. This is achieved by checking
7527 * how much we previously wrote - if it was more than half of
7528 * what we wanted, schedule the next write much sooner.
7530 if (l2arc_feed_again && wrote > (wanted / 2))
7531 interval = (hz * l2arc_feed_min_ms) / 1000;
7533 interval = hz * l2arc_feed_secs;
7535 now = ddi_get_lbolt();
7536 next = MAX(now, MIN(now + interval, began + interval));
7542 * Cycle through L2ARC devices. This is how L2ARC load balances.
7543 * If a device is returned, this also returns holding the spa config lock.
7545 static l2arc_dev_t *
7546 l2arc_dev_get_next(void)
7548 l2arc_dev_t *first, *next = NULL;
7551 * Lock out the removal of spas (spa_namespace_lock), then removal
7552 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
7553 * both locks will be dropped and a spa config lock held instead.
7555 mutex_enter(&spa_namespace_lock);
7556 mutex_enter(&l2arc_dev_mtx);
7558 /* if there are no vdevs, there is nothing to do */
7559 if (l2arc_ndev == 0)
7563 next = l2arc_dev_last;
7565 /* loop around the list looking for a non-faulted vdev */
7567 next = list_head(l2arc_dev_list);
7569 next = list_next(l2arc_dev_list, next);
7571 next = list_head(l2arc_dev_list);
7574 /* if we have come back to the start, bail out */
7577 else if (next == first)
7580 } while (vdev_is_dead(next->l2ad_vdev));
7582 /* if we were unable to find any usable vdevs, return NULL */
7583 if (vdev_is_dead(next->l2ad_vdev))
7586 l2arc_dev_last = next;
7589 mutex_exit(&l2arc_dev_mtx);
7592 * Grab the config lock to prevent the 'next' device from being
7593 * removed while we are writing to it.
7596 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
7597 mutex_exit(&spa_namespace_lock);
7603 * Free buffers that were tagged for destruction.
7606 l2arc_do_free_on_write()
7609 l2arc_data_free_t *df, *df_prev;
7611 mutex_enter(&l2arc_free_on_write_mtx);
7612 buflist = l2arc_free_on_write;
7614 for (df = list_tail(buflist); df; df = df_prev) {
7615 df_prev = list_prev(buflist, df);
7616 ASSERT3P(df->l2df_abd, !=, NULL);
7617 abd_free(df->l2df_abd);
7618 list_remove(buflist, df);
7619 kmem_free(df, sizeof (l2arc_data_free_t));
7622 mutex_exit(&l2arc_free_on_write_mtx);
7626 * A write to a cache device has completed. Update all headers to allow
7627 * reads from these buffers to begin.
7630 l2arc_write_done(zio_t *zio)
7632 l2arc_write_callback_t *cb;
7635 arc_buf_hdr_t *head, *hdr, *hdr_prev;
7636 kmutex_t *hash_lock;
7637 int64_t bytes_dropped = 0;
7639 cb = zio->io_private;
7640 ASSERT3P(cb, !=, NULL);
7641 dev = cb->l2wcb_dev;
7642 ASSERT3P(dev, !=, NULL);
7643 head = cb->l2wcb_head;
7644 ASSERT3P(head, !=, NULL);
7645 buflist = &dev->l2ad_buflist;
7646 ASSERT3P(buflist, !=, NULL);
7647 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
7648 l2arc_write_callback_t *, cb);
7650 if (zio->io_error != 0)
7651 ARCSTAT_BUMP(arcstat_l2_writes_error);
7654 * All writes completed, or an error was hit.
7657 mutex_enter(&dev->l2ad_mtx);
7658 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
7659 hdr_prev = list_prev(buflist, hdr);
7661 hash_lock = HDR_LOCK(hdr);
7664 * We cannot use mutex_enter or else we can deadlock
7665 * with l2arc_write_buffers (due to swapping the order
7666 * the hash lock and l2ad_mtx are taken).
7668 if (!mutex_tryenter(hash_lock)) {
7670 * Missed the hash lock. We must retry so we
7671 * don't leave the ARC_FLAG_L2_WRITING bit set.
7673 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
7676 * We don't want to rescan the headers we've
7677 * already marked as having been written out, so
7678 * we reinsert the head node so we can pick up
7679 * where we left off.
7681 list_remove(buflist, head);
7682 list_insert_after(buflist, hdr, head);
7684 mutex_exit(&dev->l2ad_mtx);
7687 * We wait for the hash lock to become available
7688 * to try and prevent busy waiting, and increase
7689 * the chance we'll be able to acquire the lock
7690 * the next time around.
7692 mutex_enter(hash_lock);
7693 mutex_exit(hash_lock);
7698 * We could not have been moved into the arc_l2c_only
7699 * state while in-flight due to our ARC_FLAG_L2_WRITING
7700 * bit being set. Let's just ensure that's being enforced.
7702 ASSERT(HDR_HAS_L1HDR(hdr));
7704 if (zio->io_error != 0) {
7706 * Error - drop L2ARC entry.
7708 list_remove(buflist, hdr);
7710 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
7712 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr));
7713 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
7715 bytes_dropped += arc_hdr_size(hdr);
7716 (void) refcount_remove_many(&dev->l2ad_alloc,
7717 arc_hdr_size(hdr), hdr);
7721 * Allow ARC to begin reads and ghost list evictions to
7724 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
7726 mutex_exit(hash_lock);
7729 atomic_inc_64(&l2arc_writes_done);
7730 list_remove(buflist, head);
7731 ASSERT(!HDR_HAS_L1HDR(head));
7732 kmem_cache_free(hdr_l2only_cache, head);
7733 mutex_exit(&dev->l2ad_mtx);
7735 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
7737 l2arc_do_free_on_write();
7739 kmem_free(cb, sizeof (l2arc_write_callback_t));
7743 * A read to a cache device completed. Validate buffer contents before
7744 * handing over to the regular ARC routines.
7747 l2arc_read_done(zio_t *zio)
7749 l2arc_read_callback_t *cb;
7751 kmutex_t *hash_lock;
7752 boolean_t valid_cksum;
7754 ASSERT3P(zio->io_vd, !=, NULL);
7755 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
7757 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
7759 cb = zio->io_private;
7760 ASSERT3P(cb, !=, NULL);
7761 hdr = cb->l2rcb_hdr;
7762 ASSERT3P(hdr, !=, NULL);
7764 hash_lock = HDR_LOCK(hdr);
7765 mutex_enter(hash_lock);
7766 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
7769 * If the data was read into a temporary buffer,
7770 * move it and free the buffer.
7772 if (cb->l2rcb_abd != NULL) {
7773 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
7774 if (zio->io_error == 0) {
7775 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd,
7780 * The following must be done regardless of whether
7781 * there was an error:
7782 * - free the temporary buffer
7783 * - point zio to the real ARC buffer
7784 * - set zio size accordingly
7785 * These are required because zio is either re-used for
7786 * an I/O of the block in the case of the error
7787 * or the zio is passed to arc_read_done() and it
7790 abd_free(cb->l2rcb_abd);
7791 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
7792 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
7795 ASSERT3P(zio->io_abd, !=, NULL);
7798 * Check this survived the L2ARC journey.
7800 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd);
7801 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
7802 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
7804 valid_cksum = arc_cksum_is_equal(hdr, zio);
7805 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
7806 mutex_exit(hash_lock);
7807 zio->io_private = hdr;
7810 mutex_exit(hash_lock);
7812 * Buffer didn't survive caching. Increment stats and
7813 * reissue to the original storage device.
7815 if (zio->io_error != 0) {
7816 ARCSTAT_BUMP(arcstat_l2_io_error);
7818 zio->io_error = SET_ERROR(EIO);
7821 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
7824 * If there's no waiter, issue an async i/o to the primary
7825 * storage now. If there *is* a waiter, the caller must
7826 * issue the i/o in a context where it's OK to block.
7828 if (zio->io_waiter == NULL) {
7829 zio_t *pio = zio_unique_parent(zio);
7831 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
7833 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
7834 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done,
7835 hdr, zio->io_priority, cb->l2rcb_flags,
7840 kmem_free(cb, sizeof (l2arc_read_callback_t));
7844 * This is the list priority from which the L2ARC will search for pages to
7845 * cache. This is used within loops (0..3) to cycle through lists in the
7846 * desired order. This order can have a significant effect on cache
7849 * Currently the metadata lists are hit first, MFU then MRU, followed by
7850 * the data lists. This function returns a locked list, and also returns
7853 static multilist_sublist_t *
7854 l2arc_sublist_lock(int list_num)
7856 multilist_t *ml = NULL;
7859 ASSERT(list_num >= 0 && list_num <= 3);
7863 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
7866 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
7869 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
7872 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
7877 * Return a randomly-selected sublist. This is acceptable
7878 * because the caller feeds only a little bit of data for each
7879 * call (8MB). Subsequent calls will result in different
7880 * sublists being selected.
7882 idx = multilist_get_random_index(ml);
7883 return (multilist_sublist_lock(ml, idx));
7887 * Evict buffers from the device write hand to the distance specified in
7888 * bytes. This distance may span populated buffers, it may span nothing.
7889 * This is clearing a region on the L2ARC device ready for writing.
7890 * If the 'all' boolean is set, every buffer is evicted.
7893 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
7896 arc_buf_hdr_t *hdr, *hdr_prev;
7897 kmutex_t *hash_lock;
7900 buflist = &dev->l2ad_buflist;
7902 if (!all && dev->l2ad_first) {
7904 * This is the first sweep through the device. There is
7910 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
7912 * When nearing the end of the device, evict to the end
7913 * before the device write hand jumps to the start.
7915 taddr = dev->l2ad_end;
7917 taddr = dev->l2ad_hand + distance;
7919 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
7920 uint64_t, taddr, boolean_t, all);
7923 mutex_enter(&dev->l2ad_mtx);
7924 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
7925 hdr_prev = list_prev(buflist, hdr);
7927 hash_lock = HDR_LOCK(hdr);
7930 * We cannot use mutex_enter or else we can deadlock
7931 * with l2arc_write_buffers (due to swapping the order
7932 * the hash lock and l2ad_mtx are taken).
7934 if (!mutex_tryenter(hash_lock)) {
7936 * Missed the hash lock. Retry.
7938 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
7939 mutex_exit(&dev->l2ad_mtx);
7940 mutex_enter(hash_lock);
7941 mutex_exit(hash_lock);
7946 * A header can't be on this list if it doesn't have L2 header.
7948 ASSERT(HDR_HAS_L2HDR(hdr));
7950 /* Ensure this header has finished being written. */
7951 ASSERT(!HDR_L2_WRITING(hdr));
7952 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
7954 if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
7955 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
7957 * We've evicted to the target address,
7958 * or the end of the device.
7960 mutex_exit(hash_lock);
7964 if (!HDR_HAS_L1HDR(hdr)) {
7965 ASSERT(!HDR_L2_READING(hdr));
7967 * This doesn't exist in the ARC. Destroy.
7968 * arc_hdr_destroy() will call list_remove()
7969 * and decrement arcstat_l2_lsize.
7971 arc_change_state(arc_anon, hdr, hash_lock);
7972 arc_hdr_destroy(hdr);
7974 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
7975 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
7977 * Invalidate issued or about to be issued
7978 * reads, since we may be about to write
7979 * over this location.
7981 if (HDR_L2_READING(hdr)) {
7982 ARCSTAT_BUMP(arcstat_l2_evict_reading);
7983 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
7986 arc_hdr_l2hdr_destroy(hdr);
7988 mutex_exit(hash_lock);
7990 mutex_exit(&dev->l2ad_mtx);
7994 * Find and write ARC buffers to the L2ARC device.
7996 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7997 * for reading until they have completed writing.
7998 * The headroom_boost is an in-out parameter used to maintain headroom boost
7999 * state between calls to this function.
8001 * Returns the number of bytes actually written (which may be smaller than
8002 * the delta by which the device hand has changed due to alignment).
8005 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
8007 arc_buf_hdr_t *hdr, *hdr_prev, *head;
8008 uint64_t write_asize, write_psize, write_lsize, headroom;
8010 l2arc_write_callback_t *cb;
8012 uint64_t guid = spa_load_guid(spa);
8015 ASSERT3P(dev->l2ad_vdev, !=, NULL);
8018 write_lsize = write_asize = write_psize = 0;
8020 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
8021 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
8023 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
8025 * Copy buffers for L2ARC writing.
8027 for (try = 0; try <= 3; try++) {
8028 multilist_sublist_t *mls = l2arc_sublist_lock(try);
8029 uint64_t passed_sz = 0;
8031 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
8034 * L2ARC fast warmup.
8036 * Until the ARC is warm and starts to evict, read from the
8037 * head of the ARC lists rather than the tail.
8039 if (arc_warm == B_FALSE)
8040 hdr = multilist_sublist_head(mls);
8042 hdr = multilist_sublist_tail(mls);
8044 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
8046 headroom = target_sz * l2arc_headroom;
8047 if (zfs_compressed_arc_enabled)
8048 headroom = (headroom * l2arc_headroom_boost) / 100;
8050 for (; hdr; hdr = hdr_prev) {
8051 kmutex_t *hash_lock;
8053 if (arc_warm == B_FALSE)
8054 hdr_prev = multilist_sublist_next(mls, hdr);
8056 hdr_prev = multilist_sublist_prev(mls, hdr);
8057 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned,
8058 HDR_GET_LSIZE(hdr));
8060 hash_lock = HDR_LOCK(hdr);
8061 if (!mutex_tryenter(hash_lock)) {
8062 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
8064 * Skip this buffer rather than waiting.
8069 passed_sz += HDR_GET_LSIZE(hdr);
8070 if (passed_sz > headroom) {
8074 mutex_exit(hash_lock);
8075 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
8079 if (!l2arc_write_eligible(guid, hdr)) {
8080 mutex_exit(hash_lock);
8085 * We rely on the L1 portion of the header below, so
8086 * it's invalid for this header to have been evicted out
8087 * of the ghost cache, prior to being written out. The
8088 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8090 ASSERT(HDR_HAS_L1HDR(hdr));
8092 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8093 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8094 ASSERT3U(arc_hdr_size(hdr), >, 0);
8095 uint64_t psize = arc_hdr_size(hdr);
8096 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
8099 if ((write_asize + asize) > target_sz) {
8101 mutex_exit(hash_lock);
8102 ARCSTAT_BUMP(arcstat_l2_write_full);
8108 * Insert a dummy header on the buflist so
8109 * l2arc_write_done() can find where the
8110 * write buffers begin without searching.
8112 mutex_enter(&dev->l2ad_mtx);
8113 list_insert_head(&dev->l2ad_buflist, head);
8114 mutex_exit(&dev->l2ad_mtx);
8117 sizeof (l2arc_write_callback_t), KM_SLEEP);
8118 cb->l2wcb_dev = dev;
8119 cb->l2wcb_head = head;
8120 pio = zio_root(spa, l2arc_write_done, cb,
8122 ARCSTAT_BUMP(arcstat_l2_write_pios);
8125 hdr->b_l2hdr.b_dev = dev;
8126 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
8127 arc_hdr_set_flags(hdr,
8128 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
8130 mutex_enter(&dev->l2ad_mtx);
8131 list_insert_head(&dev->l2ad_buflist, hdr);
8132 mutex_exit(&dev->l2ad_mtx);
8134 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr);
8137 * Normally the L2ARC can use the hdr's data, but if
8138 * we're sharing data between the hdr and one of its
8139 * bufs, L2ARC needs its own copy of the data so that
8140 * the ZIO below can't race with the buf consumer.
8141 * Another case where we need to create a copy of the
8142 * data is when the buffer size is not device-aligned
8143 * and we need to pad the block to make it such.
8144 * That also keeps the clock hand suitably aligned.
8146 * To ensure that the copy will be available for the
8147 * lifetime of the ZIO and be cleaned up afterwards, we
8148 * add it to the l2arc_free_on_write queue.
8151 if (!HDR_SHARED_DATA(hdr) && psize == asize) {
8152 to_write = hdr->b_l1hdr.b_pabd;
8154 to_write = abd_alloc_for_io(asize,
8155 HDR_ISTYPE_METADATA(hdr));
8156 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
8157 if (asize != psize) {
8158 abd_zero_off(to_write, psize,
8161 l2arc_free_abd_on_write(to_write, asize,
8164 wzio = zio_write_phys(pio, dev->l2ad_vdev,
8165 hdr->b_l2hdr.b_daddr, asize, to_write,
8166 ZIO_CHECKSUM_OFF, NULL, hdr,
8167 ZIO_PRIORITY_ASYNC_WRITE,
8168 ZIO_FLAG_CANFAIL, B_FALSE);
8170 write_lsize += HDR_GET_LSIZE(hdr);
8171 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
8174 write_psize += psize;
8175 write_asize += asize;
8176 dev->l2ad_hand += asize;
8178 mutex_exit(hash_lock);
8180 (void) zio_nowait(wzio);
8183 multilist_sublist_unlock(mls);
8189 /* No buffers selected for writing? */
8191 ASSERT0(write_lsize);
8192 ASSERT(!HDR_HAS_L1HDR(head));
8193 kmem_cache_free(hdr_l2only_cache, head);
8197 ASSERT3U(write_psize, <=, target_sz);
8198 ARCSTAT_BUMP(arcstat_l2_writes_sent);
8199 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
8200 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
8201 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
8202 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
8205 * Bump device hand to the device start if it is approaching the end.
8206 * l2arc_evict() will already have evicted ahead for this case.
8208 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
8209 dev->l2ad_hand = dev->l2ad_start;
8210 dev->l2ad_first = B_FALSE;
8213 dev->l2ad_writing = B_TRUE;
8214 (void) zio_wait(pio);
8215 dev->l2ad_writing = B_FALSE;
8217 return (write_asize);
8221 * This thread feeds the L2ARC at regular intervals. This is the beating
8222 * heart of the L2ARC.
8226 l2arc_feed_thread(void *unused __unused)
8231 uint64_t size, wrote;
8232 clock_t begin, next = ddi_get_lbolt();
8234 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
8236 mutex_enter(&l2arc_feed_thr_lock);
8238 while (l2arc_thread_exit == 0) {
8239 CALLB_CPR_SAFE_BEGIN(&cpr);
8240 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
8241 next - ddi_get_lbolt());
8242 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
8243 next = ddi_get_lbolt() + hz;
8246 * Quick check for L2ARC devices.
8248 mutex_enter(&l2arc_dev_mtx);
8249 if (l2arc_ndev == 0) {
8250 mutex_exit(&l2arc_dev_mtx);
8253 mutex_exit(&l2arc_dev_mtx);
8254 begin = ddi_get_lbolt();
8257 * This selects the next l2arc device to write to, and in
8258 * doing so the next spa to feed from: dev->l2ad_spa. This
8259 * will return NULL if there are now no l2arc devices or if
8260 * they are all faulted.
8262 * If a device is returned, its spa's config lock is also
8263 * held to prevent device removal. l2arc_dev_get_next()
8264 * will grab and release l2arc_dev_mtx.
8266 if ((dev = l2arc_dev_get_next()) == NULL)
8269 spa = dev->l2ad_spa;
8270 ASSERT3P(spa, !=, NULL);
8273 * If the pool is read-only then force the feed thread to
8274 * sleep a little longer.
8276 if (!spa_writeable(spa)) {
8277 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
8278 spa_config_exit(spa, SCL_L2ARC, dev);
8283 * Avoid contributing to memory pressure.
8285 if (arc_reclaim_needed()) {
8286 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
8287 spa_config_exit(spa, SCL_L2ARC, dev);
8291 ARCSTAT_BUMP(arcstat_l2_feeds);
8293 size = l2arc_write_size();
8296 * Evict L2ARC buffers that will be overwritten.
8298 l2arc_evict(dev, size, B_FALSE);
8301 * Write ARC buffers.
8303 wrote = l2arc_write_buffers(spa, dev, size);
8306 * Calculate interval between writes.
8308 next = l2arc_write_interval(begin, size, wrote);
8309 spa_config_exit(spa, SCL_L2ARC, dev);
8312 l2arc_thread_exit = 0;
8313 cv_broadcast(&l2arc_feed_thr_cv);
8314 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
8319 l2arc_vdev_present(vdev_t *vd)
8323 mutex_enter(&l2arc_dev_mtx);
8324 for (dev = list_head(l2arc_dev_list); dev != NULL;
8325 dev = list_next(l2arc_dev_list, dev)) {
8326 if (dev->l2ad_vdev == vd)
8329 mutex_exit(&l2arc_dev_mtx);
8331 return (dev != NULL);
8335 * Add a vdev for use by the L2ARC. By this point the spa has already
8336 * validated the vdev and opened it.
8339 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
8341 l2arc_dev_t *adddev;
8343 ASSERT(!l2arc_vdev_present(vd));
8345 vdev_ashift_optimize(vd);
8348 * Create a new l2arc device entry.
8350 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
8351 adddev->l2ad_spa = spa;
8352 adddev->l2ad_vdev = vd;
8353 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
8354 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
8355 adddev->l2ad_hand = adddev->l2ad_start;
8356 adddev->l2ad_first = B_TRUE;
8357 adddev->l2ad_writing = B_FALSE;
8359 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
8361 * This is a list of all ARC buffers that are still valid on the
8364 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
8365 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
8367 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
8368 refcount_create(&adddev->l2ad_alloc);
8371 * Add device to global list
8373 mutex_enter(&l2arc_dev_mtx);
8374 list_insert_head(l2arc_dev_list, adddev);
8375 atomic_inc_64(&l2arc_ndev);
8376 mutex_exit(&l2arc_dev_mtx);
8380 * Remove a vdev from the L2ARC.
8383 l2arc_remove_vdev(vdev_t *vd)
8385 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
8388 * Find the device by vdev
8390 mutex_enter(&l2arc_dev_mtx);
8391 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
8392 nextdev = list_next(l2arc_dev_list, dev);
8393 if (vd == dev->l2ad_vdev) {
8398 ASSERT3P(remdev, !=, NULL);
8401 * Remove device from global list
8403 list_remove(l2arc_dev_list, remdev);
8404 l2arc_dev_last = NULL; /* may have been invalidated */
8405 atomic_dec_64(&l2arc_ndev);
8406 mutex_exit(&l2arc_dev_mtx);
8409 * Clear all buflists and ARC references. L2ARC device flush.
8411 l2arc_evict(remdev, 0, B_TRUE);
8412 list_destroy(&remdev->l2ad_buflist);
8413 mutex_destroy(&remdev->l2ad_mtx);
8414 refcount_destroy(&remdev->l2ad_alloc);
8415 kmem_free(remdev, sizeof (l2arc_dev_t));
8421 l2arc_thread_exit = 0;
8423 l2arc_writes_sent = 0;
8424 l2arc_writes_done = 0;
8426 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
8427 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
8428 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
8429 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
8431 l2arc_dev_list = &L2ARC_dev_list;
8432 l2arc_free_on_write = &L2ARC_free_on_write;
8433 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
8434 offsetof(l2arc_dev_t, l2ad_node));
8435 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
8436 offsetof(l2arc_data_free_t, l2df_list_node));
8443 * This is called from dmu_fini(), which is called from spa_fini();
8444 * Because of this, we can assume that all l2arc devices have
8445 * already been removed when the pools themselves were removed.
8448 l2arc_do_free_on_write();
8450 mutex_destroy(&l2arc_feed_thr_lock);
8451 cv_destroy(&l2arc_feed_thr_cv);
8452 mutex_destroy(&l2arc_dev_mtx);
8453 mutex_destroy(&l2arc_free_on_write_mtx);
8455 list_destroy(l2arc_dev_list);
8456 list_destroy(l2arc_free_on_write);
8462 if (!(spa_mode_global & FWRITE))
8465 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
8466 TS_RUN, minclsyspri);
8472 if (!(spa_mode_global & FWRITE))
8475 mutex_enter(&l2arc_feed_thr_lock);
8476 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
8477 l2arc_thread_exit = 1;
8478 while (l2arc_thread_exit != 0)
8479 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
8480 mutex_exit(&l2arc_feed_thr_lock);