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 = 1000;
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, "Enable compressed ARC");
441 * We don't have a tunable for arc_free_target due to the dependency on
442 * pagedaemon initialisation.
444 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
445 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
446 sysctl_vfs_zfs_arc_free_target, "IU",
447 "Desired number of free pages below which ARC triggers reclaim");
450 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
455 val = zfs_arc_free_target;
456 err = sysctl_handle_int(oidp, &val, 0, req);
457 if (err != 0 || req->newptr == NULL)
462 if (val > vm_cnt.v_page_count)
465 zfs_arc_free_target = val;
471 * Must be declared here, before the definition of corresponding kstat
472 * macro which uses the same names will confuse the compiler.
474 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
475 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
476 sysctl_vfs_zfs_arc_meta_limit, "QU",
477 "ARC metadata limit");
481 * Note that buffers can be in one of 6 states:
482 * ARC_anon - anonymous (discussed below)
483 * ARC_mru - recently used, currently cached
484 * ARC_mru_ghost - recentely used, no longer in cache
485 * ARC_mfu - frequently used, currently cached
486 * ARC_mfu_ghost - frequently used, no longer in cache
487 * ARC_l2c_only - exists in L2ARC but not other states
488 * When there are no active references to the buffer, they are
489 * are linked onto a list in one of these arc states. These are
490 * the only buffers that can be evicted or deleted. Within each
491 * state there are multiple lists, one for meta-data and one for
492 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
493 * etc.) is tracked separately so that it can be managed more
494 * explicitly: favored over data, limited explicitly.
496 * Anonymous buffers are buffers that are not associated with
497 * a DVA. These are buffers that hold dirty block copies
498 * before they are written to stable storage. By definition,
499 * they are "ref'd" and are considered part of arc_mru
500 * that cannot be freed. Generally, they will aquire a DVA
501 * as they are written and migrate onto the arc_mru list.
503 * The ARC_l2c_only state is for buffers that are in the second
504 * level ARC but no longer in any of the ARC_m* lists. The second
505 * level ARC itself may also contain buffers that are in any of
506 * the ARC_m* states - meaning that a buffer can exist in two
507 * places. The reason for the ARC_l2c_only state is to keep the
508 * buffer header in the hash table, so that reads that hit the
509 * second level ARC benefit from these fast lookups.
512 typedef struct arc_state {
514 * list of evictable buffers
516 multilist_t *arcs_list[ARC_BUFC_NUMTYPES];
518 * total amount of evictable data in this state
520 refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
522 * total amount of data in this state; this includes: evictable,
523 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
525 refcount_t arcs_size;
527 * supports the "dbufs" kstat
529 arc_state_type_t arcs_state;
533 * Percentage that can be consumed by dnodes of ARC meta buffers.
535 int zfs_arc_meta_prune = 10000;
536 unsigned long zfs_arc_dnode_limit_percent = 10;
537 int zfs_arc_meta_strategy = ARC_STRATEGY_META_BALANCED;
538 int zfs_arc_meta_adjust_restarts = 4096;
541 static arc_state_t ARC_anon;
542 static arc_state_t ARC_mru;
543 static arc_state_t ARC_mru_ghost;
544 static arc_state_t ARC_mfu;
545 static arc_state_t ARC_mfu_ghost;
546 static arc_state_t ARC_l2c_only;
548 typedef struct arc_stats {
549 kstat_named_t arcstat_hits;
550 kstat_named_t arcstat_misses;
551 kstat_named_t arcstat_demand_data_hits;
552 kstat_named_t arcstat_demand_data_misses;
553 kstat_named_t arcstat_demand_metadata_hits;
554 kstat_named_t arcstat_demand_metadata_misses;
555 kstat_named_t arcstat_prefetch_data_hits;
556 kstat_named_t arcstat_prefetch_data_misses;
557 kstat_named_t arcstat_prefetch_metadata_hits;
558 kstat_named_t arcstat_prefetch_metadata_misses;
559 kstat_named_t arcstat_mru_hits;
560 kstat_named_t arcstat_mru_ghost_hits;
561 kstat_named_t arcstat_mfu_hits;
562 kstat_named_t arcstat_mfu_ghost_hits;
563 kstat_named_t arcstat_allocated;
564 kstat_named_t arcstat_deleted;
566 * Number of buffers that could not be evicted because the hash lock
567 * was held by another thread. The lock may not necessarily be held
568 * by something using the same buffer, since hash locks are shared
569 * by multiple buffers.
571 kstat_named_t arcstat_mutex_miss;
573 * Number of buffers skipped when updating the access state due to the
574 * header having already been released after acquiring the hash lock.
576 kstat_named_t arcstat_access_skip;
578 * Number of buffers skipped because they have I/O in progress, are
579 * indirect prefetch buffers that have not lived long enough, or are
580 * not from the spa we're trying to evict from.
582 kstat_named_t arcstat_evict_skip;
584 * Number of times arc_evict_state() was unable to evict enough
585 * buffers to reach it's target amount.
587 kstat_named_t arcstat_evict_not_enough;
588 kstat_named_t arcstat_evict_l2_cached;
589 kstat_named_t arcstat_evict_l2_eligible;
590 kstat_named_t arcstat_evict_l2_ineligible;
591 kstat_named_t arcstat_evict_l2_skip;
592 kstat_named_t arcstat_hash_elements;
593 kstat_named_t arcstat_hash_elements_max;
594 kstat_named_t arcstat_hash_collisions;
595 kstat_named_t arcstat_hash_chains;
596 kstat_named_t arcstat_hash_chain_max;
597 kstat_named_t arcstat_p;
598 kstat_named_t arcstat_c;
599 kstat_named_t arcstat_c_min;
600 kstat_named_t arcstat_c_max;
601 /* Not updated directly; only synced in arc_kstat_update. */
602 kstat_named_t arcstat_size;
604 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
605 * Note that the compressed bytes may match the uncompressed bytes
606 * if the block is either not compressed or compressed arc is disabled.
608 kstat_named_t arcstat_compressed_size;
610 * Uncompressed size of the data stored in b_pabd. If compressed
611 * arc is disabled then this value will be identical to the stat
614 kstat_named_t arcstat_uncompressed_size;
616 * Number of bytes stored in all the arc_buf_t's. This is classified
617 * as "overhead" since this data is typically short-lived and will
618 * be evicted from the arc when it becomes unreferenced unless the
619 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
620 * values have been set (see comment in dbuf.c for more information).
622 kstat_named_t arcstat_overhead_size;
624 * Number of bytes consumed by internal ARC structures necessary
625 * for tracking purposes; these structures are not actually
626 * backed by ARC buffers. This includes arc_buf_hdr_t structures
627 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
628 * caches), and arc_buf_t structures (allocated via arc_buf_t
630 * Not updated directly; only synced in arc_kstat_update.
632 kstat_named_t arcstat_hdr_size;
634 * Number of bytes consumed by ARC buffers of type equal to
635 * ARC_BUFC_DATA. This is generally consumed by buffers backing
636 * on disk user data (e.g. plain file contents).
637 * Not updated directly; only synced in arc_kstat_update.
639 kstat_named_t arcstat_data_size;
641 * Number of bytes consumed by ARC buffers of type equal to
642 * ARC_BUFC_METADATA. This is generally consumed by buffers
643 * backing on disk data that is used for internal ZFS
644 * structures (e.g. ZAP, dnode, indirect blocks, etc).
645 * Not updated directly; only synced in arc_kstat_update.
647 kstat_named_t arcstat_metadata_size;
649 * Number of bytes consumed by dmu_buf_impl_t objects.
651 kstat_named_t arcstat_dbuf_size;
653 * Number of bytes consumed by dnode_t objects.
655 kstat_named_t arcstat_dnode_size;
657 * Number of bytes consumed by bonus buffers.
659 kstat_named_t arcstat_bonus_size;
661 * Total number of bytes consumed by ARC buffers residing in the
662 * arc_anon state. This includes *all* buffers in the arc_anon
663 * state; e.g. data, metadata, evictable, and unevictable buffers
664 * are all included in this value.
665 * Not updated directly; only synced in arc_kstat_update.
667 kstat_named_t arcstat_anon_size;
669 * Number of bytes consumed by ARC buffers that meet the
670 * following criteria: backing buffers of type ARC_BUFC_DATA,
671 * residing in the arc_anon state, and are eligible for eviction
672 * (e.g. have no outstanding holds on the buffer).
673 * Not updated directly; only synced in arc_kstat_update.
675 kstat_named_t arcstat_anon_evictable_data;
677 * Number of bytes consumed by ARC buffers that meet the
678 * following criteria: backing buffers of type ARC_BUFC_METADATA,
679 * residing in the arc_anon state, and are eligible for eviction
680 * (e.g. have no outstanding holds on the buffer).
681 * Not updated directly; only synced in arc_kstat_update.
683 kstat_named_t arcstat_anon_evictable_metadata;
685 * Total number of bytes consumed by ARC buffers residing in the
686 * arc_mru state. This includes *all* buffers in the arc_mru
687 * state; e.g. data, metadata, evictable, and unevictable buffers
688 * are all included in this value.
689 * Not updated directly; only synced in arc_kstat_update.
691 kstat_named_t arcstat_mru_size;
693 * Number of bytes consumed by ARC buffers that meet the
694 * following criteria: backing buffers of type ARC_BUFC_DATA,
695 * residing in the arc_mru state, and are eligible for eviction
696 * (e.g. have no outstanding holds on the buffer).
697 * Not updated directly; only synced in arc_kstat_update.
699 kstat_named_t arcstat_mru_evictable_data;
701 * Number of bytes consumed by ARC buffers that meet the
702 * following criteria: backing buffers of type ARC_BUFC_METADATA,
703 * residing in the arc_mru state, and are eligible for eviction
704 * (e.g. have no outstanding holds on the buffer).
705 * Not updated directly; only synced in arc_kstat_update.
707 kstat_named_t arcstat_mru_evictable_metadata;
709 * Total number of bytes that *would have been* consumed by ARC
710 * buffers in the arc_mru_ghost state. The key thing to note
711 * here, is the fact that this size doesn't actually indicate
712 * RAM consumption. The ghost lists only consist of headers and
713 * don't actually have ARC buffers linked off of these headers.
714 * Thus, *if* the headers had associated ARC buffers, these
715 * buffers *would have* consumed this number of bytes.
716 * Not updated directly; only synced in arc_kstat_update.
718 kstat_named_t arcstat_mru_ghost_size;
720 * Number of bytes that *would have been* consumed by ARC
721 * buffers that are eligible for eviction, of type
722 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
723 * Not updated directly; only synced in arc_kstat_update.
725 kstat_named_t arcstat_mru_ghost_evictable_data;
727 * Number of bytes that *would have been* consumed by ARC
728 * buffers that are eligible for eviction, of type
729 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
730 * Not updated directly; only synced in arc_kstat_update.
732 kstat_named_t arcstat_mru_ghost_evictable_metadata;
734 * Total number of bytes consumed by ARC buffers residing in the
735 * arc_mfu state. This includes *all* buffers in the arc_mfu
736 * state; e.g. data, metadata, evictable, and unevictable buffers
737 * are all included in this value.
738 * Not updated directly; only synced in arc_kstat_update.
740 kstat_named_t arcstat_mfu_size;
742 * Number of bytes consumed by ARC buffers that are eligible for
743 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
745 * Not updated directly; only synced in arc_kstat_update.
747 kstat_named_t arcstat_mfu_evictable_data;
749 * Number of bytes consumed by ARC buffers that are eligible for
750 * eviction, of type ARC_BUFC_METADATA, and reside in the
752 * Not updated directly; only synced in arc_kstat_update.
754 kstat_named_t arcstat_mfu_evictable_metadata;
756 * Total number of bytes that *would have been* consumed by ARC
757 * buffers in the arc_mfu_ghost state. See the comment above
758 * arcstat_mru_ghost_size for more details.
759 * Not updated directly; only synced in arc_kstat_update.
761 kstat_named_t arcstat_mfu_ghost_size;
763 * Number of bytes that *would have been* consumed by ARC
764 * buffers that are eligible for eviction, of type
765 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
766 * Not updated directly; only synced in arc_kstat_update.
768 kstat_named_t arcstat_mfu_ghost_evictable_data;
770 * Number of bytes that *would have been* consumed by ARC
771 * buffers that are eligible for eviction, of type
772 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
773 * Not updated directly; only synced in arc_kstat_update.
775 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
776 kstat_named_t arcstat_l2_hits;
777 kstat_named_t arcstat_l2_misses;
778 kstat_named_t arcstat_l2_feeds;
779 kstat_named_t arcstat_l2_rw_clash;
780 kstat_named_t arcstat_l2_read_bytes;
781 kstat_named_t arcstat_l2_write_bytes;
782 kstat_named_t arcstat_l2_writes_sent;
783 kstat_named_t arcstat_l2_writes_done;
784 kstat_named_t arcstat_l2_writes_error;
785 kstat_named_t arcstat_l2_writes_lock_retry;
786 kstat_named_t arcstat_l2_evict_lock_retry;
787 kstat_named_t arcstat_l2_evict_reading;
788 kstat_named_t arcstat_l2_evict_l1cached;
789 kstat_named_t arcstat_l2_free_on_write;
790 kstat_named_t arcstat_l2_abort_lowmem;
791 kstat_named_t arcstat_l2_cksum_bad;
792 kstat_named_t arcstat_l2_io_error;
793 kstat_named_t arcstat_l2_lsize;
794 kstat_named_t arcstat_l2_psize;
795 /* Not updated directly; only synced in arc_kstat_update. */
796 kstat_named_t arcstat_l2_hdr_size;
797 kstat_named_t arcstat_l2_write_trylock_fail;
798 kstat_named_t arcstat_l2_write_passed_headroom;
799 kstat_named_t arcstat_l2_write_spa_mismatch;
800 kstat_named_t arcstat_l2_write_in_l2;
801 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
802 kstat_named_t arcstat_l2_write_not_cacheable;
803 kstat_named_t arcstat_l2_write_full;
804 kstat_named_t arcstat_l2_write_buffer_iter;
805 kstat_named_t arcstat_l2_write_pios;
806 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
807 kstat_named_t arcstat_l2_write_buffer_list_iter;
808 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
809 kstat_named_t arcstat_memory_throttle_count;
810 kstat_named_t arcstat_memory_direct_count;
811 kstat_named_t arcstat_memory_indirect_count;
812 kstat_named_t arcstat_memory_all_bytes;
813 kstat_named_t arcstat_memory_free_bytes;
814 kstat_named_t arcstat_memory_available_bytes;
815 kstat_named_t arcstat_no_grow;
816 kstat_named_t arcstat_tempreserve;
817 kstat_named_t arcstat_loaned_bytes;
818 kstat_named_t arcstat_prune;
819 /* Not updated directly; only synced in arc_kstat_update. */
820 kstat_named_t arcstat_meta_used;
821 kstat_named_t arcstat_meta_limit;
822 kstat_named_t arcstat_dnode_limit;
823 kstat_named_t arcstat_meta_max;
824 kstat_named_t arcstat_meta_min;
825 kstat_named_t arcstat_async_upgrade_sync;
826 kstat_named_t arcstat_demand_hit_predictive_prefetch;
827 kstat_named_t arcstat_demand_hit_prescient_prefetch;
830 static arc_stats_t arc_stats = {
831 { "hits", KSTAT_DATA_UINT64 },
832 { "misses", KSTAT_DATA_UINT64 },
833 { "demand_data_hits", KSTAT_DATA_UINT64 },
834 { "demand_data_misses", KSTAT_DATA_UINT64 },
835 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
836 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
837 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
838 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
839 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
840 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
841 { "mru_hits", KSTAT_DATA_UINT64 },
842 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
843 { "mfu_hits", KSTAT_DATA_UINT64 },
844 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
845 { "allocated", KSTAT_DATA_UINT64 },
846 { "deleted", KSTAT_DATA_UINT64 },
847 { "mutex_miss", KSTAT_DATA_UINT64 },
848 { "access_skip", KSTAT_DATA_UINT64 },
849 { "evict_skip", KSTAT_DATA_UINT64 },
850 { "evict_not_enough", KSTAT_DATA_UINT64 },
851 { "evict_l2_cached", KSTAT_DATA_UINT64 },
852 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
853 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
854 { "evict_l2_skip", KSTAT_DATA_UINT64 },
855 { "hash_elements", KSTAT_DATA_UINT64 },
856 { "hash_elements_max", KSTAT_DATA_UINT64 },
857 { "hash_collisions", KSTAT_DATA_UINT64 },
858 { "hash_chains", KSTAT_DATA_UINT64 },
859 { "hash_chain_max", KSTAT_DATA_UINT64 },
860 { "p", KSTAT_DATA_UINT64 },
861 { "c", KSTAT_DATA_UINT64 },
862 { "c_min", KSTAT_DATA_UINT64 },
863 { "c_max", KSTAT_DATA_UINT64 },
864 { "size", KSTAT_DATA_UINT64 },
865 { "compressed_size", KSTAT_DATA_UINT64 },
866 { "uncompressed_size", KSTAT_DATA_UINT64 },
867 { "overhead_size", KSTAT_DATA_UINT64 },
868 { "hdr_size", KSTAT_DATA_UINT64 },
869 { "data_size", KSTAT_DATA_UINT64 },
870 { "metadata_size", KSTAT_DATA_UINT64 },
871 { "dbuf_size", KSTAT_DATA_UINT64 },
872 { "dnode_size", KSTAT_DATA_UINT64 },
873 { "bonus_size", KSTAT_DATA_UINT64 },
874 { "anon_size", KSTAT_DATA_UINT64 },
875 { "anon_evictable_data", KSTAT_DATA_UINT64 },
876 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
877 { "mru_size", KSTAT_DATA_UINT64 },
878 { "mru_evictable_data", KSTAT_DATA_UINT64 },
879 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
880 { "mru_ghost_size", KSTAT_DATA_UINT64 },
881 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
882 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
883 { "mfu_size", KSTAT_DATA_UINT64 },
884 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
885 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
886 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
887 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
888 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
889 { "l2_hits", KSTAT_DATA_UINT64 },
890 { "l2_misses", KSTAT_DATA_UINT64 },
891 { "l2_feeds", KSTAT_DATA_UINT64 },
892 { "l2_rw_clash", KSTAT_DATA_UINT64 },
893 { "l2_read_bytes", KSTAT_DATA_UINT64 },
894 { "l2_write_bytes", KSTAT_DATA_UINT64 },
895 { "l2_writes_sent", KSTAT_DATA_UINT64 },
896 { "l2_writes_done", KSTAT_DATA_UINT64 },
897 { "l2_writes_error", KSTAT_DATA_UINT64 },
898 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
899 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
900 { "l2_evict_reading", KSTAT_DATA_UINT64 },
901 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
902 { "l2_free_on_write", KSTAT_DATA_UINT64 },
903 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
904 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
905 { "l2_io_error", KSTAT_DATA_UINT64 },
906 { "l2_size", KSTAT_DATA_UINT64 },
907 { "l2_asize", KSTAT_DATA_UINT64 },
908 { "l2_hdr_size", KSTAT_DATA_UINT64 },
909 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
910 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
911 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
912 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
913 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
914 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
915 { "l2_write_full", KSTAT_DATA_UINT64 },
916 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
917 { "l2_write_pios", KSTAT_DATA_UINT64 },
918 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
919 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
920 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
921 { "memory_throttle_count", KSTAT_DATA_UINT64 },
922 { "memory_direct_count", KSTAT_DATA_UINT64 },
923 { "memory_indirect_count", KSTAT_DATA_UINT64 },
924 { "memory_all_bytes", KSTAT_DATA_UINT64 },
925 { "memory_free_bytes", KSTAT_DATA_UINT64 },
926 { "memory_available_bytes", KSTAT_DATA_UINT64 },
927 { "arc_no_grow", KSTAT_DATA_UINT64 },
928 { "arc_tempreserve", KSTAT_DATA_UINT64 },
929 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
930 { "arc_prune", KSTAT_DATA_UINT64 },
931 { "arc_meta_used", KSTAT_DATA_UINT64 },
932 { "arc_meta_limit", KSTAT_DATA_UINT64 },
933 { "arc_dnode_limit", KSTAT_DATA_UINT64 },
934 { "arc_meta_max", KSTAT_DATA_UINT64 },
935 { "arc_meta_min", KSTAT_DATA_UINT64 },
936 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
937 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
938 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
941 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
943 #define ARCSTAT_INCR(stat, val) \
944 atomic_add_64(&arc_stats.stat.value.ui64, (val))
946 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
947 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
949 #define ARCSTAT_MAX(stat, val) { \
951 while ((val) > (m = arc_stats.stat.value.ui64) && \
952 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
956 #define ARCSTAT_MAXSTAT(stat) \
957 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
960 * We define a macro to allow ARC hits/misses to be easily broken down by
961 * two separate conditions, giving a total of four different subtypes for
962 * each of hits and misses (so eight statistics total).
964 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
967 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
969 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
973 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
975 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
980 static arc_state_t *arc_anon;
981 static arc_state_t *arc_mru;
982 static arc_state_t *arc_mru_ghost;
983 static arc_state_t *arc_mfu;
984 static arc_state_t *arc_mfu_ghost;
985 static arc_state_t *arc_l2c_only;
988 * There are several ARC variables that are critical to export as kstats --
989 * but we don't want to have to grovel around in the kstat whenever we wish to
990 * manipulate them. For these variables, we therefore define them to be in
991 * terms of the statistic variable. This assures that we are not introducing
992 * the possibility of inconsistency by having shadow copies of the variables,
993 * while still allowing the code to be readable.
995 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
996 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
997 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
998 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
999 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
1000 #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
1001 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
1002 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
1003 #define arc_dbuf_size ARCSTAT(arcstat_dbuf_size) /* dbuf metadata */
1004 #define arc_dnode_size ARCSTAT(arcstat_dnode_size) /* dnode metadata */
1005 #define arc_bonus_size ARCSTAT(arcstat_bonus_size) /* bonus buffer metadata */
1007 /* compressed size of entire arc */
1008 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
1009 /* uncompressed size of entire arc */
1010 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
1011 /* number of bytes in the arc from arc_buf_t's */
1012 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
1015 * There are also some ARC variables that we want to export, but that are
1016 * updated so often that having the canonical representation be the statistic
1017 * variable causes a performance bottleneck. We want to use aggsum_t's for these
1018 * instead, but still be able to export the kstat in the same way as before.
1019 * The solution is to always use the aggsum version, except in the kstat update
1023 aggsum_t arc_meta_used;
1024 aggsum_t astat_data_size;
1025 aggsum_t astat_metadata_size;
1026 aggsum_t astat_hdr_size;
1027 aggsum_t astat_bonus_size;
1028 aggsum_t astat_dnode_size;
1029 aggsum_t astat_dbuf_size;
1030 aggsum_t astat_l2_hdr_size;
1032 static list_t arc_prune_list;
1033 static kmutex_t arc_prune_mtx;
1034 static taskq_t *arc_prune_taskq;
1036 static int arc_no_grow; /* Don't try to grow cache size */
1037 static uint64_t arc_tempreserve;
1038 static uint64_t arc_loaned_bytes;
1040 typedef struct arc_callback arc_callback_t;
1042 struct arc_callback {
1044 arc_read_done_func_t *acb_done;
1046 boolean_t acb_compressed;
1047 zio_t *acb_zio_dummy;
1048 zio_t *acb_zio_head;
1049 arc_callback_t *acb_next;
1052 typedef struct arc_write_callback arc_write_callback_t;
1054 struct arc_write_callback {
1056 arc_write_done_func_t *awcb_ready;
1057 arc_write_done_func_t *awcb_children_ready;
1058 arc_write_done_func_t *awcb_physdone;
1059 arc_write_done_func_t *awcb_done;
1060 arc_buf_t *awcb_buf;
1064 * ARC buffers are separated into multiple structs as a memory saving measure:
1065 * - Common fields struct, always defined, and embedded within it:
1066 * - L2-only fields, always allocated but undefined when not in L2ARC
1067 * - L1-only fields, only allocated when in L1ARC
1069 * Buffer in L1 Buffer only in L2
1070 * +------------------------+ +------------------------+
1071 * | arc_buf_hdr_t | | arc_buf_hdr_t |
1075 * +------------------------+ +------------------------+
1076 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
1077 * | (undefined if L1-only) | | |
1078 * +------------------------+ +------------------------+
1079 * | l1arc_buf_hdr_t |
1084 * +------------------------+
1086 * Because it's possible for the L2ARC to become extremely large, we can wind
1087 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
1088 * is minimized by only allocating the fields necessary for an L1-cached buffer
1089 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
1090 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
1091 * words in pointers. arc_hdr_realloc() is used to switch a header between
1092 * these two allocation states.
1094 typedef struct l1arc_buf_hdr {
1095 kmutex_t b_freeze_lock;
1096 zio_cksum_t *b_freeze_cksum;
1099 * Used for debugging with kmem_flags - by allocating and freeing
1100 * b_thawed when the buffer is thawed, we get a record of the stack
1101 * trace that thawed it.
1108 /* for waiting on writes to complete */
1112 /* protected by arc state mutex */
1113 arc_state_t *b_state;
1114 multilist_node_t b_arc_node;
1116 /* updated atomically */
1117 clock_t b_arc_access;
1118 uint32_t b_mru_hits;
1119 uint32_t b_mru_ghost_hits;
1120 uint32_t b_mfu_hits;
1121 uint32_t b_mfu_ghost_hits;
1124 /* self protecting */
1125 refcount_t b_refcnt;
1127 arc_callback_t *b_acb;
1131 typedef struct l2arc_dev l2arc_dev_t;
1133 typedef struct l2arc_buf_hdr {
1134 /* protected by arc_buf_hdr mutex */
1135 l2arc_dev_t *b_dev; /* L2ARC device */
1136 uint64_t b_daddr; /* disk address, offset byte */
1139 list_node_t b_l2node;
1142 struct arc_buf_hdr {
1143 /* protected by hash lock */
1147 arc_buf_contents_t b_type;
1148 arc_buf_hdr_t *b_hash_next;
1149 arc_flags_t b_flags;
1152 * This field stores the size of the data buffer after
1153 * compression, and is set in the arc's zio completion handlers.
1154 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
1156 * While the block pointers can store up to 32MB in their psize
1157 * field, we can only store up to 32MB minus 512B. This is due
1158 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
1159 * a field of zeros represents 512B in the bp). We can't use a
1160 * bias of 1 since we need to reserve a psize of zero, here, to
1161 * represent holes and embedded blocks.
1163 * This isn't a problem in practice, since the maximum size of a
1164 * buffer is limited to 16MB, so we never need to store 32MB in
1165 * this field. Even in the upstream illumos code base, the
1166 * maximum size of a buffer is limited to 16MB.
1171 * This field stores the size of the data buffer before
1172 * compression, and cannot change once set. It is in units
1173 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
1175 uint16_t b_lsize; /* immutable */
1176 uint64_t b_spa; /* immutable */
1178 /* L2ARC fields. Undefined when not in L2ARC. */
1179 l2arc_buf_hdr_t b_l2hdr;
1180 /* L1ARC fields. Undefined when in l2arc_only state */
1181 l1arc_buf_hdr_t b_l1hdr;
1184 #if defined(__FreeBSD__) && defined(_KERNEL)
1186 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
1191 val = arc_meta_limit;
1192 err = sysctl_handle_64(oidp, &val, 0, req);
1193 if (err != 0 || req->newptr == NULL)
1196 if (val <= 0 || val > arc_c_max)
1199 arc_meta_limit = val;
1204 sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS)
1209 val = arc_no_grow_shift;
1210 err = sysctl_handle_32(oidp, &val, 0, req);
1211 if (err != 0 || req->newptr == NULL)
1214 if (val >= arc_shrink_shift)
1217 arc_no_grow_shift = val;
1222 sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
1228 err = sysctl_handle_64(oidp, &val, 0, req);
1229 if (err != 0 || req->newptr == NULL)
1232 if (zfs_arc_max == 0) {
1233 /* Loader tunable so blindly set */
1238 if (val < arc_abs_min || val > kmem_size())
1240 if (val < arc_c_min)
1242 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
1248 arc_p = (arc_c >> 1);
1250 if (zfs_arc_meta_limit == 0) {
1251 /* limit meta-data to 1/4 of the arc capacity */
1252 arc_meta_limit = arc_c_max / 4;
1255 /* if kmem_flags are set, lets try to use less memory */
1256 if (kmem_debugging())
1259 zfs_arc_max = arc_c;
1265 sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
1271 err = sysctl_handle_64(oidp, &val, 0, req);
1272 if (err != 0 || req->newptr == NULL)
1275 if (zfs_arc_min == 0) {
1276 /* Loader tunable so blindly set */
1281 if (val < arc_abs_min || val > arc_c_max)
1286 if (zfs_arc_meta_min == 0)
1287 arc_meta_min = arc_c_min / 2;
1289 if (arc_c < arc_c_min)
1292 zfs_arc_min = arc_c_min;
1298 #define GHOST_STATE(state) \
1299 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
1300 (state) == arc_l2c_only)
1302 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1303 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1304 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1305 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1306 #define HDR_PRESCIENT_PREFETCH(hdr) \
1307 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
1308 #define HDR_COMPRESSION_ENABLED(hdr) \
1309 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1311 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1312 #define HDR_L2_READING(hdr) \
1313 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1314 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1315 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1316 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1317 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1318 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1320 #define HDR_ISTYPE_METADATA(hdr) \
1321 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1322 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1324 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1325 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1327 /* For storing compression mode in b_flags */
1328 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1330 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1331 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1332 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1333 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1335 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1336 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1337 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1343 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1344 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1347 * Hash table routines
1350 #define HT_LOCK_PAD CACHE_LINE_SIZE
1355 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1359 #define BUF_LOCKS 256
1360 typedef struct buf_hash_table {
1362 arc_buf_hdr_t **ht_table;
1363 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1366 static buf_hash_table_t buf_hash_table;
1368 #define BUF_HASH_INDEX(spa, dva, birth) \
1369 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1370 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1371 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1372 #define HDR_LOCK(hdr) \
1373 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1375 uint64_t zfs_crc64_table[256];
1381 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1382 #define L2ARC_HEADROOM 2 /* num of writes */
1384 * If we discover during ARC scan any buffers to be compressed, we boost
1385 * our headroom for the next scanning cycle by this percentage multiple.
1387 #define L2ARC_HEADROOM_BOOST 200
1388 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1389 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1391 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1392 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1394 /* L2ARC Performance Tunables */
1395 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1396 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1397 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1398 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1399 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1400 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1401 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1402 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1403 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1405 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1406 &l2arc_write_max, 0, "max write size");
1407 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1408 &l2arc_write_boost, 0, "extra write during warmup");
1409 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1410 &l2arc_headroom, 0, "number of dev writes");
1411 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1412 &l2arc_feed_secs, 0, "interval seconds");
1413 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1414 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1416 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1417 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1418 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1419 &l2arc_feed_again, 0, "turbo warmup");
1420 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1421 &l2arc_norw, 0, "no reads during writes");
1423 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1424 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1425 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD,
1426 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1427 "size of anonymous state");
1428 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD,
1429 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1430 "size of anonymous state");
1432 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1433 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1434 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD,
1435 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1436 "size of metadata in mru state");
1437 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD,
1438 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1439 "size of data in mru state");
1441 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1442 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1443 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD,
1444 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1445 "size of metadata in mru ghost state");
1446 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD,
1447 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1448 "size of data in mru ghost state");
1450 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1451 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1452 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD,
1453 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1454 "size of metadata in mfu state");
1455 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD,
1456 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1457 "size of data in mfu state");
1459 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1460 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1461 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD,
1462 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1463 "size of metadata in mfu ghost state");
1464 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD,
1465 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1466 "size of data in mfu ghost state");
1468 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1469 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1471 SYSCTL_UINT(_vfs_zfs, OID_AUTO, arc_min_prefetch_ms, CTLFLAG_RW,
1472 &zfs_arc_min_prefetch_ms, 0, "Min life of prefetch block in ms");
1473 SYSCTL_UINT(_vfs_zfs, OID_AUTO, arc_min_prescient_prefetch_ms, CTLFLAG_RW,
1474 &zfs_arc_min_prescient_prefetch_ms, 0, "Min life of prescient prefetched block in ms");
1480 vdev_t *l2ad_vdev; /* vdev */
1481 spa_t *l2ad_spa; /* spa */
1482 uint64_t l2ad_hand; /* next write location */
1483 uint64_t l2ad_start; /* first addr on device */
1484 uint64_t l2ad_end; /* last addr on device */
1485 boolean_t l2ad_first; /* first sweep through */
1486 boolean_t l2ad_writing; /* currently writing */
1487 kmutex_t l2ad_mtx; /* lock for buffer list */
1488 list_t l2ad_buflist; /* buffer list */
1489 list_node_t l2ad_node; /* device list node */
1490 refcount_t l2ad_alloc; /* allocated bytes */
1493 static list_t L2ARC_dev_list; /* device list */
1494 static list_t *l2arc_dev_list; /* device list pointer */
1495 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1496 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1497 static list_t L2ARC_free_on_write; /* free after write buf list */
1498 static list_t *l2arc_free_on_write; /* free after write list ptr */
1499 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1500 static uint64_t l2arc_ndev; /* number of devices */
1502 typedef struct l2arc_read_callback {
1503 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1504 blkptr_t l2rcb_bp; /* original blkptr */
1505 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1506 int l2rcb_flags; /* original flags */
1507 abd_t *l2rcb_abd; /* temporary buffer */
1508 } l2arc_read_callback_t;
1510 typedef struct l2arc_write_callback {
1511 l2arc_dev_t *l2wcb_dev; /* device info */
1512 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1513 } l2arc_write_callback_t;
1515 typedef struct l2arc_data_free {
1516 /* protected by l2arc_free_on_write_mtx */
1519 arc_buf_contents_t l2df_type;
1520 list_node_t l2df_list_node;
1521 } l2arc_data_free_t;
1523 static kmutex_t l2arc_feed_thr_lock;
1524 static kcondvar_t l2arc_feed_thr_cv;
1525 static uint8_t l2arc_thread_exit;
1527 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1528 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1529 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1530 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1531 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1532 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1533 static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1534 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *);
1535 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1536 static boolean_t arc_is_overflowing();
1537 static void arc_buf_watch(arc_buf_t *);
1538 static void arc_prune_async(int64_t);
1540 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1541 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1542 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1543 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1545 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1546 static void l2arc_read_done(zio_t *);
1549 l2arc_trim(const arc_buf_hdr_t *hdr)
1551 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1553 ASSERT(HDR_HAS_L2HDR(hdr));
1554 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1556 if (HDR_GET_PSIZE(hdr) != 0) {
1557 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1558 HDR_GET_PSIZE(hdr), 0);
1563 * We use Cityhash for this. It's fast, and has good hash properties without
1564 * requiring any large static buffers.
1567 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1569 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
1572 #define HDR_EMPTY(hdr) \
1573 ((hdr)->b_dva.dva_word[0] == 0 && \
1574 (hdr)->b_dva.dva_word[1] == 0)
1576 #define HDR_EQUAL(spa, dva, birth, hdr) \
1577 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1578 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1579 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1582 buf_discard_identity(arc_buf_hdr_t *hdr)
1584 hdr->b_dva.dva_word[0] = 0;
1585 hdr->b_dva.dva_word[1] = 0;
1589 static arc_buf_hdr_t *
1590 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1592 const dva_t *dva = BP_IDENTITY(bp);
1593 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1594 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1595 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1598 mutex_enter(hash_lock);
1599 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1600 hdr = hdr->b_hash_next) {
1601 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1606 mutex_exit(hash_lock);
1612 * Insert an entry into the hash table. If there is already an element
1613 * equal to elem in the hash table, then the already existing element
1614 * will be returned and the new element will not be inserted.
1615 * Otherwise returns NULL.
1616 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1618 static arc_buf_hdr_t *
1619 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1621 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1622 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1623 arc_buf_hdr_t *fhdr;
1626 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1627 ASSERT(hdr->b_birth != 0);
1628 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1630 if (lockp != NULL) {
1632 mutex_enter(hash_lock);
1634 ASSERT(MUTEX_HELD(hash_lock));
1637 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1638 fhdr = fhdr->b_hash_next, i++) {
1639 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1643 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1644 buf_hash_table.ht_table[idx] = hdr;
1645 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1647 /* collect some hash table performance data */
1649 ARCSTAT_BUMP(arcstat_hash_collisions);
1651 ARCSTAT_BUMP(arcstat_hash_chains);
1653 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1656 ARCSTAT_BUMP(arcstat_hash_elements);
1657 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1663 buf_hash_remove(arc_buf_hdr_t *hdr)
1665 arc_buf_hdr_t *fhdr, **hdrp;
1666 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1668 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1669 ASSERT(HDR_IN_HASH_TABLE(hdr));
1671 hdrp = &buf_hash_table.ht_table[idx];
1672 while ((fhdr = *hdrp) != hdr) {
1673 ASSERT3P(fhdr, !=, NULL);
1674 hdrp = &fhdr->b_hash_next;
1676 *hdrp = hdr->b_hash_next;
1677 hdr->b_hash_next = NULL;
1678 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1680 /* collect some hash table performance data */
1681 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1683 if (buf_hash_table.ht_table[idx] &&
1684 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1685 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1689 * Global data structures and functions for the buf kmem cache.
1691 static kmem_cache_t *hdr_full_cache;
1692 static kmem_cache_t *hdr_l2only_cache;
1693 static kmem_cache_t *buf_cache;
1700 kmem_free(buf_hash_table.ht_table,
1701 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1702 for (i = 0; i < BUF_LOCKS; i++)
1703 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1704 kmem_cache_destroy(hdr_full_cache);
1705 kmem_cache_destroy(hdr_l2only_cache);
1706 kmem_cache_destroy(buf_cache);
1710 * Constructor callback - called when the cache is empty
1711 * and a new buf is requested.
1715 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1717 arc_buf_hdr_t *hdr = vbuf;
1719 bzero(hdr, HDR_FULL_SIZE);
1720 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1721 refcount_create(&hdr->b_l1hdr.b_refcnt);
1722 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1723 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1724 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1731 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1733 arc_buf_hdr_t *hdr = vbuf;
1735 bzero(hdr, HDR_L2ONLY_SIZE);
1736 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1743 buf_cons(void *vbuf, void *unused, int kmflag)
1745 arc_buf_t *buf = vbuf;
1747 bzero(buf, sizeof (arc_buf_t));
1748 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1749 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1755 * Destructor callback - called when a cached buf is
1756 * no longer required.
1760 hdr_full_dest(void *vbuf, void *unused)
1762 arc_buf_hdr_t *hdr = vbuf;
1764 ASSERT(HDR_EMPTY(hdr));
1765 cv_destroy(&hdr->b_l1hdr.b_cv);
1766 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1767 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1768 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1769 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1774 hdr_l2only_dest(void *vbuf, void *unused)
1776 arc_buf_hdr_t *hdr = vbuf;
1778 ASSERT(HDR_EMPTY(hdr));
1779 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1784 buf_dest(void *vbuf, void *unused)
1786 arc_buf_t *buf = vbuf;
1788 mutex_destroy(&buf->b_evict_lock);
1789 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1793 * Reclaim callback -- invoked when memory is low.
1797 hdr_recl(void *unused)
1799 dprintf("hdr_recl called\n");
1801 * umem calls the reclaim func when we destroy the buf cache,
1802 * which is after we do arc_fini().
1805 cv_signal(&arc_reclaim_thread_cv);
1812 uint64_t hsize = 1ULL << 12;
1816 * The hash table is big enough to fill all of physical memory
1817 * with an average block size of zfs_arc_average_blocksize (default 8K).
1818 * By default, the table will take up
1819 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1821 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1824 buf_hash_table.ht_mask = hsize - 1;
1825 buf_hash_table.ht_table =
1826 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1827 if (buf_hash_table.ht_table == NULL) {
1828 ASSERT(hsize > (1ULL << 8));
1833 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1834 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1835 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1836 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1838 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1839 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1841 for (i = 0; i < 256; i++)
1842 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1843 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1845 for (i = 0; i < BUF_LOCKS; i++) {
1846 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1847 NULL, MUTEX_DEFAULT, NULL);
1852 * This is the size that the buf occupies in memory. If the buf is compressed,
1853 * it will correspond to the compressed size. You should use this method of
1854 * getting the buf size unless you explicitly need the logical size.
1857 arc_buf_size(arc_buf_t *buf)
1859 return (ARC_BUF_COMPRESSED(buf) ?
1860 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1864 arc_buf_lsize(arc_buf_t *buf)
1866 return (HDR_GET_LSIZE(buf->b_hdr));
1870 arc_get_compression(arc_buf_t *buf)
1872 return (ARC_BUF_COMPRESSED(buf) ?
1873 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1876 #define ARC_MINTIME (hz>>4) /* 62 ms */
1878 static inline boolean_t
1879 arc_buf_is_shared(arc_buf_t *buf)
1881 boolean_t shared = (buf->b_data != NULL &&
1882 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1883 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1884 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1885 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1886 IMPLY(shared, ARC_BUF_SHARED(buf));
1887 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1890 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1891 * already being shared" requirement prevents us from doing that.
1898 * Free the checksum associated with this header. If there is no checksum, this
1902 arc_cksum_free(arc_buf_hdr_t *hdr)
1904 ASSERT(HDR_HAS_L1HDR(hdr));
1905 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1906 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1907 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1908 hdr->b_l1hdr.b_freeze_cksum = NULL;
1910 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1914 * Return true iff at least one of the bufs on hdr is not compressed.
1917 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1919 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1920 if (!ARC_BUF_COMPRESSED(b)) {
1928 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1929 * matches the checksum that is stored in the hdr. If there is no checksum,
1930 * or if the buf is compressed, this is a no-op.
1933 arc_cksum_verify(arc_buf_t *buf)
1935 arc_buf_hdr_t *hdr = buf->b_hdr;
1938 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1941 if (ARC_BUF_COMPRESSED(buf)) {
1942 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1943 arc_hdr_has_uncompressed_buf(hdr));
1947 ASSERT(HDR_HAS_L1HDR(hdr));
1949 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1950 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1951 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1955 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1956 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1957 panic("buffer modified while frozen!");
1958 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1962 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1964 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1965 boolean_t valid_cksum;
1967 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1968 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1971 * We rely on the blkptr's checksum to determine if the block
1972 * is valid or not. When compressed arc is enabled, the l2arc
1973 * writes the block to the l2arc just as it appears in the pool.
1974 * This allows us to use the blkptr's checksum to validate the
1975 * data that we just read off of the l2arc without having to store
1976 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1977 * arc is disabled, then the data written to the l2arc is always
1978 * uncompressed and won't match the block as it exists in the main
1979 * pool. When this is the case, we must first compress it if it is
1980 * compressed on the main pool before we can validate the checksum.
1982 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1983 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1984 uint64_t lsize = HDR_GET_LSIZE(hdr);
1987 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
1988 csize = zio_compress_data(compress, zio->io_abd,
1989 abd_to_buf(cdata), lsize);
1991 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1992 if (csize < HDR_GET_PSIZE(hdr)) {
1994 * Compressed blocks are always a multiple of the
1995 * smallest ashift in the pool. Ideally, we would
1996 * like to round up the csize to the next
1997 * spa_min_ashift but that value may have changed
1998 * since the block was last written. Instead,
1999 * we rely on the fact that the hdr's psize
2000 * was set to the psize of the block when it was
2001 * last written. We set the csize to that value
2002 * and zero out any part that should not contain
2005 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
2006 csize = HDR_GET_PSIZE(hdr);
2008 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
2012 * Block pointers always store the checksum for the logical data.
2013 * If the block pointer has the gang bit set, then the checksum
2014 * it represents is for the reconstituted data and not for an
2015 * individual gang member. The zio pipeline, however, must be able to
2016 * determine the checksum of each of the gang constituents so it
2017 * treats the checksum comparison differently than what we need
2018 * for l2arc blocks. This prevents us from using the
2019 * zio_checksum_error() interface directly. Instead we must call the
2020 * zio_checksum_error_impl() so that we can ensure the checksum is
2021 * generated using the correct checksum algorithm and accounts for the
2022 * logical I/O size and not just a gang fragment.
2024 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
2025 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
2026 zio->io_offset, NULL) == 0);
2027 zio_pop_transforms(zio);
2028 return (valid_cksum);
2032 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
2033 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
2034 * isn't modified later on. If buf is compressed or there is already a checksum
2035 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
2038 arc_cksum_compute(arc_buf_t *buf)
2040 arc_buf_hdr_t *hdr = buf->b_hdr;
2042 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2045 ASSERT(HDR_HAS_L1HDR(hdr));
2047 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
2048 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
2049 ASSERT(arc_hdr_has_uncompressed_buf(hdr));
2050 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2052 } else if (ARC_BUF_COMPRESSED(buf)) {
2053 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2057 ASSERT(!ARC_BUF_COMPRESSED(buf));
2058 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
2060 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
2061 hdr->b_l1hdr.b_freeze_cksum);
2062 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2070 typedef struct procctl {
2078 arc_buf_unwatch(arc_buf_t *buf)
2085 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2086 ctl.prwatch.pr_size = 0;
2087 ctl.prwatch.pr_wflags = 0;
2088 result = write(arc_procfd, &ctl, sizeof (ctl));
2089 ASSERT3U(result, ==, sizeof (ctl));
2096 arc_buf_watch(arc_buf_t *buf)
2103 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2104 ctl.prwatch.pr_size = arc_buf_size(buf);
2105 ctl.prwatch.pr_wflags = WA_WRITE;
2106 result = write(arc_procfd, &ctl, sizeof (ctl));
2107 ASSERT3U(result, ==, sizeof (ctl));
2111 #endif /* illumos */
2113 static arc_buf_contents_t
2114 arc_buf_type(arc_buf_hdr_t *hdr)
2116 arc_buf_contents_t type;
2117 if (HDR_ISTYPE_METADATA(hdr)) {
2118 type = ARC_BUFC_METADATA;
2120 type = ARC_BUFC_DATA;
2122 VERIFY3U(hdr->b_type, ==, type);
2127 arc_is_metadata(arc_buf_t *buf)
2129 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
2133 arc_bufc_to_flags(arc_buf_contents_t type)
2137 /* metadata field is 0 if buffer contains normal data */
2139 case ARC_BUFC_METADATA:
2140 return (ARC_FLAG_BUFC_METADATA);
2144 panic("undefined ARC buffer type!");
2145 return ((uint32_t)-1);
2149 arc_buf_thaw(arc_buf_t *buf)
2151 arc_buf_hdr_t *hdr = buf->b_hdr;
2153 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2154 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2156 arc_cksum_verify(buf);
2159 * Compressed buffers do not manipulate the b_freeze_cksum or
2160 * allocate b_thawed.
2162 if (ARC_BUF_COMPRESSED(buf)) {
2163 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2164 arc_hdr_has_uncompressed_buf(hdr));
2168 ASSERT(HDR_HAS_L1HDR(hdr));
2169 arc_cksum_free(hdr);
2171 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
2173 if (zfs_flags & ZFS_DEBUG_MODIFY) {
2174 if (hdr->b_l1hdr.b_thawed != NULL)
2175 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2176 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
2180 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2183 arc_buf_unwatch(buf);
2188 arc_buf_freeze(arc_buf_t *buf)
2190 arc_buf_hdr_t *hdr = buf->b_hdr;
2191 kmutex_t *hash_lock;
2193 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2196 if (ARC_BUF_COMPRESSED(buf)) {
2197 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2198 arc_hdr_has_uncompressed_buf(hdr));
2202 hash_lock = HDR_LOCK(hdr);
2203 mutex_enter(hash_lock);
2205 ASSERT(HDR_HAS_L1HDR(hdr));
2206 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
2207 hdr->b_l1hdr.b_state == arc_anon);
2208 arc_cksum_compute(buf);
2209 mutex_exit(hash_lock);
2213 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
2214 * the following functions should be used to ensure that the flags are
2215 * updated in a thread-safe way. When manipulating the flags either
2216 * the hash_lock must be held or the hdr must be undiscoverable. This
2217 * ensures that we're not racing with any other threads when updating
2221 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2223 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2224 hdr->b_flags |= flags;
2228 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2230 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2231 hdr->b_flags &= ~flags;
2235 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
2236 * done in a special way since we have to clear and set bits
2237 * at the same time. Consumers that wish to set the compression bits
2238 * must use this function to ensure that the flags are updated in
2239 * thread-safe manner.
2242 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
2244 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2247 * Holes and embedded blocks will always have a psize = 0 so
2248 * we ignore the compression of the blkptr and set the
2249 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
2250 * Holes and embedded blocks remain anonymous so we don't
2251 * want to uncompress them. Mark them as uncompressed.
2253 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
2254 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2255 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
2256 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
2257 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2259 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2260 HDR_SET_COMPRESS(hdr, cmp);
2261 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
2262 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
2267 * Looks for another buf on the same hdr which has the data decompressed, copies
2268 * from it, and returns true. If no such buf exists, returns false.
2271 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
2273 arc_buf_hdr_t *hdr = buf->b_hdr;
2274 boolean_t copied = B_FALSE;
2276 ASSERT(HDR_HAS_L1HDR(hdr));
2277 ASSERT3P(buf->b_data, !=, NULL);
2278 ASSERT(!ARC_BUF_COMPRESSED(buf));
2280 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
2281 from = from->b_next) {
2282 /* can't use our own data buffer */
2287 if (!ARC_BUF_COMPRESSED(from)) {
2288 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
2295 * There were no decompressed bufs, so there should not be a
2296 * checksum on the hdr either.
2298 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
2304 * Given a buf that has a data buffer attached to it, this function will
2305 * efficiently fill the buf with data of the specified compression setting from
2306 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2307 * are already sharing a data buf, no copy is performed.
2309 * If the buf is marked as compressed but uncompressed data was requested, this
2310 * will allocate a new data buffer for the buf, remove that flag, and fill the
2311 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2312 * uncompressed data, and (since we haven't added support for it yet) if you
2313 * want compressed data your buf must already be marked as compressed and have
2314 * the correct-sized data buffer.
2317 arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
2319 arc_buf_hdr_t *hdr = buf->b_hdr;
2320 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
2321 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2323 ASSERT3P(buf->b_data, !=, NULL);
2324 IMPLY(compressed, hdr_compressed);
2325 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2327 if (hdr_compressed == compressed) {
2328 if (!arc_buf_is_shared(buf)) {
2329 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2333 ASSERT(hdr_compressed);
2334 ASSERT(!compressed);
2335 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2338 * If the buf is sharing its data with the hdr, unlink it and
2339 * allocate a new data buffer for the buf.
2341 if (arc_buf_is_shared(buf)) {
2342 ASSERT(ARC_BUF_COMPRESSED(buf));
2344 /* We need to give the buf it's own b_data */
2345 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2347 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2348 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2350 /* Previously overhead was 0; just add new overhead */
2351 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2352 } else if (ARC_BUF_COMPRESSED(buf)) {
2353 /* We need to reallocate the buf's b_data */
2354 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2357 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2359 /* We increased the size of b_data; update overhead */
2360 ARCSTAT_INCR(arcstat_overhead_size,
2361 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2365 * Regardless of the buf's previous compression settings, it
2366 * should not be compressed at the end of this function.
2368 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2371 * Try copying the data from another buf which already has a
2372 * decompressed version. If that's not possible, it's time to
2373 * bite the bullet and decompress the data from the hdr.
2375 if (arc_buf_try_copy_decompressed_data(buf)) {
2376 /* Skip byteswapping and checksumming (already done) */
2377 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2380 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2381 hdr->b_l1hdr.b_pabd, buf->b_data,
2382 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2385 * Absent hardware errors or software bugs, this should
2386 * be impossible, but log it anyway so we can debug it.
2390 "hdr %p, compress %d, psize %d, lsize %d",
2391 hdr, HDR_GET_COMPRESS(hdr),
2392 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2393 return (SET_ERROR(EIO));
2398 /* Byteswap the buf's data if necessary */
2399 if (bswap != DMU_BSWAP_NUMFUNCS) {
2400 ASSERT(!HDR_SHARED_DATA(hdr));
2401 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2402 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2405 /* Compute the hdr's checksum if necessary */
2406 arc_cksum_compute(buf);
2412 arc_decompress(arc_buf_t *buf)
2414 return (arc_buf_fill(buf, B_FALSE));
2418 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
2421 arc_hdr_size(arc_buf_hdr_t *hdr)
2425 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2426 HDR_GET_PSIZE(hdr) > 0) {
2427 size = HDR_GET_PSIZE(hdr);
2429 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2430 size = HDR_GET_LSIZE(hdr);
2436 * Increment the amount of evictable space in the arc_state_t's refcount.
2437 * We account for the space used by the hdr and the arc buf individually
2438 * so that we can add and remove them from the refcount individually.
2441 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2443 arc_buf_contents_t type = arc_buf_type(hdr);
2445 ASSERT(HDR_HAS_L1HDR(hdr));
2447 if (GHOST_STATE(state)) {
2448 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2449 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2450 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2451 (void) refcount_add_many(&state->arcs_esize[type],
2452 HDR_GET_LSIZE(hdr), hdr);
2456 ASSERT(!GHOST_STATE(state));
2457 if (hdr->b_l1hdr.b_pabd != NULL) {
2458 (void) refcount_add_many(&state->arcs_esize[type],
2459 arc_hdr_size(hdr), hdr);
2461 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2462 buf = buf->b_next) {
2463 if (arc_buf_is_shared(buf))
2465 (void) refcount_add_many(&state->arcs_esize[type],
2466 arc_buf_size(buf), buf);
2471 * Decrement the amount of evictable space in the arc_state_t's refcount.
2472 * We account for the space used by the hdr and the arc buf individually
2473 * so that we can add and remove them from the refcount individually.
2476 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2478 arc_buf_contents_t type = arc_buf_type(hdr);
2480 ASSERT(HDR_HAS_L1HDR(hdr));
2482 if (GHOST_STATE(state)) {
2483 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2484 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2485 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2486 (void) refcount_remove_many(&state->arcs_esize[type],
2487 HDR_GET_LSIZE(hdr), hdr);
2491 ASSERT(!GHOST_STATE(state));
2492 if (hdr->b_l1hdr.b_pabd != NULL) {
2493 (void) refcount_remove_many(&state->arcs_esize[type],
2494 arc_hdr_size(hdr), hdr);
2496 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2497 buf = buf->b_next) {
2498 if (arc_buf_is_shared(buf))
2500 (void) refcount_remove_many(&state->arcs_esize[type],
2501 arc_buf_size(buf), buf);
2506 * Add a reference to this hdr indicating that someone is actively
2507 * referencing that memory. When the refcount transitions from 0 to 1,
2508 * we remove it from the respective arc_state_t list to indicate that
2509 * it is not evictable.
2512 add_reference(arc_buf_hdr_t *hdr, void *tag)
2514 ASSERT(HDR_HAS_L1HDR(hdr));
2515 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2516 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2517 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2518 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2521 arc_state_t *state = hdr->b_l1hdr.b_state;
2523 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2524 (state != arc_anon)) {
2525 /* We don't use the L2-only state list. */
2526 if (state != arc_l2c_only) {
2527 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2529 arc_evictable_space_decrement(hdr, state);
2531 /* remove the prefetch flag if we get a reference */
2532 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2537 * Remove a reference from this hdr. When the reference transitions from
2538 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2539 * list making it eligible for eviction.
2542 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2545 arc_state_t *state = hdr->b_l1hdr.b_state;
2547 ASSERT(HDR_HAS_L1HDR(hdr));
2548 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2549 ASSERT(!GHOST_STATE(state));
2552 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2553 * check to prevent usage of the arc_l2c_only list.
2555 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2556 (state != arc_anon)) {
2557 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2558 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2559 arc_evictable_space_increment(hdr, state);
2565 * Returns detailed information about a specific arc buffer. When the
2566 * state_index argument is set the function will calculate the arc header
2567 * list position for its arc state. Since this requires a linear traversal
2568 * callers are strongly encourage not to do this. However, it can be helpful
2569 * for targeted analysis so the functionality is provided.
2572 arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
2574 arc_buf_hdr_t *hdr = ab->b_hdr;
2575 l1arc_buf_hdr_t *l1hdr = NULL;
2576 l2arc_buf_hdr_t *l2hdr = NULL;
2577 arc_state_t *state = NULL;
2579 memset(abi, 0, sizeof (arc_buf_info_t));
2584 abi->abi_flags = hdr->b_flags;
2586 if (HDR_HAS_L1HDR(hdr)) {
2587 l1hdr = &hdr->b_l1hdr;
2588 state = l1hdr->b_state;
2590 if (HDR_HAS_L2HDR(hdr))
2591 l2hdr = &hdr->b_l2hdr;
2594 abi->abi_bufcnt = l1hdr->b_bufcnt;
2595 abi->abi_access = l1hdr->b_arc_access;
2596 abi->abi_mru_hits = l1hdr->b_mru_hits;
2597 abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits;
2598 abi->abi_mfu_hits = l1hdr->b_mfu_hits;
2599 abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits;
2600 abi->abi_holds = refcount_count(&l1hdr->b_refcnt);
2604 abi->abi_l2arc_dattr = l2hdr->b_daddr;
2605 abi->abi_l2arc_hits = l2hdr->b_hits;
2608 abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
2609 abi->abi_state_contents = arc_buf_type(hdr);
2610 abi->abi_size = arc_hdr_size(hdr);
2614 * Move the supplied buffer to the indicated state. The hash lock
2615 * for the buffer must be held by the caller.
2618 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2619 kmutex_t *hash_lock)
2621 arc_state_t *old_state;
2624 boolean_t update_old, update_new;
2625 arc_buf_contents_t buftype = arc_buf_type(hdr);
2628 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2629 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2630 * L1 hdr doesn't always exist when we change state to arc_anon before
2631 * destroying a header, in which case reallocating to add the L1 hdr is
2634 if (HDR_HAS_L1HDR(hdr)) {
2635 old_state = hdr->b_l1hdr.b_state;
2636 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2637 bufcnt = hdr->b_l1hdr.b_bufcnt;
2638 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL);
2640 old_state = arc_l2c_only;
2643 update_old = B_FALSE;
2645 update_new = update_old;
2647 ASSERT(MUTEX_HELD(hash_lock));
2648 ASSERT3P(new_state, !=, old_state);
2649 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2650 ASSERT(old_state != arc_anon || bufcnt <= 1);
2653 * If this buffer is evictable, transfer it from the
2654 * old state list to the new state list.
2657 if (old_state != arc_anon && old_state != arc_l2c_only) {
2658 ASSERT(HDR_HAS_L1HDR(hdr));
2659 multilist_remove(old_state->arcs_list[buftype], hdr);
2661 if (GHOST_STATE(old_state)) {
2663 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2664 update_old = B_TRUE;
2666 arc_evictable_space_decrement(hdr, old_state);
2668 if (new_state != arc_anon && new_state != arc_l2c_only) {
2671 * An L1 header always exists here, since if we're
2672 * moving to some L1-cached state (i.e. not l2c_only or
2673 * anonymous), we realloc the header to add an L1hdr
2676 ASSERT(HDR_HAS_L1HDR(hdr));
2677 multilist_insert(new_state->arcs_list[buftype], hdr);
2679 if (GHOST_STATE(new_state)) {
2681 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2682 update_new = B_TRUE;
2684 arc_evictable_space_increment(hdr, new_state);
2688 ASSERT(!HDR_EMPTY(hdr));
2689 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2690 buf_hash_remove(hdr);
2692 /* adjust state sizes (ignore arc_l2c_only) */
2694 if (update_new && new_state != arc_l2c_only) {
2695 ASSERT(HDR_HAS_L1HDR(hdr));
2696 if (GHOST_STATE(new_state)) {
2700 * When moving a header to a ghost state, we first
2701 * remove all arc buffers. Thus, we'll have a
2702 * bufcnt of zero, and no arc buffer to use for
2703 * the reference. As a result, we use the arc
2704 * header pointer for the reference.
2706 (void) refcount_add_many(&new_state->arcs_size,
2707 HDR_GET_LSIZE(hdr), hdr);
2708 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2710 uint32_t buffers = 0;
2713 * Each individual buffer holds a unique reference,
2714 * thus we must remove each of these references one
2717 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2718 buf = buf->b_next) {
2719 ASSERT3U(bufcnt, !=, 0);
2723 * When the arc_buf_t is sharing the data
2724 * block with the hdr, the owner of the
2725 * reference belongs to the hdr. Only
2726 * add to the refcount if the arc_buf_t is
2729 if (arc_buf_is_shared(buf))
2732 (void) refcount_add_many(&new_state->arcs_size,
2733 arc_buf_size(buf), buf);
2735 ASSERT3U(bufcnt, ==, buffers);
2737 if (hdr->b_l1hdr.b_pabd != NULL) {
2738 (void) refcount_add_many(&new_state->arcs_size,
2739 arc_hdr_size(hdr), hdr);
2741 ASSERT(GHOST_STATE(old_state));
2746 if (update_old && old_state != arc_l2c_only) {
2747 ASSERT(HDR_HAS_L1HDR(hdr));
2748 if (GHOST_STATE(old_state)) {
2750 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2753 * When moving a header off of a ghost state,
2754 * the header will not contain any arc buffers.
2755 * We use the arc header pointer for the reference
2756 * which is exactly what we did when we put the
2757 * header on the ghost state.
2760 (void) refcount_remove_many(&old_state->arcs_size,
2761 HDR_GET_LSIZE(hdr), hdr);
2763 uint32_t buffers = 0;
2766 * Each individual buffer holds a unique reference,
2767 * thus we must remove each of these references one
2770 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2771 buf = buf->b_next) {
2772 ASSERT3U(bufcnt, !=, 0);
2776 * When the arc_buf_t is sharing the data
2777 * block with the hdr, the owner of the
2778 * reference belongs to the hdr. Only
2779 * add to the refcount if the arc_buf_t is
2782 if (arc_buf_is_shared(buf))
2785 (void) refcount_remove_many(
2786 &old_state->arcs_size, arc_buf_size(buf),
2789 ASSERT3U(bufcnt, ==, buffers);
2790 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2791 (void) refcount_remove_many(
2792 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2796 if (HDR_HAS_L1HDR(hdr))
2797 hdr->b_l1hdr.b_state = new_state;
2800 * L2 headers should never be on the L2 state list since they don't
2801 * have L1 headers allocated.
2803 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2804 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2808 arc_space_consume(uint64_t space, arc_space_type_t type)
2810 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2813 case ARC_SPACE_DATA:
2814 aggsum_add(&astat_data_size, space);
2816 case ARC_SPACE_META:
2817 aggsum_add(&astat_metadata_size, space);
2819 case ARC_SPACE_BONUS:
2820 aggsum_add(&astat_bonus_size, space);
2822 case ARC_SPACE_DNODE:
2823 aggsum_add(&astat_dnode_size, space);
2825 case ARC_SPACE_DBUF:
2826 aggsum_add(&astat_dbuf_size, space);
2828 case ARC_SPACE_HDRS:
2829 aggsum_add(&astat_hdr_size, space);
2831 case ARC_SPACE_L2HDRS:
2832 aggsum_add(&astat_l2_hdr_size, space);
2836 if (type != ARC_SPACE_DATA)
2837 aggsum_add(&arc_meta_used, space);
2839 aggsum_add(&arc_size, space);
2843 arc_space_return(uint64_t space, arc_space_type_t type)
2845 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2848 case ARC_SPACE_DATA:
2849 aggsum_add(&astat_data_size, -space);
2851 case ARC_SPACE_META:
2852 aggsum_add(&astat_metadata_size, -space);
2854 case ARC_SPACE_BONUS:
2855 aggsum_add(&astat_bonus_size, -space);
2857 case ARC_SPACE_DNODE:
2858 aggsum_add(&astat_dnode_size, -space);
2860 case ARC_SPACE_DBUF:
2861 aggsum_add(&astat_dbuf_size, -space);
2863 case ARC_SPACE_HDRS:
2864 aggsum_add(&astat_hdr_size, -space);
2866 case ARC_SPACE_L2HDRS:
2867 aggsum_add(&astat_l2_hdr_size, -space);
2871 if (type != ARC_SPACE_DATA) {
2872 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2874 * We use the upper bound here rather than the precise value
2875 * because the arc_meta_max value doesn't need to be
2876 * precise. It's only consumed by humans via arcstats.
2878 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2879 arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2880 aggsum_add(&arc_meta_used, -space);
2883 ASSERT(aggsum_compare(&arc_size, space) >= 0);
2884 aggsum_add(&arc_size, -space);
2888 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2889 * with the hdr's b_pabd.
2892 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2895 * The criteria for sharing a hdr's data are:
2896 * 1. the hdr's compression matches the buf's compression
2897 * 2. the hdr doesn't need to be byteswapped
2898 * 3. the hdr isn't already being shared
2899 * 4. the buf is either compressed or it is the last buf in the hdr list
2901 * Criterion #4 maintains the invariant that shared uncompressed
2902 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2903 * might ask, "if a compressed buf is allocated first, won't that be the
2904 * last thing in the list?", but in that case it's impossible to create
2905 * a shared uncompressed buf anyway (because the hdr must be compressed
2906 * to have the compressed buf). You might also think that #3 is
2907 * sufficient to make this guarantee, however it's possible
2908 * (specifically in the rare L2ARC write race mentioned in
2909 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2910 * is sharable, but wasn't at the time of its allocation. Rather than
2911 * allow a new shared uncompressed buf to be created and then shuffle
2912 * the list around to make it the last element, this simply disallows
2913 * sharing if the new buf isn't the first to be added.
2915 ASSERT3P(buf->b_hdr, ==, hdr);
2916 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF;
2917 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2918 return (buf_compressed == hdr_compressed &&
2919 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2920 !HDR_SHARED_DATA(hdr) &&
2921 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2925 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2926 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2927 * copy was made successfully, or an error code otherwise.
2930 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2931 boolean_t fill, arc_buf_t **ret)
2935 ASSERT(HDR_HAS_L1HDR(hdr));
2936 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2937 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2938 hdr->b_type == ARC_BUFC_METADATA);
2939 ASSERT3P(ret, !=, NULL);
2940 ASSERT3P(*ret, ==, NULL);
2942 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2945 buf->b_next = hdr->b_l1hdr.b_buf;
2948 add_reference(hdr, tag);
2951 * We're about to change the hdr's b_flags. We must either
2952 * hold the hash_lock or be undiscoverable.
2954 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2957 * Only honor requests for compressed bufs if the hdr is actually
2960 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
2961 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2964 * If the hdr's data can be shared then we share the data buffer and
2965 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2966 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2967 * buffer to store the buf's data.
2969 * There are two additional restrictions here because we're sharing
2970 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2971 * actively involved in an L2ARC write, because if this buf is used by
2972 * an arc_write() then the hdr's data buffer will be released when the
2973 * write completes, even though the L2ARC write might still be using it.
2974 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2975 * need to be ABD-aware.
2977 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2978 abd_is_linear(hdr->b_l1hdr.b_pabd);
2980 /* Set up b_data and sharing */
2982 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2983 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2984 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2987 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2988 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2990 VERIFY3P(buf->b_data, !=, NULL);
2992 hdr->b_l1hdr.b_buf = buf;
2993 hdr->b_l1hdr.b_bufcnt += 1;
2996 * If the user wants the data from the hdr, we need to either copy or
2997 * decompress the data.
3000 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
3006 static char *arc_onloan_tag = "onloan";
3009 arc_loaned_bytes_update(int64_t delta)
3011 atomic_add_64(&arc_loaned_bytes, delta);
3013 /* assert that it did not wrap around */
3014 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
3018 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
3019 * flight data by arc_tempreserve_space() until they are "returned". Loaned
3020 * buffers must be returned to the arc before they can be used by the DMU or
3024 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
3026 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
3027 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
3029 arc_loaned_bytes_update(arc_buf_size(buf));
3035 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
3036 enum zio_compress compression_type)
3038 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
3039 psize, lsize, compression_type);
3041 arc_loaned_bytes_update(arc_buf_size(buf));
3048 * Return a loaned arc buffer to the arc.
3051 arc_return_buf(arc_buf_t *buf, void *tag)
3053 arc_buf_hdr_t *hdr = buf->b_hdr;
3055 ASSERT3P(buf->b_data, !=, NULL);
3056 ASSERT(HDR_HAS_L1HDR(hdr));
3057 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
3058 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
3060 arc_loaned_bytes_update(-arc_buf_size(buf));
3063 /* Detach an arc_buf from a dbuf (tag) */
3065 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
3067 arc_buf_hdr_t *hdr = buf->b_hdr;
3069 ASSERT3P(buf->b_data, !=, NULL);
3070 ASSERT(HDR_HAS_L1HDR(hdr));
3071 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
3072 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
3074 arc_loaned_bytes_update(arc_buf_size(buf));
3078 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
3080 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
3083 df->l2df_size = size;
3084 df->l2df_type = type;
3085 mutex_enter(&l2arc_free_on_write_mtx);
3086 list_insert_head(l2arc_free_on_write, df);
3087 mutex_exit(&l2arc_free_on_write_mtx);
3091 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
3093 arc_state_t *state = hdr->b_l1hdr.b_state;
3094 arc_buf_contents_t type = arc_buf_type(hdr);
3095 uint64_t size = arc_hdr_size(hdr);
3097 /* protected by hash lock, if in the hash table */
3098 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3099 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3100 ASSERT(state != arc_anon && state != arc_l2c_only);
3102 (void) refcount_remove_many(&state->arcs_esize[type],
3105 (void) refcount_remove_many(&state->arcs_size, size, hdr);
3106 if (type == ARC_BUFC_METADATA) {
3107 arc_space_return(size, ARC_SPACE_META);
3109 ASSERT(type == ARC_BUFC_DATA);
3110 arc_space_return(size, ARC_SPACE_DATA);
3113 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
3117 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
3118 * data buffer, we transfer the refcount ownership to the hdr and update
3119 * the appropriate kstats.
3122 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3124 arc_state_t *state = hdr->b_l1hdr.b_state;
3126 ASSERT(arc_can_share(hdr, buf));
3127 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3128 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3131 * Start sharing the data buffer. We transfer the
3132 * refcount ownership to the hdr since it always owns
3133 * the refcount whenever an arc_buf_t is shared.
3135 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
3136 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
3137 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
3138 HDR_ISTYPE_METADATA(hdr));
3139 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
3140 buf->b_flags |= ARC_BUF_FLAG_SHARED;
3143 * Since we've transferred ownership to the hdr we need
3144 * to increment its compressed and uncompressed kstats and
3145 * decrement the overhead size.
3147 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3148 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3149 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
3153 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3155 arc_state_t *state = hdr->b_l1hdr.b_state;
3157 ASSERT(arc_buf_is_shared(buf));
3158 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3159 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3162 * We are no longer sharing this buffer so we need
3163 * to transfer its ownership to the rightful owner.
3165 refcount_transfer_ownership(&state->arcs_size, hdr, buf);
3166 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3167 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3168 abd_put(hdr->b_l1hdr.b_pabd);
3169 hdr->b_l1hdr.b_pabd = NULL;
3170 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3173 * Since the buffer is no longer shared between
3174 * the arc buf and the hdr, count it as overhead.
3176 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3177 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3178 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3182 * Remove an arc_buf_t from the hdr's buf list and return the last
3183 * arc_buf_t on the list. If no buffers remain on the list then return
3187 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3189 ASSERT(HDR_HAS_L1HDR(hdr));
3190 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3192 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3193 arc_buf_t *lastbuf = NULL;
3196 * Remove the buf from the hdr list and locate the last
3197 * remaining buffer on the list.
3199 while (*bufp != NULL) {
3201 *bufp = buf->b_next;
3204 * If we've removed a buffer in the middle of
3205 * the list then update the lastbuf and update
3208 if (*bufp != NULL) {
3210 bufp = &(*bufp)->b_next;
3214 ASSERT3P(lastbuf, !=, buf);
3215 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3216 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3217 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3223 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3227 arc_buf_destroy_impl(arc_buf_t *buf)
3229 arc_buf_hdr_t *hdr = buf->b_hdr;
3232 * Free up the data associated with the buf but only if we're not
3233 * sharing this with the hdr. If we are sharing it with the hdr, the
3234 * hdr is responsible for doing the free.
3236 if (buf->b_data != NULL) {
3238 * We're about to change the hdr's b_flags. We must either
3239 * hold the hash_lock or be undiscoverable.
3241 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3243 arc_cksum_verify(buf);
3245 arc_buf_unwatch(buf);
3248 if (arc_buf_is_shared(buf)) {
3249 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3251 uint64_t size = arc_buf_size(buf);
3252 arc_free_data_buf(hdr, buf->b_data, size, buf);
3253 ARCSTAT_INCR(arcstat_overhead_size, -size);
3257 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3258 hdr->b_l1hdr.b_bufcnt -= 1;
3261 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3263 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3265 * If the current arc_buf_t is sharing its data buffer with the
3266 * hdr, then reassign the hdr's b_pabd to share it with the new
3267 * buffer at the end of the list. The shared buffer is always
3268 * the last one on the hdr's buffer list.
3270 * There is an equivalent case for compressed bufs, but since
3271 * they aren't guaranteed to be the last buf in the list and
3272 * that is an exceedingly rare case, we just allow that space be
3273 * wasted temporarily.
3275 if (lastbuf != NULL) {
3276 /* Only one buf can be shared at once */
3277 VERIFY(!arc_buf_is_shared(lastbuf));
3278 /* hdr is uncompressed so can't have compressed buf */
3279 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3281 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3282 arc_hdr_free_pabd(hdr);
3285 * We must setup a new shared block between the
3286 * last buffer and the hdr. The data would have
3287 * been allocated by the arc buf so we need to transfer
3288 * ownership to the hdr since it's now being shared.
3290 arc_share_buf(hdr, lastbuf);
3292 } else if (HDR_SHARED_DATA(hdr)) {
3294 * Uncompressed shared buffers are always at the end
3295 * of the list. Compressed buffers don't have the
3296 * same requirements. This makes it hard to
3297 * simply assert that the lastbuf is shared so
3298 * we rely on the hdr's compression flags to determine
3299 * if we have a compressed, shared buffer.
3301 ASSERT3P(lastbuf, !=, NULL);
3302 ASSERT(arc_buf_is_shared(lastbuf) ||
3303 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
3307 * Free the checksum if we're removing the last uncompressed buf from
3310 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3311 arc_cksum_free(hdr);
3314 /* clean up the buf */
3316 kmem_cache_free(buf_cache, buf);
3320 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr)
3322 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3323 ASSERT(HDR_HAS_L1HDR(hdr));
3324 ASSERT(!HDR_SHARED_DATA(hdr));
3326 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3327 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
3328 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3329 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3331 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3332 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3336 arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
3338 ASSERT(HDR_HAS_L1HDR(hdr));
3339 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3342 * If the hdr is currently being written to the l2arc then
3343 * we defer freeing the data by adding it to the l2arc_free_on_write
3344 * list. The l2arc will free the data once it's finished
3345 * writing it to the l2arc device.
3347 if (HDR_L2_WRITING(hdr)) {
3348 arc_hdr_free_on_write(hdr);
3349 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3351 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
3352 arc_hdr_size(hdr), hdr);
3354 hdr->b_l1hdr.b_pabd = NULL;
3355 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3357 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3358 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3361 static arc_buf_hdr_t *
3362 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3363 enum zio_compress compression_type, arc_buf_contents_t type)
3367 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3369 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3370 ASSERT(HDR_EMPTY(hdr));
3371 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3372 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
3373 HDR_SET_PSIZE(hdr, psize);
3374 HDR_SET_LSIZE(hdr, lsize);
3378 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3379 arc_hdr_set_compress(hdr, compression_type);
3381 hdr->b_l1hdr.b_state = arc_anon;
3382 hdr->b_l1hdr.b_arc_access = 0;
3383 hdr->b_l1hdr.b_bufcnt = 0;
3384 hdr->b_l1hdr.b_buf = NULL;
3387 * Allocate the hdr's buffer. This will contain either
3388 * the compressed or uncompressed data depending on the block
3389 * it references and compressed arc enablement.
3391 arc_hdr_alloc_pabd(hdr);
3392 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3398 * Transition between the two allocation states for the arc_buf_hdr struct.
3399 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3400 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3401 * version is used when a cache buffer is only in the L2ARC in order to reduce
3404 static arc_buf_hdr_t *
3405 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3407 ASSERT(HDR_HAS_L2HDR(hdr));
3409 arc_buf_hdr_t *nhdr;
3410 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3412 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3413 (old == hdr_l2only_cache && new == hdr_full_cache));
3415 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3417 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3418 buf_hash_remove(hdr);
3420 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3422 if (new == hdr_full_cache) {
3423 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3425 * arc_access and arc_change_state need to be aware that a
3426 * header has just come out of L2ARC, so we set its state to
3427 * l2c_only even though it's about to change.
3429 nhdr->b_l1hdr.b_state = arc_l2c_only;
3431 /* Verify previous threads set to NULL before freeing */
3432 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3434 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3435 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3436 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3439 * If we've reached here, We must have been called from
3440 * arc_evict_hdr(), as such we should have already been
3441 * removed from any ghost list we were previously on
3442 * (which protects us from racing with arc_evict_state),
3443 * thus no locking is needed during this check.
3445 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3448 * A buffer must not be moved into the arc_l2c_only
3449 * state if it's not finished being written out to the
3450 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3451 * might try to be accessed, even though it was removed.
3453 VERIFY(!HDR_L2_WRITING(hdr));
3454 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3457 if (hdr->b_l1hdr.b_thawed != NULL) {
3458 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3459 hdr->b_l1hdr.b_thawed = NULL;
3463 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3466 * The header has been reallocated so we need to re-insert it into any
3469 (void) buf_hash_insert(nhdr, NULL);
3471 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3473 mutex_enter(&dev->l2ad_mtx);
3476 * We must place the realloc'ed header back into the list at
3477 * the same spot. Otherwise, if it's placed earlier in the list,
3478 * l2arc_write_buffers() could find it during the function's
3479 * write phase, and try to write it out to the l2arc.
3481 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3482 list_remove(&dev->l2ad_buflist, hdr);
3484 mutex_exit(&dev->l2ad_mtx);
3487 * Since we're using the pointer address as the tag when
3488 * incrementing and decrementing the l2ad_alloc refcount, we
3489 * must remove the old pointer (that we're about to destroy) and
3490 * add the new pointer to the refcount. Otherwise we'd remove
3491 * the wrong pointer address when calling arc_hdr_destroy() later.
3494 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
3495 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
3497 buf_discard_identity(hdr);
3498 kmem_cache_free(old, hdr);
3504 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3505 * The buf is returned thawed since we expect the consumer to modify it.
3508 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3510 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3511 ZIO_COMPRESS_OFF, type);
3512 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3514 arc_buf_t *buf = NULL;
3515 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf));
3522 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3523 * for bufs containing metadata.
3526 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3527 enum zio_compress compression_type)
3529 ASSERT3U(lsize, >, 0);
3530 ASSERT3U(lsize, >=, psize);
3531 ASSERT(compression_type > ZIO_COMPRESS_OFF);
3532 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3534 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3535 compression_type, ARC_BUFC_DATA);
3536 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3538 arc_buf_t *buf = NULL;
3539 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3541 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3543 if (!arc_buf_is_shared(buf)) {
3545 * To ensure that the hdr has the correct data in it if we call
3546 * arc_decompress() on this buf before it's been written to
3547 * disk, it's easiest if we just set up sharing between the
3550 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3551 arc_hdr_free_pabd(hdr);
3552 arc_share_buf(hdr, buf);
3559 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3561 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3562 l2arc_dev_t *dev = l2hdr->b_dev;
3563 uint64_t psize = arc_hdr_size(hdr);
3565 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3566 ASSERT(HDR_HAS_L2HDR(hdr));
3568 list_remove(&dev->l2ad_buflist, hdr);
3570 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3571 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3573 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3575 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3576 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3580 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3582 if (HDR_HAS_L1HDR(hdr)) {
3583 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3584 hdr->b_l1hdr.b_bufcnt > 0);
3585 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3586 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3588 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3589 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3591 if (!HDR_EMPTY(hdr))
3592 buf_discard_identity(hdr);
3594 if (HDR_HAS_L2HDR(hdr)) {
3595 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3596 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3599 mutex_enter(&dev->l2ad_mtx);
3602 * Even though we checked this conditional above, we
3603 * need to check this again now that we have the
3604 * l2ad_mtx. This is because we could be racing with
3605 * another thread calling l2arc_evict() which might have
3606 * destroyed this header's L2 portion as we were waiting
3607 * to acquire the l2ad_mtx. If that happens, we don't
3608 * want to re-destroy the header's L2 portion.
3610 if (HDR_HAS_L2HDR(hdr)) {
3612 arc_hdr_l2hdr_destroy(hdr);
3616 mutex_exit(&dev->l2ad_mtx);
3619 if (HDR_HAS_L1HDR(hdr)) {
3620 arc_cksum_free(hdr);
3622 while (hdr->b_l1hdr.b_buf != NULL)
3623 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3626 if (hdr->b_l1hdr.b_thawed != NULL) {
3627 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3628 hdr->b_l1hdr.b_thawed = NULL;
3632 if (hdr->b_l1hdr.b_pabd != NULL) {
3633 arc_hdr_free_pabd(hdr);
3637 ASSERT3P(hdr->b_hash_next, ==, NULL);
3638 if (HDR_HAS_L1HDR(hdr)) {
3639 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3640 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3641 kmem_cache_free(hdr_full_cache, hdr);
3643 kmem_cache_free(hdr_l2only_cache, hdr);
3648 arc_buf_destroy(arc_buf_t *buf, void* tag)
3650 arc_buf_hdr_t *hdr = buf->b_hdr;
3651 kmutex_t *hash_lock = HDR_LOCK(hdr);
3653 if (hdr->b_l1hdr.b_state == arc_anon) {
3654 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3655 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3656 VERIFY0(remove_reference(hdr, NULL, tag));
3657 arc_hdr_destroy(hdr);
3661 mutex_enter(hash_lock);
3662 ASSERT3P(hdr, ==, buf->b_hdr);
3663 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3664 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3665 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3666 ASSERT3P(buf->b_data, !=, NULL);
3668 (void) remove_reference(hdr, hash_lock, tag);
3669 arc_buf_destroy_impl(buf);
3670 mutex_exit(hash_lock);
3674 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3675 * state of the header is dependent on its state prior to entering this
3676 * function. The following transitions are possible:
3678 * - arc_mru -> arc_mru_ghost
3679 * - arc_mfu -> arc_mfu_ghost
3680 * - arc_mru_ghost -> arc_l2c_only
3681 * - arc_mru_ghost -> deleted
3682 * - arc_mfu_ghost -> arc_l2c_only
3683 * - arc_mfu_ghost -> deleted
3686 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3688 arc_state_t *evicted_state, *state;
3689 int64_t bytes_evicted = 0;
3690 int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3691 zfs_arc_min_prescient_prefetch_ms : zfs_arc_min_prefetch_ms;
3693 ASSERT(MUTEX_HELD(hash_lock));
3694 ASSERT(HDR_HAS_L1HDR(hdr));
3696 state = hdr->b_l1hdr.b_state;
3697 if (GHOST_STATE(state)) {
3698 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3699 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3702 * l2arc_write_buffers() relies on a header's L1 portion
3703 * (i.e. its b_pabd field) during it's write phase.
3704 * Thus, we cannot push a header onto the arc_l2c_only
3705 * state (removing it's L1 piece) until the header is
3706 * done being written to the l2arc.
3708 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3709 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3710 return (bytes_evicted);
3713 ARCSTAT_BUMP(arcstat_deleted);
3714 bytes_evicted += HDR_GET_LSIZE(hdr);
3716 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3718 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3719 if (HDR_HAS_L2HDR(hdr)) {
3721 * This buffer is cached on the 2nd Level ARC;
3722 * don't destroy the header.
3724 arc_change_state(arc_l2c_only, hdr, hash_lock);
3726 * dropping from L1+L2 cached to L2-only,
3727 * realloc to remove the L1 header.
3729 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3732 arc_change_state(arc_anon, hdr, hash_lock);
3733 arc_hdr_destroy(hdr);
3735 return (bytes_evicted);
3738 ASSERT(state == arc_mru || state == arc_mfu);
3739 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3741 /* prefetch buffers have a minimum lifespan */
3742 if (HDR_IO_IN_PROGRESS(hdr) ||
3743 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3744 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < min_lifetime * hz)) {
3745 ARCSTAT_BUMP(arcstat_evict_skip);
3746 return (bytes_evicted);
3749 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3750 while (hdr->b_l1hdr.b_buf) {
3751 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3752 if (!mutex_tryenter(&buf->b_evict_lock)) {
3753 ARCSTAT_BUMP(arcstat_mutex_miss);
3756 if (buf->b_data != NULL)
3757 bytes_evicted += HDR_GET_LSIZE(hdr);
3758 mutex_exit(&buf->b_evict_lock);
3759 arc_buf_destroy_impl(buf);
3762 if (HDR_HAS_L2HDR(hdr)) {
3763 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3765 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3766 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3767 HDR_GET_LSIZE(hdr));
3769 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3770 HDR_GET_LSIZE(hdr));
3774 if (hdr->b_l1hdr.b_bufcnt == 0) {
3775 arc_cksum_free(hdr);
3777 bytes_evicted += arc_hdr_size(hdr);
3780 * If this hdr is being evicted and has a compressed
3781 * buffer then we discard it here before we change states.
3782 * This ensures that the accounting is updated correctly
3783 * in arc_free_data_impl().
3785 arc_hdr_free_pabd(hdr);
3787 arc_change_state(evicted_state, hdr, hash_lock);
3788 ASSERT(HDR_IN_HASH_TABLE(hdr));
3789 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3790 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3793 return (bytes_evicted);
3797 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3798 uint64_t spa, int64_t bytes)
3800 multilist_sublist_t *mls;
3801 uint64_t bytes_evicted = 0;
3803 kmutex_t *hash_lock;
3804 int evict_count = 0;
3806 ASSERT3P(marker, !=, NULL);
3807 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3809 mls = multilist_sublist_lock(ml, idx);
3811 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3812 hdr = multilist_sublist_prev(mls, marker)) {
3813 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3814 (evict_count >= zfs_arc_evict_batch_limit))
3818 * To keep our iteration location, move the marker
3819 * forward. Since we're not holding hdr's hash lock, we
3820 * must be very careful and not remove 'hdr' from the
3821 * sublist. Otherwise, other consumers might mistake the
3822 * 'hdr' as not being on a sublist when they call the
3823 * multilist_link_active() function (they all rely on
3824 * the hash lock protecting concurrent insertions and
3825 * removals). multilist_sublist_move_forward() was
3826 * specifically implemented to ensure this is the case
3827 * (only 'marker' will be removed and re-inserted).
3829 multilist_sublist_move_forward(mls, marker);
3832 * The only case where the b_spa field should ever be
3833 * zero, is the marker headers inserted by
3834 * arc_evict_state(). It's possible for multiple threads
3835 * to be calling arc_evict_state() concurrently (e.g.
3836 * dsl_pool_close() and zio_inject_fault()), so we must
3837 * skip any markers we see from these other threads.
3839 if (hdr->b_spa == 0)
3842 /* we're only interested in evicting buffers of a certain spa */
3843 if (spa != 0 && hdr->b_spa != spa) {
3844 ARCSTAT_BUMP(arcstat_evict_skip);
3848 hash_lock = HDR_LOCK(hdr);
3851 * We aren't calling this function from any code path
3852 * that would already be holding a hash lock, so we're
3853 * asserting on this assumption to be defensive in case
3854 * this ever changes. Without this check, it would be
3855 * possible to incorrectly increment arcstat_mutex_miss
3856 * below (e.g. if the code changed such that we called
3857 * this function with a hash lock held).
3859 ASSERT(!MUTEX_HELD(hash_lock));
3861 if (mutex_tryenter(hash_lock)) {
3862 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3863 mutex_exit(hash_lock);
3865 bytes_evicted += evicted;
3868 * If evicted is zero, arc_evict_hdr() must have
3869 * decided to skip this header, don't increment
3870 * evict_count in this case.
3876 * If arc_size isn't overflowing, signal any
3877 * threads that might happen to be waiting.
3879 * For each header evicted, we wake up a single
3880 * thread. If we used cv_broadcast, we could
3881 * wake up "too many" threads causing arc_size
3882 * to significantly overflow arc_c; since
3883 * arc_get_data_impl() doesn't check for overflow
3884 * when it's woken up (it doesn't because it's
3885 * possible for the ARC to be overflowing while
3886 * full of un-evictable buffers, and the
3887 * function should proceed in this case).
3889 * If threads are left sleeping, due to not
3890 * using cv_broadcast, they will be woken up
3891 * just before arc_reclaim_thread() sleeps.
3893 mutex_enter(&arc_reclaim_lock);
3894 if (!arc_is_overflowing())
3895 cv_signal(&arc_reclaim_waiters_cv);
3896 mutex_exit(&arc_reclaim_lock);
3898 ARCSTAT_BUMP(arcstat_mutex_miss);
3902 multilist_sublist_unlock(mls);
3904 return (bytes_evicted);
3908 * Evict buffers from the given arc state, until we've removed the
3909 * specified number of bytes. Move the removed buffers to the
3910 * appropriate evict state.
3912 * This function makes a "best effort". It skips over any buffers
3913 * it can't get a hash_lock on, and so, may not catch all candidates.
3914 * It may also return without evicting as much space as requested.
3916 * If bytes is specified using the special value ARC_EVICT_ALL, this
3917 * will evict all available (i.e. unlocked and evictable) buffers from
3918 * the given arc state; which is used by arc_flush().
3921 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3922 arc_buf_contents_t type)
3924 uint64_t total_evicted = 0;
3925 multilist_t *ml = state->arcs_list[type];
3927 arc_buf_hdr_t **markers;
3929 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3931 num_sublists = multilist_get_num_sublists(ml);
3934 * If we've tried to evict from each sublist, made some
3935 * progress, but still have not hit the target number of bytes
3936 * to evict, we want to keep trying. The markers allow us to
3937 * pick up where we left off for each individual sublist, rather
3938 * than starting from the tail each time.
3940 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3941 for (int i = 0; i < num_sublists; i++) {
3942 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3945 * A b_spa of 0 is used to indicate that this header is
3946 * a marker. This fact is used in arc_adjust_type() and
3947 * arc_evict_state_impl().
3949 markers[i]->b_spa = 0;
3951 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3952 multilist_sublist_insert_tail(mls, markers[i]);
3953 multilist_sublist_unlock(mls);
3957 * While we haven't hit our target number of bytes to evict, or
3958 * we're evicting all available buffers.
3960 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3961 int sublist_idx = multilist_get_random_index(ml);
3962 uint64_t scan_evicted = 0;
3965 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
3966 * Request that 10% of the LRUs be scanned by the superblock
3969 if (type == ARC_BUFC_DATA && aggsum_compare(&astat_dnode_size,
3970 arc_dnode_limit) > 0) {
3971 arc_prune_async((aggsum_upper_bound(&astat_dnode_size) -
3972 arc_dnode_limit) / sizeof (dnode_t) /
3973 zfs_arc_dnode_reduce_percent);
3977 * Start eviction using a randomly selected sublist,
3978 * this is to try and evenly balance eviction across all
3979 * sublists. Always starting at the same sublist
3980 * (e.g. index 0) would cause evictions to favor certain
3981 * sublists over others.
3983 for (int i = 0; i < num_sublists; i++) {
3984 uint64_t bytes_remaining;
3985 uint64_t bytes_evicted;
3987 if (bytes == ARC_EVICT_ALL)
3988 bytes_remaining = ARC_EVICT_ALL;
3989 else if (total_evicted < bytes)
3990 bytes_remaining = bytes - total_evicted;
3994 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3995 markers[sublist_idx], spa, bytes_remaining);
3997 scan_evicted += bytes_evicted;
3998 total_evicted += bytes_evicted;
4000 /* we've reached the end, wrap to the beginning */
4001 if (++sublist_idx >= num_sublists)
4006 * If we didn't evict anything during this scan, we have
4007 * no reason to believe we'll evict more during another
4008 * scan, so break the loop.
4010 if (scan_evicted == 0) {
4011 /* This isn't possible, let's make that obvious */
4012 ASSERT3S(bytes, !=, 0);
4015 * When bytes is ARC_EVICT_ALL, the only way to
4016 * break the loop is when scan_evicted is zero.
4017 * In that case, we actually have evicted enough,
4018 * so we don't want to increment the kstat.
4020 if (bytes != ARC_EVICT_ALL) {
4021 ASSERT3S(total_evicted, <, bytes);
4022 ARCSTAT_BUMP(arcstat_evict_not_enough);
4029 for (int i = 0; i < num_sublists; i++) {
4030 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4031 multilist_sublist_remove(mls, markers[i]);
4032 multilist_sublist_unlock(mls);
4034 kmem_cache_free(hdr_full_cache, markers[i]);
4036 kmem_free(markers, sizeof (*markers) * num_sublists);
4038 return (total_evicted);
4042 * Flush all "evictable" data of the given type from the arc state
4043 * specified. This will not evict any "active" buffers (i.e. referenced).
4045 * When 'retry' is set to B_FALSE, the function will make a single pass
4046 * over the state and evict any buffers that it can. Since it doesn't
4047 * continually retry the eviction, it might end up leaving some buffers
4048 * in the ARC due to lock misses.
4050 * When 'retry' is set to B_TRUE, the function will continually retry the
4051 * eviction until *all* evictable buffers have been removed from the
4052 * state. As a result, if concurrent insertions into the state are
4053 * allowed (e.g. if the ARC isn't shutting down), this function might
4054 * wind up in an infinite loop, continually trying to evict buffers.
4057 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4060 uint64_t evicted = 0;
4062 while (refcount_count(&state->arcs_esize[type]) != 0) {
4063 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
4073 * Helper function for arc_prune_async() it is responsible for safely
4074 * handling the execution of a registered arc_prune_func_t.
4077 arc_prune_task(void *ptr)
4079 arc_prune_t *ap = (arc_prune_t *)ptr;
4080 arc_prune_func_t *func = ap->p_pfunc;
4083 func(ap->p_adjust, ap->p_private);
4085 refcount_remove(&ap->p_refcnt, func);
4089 * Notify registered consumers they must drop holds on a portion of the ARC
4090 * buffered they reference. This provides a mechanism to ensure the ARC can
4091 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
4092 * is analogous to dnlc_reduce_cache() but more generic.
4094 * This operation is performed asynchronously so it may be safely called
4095 * in the context of the arc_reclaim_thread(). A reference is taken here
4096 * for each registered arc_prune_t and the arc_prune_task() is responsible
4097 * for releasing it once the registered arc_prune_func_t has completed.
4100 arc_prune_async(int64_t adjust)
4104 mutex_enter(&arc_prune_mtx);
4105 for (ap = list_head(&arc_prune_list); ap != NULL;
4106 ap = list_next(&arc_prune_list, ap)) {
4108 if (refcount_count(&ap->p_refcnt) >= 2)
4111 refcount_add(&ap->p_refcnt, ap->p_pfunc);
4112 ap->p_adjust = adjust;
4113 if (taskq_dispatch(arc_prune_taskq, arc_prune_task,
4114 ap, TQ_SLEEP) == TASKQID_INVALID) {
4115 refcount_remove(&ap->p_refcnt, ap->p_pfunc);
4118 ARCSTAT_BUMP(arcstat_prune);
4120 mutex_exit(&arc_prune_mtx);
4124 * Evict the specified number of bytes from the state specified,
4125 * restricting eviction to the spa and type given. This function
4126 * prevents us from trying to evict more from a state's list than
4127 * is "evictable", and to skip evicting altogether when passed a
4128 * negative value for "bytes". In contrast, arc_evict_state() will
4129 * evict everything it can, when passed a negative value for "bytes".
4132 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
4133 arc_buf_contents_t type)
4137 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
4138 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
4139 return (arc_evict_state(state, spa, delta, type));
4146 * The goal of this function is to evict enough meta data buffers from the
4147 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
4148 * more complicated than it appears because it is common for data buffers
4149 * to have holds on meta data buffers. In addition, dnode meta data buffers
4150 * will be held by the dnodes in the block preventing them from being freed.
4151 * This means we can't simply traverse the ARC and expect to always find
4152 * enough unheld meta data buffer to release.
4154 * Therefore, this function has been updated to make alternating passes
4155 * over the ARC releasing data buffers and then newly unheld meta data
4156 * buffers. This ensures forward progress is maintained and meta_used
4157 * will decrease. Normally this is sufficient, but if required the ARC
4158 * will call the registered prune callbacks causing dentry and inodes to
4159 * be dropped from the VFS cache. This will make dnode meta data buffers
4160 * available for reclaim.
4163 arc_adjust_meta_balanced(uint64_t meta_used)
4165 int64_t delta, prune = 0, adjustmnt;
4166 uint64_t total_evicted = 0;
4167 arc_buf_contents_t type = ARC_BUFC_DATA;
4168 int restarts = MAX(zfs_arc_meta_adjust_restarts, 0);
4172 * This slightly differs than the way we evict from the mru in
4173 * arc_adjust because we don't have a "target" value (i.e. no
4174 * "meta" arc_p). As a result, I think we can completely
4175 * cannibalize the metadata in the MRU before we evict the
4176 * metadata from the MFU. I think we probably need to implement a
4177 * "metadata arc_p" value to do this properly.
4179 adjustmnt = meta_used - arc_meta_limit;
4181 if (adjustmnt > 0 && refcount_count(&arc_mru->arcs_esize[type]) > 0) {
4182 delta = MIN(refcount_count(&arc_mru->arcs_esize[type]),
4184 total_evicted += arc_adjust_impl(arc_mru, 0, delta, type);
4189 * We can't afford to recalculate adjustmnt here. If we do,
4190 * new metadata buffers can sneak into the MRU or ANON lists,
4191 * thus penalize the MFU metadata. Although the fudge factor is
4192 * small, it has been empirically shown to be significant for
4193 * certain workloads (e.g. creating many empty directories). As
4194 * such, we use the original calculation for adjustmnt, and
4195 * simply decrement the amount of data evicted from the MRU.
4198 if (adjustmnt > 0 && refcount_count(&arc_mfu->arcs_esize[type]) > 0) {
4199 delta = MIN(refcount_count(&arc_mfu->arcs_esize[type]),
4201 total_evicted += arc_adjust_impl(arc_mfu, 0, delta, type);
4204 adjustmnt = meta_used - arc_meta_limit;
4206 if (adjustmnt > 0 &&
4207 refcount_count(&arc_mru_ghost->arcs_esize[type]) > 0) {
4208 delta = MIN(adjustmnt,
4209 refcount_count(&arc_mru_ghost->arcs_esize[type]));
4210 total_evicted += arc_adjust_impl(arc_mru_ghost, 0, delta, type);
4214 if (adjustmnt > 0 &&
4215 refcount_count(&arc_mfu_ghost->arcs_esize[type]) > 0) {
4216 delta = MIN(adjustmnt,
4217 refcount_count(&arc_mfu_ghost->arcs_esize[type]));
4218 total_evicted += arc_adjust_impl(arc_mfu_ghost, 0, delta, type);
4222 * If after attempting to make the requested adjustment to the ARC
4223 * the meta limit is still being exceeded then request that the
4224 * higher layers drop some cached objects which have holds on ARC
4225 * meta buffers. Requests to the upper layers will be made with
4226 * increasingly large scan sizes until the ARC is below the limit.
4228 if (meta_used > arc_meta_limit) {
4229 if (type == ARC_BUFC_DATA) {
4230 type = ARC_BUFC_METADATA;
4232 type = ARC_BUFC_DATA;
4234 if (zfs_arc_meta_prune) {
4235 prune += zfs_arc_meta_prune;
4236 arc_prune_async(prune);
4245 return (total_evicted);
4249 * Evict metadata buffers from the cache, such that arc_meta_used is
4250 * capped by the arc_meta_limit tunable.
4253 arc_adjust_meta_only(uint64_t meta_used)
4255 uint64_t total_evicted = 0;
4259 * If we're over the meta limit, we want to evict enough
4260 * metadata to get back under the meta limit. We don't want to
4261 * evict so much that we drop the MRU below arc_p, though. If
4262 * we're over the meta limit more than we're over arc_p, we
4263 * evict some from the MRU here, and some from the MFU below.
4265 target = MIN((int64_t)(meta_used - arc_meta_limit),
4266 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4267 refcount_count(&arc_mru->arcs_size) - arc_p));
4269 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4272 * Similar to the above, we want to evict enough bytes to get us
4273 * below the meta limit, but not so much as to drop us below the
4274 * space allotted to the MFU (which is defined as arc_c - arc_p).
4276 target = MIN((int64_t)(meta_used - arc_meta_limit),
4277 (int64_t)(refcount_count(&arc_mfu->arcs_size) -
4280 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4282 return (total_evicted);
4286 arc_adjust_meta(uint64_t meta_used)
4288 if (zfs_arc_meta_strategy == ARC_STRATEGY_META_ONLY)
4289 return (arc_adjust_meta_only(meta_used));
4291 return (arc_adjust_meta_balanced(meta_used));
4295 * Return the type of the oldest buffer in the given arc state
4297 * This function will select a random sublist of type ARC_BUFC_DATA and
4298 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4299 * is compared, and the type which contains the "older" buffer will be
4302 static arc_buf_contents_t
4303 arc_adjust_type(arc_state_t *state)
4305 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
4306 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
4307 int data_idx = multilist_get_random_index(data_ml);
4308 int meta_idx = multilist_get_random_index(meta_ml);
4309 multilist_sublist_t *data_mls;
4310 multilist_sublist_t *meta_mls;
4311 arc_buf_contents_t type;
4312 arc_buf_hdr_t *data_hdr;
4313 arc_buf_hdr_t *meta_hdr;
4316 * We keep the sublist lock until we're finished, to prevent
4317 * the headers from being destroyed via arc_evict_state().
4319 data_mls = multilist_sublist_lock(data_ml, data_idx);
4320 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4323 * These two loops are to ensure we skip any markers that
4324 * might be at the tail of the lists due to arc_evict_state().
4327 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4328 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4329 if (data_hdr->b_spa != 0)
4333 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4334 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4335 if (meta_hdr->b_spa != 0)
4339 if (data_hdr == NULL && meta_hdr == NULL) {
4340 type = ARC_BUFC_DATA;
4341 } else if (data_hdr == NULL) {
4342 ASSERT3P(meta_hdr, !=, NULL);
4343 type = ARC_BUFC_METADATA;
4344 } else if (meta_hdr == NULL) {
4345 ASSERT3P(data_hdr, !=, NULL);
4346 type = ARC_BUFC_DATA;
4348 ASSERT3P(data_hdr, !=, NULL);
4349 ASSERT3P(meta_hdr, !=, NULL);
4351 /* The headers can't be on the sublist without an L1 header */
4352 ASSERT(HDR_HAS_L1HDR(data_hdr));
4353 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4355 if (data_hdr->b_l1hdr.b_arc_access <
4356 meta_hdr->b_l1hdr.b_arc_access) {
4357 type = ARC_BUFC_DATA;
4359 type = ARC_BUFC_METADATA;
4363 multilist_sublist_unlock(meta_mls);
4364 multilist_sublist_unlock(data_mls);
4370 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4375 uint64_t total_evicted = 0;
4378 uint64_t asize = aggsum_value(&arc_size);
4379 uint64_t ameta = aggsum_value(&arc_meta_used);
4382 * If we're over arc_meta_limit, we want to correct that before
4383 * potentially evicting data buffers below.
4385 total_evicted += arc_adjust_meta(ameta);
4390 * If we're over the target cache size, we want to evict enough
4391 * from the list to get back to our target size. We don't want
4392 * to evict too much from the MRU, such that it drops below
4393 * arc_p. So, if we're over our target cache size more than
4394 * the MRU is over arc_p, we'll evict enough to get back to
4395 * arc_p here, and then evict more from the MFU below.
4397 target = MIN((int64_t)(asize - arc_c),
4398 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4399 refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4402 * If we're below arc_meta_min, always prefer to evict data.
4403 * Otherwise, try to satisfy the requested number of bytes to
4404 * evict from the type which contains older buffers; in an
4405 * effort to keep newer buffers in the cache regardless of their
4406 * type. If we cannot satisfy the number of bytes from this
4407 * type, spill over into the next type.
4409 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4410 ameta > arc_meta_min) {
4411 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4412 total_evicted += bytes;
4415 * If we couldn't evict our target number of bytes from
4416 * metadata, we try to get the rest from data.
4421 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4423 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4424 total_evicted += bytes;
4427 * If we couldn't evict our target number of bytes from
4428 * data, we try to get the rest from metadata.
4433 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4439 * Now that we've tried to evict enough from the MRU to get its
4440 * size back to arc_p, if we're still above the target cache
4441 * size, we evict the rest from the MFU.
4443 target = asize - arc_c;
4445 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4446 ameta > arc_meta_min) {
4447 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4448 total_evicted += bytes;
4451 * If we couldn't evict our target number of bytes from
4452 * metadata, we try to get the rest from data.
4457 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4459 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4460 total_evicted += bytes;
4463 * If we couldn't evict our target number of bytes from
4464 * data, we try to get the rest from data.
4469 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4473 * Adjust ghost lists
4475 * In addition to the above, the ARC also defines target values
4476 * for the ghost lists. The sum of the mru list and mru ghost
4477 * list should never exceed the target size of the cache, and
4478 * the sum of the mru list, mfu list, mru ghost list, and mfu
4479 * ghost list should never exceed twice the target size of the
4480 * cache. The following logic enforces these limits on the ghost
4481 * caches, and evicts from them as needed.
4483 target = refcount_count(&arc_mru->arcs_size) +
4484 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4486 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4487 total_evicted += bytes;
4492 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4495 * We assume the sum of the mru list and mfu list is less than
4496 * or equal to arc_c (we enforced this above), which means we
4497 * can use the simpler of the two equations below:
4499 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4500 * mru ghost + mfu ghost <= arc_c
4502 target = refcount_count(&arc_mru_ghost->arcs_size) +
4503 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4505 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4506 total_evicted += bytes;
4511 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4513 return (total_evicted);
4517 arc_flush(spa_t *spa, boolean_t retry)
4522 * If retry is B_TRUE, a spa must not be specified since we have
4523 * no good way to determine if all of a spa's buffers have been
4524 * evicted from an arc state.
4526 ASSERT(!retry || spa == 0);
4529 guid = spa_load_guid(spa);
4531 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4532 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4534 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4535 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4537 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4538 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4540 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4541 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4545 arc_shrink(int64_t to_free)
4547 uint64_t asize = aggsum_value(&arc_size);
4548 if (arc_c > arc_c_min) {
4549 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
4550 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
4551 if (arc_c > arc_c_min + to_free)
4552 atomic_add_64(&arc_c, -to_free);
4556 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4558 arc_c = MAX(asize, arc_c_min);
4560 arc_p = (arc_c >> 1);
4562 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
4565 ASSERT(arc_c >= arc_c_min);
4566 ASSERT((int64_t)arc_p >= 0);
4569 if (asize > arc_c) {
4570 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, asize,
4572 (void) arc_adjust();
4576 typedef enum free_memory_reason_t {
4581 FMR_PAGES_PP_MAXIMUM,
4584 } free_memory_reason_t;
4586 int64_t last_free_memory;
4587 free_memory_reason_t last_free_reason;
4590 * Additional reserve of pages for pp_reserve.
4592 int64_t arc_pages_pp_reserve = 64;
4595 * Additional reserve of pages for swapfs.
4597 int64_t arc_swapfs_reserve = 64;
4600 * Return the amount of memory that can be consumed before reclaim will be
4601 * needed. Positive if there is sufficient free memory, negative indicates
4602 * the amount of memory that needs to be freed up.
4605 arc_available_memory(void)
4607 int64_t lowest = INT64_MAX;
4609 free_memory_reason_t r = FMR_UNKNOWN;
4614 * Cooperate with pagedaemon when it's time for it to scan
4615 * and reclaim some pages.
4617 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
4625 n = PAGESIZE * (-needfree);
4633 * check that we're out of range of the pageout scanner. It starts to
4634 * schedule paging if freemem is less than lotsfree and needfree.
4635 * lotsfree is the high-water mark for pageout, and needfree is the
4636 * number of needed free pages. We add extra pages here to make sure
4637 * the scanner doesn't start up while we're freeing memory.
4639 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4646 * check to make sure that swapfs has enough space so that anon
4647 * reservations can still succeed. anon_resvmem() checks that the
4648 * availrmem is greater than swapfs_minfree, and the number of reserved
4649 * swap pages. We also add a bit of extra here just to prevent
4650 * circumstances from getting really dire.
4652 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4653 desfree - arc_swapfs_reserve);
4656 r = FMR_SWAPFS_MINFREE;
4661 * Check that we have enough availrmem that memory locking (e.g., via
4662 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4663 * stores the number of pages that cannot be locked; when availrmem
4664 * drops below pages_pp_maximum, page locking mechanisms such as
4665 * page_pp_lock() will fail.)
4667 n = PAGESIZE * (availrmem - pages_pp_maximum -
4668 arc_pages_pp_reserve);
4671 r = FMR_PAGES_PP_MAXIMUM;
4674 #endif /* __FreeBSD__ */
4675 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
4677 * If we're on an i386 platform, it's possible that we'll exhaust the
4678 * kernel heap space before we ever run out of available physical
4679 * memory. Most checks of the size of the heap_area compare against
4680 * tune.t_minarmem, which is the minimum available real memory that we
4681 * can have in the system. However, this is generally fixed at 25 pages
4682 * which is so low that it's useless. In this comparison, we seek to
4683 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4684 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4687 n = uma_avail() - (long)(uma_limit() / 4);
4695 * If zio data pages are being allocated out of a separate heap segment,
4696 * then enforce that the size of available vmem for this arena remains
4697 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4699 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4700 * memory (in the zio_arena) free, which can avoid memory
4701 * fragmentation issues.
4703 if (zio_arena != NULL) {
4704 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4705 (vmem_size(zio_arena, VMEM_ALLOC) >>
4706 arc_zio_arena_free_shift);
4714 /* Every 100 calls, free a small amount */
4715 if (spa_get_random(100) == 0)
4717 #endif /* _KERNEL */
4719 last_free_memory = lowest;
4720 last_free_reason = r;
4721 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
4727 * Determine if the system is under memory pressure and is asking
4728 * to reclaim memory. A return value of B_TRUE indicates that the system
4729 * is under memory pressure and that the arc should adjust accordingly.
4732 arc_reclaim_needed(void)
4734 return (arc_available_memory() < 0);
4737 extern kmem_cache_t *zio_buf_cache[];
4738 extern kmem_cache_t *zio_data_buf_cache[];
4739 extern kmem_cache_t *range_seg_cache;
4740 extern kmem_cache_t *abd_chunk_cache;
4742 static __noinline void
4743 arc_kmem_reap_now(void)
4746 kmem_cache_t *prev_cache = NULL;
4747 kmem_cache_t *prev_data_cache = NULL;
4749 DTRACE_PROBE(arc__kmem_reap_start);
4751 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) {
4753 * We are exceeding our meta-data cache limit.
4754 * Purge some DNLC entries to release holds on meta-data.
4756 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4760 * Reclaim unused memory from all kmem caches.
4767 * If a kmem reap is already active, don't schedule more. We must
4768 * check for this because kmem_cache_reap_soon() won't actually
4769 * block on the cache being reaped (this is to prevent callers from
4770 * becoming implicitly blocked by a system-wide kmem reap -- which,
4771 * on a system with many, many full magazines, can take minutes).
4773 if (kmem_cache_reap_active())
4776 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4777 if (zio_buf_cache[i] != prev_cache) {
4778 prev_cache = zio_buf_cache[i];
4779 kmem_cache_reap_soon(zio_buf_cache[i]);
4781 if (zio_data_buf_cache[i] != prev_data_cache) {
4782 prev_data_cache = zio_data_buf_cache[i];
4783 kmem_cache_reap_soon(zio_data_buf_cache[i]);
4786 kmem_cache_reap_soon(abd_chunk_cache);
4787 kmem_cache_reap_soon(buf_cache);
4788 kmem_cache_reap_soon(hdr_full_cache);
4789 kmem_cache_reap_soon(hdr_l2only_cache);
4790 kmem_cache_reap_soon(range_seg_cache);
4793 if (zio_arena != NULL) {
4795 * Ask the vmem arena to reclaim unused memory from its
4798 vmem_qcache_reap(zio_arena);
4801 DTRACE_PROBE(arc__kmem_reap_end);
4805 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4806 * enough data and signal them to proceed. When this happens, the threads in
4807 * arc_get_data_impl() are sleeping while holding the hash lock for their
4808 * particular arc header. Thus, we must be careful to never sleep on a
4809 * hash lock in this thread. This is to prevent the following deadlock:
4811 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4812 * waiting for the reclaim thread to signal it.
4814 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4815 * fails, and goes to sleep forever.
4817 * This possible deadlock is avoided by always acquiring a hash lock
4818 * using mutex_tryenter() from arc_reclaim_thread().
4822 arc_reclaim_thread(void *unused __unused)
4824 hrtime_t growtime = 0;
4825 hrtime_t kmem_reap_time = 0;
4828 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4830 mutex_enter(&arc_reclaim_lock);
4831 while (!arc_reclaim_thread_exit) {
4832 uint64_t evicted = 0;
4835 * This is necessary in order for the mdb ::arc dcmd to
4836 * show up to date information. Since the ::arc command
4837 * does not call the kstat's update function, without
4838 * this call, the command may show stale stats for the
4839 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4840 * with this change, the data might be up to 1 second
4841 * out of date; but that should suffice. The arc_state_t
4842 * structures can be queried directly if more accurate
4843 * information is needed.
4845 if (arc_ksp != NULL)
4846 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4848 mutex_exit(&arc_reclaim_lock);
4851 * We call arc_adjust() before (possibly) calling
4852 * arc_kmem_reap_now(), so that we can wake up
4853 * arc_get_data_impl() sooner.
4855 evicted = arc_adjust();
4857 int64_t free_memory = arc_available_memory();
4858 if (free_memory < 0) {
4859 hrtime_t curtime = gethrtime();
4860 arc_no_grow = B_TRUE;
4864 * Wait at least zfs_grow_retry (default 60) seconds
4865 * before considering growing.
4867 growtime = curtime + SEC2NSEC(arc_grow_retry);
4870 * Wait at least arc_kmem_cache_reap_retry_ms
4871 * between arc_kmem_reap_now() calls. Without
4872 * this check it is possible to end up in a
4873 * situation where we spend lots of time
4874 * reaping caches, while we're near arc_c_min.
4876 if (curtime >= kmem_reap_time) {
4877 arc_kmem_reap_now();
4878 kmem_reap_time = gethrtime() +
4879 MSEC2NSEC(arc_kmem_cache_reap_retry_ms);
4883 * If we are still low on memory, shrink the ARC
4884 * so that we have arc_shrink_min free space.
4886 free_memory = arc_available_memory();
4889 (arc_c >> arc_shrink_shift) - free_memory;
4893 to_free = MAX(to_free, ptob(needfree));
4896 arc_shrink(to_free);
4898 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4899 arc_no_grow = B_TRUE;
4900 } else if (gethrtime() >= growtime) {
4901 arc_no_grow = B_FALSE;
4904 mutex_enter(&arc_reclaim_lock);
4907 * If evicted is zero, we couldn't evict anything via
4908 * arc_adjust(). This could be due to hash lock
4909 * collisions, but more likely due to the majority of
4910 * arc buffers being unevictable. Therefore, even if
4911 * arc_size is above arc_c, another pass is unlikely to
4912 * be helpful and could potentially cause us to enter an
4915 if (aggsum_compare(&arc_size, arc_c) <= 0|| evicted == 0) {
4917 * We're either no longer overflowing, or we
4918 * can't evict anything more, so we should wake
4919 * up any threads before we go to sleep.
4921 cv_broadcast(&arc_reclaim_waiters_cv);
4924 * Block until signaled, or after one second (we
4925 * might need to perform arc_kmem_reap_now()
4926 * even if we aren't being signalled)
4928 CALLB_CPR_SAFE_BEGIN(&cpr);
4929 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4930 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4931 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4935 arc_reclaim_thread_exit = B_FALSE;
4936 cv_broadcast(&arc_reclaim_thread_cv);
4937 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
4941 static u_int arc_dnlc_evicts_arg;
4942 extern struct vfsops zfs_vfsops;
4945 arc_dnlc_evicts_thread(void *dummy __unused)
4950 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG);
4952 mutex_enter(&arc_dnlc_evicts_lock);
4953 while (!arc_dnlc_evicts_thread_exit) {
4954 CALLB_CPR_SAFE_BEGIN(&cpr);
4955 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
4956 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock);
4957 if (arc_dnlc_evicts_arg != 0) {
4958 percent = arc_dnlc_evicts_arg;
4959 mutex_exit(&arc_dnlc_evicts_lock);
4961 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops);
4963 mutex_enter(&arc_dnlc_evicts_lock);
4965 * Clear our token only after vnlru_free()
4966 * pass is done, to avoid false queueing of
4969 arc_dnlc_evicts_arg = 0;
4972 arc_dnlc_evicts_thread_exit = FALSE;
4973 cv_broadcast(&arc_dnlc_evicts_cv);
4974 CALLB_CPR_EXIT(&cpr);
4979 dnlc_reduce_cache(void *arg)
4983 percent = (u_int)(uintptr_t)arg;
4984 mutex_enter(&arc_dnlc_evicts_lock);
4985 if (arc_dnlc_evicts_arg == 0) {
4986 arc_dnlc_evicts_arg = percent;
4987 cv_broadcast(&arc_dnlc_evicts_cv);
4989 mutex_exit(&arc_dnlc_evicts_lock);
4993 * Adapt arc info given the number of bytes we are trying to add and
4994 * the state that we are comming from. This function is only called
4995 * when we are adding new content to the cache.
4998 arc_adapt(int bytes, arc_state_t *state)
5001 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
5002 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
5003 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
5005 if (state == arc_l2c_only)
5010 * Adapt the target size of the MRU list:
5011 * - if we just hit in the MRU ghost list, then increase
5012 * the target size of the MRU list.
5013 * - if we just hit in the MFU ghost list, then increase
5014 * the target size of the MFU list by decreasing the
5015 * target size of the MRU list.
5017 if (state == arc_mru_ghost) {
5018 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
5019 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
5021 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
5022 } else if (state == arc_mfu_ghost) {
5025 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
5026 mult = MIN(mult, 10);
5028 delta = MIN(bytes * mult, arc_p);
5029 arc_p = MAX(arc_p_min, arc_p - delta);
5031 ASSERT((int64_t)arc_p >= 0);
5033 if (arc_reclaim_needed()) {
5034 cv_signal(&arc_reclaim_thread_cv);
5041 if (arc_c >= arc_c_max)
5045 * If we're within (2 * maxblocksize) bytes of the target
5046 * cache size, increment the target cache size
5048 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >
5050 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
5051 atomic_add_64(&arc_c, (int64_t)bytes);
5052 if (arc_c > arc_c_max)
5054 else if (state == arc_anon)
5055 atomic_add_64(&arc_p, (int64_t)bytes);
5059 ASSERT((int64_t)arc_p >= 0);
5063 * Check if arc_size has grown past our upper threshold, determined by
5064 * zfs_arc_overflow_shift.
5067 arc_is_overflowing(void)
5069 /* Always allow at least one block of overflow */
5070 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
5071 arc_c >> zfs_arc_overflow_shift);
5074 * We just compare the lower bound here for performance reasons. Our
5075 * primary goals are to make sure that the arc never grows without
5076 * bound, and that it can reach its maximum size. This check
5077 * accomplishes both goals. The maximum amount we could run over by is
5078 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
5079 * in the ARC. In practice, that's in the tens of MB, which is low
5080 * enough to be safe.
5082 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
5086 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5088 arc_buf_contents_t type = arc_buf_type(hdr);
5090 arc_get_data_impl(hdr, size, tag);
5091 if (type == ARC_BUFC_METADATA) {
5092 return (abd_alloc(size, B_TRUE));
5094 ASSERT(type == ARC_BUFC_DATA);
5095 return (abd_alloc(size, B_FALSE));
5100 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5102 arc_buf_contents_t type = arc_buf_type(hdr);
5104 arc_get_data_impl(hdr, size, tag);
5105 if (type == ARC_BUFC_METADATA) {
5106 return (zio_buf_alloc(size));
5108 ASSERT(type == ARC_BUFC_DATA);
5109 return (zio_data_buf_alloc(size));
5114 * Allocate a block and return it to the caller. If we are hitting the
5115 * hard limit for the cache size, we must sleep, waiting for the eviction
5116 * thread to catch up. If we're past the target size but below the hard
5117 * limit, we'll only signal the reclaim thread and continue on.
5120 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5122 arc_state_t *state = hdr->b_l1hdr.b_state;
5123 arc_buf_contents_t type = arc_buf_type(hdr);
5125 arc_adapt(size, state);
5128 * If arc_size is currently overflowing, and has grown past our
5129 * upper limit, we must be adding data faster than the evict
5130 * thread can evict. Thus, to ensure we don't compound the
5131 * problem by adding more data and forcing arc_size to grow even
5132 * further past it's target size, we halt and wait for the
5133 * eviction thread to catch up.
5135 * It's also possible that the reclaim thread is unable to evict
5136 * enough buffers to get arc_size below the overflow limit (e.g.
5137 * due to buffers being un-evictable, or hash lock collisions).
5138 * In this case, we want to proceed regardless if we're
5139 * overflowing; thus we don't use a while loop here.
5141 if (arc_is_overflowing()) {
5142 mutex_enter(&arc_reclaim_lock);
5145 * Now that we've acquired the lock, we may no longer be
5146 * over the overflow limit, lets check.
5148 * We're ignoring the case of spurious wake ups. If that
5149 * were to happen, it'd let this thread consume an ARC
5150 * buffer before it should have (i.e. before we're under
5151 * the overflow limit and were signalled by the reclaim
5152 * thread). As long as that is a rare occurrence, it
5153 * shouldn't cause any harm.
5155 if (arc_is_overflowing()) {
5156 cv_signal(&arc_reclaim_thread_cv);
5157 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
5160 mutex_exit(&arc_reclaim_lock);
5163 VERIFY3U(hdr->b_type, ==, type);
5164 if (type == ARC_BUFC_METADATA) {
5165 arc_space_consume(size, ARC_SPACE_META);
5167 arc_space_consume(size, ARC_SPACE_DATA);
5171 * Update the state size. Note that ghost states have a
5172 * "ghost size" and so don't need to be updated.
5174 if (!GHOST_STATE(state)) {
5176 (void) refcount_add_many(&state->arcs_size, size, tag);
5179 * If this is reached via arc_read, the link is
5180 * protected by the hash lock. If reached via
5181 * arc_buf_alloc, the header should not be accessed by
5182 * any other thread. And, if reached via arc_read_done,
5183 * the hash lock will protect it if it's found in the
5184 * hash table; otherwise no other thread should be
5185 * trying to [add|remove]_reference it.
5187 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5188 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5189 (void) refcount_add_many(&state->arcs_esize[type],
5194 * If we are growing the cache, and we are adding anonymous
5195 * data, and we have outgrown arc_p, update arc_p
5197 if (aggsum_compare(&arc_size, arc_c) < 0 &&
5198 hdr->b_l1hdr.b_state == arc_anon &&
5199 (refcount_count(&arc_anon->arcs_size) +
5200 refcount_count(&arc_mru->arcs_size) > arc_p))
5201 arc_p = MIN(arc_c, arc_p + size);
5203 ARCSTAT_BUMP(arcstat_allocated);
5207 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
5209 arc_free_data_impl(hdr, size, tag);
5214 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
5216 arc_buf_contents_t type = arc_buf_type(hdr);
5218 arc_free_data_impl(hdr, size, tag);
5219 if (type == ARC_BUFC_METADATA) {
5220 zio_buf_free(buf, size);
5222 ASSERT(type == ARC_BUFC_DATA);
5223 zio_data_buf_free(buf, size);
5228 * Free the arc data buffer.
5231 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5233 arc_state_t *state = hdr->b_l1hdr.b_state;
5234 arc_buf_contents_t type = arc_buf_type(hdr);
5236 /* protected by hash lock, if in the hash table */
5237 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5238 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5239 ASSERT(state != arc_anon && state != arc_l2c_only);
5241 (void) refcount_remove_many(&state->arcs_esize[type],
5244 (void) refcount_remove_many(&state->arcs_size, size, tag);
5246 VERIFY3U(hdr->b_type, ==, type);
5247 if (type == ARC_BUFC_METADATA) {
5248 arc_space_return(size, ARC_SPACE_META);
5250 ASSERT(type == ARC_BUFC_DATA);
5251 arc_space_return(size, ARC_SPACE_DATA);
5256 * This routine is called whenever a buffer is accessed.
5257 * NOTE: the hash lock is dropped in this function.
5260 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
5264 ASSERT(MUTEX_HELD(hash_lock));
5265 ASSERT(HDR_HAS_L1HDR(hdr));
5267 if (hdr->b_l1hdr.b_state == arc_anon) {
5269 * This buffer is not in the cache, and does not
5270 * appear in our "ghost" list. Add the new buffer
5274 ASSERT0(hdr->b_l1hdr.b_arc_access);
5275 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5276 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5277 arc_change_state(arc_mru, hdr, hash_lock);
5279 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5280 now = ddi_get_lbolt();
5283 * If this buffer is here because of a prefetch, then either:
5284 * - clear the flag if this is a "referencing" read
5285 * (any subsequent access will bump this into the MFU state).
5287 * - move the buffer to the head of the list if this is
5288 * another prefetch (to make it less likely to be evicted).
5290 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5291 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5292 /* link protected by hash lock */
5293 ASSERT(multilist_link_active(
5294 &hdr->b_l1hdr.b_arc_node));
5296 arc_hdr_clear_flags(hdr,
5298 ARC_FLAG_PRESCIENT_PREFETCH);
5299 ARCSTAT_BUMP(arcstat_mru_hits);
5301 hdr->b_l1hdr.b_arc_access = now;
5306 * This buffer has been "accessed" only once so far,
5307 * but it is still in the cache. Move it to the MFU
5310 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
5312 * More than 125ms have passed since we
5313 * instantiated this buffer. Move it to the
5314 * most frequently used state.
5316 hdr->b_l1hdr.b_arc_access = now;
5317 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5318 arc_change_state(arc_mfu, hdr, hash_lock);
5320 atomic_inc_32(&hdr->b_l1hdr.b_mru_hits);
5321 ARCSTAT_BUMP(arcstat_mru_hits);
5322 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5323 arc_state_t *new_state;
5325 * This buffer has been "accessed" recently, but
5326 * was evicted from the cache. Move it to the
5330 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5331 new_state = arc_mru;
5332 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
5333 arc_hdr_clear_flags(hdr,
5335 ARC_FLAG_PRESCIENT_PREFETCH);
5337 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5339 new_state = arc_mfu;
5340 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5343 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5344 arc_change_state(new_state, hdr, hash_lock);
5346 atomic_inc_32(&hdr->b_l1hdr.b_mru_ghost_hits);
5347 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5348 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5350 * This buffer has been accessed more than once and is
5351 * still in the cache. Keep it in the MFU state.
5353 * NOTE: an add_reference() that occurred when we did
5354 * the arc_read() will have kicked this off the list.
5355 * If it was a prefetch, we will explicitly move it to
5356 * the head of the list now.
5359 atomic_inc_32(&hdr->b_l1hdr.b_mfu_hits);
5360 ARCSTAT_BUMP(arcstat_mfu_hits);
5361 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5362 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5363 arc_state_t *new_state = arc_mfu;
5365 * This buffer has been accessed more than once but has
5366 * been evicted from the cache. Move it back to the
5370 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5372 * This is a prefetch access...
5373 * move this block back to the MRU state.
5375 new_state = arc_mru;
5378 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5379 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5380 arc_change_state(new_state, hdr, hash_lock);
5382 atomic_inc_32(&hdr->b_l1hdr.b_mfu_ghost_hits);
5383 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5384 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5386 * This buffer is on the 2nd Level ARC.
5389 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5390 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5391 arc_change_state(arc_mfu, hdr, hash_lock);
5393 ASSERT(!"invalid arc state");
5398 * This routine is called by dbuf_hold() to update the arc_access() state
5399 * which otherwise would be skipped for entries in the dbuf cache.
5402 arc_buf_access(arc_buf_t *buf)
5404 mutex_enter(&buf->b_evict_lock);
5405 arc_buf_hdr_t *hdr = buf->b_hdr;
5408 * Avoid taking the hash_lock when possible as an optimization.
5409 * The header must be checked again under the hash_lock in order
5410 * to handle the case where it is concurrently being released.
5412 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5413 mutex_exit(&buf->b_evict_lock);
5414 ARCSTAT_BUMP(arcstat_access_skip);
5418 kmutex_t *hash_lock = HDR_LOCK(hdr);
5419 mutex_enter(hash_lock);
5421 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5422 mutex_exit(hash_lock);
5423 mutex_exit(&buf->b_evict_lock);
5424 ARCSTAT_BUMP(arcstat_access_skip);
5428 mutex_exit(&buf->b_evict_lock);
5430 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5431 hdr->b_l1hdr.b_state == arc_mfu);
5433 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5434 arc_access(hdr, hash_lock);
5435 mutex_exit(hash_lock);
5437 ARCSTAT_BUMP(arcstat_hits);
5438 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5439 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5442 /* a generic arc_read_done_func_t which you can use */
5445 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5446 arc_buf_t *buf, void *arg)
5451 bcopy(buf->b_data, arg, arc_buf_size(buf));
5452 arc_buf_destroy(buf, arg);
5455 /* a generic arc_read_done_func_t */
5458 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5459 arc_buf_t *buf, void *arg)
5461 arc_buf_t **bufp = arg;
5463 ASSERT(zio == NULL || zio->io_error != 0);
5466 ASSERT(zio == NULL || zio->io_error == 0);
5468 ASSERT(buf->b_data != NULL);
5473 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5475 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5476 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5477 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
5479 if (HDR_COMPRESSION_ENABLED(hdr)) {
5480 ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
5481 BP_GET_COMPRESS(bp));
5483 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5484 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5489 arc_read_done(zio_t *zio)
5491 arc_buf_hdr_t *hdr = zio->io_private;
5492 kmutex_t *hash_lock = NULL;
5493 arc_callback_t *callback_list;
5494 arc_callback_t *acb;
5495 boolean_t freeable = B_FALSE;
5496 boolean_t no_zio_error = (zio->io_error == 0);
5499 * The hdr was inserted into hash-table and removed from lists
5500 * prior to starting I/O. We should find this header, since
5501 * it's in the hash table, and it should be legit since it's
5502 * not possible to evict it during the I/O. The only possible
5503 * reason for it not to be found is if we were freed during the
5506 if (HDR_IN_HASH_TABLE(hdr)) {
5507 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5508 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5509 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5510 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5511 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5513 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
5516 ASSERT((found == hdr &&
5517 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5518 (found == hdr && HDR_L2_READING(hdr)));
5519 ASSERT3P(hash_lock, !=, NULL);
5523 /* byteswap if necessary */
5524 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5525 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5526 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5528 hdr->b_l1hdr.b_byteswap =
5529 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5532 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5536 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5537 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5538 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5540 callback_list = hdr->b_l1hdr.b_acb;
5541 ASSERT3P(callback_list, !=, NULL);
5543 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) {
5545 * Only call arc_access on anonymous buffers. This is because
5546 * if we've issued an I/O for an evicted buffer, we've already
5547 * called arc_access (to prevent any simultaneous readers from
5548 * getting confused).
5550 arc_access(hdr, hash_lock);
5554 * If a read request has a callback (i.e. acb_done is not NULL), then we
5555 * make a buf containing the data according to the parameters which were
5556 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5557 * aren't needlessly decompressing the data multiple times.
5559 int callback_cnt = 0;
5560 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5567 int error = arc_buf_alloc_impl(hdr, acb->acb_private,
5568 acb->acb_compressed, zio->io_error == 0,
5572 * Decompression failed. Set io_error
5573 * so that when we call acb_done (below),
5574 * we will indicate that the read failed.
5575 * Note that in the unusual case where one
5576 * callback is compressed and another
5577 * uncompressed, we will mark all of them
5578 * as failed, even though the uncompressed
5579 * one can't actually fail. In this case,
5580 * the hdr will not be anonymous, because
5581 * if there are multiple callbacks, it's
5582 * because multiple threads found the same
5583 * arc buf in the hash table.
5585 zio->io_error = error;
5590 * If there are multiple callbacks, we must have the hash lock,
5591 * because the only way for multiple threads to find this hdr is
5592 * in the hash table. This ensures that if there are multiple
5593 * callbacks, the hdr is not anonymous. If it were anonymous,
5594 * we couldn't use arc_buf_destroy() in the error case below.
5596 ASSERT(callback_cnt < 2 || hash_lock != NULL);
5598 hdr->b_l1hdr.b_acb = NULL;
5599 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5600 if (callback_cnt == 0) {
5601 ASSERT(HDR_PREFETCH(hdr));
5602 ASSERT0(hdr->b_l1hdr.b_bufcnt);
5603 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5606 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5607 callback_list != NULL);
5610 arc_hdr_verify(hdr, zio->io_bp);
5612 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5613 if (hdr->b_l1hdr.b_state != arc_anon)
5614 arc_change_state(arc_anon, hdr, hash_lock);
5615 if (HDR_IN_HASH_TABLE(hdr))
5616 buf_hash_remove(hdr);
5617 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5621 * Broadcast before we drop the hash_lock to avoid the possibility
5622 * that the hdr (and hence the cv) might be freed before we get to
5623 * the cv_broadcast().
5625 cv_broadcast(&hdr->b_l1hdr.b_cv);
5627 if (hash_lock != NULL) {
5628 mutex_exit(hash_lock);
5631 * This block was freed while we waited for the read to
5632 * complete. It has been removed from the hash table and
5633 * moved to the anonymous state (so that it won't show up
5636 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5637 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5640 /* execute each callback and free its structure */
5641 while ((acb = callback_list) != NULL) {
5642 if (acb->acb_done != NULL) {
5643 if (zio->io_error != 0 && acb->acb_buf != NULL) {
5645 * If arc_buf_alloc_impl() fails during
5646 * decompression, the buf will still be
5647 * allocated, and needs to be freed here.
5649 arc_buf_destroy(acb->acb_buf, acb->acb_private);
5650 acb->acb_buf = NULL;
5652 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5653 acb->acb_buf, acb->acb_private);
5656 if (acb->acb_zio_dummy != NULL) {
5657 acb->acb_zio_dummy->io_error = zio->io_error;
5658 zio_nowait(acb->acb_zio_dummy);
5661 callback_list = acb->acb_next;
5662 kmem_free(acb, sizeof (arc_callback_t));
5666 arc_hdr_destroy(hdr);
5670 * "Read" the block at the specified DVA (in bp) via the
5671 * cache. If the block is found in the cache, invoke the provided
5672 * callback immediately and return. Note that the `zio' parameter
5673 * in the callback will be NULL in this case, since no IO was
5674 * required. If the block is not in the cache pass the read request
5675 * on to the spa with a substitute callback function, so that the
5676 * requested block will be added to the cache.
5678 * If a read request arrives for a block that has a read in-progress,
5679 * either wait for the in-progress read to complete (and return the
5680 * results); or, if this is a read with a "done" func, add a record
5681 * to the read to invoke the "done" func when the read completes,
5682 * and return; or just return.
5684 * arc_read_done() will invoke all the requested "done" functions
5685 * for readers of this block.
5688 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_read_done_func_t *done,
5689 void *private, zio_priority_t priority, int zio_flags,
5690 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5692 arc_buf_hdr_t *hdr = NULL;
5693 kmutex_t *hash_lock = NULL;
5695 uint64_t guid = spa_load_guid(spa);
5696 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0;
5699 ASSERT(!BP_IS_EMBEDDED(bp) ||
5700 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5703 if (!BP_IS_EMBEDDED(bp)) {
5705 * Embedded BP's have no DVA and require no I/O to "read".
5706 * Create an anonymous arc buf to back it.
5708 hdr = buf_hash_find(guid, bp, &hash_lock);
5711 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) {
5712 arc_buf_t *buf = NULL;
5713 *arc_flags |= ARC_FLAG_CACHED;
5715 if (HDR_IO_IN_PROGRESS(hdr)) {
5716 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
5718 ASSERT3P(head_zio, !=, NULL);
5719 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5720 priority == ZIO_PRIORITY_SYNC_READ) {
5722 * This is a sync read that needs to wait for
5723 * an in-flight async read. Request that the
5724 * zio have its priority upgraded.
5726 zio_change_priority(head_zio, priority);
5727 DTRACE_PROBE1(arc__async__upgrade__sync,
5728 arc_buf_hdr_t *, hdr);
5729 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
5731 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5732 arc_hdr_clear_flags(hdr,
5733 ARC_FLAG_PREDICTIVE_PREFETCH);
5736 if (*arc_flags & ARC_FLAG_WAIT) {
5737 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5738 mutex_exit(hash_lock);
5741 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5744 arc_callback_t *acb = NULL;
5746 acb = kmem_zalloc(sizeof (arc_callback_t),
5748 acb->acb_done = done;
5749 acb->acb_private = private;
5750 acb->acb_compressed = compressed_read;
5752 acb->acb_zio_dummy = zio_null(pio,
5753 spa, NULL, NULL, NULL, zio_flags);
5755 ASSERT3P(acb->acb_done, !=, NULL);
5756 acb->acb_zio_head = head_zio;
5757 acb->acb_next = hdr->b_l1hdr.b_acb;
5758 hdr->b_l1hdr.b_acb = acb;
5759 mutex_exit(hash_lock);
5762 mutex_exit(hash_lock);
5766 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5767 hdr->b_l1hdr.b_state == arc_mfu);
5770 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5772 * This is a demand read which does not have to
5773 * wait for i/o because we did a predictive
5774 * prefetch i/o for it, which has completed.
5777 arc__demand__hit__predictive__prefetch,
5778 arc_buf_hdr_t *, hdr);
5780 arcstat_demand_hit_predictive_prefetch);
5781 arc_hdr_clear_flags(hdr,
5782 ARC_FLAG_PREDICTIVE_PREFETCH);
5785 if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
5787 arcstat_demand_hit_prescient_prefetch);
5788 arc_hdr_clear_flags(hdr,
5789 ARC_FLAG_PRESCIENT_PREFETCH);
5792 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5793 /* Get a buf with the desired data in it. */
5794 rc = arc_buf_alloc_impl(hdr, private,
5795 compressed_read, B_TRUE, &buf);
5797 arc_buf_destroy(buf, private);
5800 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
5801 rc == 0 || rc != ENOENT);
5802 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
5803 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5804 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5806 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5807 arc_access(hdr, hash_lock);
5808 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5809 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5810 if (*arc_flags & ARC_FLAG_L2CACHE)
5811 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5812 mutex_exit(hash_lock);
5813 ARCSTAT_BUMP(arcstat_hits);
5814 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5815 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5816 data, metadata, hits);
5819 done(NULL, zb, bp, buf, private);
5821 uint64_t lsize = BP_GET_LSIZE(bp);
5822 uint64_t psize = BP_GET_PSIZE(bp);
5823 arc_callback_t *acb;
5826 boolean_t devw = B_FALSE;
5830 /* this block is not in the cache */
5831 arc_buf_hdr_t *exists = NULL;
5832 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5833 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5834 BP_GET_COMPRESS(bp), type);
5836 if (!BP_IS_EMBEDDED(bp)) {
5837 hdr->b_dva = *BP_IDENTITY(bp);
5838 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5839 exists = buf_hash_insert(hdr, &hash_lock);
5841 if (exists != NULL) {
5842 /* somebody beat us to the hash insert */
5843 mutex_exit(hash_lock);
5844 buf_discard_identity(hdr);
5845 arc_hdr_destroy(hdr);
5846 goto top; /* restart the IO request */
5850 * This block is in the ghost cache. If it was L2-only
5851 * (and thus didn't have an L1 hdr), we realloc the
5852 * header to add an L1 hdr.
5854 if (!HDR_HAS_L1HDR(hdr)) {
5855 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5858 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5859 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5860 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5861 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5862 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5863 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5866 * This is a delicate dance that we play here.
5867 * This hdr is in the ghost list so we access it
5868 * to move it out of the ghost list before we
5869 * initiate the read. If it's a prefetch then
5870 * it won't have a callback so we'll remove the
5871 * reference that arc_buf_alloc_impl() created. We
5872 * do this after we've called arc_access() to
5873 * avoid hitting an assert in remove_reference().
5875 arc_access(hdr, hash_lock);
5876 arc_hdr_alloc_pabd(hdr);
5878 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5879 size = arc_hdr_size(hdr);
5882 * If compression is enabled on the hdr, then will do
5883 * RAW I/O and will store the compressed data in the hdr's
5884 * data block. Otherwise, the hdr's data block will contain
5885 * the uncompressed data.
5887 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5888 zio_flags |= ZIO_FLAG_RAW;
5891 if (*arc_flags & ARC_FLAG_PREFETCH)
5892 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5893 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5894 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5896 if (*arc_flags & ARC_FLAG_L2CACHE)
5897 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5898 if (BP_GET_LEVEL(bp) > 0)
5899 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5900 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5901 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5902 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5904 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5905 acb->acb_done = done;
5906 acb->acb_private = private;
5907 acb->acb_compressed = compressed_read;
5909 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5910 hdr->b_l1hdr.b_acb = acb;
5911 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5913 if (HDR_HAS_L2HDR(hdr) &&
5914 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5915 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5916 addr = hdr->b_l2hdr.b_daddr;
5918 * Lock out L2ARC device removal.
5920 if (vdev_is_dead(vd) ||
5921 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5926 * We count both async reads and scrub IOs as asynchronous so
5927 * that both can be upgraded in the event of a cache hit while
5928 * the read IO is still in-flight.
5930 if (priority == ZIO_PRIORITY_ASYNC_READ ||
5931 priority == ZIO_PRIORITY_SCRUB)
5932 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5934 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5937 * At this point, we have a level 1 cache miss. Try again in
5938 * L2ARC if possible.
5940 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5942 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5943 uint64_t, lsize, zbookmark_phys_t *, zb);
5944 ARCSTAT_BUMP(arcstat_misses);
5945 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5946 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5947 data, metadata, misses);
5952 racct_add_force(curproc, RACCT_READBPS, size);
5953 racct_add_force(curproc, RACCT_READIOPS, 1);
5954 PROC_UNLOCK(curproc);
5957 curthread->td_ru.ru_inblock++;
5960 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5962 * Read from the L2ARC if the following are true:
5963 * 1. The L2ARC vdev was previously cached.
5964 * 2. This buffer still has L2ARC metadata.
5965 * 3. This buffer isn't currently writing to the L2ARC.
5966 * 4. The L2ARC entry wasn't evicted, which may
5967 * also have invalidated the vdev.
5968 * 5. This isn't prefetch and l2arc_noprefetch is set.
5970 if (HDR_HAS_L2HDR(hdr) &&
5971 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5972 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5973 l2arc_read_callback_t *cb;
5977 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5978 ARCSTAT_BUMP(arcstat_l2_hits);
5979 atomic_inc_32(&hdr->b_l2hdr.b_hits);
5981 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5983 cb->l2rcb_hdr = hdr;
5986 cb->l2rcb_flags = zio_flags;
5988 asize = vdev_psize_to_asize(vd, size);
5989 if (asize != size) {
5990 abd = abd_alloc_for_io(asize,
5991 HDR_ISTYPE_METADATA(hdr));
5992 cb->l2rcb_abd = abd;
5994 abd = hdr->b_l1hdr.b_pabd;
5997 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5998 addr + asize <= vd->vdev_psize -
5999 VDEV_LABEL_END_SIZE);
6002 * l2arc read. The SCL_L2ARC lock will be
6003 * released by l2arc_read_done().
6004 * Issue a null zio if the underlying buffer
6005 * was squashed to zero size by compression.
6007 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
6008 ZIO_COMPRESS_EMPTY);
6009 rzio = zio_read_phys(pio, vd, addr,
6012 l2arc_read_done, cb, priority,
6013 zio_flags | ZIO_FLAG_DONT_CACHE |
6015 ZIO_FLAG_DONT_PROPAGATE |
6016 ZIO_FLAG_DONT_RETRY, B_FALSE);
6017 acb->acb_zio_head = rzio;
6019 if (hash_lock != NULL)
6020 mutex_exit(hash_lock);
6022 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
6024 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
6026 if (*arc_flags & ARC_FLAG_NOWAIT) {
6031 ASSERT(*arc_flags & ARC_FLAG_WAIT);
6032 if (zio_wait(rzio) == 0)
6035 /* l2arc read error; goto zio_read() */
6036 if (hash_lock != NULL)
6037 mutex_enter(hash_lock);
6039 DTRACE_PROBE1(l2arc__miss,
6040 arc_buf_hdr_t *, hdr);
6041 ARCSTAT_BUMP(arcstat_l2_misses);
6042 if (HDR_L2_WRITING(hdr))
6043 ARCSTAT_BUMP(arcstat_l2_rw_clash);
6044 spa_config_exit(spa, SCL_L2ARC, vd);
6048 spa_config_exit(spa, SCL_L2ARC, vd);
6049 if (l2arc_ndev != 0) {
6050 DTRACE_PROBE1(l2arc__miss,
6051 arc_buf_hdr_t *, hdr);
6052 ARCSTAT_BUMP(arcstat_l2_misses);
6056 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size,
6057 arc_read_done, hdr, priority, zio_flags, zb);
6058 acb->acb_zio_head = rzio;
6060 if (hash_lock != NULL)
6061 mutex_exit(hash_lock);
6063 if (*arc_flags & ARC_FLAG_WAIT)
6064 return (zio_wait(rzio));
6066 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6073 arc_add_prune_callback(arc_prune_func_t *func, void *private)
6077 p = kmem_alloc(sizeof (*p), KM_SLEEP);
6079 p->p_private = private;
6080 list_link_init(&p->p_node);
6081 refcount_create(&p->p_refcnt);
6083 mutex_enter(&arc_prune_mtx);
6084 refcount_add(&p->p_refcnt, &arc_prune_list);
6085 list_insert_head(&arc_prune_list, p);
6086 mutex_exit(&arc_prune_mtx);
6092 arc_remove_prune_callback(arc_prune_t *p)
6094 boolean_t wait = B_FALSE;
6095 mutex_enter(&arc_prune_mtx);
6096 list_remove(&arc_prune_list, p);
6097 if (refcount_remove(&p->p_refcnt, &arc_prune_list) > 0)
6099 mutex_exit(&arc_prune_mtx);
6101 /* wait for arc_prune_task to finish */
6103 taskq_wait(arc_prune_taskq);
6104 ASSERT0(refcount_count(&p->p_refcnt));
6105 refcount_destroy(&p->p_refcnt);
6106 kmem_free(p, sizeof (*p));
6110 * Notify the arc that a block was freed, and thus will never be used again.
6113 arc_freed(spa_t *spa, const blkptr_t *bp)
6116 kmutex_t *hash_lock;
6117 uint64_t guid = spa_load_guid(spa);
6119 ASSERT(!BP_IS_EMBEDDED(bp));
6121 hdr = buf_hash_find(guid, bp, &hash_lock);
6126 * We might be trying to free a block that is still doing I/O
6127 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
6128 * dmu_sync-ed block). If this block is being prefetched, then it
6129 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
6130 * until the I/O completes. A block may also have a reference if it is
6131 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6132 * have written the new block to its final resting place on disk but
6133 * without the dedup flag set. This would have left the hdr in the MRU
6134 * state and discoverable. When the txg finally syncs it detects that
6135 * the block was overridden in open context and issues an override I/O.
6136 * Since this is a dedup block, the override I/O will determine if the
6137 * block is already in the DDT. If so, then it will replace the io_bp
6138 * with the bp from the DDT and allow the I/O to finish. When the I/O
6139 * reaches the done callback, dbuf_write_override_done, it will
6140 * check to see if the io_bp and io_bp_override are identical.
6141 * If they are not, then it indicates that the bp was replaced with
6142 * the bp in the DDT and the override bp is freed. This allows
6143 * us to arrive here with a reference on a block that is being
6144 * freed. So if we have an I/O in progress, or a reference to
6145 * this hdr, then we don't destroy the hdr.
6147 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
6148 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
6149 arc_change_state(arc_anon, hdr, hash_lock);
6150 arc_hdr_destroy(hdr);
6151 mutex_exit(hash_lock);
6153 mutex_exit(hash_lock);
6159 * Release this buffer from the cache, making it an anonymous buffer. This
6160 * must be done after a read and prior to modifying the buffer contents.
6161 * If the buffer has more than one reference, we must make
6162 * a new hdr for the buffer.
6165 arc_release(arc_buf_t *buf, void *tag)
6167 arc_buf_hdr_t *hdr = buf->b_hdr;
6170 * It would be nice to assert that if it's DMU metadata (level >
6171 * 0 || it's the dnode file), then it must be syncing context.
6172 * But we don't know that information at this level.
6175 mutex_enter(&buf->b_evict_lock);
6177 ASSERT(HDR_HAS_L1HDR(hdr));
6180 * We don't grab the hash lock prior to this check, because if
6181 * the buffer's header is in the arc_anon state, it won't be
6182 * linked into the hash table.
6184 if (hdr->b_l1hdr.b_state == arc_anon) {
6185 mutex_exit(&buf->b_evict_lock);
6186 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6187 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6188 ASSERT(!HDR_HAS_L2HDR(hdr));
6189 ASSERT(HDR_EMPTY(hdr));
6190 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6191 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6192 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
6194 hdr->b_l1hdr.b_arc_access = 0;
6197 * If the buf is being overridden then it may already
6198 * have a hdr that is not empty.
6200 buf_discard_identity(hdr);
6206 kmutex_t *hash_lock = HDR_LOCK(hdr);
6207 mutex_enter(hash_lock);
6210 * This assignment is only valid as long as the hash_lock is
6211 * held, we must be careful not to reference state or the
6212 * b_state field after dropping the lock.
6214 arc_state_t *state = hdr->b_l1hdr.b_state;
6215 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6216 ASSERT3P(state, !=, arc_anon);
6218 /* this buffer is not on any list */
6219 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6221 if (HDR_HAS_L2HDR(hdr)) {
6222 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6225 * We have to recheck this conditional again now that
6226 * we're holding the l2ad_mtx to prevent a race with
6227 * another thread which might be concurrently calling
6228 * l2arc_evict(). In that case, l2arc_evict() might have
6229 * destroyed the header's L2 portion as we were waiting
6230 * to acquire the l2ad_mtx.
6232 if (HDR_HAS_L2HDR(hdr)) {
6234 arc_hdr_l2hdr_destroy(hdr);
6237 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6241 * Do we have more than one buf?
6243 if (hdr->b_l1hdr.b_bufcnt > 1) {
6244 arc_buf_hdr_t *nhdr;
6245 uint64_t spa = hdr->b_spa;
6246 uint64_t psize = HDR_GET_PSIZE(hdr);
6247 uint64_t lsize = HDR_GET_LSIZE(hdr);
6248 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
6249 arc_buf_contents_t type = arc_buf_type(hdr);
6250 VERIFY3U(hdr->b_type, ==, type);
6252 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6253 (void) remove_reference(hdr, hash_lock, tag);
6255 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
6256 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6257 ASSERT(ARC_BUF_LAST(buf));
6261 * Pull the data off of this hdr and attach it to
6262 * a new anonymous hdr. Also find the last buffer
6263 * in the hdr's buffer list.
6265 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6266 ASSERT3P(lastbuf, !=, NULL);
6269 * If the current arc_buf_t and the hdr are sharing their data
6270 * buffer, then we must stop sharing that block.
6272 if (arc_buf_is_shared(buf)) {
6273 VERIFY(!arc_buf_is_shared(lastbuf));
6276 * First, sever the block sharing relationship between
6277 * buf and the arc_buf_hdr_t.
6279 arc_unshare_buf(hdr, buf);
6282 * Now we need to recreate the hdr's b_pabd. Since we
6283 * have lastbuf handy, we try to share with it, but if
6284 * we can't then we allocate a new b_pabd and copy the
6285 * data from buf into it.
6287 if (arc_can_share(hdr, lastbuf)) {
6288 arc_share_buf(hdr, lastbuf);
6290 arc_hdr_alloc_pabd(hdr);
6291 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6292 buf->b_data, psize);
6294 VERIFY3P(lastbuf->b_data, !=, NULL);
6295 } else if (HDR_SHARED_DATA(hdr)) {
6297 * Uncompressed shared buffers are always at the end
6298 * of the list. Compressed buffers don't have the
6299 * same requirements. This makes it hard to
6300 * simply assert that the lastbuf is shared so
6301 * we rely on the hdr's compression flags to determine
6302 * if we have a compressed, shared buffer.
6304 ASSERT(arc_buf_is_shared(lastbuf) ||
6305 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
6306 ASSERT(!ARC_BUF_SHARED(buf));
6308 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
6309 ASSERT3P(state, !=, arc_l2c_only);
6311 (void) refcount_remove_many(&state->arcs_size,
6312 arc_buf_size(buf), buf);
6314 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6315 ASSERT3P(state, !=, arc_l2c_only);
6316 (void) refcount_remove_many(&state->arcs_esize[type],
6317 arc_buf_size(buf), buf);
6320 hdr->b_l1hdr.b_bufcnt -= 1;
6321 arc_cksum_verify(buf);
6323 arc_buf_unwatch(buf);
6326 mutex_exit(hash_lock);
6329 * Allocate a new hdr. The new hdr will contain a b_pabd
6330 * buffer which will be freed in arc_write().
6332 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
6333 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6334 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6335 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
6336 VERIFY3U(nhdr->b_type, ==, type);
6337 ASSERT(!HDR_SHARED_DATA(nhdr));
6339 nhdr->b_l1hdr.b_buf = buf;
6340 nhdr->b_l1hdr.b_bufcnt = 1;
6341 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6344 mutex_exit(&buf->b_evict_lock);
6345 (void) refcount_add_many(&arc_anon->arcs_size,
6346 arc_buf_size(buf), buf);
6348 mutex_exit(&buf->b_evict_lock);
6349 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6350 /* protected by hash lock, or hdr is on arc_anon */
6351 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6352 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6353 arc_change_state(arc_anon, hdr, hash_lock);
6354 hdr->b_l1hdr.b_arc_access = 0;
6355 mutex_exit(hash_lock);
6357 buf_discard_identity(hdr);
6363 arc_released(arc_buf_t *buf)
6367 mutex_enter(&buf->b_evict_lock);
6368 released = (buf->b_data != NULL &&
6369 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6370 mutex_exit(&buf->b_evict_lock);
6376 arc_referenced(arc_buf_t *buf)
6380 mutex_enter(&buf->b_evict_lock);
6381 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6382 mutex_exit(&buf->b_evict_lock);
6383 return (referenced);
6388 arc_write_ready(zio_t *zio)
6390 arc_write_callback_t *callback = zio->io_private;
6391 arc_buf_t *buf = callback->awcb_buf;
6392 arc_buf_hdr_t *hdr = buf->b_hdr;
6393 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
6395 ASSERT(HDR_HAS_L1HDR(hdr));
6396 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6397 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6400 * If we're reexecuting this zio because the pool suspended, then
6401 * cleanup any state that was previously set the first time the
6402 * callback was invoked.
6404 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6405 arc_cksum_free(hdr);
6407 arc_buf_unwatch(buf);
6409 if (hdr->b_l1hdr.b_pabd != NULL) {
6410 if (arc_buf_is_shared(buf)) {
6411 arc_unshare_buf(hdr, buf);
6413 arc_hdr_free_pabd(hdr);
6417 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6418 ASSERT(!HDR_SHARED_DATA(hdr));
6419 ASSERT(!arc_buf_is_shared(buf));
6421 callback->awcb_ready(zio, buf, callback->awcb_private);
6423 if (HDR_IO_IN_PROGRESS(hdr))
6424 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6426 arc_cksum_compute(buf);
6427 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6429 enum zio_compress compress;
6430 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6431 compress = ZIO_COMPRESS_OFF;
6433 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
6434 compress = BP_GET_COMPRESS(zio->io_bp);
6436 HDR_SET_PSIZE(hdr, psize);
6437 arc_hdr_set_compress(hdr, compress);
6441 * Fill the hdr with data. If the hdr is compressed, the data we want
6442 * is available from the zio, otherwise we can take it from the buf.
6444 * We might be able to share the buf's data with the hdr here. However,
6445 * doing so would cause the ARC to be full of linear ABDs if we write a
6446 * lot of shareable data. As a compromise, we check whether scattered
6447 * ABDs are allowed, and assume that if they are then the user wants
6448 * the ARC to be primarily filled with them regardless of the data being
6449 * written. Therefore, if they're allowed then we allocate one and copy
6450 * the data into it; otherwise, we share the data directly if we can.
6452 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
6453 arc_hdr_alloc_pabd(hdr);
6456 * Ideally, we would always copy the io_abd into b_pabd, but the
6457 * user may have disabled compressed ARC, thus we must check the
6458 * hdr's compression setting rather than the io_bp's.
6460 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
6461 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=,
6463 ASSERT3U(psize, >, 0);
6465 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6467 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6469 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6473 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6474 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6475 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6477 arc_share_buf(hdr, buf);
6480 arc_hdr_verify(hdr, zio->io_bp);
6484 arc_write_children_ready(zio_t *zio)
6486 arc_write_callback_t *callback = zio->io_private;
6487 arc_buf_t *buf = callback->awcb_buf;
6489 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6493 * The SPA calls this callback for each physical write that happens on behalf
6494 * of a logical write. See the comment in dbuf_write_physdone() for details.
6497 arc_write_physdone(zio_t *zio)
6499 arc_write_callback_t *cb = zio->io_private;
6500 if (cb->awcb_physdone != NULL)
6501 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
6505 arc_write_done(zio_t *zio)
6507 arc_write_callback_t *callback = zio->io_private;
6508 arc_buf_t *buf = callback->awcb_buf;
6509 arc_buf_hdr_t *hdr = buf->b_hdr;
6511 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6513 if (zio->io_error == 0) {
6514 arc_hdr_verify(hdr, zio->io_bp);
6516 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6517 buf_discard_identity(hdr);
6519 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6520 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6523 ASSERT(HDR_EMPTY(hdr));
6527 * If the block to be written was all-zero or compressed enough to be
6528 * embedded in the BP, no write was performed so there will be no
6529 * dva/birth/checksum. The buffer must therefore remain anonymous
6532 if (!HDR_EMPTY(hdr)) {
6533 arc_buf_hdr_t *exists;
6534 kmutex_t *hash_lock;
6536 ASSERT3U(zio->io_error, ==, 0);
6538 arc_cksum_verify(buf);
6540 exists = buf_hash_insert(hdr, &hash_lock);
6541 if (exists != NULL) {
6543 * This can only happen if we overwrite for
6544 * sync-to-convergence, because we remove
6545 * buffers from the hash table when we arc_free().
6547 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6548 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6549 panic("bad overwrite, hdr=%p exists=%p",
6550 (void *)hdr, (void *)exists);
6551 ASSERT(refcount_is_zero(
6552 &exists->b_l1hdr.b_refcnt));
6553 arc_change_state(arc_anon, exists, hash_lock);
6554 mutex_exit(hash_lock);
6555 arc_hdr_destroy(exists);
6556 exists = buf_hash_insert(hdr, &hash_lock);
6557 ASSERT3P(exists, ==, NULL);
6558 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6560 ASSERT(zio->io_prop.zp_nopwrite);
6561 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6562 panic("bad nopwrite, hdr=%p exists=%p",
6563 (void *)hdr, (void *)exists);
6566 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6567 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6568 ASSERT(BP_GET_DEDUP(zio->io_bp));
6569 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6572 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6573 /* if it's not anon, we are doing a scrub */
6574 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6575 arc_access(hdr, hash_lock);
6576 mutex_exit(hash_lock);
6578 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6581 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6582 callback->awcb_done(zio, buf, callback->awcb_private);
6584 abd_put(zio->io_abd);
6585 kmem_free(callback, sizeof (arc_write_callback_t));
6589 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
6590 boolean_t l2arc, const zio_prop_t *zp, arc_write_done_func_t *ready,
6591 arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
6592 arc_write_done_func_t *done, void *private, zio_priority_t priority,
6593 int zio_flags, const zbookmark_phys_t *zb)
6595 arc_buf_hdr_t *hdr = buf->b_hdr;
6596 arc_write_callback_t *callback;
6598 zio_prop_t localprop = *zp;
6600 ASSERT3P(ready, !=, NULL);
6601 ASSERT3P(done, !=, NULL);
6602 ASSERT(!HDR_IO_ERROR(hdr));
6603 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6604 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6605 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6607 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6608 if (ARC_BUF_COMPRESSED(buf)) {
6610 * We're writing a pre-compressed buffer. Make the
6611 * compression algorithm requested by the zio_prop_t match
6612 * the pre-compressed buffer's compression algorithm.
6614 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6616 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6617 zio_flags |= ZIO_FLAG_RAW;
6619 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6620 callback->awcb_ready = ready;
6621 callback->awcb_children_ready = children_ready;
6622 callback->awcb_physdone = physdone;
6623 callback->awcb_done = done;
6624 callback->awcb_private = private;
6625 callback->awcb_buf = buf;
6628 * The hdr's b_pabd is now stale, free it now. A new data block
6629 * will be allocated when the zio pipeline calls arc_write_ready().
6631 if (hdr->b_l1hdr.b_pabd != NULL) {
6633 * If the buf is currently sharing the data block with
6634 * the hdr then we need to break that relationship here.
6635 * The hdr will remain with a NULL data pointer and the
6636 * buf will take sole ownership of the block.
6638 if (arc_buf_is_shared(buf)) {
6639 arc_unshare_buf(hdr, buf);
6641 arc_hdr_free_pabd(hdr);
6643 VERIFY3P(buf->b_data, !=, NULL);
6644 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6646 ASSERT(!arc_buf_is_shared(buf));
6647 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6649 zio = zio_write(pio, spa, txg, bp,
6650 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6651 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6652 (children_ready != NULL) ? arc_write_children_ready : NULL,
6653 arc_write_physdone, arc_write_done, callback,
6654 priority, zio_flags, zb);
6660 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
6663 uint64_t available_memory = ptob(freemem);
6665 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
6666 available_memory = MIN(available_memory, uma_avail());
6669 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
6672 if (txg > spa->spa_lowmem_last_txg) {
6673 spa->spa_lowmem_last_txg = txg;
6674 spa->spa_lowmem_page_load = 0;
6677 * If we are in pageout, we know that memory is already tight,
6678 * the arc is already going to be evicting, so we just want to
6679 * continue to let page writes occur as quickly as possible.
6681 if (curproc == pageproc) {
6682 if (spa->spa_lowmem_page_load >
6683 MAX(ptob(minfree), available_memory) / 4)
6684 return (SET_ERROR(ERESTART));
6685 /* Note: reserve is inflated, so we deflate */
6686 atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8);
6688 } else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) {
6689 /* memory is low, delay before restarting */
6690 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
6691 return (SET_ERROR(EAGAIN));
6693 spa->spa_lowmem_page_load = 0;
6694 #endif /* _KERNEL */
6699 arc_tempreserve_clear(uint64_t reserve)
6701 atomic_add_64(&arc_tempreserve, -reserve);
6702 ASSERT((int64_t)arc_tempreserve >= 0);
6706 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
6711 if (reserve > arc_c/4 && !arc_no_grow) {
6712 arc_c = MIN(arc_c_max, reserve * 4);
6713 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
6715 if (reserve > arc_c)
6716 return (SET_ERROR(ENOMEM));
6719 * Don't count loaned bufs as in flight dirty data to prevent long
6720 * network delays from blocking transactions that are ready to be
6721 * assigned to a txg.
6724 /* assert that it has not wrapped around */
6725 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6727 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
6728 arc_loaned_bytes), 0);
6731 * Writes will, almost always, require additional memory allocations
6732 * in order to compress/encrypt/etc the data. We therefore need to
6733 * make sure that there is sufficient available memory for this.
6735 error = arc_memory_throttle(spa, reserve, txg);
6740 * Throttle writes when the amount of dirty data in the cache
6741 * gets too large. We try to keep the cache less than half full
6742 * of dirty blocks so that our sync times don't grow too large.
6744 * In the case of one pool being built on another pool, we want
6745 * to make sure we don't end up throttling the lower (backing)
6746 * pool when the upper pool is the majority contributor to dirty
6747 * data. To insure we make forward progress during throttling, we
6748 * also check the current pool's net dirty data and only throttle
6749 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
6750 * data in the cache.
6752 * Note: if two requests come in concurrently, we might let them
6753 * both succeed, when one of them should fail. Not a huge deal.
6755 uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
6756 uint64_t spa_dirty_anon = spa_dirty_data(spa);
6758 if (total_dirty > arc_c * zfs_arc_dirty_limit_percent / 100 &&
6759 anon_size > arc_c * zfs_arc_anon_limit_percent / 100 &&
6760 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
6761 uint64_t meta_esize =
6762 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6763 uint64_t data_esize =
6764 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6765 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6766 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6767 arc_tempreserve >> 10, meta_esize >> 10,
6768 data_esize >> 10, reserve >> 10, arc_c >> 10);
6769 return (SET_ERROR(ERESTART));
6771 atomic_add_64(&arc_tempreserve, reserve);
6776 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6777 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6779 size->value.ui64 = refcount_count(&state->arcs_size);
6780 evict_data->value.ui64 =
6781 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6782 evict_metadata->value.ui64 =
6783 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6787 arc_kstat_update(kstat_t *ksp, int rw)
6789 arc_stats_t *as = ksp->ks_data;
6791 if (rw == KSTAT_WRITE) {
6794 arc_kstat_update_state(arc_anon,
6795 &as->arcstat_anon_size,
6796 &as->arcstat_anon_evictable_data,
6797 &as->arcstat_anon_evictable_metadata);
6798 arc_kstat_update_state(arc_mru,
6799 &as->arcstat_mru_size,
6800 &as->arcstat_mru_evictable_data,
6801 &as->arcstat_mru_evictable_metadata);
6802 arc_kstat_update_state(arc_mru_ghost,
6803 &as->arcstat_mru_ghost_size,
6804 &as->arcstat_mru_ghost_evictable_data,
6805 &as->arcstat_mru_ghost_evictable_metadata);
6806 arc_kstat_update_state(arc_mfu,
6807 &as->arcstat_mfu_size,
6808 &as->arcstat_mfu_evictable_data,
6809 &as->arcstat_mfu_evictable_metadata);
6810 arc_kstat_update_state(arc_mfu_ghost,
6811 &as->arcstat_mfu_ghost_size,
6812 &as->arcstat_mfu_ghost_evictable_data,
6813 &as->arcstat_mfu_ghost_evictable_metadata);
6815 ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
6816 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
6817 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
6818 ARCSTAT(arcstat_metadata_size) =
6819 aggsum_value(&astat_metadata_size);
6820 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
6821 ARCSTAT(arcstat_bonus_size) = aggsum_value(&astat_bonus_size);
6822 ARCSTAT(arcstat_dnode_size) = aggsum_value(&astat_dnode_size);
6823 ARCSTAT(arcstat_dbuf_size) = aggsum_value(&astat_dbuf_size);
6824 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
6831 * This function *must* return indices evenly distributed between all
6832 * sublists of the multilist. This is needed due to how the ARC eviction
6833 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6834 * distributed between all sublists and uses this assumption when
6835 * deciding which sublist to evict from and how much to evict from it.
6838 arc_state_multilist_index_func(multilist_t *ml, void *obj)
6840 arc_buf_hdr_t *hdr = obj;
6843 * We rely on b_dva to generate evenly distributed index
6844 * numbers using buf_hash below. So, as an added precaution,
6845 * let's make sure we never add empty buffers to the arc lists.
6847 ASSERT(!HDR_EMPTY(hdr));
6850 * The assumption here, is the hash value for a given
6851 * arc_buf_hdr_t will remain constant throughout it's lifetime
6852 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
6853 * Thus, we don't need to store the header's sublist index
6854 * on insertion, as this index can be recalculated on removal.
6856 * Also, the low order bits of the hash value are thought to be
6857 * distributed evenly. Otherwise, in the case that the multilist
6858 * has a power of two number of sublists, each sublists' usage
6859 * would not be evenly distributed.
6861 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
6862 multilist_get_num_sublists(ml));
6866 static eventhandler_tag arc_event_lowmem = NULL;
6869 arc_lowmem(void *arg __unused, int howto __unused)
6872 mutex_enter(&arc_reclaim_lock);
6873 DTRACE_PROBE1(arc__needfree, int64_t, ((int64_t)freemem - zfs_arc_free_target) * PAGESIZE);
6874 cv_signal(&arc_reclaim_thread_cv);
6877 * It is unsafe to block here in arbitrary threads, because we can come
6878 * here from ARC itself and may hold ARC locks and thus risk a deadlock
6879 * with ARC reclaim thread.
6881 if (curproc == pageproc)
6882 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
6883 mutex_exit(&arc_reclaim_lock);
6888 arc_state_init(void)
6890 arc_anon = &ARC_anon;
6892 arc_mru_ghost = &ARC_mru_ghost;
6894 arc_mfu_ghost = &ARC_mfu_ghost;
6895 arc_l2c_only = &ARC_l2c_only;
6897 arc_mru->arcs_list[ARC_BUFC_METADATA] =
6898 multilist_create(sizeof (arc_buf_hdr_t),
6899 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6900 arc_state_multilist_index_func);
6901 arc_mru->arcs_list[ARC_BUFC_DATA] =
6902 multilist_create(sizeof (arc_buf_hdr_t),
6903 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6904 arc_state_multilist_index_func);
6905 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
6906 multilist_create(sizeof (arc_buf_hdr_t),
6907 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6908 arc_state_multilist_index_func);
6909 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
6910 multilist_create(sizeof (arc_buf_hdr_t),
6911 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6912 arc_state_multilist_index_func);
6913 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
6914 multilist_create(sizeof (arc_buf_hdr_t),
6915 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6916 arc_state_multilist_index_func);
6917 arc_mfu->arcs_list[ARC_BUFC_DATA] =
6918 multilist_create(sizeof (arc_buf_hdr_t),
6919 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6920 arc_state_multilist_index_func);
6921 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
6922 multilist_create(sizeof (arc_buf_hdr_t),
6923 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6924 arc_state_multilist_index_func);
6925 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
6926 multilist_create(sizeof (arc_buf_hdr_t),
6927 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6928 arc_state_multilist_index_func);
6929 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
6930 multilist_create(sizeof (arc_buf_hdr_t),
6931 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6932 arc_state_multilist_index_func);
6933 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
6934 multilist_create(sizeof (arc_buf_hdr_t),
6935 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6936 arc_state_multilist_index_func);
6938 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6939 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6940 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6941 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6942 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6943 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6944 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6945 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6946 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6947 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6948 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6949 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6951 refcount_create(&arc_anon->arcs_size);
6952 refcount_create(&arc_mru->arcs_size);
6953 refcount_create(&arc_mru_ghost->arcs_size);
6954 refcount_create(&arc_mfu->arcs_size);
6955 refcount_create(&arc_mfu_ghost->arcs_size);
6956 refcount_create(&arc_l2c_only->arcs_size);
6958 aggsum_init(&arc_meta_used, 0);
6959 aggsum_init(&arc_size, 0);
6960 aggsum_init(&astat_data_size, 0);
6961 aggsum_init(&astat_metadata_size, 0);
6962 aggsum_init(&astat_hdr_size, 0);
6963 aggsum_init(&astat_bonus_size, 0);
6964 aggsum_init(&astat_dnode_size, 0);
6965 aggsum_init(&astat_dbuf_size, 0);
6966 aggsum_init(&astat_l2_hdr_size, 0);
6970 arc_state_fini(void)
6972 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6973 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6974 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6975 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6976 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6977 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6978 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6979 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6980 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6981 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6982 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6983 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6985 refcount_destroy(&arc_anon->arcs_size);
6986 refcount_destroy(&arc_mru->arcs_size);
6987 refcount_destroy(&arc_mru_ghost->arcs_size);
6988 refcount_destroy(&arc_mfu->arcs_size);
6989 refcount_destroy(&arc_mfu_ghost->arcs_size);
6990 refcount_destroy(&arc_l2c_only->arcs_size);
6992 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
6993 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
6994 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
6995 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
6996 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
6997 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
6998 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
6999 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7011 int i, prefetch_tunable_set = 0;
7014 * allmem is "all memory that we could possibly use".
7018 uint64_t allmem = ptob(physmem - swapfs_minfree);
7020 uint64_t allmem = (physmem * PAGESIZE) / 2;
7023 uint64_t allmem = kmem_size();
7027 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
7028 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
7029 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
7031 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
7032 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL);
7034 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
7035 arc_c_min = MAX(allmem / 32, arc_abs_min);
7036 /* set max to 5/8 of all memory, or all but 1GB, whichever is more */
7037 if (allmem >= 1 << 30)
7038 arc_c_max = allmem - (1 << 30);
7040 arc_c_max = arc_c_min;
7041 arc_c_max = MAX(allmem * 5 / 8, arc_c_max);
7044 * In userland, there's only the memory pressure that we artificially
7045 * create (see arc_available_memory()). Don't let arc_c get too
7046 * small, because it can cause transactions to be larger than
7047 * arc_c, causing arc_tempreserve_space() to fail.
7050 arc_c_min = arc_c_max / 2;
7055 * Allow the tunables to override our calculations if they are
7058 if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) {
7059 arc_c_max = zfs_arc_max;
7060 arc_c_min = MIN(arc_c_min, arc_c_max);
7062 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
7063 arc_c_min = zfs_arc_min;
7067 arc_p = (arc_c >> 1);
7069 /* limit meta-data to 1/4 of the arc capacity */
7070 arc_meta_limit = arc_c_max / 4;
7074 * Metadata is stored in the kernel's heap. Don't let us
7075 * use more than half the heap for the ARC.
7078 arc_meta_limit = MIN(arc_meta_limit, uma_limit() / 2);
7079 arc_dnode_limit = arc_meta_limit / 10;
7081 arc_meta_limit = MIN(arc_meta_limit,
7082 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
7086 /* Allow the tunable to override if it is reasonable */
7087 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
7088 arc_meta_limit = zfs_arc_meta_limit;
7090 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
7091 arc_c_min = arc_meta_limit / 2;
7093 if (zfs_arc_meta_min > 0) {
7094 arc_meta_min = zfs_arc_meta_min;
7096 arc_meta_min = arc_c_min / 2;
7099 /* Valid range: <arc_meta_min> - <arc_c_max> */
7100 if ((zfs_arc_dnode_limit) && (zfs_arc_dnode_limit != arc_dnode_limit) &&
7101 (zfs_arc_dnode_limit >= zfs_arc_meta_min) &&
7102 (zfs_arc_dnode_limit <= arc_c_max))
7103 arc_dnode_limit = zfs_arc_dnode_limit;
7105 if (zfs_arc_grow_retry > 0)
7106 arc_grow_retry = zfs_arc_grow_retry;
7108 if (zfs_arc_shrink_shift > 0)
7109 arc_shrink_shift = zfs_arc_shrink_shift;
7111 if (zfs_arc_no_grow_shift > 0)
7112 arc_no_grow_shift = zfs_arc_no_grow_shift;
7114 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
7116 if (arc_no_grow_shift >= arc_shrink_shift)
7117 arc_no_grow_shift = arc_shrink_shift - 1;
7119 if (zfs_arc_p_min_shift > 0)
7120 arc_p_min_shift = zfs_arc_p_min_shift;
7122 /* if kmem_flags are set, lets try to use less memory */
7123 if (kmem_debugging())
7125 if (arc_c < arc_c_min)
7128 zfs_arc_min = arc_c_min;
7129 zfs_arc_max = arc_c_max;
7134 list_create(&arc_prune_list, sizeof (arc_prune_t),
7135 offsetof(arc_prune_t, p_node));
7136 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
7138 arc_prune_taskq = taskq_create("arc_prune", max_ncpus, minclsyspri,
7139 max_ncpus, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
7141 arc_reclaim_thread_exit = B_FALSE;
7142 arc_dnlc_evicts_thread_exit = FALSE;
7144 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
7145 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
7147 if (arc_ksp != NULL) {
7148 arc_ksp->ks_data = &arc_stats;
7149 arc_ksp->ks_update = arc_kstat_update;
7150 kstat_install(arc_ksp);
7153 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
7154 TS_RUN, minclsyspri);
7157 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
7158 EVENTHANDLER_PRI_FIRST);
7161 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0,
7162 TS_RUN, minclsyspri);
7168 * Calculate maximum amount of dirty data per pool.
7170 * If it has been set by /etc/system, take that.
7171 * Otherwise, use a percentage of physical memory defined by
7172 * zfs_dirty_data_max_percent (default 10%) with a cap at
7173 * zfs_dirty_data_max_max (default 4GB).
7175 if (zfs_dirty_data_max == 0) {
7176 zfs_dirty_data_max = ptob(physmem) *
7177 zfs_dirty_data_max_percent / 100;
7178 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
7179 zfs_dirty_data_max_max);
7183 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
7184 prefetch_tunable_set = 1;
7187 if (prefetch_tunable_set == 0) {
7188 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
7190 printf(" add \"vfs.zfs.prefetch_disable=0\" "
7191 "to /boot/loader.conf.\n");
7192 zfs_prefetch_disable = 1;
7195 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
7196 prefetch_tunable_set == 0) {
7197 printf("ZFS NOTICE: Prefetch is disabled by default if less "
7198 "than 4GB of RAM is present;\n"
7199 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
7200 "to /boot/loader.conf.\n");
7201 zfs_prefetch_disable = 1;
7204 /* Warn about ZFS memory and address space requirements. */
7205 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
7206 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
7207 "expect unstable behavior.\n");
7209 if (allmem < 512 * (1 << 20)) {
7210 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
7211 "expect unstable behavior.\n");
7212 printf(" Consider tuning vm.kmem_size and "
7213 "vm.kmem_size_max\n");
7214 printf(" in /boot/loader.conf.\n");
7225 if (arc_event_lowmem != NULL)
7226 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
7229 mutex_enter(&arc_reclaim_lock);
7230 arc_reclaim_thread_exit = B_TRUE;
7232 * The reclaim thread will set arc_reclaim_thread_exit back to
7233 * B_FALSE when it is finished exiting; we're waiting for that.
7235 while (arc_reclaim_thread_exit) {
7236 cv_signal(&arc_reclaim_thread_cv);
7237 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
7239 mutex_exit(&arc_reclaim_lock);
7241 /* Use B_TRUE to ensure *all* buffers are evicted */
7242 arc_flush(NULL, B_TRUE);
7244 mutex_enter(&arc_dnlc_evicts_lock);
7245 arc_dnlc_evicts_thread_exit = TRUE;
7247 * The user evicts thread will set arc_user_evicts_thread_exit
7248 * to FALSE when it is finished exiting; we're waiting for that.
7250 while (arc_dnlc_evicts_thread_exit) {
7251 cv_signal(&arc_dnlc_evicts_cv);
7252 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
7254 mutex_exit(&arc_dnlc_evicts_lock);
7258 if (arc_ksp != NULL) {
7259 kstat_delete(arc_ksp);
7263 taskq_wait(arc_prune_taskq);
7264 taskq_destroy(arc_prune_taskq);
7266 mutex_enter(&arc_prune_mtx);
7267 while ((p = list_head(&arc_prune_list)) != NULL) {
7268 list_remove(&arc_prune_list, p);
7269 refcount_remove(&p->p_refcnt, &arc_prune_list);
7270 refcount_destroy(&p->p_refcnt);
7271 kmem_free(p, sizeof (*p));
7273 mutex_exit(&arc_prune_mtx);
7275 list_destroy(&arc_prune_list);
7276 mutex_destroy(&arc_prune_mtx);
7277 mutex_destroy(&arc_reclaim_lock);
7278 cv_destroy(&arc_reclaim_thread_cv);
7279 cv_destroy(&arc_reclaim_waiters_cv);
7281 mutex_destroy(&arc_dnlc_evicts_lock);
7282 cv_destroy(&arc_dnlc_evicts_cv);
7287 ASSERT0(arc_loaned_bytes);
7293 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7294 * It uses dedicated storage devices to hold cached data, which are populated
7295 * using large infrequent writes. The main role of this cache is to boost
7296 * the performance of random read workloads. The intended L2ARC devices
7297 * include short-stroked disks, solid state disks, and other media with
7298 * substantially faster read latency than disk.
7300 * +-----------------------+
7302 * +-----------------------+
7305 * l2arc_feed_thread() arc_read()
7309 * +---------------+ |
7311 * +---------------+ |
7316 * +-------+ +-------+
7318 * | cache | | cache |
7319 * +-------+ +-------+
7320 * +=========+ .-----.
7321 * : L2ARC : |-_____-|
7322 * : devices : | Disks |
7323 * +=========+ `-_____-'
7325 * Read requests are satisfied from the following sources, in order:
7328 * 2) vdev cache of L2ARC devices
7330 * 4) vdev cache of disks
7333 * Some L2ARC device types exhibit extremely slow write performance.
7334 * To accommodate for this there are some significant differences between
7335 * the L2ARC and traditional cache design:
7337 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
7338 * the ARC behave as usual, freeing buffers and placing headers on ghost
7339 * lists. The ARC does not send buffers to the L2ARC during eviction as
7340 * this would add inflated write latencies for all ARC memory pressure.
7342 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7343 * It does this by periodically scanning buffers from the eviction-end of
7344 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7345 * not already there. It scans until a headroom of buffers is satisfied,
7346 * which itself is a buffer for ARC eviction. If a compressible buffer is
7347 * found during scanning and selected for writing to an L2ARC device, we
7348 * temporarily boost scanning headroom during the next scan cycle to make
7349 * sure we adapt to compression effects (which might significantly reduce
7350 * the data volume we write to L2ARC). The thread that does this is
7351 * l2arc_feed_thread(), illustrated below; example sizes are included to
7352 * provide a better sense of ratio than this diagram:
7355 * +---------------------+----------+
7356 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
7357 * +---------------------+----------+ | o L2ARC eligible
7358 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
7359 * +---------------------+----------+ |
7360 * 15.9 Gbytes ^ 32 Mbytes |
7362 * l2arc_feed_thread()
7364 * l2arc write hand <--[oooo]--'
7368 * +==============================+
7369 * L2ARC dev |####|#|###|###| |####| ... |
7370 * +==============================+
7373 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
7374 * evicted, then the L2ARC has cached a buffer much sooner than it probably
7375 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
7376 * safe to say that this is an uncommon case, since buffers at the end of
7377 * the ARC lists have moved there due to inactivity.
7379 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
7380 * then the L2ARC simply misses copying some buffers. This serves as a
7381 * pressure valve to prevent heavy read workloads from both stalling the ARC
7382 * with waits and clogging the L2ARC with writes. This also helps prevent
7383 * the potential for the L2ARC to churn if it attempts to cache content too
7384 * quickly, such as during backups of the entire pool.
7386 * 5. After system boot and before the ARC has filled main memory, there are
7387 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
7388 * lists can remain mostly static. Instead of searching from tail of these
7389 * lists as pictured, the l2arc_feed_thread() will search from the list heads
7390 * for eligible buffers, greatly increasing its chance of finding them.
7392 * The L2ARC device write speed is also boosted during this time so that
7393 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
7394 * there are no L2ARC reads, and no fear of degrading read performance
7395 * through increased writes.
7397 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
7398 * the vdev queue can aggregate them into larger and fewer writes. Each
7399 * device is written to in a rotor fashion, sweeping writes through
7400 * available space then repeating.
7402 * 7. The L2ARC does not store dirty content. It never needs to flush
7403 * write buffers back to disk based storage.
7405 * 8. If an ARC buffer is written (and dirtied) which also exists in the
7406 * L2ARC, the now stale L2ARC buffer is immediately dropped.
7408 * The performance of the L2ARC can be tweaked by a number of tunables, which
7409 * may be necessary for different workloads:
7411 * l2arc_write_max max write bytes per interval
7412 * l2arc_write_boost extra write bytes during device warmup
7413 * l2arc_noprefetch skip caching prefetched buffers
7414 * l2arc_headroom number of max device writes to precache
7415 * l2arc_headroom_boost when we find compressed buffers during ARC
7416 * scanning, we multiply headroom by this
7417 * percentage factor for the next scan cycle,
7418 * since more compressed buffers are likely to
7420 * l2arc_feed_secs seconds between L2ARC writing
7422 * Tunables may be removed or added as future performance improvements are
7423 * integrated, and also may become zpool properties.
7425 * There are three key functions that control how the L2ARC warms up:
7427 * l2arc_write_eligible() check if a buffer is eligible to cache
7428 * l2arc_write_size() calculate how much to write
7429 * l2arc_write_interval() calculate sleep delay between writes
7431 * These three functions determine what to write, how much, and how quickly
7436 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
7439 * A buffer is *not* eligible for the L2ARC if it:
7440 * 1. belongs to a different spa.
7441 * 2. is already cached on the L2ARC.
7442 * 3. has an I/O in progress (it may be an incomplete read).
7443 * 4. is flagged not eligible (zfs property).
7445 if (hdr->b_spa != spa_guid) {
7446 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
7449 if (HDR_HAS_L2HDR(hdr)) {
7450 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
7453 if (HDR_IO_IN_PROGRESS(hdr)) {
7454 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
7457 if (!HDR_L2CACHE(hdr)) {
7458 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
7466 l2arc_write_size(void)
7471 * Make sure our globals have meaningful values in case the user
7474 size = l2arc_write_max;
7476 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
7477 "be greater than zero, resetting it to the default (%d)",
7479 size = l2arc_write_max = L2ARC_WRITE_SIZE;
7482 if (arc_warm == B_FALSE)
7483 size += l2arc_write_boost;
7490 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
7492 clock_t interval, next, now;
7495 * If the ARC lists are busy, increase our write rate; if the
7496 * lists are stale, idle back. This is achieved by checking
7497 * how much we previously wrote - if it was more than half of
7498 * what we wanted, schedule the next write much sooner.
7500 if (l2arc_feed_again && wrote > (wanted / 2))
7501 interval = (hz * l2arc_feed_min_ms) / 1000;
7503 interval = hz * l2arc_feed_secs;
7505 now = ddi_get_lbolt();
7506 next = MAX(now, MIN(now + interval, began + interval));
7512 * Cycle through L2ARC devices. This is how L2ARC load balances.
7513 * If a device is returned, this also returns holding the spa config lock.
7515 static l2arc_dev_t *
7516 l2arc_dev_get_next(void)
7518 l2arc_dev_t *first, *next = NULL;
7521 * Lock out the removal of spas (spa_namespace_lock), then removal
7522 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
7523 * both locks will be dropped and a spa config lock held instead.
7525 mutex_enter(&spa_namespace_lock);
7526 mutex_enter(&l2arc_dev_mtx);
7528 /* if there are no vdevs, there is nothing to do */
7529 if (l2arc_ndev == 0)
7533 next = l2arc_dev_last;
7535 /* loop around the list looking for a non-faulted vdev */
7537 next = list_head(l2arc_dev_list);
7539 next = list_next(l2arc_dev_list, next);
7541 next = list_head(l2arc_dev_list);
7544 /* if we have come back to the start, bail out */
7547 else if (next == first)
7550 } while (vdev_is_dead(next->l2ad_vdev));
7552 /* if we were unable to find any usable vdevs, return NULL */
7553 if (vdev_is_dead(next->l2ad_vdev))
7556 l2arc_dev_last = next;
7559 mutex_exit(&l2arc_dev_mtx);
7562 * Grab the config lock to prevent the 'next' device from being
7563 * removed while we are writing to it.
7566 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
7567 mutex_exit(&spa_namespace_lock);
7573 * Free buffers that were tagged for destruction.
7576 l2arc_do_free_on_write()
7579 l2arc_data_free_t *df, *df_prev;
7581 mutex_enter(&l2arc_free_on_write_mtx);
7582 buflist = l2arc_free_on_write;
7584 for (df = list_tail(buflist); df; df = df_prev) {
7585 df_prev = list_prev(buflist, df);
7586 ASSERT3P(df->l2df_abd, !=, NULL);
7587 abd_free(df->l2df_abd);
7588 list_remove(buflist, df);
7589 kmem_free(df, sizeof (l2arc_data_free_t));
7592 mutex_exit(&l2arc_free_on_write_mtx);
7596 * A write to a cache device has completed. Update all headers to allow
7597 * reads from these buffers to begin.
7600 l2arc_write_done(zio_t *zio)
7602 l2arc_write_callback_t *cb;
7605 arc_buf_hdr_t *head, *hdr, *hdr_prev;
7606 kmutex_t *hash_lock;
7607 int64_t bytes_dropped = 0;
7609 cb = zio->io_private;
7610 ASSERT3P(cb, !=, NULL);
7611 dev = cb->l2wcb_dev;
7612 ASSERT3P(dev, !=, NULL);
7613 head = cb->l2wcb_head;
7614 ASSERT3P(head, !=, NULL);
7615 buflist = &dev->l2ad_buflist;
7616 ASSERT3P(buflist, !=, NULL);
7617 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
7618 l2arc_write_callback_t *, cb);
7620 if (zio->io_error != 0)
7621 ARCSTAT_BUMP(arcstat_l2_writes_error);
7624 * All writes completed, or an error was hit.
7627 mutex_enter(&dev->l2ad_mtx);
7628 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
7629 hdr_prev = list_prev(buflist, hdr);
7631 hash_lock = HDR_LOCK(hdr);
7634 * We cannot use mutex_enter or else we can deadlock
7635 * with l2arc_write_buffers (due to swapping the order
7636 * the hash lock and l2ad_mtx are taken).
7638 if (!mutex_tryenter(hash_lock)) {
7640 * Missed the hash lock. We must retry so we
7641 * don't leave the ARC_FLAG_L2_WRITING bit set.
7643 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
7646 * We don't want to rescan the headers we've
7647 * already marked as having been written out, so
7648 * we reinsert the head node so we can pick up
7649 * where we left off.
7651 list_remove(buflist, head);
7652 list_insert_after(buflist, hdr, head);
7654 mutex_exit(&dev->l2ad_mtx);
7657 * We wait for the hash lock to become available
7658 * to try and prevent busy waiting, and increase
7659 * the chance we'll be able to acquire the lock
7660 * the next time around.
7662 mutex_enter(hash_lock);
7663 mutex_exit(hash_lock);
7668 * We could not have been moved into the arc_l2c_only
7669 * state while in-flight due to our ARC_FLAG_L2_WRITING
7670 * bit being set. Let's just ensure that's being enforced.
7672 ASSERT(HDR_HAS_L1HDR(hdr));
7674 if (zio->io_error != 0) {
7676 * Error - drop L2ARC entry.
7678 list_remove(buflist, hdr);
7680 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
7682 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr));
7683 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
7685 bytes_dropped += arc_hdr_size(hdr);
7686 (void) refcount_remove_many(&dev->l2ad_alloc,
7687 arc_hdr_size(hdr), hdr);
7691 * Allow ARC to begin reads and ghost list evictions to
7694 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
7696 mutex_exit(hash_lock);
7699 atomic_inc_64(&l2arc_writes_done);
7700 list_remove(buflist, head);
7701 ASSERT(!HDR_HAS_L1HDR(head));
7702 kmem_cache_free(hdr_l2only_cache, head);
7703 mutex_exit(&dev->l2ad_mtx);
7705 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
7707 l2arc_do_free_on_write();
7709 kmem_free(cb, sizeof (l2arc_write_callback_t));
7713 * A read to a cache device completed. Validate buffer contents before
7714 * handing over to the regular ARC routines.
7717 l2arc_read_done(zio_t *zio)
7719 l2arc_read_callback_t *cb;
7721 kmutex_t *hash_lock;
7722 boolean_t valid_cksum;
7724 ASSERT3P(zio->io_vd, !=, NULL);
7725 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
7727 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
7729 cb = zio->io_private;
7730 ASSERT3P(cb, !=, NULL);
7731 hdr = cb->l2rcb_hdr;
7732 ASSERT3P(hdr, !=, NULL);
7734 hash_lock = HDR_LOCK(hdr);
7735 mutex_enter(hash_lock);
7736 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
7739 * If the data was read into a temporary buffer,
7740 * move it and free the buffer.
7742 if (cb->l2rcb_abd != NULL) {
7743 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
7744 if (zio->io_error == 0) {
7745 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd,
7750 * The following must be done regardless of whether
7751 * there was an error:
7752 * - free the temporary buffer
7753 * - point zio to the real ARC buffer
7754 * - set zio size accordingly
7755 * These are required because zio is either re-used for
7756 * an I/O of the block in the case of the error
7757 * or the zio is passed to arc_read_done() and it
7760 abd_free(cb->l2rcb_abd);
7761 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
7762 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
7765 ASSERT3P(zio->io_abd, !=, NULL);
7768 * Check this survived the L2ARC journey.
7770 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd);
7771 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
7772 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
7774 valid_cksum = arc_cksum_is_equal(hdr, zio);
7775 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
7776 mutex_exit(hash_lock);
7777 zio->io_private = hdr;
7780 mutex_exit(hash_lock);
7782 * Buffer didn't survive caching. Increment stats and
7783 * reissue to the original storage device.
7785 if (zio->io_error != 0) {
7786 ARCSTAT_BUMP(arcstat_l2_io_error);
7788 zio->io_error = SET_ERROR(EIO);
7791 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
7794 * If there's no waiter, issue an async i/o to the primary
7795 * storage now. If there *is* a waiter, the caller must
7796 * issue the i/o in a context where it's OK to block.
7798 if (zio->io_waiter == NULL) {
7799 zio_t *pio = zio_unique_parent(zio);
7801 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
7803 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
7804 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done,
7805 hdr, zio->io_priority, cb->l2rcb_flags,
7810 kmem_free(cb, sizeof (l2arc_read_callback_t));
7814 * This is the list priority from which the L2ARC will search for pages to
7815 * cache. This is used within loops (0..3) to cycle through lists in the
7816 * desired order. This order can have a significant effect on cache
7819 * Currently the metadata lists are hit first, MFU then MRU, followed by
7820 * the data lists. This function returns a locked list, and also returns
7823 static multilist_sublist_t *
7824 l2arc_sublist_lock(int list_num)
7826 multilist_t *ml = NULL;
7829 ASSERT(list_num >= 0 && list_num <= 3);
7833 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
7836 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
7839 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
7842 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
7847 * Return a randomly-selected sublist. This is acceptable
7848 * because the caller feeds only a little bit of data for each
7849 * call (8MB). Subsequent calls will result in different
7850 * sublists being selected.
7852 idx = multilist_get_random_index(ml);
7853 return (multilist_sublist_lock(ml, idx));
7857 * Evict buffers from the device write hand to the distance specified in
7858 * bytes. This distance may span populated buffers, it may span nothing.
7859 * This is clearing a region on the L2ARC device ready for writing.
7860 * If the 'all' boolean is set, every buffer is evicted.
7863 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
7866 arc_buf_hdr_t *hdr, *hdr_prev;
7867 kmutex_t *hash_lock;
7870 buflist = &dev->l2ad_buflist;
7872 if (!all && dev->l2ad_first) {
7874 * This is the first sweep through the device. There is
7880 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
7882 * When nearing the end of the device, evict to the end
7883 * before the device write hand jumps to the start.
7885 taddr = dev->l2ad_end;
7887 taddr = dev->l2ad_hand + distance;
7889 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
7890 uint64_t, taddr, boolean_t, all);
7893 mutex_enter(&dev->l2ad_mtx);
7894 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
7895 hdr_prev = list_prev(buflist, hdr);
7897 hash_lock = HDR_LOCK(hdr);
7900 * We cannot use mutex_enter or else we can deadlock
7901 * with l2arc_write_buffers (due to swapping the order
7902 * the hash lock and l2ad_mtx are taken).
7904 if (!mutex_tryenter(hash_lock)) {
7906 * Missed the hash lock. Retry.
7908 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
7909 mutex_exit(&dev->l2ad_mtx);
7910 mutex_enter(hash_lock);
7911 mutex_exit(hash_lock);
7916 * A header can't be on this list if it doesn't have L2 header.
7918 ASSERT(HDR_HAS_L2HDR(hdr));
7920 /* Ensure this header has finished being written. */
7921 ASSERT(!HDR_L2_WRITING(hdr));
7922 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
7924 if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
7925 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
7927 * We've evicted to the target address,
7928 * or the end of the device.
7930 mutex_exit(hash_lock);
7934 if (!HDR_HAS_L1HDR(hdr)) {
7935 ASSERT(!HDR_L2_READING(hdr));
7937 * This doesn't exist in the ARC. Destroy.
7938 * arc_hdr_destroy() will call list_remove()
7939 * and decrement arcstat_l2_lsize.
7941 arc_change_state(arc_anon, hdr, hash_lock);
7942 arc_hdr_destroy(hdr);
7944 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
7945 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
7947 * Invalidate issued or about to be issued
7948 * reads, since we may be about to write
7949 * over this location.
7951 if (HDR_L2_READING(hdr)) {
7952 ARCSTAT_BUMP(arcstat_l2_evict_reading);
7953 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
7956 arc_hdr_l2hdr_destroy(hdr);
7958 mutex_exit(hash_lock);
7960 mutex_exit(&dev->l2ad_mtx);
7964 * Find and write ARC buffers to the L2ARC device.
7966 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7967 * for reading until they have completed writing.
7968 * The headroom_boost is an in-out parameter used to maintain headroom boost
7969 * state between calls to this function.
7971 * Returns the number of bytes actually written (which may be smaller than
7972 * the delta by which the device hand has changed due to alignment).
7975 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
7977 arc_buf_hdr_t *hdr, *hdr_prev, *head;
7978 uint64_t write_asize, write_psize, write_lsize, headroom;
7980 l2arc_write_callback_t *cb;
7982 uint64_t guid = spa_load_guid(spa);
7985 ASSERT3P(dev->l2ad_vdev, !=, NULL);
7988 write_lsize = write_asize = write_psize = 0;
7990 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
7991 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
7993 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
7995 * Copy buffers for L2ARC writing.
7997 for (try = 0; try <= 3; try++) {
7998 multilist_sublist_t *mls = l2arc_sublist_lock(try);
7999 uint64_t passed_sz = 0;
8001 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
8004 * L2ARC fast warmup.
8006 * Until the ARC is warm and starts to evict, read from the
8007 * head of the ARC lists rather than the tail.
8009 if (arc_warm == B_FALSE)
8010 hdr = multilist_sublist_head(mls);
8012 hdr = multilist_sublist_tail(mls);
8014 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
8016 headroom = target_sz * l2arc_headroom;
8017 if (zfs_compressed_arc_enabled)
8018 headroom = (headroom * l2arc_headroom_boost) / 100;
8020 for (; hdr; hdr = hdr_prev) {
8021 kmutex_t *hash_lock;
8023 if (arc_warm == B_FALSE)
8024 hdr_prev = multilist_sublist_next(mls, hdr);
8026 hdr_prev = multilist_sublist_prev(mls, hdr);
8027 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned,
8028 HDR_GET_LSIZE(hdr));
8030 hash_lock = HDR_LOCK(hdr);
8031 if (!mutex_tryenter(hash_lock)) {
8032 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
8034 * Skip this buffer rather than waiting.
8039 passed_sz += HDR_GET_LSIZE(hdr);
8040 if (passed_sz > headroom) {
8044 mutex_exit(hash_lock);
8045 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
8049 if (!l2arc_write_eligible(guid, hdr)) {
8050 mutex_exit(hash_lock);
8055 * We rely on the L1 portion of the header below, so
8056 * it's invalid for this header to have been evicted out
8057 * of the ghost cache, prior to being written out. The
8058 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8060 ASSERT(HDR_HAS_L1HDR(hdr));
8062 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8063 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8064 ASSERT3U(arc_hdr_size(hdr), >, 0);
8065 uint64_t psize = arc_hdr_size(hdr);
8066 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
8069 if ((write_asize + asize) > target_sz) {
8071 mutex_exit(hash_lock);
8072 ARCSTAT_BUMP(arcstat_l2_write_full);
8078 * Insert a dummy header on the buflist so
8079 * l2arc_write_done() can find where the
8080 * write buffers begin without searching.
8082 mutex_enter(&dev->l2ad_mtx);
8083 list_insert_head(&dev->l2ad_buflist, head);
8084 mutex_exit(&dev->l2ad_mtx);
8087 sizeof (l2arc_write_callback_t), KM_SLEEP);
8088 cb->l2wcb_dev = dev;
8089 cb->l2wcb_head = head;
8090 pio = zio_root(spa, l2arc_write_done, cb,
8092 ARCSTAT_BUMP(arcstat_l2_write_pios);
8095 hdr->b_l2hdr.b_dev = dev;
8096 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
8097 arc_hdr_set_flags(hdr,
8098 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
8100 mutex_enter(&dev->l2ad_mtx);
8101 list_insert_head(&dev->l2ad_buflist, hdr);
8102 mutex_exit(&dev->l2ad_mtx);
8104 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr);
8107 * Normally the L2ARC can use the hdr's data, but if
8108 * we're sharing data between the hdr and one of its
8109 * bufs, L2ARC needs its own copy of the data so that
8110 * the ZIO below can't race with the buf consumer.
8111 * Another case where we need to create a copy of the
8112 * data is when the buffer size is not device-aligned
8113 * and we need to pad the block to make it such.
8114 * That also keeps the clock hand suitably aligned.
8116 * To ensure that the copy will be available for the
8117 * lifetime of the ZIO and be cleaned up afterwards, we
8118 * add it to the l2arc_free_on_write queue.
8121 if (!HDR_SHARED_DATA(hdr) && psize == asize) {
8122 to_write = hdr->b_l1hdr.b_pabd;
8124 to_write = abd_alloc_for_io(asize,
8125 HDR_ISTYPE_METADATA(hdr));
8126 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
8127 if (asize != psize) {
8128 abd_zero_off(to_write, psize,
8131 l2arc_free_abd_on_write(to_write, asize,
8134 wzio = zio_write_phys(pio, dev->l2ad_vdev,
8135 hdr->b_l2hdr.b_daddr, asize, to_write,
8136 ZIO_CHECKSUM_OFF, NULL, hdr,
8137 ZIO_PRIORITY_ASYNC_WRITE,
8138 ZIO_FLAG_CANFAIL, B_FALSE);
8140 write_lsize += HDR_GET_LSIZE(hdr);
8141 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
8144 write_psize += psize;
8145 write_asize += asize;
8146 dev->l2ad_hand += asize;
8148 mutex_exit(hash_lock);
8150 (void) zio_nowait(wzio);
8153 multilist_sublist_unlock(mls);
8159 /* No buffers selected for writing? */
8161 ASSERT0(write_lsize);
8162 ASSERT(!HDR_HAS_L1HDR(head));
8163 kmem_cache_free(hdr_l2only_cache, head);
8167 ASSERT3U(write_psize, <=, target_sz);
8168 ARCSTAT_BUMP(arcstat_l2_writes_sent);
8169 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
8170 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
8171 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
8172 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
8175 * Bump device hand to the device start if it is approaching the end.
8176 * l2arc_evict() will already have evicted ahead for this case.
8178 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
8179 dev->l2ad_hand = dev->l2ad_start;
8180 dev->l2ad_first = B_FALSE;
8183 dev->l2ad_writing = B_TRUE;
8184 (void) zio_wait(pio);
8185 dev->l2ad_writing = B_FALSE;
8187 return (write_asize);
8191 * This thread feeds the L2ARC at regular intervals. This is the beating
8192 * heart of the L2ARC.
8196 l2arc_feed_thread(void *unused __unused)
8201 uint64_t size, wrote;
8202 clock_t begin, next = ddi_get_lbolt();
8204 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
8206 mutex_enter(&l2arc_feed_thr_lock);
8208 while (l2arc_thread_exit == 0) {
8209 CALLB_CPR_SAFE_BEGIN(&cpr);
8210 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
8211 next - ddi_get_lbolt());
8212 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
8213 next = ddi_get_lbolt() + hz;
8216 * Quick check for L2ARC devices.
8218 mutex_enter(&l2arc_dev_mtx);
8219 if (l2arc_ndev == 0) {
8220 mutex_exit(&l2arc_dev_mtx);
8223 mutex_exit(&l2arc_dev_mtx);
8224 begin = ddi_get_lbolt();
8227 * This selects the next l2arc device to write to, and in
8228 * doing so the next spa to feed from: dev->l2ad_spa. This
8229 * will return NULL if there are now no l2arc devices or if
8230 * they are all faulted.
8232 * If a device is returned, its spa's config lock is also
8233 * held to prevent device removal. l2arc_dev_get_next()
8234 * will grab and release l2arc_dev_mtx.
8236 if ((dev = l2arc_dev_get_next()) == NULL)
8239 spa = dev->l2ad_spa;
8240 ASSERT3P(spa, !=, NULL);
8243 * If the pool is read-only then force the feed thread to
8244 * sleep a little longer.
8246 if (!spa_writeable(spa)) {
8247 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
8248 spa_config_exit(spa, SCL_L2ARC, dev);
8253 * Avoid contributing to memory pressure.
8255 if (arc_reclaim_needed()) {
8256 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
8257 spa_config_exit(spa, SCL_L2ARC, dev);
8261 ARCSTAT_BUMP(arcstat_l2_feeds);
8263 size = l2arc_write_size();
8266 * Evict L2ARC buffers that will be overwritten.
8268 l2arc_evict(dev, size, B_FALSE);
8271 * Write ARC buffers.
8273 wrote = l2arc_write_buffers(spa, dev, size);
8276 * Calculate interval between writes.
8278 next = l2arc_write_interval(begin, size, wrote);
8279 spa_config_exit(spa, SCL_L2ARC, dev);
8282 l2arc_thread_exit = 0;
8283 cv_broadcast(&l2arc_feed_thr_cv);
8284 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
8289 l2arc_vdev_present(vdev_t *vd)
8293 mutex_enter(&l2arc_dev_mtx);
8294 for (dev = list_head(l2arc_dev_list); dev != NULL;
8295 dev = list_next(l2arc_dev_list, dev)) {
8296 if (dev->l2ad_vdev == vd)
8299 mutex_exit(&l2arc_dev_mtx);
8301 return (dev != NULL);
8305 * Add a vdev for use by the L2ARC. By this point the spa has already
8306 * validated the vdev and opened it.
8309 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
8311 l2arc_dev_t *adddev;
8313 ASSERT(!l2arc_vdev_present(vd));
8315 vdev_ashift_optimize(vd);
8318 * Create a new l2arc device entry.
8320 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
8321 adddev->l2ad_spa = spa;
8322 adddev->l2ad_vdev = vd;
8323 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
8324 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
8325 adddev->l2ad_hand = adddev->l2ad_start;
8326 adddev->l2ad_first = B_TRUE;
8327 adddev->l2ad_writing = B_FALSE;
8329 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
8331 * This is a list of all ARC buffers that are still valid on the
8334 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
8335 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
8337 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
8338 refcount_create(&adddev->l2ad_alloc);
8341 * Add device to global list
8343 mutex_enter(&l2arc_dev_mtx);
8344 list_insert_head(l2arc_dev_list, adddev);
8345 atomic_inc_64(&l2arc_ndev);
8346 mutex_exit(&l2arc_dev_mtx);
8350 * Remove a vdev from the L2ARC.
8353 l2arc_remove_vdev(vdev_t *vd)
8355 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
8358 * Find the device by vdev
8360 mutex_enter(&l2arc_dev_mtx);
8361 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
8362 nextdev = list_next(l2arc_dev_list, dev);
8363 if (vd == dev->l2ad_vdev) {
8368 ASSERT3P(remdev, !=, NULL);
8371 * Remove device from global list
8373 list_remove(l2arc_dev_list, remdev);
8374 l2arc_dev_last = NULL; /* may have been invalidated */
8375 atomic_dec_64(&l2arc_ndev);
8376 mutex_exit(&l2arc_dev_mtx);
8379 * Clear all buflists and ARC references. L2ARC device flush.
8381 l2arc_evict(remdev, 0, B_TRUE);
8382 list_destroy(&remdev->l2ad_buflist);
8383 mutex_destroy(&remdev->l2ad_mtx);
8384 refcount_destroy(&remdev->l2ad_alloc);
8385 kmem_free(remdev, sizeof (l2arc_dev_t));
8391 l2arc_thread_exit = 0;
8393 l2arc_writes_sent = 0;
8394 l2arc_writes_done = 0;
8396 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
8397 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
8398 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
8399 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
8401 l2arc_dev_list = &L2ARC_dev_list;
8402 l2arc_free_on_write = &L2ARC_free_on_write;
8403 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
8404 offsetof(l2arc_dev_t, l2ad_node));
8405 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
8406 offsetof(l2arc_data_free_t, l2df_list_node));
8413 * This is called from dmu_fini(), which is called from spa_fini();
8414 * Because of this, we can assume that all l2arc devices have
8415 * already been removed when the pools themselves were removed.
8418 l2arc_do_free_on_write();
8420 mutex_destroy(&l2arc_feed_thr_lock);
8421 cv_destroy(&l2arc_feed_thr_cv);
8422 mutex_destroy(&l2arc_dev_mtx);
8423 mutex_destroy(&l2arc_free_on_write_mtx);
8425 list_destroy(l2arc_dev_list);
8426 list_destroy(l2arc_free_on_write);
8432 if (!(spa_mode_global & FWRITE))
8435 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
8436 TS_RUN, minclsyspri);
8442 if (!(spa_mode_global & FWRITE))
8445 mutex_enter(&l2arc_feed_thr_lock);
8446 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
8447 l2arc_thread_exit = 1;
8448 while (l2arc_thread_exit != 0)
8449 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
8450 mutex_exit(&l2arc_feed_thr_lock);