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 <sys/zthr.h>
285 #include <zfs_fletcher.h>
287 #include <sys/aggsum.h>
288 #include <sys/cityhash.h>
290 #include <machine/vmparam.h>
294 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
295 boolean_t arc_watch = B_FALSE;
301 * This thread's job is to keep enough free memory in the system, by
302 * calling arc_kmem_reap_now() plus arc_shrink(), which improves
303 * arc_available_memory().
305 static zthr_t *arc_reap_zthr;
308 * This thread's job is to keep arc_size under arc_c, by calling
309 * arc_adjust(), which improves arc_is_overflowing().
311 static zthr_t *arc_adjust_zthr;
313 static kmutex_t arc_adjust_lock;
314 static kcondvar_t arc_adjust_waiters_cv;
315 static boolean_t arc_adjust_needed = B_FALSE;
317 static kmutex_t arc_dnlc_evicts_lock;
318 static kcondvar_t arc_dnlc_evicts_cv;
319 static boolean_t arc_dnlc_evicts_thread_exit;
321 uint_t arc_reduce_dnlc_percent = 3;
324 * The number of headers to evict in arc_evict_state_impl() before
325 * dropping the sublist lock and evicting from another sublist. A lower
326 * value means we're more likely to evict the "correct" header (i.e. the
327 * oldest header in the arc state), but comes with higher overhead
328 * (i.e. more invocations of arc_evict_state_impl()).
330 int zfs_arc_evict_batch_limit = 10;
332 /* number of seconds before growing cache again */
333 int arc_grow_retry = 60;
336 * Minimum time between calls to arc_kmem_reap_soon(). Note that this will
337 * be converted to ticks, so with the default hz=100, a setting of 15 ms
338 * will actually wait 2 ticks, or 20ms.
340 int arc_kmem_cache_reap_retry_ms = 1000;
342 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
343 int zfs_arc_overflow_shift = 8;
345 /* shift of arc_c for calculating both min and max arc_p */
346 int arc_p_min_shift = 4;
348 /* log2(fraction of arc to reclaim) */
349 int arc_shrink_shift = 7;
352 * log2(fraction of ARC which must be free to allow growing).
353 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
354 * when reading a new block into the ARC, we will evict an equal-sized block
357 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
358 * we will still not allow it to grow.
360 int arc_no_grow_shift = 5;
364 * minimum lifespan of a prefetch block in clock ticks
365 * (initialized in arc_init())
367 static int zfs_arc_min_prefetch_ms = 1;
368 static int zfs_arc_min_prescient_prefetch_ms = 6;
371 * If this percent of memory is free, don't throttle.
373 int arc_lotsfree_percent = 10;
375 static boolean_t arc_initialized;
376 extern boolean_t zfs_prefetch_disable;
379 * The arc has filled available memory and has now warmed up.
381 static boolean_t arc_warm;
384 * log2 fraction of the zio arena to keep free.
386 int arc_zio_arena_free_shift = 2;
389 * These tunables are for performance analysis.
391 uint64_t zfs_arc_max;
392 uint64_t zfs_arc_min;
393 uint64_t zfs_arc_meta_limit = 0;
394 uint64_t zfs_arc_meta_min = 0;
395 uint64_t zfs_arc_dnode_limit = 0;
396 uint64_t zfs_arc_dnode_reduce_percent = 10;
397 int zfs_arc_grow_retry = 0;
398 int zfs_arc_shrink_shift = 0;
399 int zfs_arc_no_grow_shift = 0;
400 int zfs_arc_p_min_shift = 0;
401 uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
402 u_int zfs_arc_free_target = 0;
404 /* Absolute min for arc min / max is 16MB. */
405 static uint64_t arc_abs_min = 16 << 20;
408 * ARC dirty data constraints for arc_tempreserve_space() throttle
410 uint_t zfs_arc_dirty_limit_percent = 50; /* total dirty data limit */
411 uint_t zfs_arc_anon_limit_percent = 25; /* anon block dirty limit */
412 uint_t zfs_arc_pool_dirty_percent = 20; /* each pool's anon allowance */
414 boolean_t zfs_compressed_arc_enabled = B_TRUE;
416 static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
417 static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
418 static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS);
419 static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS);
420 static int sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS);
422 #if defined(__FreeBSD__) && defined(_KERNEL)
424 arc_free_target_init(void *unused __unused)
427 zfs_arc_free_target = vm_cnt.v_free_target;
429 SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
430 arc_free_target_init, NULL);
432 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
433 TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
434 TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
435 TUNABLE_INT("vfs.zfs.arc_grow_retry", &zfs_arc_grow_retry);
436 TUNABLE_INT("vfs.zfs.arc_no_grow_shift", &zfs_arc_no_grow_shift);
437 SYSCTL_DECL(_vfs_zfs);
438 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN,
439 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size");
440 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN,
441 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size");
442 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_no_grow_shift, CTLTYPE_U32 | CTLFLAG_RWTUN,
443 0, sizeof(uint32_t), sysctl_vfs_zfs_arc_no_grow_shift, "U",
444 "log2(fraction of ARC which must be free to allow growing)");
445 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
446 &zfs_arc_average_blocksize, 0,
447 "ARC average blocksize");
448 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
449 &arc_shrink_shift, 0,
450 "log2(fraction of arc to reclaim)");
451 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_grow_retry, CTLFLAG_RW,
453 "Wait in seconds before considering growing ARC");
454 SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN,
455 &zfs_compressed_arc_enabled, 0,
456 "Enable compressed ARC");
457 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_kmem_cache_reap_retry_ms, CTLFLAG_RWTUN,
458 &arc_kmem_cache_reap_retry_ms, 0,
459 "Interval between ARC kmem_cache reapings");
462 * We don't have a tunable for arc_free_target due to the dependency on
463 * pagedaemon initialisation.
465 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
466 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
467 sysctl_vfs_zfs_arc_free_target, "IU",
468 "Desired number of free pages below which ARC triggers reclaim");
471 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
476 val = zfs_arc_free_target;
477 err = sysctl_handle_int(oidp, &val, 0, req);
478 if (err != 0 || req->newptr == NULL)
483 if (val > vm_cnt.v_page_count)
486 zfs_arc_free_target = val;
492 * Must be declared here, before the definition of corresponding kstat
493 * macro which uses the same names will confuse the compiler.
495 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
496 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
497 sysctl_vfs_zfs_arc_meta_limit, "QU",
498 "ARC metadata limit");
502 * Note that buffers can be in one of 6 states:
503 * ARC_anon - anonymous (discussed below)
504 * ARC_mru - recently used, currently cached
505 * ARC_mru_ghost - recentely used, no longer in cache
506 * ARC_mfu - frequently used, currently cached
507 * ARC_mfu_ghost - frequently used, no longer in cache
508 * ARC_l2c_only - exists in L2ARC but not other states
509 * When there are no active references to the buffer, they are
510 * are linked onto a list in one of these arc states. These are
511 * the only buffers that can be evicted or deleted. Within each
512 * state there are multiple lists, one for meta-data and one for
513 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
514 * etc.) is tracked separately so that it can be managed more
515 * explicitly: favored over data, limited explicitly.
517 * Anonymous buffers are buffers that are not associated with
518 * a DVA. These are buffers that hold dirty block copies
519 * before they are written to stable storage. By definition,
520 * they are "ref'd" and are considered part of arc_mru
521 * that cannot be freed. Generally, they will aquire a DVA
522 * as they are written and migrate onto the arc_mru list.
524 * The ARC_l2c_only state is for buffers that are in the second
525 * level ARC but no longer in any of the ARC_m* lists. The second
526 * level ARC itself may also contain buffers that are in any of
527 * the ARC_m* states - meaning that a buffer can exist in two
528 * places. The reason for the ARC_l2c_only state is to keep the
529 * buffer header in the hash table, so that reads that hit the
530 * second level ARC benefit from these fast lookups.
533 typedef struct arc_state {
535 * list of evictable buffers
537 multilist_t *arcs_list[ARC_BUFC_NUMTYPES];
539 * total amount of evictable data in this state
541 refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
543 * total amount of data in this state; this includes: evictable,
544 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
546 refcount_t arcs_size;
548 * supports the "dbufs" kstat
550 arc_state_type_t arcs_state;
554 * Percentage that can be consumed by dnodes of ARC meta buffers.
556 int zfs_arc_meta_prune = 10000;
557 unsigned long zfs_arc_dnode_limit_percent = 10;
558 int zfs_arc_meta_strategy = ARC_STRATEGY_META_ONLY;
559 int zfs_arc_meta_adjust_restarts = 4096;
561 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_meta_strategy, CTLFLAG_RWTUN,
562 &zfs_arc_meta_strategy, 0,
563 "ARC metadata reclamation strategy "
564 "(0 = metadata only, 1 = balance data and metadata)");
567 static arc_state_t ARC_anon;
568 static arc_state_t ARC_mru;
569 static arc_state_t ARC_mru_ghost;
570 static arc_state_t ARC_mfu;
571 static arc_state_t ARC_mfu_ghost;
572 static arc_state_t ARC_l2c_only;
574 typedef struct arc_stats {
575 kstat_named_t arcstat_hits;
576 kstat_named_t arcstat_misses;
577 kstat_named_t arcstat_demand_data_hits;
578 kstat_named_t arcstat_demand_data_misses;
579 kstat_named_t arcstat_demand_metadata_hits;
580 kstat_named_t arcstat_demand_metadata_misses;
581 kstat_named_t arcstat_prefetch_data_hits;
582 kstat_named_t arcstat_prefetch_data_misses;
583 kstat_named_t arcstat_prefetch_metadata_hits;
584 kstat_named_t arcstat_prefetch_metadata_misses;
585 kstat_named_t arcstat_mru_hits;
586 kstat_named_t arcstat_mru_ghost_hits;
587 kstat_named_t arcstat_mfu_hits;
588 kstat_named_t arcstat_mfu_ghost_hits;
589 kstat_named_t arcstat_allocated;
590 kstat_named_t arcstat_deleted;
592 * Number of buffers that could not be evicted because the hash lock
593 * was held by another thread. The lock may not necessarily be held
594 * by something using the same buffer, since hash locks are shared
595 * by multiple buffers.
597 kstat_named_t arcstat_mutex_miss;
599 * Number of buffers skipped when updating the access state due to the
600 * header having already been released after acquiring the hash lock.
602 kstat_named_t arcstat_access_skip;
604 * Number of buffers skipped because they have I/O in progress, are
605 * indirect prefetch buffers that have not lived long enough, or are
606 * not from the spa we're trying to evict from.
608 kstat_named_t arcstat_evict_skip;
610 * Number of times arc_evict_state() was unable to evict enough
611 * buffers to reach it's target amount.
613 kstat_named_t arcstat_evict_not_enough;
614 kstat_named_t arcstat_evict_l2_cached;
615 kstat_named_t arcstat_evict_l2_eligible;
616 kstat_named_t arcstat_evict_l2_ineligible;
617 kstat_named_t arcstat_evict_l2_skip;
618 kstat_named_t arcstat_hash_elements;
619 kstat_named_t arcstat_hash_elements_max;
620 kstat_named_t arcstat_hash_collisions;
621 kstat_named_t arcstat_hash_chains;
622 kstat_named_t arcstat_hash_chain_max;
623 kstat_named_t arcstat_p;
624 kstat_named_t arcstat_c;
625 kstat_named_t arcstat_c_min;
626 kstat_named_t arcstat_c_max;
627 /* Not updated directly; only synced in arc_kstat_update. */
628 kstat_named_t arcstat_size;
630 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
631 * Note that the compressed bytes may match the uncompressed bytes
632 * if the block is either not compressed or compressed arc is disabled.
634 kstat_named_t arcstat_compressed_size;
636 * Uncompressed size of the data stored in b_pabd. If compressed
637 * arc is disabled then this value will be identical to the stat
640 kstat_named_t arcstat_uncompressed_size;
642 * Number of bytes stored in all the arc_buf_t's. This is classified
643 * as "overhead" since this data is typically short-lived and will
644 * be evicted from the arc when it becomes unreferenced unless the
645 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
646 * values have been set (see comment in dbuf.c for more information).
648 kstat_named_t arcstat_overhead_size;
650 * Number of bytes consumed by internal ARC structures necessary
651 * for tracking purposes; these structures are not actually
652 * backed by ARC buffers. This includes arc_buf_hdr_t structures
653 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
654 * caches), and arc_buf_t structures (allocated via arc_buf_t
656 * Not updated directly; only synced in arc_kstat_update.
658 kstat_named_t arcstat_hdr_size;
660 * Number of bytes consumed by ARC buffers of type equal to
661 * ARC_BUFC_DATA. This is generally consumed by buffers backing
662 * on disk user data (e.g. plain file contents).
663 * Not updated directly; only synced in arc_kstat_update.
665 kstat_named_t arcstat_data_size;
667 * Number of bytes consumed by ARC buffers of type equal to
668 * ARC_BUFC_METADATA. This is generally consumed by buffers
669 * backing on disk data that is used for internal ZFS
670 * structures (e.g. ZAP, dnode, indirect blocks, etc).
671 * Not updated directly; only synced in arc_kstat_update.
673 kstat_named_t arcstat_metadata_size;
675 * Number of bytes consumed by dmu_buf_impl_t objects.
677 kstat_named_t arcstat_dbuf_size;
679 * Number of bytes consumed by dnode_t objects.
681 kstat_named_t arcstat_dnode_size;
683 * Number of bytes consumed by bonus buffers.
685 kstat_named_t arcstat_bonus_size;
686 #if defined(__FreeBSD__) && defined(COMPAT_FREEBSD11)
688 * Sum of the previous three counters, provided for compatibility.
690 kstat_named_t arcstat_other_size;
693 * Total number of bytes consumed by ARC buffers residing in the
694 * arc_anon state. This includes *all* buffers in the arc_anon
695 * state; e.g. data, metadata, evictable, and unevictable buffers
696 * are all included in this value.
697 * Not updated directly; only synced in arc_kstat_update.
699 kstat_named_t arcstat_anon_size;
701 * Number of bytes consumed by ARC buffers that meet the
702 * following criteria: backing buffers of type ARC_BUFC_DATA,
703 * residing in the arc_anon 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_anon_evictable_data;
709 * Number of bytes consumed by ARC buffers that meet the
710 * following criteria: backing buffers of type ARC_BUFC_METADATA,
711 * residing in the arc_anon state, and are eligible for eviction
712 * (e.g. have no outstanding holds on the buffer).
713 * Not updated directly; only synced in arc_kstat_update.
715 kstat_named_t arcstat_anon_evictable_metadata;
717 * Total number of bytes consumed by ARC buffers residing in the
718 * arc_mru state. This includes *all* buffers in the arc_mru
719 * state; e.g. data, metadata, evictable, and unevictable buffers
720 * are all included in this value.
721 * Not updated directly; only synced in arc_kstat_update.
723 kstat_named_t arcstat_mru_size;
725 * Number of bytes consumed by ARC buffers that meet the
726 * following criteria: backing buffers of type ARC_BUFC_DATA,
727 * residing in the arc_mru state, and are eligible for eviction
728 * (e.g. have no outstanding holds on the buffer).
729 * Not updated directly; only synced in arc_kstat_update.
731 kstat_named_t arcstat_mru_evictable_data;
733 * Number of bytes consumed by ARC buffers that meet the
734 * following criteria: backing buffers of type ARC_BUFC_METADATA,
735 * residing in the arc_mru state, and are eligible for eviction
736 * (e.g. have no outstanding holds on the buffer).
737 * Not updated directly; only synced in arc_kstat_update.
739 kstat_named_t arcstat_mru_evictable_metadata;
741 * Total number of bytes that *would have been* consumed by ARC
742 * buffers in the arc_mru_ghost state. The key thing to note
743 * here, is the fact that this size doesn't actually indicate
744 * RAM consumption. The ghost lists only consist of headers and
745 * don't actually have ARC buffers linked off of these headers.
746 * Thus, *if* the headers had associated ARC buffers, these
747 * buffers *would have* consumed this number of bytes.
748 * Not updated directly; only synced in arc_kstat_update.
750 kstat_named_t arcstat_mru_ghost_size;
752 * Number of bytes that *would have been* consumed by ARC
753 * buffers that are eligible for eviction, of type
754 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
755 * Not updated directly; only synced in arc_kstat_update.
757 kstat_named_t arcstat_mru_ghost_evictable_data;
759 * Number of bytes that *would have been* consumed by ARC
760 * buffers that are eligible for eviction, of type
761 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
762 * Not updated directly; only synced in arc_kstat_update.
764 kstat_named_t arcstat_mru_ghost_evictable_metadata;
766 * Total number of bytes consumed by ARC buffers residing in the
767 * arc_mfu state. This includes *all* buffers in the arc_mfu
768 * state; e.g. data, metadata, evictable, and unevictable buffers
769 * are all included in this value.
770 * Not updated directly; only synced in arc_kstat_update.
772 kstat_named_t arcstat_mfu_size;
774 * Number of bytes consumed by ARC buffers that are eligible for
775 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
777 * Not updated directly; only synced in arc_kstat_update.
779 kstat_named_t arcstat_mfu_evictable_data;
781 * Number of bytes consumed by ARC buffers that are eligible for
782 * eviction, of type ARC_BUFC_METADATA, and reside in the
784 * Not updated directly; only synced in arc_kstat_update.
786 kstat_named_t arcstat_mfu_evictable_metadata;
788 * Total number of bytes that *would have been* consumed by ARC
789 * buffers in the arc_mfu_ghost state. See the comment above
790 * arcstat_mru_ghost_size for more details.
791 * Not updated directly; only synced in arc_kstat_update.
793 kstat_named_t arcstat_mfu_ghost_size;
795 * Number of bytes that *would have been* consumed by ARC
796 * buffers that are eligible for eviction, of type
797 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
798 * Not updated directly; only synced in arc_kstat_update.
800 kstat_named_t arcstat_mfu_ghost_evictable_data;
802 * Number of bytes that *would have been* consumed by ARC
803 * buffers that are eligible for eviction, of type
804 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
805 * Not updated directly; only synced in arc_kstat_update.
807 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
808 kstat_named_t arcstat_l2_hits;
809 kstat_named_t arcstat_l2_misses;
810 kstat_named_t arcstat_l2_feeds;
811 kstat_named_t arcstat_l2_rw_clash;
812 kstat_named_t arcstat_l2_read_bytes;
813 kstat_named_t arcstat_l2_write_bytes;
814 kstat_named_t arcstat_l2_writes_sent;
815 kstat_named_t arcstat_l2_writes_done;
816 kstat_named_t arcstat_l2_writes_error;
817 kstat_named_t arcstat_l2_writes_lock_retry;
818 kstat_named_t arcstat_l2_evict_lock_retry;
819 kstat_named_t arcstat_l2_evict_reading;
820 kstat_named_t arcstat_l2_evict_l1cached;
821 kstat_named_t arcstat_l2_free_on_write;
822 kstat_named_t arcstat_l2_abort_lowmem;
823 kstat_named_t arcstat_l2_cksum_bad;
824 kstat_named_t arcstat_l2_io_error;
825 kstat_named_t arcstat_l2_lsize;
826 kstat_named_t arcstat_l2_psize;
827 /* Not updated directly; only synced in arc_kstat_update. */
828 kstat_named_t arcstat_l2_hdr_size;
829 kstat_named_t arcstat_l2_write_trylock_fail;
830 kstat_named_t arcstat_l2_write_passed_headroom;
831 kstat_named_t arcstat_l2_write_spa_mismatch;
832 kstat_named_t arcstat_l2_write_in_l2;
833 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
834 kstat_named_t arcstat_l2_write_not_cacheable;
835 kstat_named_t arcstat_l2_write_full;
836 kstat_named_t arcstat_l2_write_buffer_iter;
837 kstat_named_t arcstat_l2_write_pios;
838 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
839 kstat_named_t arcstat_l2_write_buffer_list_iter;
840 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
841 kstat_named_t arcstat_memory_throttle_count;
842 kstat_named_t arcstat_memory_direct_count;
843 kstat_named_t arcstat_memory_indirect_count;
844 kstat_named_t arcstat_memory_all_bytes;
845 kstat_named_t arcstat_memory_free_bytes;
846 kstat_named_t arcstat_memory_available_bytes;
847 kstat_named_t arcstat_no_grow;
848 kstat_named_t arcstat_tempreserve;
849 kstat_named_t arcstat_loaned_bytes;
850 kstat_named_t arcstat_prune;
851 /* Not updated directly; only synced in arc_kstat_update. */
852 kstat_named_t arcstat_meta_used;
853 kstat_named_t arcstat_meta_limit;
854 kstat_named_t arcstat_dnode_limit;
855 kstat_named_t arcstat_meta_max;
856 kstat_named_t arcstat_meta_min;
857 kstat_named_t arcstat_async_upgrade_sync;
858 kstat_named_t arcstat_demand_hit_predictive_prefetch;
859 kstat_named_t arcstat_demand_hit_prescient_prefetch;
862 static arc_stats_t arc_stats = {
863 { "hits", KSTAT_DATA_UINT64 },
864 { "misses", KSTAT_DATA_UINT64 },
865 { "demand_data_hits", KSTAT_DATA_UINT64 },
866 { "demand_data_misses", KSTAT_DATA_UINT64 },
867 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
868 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
869 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
870 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
871 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
872 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
873 { "mru_hits", KSTAT_DATA_UINT64 },
874 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
875 { "mfu_hits", KSTAT_DATA_UINT64 },
876 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
877 { "allocated", KSTAT_DATA_UINT64 },
878 { "deleted", KSTAT_DATA_UINT64 },
879 { "mutex_miss", KSTAT_DATA_UINT64 },
880 { "access_skip", KSTAT_DATA_UINT64 },
881 { "evict_skip", KSTAT_DATA_UINT64 },
882 { "evict_not_enough", KSTAT_DATA_UINT64 },
883 { "evict_l2_cached", KSTAT_DATA_UINT64 },
884 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
885 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
886 { "evict_l2_skip", KSTAT_DATA_UINT64 },
887 { "hash_elements", KSTAT_DATA_UINT64 },
888 { "hash_elements_max", KSTAT_DATA_UINT64 },
889 { "hash_collisions", KSTAT_DATA_UINT64 },
890 { "hash_chains", KSTAT_DATA_UINT64 },
891 { "hash_chain_max", KSTAT_DATA_UINT64 },
892 { "p", KSTAT_DATA_UINT64 },
893 { "c", KSTAT_DATA_UINT64 },
894 { "c_min", KSTAT_DATA_UINT64 },
895 { "c_max", KSTAT_DATA_UINT64 },
896 { "size", KSTAT_DATA_UINT64 },
897 { "compressed_size", KSTAT_DATA_UINT64 },
898 { "uncompressed_size", KSTAT_DATA_UINT64 },
899 { "overhead_size", KSTAT_DATA_UINT64 },
900 { "hdr_size", KSTAT_DATA_UINT64 },
901 { "data_size", KSTAT_DATA_UINT64 },
902 { "metadata_size", KSTAT_DATA_UINT64 },
903 { "dbuf_size", KSTAT_DATA_UINT64 },
904 { "dnode_size", KSTAT_DATA_UINT64 },
905 { "bonus_size", KSTAT_DATA_UINT64 },
906 #if defined(__FreeBSD__) && defined(COMPAT_FREEBSD11)
907 { "other_size", KSTAT_DATA_UINT64 },
909 { "anon_size", KSTAT_DATA_UINT64 },
910 { "anon_evictable_data", KSTAT_DATA_UINT64 },
911 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
912 { "mru_size", KSTAT_DATA_UINT64 },
913 { "mru_evictable_data", KSTAT_DATA_UINT64 },
914 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
915 { "mru_ghost_size", KSTAT_DATA_UINT64 },
916 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
917 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
918 { "mfu_size", KSTAT_DATA_UINT64 },
919 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
920 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
921 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
922 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
923 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
924 { "l2_hits", KSTAT_DATA_UINT64 },
925 { "l2_misses", KSTAT_DATA_UINT64 },
926 { "l2_feeds", KSTAT_DATA_UINT64 },
927 { "l2_rw_clash", KSTAT_DATA_UINT64 },
928 { "l2_read_bytes", KSTAT_DATA_UINT64 },
929 { "l2_write_bytes", KSTAT_DATA_UINT64 },
930 { "l2_writes_sent", KSTAT_DATA_UINT64 },
931 { "l2_writes_done", KSTAT_DATA_UINT64 },
932 { "l2_writes_error", KSTAT_DATA_UINT64 },
933 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
934 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
935 { "l2_evict_reading", KSTAT_DATA_UINT64 },
936 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
937 { "l2_free_on_write", KSTAT_DATA_UINT64 },
938 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
939 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
940 { "l2_io_error", KSTAT_DATA_UINT64 },
941 { "l2_size", KSTAT_DATA_UINT64 },
942 { "l2_asize", KSTAT_DATA_UINT64 },
943 { "l2_hdr_size", KSTAT_DATA_UINT64 },
944 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
945 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
946 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
947 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
948 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
949 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
950 { "l2_write_full", KSTAT_DATA_UINT64 },
951 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
952 { "l2_write_pios", KSTAT_DATA_UINT64 },
953 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
954 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
955 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
956 { "memory_throttle_count", KSTAT_DATA_UINT64 },
957 { "memory_direct_count", KSTAT_DATA_UINT64 },
958 { "memory_indirect_count", KSTAT_DATA_UINT64 },
959 { "memory_all_bytes", KSTAT_DATA_UINT64 },
960 { "memory_free_bytes", KSTAT_DATA_UINT64 },
961 { "memory_available_bytes", KSTAT_DATA_UINT64 },
962 { "arc_no_grow", KSTAT_DATA_UINT64 },
963 { "arc_tempreserve", KSTAT_DATA_UINT64 },
964 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
965 { "arc_prune", KSTAT_DATA_UINT64 },
966 { "arc_meta_used", KSTAT_DATA_UINT64 },
967 { "arc_meta_limit", KSTAT_DATA_UINT64 },
968 { "arc_dnode_limit", KSTAT_DATA_UINT64 },
969 { "arc_meta_max", KSTAT_DATA_UINT64 },
970 { "arc_meta_min", KSTAT_DATA_UINT64 },
971 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
972 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
973 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
976 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
978 #define ARCSTAT_INCR(stat, val) \
979 atomic_add_64(&arc_stats.stat.value.ui64, (val))
981 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
982 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
984 #define ARCSTAT_MAX(stat, val) { \
986 while ((val) > (m = arc_stats.stat.value.ui64) && \
987 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
991 #define ARCSTAT_MAXSTAT(stat) \
992 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
995 * We define a macro to allow ARC hits/misses to be easily broken down by
996 * two separate conditions, giving a total of four different subtypes for
997 * each of hits and misses (so eight statistics total).
999 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
1002 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
1004 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
1008 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
1010 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
1015 static arc_state_t *arc_anon;
1016 static arc_state_t *arc_mru;
1017 static arc_state_t *arc_mru_ghost;
1018 static arc_state_t *arc_mfu;
1019 static arc_state_t *arc_mfu_ghost;
1020 static arc_state_t *arc_l2c_only;
1023 * There are several ARC variables that are critical to export as kstats --
1024 * but we don't want to have to grovel around in the kstat whenever we wish to
1025 * manipulate them. For these variables, we therefore define them to be in
1026 * terms of the statistic variable. This assures that we are not introducing
1027 * the possibility of inconsistency by having shadow copies of the variables,
1028 * while still allowing the code to be readable.
1030 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
1031 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
1032 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
1033 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
1034 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
1035 #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
1036 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
1037 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
1038 #define arc_dbuf_size ARCSTAT(arcstat_dbuf_size) /* dbuf metadata */
1039 #define arc_dnode_size ARCSTAT(arcstat_dnode_size) /* dnode metadata */
1040 #define arc_bonus_size ARCSTAT(arcstat_bonus_size) /* bonus buffer metadata */
1042 /* compressed size of entire arc */
1043 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
1044 /* uncompressed size of entire arc */
1045 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
1046 /* number of bytes in the arc from arc_buf_t's */
1047 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
1050 * There are also some ARC variables that we want to export, but that are
1051 * updated so often that having the canonical representation be the statistic
1052 * variable causes a performance bottleneck. We want to use aggsum_t's for these
1053 * instead, but still be able to export the kstat in the same way as before.
1054 * The solution is to always use the aggsum version, except in the kstat update
1058 aggsum_t arc_meta_used;
1059 aggsum_t astat_data_size;
1060 aggsum_t astat_metadata_size;
1061 aggsum_t astat_hdr_size;
1062 aggsum_t astat_bonus_size;
1063 aggsum_t astat_dnode_size;
1064 aggsum_t astat_dbuf_size;
1065 aggsum_t astat_l2_hdr_size;
1067 static list_t arc_prune_list;
1068 static kmutex_t arc_prune_mtx;
1069 static taskq_t *arc_prune_taskq;
1071 static int arc_no_grow; /* Don't try to grow cache size */
1072 static hrtime_t arc_growtime;
1073 static uint64_t arc_tempreserve;
1074 static uint64_t arc_loaned_bytes;
1076 typedef struct arc_callback arc_callback_t;
1078 struct arc_callback {
1080 arc_read_done_func_t *acb_done;
1082 boolean_t acb_compressed;
1083 zio_t *acb_zio_dummy;
1084 zio_t *acb_zio_head;
1085 arc_callback_t *acb_next;
1088 typedef struct arc_write_callback arc_write_callback_t;
1090 struct arc_write_callback {
1092 arc_write_done_func_t *awcb_ready;
1093 arc_write_done_func_t *awcb_children_ready;
1094 arc_write_done_func_t *awcb_physdone;
1095 arc_write_done_func_t *awcb_done;
1096 arc_buf_t *awcb_buf;
1100 * ARC buffers are separated into multiple structs as a memory saving measure:
1101 * - Common fields struct, always defined, and embedded within it:
1102 * - L2-only fields, always allocated but undefined when not in L2ARC
1103 * - L1-only fields, only allocated when in L1ARC
1105 * Buffer in L1 Buffer only in L2
1106 * +------------------------+ +------------------------+
1107 * | arc_buf_hdr_t | | arc_buf_hdr_t |
1111 * +------------------------+ +------------------------+
1112 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
1113 * | (undefined if L1-only) | | |
1114 * +------------------------+ +------------------------+
1115 * | l1arc_buf_hdr_t |
1120 * +------------------------+
1122 * Because it's possible for the L2ARC to become extremely large, we can wind
1123 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
1124 * is minimized by only allocating the fields necessary for an L1-cached buffer
1125 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
1126 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
1127 * words in pointers. arc_hdr_realloc() is used to switch a header between
1128 * these two allocation states.
1130 typedef struct l1arc_buf_hdr {
1131 kmutex_t b_freeze_lock;
1132 zio_cksum_t *b_freeze_cksum;
1135 * Used for debugging with kmem_flags - by allocating and freeing
1136 * b_thawed when the buffer is thawed, we get a record of the stack
1137 * trace that thawed it.
1144 /* for waiting on writes to complete */
1148 /* protected by arc state mutex */
1149 arc_state_t *b_state;
1150 multilist_node_t b_arc_node;
1152 /* updated atomically */
1153 clock_t b_arc_access;
1154 uint32_t b_mru_hits;
1155 uint32_t b_mru_ghost_hits;
1156 uint32_t b_mfu_hits;
1157 uint32_t b_mfu_ghost_hits;
1160 /* self protecting */
1161 refcount_t b_refcnt;
1163 arc_callback_t *b_acb;
1167 typedef struct l2arc_dev l2arc_dev_t;
1169 typedef struct l2arc_buf_hdr {
1170 /* protected by arc_buf_hdr mutex */
1171 l2arc_dev_t *b_dev; /* L2ARC device */
1172 uint64_t b_daddr; /* disk address, offset byte */
1175 list_node_t b_l2node;
1178 struct arc_buf_hdr {
1179 /* protected by hash lock */
1183 arc_buf_contents_t b_type;
1184 arc_buf_hdr_t *b_hash_next;
1185 arc_flags_t b_flags;
1188 * This field stores the size of the data buffer after
1189 * compression, and is set in the arc's zio completion handlers.
1190 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
1192 * While the block pointers can store up to 32MB in their psize
1193 * field, we can only store up to 32MB minus 512B. This is due
1194 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
1195 * a field of zeros represents 512B in the bp). We can't use a
1196 * bias of 1 since we need to reserve a psize of zero, here, to
1197 * represent holes and embedded blocks.
1199 * This isn't a problem in practice, since the maximum size of a
1200 * buffer is limited to 16MB, so we never need to store 32MB in
1201 * this field. Even in the upstream illumos code base, the
1202 * maximum size of a buffer is limited to 16MB.
1207 * This field stores the size of the data buffer before
1208 * compression, and cannot change once set. It is in units
1209 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
1211 uint16_t b_lsize; /* immutable */
1212 uint64_t b_spa; /* immutable */
1214 /* L2ARC fields. Undefined when not in L2ARC. */
1215 l2arc_buf_hdr_t b_l2hdr;
1216 /* L1ARC fields. Undefined when in l2arc_only state */
1217 l1arc_buf_hdr_t b_l1hdr;
1220 #if defined(__FreeBSD__) && defined(_KERNEL)
1222 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
1227 val = arc_meta_limit;
1228 err = sysctl_handle_64(oidp, &val, 0, req);
1229 if (err != 0 || req->newptr == NULL)
1232 if (val <= 0 || val > arc_c_max)
1235 arc_meta_limit = val;
1237 mutex_enter(&arc_adjust_lock);
1238 arc_adjust_needed = B_TRUE;
1239 mutex_exit(&arc_adjust_lock);
1240 zthr_wakeup(arc_adjust_zthr);
1246 sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS)
1251 val = arc_no_grow_shift;
1252 err = sysctl_handle_32(oidp, &val, 0, req);
1253 if (err != 0 || req->newptr == NULL)
1256 if (val >= arc_shrink_shift)
1259 arc_no_grow_shift = val;
1264 sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
1270 err = sysctl_handle_64(oidp, &val, 0, req);
1271 if (err != 0 || req->newptr == NULL)
1274 if (zfs_arc_max == 0) {
1275 /* Loader tunable so blindly set */
1280 if (val < arc_abs_min || val > kmem_size())
1282 if (val < arc_c_min)
1284 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
1290 arc_p = (arc_c >> 1);
1292 if (zfs_arc_meta_limit == 0) {
1293 /* limit meta-data to 1/4 of the arc capacity */
1294 arc_meta_limit = arc_c_max / 4;
1297 /* if kmem_flags are set, lets try to use less memory */
1298 if (kmem_debugging())
1301 zfs_arc_max = arc_c;
1303 mutex_enter(&arc_adjust_lock);
1304 arc_adjust_needed = B_TRUE;
1305 mutex_exit(&arc_adjust_lock);
1306 zthr_wakeup(arc_adjust_zthr);
1312 sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
1318 err = sysctl_handle_64(oidp, &val, 0, req);
1319 if (err != 0 || req->newptr == NULL)
1322 if (zfs_arc_min == 0) {
1323 /* Loader tunable so blindly set */
1328 if (val < arc_abs_min || val > arc_c_max)
1333 if (zfs_arc_meta_min == 0)
1334 arc_meta_min = arc_c_min / 2;
1336 if (arc_c < arc_c_min)
1339 zfs_arc_min = arc_c_min;
1345 #define GHOST_STATE(state) \
1346 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
1347 (state) == arc_l2c_only)
1349 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1350 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1351 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1352 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1353 #define HDR_PRESCIENT_PREFETCH(hdr) \
1354 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
1355 #define HDR_COMPRESSION_ENABLED(hdr) \
1356 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1358 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1359 #define HDR_L2_READING(hdr) \
1360 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1361 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1362 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1363 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1364 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1365 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1367 #define HDR_ISTYPE_METADATA(hdr) \
1368 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1369 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1371 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1372 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1374 /* For storing compression mode in b_flags */
1375 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1377 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1378 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1379 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1380 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1382 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1383 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1384 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1390 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1391 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1394 * Hash table routines
1397 #define HT_LOCK_PAD CACHE_LINE_SIZE
1402 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1406 #define BUF_LOCKS 256
1407 typedef struct buf_hash_table {
1409 arc_buf_hdr_t **ht_table;
1410 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1413 static buf_hash_table_t buf_hash_table;
1415 #define BUF_HASH_INDEX(spa, dva, birth) \
1416 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1417 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1418 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1419 #define HDR_LOCK(hdr) \
1420 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1422 uint64_t zfs_crc64_table[256];
1428 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1429 #define L2ARC_HEADROOM 2 /* num of writes */
1431 * If we discover during ARC scan any buffers to be compressed, we boost
1432 * our headroom for the next scanning cycle by this percentage multiple.
1434 #define L2ARC_HEADROOM_BOOST 200
1435 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1436 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1438 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1439 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1441 /* L2ARC Performance Tunables */
1442 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1443 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1444 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1445 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1446 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1447 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1448 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1449 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1450 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1452 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RWTUN,
1453 &l2arc_write_max, 0, "max write size");
1454 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RWTUN,
1455 &l2arc_write_boost, 0, "extra write during warmup");
1456 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RWTUN,
1457 &l2arc_headroom, 0, "number of dev writes");
1458 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RWTUN,
1459 &l2arc_feed_secs, 0, "interval seconds");
1460 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RWTUN,
1461 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1463 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RWTUN,
1464 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1465 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RWTUN,
1466 &l2arc_feed_again, 0, "turbo warmup");
1467 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RWTUN,
1468 &l2arc_norw, 0, "no reads during writes");
1470 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1471 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1472 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD,
1473 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1474 "size of anonymous state");
1475 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD,
1476 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1477 "size of anonymous state");
1479 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1480 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1481 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD,
1482 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1483 "size of metadata in mru state");
1484 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD,
1485 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1486 "size of data in mru state");
1488 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1489 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1490 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD,
1491 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1492 "size of metadata in mru ghost state");
1493 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD,
1494 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1495 "size of data in mru ghost state");
1497 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1498 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1499 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD,
1500 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1501 "size of metadata in mfu state");
1502 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD,
1503 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1504 "size of data in mfu state");
1506 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1507 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1508 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD,
1509 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1510 "size of metadata in mfu ghost state");
1511 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD,
1512 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1513 "size of data in mfu ghost state");
1515 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1516 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1518 SYSCTL_UINT(_vfs_zfs, OID_AUTO, arc_min_prefetch_ms, CTLFLAG_RW,
1519 &zfs_arc_min_prefetch_ms, 0, "Min life of prefetch block in ms");
1520 SYSCTL_UINT(_vfs_zfs, OID_AUTO, arc_min_prescient_prefetch_ms, CTLFLAG_RW,
1521 &zfs_arc_min_prescient_prefetch_ms, 0, "Min life of prescient prefetched block in ms");
1527 vdev_t *l2ad_vdev; /* vdev */
1528 spa_t *l2ad_spa; /* spa */
1529 uint64_t l2ad_hand; /* next write location */
1530 uint64_t l2ad_start; /* first addr on device */
1531 uint64_t l2ad_end; /* last addr on device */
1532 boolean_t l2ad_first; /* first sweep through */
1533 boolean_t l2ad_writing; /* currently writing */
1534 kmutex_t l2ad_mtx; /* lock for buffer list */
1535 list_t l2ad_buflist; /* buffer list */
1536 list_node_t l2ad_node; /* device list node */
1537 refcount_t l2ad_alloc; /* allocated bytes */
1540 static list_t L2ARC_dev_list; /* device list */
1541 static list_t *l2arc_dev_list; /* device list pointer */
1542 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1543 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1544 static list_t L2ARC_free_on_write; /* free after write buf list */
1545 static list_t *l2arc_free_on_write; /* free after write list ptr */
1546 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1547 static uint64_t l2arc_ndev; /* number of devices */
1549 typedef struct l2arc_read_callback {
1550 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1551 blkptr_t l2rcb_bp; /* original blkptr */
1552 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1553 int l2rcb_flags; /* original flags */
1554 abd_t *l2rcb_abd; /* temporary buffer */
1555 } l2arc_read_callback_t;
1557 typedef struct l2arc_write_callback {
1558 l2arc_dev_t *l2wcb_dev; /* device info */
1559 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1560 } l2arc_write_callback_t;
1562 typedef struct l2arc_data_free {
1563 /* protected by l2arc_free_on_write_mtx */
1566 arc_buf_contents_t l2df_type;
1567 list_node_t l2df_list_node;
1568 } l2arc_data_free_t;
1570 static kmutex_t l2arc_feed_thr_lock;
1571 static kcondvar_t l2arc_feed_thr_cv;
1572 static uint8_t l2arc_thread_exit;
1574 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *, boolean_t);
1575 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1576 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *, boolean_t);
1577 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1578 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1579 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1580 static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1581 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *, boolean_t);
1582 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1583 static boolean_t arc_is_overflowing();
1584 static void arc_buf_watch(arc_buf_t *);
1585 static void arc_prune_async(int64_t);
1587 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1588 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1589 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1590 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1592 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1593 static void l2arc_read_done(zio_t *);
1596 l2arc_trim(const arc_buf_hdr_t *hdr)
1598 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1600 ASSERT(HDR_HAS_L2HDR(hdr));
1601 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1603 if (HDR_GET_PSIZE(hdr) != 0) {
1604 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1605 HDR_GET_PSIZE(hdr), 0);
1610 * We use Cityhash for this. It's fast, and has good hash properties without
1611 * requiring any large static buffers.
1614 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1616 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
1619 #define HDR_EMPTY(hdr) \
1620 ((hdr)->b_dva.dva_word[0] == 0 && \
1621 (hdr)->b_dva.dva_word[1] == 0)
1623 #define HDR_EQUAL(spa, dva, birth, hdr) \
1624 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1625 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1626 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1629 buf_discard_identity(arc_buf_hdr_t *hdr)
1631 hdr->b_dva.dva_word[0] = 0;
1632 hdr->b_dva.dva_word[1] = 0;
1636 static arc_buf_hdr_t *
1637 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1639 const dva_t *dva = BP_IDENTITY(bp);
1640 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1641 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1642 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1645 mutex_enter(hash_lock);
1646 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1647 hdr = hdr->b_hash_next) {
1648 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1653 mutex_exit(hash_lock);
1659 * Insert an entry into the hash table. If there is already an element
1660 * equal to elem in the hash table, then the already existing element
1661 * will be returned and the new element will not be inserted.
1662 * Otherwise returns NULL.
1663 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1665 static arc_buf_hdr_t *
1666 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1668 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1669 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1670 arc_buf_hdr_t *fhdr;
1673 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1674 ASSERT(hdr->b_birth != 0);
1675 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1677 if (lockp != NULL) {
1679 mutex_enter(hash_lock);
1681 ASSERT(MUTEX_HELD(hash_lock));
1684 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1685 fhdr = fhdr->b_hash_next, i++) {
1686 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1690 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1691 buf_hash_table.ht_table[idx] = hdr;
1692 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1694 /* collect some hash table performance data */
1696 ARCSTAT_BUMP(arcstat_hash_collisions);
1698 ARCSTAT_BUMP(arcstat_hash_chains);
1700 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1703 ARCSTAT_BUMP(arcstat_hash_elements);
1704 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1710 buf_hash_remove(arc_buf_hdr_t *hdr)
1712 arc_buf_hdr_t *fhdr, **hdrp;
1713 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1715 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1716 ASSERT(HDR_IN_HASH_TABLE(hdr));
1718 hdrp = &buf_hash_table.ht_table[idx];
1719 while ((fhdr = *hdrp) != hdr) {
1720 ASSERT3P(fhdr, !=, NULL);
1721 hdrp = &fhdr->b_hash_next;
1723 *hdrp = hdr->b_hash_next;
1724 hdr->b_hash_next = NULL;
1725 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1727 /* collect some hash table performance data */
1728 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1730 if (buf_hash_table.ht_table[idx] &&
1731 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1732 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1736 * Global data structures and functions for the buf kmem cache.
1738 static kmem_cache_t *hdr_full_cache;
1739 static kmem_cache_t *hdr_l2only_cache;
1740 static kmem_cache_t *buf_cache;
1747 kmem_free(buf_hash_table.ht_table,
1748 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1749 for (i = 0; i < BUF_LOCKS; i++)
1750 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1751 kmem_cache_destroy(hdr_full_cache);
1752 kmem_cache_destroy(hdr_l2only_cache);
1753 kmem_cache_destroy(buf_cache);
1757 * Constructor callback - called when the cache is empty
1758 * and a new buf is requested.
1762 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1764 arc_buf_hdr_t *hdr = vbuf;
1766 bzero(hdr, HDR_FULL_SIZE);
1767 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1768 refcount_create(&hdr->b_l1hdr.b_refcnt);
1769 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1770 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1771 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1778 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1780 arc_buf_hdr_t *hdr = vbuf;
1782 bzero(hdr, HDR_L2ONLY_SIZE);
1783 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1790 buf_cons(void *vbuf, void *unused, int kmflag)
1792 arc_buf_t *buf = vbuf;
1794 bzero(buf, sizeof (arc_buf_t));
1795 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1796 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1802 * Destructor callback - called when a cached buf is
1803 * no longer required.
1807 hdr_full_dest(void *vbuf, void *unused)
1809 arc_buf_hdr_t *hdr = vbuf;
1811 ASSERT(HDR_EMPTY(hdr));
1812 cv_destroy(&hdr->b_l1hdr.b_cv);
1813 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1814 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1815 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1816 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1821 hdr_l2only_dest(void *vbuf, void *unused)
1823 arc_buf_hdr_t *hdr = vbuf;
1825 ASSERT(HDR_EMPTY(hdr));
1826 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1831 buf_dest(void *vbuf, void *unused)
1833 arc_buf_t *buf = vbuf;
1835 mutex_destroy(&buf->b_evict_lock);
1836 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1840 * Reclaim callback -- invoked when memory is low.
1844 hdr_recl(void *unused)
1846 dprintf("hdr_recl called\n");
1848 * umem calls the reclaim func when we destroy the buf cache,
1849 * which is after we do arc_fini().
1851 if (arc_initialized)
1852 zthr_wakeup(arc_reap_zthr);
1859 uint64_t hsize = 1ULL << 12;
1863 * The hash table is big enough to fill all of physical memory
1864 * with an average block size of zfs_arc_average_blocksize (default 8K).
1865 * By default, the table will take up
1866 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1868 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1871 buf_hash_table.ht_mask = hsize - 1;
1872 buf_hash_table.ht_table =
1873 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1874 if (buf_hash_table.ht_table == NULL) {
1875 ASSERT(hsize > (1ULL << 8));
1880 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1881 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1882 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1883 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1885 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1886 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1888 for (i = 0; i < 256; i++)
1889 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1890 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1892 for (i = 0; i < BUF_LOCKS; i++) {
1893 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1894 NULL, MUTEX_DEFAULT, NULL);
1899 * This is the size that the buf occupies in memory. If the buf is compressed,
1900 * it will correspond to the compressed size. You should use this method of
1901 * getting the buf size unless you explicitly need the logical size.
1904 arc_buf_size(arc_buf_t *buf)
1906 return (ARC_BUF_COMPRESSED(buf) ?
1907 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1911 arc_buf_lsize(arc_buf_t *buf)
1913 return (HDR_GET_LSIZE(buf->b_hdr));
1917 arc_get_compression(arc_buf_t *buf)
1919 return (ARC_BUF_COMPRESSED(buf) ?
1920 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1923 #define ARC_MINTIME (hz>>4) /* 62 ms */
1925 static inline boolean_t
1926 arc_buf_is_shared(arc_buf_t *buf)
1928 boolean_t shared = (buf->b_data != NULL &&
1929 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1930 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1931 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1932 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1933 IMPLY(shared, ARC_BUF_SHARED(buf));
1934 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1937 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1938 * already being shared" requirement prevents us from doing that.
1945 * Free the checksum associated with this header. If there is no checksum, this
1949 arc_cksum_free(arc_buf_hdr_t *hdr)
1951 ASSERT(HDR_HAS_L1HDR(hdr));
1952 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1953 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1954 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1955 hdr->b_l1hdr.b_freeze_cksum = NULL;
1957 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1961 * Return true iff at least one of the bufs on hdr is not compressed.
1964 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1966 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1967 if (!ARC_BUF_COMPRESSED(b)) {
1975 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1976 * matches the checksum that is stored in the hdr. If there is no checksum,
1977 * or if the buf is compressed, this is a no-op.
1980 arc_cksum_verify(arc_buf_t *buf)
1982 arc_buf_hdr_t *hdr = buf->b_hdr;
1985 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1988 if (ARC_BUF_COMPRESSED(buf)) {
1989 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1990 arc_hdr_has_uncompressed_buf(hdr));
1994 ASSERT(HDR_HAS_L1HDR(hdr));
1996 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1997 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1998 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2002 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
2003 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
2004 panic("buffer modified while frozen!");
2005 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2009 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
2011 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
2012 boolean_t valid_cksum;
2014 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
2015 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
2018 * We rely on the blkptr's checksum to determine if the block
2019 * is valid or not. When compressed arc is enabled, the l2arc
2020 * writes the block to the l2arc just as it appears in the pool.
2021 * This allows us to use the blkptr's checksum to validate the
2022 * data that we just read off of the l2arc without having to store
2023 * a separate checksum in the arc_buf_hdr_t. However, if compressed
2024 * arc is disabled, then the data written to the l2arc is always
2025 * uncompressed and won't match the block as it exists in the main
2026 * pool. When this is the case, we must first compress it if it is
2027 * compressed on the main pool before we can validate the checksum.
2029 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
2030 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2031 uint64_t lsize = HDR_GET_LSIZE(hdr);
2034 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
2035 csize = zio_compress_data(compress, zio->io_abd,
2036 abd_to_buf(cdata), lsize);
2038 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
2039 if (csize < HDR_GET_PSIZE(hdr)) {
2041 * Compressed blocks are always a multiple of the
2042 * smallest ashift in the pool. Ideally, we would
2043 * like to round up the csize to the next
2044 * spa_min_ashift but that value may have changed
2045 * since the block was last written. Instead,
2046 * we rely on the fact that the hdr's psize
2047 * was set to the psize of the block when it was
2048 * last written. We set the csize to that value
2049 * and zero out any part that should not contain
2052 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
2053 csize = HDR_GET_PSIZE(hdr);
2055 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
2059 * Block pointers always store the checksum for the logical data.
2060 * If the block pointer has the gang bit set, then the checksum
2061 * it represents is for the reconstituted data and not for an
2062 * individual gang member. The zio pipeline, however, must be able to
2063 * determine the checksum of each of the gang constituents so it
2064 * treats the checksum comparison differently than what we need
2065 * for l2arc blocks. This prevents us from using the
2066 * zio_checksum_error() interface directly. Instead we must call the
2067 * zio_checksum_error_impl() so that we can ensure the checksum is
2068 * generated using the correct checksum algorithm and accounts for the
2069 * logical I/O size and not just a gang fragment.
2071 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
2072 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
2073 zio->io_offset, NULL) == 0);
2074 zio_pop_transforms(zio);
2075 return (valid_cksum);
2079 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
2080 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
2081 * isn't modified later on. If buf is compressed or there is already a checksum
2082 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
2085 arc_cksum_compute(arc_buf_t *buf)
2087 arc_buf_hdr_t *hdr = buf->b_hdr;
2089 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2092 ASSERT(HDR_HAS_L1HDR(hdr));
2094 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
2095 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
2096 ASSERT(arc_hdr_has_uncompressed_buf(hdr));
2097 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2099 } else if (ARC_BUF_COMPRESSED(buf)) {
2100 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2104 ASSERT(!ARC_BUF_COMPRESSED(buf));
2105 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
2107 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
2108 hdr->b_l1hdr.b_freeze_cksum);
2109 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2117 typedef struct procctl {
2125 arc_buf_unwatch(arc_buf_t *buf)
2132 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2133 ctl.prwatch.pr_size = 0;
2134 ctl.prwatch.pr_wflags = 0;
2135 result = write(arc_procfd, &ctl, sizeof (ctl));
2136 ASSERT3U(result, ==, sizeof (ctl));
2143 arc_buf_watch(arc_buf_t *buf)
2150 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2151 ctl.prwatch.pr_size = arc_buf_size(buf);
2152 ctl.prwatch.pr_wflags = WA_WRITE;
2153 result = write(arc_procfd, &ctl, sizeof (ctl));
2154 ASSERT3U(result, ==, sizeof (ctl));
2158 #endif /* illumos */
2160 static arc_buf_contents_t
2161 arc_buf_type(arc_buf_hdr_t *hdr)
2163 arc_buf_contents_t type;
2164 if (HDR_ISTYPE_METADATA(hdr)) {
2165 type = ARC_BUFC_METADATA;
2167 type = ARC_BUFC_DATA;
2169 VERIFY3U(hdr->b_type, ==, type);
2174 arc_is_metadata(arc_buf_t *buf)
2176 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
2180 arc_bufc_to_flags(arc_buf_contents_t type)
2184 /* metadata field is 0 if buffer contains normal data */
2186 case ARC_BUFC_METADATA:
2187 return (ARC_FLAG_BUFC_METADATA);
2191 panic("undefined ARC buffer type!");
2192 return ((uint32_t)-1);
2196 arc_buf_thaw(arc_buf_t *buf)
2198 arc_buf_hdr_t *hdr = buf->b_hdr;
2200 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2201 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2203 arc_cksum_verify(buf);
2206 * Compressed buffers do not manipulate the b_freeze_cksum or
2207 * allocate b_thawed.
2209 if (ARC_BUF_COMPRESSED(buf)) {
2210 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2211 arc_hdr_has_uncompressed_buf(hdr));
2215 ASSERT(HDR_HAS_L1HDR(hdr));
2216 arc_cksum_free(hdr);
2218 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
2220 if (zfs_flags & ZFS_DEBUG_MODIFY) {
2221 if (hdr->b_l1hdr.b_thawed != NULL)
2222 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2223 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
2227 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2230 arc_buf_unwatch(buf);
2235 arc_buf_freeze(arc_buf_t *buf)
2237 arc_buf_hdr_t *hdr = buf->b_hdr;
2238 kmutex_t *hash_lock;
2240 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2243 if (ARC_BUF_COMPRESSED(buf)) {
2244 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2245 arc_hdr_has_uncompressed_buf(hdr));
2249 hash_lock = HDR_LOCK(hdr);
2250 mutex_enter(hash_lock);
2252 ASSERT(HDR_HAS_L1HDR(hdr));
2253 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
2254 hdr->b_l1hdr.b_state == arc_anon);
2255 arc_cksum_compute(buf);
2256 mutex_exit(hash_lock);
2260 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
2261 * the following functions should be used to ensure that the flags are
2262 * updated in a thread-safe way. When manipulating the flags either
2263 * the hash_lock must be held or the hdr must be undiscoverable. This
2264 * ensures that we're not racing with any other threads when updating
2268 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2270 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2271 hdr->b_flags |= flags;
2275 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2277 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2278 hdr->b_flags &= ~flags;
2282 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
2283 * done in a special way since we have to clear and set bits
2284 * at the same time. Consumers that wish to set the compression bits
2285 * must use this function to ensure that the flags are updated in
2286 * thread-safe manner.
2289 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
2291 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2294 * Holes and embedded blocks will always have a psize = 0 so
2295 * we ignore the compression of the blkptr and set the
2296 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
2297 * Holes and embedded blocks remain anonymous so we don't
2298 * want to uncompress them. Mark them as uncompressed.
2300 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
2301 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2302 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
2303 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
2304 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2306 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2307 HDR_SET_COMPRESS(hdr, cmp);
2308 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
2309 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
2314 * Looks for another buf on the same hdr which has the data decompressed, copies
2315 * from it, and returns true. If no such buf exists, returns false.
2318 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
2320 arc_buf_hdr_t *hdr = buf->b_hdr;
2321 boolean_t copied = B_FALSE;
2323 ASSERT(HDR_HAS_L1HDR(hdr));
2324 ASSERT3P(buf->b_data, !=, NULL);
2325 ASSERT(!ARC_BUF_COMPRESSED(buf));
2327 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
2328 from = from->b_next) {
2329 /* can't use our own data buffer */
2334 if (!ARC_BUF_COMPRESSED(from)) {
2335 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
2342 * There were no decompressed bufs, so there should not be a
2343 * checksum on the hdr either.
2345 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
2351 * Given a buf that has a data buffer attached to it, this function will
2352 * efficiently fill the buf with data of the specified compression setting from
2353 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2354 * are already sharing a data buf, no copy is performed.
2356 * If the buf is marked as compressed but uncompressed data was requested, this
2357 * will allocate a new data buffer for the buf, remove that flag, and fill the
2358 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2359 * uncompressed data, and (since we haven't added support for it yet) if you
2360 * want compressed data your buf must already be marked as compressed and have
2361 * the correct-sized data buffer.
2364 arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
2366 arc_buf_hdr_t *hdr = buf->b_hdr;
2367 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
2368 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2370 ASSERT3P(buf->b_data, !=, NULL);
2371 IMPLY(compressed, hdr_compressed);
2372 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2374 if (hdr_compressed == compressed) {
2375 if (!arc_buf_is_shared(buf)) {
2376 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2380 ASSERT(hdr_compressed);
2381 ASSERT(!compressed);
2382 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2385 * If the buf is sharing its data with the hdr, unlink it and
2386 * allocate a new data buffer for the buf.
2388 if (arc_buf_is_shared(buf)) {
2389 ASSERT(ARC_BUF_COMPRESSED(buf));
2391 /* We need to give the buf it's own b_data */
2392 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2394 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2395 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2397 /* Previously overhead was 0; just add new overhead */
2398 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2399 } else if (ARC_BUF_COMPRESSED(buf)) {
2400 /* We need to reallocate the buf's b_data */
2401 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2404 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2406 /* We increased the size of b_data; update overhead */
2407 ARCSTAT_INCR(arcstat_overhead_size,
2408 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2412 * Regardless of the buf's previous compression settings, it
2413 * should not be compressed at the end of this function.
2415 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2418 * Try copying the data from another buf which already has a
2419 * decompressed version. If that's not possible, it's time to
2420 * bite the bullet and decompress the data from the hdr.
2422 if (arc_buf_try_copy_decompressed_data(buf)) {
2423 /* Skip byteswapping and checksumming (already done) */
2424 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2427 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2428 hdr->b_l1hdr.b_pabd, buf->b_data,
2429 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2432 * Absent hardware errors or software bugs, this should
2433 * be impossible, but log it anyway so we can debug it.
2437 "hdr %p, compress %d, psize %d, lsize %d",
2438 hdr, HDR_GET_COMPRESS(hdr),
2439 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2440 return (SET_ERROR(EIO));
2445 /* Byteswap the buf's data if necessary */
2446 if (bswap != DMU_BSWAP_NUMFUNCS) {
2447 ASSERT(!HDR_SHARED_DATA(hdr));
2448 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2449 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2452 /* Compute the hdr's checksum if necessary */
2453 arc_cksum_compute(buf);
2459 arc_decompress(arc_buf_t *buf)
2461 return (arc_buf_fill(buf, B_FALSE));
2465 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
2468 arc_hdr_size(arc_buf_hdr_t *hdr)
2472 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2473 HDR_GET_PSIZE(hdr) > 0) {
2474 size = HDR_GET_PSIZE(hdr);
2476 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2477 size = HDR_GET_LSIZE(hdr);
2483 * Increment the amount of evictable space in the arc_state_t's refcount.
2484 * We account for the space used by the hdr and the arc buf individually
2485 * so that we can add and remove them from the refcount individually.
2488 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2490 arc_buf_contents_t type = arc_buf_type(hdr);
2492 ASSERT(HDR_HAS_L1HDR(hdr));
2494 if (GHOST_STATE(state)) {
2495 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2496 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2497 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2498 (void) refcount_add_many(&state->arcs_esize[type],
2499 HDR_GET_LSIZE(hdr), hdr);
2503 ASSERT(!GHOST_STATE(state));
2504 if (hdr->b_l1hdr.b_pabd != NULL) {
2505 (void) refcount_add_many(&state->arcs_esize[type],
2506 arc_hdr_size(hdr), hdr);
2508 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2509 buf = buf->b_next) {
2510 if (arc_buf_is_shared(buf))
2512 (void) refcount_add_many(&state->arcs_esize[type],
2513 arc_buf_size(buf), buf);
2518 * Decrement the amount of evictable space in the arc_state_t's refcount.
2519 * We account for the space used by the hdr and the arc buf individually
2520 * so that we can add and remove them from the refcount individually.
2523 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2525 arc_buf_contents_t type = arc_buf_type(hdr);
2527 ASSERT(HDR_HAS_L1HDR(hdr));
2529 if (GHOST_STATE(state)) {
2530 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2531 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2532 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2533 (void) refcount_remove_many(&state->arcs_esize[type],
2534 HDR_GET_LSIZE(hdr), hdr);
2538 ASSERT(!GHOST_STATE(state));
2539 if (hdr->b_l1hdr.b_pabd != NULL) {
2540 (void) refcount_remove_many(&state->arcs_esize[type],
2541 arc_hdr_size(hdr), hdr);
2543 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2544 buf = buf->b_next) {
2545 if (arc_buf_is_shared(buf))
2547 (void) refcount_remove_many(&state->arcs_esize[type],
2548 arc_buf_size(buf), buf);
2553 * Add a reference to this hdr indicating that someone is actively
2554 * referencing that memory. When the refcount transitions from 0 to 1,
2555 * we remove it from the respective arc_state_t list to indicate that
2556 * it is not evictable.
2559 add_reference(arc_buf_hdr_t *hdr, void *tag)
2561 ASSERT(HDR_HAS_L1HDR(hdr));
2562 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2563 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2564 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2565 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2568 arc_state_t *state = hdr->b_l1hdr.b_state;
2570 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2571 (state != arc_anon)) {
2572 /* We don't use the L2-only state list. */
2573 if (state != arc_l2c_only) {
2574 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2576 arc_evictable_space_decrement(hdr, state);
2578 /* remove the prefetch flag if we get a reference */
2579 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2584 * Remove a reference from this hdr. When the reference transitions from
2585 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2586 * list making it eligible for eviction.
2589 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2592 arc_state_t *state = hdr->b_l1hdr.b_state;
2594 ASSERT(HDR_HAS_L1HDR(hdr));
2595 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2596 ASSERT(!GHOST_STATE(state));
2599 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2600 * check to prevent usage of the arc_l2c_only list.
2602 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2603 (state != arc_anon)) {
2604 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2605 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2606 arc_evictable_space_increment(hdr, state);
2612 * Returns detailed information about a specific arc buffer. When the
2613 * state_index argument is set the function will calculate the arc header
2614 * list position for its arc state. Since this requires a linear traversal
2615 * callers are strongly encourage not to do this. However, it can be helpful
2616 * for targeted analysis so the functionality is provided.
2619 arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
2621 arc_buf_hdr_t *hdr = ab->b_hdr;
2622 l1arc_buf_hdr_t *l1hdr = NULL;
2623 l2arc_buf_hdr_t *l2hdr = NULL;
2624 arc_state_t *state = NULL;
2626 memset(abi, 0, sizeof (arc_buf_info_t));
2631 abi->abi_flags = hdr->b_flags;
2633 if (HDR_HAS_L1HDR(hdr)) {
2634 l1hdr = &hdr->b_l1hdr;
2635 state = l1hdr->b_state;
2637 if (HDR_HAS_L2HDR(hdr))
2638 l2hdr = &hdr->b_l2hdr;
2641 abi->abi_bufcnt = l1hdr->b_bufcnt;
2642 abi->abi_access = l1hdr->b_arc_access;
2643 abi->abi_mru_hits = l1hdr->b_mru_hits;
2644 abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits;
2645 abi->abi_mfu_hits = l1hdr->b_mfu_hits;
2646 abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits;
2647 abi->abi_holds = refcount_count(&l1hdr->b_refcnt);
2651 abi->abi_l2arc_dattr = l2hdr->b_daddr;
2652 abi->abi_l2arc_hits = l2hdr->b_hits;
2655 abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
2656 abi->abi_state_contents = arc_buf_type(hdr);
2657 abi->abi_size = arc_hdr_size(hdr);
2661 * Move the supplied buffer to the indicated state. The hash lock
2662 * for the buffer must be held by the caller.
2665 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2666 kmutex_t *hash_lock)
2668 arc_state_t *old_state;
2671 boolean_t update_old, update_new;
2672 arc_buf_contents_t buftype = arc_buf_type(hdr);
2675 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2676 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2677 * L1 hdr doesn't always exist when we change state to arc_anon before
2678 * destroying a header, in which case reallocating to add the L1 hdr is
2681 if (HDR_HAS_L1HDR(hdr)) {
2682 old_state = hdr->b_l1hdr.b_state;
2683 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2684 bufcnt = hdr->b_l1hdr.b_bufcnt;
2685 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL);
2687 old_state = arc_l2c_only;
2690 update_old = B_FALSE;
2692 update_new = update_old;
2694 ASSERT(MUTEX_HELD(hash_lock));
2695 ASSERT3P(new_state, !=, old_state);
2696 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2697 ASSERT(old_state != arc_anon || bufcnt <= 1);
2700 * If this buffer is evictable, transfer it from the
2701 * old state list to the new state list.
2704 if (old_state != arc_anon && old_state != arc_l2c_only) {
2705 ASSERT(HDR_HAS_L1HDR(hdr));
2706 multilist_remove(old_state->arcs_list[buftype], hdr);
2708 if (GHOST_STATE(old_state)) {
2710 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2711 update_old = B_TRUE;
2713 arc_evictable_space_decrement(hdr, old_state);
2715 if (new_state != arc_anon && new_state != arc_l2c_only) {
2718 * An L1 header always exists here, since if we're
2719 * moving to some L1-cached state (i.e. not l2c_only or
2720 * anonymous), we realloc the header to add an L1hdr
2723 ASSERT(HDR_HAS_L1HDR(hdr));
2724 multilist_insert(new_state->arcs_list[buftype], hdr);
2726 if (GHOST_STATE(new_state)) {
2728 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2729 update_new = B_TRUE;
2731 arc_evictable_space_increment(hdr, new_state);
2735 ASSERT(!HDR_EMPTY(hdr));
2736 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2737 buf_hash_remove(hdr);
2739 /* adjust state sizes (ignore arc_l2c_only) */
2741 if (update_new && new_state != arc_l2c_only) {
2742 ASSERT(HDR_HAS_L1HDR(hdr));
2743 if (GHOST_STATE(new_state)) {
2747 * When moving a header to a ghost state, we first
2748 * remove all arc buffers. Thus, we'll have a
2749 * bufcnt of zero, and no arc buffer to use for
2750 * the reference. As a result, we use the arc
2751 * header pointer for the reference.
2753 (void) refcount_add_many(&new_state->arcs_size,
2754 HDR_GET_LSIZE(hdr), hdr);
2755 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2757 uint32_t buffers = 0;
2760 * Each individual buffer holds a unique reference,
2761 * thus we must remove each of these references one
2764 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2765 buf = buf->b_next) {
2766 ASSERT3U(bufcnt, !=, 0);
2770 * When the arc_buf_t is sharing the data
2771 * block with the hdr, the owner of the
2772 * reference belongs to the hdr. Only
2773 * add to the refcount if the arc_buf_t is
2776 if (arc_buf_is_shared(buf))
2779 (void) refcount_add_many(&new_state->arcs_size,
2780 arc_buf_size(buf), buf);
2782 ASSERT3U(bufcnt, ==, buffers);
2784 if (hdr->b_l1hdr.b_pabd != NULL) {
2785 (void) refcount_add_many(&new_state->arcs_size,
2786 arc_hdr_size(hdr), hdr);
2788 ASSERT(GHOST_STATE(old_state));
2793 if (update_old && old_state != arc_l2c_only) {
2794 ASSERT(HDR_HAS_L1HDR(hdr));
2795 if (GHOST_STATE(old_state)) {
2797 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2800 * When moving a header off of a ghost state,
2801 * the header will not contain any arc buffers.
2802 * We use the arc header pointer for the reference
2803 * which is exactly what we did when we put the
2804 * header on the ghost state.
2807 (void) refcount_remove_many(&old_state->arcs_size,
2808 HDR_GET_LSIZE(hdr), hdr);
2810 uint32_t buffers = 0;
2813 * Each individual buffer holds a unique reference,
2814 * thus we must remove each of these references one
2817 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2818 buf = buf->b_next) {
2819 ASSERT3U(bufcnt, !=, 0);
2823 * When the arc_buf_t is sharing the data
2824 * block with the hdr, the owner of the
2825 * reference belongs to the hdr. Only
2826 * add to the refcount if the arc_buf_t is
2829 if (arc_buf_is_shared(buf))
2832 (void) refcount_remove_many(
2833 &old_state->arcs_size, arc_buf_size(buf),
2836 ASSERT3U(bufcnt, ==, buffers);
2837 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2838 (void) refcount_remove_many(
2839 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2843 if (HDR_HAS_L1HDR(hdr))
2844 hdr->b_l1hdr.b_state = new_state;
2847 * L2 headers should never be on the L2 state list since they don't
2848 * have L1 headers allocated.
2850 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2851 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2855 arc_space_consume(uint64_t space, arc_space_type_t type)
2857 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2860 case ARC_SPACE_DATA:
2861 aggsum_add(&astat_data_size, space);
2863 case ARC_SPACE_META:
2864 aggsum_add(&astat_metadata_size, space);
2866 case ARC_SPACE_BONUS:
2867 aggsum_add(&astat_bonus_size, space);
2869 case ARC_SPACE_DNODE:
2870 aggsum_add(&astat_dnode_size, space);
2872 case ARC_SPACE_DBUF:
2873 aggsum_add(&astat_dbuf_size, space);
2875 case ARC_SPACE_HDRS:
2876 aggsum_add(&astat_hdr_size, space);
2878 case ARC_SPACE_L2HDRS:
2879 aggsum_add(&astat_l2_hdr_size, space);
2883 if (type != ARC_SPACE_DATA)
2884 aggsum_add(&arc_meta_used, space);
2886 aggsum_add(&arc_size, space);
2890 arc_space_return(uint64_t space, arc_space_type_t type)
2892 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2895 case ARC_SPACE_DATA:
2896 aggsum_add(&astat_data_size, -space);
2898 case ARC_SPACE_META:
2899 aggsum_add(&astat_metadata_size, -space);
2901 case ARC_SPACE_BONUS:
2902 aggsum_add(&astat_bonus_size, -space);
2904 case ARC_SPACE_DNODE:
2905 aggsum_add(&astat_dnode_size, -space);
2907 case ARC_SPACE_DBUF:
2908 aggsum_add(&astat_dbuf_size, -space);
2910 case ARC_SPACE_HDRS:
2911 aggsum_add(&astat_hdr_size, -space);
2913 case ARC_SPACE_L2HDRS:
2914 aggsum_add(&astat_l2_hdr_size, -space);
2918 if (type != ARC_SPACE_DATA) {
2919 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2921 * We use the upper bound here rather than the precise value
2922 * because the arc_meta_max value doesn't need to be
2923 * precise. It's only consumed by humans via arcstats.
2925 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2926 arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2927 aggsum_add(&arc_meta_used, -space);
2930 ASSERT(aggsum_compare(&arc_size, space) >= 0);
2931 aggsum_add(&arc_size, -space);
2935 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2936 * with the hdr's b_pabd.
2939 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2942 * The criteria for sharing a hdr's data are:
2943 * 1. the hdr's compression matches the buf's compression
2944 * 2. the hdr doesn't need to be byteswapped
2945 * 3. the hdr isn't already being shared
2946 * 4. the buf is either compressed or it is the last buf in the hdr list
2948 * Criterion #4 maintains the invariant that shared uncompressed
2949 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2950 * might ask, "if a compressed buf is allocated first, won't that be the
2951 * last thing in the list?", but in that case it's impossible to create
2952 * a shared uncompressed buf anyway (because the hdr must be compressed
2953 * to have the compressed buf). You might also think that #3 is
2954 * sufficient to make this guarantee, however it's possible
2955 * (specifically in the rare L2ARC write race mentioned in
2956 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2957 * is sharable, but wasn't at the time of its allocation. Rather than
2958 * allow a new shared uncompressed buf to be created and then shuffle
2959 * the list around to make it the last element, this simply disallows
2960 * sharing if the new buf isn't the first to be added.
2962 ASSERT3P(buf->b_hdr, ==, hdr);
2963 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF;
2964 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2965 return (buf_compressed == hdr_compressed &&
2966 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2967 !HDR_SHARED_DATA(hdr) &&
2968 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2972 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2973 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2974 * copy was made successfully, or an error code otherwise.
2977 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2978 boolean_t fill, arc_buf_t **ret)
2982 ASSERT(HDR_HAS_L1HDR(hdr));
2983 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2984 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2985 hdr->b_type == ARC_BUFC_METADATA);
2986 ASSERT3P(ret, !=, NULL);
2987 ASSERT3P(*ret, ==, NULL);
2989 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2992 buf->b_next = hdr->b_l1hdr.b_buf;
2995 add_reference(hdr, tag);
2998 * We're about to change the hdr's b_flags. We must either
2999 * hold the hash_lock or be undiscoverable.
3001 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3004 * Only honor requests for compressed bufs if the hdr is actually
3007 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
3008 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
3011 * If the hdr's data can be shared then we share the data buffer and
3012 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
3013 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
3014 * buffer to store the buf's data.
3016 * There are two additional restrictions here because we're sharing
3017 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
3018 * actively involved in an L2ARC write, because if this buf is used by
3019 * an arc_write() then the hdr's data buffer will be released when the
3020 * write completes, even though the L2ARC write might still be using it.
3021 * Second, the hdr's ABD must be linear so that the buf's user doesn't
3022 * need to be ABD-aware.
3024 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
3025 abd_is_linear(hdr->b_l1hdr.b_pabd);
3027 /* Set up b_data and sharing */
3029 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
3030 buf->b_flags |= ARC_BUF_FLAG_SHARED;
3031 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
3034 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
3035 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3037 VERIFY3P(buf->b_data, !=, NULL);
3039 hdr->b_l1hdr.b_buf = buf;
3040 hdr->b_l1hdr.b_bufcnt += 1;
3043 * If the user wants the data from the hdr, we need to either copy or
3044 * decompress the data.
3047 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
3053 static char *arc_onloan_tag = "onloan";
3056 arc_loaned_bytes_update(int64_t delta)
3058 atomic_add_64(&arc_loaned_bytes, delta);
3060 /* assert that it did not wrap around */
3061 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
3065 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
3066 * flight data by arc_tempreserve_space() until they are "returned". Loaned
3067 * buffers must be returned to the arc before they can be used by the DMU or
3071 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
3073 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
3074 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
3076 arc_loaned_bytes_update(arc_buf_size(buf));
3082 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
3083 enum zio_compress compression_type)
3085 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
3086 psize, lsize, compression_type);
3088 arc_loaned_bytes_update(arc_buf_size(buf));
3095 * Return a loaned arc buffer to the arc.
3098 arc_return_buf(arc_buf_t *buf, void *tag)
3100 arc_buf_hdr_t *hdr = buf->b_hdr;
3102 ASSERT3P(buf->b_data, !=, NULL);
3103 ASSERT(HDR_HAS_L1HDR(hdr));
3104 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
3105 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
3107 arc_loaned_bytes_update(-arc_buf_size(buf));
3110 /* Detach an arc_buf from a dbuf (tag) */
3112 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
3114 arc_buf_hdr_t *hdr = buf->b_hdr;
3116 ASSERT3P(buf->b_data, !=, NULL);
3117 ASSERT(HDR_HAS_L1HDR(hdr));
3118 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
3119 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
3121 arc_loaned_bytes_update(arc_buf_size(buf));
3125 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
3127 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
3130 df->l2df_size = size;
3131 df->l2df_type = type;
3132 mutex_enter(&l2arc_free_on_write_mtx);
3133 list_insert_head(l2arc_free_on_write, df);
3134 mutex_exit(&l2arc_free_on_write_mtx);
3138 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
3140 arc_state_t *state = hdr->b_l1hdr.b_state;
3141 arc_buf_contents_t type = arc_buf_type(hdr);
3142 uint64_t size = arc_hdr_size(hdr);
3144 /* protected by hash lock, if in the hash table */
3145 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3146 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3147 ASSERT(state != arc_anon && state != arc_l2c_only);
3149 (void) refcount_remove_many(&state->arcs_esize[type],
3152 (void) refcount_remove_many(&state->arcs_size, size, hdr);
3153 if (type == ARC_BUFC_METADATA) {
3154 arc_space_return(size, ARC_SPACE_META);
3156 ASSERT(type == ARC_BUFC_DATA);
3157 arc_space_return(size, ARC_SPACE_DATA);
3160 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
3164 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
3165 * data buffer, we transfer the refcount ownership to the hdr and update
3166 * the appropriate kstats.
3169 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3171 arc_state_t *state = hdr->b_l1hdr.b_state;
3173 ASSERT(arc_can_share(hdr, buf));
3174 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3175 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3178 * Start sharing the data buffer. We transfer the
3179 * refcount ownership to the hdr since it always owns
3180 * the refcount whenever an arc_buf_t is shared.
3182 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
3183 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
3184 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
3185 HDR_ISTYPE_METADATA(hdr));
3186 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
3187 buf->b_flags |= ARC_BUF_FLAG_SHARED;
3190 * Since we've transferred ownership to the hdr we need
3191 * to increment its compressed and uncompressed kstats and
3192 * decrement the overhead size.
3194 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3195 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3196 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
3200 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3202 arc_state_t *state = hdr->b_l1hdr.b_state;
3204 ASSERT(arc_buf_is_shared(buf));
3205 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3206 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3209 * We are no longer sharing this buffer so we need
3210 * to transfer its ownership to the rightful owner.
3212 refcount_transfer_ownership(&state->arcs_size, hdr, buf);
3213 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3214 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3215 abd_put(hdr->b_l1hdr.b_pabd);
3216 hdr->b_l1hdr.b_pabd = NULL;
3217 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3220 * Since the buffer is no longer shared between
3221 * the arc buf and the hdr, count it as overhead.
3223 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3224 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3225 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3229 * Remove an arc_buf_t from the hdr's buf list and return the last
3230 * arc_buf_t on the list. If no buffers remain on the list then return
3234 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3236 ASSERT(HDR_HAS_L1HDR(hdr));
3237 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3239 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3240 arc_buf_t *lastbuf = NULL;
3243 * Remove the buf from the hdr list and locate the last
3244 * remaining buffer on the list.
3246 while (*bufp != NULL) {
3248 *bufp = buf->b_next;
3251 * If we've removed a buffer in the middle of
3252 * the list then update the lastbuf and update
3255 if (*bufp != NULL) {
3257 bufp = &(*bufp)->b_next;
3261 ASSERT3P(lastbuf, !=, buf);
3262 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3263 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3264 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3270 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3274 arc_buf_destroy_impl(arc_buf_t *buf)
3276 arc_buf_hdr_t *hdr = buf->b_hdr;
3279 * Free up the data associated with the buf but only if we're not
3280 * sharing this with the hdr. If we are sharing it with the hdr, the
3281 * hdr is responsible for doing the free.
3283 if (buf->b_data != NULL) {
3285 * We're about to change the hdr's b_flags. We must either
3286 * hold the hash_lock or be undiscoverable.
3288 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3290 arc_cksum_verify(buf);
3292 arc_buf_unwatch(buf);
3295 if (arc_buf_is_shared(buf)) {
3296 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3298 uint64_t size = arc_buf_size(buf);
3299 arc_free_data_buf(hdr, buf->b_data, size, buf);
3300 ARCSTAT_INCR(arcstat_overhead_size, -size);
3304 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3305 hdr->b_l1hdr.b_bufcnt -= 1;
3308 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3310 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3312 * If the current arc_buf_t is sharing its data buffer with the
3313 * hdr, then reassign the hdr's b_pabd to share it with the new
3314 * buffer at the end of the list. The shared buffer is always
3315 * the last one on the hdr's buffer list.
3317 * There is an equivalent case for compressed bufs, but since
3318 * they aren't guaranteed to be the last buf in the list and
3319 * that is an exceedingly rare case, we just allow that space be
3320 * wasted temporarily.
3322 if (lastbuf != NULL) {
3323 /* Only one buf can be shared at once */
3324 VERIFY(!arc_buf_is_shared(lastbuf));
3325 /* hdr is uncompressed so can't have compressed buf */
3326 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3328 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3329 arc_hdr_free_pabd(hdr);
3332 * We must setup a new shared block between the
3333 * last buffer and the hdr. The data would have
3334 * been allocated by the arc buf so we need to transfer
3335 * ownership to the hdr since it's now being shared.
3337 arc_share_buf(hdr, lastbuf);
3339 } else if (HDR_SHARED_DATA(hdr)) {
3341 * Uncompressed shared buffers are always at the end
3342 * of the list. Compressed buffers don't have the
3343 * same requirements. This makes it hard to
3344 * simply assert that the lastbuf is shared so
3345 * we rely on the hdr's compression flags to determine
3346 * if we have a compressed, shared buffer.
3348 ASSERT3P(lastbuf, !=, NULL);
3349 ASSERT(arc_buf_is_shared(lastbuf) ||
3350 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
3354 * Free the checksum if we're removing the last uncompressed buf from
3357 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3358 arc_cksum_free(hdr);
3361 /* clean up the buf */
3363 kmem_cache_free(buf_cache, buf);
3367 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr, boolean_t do_adapt)
3369 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3370 ASSERT(HDR_HAS_L1HDR(hdr));
3371 ASSERT(!HDR_SHARED_DATA(hdr));
3373 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3374 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr, do_adapt);
3375 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3376 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3378 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3379 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3383 arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
3385 ASSERT(HDR_HAS_L1HDR(hdr));
3386 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3389 * If the hdr is currently being written to the l2arc then
3390 * we defer freeing the data by adding it to the l2arc_free_on_write
3391 * list. The l2arc will free the data once it's finished
3392 * writing it to the l2arc device.
3394 if (HDR_L2_WRITING(hdr)) {
3395 arc_hdr_free_on_write(hdr);
3396 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3398 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
3399 arc_hdr_size(hdr), hdr);
3401 hdr->b_l1hdr.b_pabd = NULL;
3402 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3404 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3405 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3408 static arc_buf_hdr_t *
3409 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3410 enum zio_compress compression_type, arc_buf_contents_t type)
3414 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3416 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3417 ASSERT(HDR_EMPTY(hdr));
3418 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3419 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
3420 HDR_SET_PSIZE(hdr, psize);
3421 HDR_SET_LSIZE(hdr, lsize);
3425 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3426 arc_hdr_set_compress(hdr, compression_type);
3428 hdr->b_l1hdr.b_state = arc_anon;
3429 hdr->b_l1hdr.b_arc_access = 0;
3430 hdr->b_l1hdr.b_bufcnt = 0;
3431 hdr->b_l1hdr.b_buf = NULL;
3434 * Allocate the hdr's buffer. This will contain either
3435 * the compressed or uncompressed data depending on the block
3436 * it references and compressed arc enablement.
3438 arc_hdr_alloc_pabd(hdr, B_TRUE);
3439 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3445 * Transition between the two allocation states for the arc_buf_hdr struct.
3446 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3447 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3448 * version is used when a cache buffer is only in the L2ARC in order to reduce
3451 static arc_buf_hdr_t *
3452 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3454 ASSERT(HDR_HAS_L2HDR(hdr));
3456 arc_buf_hdr_t *nhdr;
3457 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3459 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3460 (old == hdr_l2only_cache && new == hdr_full_cache));
3462 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3464 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3465 buf_hash_remove(hdr);
3467 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3469 if (new == hdr_full_cache) {
3470 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3472 * arc_access and arc_change_state need to be aware that a
3473 * header has just come out of L2ARC, so we set its state to
3474 * l2c_only even though it's about to change.
3476 nhdr->b_l1hdr.b_state = arc_l2c_only;
3478 /* Verify previous threads set to NULL before freeing */
3479 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3481 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3482 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3483 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3486 * If we've reached here, We must have been called from
3487 * arc_evict_hdr(), as such we should have already been
3488 * removed from any ghost list we were previously on
3489 * (which protects us from racing with arc_evict_state),
3490 * thus no locking is needed during this check.
3492 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3495 * A buffer must not be moved into the arc_l2c_only
3496 * state if it's not finished being written out to the
3497 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3498 * might try to be accessed, even though it was removed.
3500 VERIFY(!HDR_L2_WRITING(hdr));
3501 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3504 if (hdr->b_l1hdr.b_thawed != NULL) {
3505 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3506 hdr->b_l1hdr.b_thawed = NULL;
3510 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3513 * The header has been reallocated so we need to re-insert it into any
3516 (void) buf_hash_insert(nhdr, NULL);
3518 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3520 mutex_enter(&dev->l2ad_mtx);
3523 * We must place the realloc'ed header back into the list at
3524 * the same spot. Otherwise, if it's placed earlier in the list,
3525 * l2arc_write_buffers() could find it during the function's
3526 * write phase, and try to write it out to the l2arc.
3528 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3529 list_remove(&dev->l2ad_buflist, hdr);
3531 mutex_exit(&dev->l2ad_mtx);
3534 * Since we're using the pointer address as the tag when
3535 * incrementing and decrementing the l2ad_alloc refcount, we
3536 * must remove the old pointer (that we're about to destroy) and
3537 * add the new pointer to the refcount. Otherwise we'd remove
3538 * the wrong pointer address when calling arc_hdr_destroy() later.
3541 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
3542 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
3544 buf_discard_identity(hdr);
3545 kmem_cache_free(old, hdr);
3551 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3552 * The buf is returned thawed since we expect the consumer to modify it.
3555 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3557 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3558 ZIO_COMPRESS_OFF, type);
3559 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3561 arc_buf_t *buf = NULL;
3562 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf));
3569 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3570 * for bufs containing metadata.
3573 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3574 enum zio_compress compression_type)
3576 ASSERT3U(lsize, >, 0);
3577 ASSERT3U(lsize, >=, psize);
3578 ASSERT(compression_type > ZIO_COMPRESS_OFF);
3579 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3581 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3582 compression_type, ARC_BUFC_DATA);
3583 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3585 arc_buf_t *buf = NULL;
3586 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3588 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3590 if (!arc_buf_is_shared(buf)) {
3592 * To ensure that the hdr has the correct data in it if we call
3593 * arc_decompress() on this buf before it's been written to
3594 * disk, it's easiest if we just set up sharing between the
3597 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3598 arc_hdr_free_pabd(hdr);
3599 arc_share_buf(hdr, buf);
3606 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3608 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3609 l2arc_dev_t *dev = l2hdr->b_dev;
3610 uint64_t psize = arc_hdr_size(hdr);
3612 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3613 ASSERT(HDR_HAS_L2HDR(hdr));
3615 list_remove(&dev->l2ad_buflist, hdr);
3617 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3618 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3620 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3622 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3623 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3627 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3629 if (HDR_HAS_L1HDR(hdr)) {
3630 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3631 hdr->b_l1hdr.b_bufcnt > 0);
3632 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3633 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3635 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3636 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3638 if (!HDR_EMPTY(hdr))
3639 buf_discard_identity(hdr);
3641 if (HDR_HAS_L2HDR(hdr)) {
3642 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3643 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3646 mutex_enter(&dev->l2ad_mtx);
3649 * Even though we checked this conditional above, we
3650 * need to check this again now that we have the
3651 * l2ad_mtx. This is because we could be racing with
3652 * another thread calling l2arc_evict() which might have
3653 * destroyed this header's L2 portion as we were waiting
3654 * to acquire the l2ad_mtx. If that happens, we don't
3655 * want to re-destroy the header's L2 portion.
3657 if (HDR_HAS_L2HDR(hdr)) {
3659 arc_hdr_l2hdr_destroy(hdr);
3663 mutex_exit(&dev->l2ad_mtx);
3666 if (HDR_HAS_L1HDR(hdr)) {
3667 arc_cksum_free(hdr);
3669 while (hdr->b_l1hdr.b_buf != NULL)
3670 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3673 if (hdr->b_l1hdr.b_thawed != NULL) {
3674 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3675 hdr->b_l1hdr.b_thawed = NULL;
3679 if (hdr->b_l1hdr.b_pabd != NULL) {
3680 arc_hdr_free_pabd(hdr);
3684 ASSERT3P(hdr->b_hash_next, ==, NULL);
3685 if (HDR_HAS_L1HDR(hdr)) {
3686 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3687 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3688 kmem_cache_free(hdr_full_cache, hdr);
3690 kmem_cache_free(hdr_l2only_cache, hdr);
3695 arc_buf_destroy(arc_buf_t *buf, void* tag)
3697 arc_buf_hdr_t *hdr = buf->b_hdr;
3698 kmutex_t *hash_lock = HDR_LOCK(hdr);
3700 if (hdr->b_l1hdr.b_state == arc_anon) {
3701 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3702 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3703 VERIFY0(remove_reference(hdr, NULL, tag));
3704 arc_hdr_destroy(hdr);
3708 mutex_enter(hash_lock);
3709 ASSERT3P(hdr, ==, buf->b_hdr);
3710 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3711 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3712 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3713 ASSERT3P(buf->b_data, !=, NULL);
3715 (void) remove_reference(hdr, hash_lock, tag);
3716 arc_buf_destroy_impl(buf);
3717 mutex_exit(hash_lock);
3721 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3722 * state of the header is dependent on its state prior to entering this
3723 * function. The following transitions are possible:
3725 * - arc_mru -> arc_mru_ghost
3726 * - arc_mfu -> arc_mfu_ghost
3727 * - arc_mru_ghost -> arc_l2c_only
3728 * - arc_mru_ghost -> deleted
3729 * - arc_mfu_ghost -> arc_l2c_only
3730 * - arc_mfu_ghost -> deleted
3733 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3735 arc_state_t *evicted_state, *state;
3736 int64_t bytes_evicted = 0;
3737 int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3738 zfs_arc_min_prescient_prefetch_ms : zfs_arc_min_prefetch_ms;
3740 ASSERT(MUTEX_HELD(hash_lock));
3741 ASSERT(HDR_HAS_L1HDR(hdr));
3743 state = hdr->b_l1hdr.b_state;
3744 if (GHOST_STATE(state)) {
3745 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3746 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3749 * l2arc_write_buffers() relies on a header's L1 portion
3750 * (i.e. its b_pabd field) during it's write phase.
3751 * Thus, we cannot push a header onto the arc_l2c_only
3752 * state (removing it's L1 piece) until the header is
3753 * done being written to the l2arc.
3755 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3756 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3757 return (bytes_evicted);
3760 ARCSTAT_BUMP(arcstat_deleted);
3761 bytes_evicted += HDR_GET_LSIZE(hdr);
3763 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3765 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3766 if (HDR_HAS_L2HDR(hdr)) {
3768 * This buffer is cached on the 2nd Level ARC;
3769 * don't destroy the header.
3771 arc_change_state(arc_l2c_only, hdr, hash_lock);
3773 * dropping from L1+L2 cached to L2-only,
3774 * realloc to remove the L1 header.
3776 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3779 arc_change_state(arc_anon, hdr, hash_lock);
3780 arc_hdr_destroy(hdr);
3782 return (bytes_evicted);
3785 ASSERT(state == arc_mru || state == arc_mfu);
3786 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3788 /* prefetch buffers have a minimum lifespan */
3789 if (HDR_IO_IN_PROGRESS(hdr) ||
3790 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3791 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < min_lifetime * hz)) {
3792 ARCSTAT_BUMP(arcstat_evict_skip);
3793 return (bytes_evicted);
3796 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3797 while (hdr->b_l1hdr.b_buf) {
3798 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3799 if (!mutex_tryenter(&buf->b_evict_lock)) {
3800 ARCSTAT_BUMP(arcstat_mutex_miss);
3803 if (buf->b_data != NULL)
3804 bytes_evicted += HDR_GET_LSIZE(hdr);
3805 mutex_exit(&buf->b_evict_lock);
3806 arc_buf_destroy_impl(buf);
3809 if (HDR_HAS_L2HDR(hdr)) {
3810 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3812 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3813 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3814 HDR_GET_LSIZE(hdr));
3816 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3817 HDR_GET_LSIZE(hdr));
3821 if (hdr->b_l1hdr.b_bufcnt == 0) {
3822 arc_cksum_free(hdr);
3824 bytes_evicted += arc_hdr_size(hdr);
3827 * If this hdr is being evicted and has a compressed
3828 * buffer then we discard it here before we change states.
3829 * This ensures that the accounting is updated correctly
3830 * in arc_free_data_impl().
3832 arc_hdr_free_pabd(hdr);
3834 arc_change_state(evicted_state, hdr, hash_lock);
3835 ASSERT(HDR_IN_HASH_TABLE(hdr));
3836 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3837 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3840 return (bytes_evicted);
3844 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3845 uint64_t spa, int64_t bytes)
3847 multilist_sublist_t *mls;
3848 uint64_t bytes_evicted = 0;
3850 kmutex_t *hash_lock;
3851 int evict_count = 0;
3853 ASSERT3P(marker, !=, NULL);
3854 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3856 mls = multilist_sublist_lock(ml, idx);
3858 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3859 hdr = multilist_sublist_prev(mls, marker)) {
3860 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3861 (evict_count >= zfs_arc_evict_batch_limit))
3865 * To keep our iteration location, move the marker
3866 * forward. Since we're not holding hdr's hash lock, we
3867 * must be very careful and not remove 'hdr' from the
3868 * sublist. Otherwise, other consumers might mistake the
3869 * 'hdr' as not being on a sublist when they call the
3870 * multilist_link_active() function (they all rely on
3871 * the hash lock protecting concurrent insertions and
3872 * removals). multilist_sublist_move_forward() was
3873 * specifically implemented to ensure this is the case
3874 * (only 'marker' will be removed and re-inserted).
3876 multilist_sublist_move_forward(mls, marker);
3879 * The only case where the b_spa field should ever be
3880 * zero, is the marker headers inserted by
3881 * arc_evict_state(). It's possible for multiple threads
3882 * to be calling arc_evict_state() concurrently (e.g.
3883 * dsl_pool_close() and zio_inject_fault()), so we must
3884 * skip any markers we see from these other threads.
3886 if (hdr->b_spa == 0)
3889 /* we're only interested in evicting buffers of a certain spa */
3890 if (spa != 0 && hdr->b_spa != spa) {
3891 ARCSTAT_BUMP(arcstat_evict_skip);
3895 hash_lock = HDR_LOCK(hdr);
3898 * We aren't calling this function from any code path
3899 * that would already be holding a hash lock, so we're
3900 * asserting on this assumption to be defensive in case
3901 * this ever changes. Without this check, it would be
3902 * possible to incorrectly increment arcstat_mutex_miss
3903 * below (e.g. if the code changed such that we called
3904 * this function with a hash lock held).
3906 ASSERT(!MUTEX_HELD(hash_lock));
3908 if (mutex_tryenter(hash_lock)) {
3909 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3910 mutex_exit(hash_lock);
3912 bytes_evicted += evicted;
3915 * If evicted is zero, arc_evict_hdr() must have
3916 * decided to skip this header, don't increment
3917 * evict_count in this case.
3923 * If arc_size isn't overflowing, signal any
3924 * threads that might happen to be waiting.
3926 * For each header evicted, we wake up a single
3927 * thread. If we used cv_broadcast, we could
3928 * wake up "too many" threads causing arc_size
3929 * to significantly overflow arc_c; since
3930 * arc_get_data_impl() doesn't check for overflow
3931 * when it's woken up (it doesn't because it's
3932 * possible for the ARC to be overflowing while
3933 * full of un-evictable buffers, and the
3934 * function should proceed in this case).
3936 * If threads are left sleeping, due to not
3937 * using cv_broadcast here, they will be woken
3938 * up via cv_broadcast in arc_adjust_cb() just
3939 * before arc_adjust_zthr sleeps.
3941 mutex_enter(&arc_adjust_lock);
3942 if (!arc_is_overflowing())
3943 cv_signal(&arc_adjust_waiters_cv);
3944 mutex_exit(&arc_adjust_lock);
3946 ARCSTAT_BUMP(arcstat_mutex_miss);
3950 multilist_sublist_unlock(mls);
3952 return (bytes_evicted);
3956 * Evict buffers from the given arc state, until we've removed the
3957 * specified number of bytes. Move the removed buffers to the
3958 * appropriate evict state.
3960 * This function makes a "best effort". It skips over any buffers
3961 * it can't get a hash_lock on, and so, may not catch all candidates.
3962 * It may also return without evicting as much space as requested.
3964 * If bytes is specified using the special value ARC_EVICT_ALL, this
3965 * will evict all available (i.e. unlocked and evictable) buffers from
3966 * the given arc state; which is used by arc_flush().
3969 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3970 arc_buf_contents_t type)
3972 uint64_t total_evicted = 0;
3973 multilist_t *ml = state->arcs_list[type];
3975 arc_buf_hdr_t **markers;
3977 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3979 num_sublists = multilist_get_num_sublists(ml);
3982 * If we've tried to evict from each sublist, made some
3983 * progress, but still have not hit the target number of bytes
3984 * to evict, we want to keep trying. The markers allow us to
3985 * pick up where we left off for each individual sublist, rather
3986 * than starting from the tail each time.
3988 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3989 for (int i = 0; i < num_sublists; i++) {
3990 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3993 * A b_spa of 0 is used to indicate that this header is
3994 * a marker. This fact is used in arc_adjust_type() and
3995 * arc_evict_state_impl().
3997 markers[i]->b_spa = 0;
3999 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4000 multilist_sublist_insert_tail(mls, markers[i]);
4001 multilist_sublist_unlock(mls);
4005 * While we haven't hit our target number of bytes to evict, or
4006 * we're evicting all available buffers.
4008 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
4009 int sublist_idx = multilist_get_random_index(ml);
4010 uint64_t scan_evicted = 0;
4013 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
4014 * Request that 10% of the LRUs be scanned by the superblock
4017 if (type == ARC_BUFC_DATA && aggsum_compare(&astat_dnode_size,
4018 arc_dnode_limit) > 0) {
4019 arc_prune_async((aggsum_upper_bound(&astat_dnode_size) -
4020 arc_dnode_limit) / sizeof (dnode_t) /
4021 zfs_arc_dnode_reduce_percent);
4025 * Start eviction using a randomly selected sublist,
4026 * this is to try and evenly balance eviction across all
4027 * sublists. Always starting at the same sublist
4028 * (e.g. index 0) would cause evictions to favor certain
4029 * sublists over others.
4031 for (int i = 0; i < num_sublists; i++) {
4032 uint64_t bytes_remaining;
4033 uint64_t bytes_evicted;
4035 if (bytes == ARC_EVICT_ALL)
4036 bytes_remaining = ARC_EVICT_ALL;
4037 else if (total_evicted < bytes)
4038 bytes_remaining = bytes - total_evicted;
4042 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
4043 markers[sublist_idx], spa, bytes_remaining);
4045 scan_evicted += bytes_evicted;
4046 total_evicted += bytes_evicted;
4048 /* we've reached the end, wrap to the beginning */
4049 if (++sublist_idx >= num_sublists)
4054 * If we didn't evict anything during this scan, we have
4055 * no reason to believe we'll evict more during another
4056 * scan, so break the loop.
4058 if (scan_evicted == 0) {
4059 /* This isn't possible, let's make that obvious */
4060 ASSERT3S(bytes, !=, 0);
4063 * When bytes is ARC_EVICT_ALL, the only way to
4064 * break the loop is when scan_evicted is zero.
4065 * In that case, we actually have evicted enough,
4066 * so we don't want to increment the kstat.
4068 if (bytes != ARC_EVICT_ALL) {
4069 ASSERT3S(total_evicted, <, bytes);
4070 ARCSTAT_BUMP(arcstat_evict_not_enough);
4077 for (int i = 0; i < num_sublists; i++) {
4078 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4079 multilist_sublist_remove(mls, markers[i]);
4080 multilist_sublist_unlock(mls);
4082 kmem_cache_free(hdr_full_cache, markers[i]);
4084 kmem_free(markers, sizeof (*markers) * num_sublists);
4086 return (total_evicted);
4090 * Flush all "evictable" data of the given type from the arc state
4091 * specified. This will not evict any "active" buffers (i.e. referenced).
4093 * When 'retry' is set to B_FALSE, the function will make a single pass
4094 * over the state and evict any buffers that it can. Since it doesn't
4095 * continually retry the eviction, it might end up leaving some buffers
4096 * in the ARC due to lock misses.
4098 * When 'retry' is set to B_TRUE, the function will continually retry the
4099 * eviction until *all* evictable buffers have been removed from the
4100 * state. As a result, if concurrent insertions into the state are
4101 * allowed (e.g. if the ARC isn't shutting down), this function might
4102 * wind up in an infinite loop, continually trying to evict buffers.
4105 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4108 uint64_t evicted = 0;
4110 while (refcount_count(&state->arcs_esize[type]) != 0) {
4111 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
4121 * Helper function for arc_prune_async() it is responsible for safely
4122 * handling the execution of a registered arc_prune_func_t.
4125 arc_prune_task(void *ptr)
4127 arc_prune_t *ap = (arc_prune_t *)ptr;
4128 arc_prune_func_t *func = ap->p_pfunc;
4131 func(ap->p_adjust, ap->p_private);
4133 refcount_remove(&ap->p_refcnt, func);
4137 * Notify registered consumers they must drop holds on a portion of the ARC
4138 * buffered they reference. This provides a mechanism to ensure the ARC can
4139 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
4140 * is analogous to dnlc_reduce_cache() but more generic.
4142 * This operation is performed asynchronously so it may be safely called
4143 * in the context of the arc_reclaim_thread(). A reference is taken here
4144 * for each registered arc_prune_t and the arc_prune_task() is responsible
4145 * for releasing it once the registered arc_prune_func_t has completed.
4148 arc_prune_async(int64_t adjust)
4152 mutex_enter(&arc_prune_mtx);
4153 for (ap = list_head(&arc_prune_list); ap != NULL;
4154 ap = list_next(&arc_prune_list, ap)) {
4156 if (refcount_count(&ap->p_refcnt) >= 2)
4159 refcount_add(&ap->p_refcnt, ap->p_pfunc);
4160 ap->p_adjust = adjust;
4161 if (taskq_dispatch(arc_prune_taskq, arc_prune_task,
4162 ap, TQ_SLEEP) == TASKQID_INVALID) {
4163 refcount_remove(&ap->p_refcnt, ap->p_pfunc);
4166 ARCSTAT_BUMP(arcstat_prune);
4168 mutex_exit(&arc_prune_mtx);
4172 * Evict the specified number of bytes from the state specified,
4173 * restricting eviction to the spa and type given. This function
4174 * prevents us from trying to evict more from a state's list than
4175 * is "evictable", and to skip evicting altogether when passed a
4176 * negative value for "bytes". In contrast, arc_evict_state() will
4177 * evict everything it can, when passed a negative value for "bytes".
4180 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
4181 arc_buf_contents_t type)
4185 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
4186 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
4187 return (arc_evict_state(state, spa, delta, type));
4194 * The goal of this function is to evict enough meta data buffers from the
4195 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
4196 * more complicated than it appears because it is common for data buffers
4197 * to have holds on meta data buffers. In addition, dnode meta data buffers
4198 * will be held by the dnodes in the block preventing them from being freed.
4199 * This means we can't simply traverse the ARC and expect to always find
4200 * enough unheld meta data buffer to release.
4202 * Therefore, this function has been updated to make alternating passes
4203 * over the ARC releasing data buffers and then newly unheld meta data
4204 * buffers. This ensures forward progress is maintained and meta_used
4205 * will decrease. Normally this is sufficient, but if required the ARC
4206 * will call the registered prune callbacks causing dentry and inodes to
4207 * be dropped from the VFS cache. This will make dnode meta data buffers
4208 * available for reclaim.
4211 arc_adjust_meta_balanced(uint64_t meta_used)
4213 int64_t delta, prune = 0, adjustmnt;
4214 uint64_t total_evicted = 0;
4215 arc_buf_contents_t type = ARC_BUFC_DATA;
4216 int restarts = MAX(zfs_arc_meta_adjust_restarts, 0);
4220 * This slightly differs than the way we evict from the mru in
4221 * arc_adjust because we don't have a "target" value (i.e. no
4222 * "meta" arc_p). As a result, I think we can completely
4223 * cannibalize the metadata in the MRU before we evict the
4224 * metadata from the MFU. I think we probably need to implement a
4225 * "metadata arc_p" value to do this properly.
4227 adjustmnt = meta_used - arc_meta_limit;
4229 if (adjustmnt > 0 && refcount_count(&arc_mru->arcs_esize[type]) > 0) {
4230 delta = MIN(refcount_count(&arc_mru->arcs_esize[type]),
4232 total_evicted += arc_adjust_impl(arc_mru, 0, delta, type);
4237 * We can't afford to recalculate adjustmnt here. If we do,
4238 * new metadata buffers can sneak into the MRU or ANON lists,
4239 * thus penalize the MFU metadata. Although the fudge factor is
4240 * small, it has been empirically shown to be significant for
4241 * certain workloads (e.g. creating many empty directories). As
4242 * such, we use the original calculation for adjustmnt, and
4243 * simply decrement the amount of data evicted from the MRU.
4246 if (adjustmnt > 0 && refcount_count(&arc_mfu->arcs_esize[type]) > 0) {
4247 delta = MIN(refcount_count(&arc_mfu->arcs_esize[type]),
4249 total_evicted += arc_adjust_impl(arc_mfu, 0, delta, type);
4252 adjustmnt = meta_used - arc_meta_limit;
4254 if (adjustmnt > 0 &&
4255 refcount_count(&arc_mru_ghost->arcs_esize[type]) > 0) {
4256 delta = MIN(adjustmnt,
4257 refcount_count(&arc_mru_ghost->arcs_esize[type]));
4258 total_evicted += arc_adjust_impl(arc_mru_ghost, 0, delta, type);
4262 if (adjustmnt > 0 &&
4263 refcount_count(&arc_mfu_ghost->arcs_esize[type]) > 0) {
4264 delta = MIN(adjustmnt,
4265 refcount_count(&arc_mfu_ghost->arcs_esize[type]));
4266 total_evicted += arc_adjust_impl(arc_mfu_ghost, 0, delta, type);
4270 * If after attempting to make the requested adjustment to the ARC
4271 * the meta limit is still being exceeded then request that the
4272 * higher layers drop some cached objects which have holds on ARC
4273 * meta buffers. Requests to the upper layers will be made with
4274 * increasingly large scan sizes until the ARC is below the limit.
4276 if (meta_used > arc_meta_limit) {
4277 if (type == ARC_BUFC_DATA) {
4278 type = ARC_BUFC_METADATA;
4280 type = ARC_BUFC_DATA;
4282 if (zfs_arc_meta_prune) {
4283 prune += zfs_arc_meta_prune;
4284 arc_prune_async(prune);
4293 return (total_evicted);
4297 * Evict metadata buffers from the cache, such that arc_meta_used is
4298 * capped by the arc_meta_limit tunable.
4301 arc_adjust_meta_only(uint64_t meta_used)
4303 uint64_t total_evicted = 0;
4307 * If we're over the meta limit, we want to evict enough
4308 * metadata to get back under the meta limit. We don't want to
4309 * evict so much that we drop the MRU below arc_p, though. If
4310 * we're over the meta limit more than we're over arc_p, we
4311 * evict some from the MRU here, and some from the MFU below.
4313 target = MIN((int64_t)(meta_used - arc_meta_limit),
4314 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4315 refcount_count(&arc_mru->arcs_size) - arc_p));
4317 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4320 * Similar to the above, we want to evict enough bytes to get us
4321 * below the meta limit, but not so much as to drop us below the
4322 * space allotted to the MFU (which is defined as arc_c - arc_p).
4324 target = MIN((int64_t)(meta_used - arc_meta_limit),
4325 (int64_t)(refcount_count(&arc_mfu->arcs_size) -
4328 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4330 return (total_evicted);
4334 arc_adjust_meta(uint64_t meta_used)
4336 if (zfs_arc_meta_strategy == ARC_STRATEGY_META_ONLY)
4337 return (arc_adjust_meta_only(meta_used));
4339 return (arc_adjust_meta_balanced(meta_used));
4343 * Return the type of the oldest buffer in the given arc state
4345 * This function will select a random sublist of type ARC_BUFC_DATA and
4346 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4347 * is compared, and the type which contains the "older" buffer will be
4350 static arc_buf_contents_t
4351 arc_adjust_type(arc_state_t *state)
4353 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
4354 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
4355 int data_idx = multilist_get_random_index(data_ml);
4356 int meta_idx = multilist_get_random_index(meta_ml);
4357 multilist_sublist_t *data_mls;
4358 multilist_sublist_t *meta_mls;
4359 arc_buf_contents_t type;
4360 arc_buf_hdr_t *data_hdr;
4361 arc_buf_hdr_t *meta_hdr;
4364 * We keep the sublist lock until we're finished, to prevent
4365 * the headers from being destroyed via arc_evict_state().
4367 data_mls = multilist_sublist_lock(data_ml, data_idx);
4368 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4371 * These two loops are to ensure we skip any markers that
4372 * might be at the tail of the lists due to arc_evict_state().
4375 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4376 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4377 if (data_hdr->b_spa != 0)
4381 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4382 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4383 if (meta_hdr->b_spa != 0)
4387 if (data_hdr == NULL && meta_hdr == NULL) {
4388 type = ARC_BUFC_DATA;
4389 } else if (data_hdr == NULL) {
4390 ASSERT3P(meta_hdr, !=, NULL);
4391 type = ARC_BUFC_METADATA;
4392 } else if (meta_hdr == NULL) {
4393 ASSERT3P(data_hdr, !=, NULL);
4394 type = ARC_BUFC_DATA;
4396 ASSERT3P(data_hdr, !=, NULL);
4397 ASSERT3P(meta_hdr, !=, NULL);
4399 /* The headers can't be on the sublist without an L1 header */
4400 ASSERT(HDR_HAS_L1HDR(data_hdr));
4401 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4403 if (data_hdr->b_l1hdr.b_arc_access <
4404 meta_hdr->b_l1hdr.b_arc_access) {
4405 type = ARC_BUFC_DATA;
4407 type = ARC_BUFC_METADATA;
4411 multilist_sublist_unlock(meta_mls);
4412 multilist_sublist_unlock(data_mls);
4418 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4423 uint64_t total_evicted = 0;
4426 uint64_t asize = aggsum_value(&arc_size);
4427 uint64_t ameta = aggsum_value(&arc_meta_used);
4430 * If we're over arc_meta_limit, we want to correct that before
4431 * potentially evicting data buffers below.
4433 total_evicted += arc_adjust_meta(ameta);
4438 * If we're over the target cache size, we want to evict enough
4439 * from the list to get back to our target size. We don't want
4440 * to evict too much from the MRU, such that it drops below
4441 * arc_p. So, if we're over our target cache size more than
4442 * the MRU is over arc_p, we'll evict enough to get back to
4443 * arc_p here, and then evict more from the MFU below.
4445 target = MIN((int64_t)(asize - arc_c),
4446 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4447 refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4450 * If we're below arc_meta_min, always prefer to evict data.
4451 * Otherwise, try to satisfy the requested number of bytes to
4452 * evict from the type which contains older buffers; in an
4453 * effort to keep newer buffers in the cache regardless of their
4454 * type. If we cannot satisfy the number of bytes from this
4455 * type, spill over into the next type.
4457 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4458 ameta > arc_meta_min) {
4459 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4460 total_evicted += bytes;
4463 * If we couldn't evict our target number of bytes from
4464 * metadata, we try to get the rest from data.
4469 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4471 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4472 total_evicted += bytes;
4475 * If we couldn't evict our target number of bytes from
4476 * data, we try to get the rest from metadata.
4481 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4485 * Re-sum ARC stats after the first round of evictions.
4487 asize = aggsum_value(&arc_size);
4488 ameta = aggsum_value(&arc_meta_used);
4493 * Now that we've tried to evict enough from the MRU to get its
4494 * size back to arc_p, if we're still above the target cache
4495 * size, we evict the rest from the MFU.
4497 target = asize - arc_c;
4499 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4500 ameta > arc_meta_min) {
4501 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4502 total_evicted += bytes;
4505 * If we couldn't evict our target number of bytes from
4506 * metadata, we try to get the rest from data.
4511 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4513 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4514 total_evicted += bytes;
4517 * If we couldn't evict our target number of bytes from
4518 * data, we try to get the rest from data.
4523 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4527 * Adjust ghost lists
4529 * In addition to the above, the ARC also defines target values
4530 * for the ghost lists. The sum of the mru list and mru ghost
4531 * list should never exceed the target size of the cache, and
4532 * the sum of the mru list, mfu list, mru ghost list, and mfu
4533 * ghost list should never exceed twice the target size of the
4534 * cache. The following logic enforces these limits on the ghost
4535 * caches, and evicts from them as needed.
4537 target = refcount_count(&arc_mru->arcs_size) +
4538 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4540 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4541 total_evicted += bytes;
4546 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4549 * We assume the sum of the mru list and mfu list is less than
4550 * or equal to arc_c (we enforced this above), which means we
4551 * can use the simpler of the two equations below:
4553 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4554 * mru ghost + mfu ghost <= arc_c
4556 target = refcount_count(&arc_mru_ghost->arcs_size) +
4557 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4559 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4560 total_evicted += bytes;
4565 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4567 return (total_evicted);
4571 arc_flush(spa_t *spa, boolean_t retry)
4576 * If retry is B_TRUE, a spa must not be specified since we have
4577 * no good way to determine if all of a spa's buffers have been
4578 * evicted from an arc state.
4580 ASSERT(!retry || spa == 0);
4583 guid = spa_load_guid(spa);
4585 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4586 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4588 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4589 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4591 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4592 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4594 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4595 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4599 arc_reduce_target_size(int64_t to_free)
4601 uint64_t asize = aggsum_value(&arc_size);
4602 if (arc_c > arc_c_min) {
4603 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
4604 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
4605 if (arc_c > arc_c_min + to_free)
4606 atomic_add_64(&arc_c, -to_free);
4610 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4612 arc_c = MAX(asize, arc_c_min);
4614 arc_p = (arc_c >> 1);
4616 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
4619 ASSERT(arc_c >= arc_c_min);
4620 ASSERT((int64_t)arc_p >= 0);
4623 if (asize > arc_c) {
4624 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, asize,
4626 /* See comment in arc_adjust_cb_check() on why lock+flag */
4627 mutex_enter(&arc_adjust_lock);
4628 arc_adjust_needed = B_TRUE;
4629 mutex_exit(&arc_adjust_lock);
4630 zthr_wakeup(arc_adjust_zthr);
4634 typedef enum free_memory_reason_t {
4639 FMR_PAGES_PP_MAXIMUM,
4642 } free_memory_reason_t;
4644 int64_t last_free_memory;
4645 free_memory_reason_t last_free_reason;
4648 * Additional reserve of pages for pp_reserve.
4650 int64_t arc_pages_pp_reserve = 64;
4653 * Additional reserve of pages for swapfs.
4655 int64_t arc_swapfs_reserve = 64;
4658 * Return the amount of memory that can be consumed before reclaim will be
4659 * needed. Positive if there is sufficient free memory, negative indicates
4660 * the amount of memory that needs to be freed up.
4663 arc_available_memory(void)
4665 int64_t lowest = INT64_MAX;
4667 free_memory_reason_t r = FMR_UNKNOWN;
4672 * Cooperate with pagedaemon when it's time for it to scan
4673 * and reclaim some pages.
4675 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
4683 n = PAGESIZE * (-needfree);
4691 * check that we're out of range of the pageout scanner. It starts to
4692 * schedule paging if freemem is less than lotsfree and needfree.
4693 * lotsfree is the high-water mark for pageout, and needfree is the
4694 * number of needed free pages. We add extra pages here to make sure
4695 * the scanner doesn't start up while we're freeing memory.
4697 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4704 * check to make sure that swapfs has enough space so that anon
4705 * reservations can still succeed. anon_resvmem() checks that the
4706 * availrmem is greater than swapfs_minfree, and the number of reserved
4707 * swap pages. We also add a bit of extra here just to prevent
4708 * circumstances from getting really dire.
4710 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4711 desfree - arc_swapfs_reserve);
4714 r = FMR_SWAPFS_MINFREE;
4719 * Check that we have enough availrmem that memory locking (e.g., via
4720 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4721 * stores the number of pages that cannot be locked; when availrmem
4722 * drops below pages_pp_maximum, page locking mechanisms such as
4723 * page_pp_lock() will fail.)
4725 n = PAGESIZE * (availrmem - pages_pp_maximum -
4726 arc_pages_pp_reserve);
4729 r = FMR_PAGES_PP_MAXIMUM;
4732 #endif /* __FreeBSD__ */
4733 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
4735 * If we're on an i386 platform, it's possible that we'll exhaust the
4736 * kernel heap space before we ever run out of available physical
4737 * memory. Most checks of the size of the heap_area compare against
4738 * tune.t_minarmem, which is the minimum available real memory that we
4739 * can have in the system. However, this is generally fixed at 25 pages
4740 * which is so low that it's useless. In this comparison, we seek to
4741 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4742 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4745 n = uma_avail() - (long)(uma_limit() / 4);
4753 * If zio data pages are being allocated out of a separate heap segment,
4754 * then enforce that the size of available vmem for this arena remains
4755 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4757 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4758 * memory (in the zio_arena) free, which can avoid memory
4759 * fragmentation issues.
4761 if (zio_arena != NULL) {
4762 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4763 (vmem_size(zio_arena, VMEM_ALLOC) >>
4764 arc_zio_arena_free_shift);
4772 /* Every 100 calls, free a small amount */
4773 if (spa_get_random(100) == 0)
4775 #endif /* _KERNEL */
4777 last_free_memory = lowest;
4778 last_free_reason = r;
4779 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
4785 * Determine if the system is under memory pressure and is asking
4786 * to reclaim memory. A return value of B_TRUE indicates that the system
4787 * is under memory pressure and that the arc should adjust accordingly.
4790 arc_reclaim_needed(void)
4792 return (arc_available_memory() < 0);
4795 extern kmem_cache_t *zio_buf_cache[];
4796 extern kmem_cache_t *zio_data_buf_cache[];
4797 extern kmem_cache_t *range_seg_cache;
4798 extern kmem_cache_t *abd_chunk_cache;
4800 static __noinline void
4801 arc_kmem_reap_soon(void)
4804 kmem_cache_t *prev_cache = NULL;
4805 kmem_cache_t *prev_data_cache = NULL;
4807 DTRACE_PROBE(arc__kmem_reap_start);
4809 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) {
4811 * We are exceeding our meta-data cache limit.
4812 * Purge some DNLC entries to release holds on meta-data.
4814 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4818 * Reclaim unused memory from all kmem caches.
4824 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4825 if (zio_buf_cache[i] != prev_cache) {
4826 prev_cache = zio_buf_cache[i];
4827 kmem_cache_reap_soon(zio_buf_cache[i]);
4829 if (zio_data_buf_cache[i] != prev_data_cache) {
4830 prev_data_cache = zio_data_buf_cache[i];
4831 kmem_cache_reap_soon(zio_data_buf_cache[i]);
4834 kmem_cache_reap_soon(abd_chunk_cache);
4835 kmem_cache_reap_soon(buf_cache);
4836 kmem_cache_reap_soon(hdr_full_cache);
4837 kmem_cache_reap_soon(hdr_l2only_cache);
4838 kmem_cache_reap_soon(range_seg_cache);
4841 if (zio_arena != NULL) {
4843 * Ask the vmem arena to reclaim unused memory from its
4846 vmem_qcache_reap(zio_arena);
4849 DTRACE_PROBE(arc__kmem_reap_end);
4854 arc_adjust_cb_check(void *arg, zthr_t *zthr)
4857 * This is necessary in order for the mdb ::arc dcmd to
4858 * show up to date information. Since the ::arc command
4859 * does not call the kstat's update function, without
4860 * this call, the command may show stale stats for the
4861 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4862 * with this change, the data might be up to 1 second
4863 * out of date(the arc_adjust_zthr has a maximum sleep
4864 * time of 1 second); but that should suffice. The
4865 * arc_state_t structures can be queried directly if more
4866 * accurate information is needed.
4868 if (arc_ksp != NULL)
4869 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4872 * We have to rely on arc_get_data_impl() to tell us when to adjust,
4873 * rather than checking if we are overflowing here, so that we are
4874 * sure to not leave arc_get_data_impl() waiting on
4875 * arc_adjust_waiters_cv. If we have become "not overflowing" since
4876 * arc_get_data_impl() checked, we need to wake it up. We could
4877 * broadcast the CV here, but arc_get_data_impl() may have not yet
4878 * gone to sleep. We would need to use a mutex to ensure that this
4879 * function doesn't broadcast until arc_get_data_impl() has gone to
4880 * sleep (e.g. the arc_adjust_lock). However, the lock ordering of
4881 * such a lock would necessarily be incorrect with respect to the
4882 * zthr_lock, which is held before this function is called, and is
4883 * held by arc_get_data_impl() when it calls zthr_wakeup().
4885 return (arc_adjust_needed);
4889 * Keep arc_size under arc_c by running arc_adjust which evicts data
4893 arc_adjust_cb(void *arg, zthr_t *zthr)
4895 uint64_t evicted = 0;
4897 /* Evict from cache */
4898 evicted = arc_adjust();
4901 * If evicted is zero, we couldn't evict anything
4902 * via arc_adjust(). This could be due to hash lock
4903 * collisions, but more likely due to the majority of
4904 * arc buffers being unevictable. Therefore, even if
4905 * arc_size is above arc_c, another pass is unlikely to
4906 * be helpful and could potentially cause us to enter an
4907 * infinite loop. Additionally, zthr_iscancelled() is
4908 * checked here so that if the arc is shutting down, the
4909 * broadcast will wake any remaining arc adjust waiters.
4911 mutex_enter(&arc_adjust_lock);
4912 arc_adjust_needed = !zthr_iscancelled(arc_adjust_zthr) &&
4913 evicted > 0 && aggsum_compare(&arc_size, arc_c) > 0;
4914 if (!arc_adjust_needed) {
4916 * We're either no longer overflowing, or we
4917 * can't evict anything more, so we should wake
4920 cv_broadcast(&arc_adjust_waiters_cv);
4922 mutex_exit(&arc_adjust_lock);
4929 arc_reap_cb_check(void *arg, zthr_t *zthr)
4931 int64_t free_memory = arc_available_memory();
4934 * If a kmem reap is already active, don't schedule more. We must
4935 * check for this because kmem_cache_reap_soon() won't actually
4936 * block on the cache being reaped (this is to prevent callers from
4937 * becoming implicitly blocked by a system-wide kmem reap -- which,
4938 * on a system with many, many full magazines, can take minutes).
4940 if (!kmem_cache_reap_active() &&
4942 arc_no_grow = B_TRUE;
4945 * Wait at least zfs_grow_retry (default 60) seconds
4946 * before considering growing.
4948 arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4950 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4951 arc_no_grow = B_TRUE;
4952 } else if (gethrtime() >= arc_growtime) {
4953 arc_no_grow = B_FALSE;
4960 * Keep enough free memory in the system by reaping the ARC's kmem
4961 * caches. To cause more slabs to be reapable, we may reduce the
4962 * target size of the cache (arc_c), causing the arc_adjust_cb()
4963 * to free more buffers.
4967 arc_reap_cb(void *arg, zthr_t *zthr)
4969 int64_t free_memory;
4972 * Kick off asynchronous kmem_reap()'s of all our caches.
4974 arc_kmem_reap_soon();
4977 * Wait at least arc_kmem_cache_reap_retry_ms between
4978 * arc_kmem_reap_soon() calls. Without this check it is possible to
4979 * end up in a situation where we spend lots of time reaping
4980 * caches, while we're near arc_c_min. Waiting here also gives the
4981 * subsequent free memory check a chance of finding that the
4982 * asynchronous reap has already freed enough memory, and we don't
4983 * need to call arc_reduce_target_size().
4985 delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000);
4988 * Reduce the target size as needed to maintain the amount of free
4989 * memory in the system at a fraction of the arc_size (1/128th by
4990 * default). If oversubscribed (free_memory < 0) then reduce the
4991 * target arc_size by the deficit amount plus the fractional
4992 * amount. If free memory is positive but less then the fractional
4993 * amount, reduce by what is needed to hit the fractional amount.
4995 free_memory = arc_available_memory();
4998 (arc_c >> arc_shrink_shift) - free_memory;
5002 to_free = MAX(to_free, ptob(needfree));
5005 arc_reduce_target_size(to_free);
5011 static u_int arc_dnlc_evicts_arg;
5012 extern struct vfsops zfs_vfsops;
5015 arc_dnlc_evicts_thread(void *dummy __unused)
5020 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG);
5022 mutex_enter(&arc_dnlc_evicts_lock);
5023 while (!arc_dnlc_evicts_thread_exit) {
5024 CALLB_CPR_SAFE_BEGIN(&cpr);
5025 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
5026 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock);
5027 if (arc_dnlc_evicts_arg != 0) {
5028 percent = arc_dnlc_evicts_arg;
5029 mutex_exit(&arc_dnlc_evicts_lock);
5031 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops);
5033 mutex_enter(&arc_dnlc_evicts_lock);
5035 * Clear our token only after vnlru_free()
5036 * pass is done, to avoid false queueing of
5039 arc_dnlc_evicts_arg = 0;
5042 arc_dnlc_evicts_thread_exit = FALSE;
5043 cv_broadcast(&arc_dnlc_evicts_cv);
5044 CALLB_CPR_EXIT(&cpr);
5049 dnlc_reduce_cache(void *arg)
5053 percent = (u_int)(uintptr_t)arg;
5054 mutex_enter(&arc_dnlc_evicts_lock);
5055 if (arc_dnlc_evicts_arg == 0) {
5056 arc_dnlc_evicts_arg = percent;
5057 cv_broadcast(&arc_dnlc_evicts_cv);
5059 mutex_exit(&arc_dnlc_evicts_lock);
5063 * Adapt arc info given the number of bytes we are trying to add and
5064 * the state that we are comming from. This function is only called
5065 * when we are adding new content to the cache.
5068 arc_adapt(int bytes, arc_state_t *state)
5071 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
5072 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
5073 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
5075 if (state == arc_l2c_only)
5080 * Adapt the target size of the MRU list:
5081 * - if we just hit in the MRU ghost list, then increase
5082 * the target size of the MRU list.
5083 * - if we just hit in the MFU ghost list, then increase
5084 * the target size of the MFU list by decreasing the
5085 * target size of the MRU list.
5087 if (state == arc_mru_ghost) {
5088 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
5089 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
5091 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
5092 } else if (state == arc_mfu_ghost) {
5095 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
5096 mult = MIN(mult, 10);
5098 delta = MIN(bytes * mult, arc_p);
5099 arc_p = MAX(arc_p_min, arc_p - delta);
5101 ASSERT((int64_t)arc_p >= 0);
5104 * Wake reap thread if we do not have any available memory
5106 if (arc_reclaim_needed()) {
5107 zthr_wakeup(arc_reap_zthr);
5114 if (arc_c >= arc_c_max)
5118 * If we're within (2 * maxblocksize) bytes of the target
5119 * cache size, increment the target cache size
5121 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >
5123 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
5124 atomic_add_64(&arc_c, (int64_t)bytes);
5125 if (arc_c > arc_c_max)
5127 else if (state == arc_anon)
5128 atomic_add_64(&arc_p, (int64_t)bytes);
5132 ASSERT((int64_t)arc_p >= 0);
5136 * Check if arc_size has grown past our upper threshold, determined by
5137 * zfs_arc_overflow_shift.
5140 arc_is_overflowing(void)
5142 /* Always allow at least one block of overflow */
5143 int64_t overflow = MAX(SPA_MAXBLOCKSIZE,
5144 arc_c >> zfs_arc_overflow_shift);
5147 * We just compare the lower bound here for performance reasons. Our
5148 * primary goals are to make sure that the arc never grows without
5149 * bound, and that it can reach its maximum size. This check
5150 * accomplishes both goals. The maximum amount we could run over by is
5151 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
5152 * in the ARC. In practice, that's in the tens of MB, which is low
5153 * enough to be safe.
5155 return (aggsum_lower_bound(&arc_size) >= (int64_t)arc_c + overflow);
5159 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag, boolean_t do_adapt)
5161 arc_buf_contents_t type = arc_buf_type(hdr);
5163 arc_get_data_impl(hdr, size, tag, do_adapt);
5164 if (type == ARC_BUFC_METADATA) {
5165 return (abd_alloc(size, B_TRUE));
5167 ASSERT(type == ARC_BUFC_DATA);
5168 return (abd_alloc(size, B_FALSE));
5173 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5175 arc_buf_contents_t type = arc_buf_type(hdr);
5177 arc_get_data_impl(hdr, size, tag, B_TRUE);
5178 if (type == ARC_BUFC_METADATA) {
5179 return (zio_buf_alloc(size));
5181 ASSERT(type == ARC_BUFC_DATA);
5182 return (zio_data_buf_alloc(size));
5187 * Allocate a block and return it to the caller. If we are hitting the
5188 * hard limit for the cache size, we must sleep, waiting for the eviction
5189 * thread to catch up. If we're past the target size but below the hard
5190 * limit, we'll only signal the reclaim thread and continue on.
5193 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag, boolean_t do_adapt)
5195 arc_state_t *state = hdr->b_l1hdr.b_state;
5196 arc_buf_contents_t type = arc_buf_type(hdr);
5199 arc_adapt(size, state);
5202 * If arc_size is currently overflowing, and has grown past our
5203 * upper limit, we must be adding data faster than the evict
5204 * thread can evict. Thus, to ensure we don't compound the
5205 * problem by adding more data and forcing arc_size to grow even
5206 * further past it's target size, we halt and wait for the
5207 * eviction thread to catch up.
5209 * It's also possible that the reclaim thread is unable to evict
5210 * enough buffers to get arc_size below the overflow limit (e.g.
5211 * due to buffers being un-evictable, or hash lock collisions).
5212 * In this case, we want to proceed regardless if we're
5213 * overflowing; thus we don't use a while loop here.
5215 if (arc_is_overflowing()) {
5216 mutex_enter(&arc_adjust_lock);
5219 * Now that we've acquired the lock, we may no longer be
5220 * over the overflow limit, lets check.
5222 * We're ignoring the case of spurious wake ups. If that
5223 * were to happen, it'd let this thread consume an ARC
5224 * buffer before it should have (i.e. before we're under
5225 * the overflow limit and were signalled by the reclaim
5226 * thread). As long as that is a rare occurrence, it
5227 * shouldn't cause any harm.
5229 if (arc_is_overflowing()) {
5230 arc_adjust_needed = B_TRUE;
5231 zthr_wakeup(arc_adjust_zthr);
5232 (void) cv_wait(&arc_adjust_waiters_cv,
5235 mutex_exit(&arc_adjust_lock);
5238 VERIFY3U(hdr->b_type, ==, type);
5239 if (type == ARC_BUFC_METADATA) {
5240 arc_space_consume(size, ARC_SPACE_META);
5242 arc_space_consume(size, ARC_SPACE_DATA);
5246 * Update the state size. Note that ghost states have a
5247 * "ghost size" and so don't need to be updated.
5249 if (!GHOST_STATE(state)) {
5251 (void) refcount_add_many(&state->arcs_size, size, tag);
5254 * If this is reached via arc_read, the link is
5255 * protected by the hash lock. If reached via
5256 * arc_buf_alloc, the header should not be accessed by
5257 * any other thread. And, if reached via arc_read_done,
5258 * the hash lock will protect it if it's found in the
5259 * hash table; otherwise no other thread should be
5260 * trying to [add|remove]_reference it.
5262 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5263 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5264 (void) refcount_add_many(&state->arcs_esize[type],
5269 * If we are growing the cache, and we are adding anonymous
5270 * data, and we have outgrown arc_p, update arc_p
5272 if (aggsum_upper_bound(&arc_size) < arc_c &&
5273 hdr->b_l1hdr.b_state == arc_anon &&
5274 (refcount_count(&arc_anon->arcs_size) +
5275 refcount_count(&arc_mru->arcs_size) > arc_p))
5276 arc_p = MIN(arc_c, arc_p + size);
5278 ARCSTAT_BUMP(arcstat_allocated);
5282 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
5284 arc_free_data_impl(hdr, size, tag);
5289 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
5291 arc_buf_contents_t type = arc_buf_type(hdr);
5293 arc_free_data_impl(hdr, size, tag);
5294 if (type == ARC_BUFC_METADATA) {
5295 zio_buf_free(buf, size);
5297 ASSERT(type == ARC_BUFC_DATA);
5298 zio_data_buf_free(buf, size);
5303 * Free the arc data buffer.
5306 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5308 arc_state_t *state = hdr->b_l1hdr.b_state;
5309 arc_buf_contents_t type = arc_buf_type(hdr);
5311 /* protected by hash lock, if in the hash table */
5312 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5313 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5314 ASSERT(state != arc_anon && state != arc_l2c_only);
5316 (void) refcount_remove_many(&state->arcs_esize[type],
5319 (void) refcount_remove_many(&state->arcs_size, size, tag);
5321 VERIFY3U(hdr->b_type, ==, type);
5322 if (type == ARC_BUFC_METADATA) {
5323 arc_space_return(size, ARC_SPACE_META);
5325 ASSERT(type == ARC_BUFC_DATA);
5326 arc_space_return(size, ARC_SPACE_DATA);
5331 * This routine is called whenever a buffer is accessed.
5332 * NOTE: the hash lock is dropped in this function.
5335 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
5339 ASSERT(MUTEX_HELD(hash_lock));
5340 ASSERT(HDR_HAS_L1HDR(hdr));
5342 if (hdr->b_l1hdr.b_state == arc_anon) {
5344 * This buffer is not in the cache, and does not
5345 * appear in our "ghost" list. Add the new buffer
5349 ASSERT0(hdr->b_l1hdr.b_arc_access);
5350 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5351 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5352 arc_change_state(arc_mru, hdr, hash_lock);
5354 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5355 now = ddi_get_lbolt();
5358 * If this buffer is here because of a prefetch, then either:
5359 * - clear the flag if this is a "referencing" read
5360 * (any subsequent access will bump this into the MFU state).
5362 * - move the buffer to the head of the list if this is
5363 * another prefetch (to make it less likely to be evicted).
5365 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5366 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5367 /* link protected by hash lock */
5368 ASSERT(multilist_link_active(
5369 &hdr->b_l1hdr.b_arc_node));
5371 arc_hdr_clear_flags(hdr,
5373 ARC_FLAG_PRESCIENT_PREFETCH);
5374 ARCSTAT_BUMP(arcstat_mru_hits);
5376 hdr->b_l1hdr.b_arc_access = now;
5381 * This buffer has been "accessed" only once so far,
5382 * but it is still in the cache. Move it to the MFU
5385 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
5387 * More than 125ms have passed since we
5388 * instantiated this buffer. Move it to the
5389 * most frequently used state.
5391 hdr->b_l1hdr.b_arc_access = now;
5392 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5393 arc_change_state(arc_mfu, hdr, hash_lock);
5395 atomic_inc_32(&hdr->b_l1hdr.b_mru_hits);
5396 ARCSTAT_BUMP(arcstat_mru_hits);
5397 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5398 arc_state_t *new_state;
5400 * This buffer has been "accessed" recently, but
5401 * was evicted from the cache. Move it to the
5405 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5406 new_state = arc_mru;
5407 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
5408 arc_hdr_clear_flags(hdr,
5410 ARC_FLAG_PRESCIENT_PREFETCH);
5412 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5414 new_state = arc_mfu;
5415 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5418 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5419 arc_change_state(new_state, hdr, hash_lock);
5421 atomic_inc_32(&hdr->b_l1hdr.b_mru_ghost_hits);
5422 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5423 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5425 * This buffer has been accessed more than once and is
5426 * still in the cache. Keep it in the MFU state.
5428 * NOTE: an add_reference() that occurred when we did
5429 * the arc_read() will have kicked this off the list.
5430 * If it was a prefetch, we will explicitly move it to
5431 * the head of the list now.
5434 atomic_inc_32(&hdr->b_l1hdr.b_mfu_hits);
5435 ARCSTAT_BUMP(arcstat_mfu_hits);
5436 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5437 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5438 arc_state_t *new_state = arc_mfu;
5440 * This buffer has been accessed more than once but has
5441 * been evicted from the cache. Move it back to the
5445 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5447 * This is a prefetch access...
5448 * move this block back to the MRU state.
5450 new_state = arc_mru;
5453 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5454 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5455 arc_change_state(new_state, hdr, hash_lock);
5457 atomic_inc_32(&hdr->b_l1hdr.b_mfu_ghost_hits);
5458 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5459 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5461 * This buffer is on the 2nd Level ARC.
5464 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5465 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5466 arc_change_state(arc_mfu, hdr, hash_lock);
5468 ASSERT(!"invalid arc state");
5473 * This routine is called by dbuf_hold() to update the arc_access() state
5474 * which otherwise would be skipped for entries in the dbuf cache.
5477 arc_buf_access(arc_buf_t *buf)
5479 mutex_enter(&buf->b_evict_lock);
5480 arc_buf_hdr_t *hdr = buf->b_hdr;
5483 * Avoid taking the hash_lock when possible as an optimization.
5484 * The header must be checked again under the hash_lock in order
5485 * to handle the case where it is concurrently being released.
5487 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5488 mutex_exit(&buf->b_evict_lock);
5489 ARCSTAT_BUMP(arcstat_access_skip);
5493 kmutex_t *hash_lock = HDR_LOCK(hdr);
5494 mutex_enter(hash_lock);
5496 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5497 mutex_exit(hash_lock);
5498 mutex_exit(&buf->b_evict_lock);
5499 ARCSTAT_BUMP(arcstat_access_skip);
5503 mutex_exit(&buf->b_evict_lock);
5505 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5506 hdr->b_l1hdr.b_state == arc_mfu);
5508 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5509 arc_access(hdr, hash_lock);
5510 mutex_exit(hash_lock);
5512 ARCSTAT_BUMP(arcstat_hits);
5513 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5514 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5517 /* a generic arc_read_done_func_t which you can use */
5520 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5521 arc_buf_t *buf, void *arg)
5526 bcopy(buf->b_data, arg, arc_buf_size(buf));
5527 arc_buf_destroy(buf, arg);
5530 /* a generic arc_read_done_func_t */
5533 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5534 arc_buf_t *buf, void *arg)
5536 arc_buf_t **bufp = arg;
5538 ASSERT(zio == NULL || zio->io_error != 0);
5541 ASSERT(zio == NULL || zio->io_error == 0);
5543 ASSERT(buf->b_data != NULL);
5548 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5550 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5551 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5552 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
5554 if (HDR_COMPRESSION_ENABLED(hdr)) {
5555 ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
5556 BP_GET_COMPRESS(bp));
5558 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5559 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5564 arc_read_done(zio_t *zio)
5566 arc_buf_hdr_t *hdr = zio->io_private;
5567 kmutex_t *hash_lock = NULL;
5568 arc_callback_t *callback_list;
5569 arc_callback_t *acb;
5570 boolean_t freeable = B_FALSE;
5571 boolean_t no_zio_error = (zio->io_error == 0);
5574 * The hdr was inserted into hash-table and removed from lists
5575 * prior to starting I/O. We should find this header, since
5576 * it's in the hash table, and it should be legit since it's
5577 * not possible to evict it during the I/O. The only possible
5578 * reason for it not to be found is if we were freed during the
5581 if (HDR_IN_HASH_TABLE(hdr)) {
5582 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5583 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5584 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5585 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5586 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5588 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
5591 ASSERT((found == hdr &&
5592 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5593 (found == hdr && HDR_L2_READING(hdr)));
5594 ASSERT3P(hash_lock, !=, NULL);
5598 /* byteswap if necessary */
5599 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5600 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5601 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5603 hdr->b_l1hdr.b_byteswap =
5604 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5607 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5611 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5612 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5613 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5615 callback_list = hdr->b_l1hdr.b_acb;
5616 ASSERT3P(callback_list, !=, NULL);
5618 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) {
5620 * Only call arc_access on anonymous buffers. This is because
5621 * if we've issued an I/O for an evicted buffer, we've already
5622 * called arc_access (to prevent any simultaneous readers from
5623 * getting confused).
5625 arc_access(hdr, hash_lock);
5629 * If a read request has a callback (i.e. acb_done is not NULL), then we
5630 * make a buf containing the data according to the parameters which were
5631 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5632 * aren't needlessly decompressing the data multiple times.
5634 int callback_cnt = 0;
5635 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5642 int error = arc_buf_alloc_impl(hdr, acb->acb_private,
5643 acb->acb_compressed, zio->io_error == 0,
5647 * Decompression failed. Set io_error
5648 * so that when we call acb_done (below),
5649 * we will indicate that the read failed.
5650 * Note that in the unusual case where one
5651 * callback is compressed and another
5652 * uncompressed, we will mark all of them
5653 * as failed, even though the uncompressed
5654 * one can't actually fail. In this case,
5655 * the hdr will not be anonymous, because
5656 * if there are multiple callbacks, it's
5657 * because multiple threads found the same
5658 * arc buf in the hash table.
5660 zio->io_error = error;
5665 * If there are multiple callbacks, we must have the hash lock,
5666 * because the only way for multiple threads to find this hdr is
5667 * in the hash table. This ensures that if there are multiple
5668 * callbacks, the hdr is not anonymous. If it were anonymous,
5669 * we couldn't use arc_buf_destroy() in the error case below.
5671 ASSERT(callback_cnt < 2 || hash_lock != NULL);
5673 hdr->b_l1hdr.b_acb = NULL;
5674 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5675 if (callback_cnt == 0) {
5676 ASSERT(HDR_PREFETCH(hdr));
5677 ASSERT0(hdr->b_l1hdr.b_bufcnt);
5678 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5681 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5682 callback_list != NULL);
5685 arc_hdr_verify(hdr, zio->io_bp);
5687 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5688 if (hdr->b_l1hdr.b_state != arc_anon)
5689 arc_change_state(arc_anon, hdr, hash_lock);
5690 if (HDR_IN_HASH_TABLE(hdr))
5691 buf_hash_remove(hdr);
5692 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5696 * Broadcast before we drop the hash_lock to avoid the possibility
5697 * that the hdr (and hence the cv) might be freed before we get to
5698 * the cv_broadcast().
5700 cv_broadcast(&hdr->b_l1hdr.b_cv);
5702 if (hash_lock != NULL) {
5703 mutex_exit(hash_lock);
5706 * This block was freed while we waited for the read to
5707 * complete. It has been removed from the hash table and
5708 * moved to the anonymous state (so that it won't show up
5711 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5712 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5715 /* execute each callback and free its structure */
5716 while ((acb = callback_list) != NULL) {
5717 if (acb->acb_done != NULL) {
5718 if (zio->io_error != 0 && acb->acb_buf != NULL) {
5720 * If arc_buf_alloc_impl() fails during
5721 * decompression, the buf will still be
5722 * allocated, and needs to be freed here.
5724 arc_buf_destroy(acb->acb_buf, acb->acb_private);
5725 acb->acb_buf = NULL;
5727 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5728 acb->acb_buf, acb->acb_private);
5731 if (acb->acb_zio_dummy != NULL) {
5732 acb->acb_zio_dummy->io_error = zio->io_error;
5733 zio_nowait(acb->acb_zio_dummy);
5736 callback_list = acb->acb_next;
5737 kmem_free(acb, sizeof (arc_callback_t));
5741 arc_hdr_destroy(hdr);
5745 * "Read" the block at the specified DVA (in bp) via the
5746 * cache. If the block is found in the cache, invoke the provided
5747 * callback immediately and return. Note that the `zio' parameter
5748 * in the callback will be NULL in this case, since no IO was
5749 * required. If the block is not in the cache pass the read request
5750 * on to the spa with a substitute callback function, so that the
5751 * requested block will be added to the cache.
5753 * If a read request arrives for a block that has a read in-progress,
5754 * either wait for the in-progress read to complete (and return the
5755 * results); or, if this is a read with a "done" func, add a record
5756 * to the read to invoke the "done" func when the read completes,
5757 * and return; or just return.
5759 * arc_read_done() will invoke all the requested "done" functions
5760 * for readers of this block.
5763 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_read_done_func_t *done,
5764 void *private, zio_priority_t priority, int zio_flags,
5765 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5767 arc_buf_hdr_t *hdr = NULL;
5768 kmutex_t *hash_lock = NULL;
5770 uint64_t guid = spa_load_guid(spa);
5771 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0;
5774 ASSERT(!BP_IS_EMBEDDED(bp) ||
5775 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5778 if (!BP_IS_EMBEDDED(bp)) {
5780 * Embedded BP's have no DVA and require no I/O to "read".
5781 * Create an anonymous arc buf to back it.
5783 hdr = buf_hash_find(guid, bp, &hash_lock);
5786 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) {
5787 arc_buf_t *buf = NULL;
5788 *arc_flags |= ARC_FLAG_CACHED;
5790 if (HDR_IO_IN_PROGRESS(hdr)) {
5791 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
5793 ASSERT3P(head_zio, !=, NULL);
5794 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5795 priority == ZIO_PRIORITY_SYNC_READ) {
5797 * This is a sync read that needs to wait for
5798 * an in-flight async read. Request that the
5799 * zio have its priority upgraded.
5801 zio_change_priority(head_zio, priority);
5802 DTRACE_PROBE1(arc__async__upgrade__sync,
5803 arc_buf_hdr_t *, hdr);
5804 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
5806 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5807 arc_hdr_clear_flags(hdr,
5808 ARC_FLAG_PREDICTIVE_PREFETCH);
5811 if (*arc_flags & ARC_FLAG_WAIT) {
5812 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5813 mutex_exit(hash_lock);
5816 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5819 arc_callback_t *acb = NULL;
5821 acb = kmem_zalloc(sizeof (arc_callback_t),
5823 acb->acb_done = done;
5824 acb->acb_private = private;
5825 acb->acb_compressed = compressed_read;
5827 acb->acb_zio_dummy = zio_null(pio,
5828 spa, NULL, NULL, NULL, zio_flags);
5830 ASSERT3P(acb->acb_done, !=, NULL);
5831 acb->acb_zio_head = head_zio;
5832 acb->acb_next = hdr->b_l1hdr.b_acb;
5833 hdr->b_l1hdr.b_acb = acb;
5834 mutex_exit(hash_lock);
5837 mutex_exit(hash_lock);
5841 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5842 hdr->b_l1hdr.b_state == arc_mfu);
5845 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5847 * This is a demand read which does not have to
5848 * wait for i/o because we did a predictive
5849 * prefetch i/o for it, which has completed.
5852 arc__demand__hit__predictive__prefetch,
5853 arc_buf_hdr_t *, hdr);
5855 arcstat_demand_hit_predictive_prefetch);
5856 arc_hdr_clear_flags(hdr,
5857 ARC_FLAG_PREDICTIVE_PREFETCH);
5860 if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
5862 arcstat_demand_hit_prescient_prefetch);
5863 arc_hdr_clear_flags(hdr,
5864 ARC_FLAG_PRESCIENT_PREFETCH);
5867 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5868 /* Get a buf with the desired data in it. */
5869 rc = arc_buf_alloc_impl(hdr, private,
5870 compressed_read, B_TRUE, &buf);
5872 arc_buf_destroy(buf, private);
5875 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
5876 rc == 0 || rc != ENOENT);
5877 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
5878 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5879 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5881 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5882 arc_access(hdr, hash_lock);
5883 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5884 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5885 if (*arc_flags & ARC_FLAG_L2CACHE)
5886 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5887 mutex_exit(hash_lock);
5888 ARCSTAT_BUMP(arcstat_hits);
5889 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5890 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5891 data, metadata, hits);
5894 done(NULL, zb, bp, buf, private);
5896 uint64_t lsize = BP_GET_LSIZE(bp);
5897 uint64_t psize = BP_GET_PSIZE(bp);
5898 arc_callback_t *acb;
5901 boolean_t devw = B_FALSE;
5905 /* this block is not in the cache */
5906 arc_buf_hdr_t *exists = NULL;
5907 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5908 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5909 BP_GET_COMPRESS(bp), type);
5911 if (!BP_IS_EMBEDDED(bp)) {
5912 hdr->b_dva = *BP_IDENTITY(bp);
5913 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5914 exists = buf_hash_insert(hdr, &hash_lock);
5916 if (exists != NULL) {
5917 /* somebody beat us to the hash insert */
5918 mutex_exit(hash_lock);
5919 buf_discard_identity(hdr);
5920 arc_hdr_destroy(hdr);
5921 goto top; /* restart the IO request */
5925 * This block is in the ghost cache. If it was L2-only
5926 * (and thus didn't have an L1 hdr), we realloc the
5927 * header to add an L1 hdr.
5929 if (!HDR_HAS_L1HDR(hdr)) {
5930 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5933 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5934 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5935 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5936 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5937 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5938 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5941 * This is a delicate dance that we play here.
5942 * This hdr is in the ghost list so we access it
5943 * to move it out of the ghost list before we
5944 * initiate the read. If it's a prefetch then
5945 * it won't have a callback so we'll remove the
5946 * reference that arc_buf_alloc_impl() created. We
5947 * do this after we've called arc_access() to
5948 * avoid hitting an assert in remove_reference().
5950 arc_adapt(arc_hdr_size(hdr), hdr->b_l1hdr.b_state);
5951 arc_access(hdr, hash_lock);
5952 arc_hdr_alloc_pabd(hdr, B_FALSE);
5954 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5955 size = arc_hdr_size(hdr);
5958 * If compression is enabled on the hdr, then will do
5959 * RAW I/O and will store the compressed data in the hdr's
5960 * data block. Otherwise, the hdr's data block will contain
5961 * the uncompressed data.
5963 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5964 zio_flags |= ZIO_FLAG_RAW;
5967 if (*arc_flags & ARC_FLAG_PREFETCH)
5968 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5969 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5970 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5972 if (*arc_flags & ARC_FLAG_L2CACHE)
5973 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5974 if (BP_GET_LEVEL(bp) > 0)
5975 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5976 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5977 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5978 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5980 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5981 acb->acb_done = done;
5982 acb->acb_private = private;
5983 acb->acb_compressed = compressed_read;
5985 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5986 hdr->b_l1hdr.b_acb = acb;
5987 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5989 if (HDR_HAS_L2HDR(hdr) &&
5990 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5991 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5992 addr = hdr->b_l2hdr.b_daddr;
5994 * Lock out L2ARC device removal.
5996 if (vdev_is_dead(vd) ||
5997 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
6002 * We count both async reads and scrub IOs as asynchronous so
6003 * that both can be upgraded in the event of a cache hit while
6004 * the read IO is still in-flight.
6006 if (priority == ZIO_PRIORITY_ASYNC_READ ||
6007 priority == ZIO_PRIORITY_SCRUB)
6008 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6010 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6013 * At this point, we have a level 1 cache miss. Try again in
6014 * L2ARC if possible.
6016 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
6018 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
6019 uint64_t, lsize, zbookmark_phys_t *, zb);
6020 ARCSTAT_BUMP(arcstat_misses);
6021 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
6022 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
6023 data, metadata, misses);
6028 racct_add_force(curproc, RACCT_READBPS, size);
6029 racct_add_force(curproc, RACCT_READIOPS, 1);
6030 PROC_UNLOCK(curproc);
6033 curthread->td_ru.ru_inblock++;
6036 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
6038 * Read from the L2ARC if the following are true:
6039 * 1. The L2ARC vdev was previously cached.
6040 * 2. This buffer still has L2ARC metadata.
6041 * 3. This buffer isn't currently writing to the L2ARC.
6042 * 4. The L2ARC entry wasn't evicted, which may
6043 * also have invalidated the vdev.
6044 * 5. This isn't prefetch and l2arc_noprefetch is set.
6046 if (HDR_HAS_L2HDR(hdr) &&
6047 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
6048 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
6049 l2arc_read_callback_t *cb;
6053 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
6054 ARCSTAT_BUMP(arcstat_l2_hits);
6055 atomic_inc_32(&hdr->b_l2hdr.b_hits);
6057 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
6059 cb->l2rcb_hdr = hdr;
6062 cb->l2rcb_flags = zio_flags;
6064 asize = vdev_psize_to_asize(vd, size);
6065 if (asize != size) {
6066 abd = abd_alloc_for_io(asize,
6067 HDR_ISTYPE_METADATA(hdr));
6068 cb->l2rcb_abd = abd;
6070 abd = hdr->b_l1hdr.b_pabd;
6073 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
6074 addr + asize <= vd->vdev_psize -
6075 VDEV_LABEL_END_SIZE);
6078 * l2arc read. The SCL_L2ARC lock will be
6079 * released by l2arc_read_done().
6080 * Issue a null zio if the underlying buffer
6081 * was squashed to zero size by compression.
6083 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
6084 ZIO_COMPRESS_EMPTY);
6085 rzio = zio_read_phys(pio, vd, addr,
6088 l2arc_read_done, cb, priority,
6089 zio_flags | ZIO_FLAG_DONT_CACHE |
6091 ZIO_FLAG_DONT_PROPAGATE |
6092 ZIO_FLAG_DONT_RETRY, B_FALSE);
6093 acb->acb_zio_head = rzio;
6095 if (hash_lock != NULL)
6096 mutex_exit(hash_lock);
6098 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
6100 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
6102 if (*arc_flags & ARC_FLAG_NOWAIT) {
6107 ASSERT(*arc_flags & ARC_FLAG_WAIT);
6108 if (zio_wait(rzio) == 0)
6111 /* l2arc read error; goto zio_read() */
6112 if (hash_lock != NULL)
6113 mutex_enter(hash_lock);
6115 DTRACE_PROBE1(l2arc__miss,
6116 arc_buf_hdr_t *, hdr);
6117 ARCSTAT_BUMP(arcstat_l2_misses);
6118 if (HDR_L2_WRITING(hdr))
6119 ARCSTAT_BUMP(arcstat_l2_rw_clash);
6120 spa_config_exit(spa, SCL_L2ARC, vd);
6124 spa_config_exit(spa, SCL_L2ARC, vd);
6125 if (l2arc_ndev != 0) {
6126 DTRACE_PROBE1(l2arc__miss,
6127 arc_buf_hdr_t *, hdr);
6128 ARCSTAT_BUMP(arcstat_l2_misses);
6132 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size,
6133 arc_read_done, hdr, priority, zio_flags, zb);
6134 acb->acb_zio_head = rzio;
6136 if (hash_lock != NULL)
6137 mutex_exit(hash_lock);
6139 if (*arc_flags & ARC_FLAG_WAIT)
6140 return (zio_wait(rzio));
6142 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6149 arc_add_prune_callback(arc_prune_func_t *func, void *private)
6153 p = kmem_alloc(sizeof (*p), KM_SLEEP);
6155 p->p_private = private;
6156 list_link_init(&p->p_node);
6157 refcount_create(&p->p_refcnt);
6159 mutex_enter(&arc_prune_mtx);
6160 refcount_add(&p->p_refcnt, &arc_prune_list);
6161 list_insert_head(&arc_prune_list, p);
6162 mutex_exit(&arc_prune_mtx);
6168 arc_remove_prune_callback(arc_prune_t *p)
6170 boolean_t wait = B_FALSE;
6171 mutex_enter(&arc_prune_mtx);
6172 list_remove(&arc_prune_list, p);
6173 if (refcount_remove(&p->p_refcnt, &arc_prune_list) > 0)
6175 mutex_exit(&arc_prune_mtx);
6177 /* wait for arc_prune_task to finish */
6179 taskq_wait(arc_prune_taskq);
6180 ASSERT0(refcount_count(&p->p_refcnt));
6181 refcount_destroy(&p->p_refcnt);
6182 kmem_free(p, sizeof (*p));
6186 * Notify the arc that a block was freed, and thus will never be used again.
6189 arc_freed(spa_t *spa, const blkptr_t *bp)
6192 kmutex_t *hash_lock;
6193 uint64_t guid = spa_load_guid(spa);
6195 ASSERT(!BP_IS_EMBEDDED(bp));
6197 hdr = buf_hash_find(guid, bp, &hash_lock);
6202 * We might be trying to free a block that is still doing I/O
6203 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
6204 * dmu_sync-ed block). If this block is being prefetched, then it
6205 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
6206 * until the I/O completes. A block may also have a reference if it is
6207 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6208 * have written the new block to its final resting place on disk but
6209 * without the dedup flag set. This would have left the hdr in the MRU
6210 * state and discoverable. When the txg finally syncs it detects that
6211 * the block was overridden in open context and issues an override I/O.
6212 * Since this is a dedup block, the override I/O will determine if the
6213 * block is already in the DDT. If so, then it will replace the io_bp
6214 * with the bp from the DDT and allow the I/O to finish. When the I/O
6215 * reaches the done callback, dbuf_write_override_done, it will
6216 * check to see if the io_bp and io_bp_override are identical.
6217 * If they are not, then it indicates that the bp was replaced with
6218 * the bp in the DDT and the override bp is freed. This allows
6219 * us to arrive here with a reference on a block that is being
6220 * freed. So if we have an I/O in progress, or a reference to
6221 * this hdr, then we don't destroy the hdr.
6223 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
6224 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
6225 arc_change_state(arc_anon, hdr, hash_lock);
6226 arc_hdr_destroy(hdr);
6227 mutex_exit(hash_lock);
6229 mutex_exit(hash_lock);
6235 * Release this buffer from the cache, making it an anonymous buffer. This
6236 * must be done after a read and prior to modifying the buffer contents.
6237 * If the buffer has more than one reference, we must make
6238 * a new hdr for the buffer.
6241 arc_release(arc_buf_t *buf, void *tag)
6243 arc_buf_hdr_t *hdr = buf->b_hdr;
6246 * It would be nice to assert that if it's DMU metadata (level >
6247 * 0 || it's the dnode file), then it must be syncing context.
6248 * But we don't know that information at this level.
6251 mutex_enter(&buf->b_evict_lock);
6253 ASSERT(HDR_HAS_L1HDR(hdr));
6256 * We don't grab the hash lock prior to this check, because if
6257 * the buffer's header is in the arc_anon state, it won't be
6258 * linked into the hash table.
6260 if (hdr->b_l1hdr.b_state == arc_anon) {
6261 mutex_exit(&buf->b_evict_lock);
6262 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6263 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6264 ASSERT(!HDR_HAS_L2HDR(hdr));
6265 ASSERT(HDR_EMPTY(hdr));
6266 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6267 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6268 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
6270 hdr->b_l1hdr.b_arc_access = 0;
6273 * If the buf is being overridden then it may already
6274 * have a hdr that is not empty.
6276 buf_discard_identity(hdr);
6282 kmutex_t *hash_lock = HDR_LOCK(hdr);
6283 mutex_enter(hash_lock);
6286 * This assignment is only valid as long as the hash_lock is
6287 * held, we must be careful not to reference state or the
6288 * b_state field after dropping the lock.
6290 arc_state_t *state = hdr->b_l1hdr.b_state;
6291 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6292 ASSERT3P(state, !=, arc_anon);
6294 /* this buffer is not on any list */
6295 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6297 if (HDR_HAS_L2HDR(hdr)) {
6298 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6301 * We have to recheck this conditional again now that
6302 * we're holding the l2ad_mtx to prevent a race with
6303 * another thread which might be concurrently calling
6304 * l2arc_evict(). In that case, l2arc_evict() might have
6305 * destroyed the header's L2 portion as we were waiting
6306 * to acquire the l2ad_mtx.
6308 if (HDR_HAS_L2HDR(hdr)) {
6310 arc_hdr_l2hdr_destroy(hdr);
6313 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6317 * Do we have more than one buf?
6319 if (hdr->b_l1hdr.b_bufcnt > 1) {
6320 arc_buf_hdr_t *nhdr;
6321 uint64_t spa = hdr->b_spa;
6322 uint64_t psize = HDR_GET_PSIZE(hdr);
6323 uint64_t lsize = HDR_GET_LSIZE(hdr);
6324 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
6325 arc_buf_contents_t type = arc_buf_type(hdr);
6326 VERIFY3U(hdr->b_type, ==, type);
6328 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6329 (void) remove_reference(hdr, hash_lock, tag);
6331 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
6332 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6333 ASSERT(ARC_BUF_LAST(buf));
6337 * Pull the data off of this hdr and attach it to
6338 * a new anonymous hdr. Also find the last buffer
6339 * in the hdr's buffer list.
6341 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6342 ASSERT3P(lastbuf, !=, NULL);
6345 * If the current arc_buf_t and the hdr are sharing their data
6346 * buffer, then we must stop sharing that block.
6348 if (arc_buf_is_shared(buf)) {
6349 VERIFY(!arc_buf_is_shared(lastbuf));
6352 * First, sever the block sharing relationship between
6353 * buf and the arc_buf_hdr_t.
6355 arc_unshare_buf(hdr, buf);
6358 * Now we need to recreate the hdr's b_pabd. Since we
6359 * have lastbuf handy, we try to share with it, but if
6360 * we can't then we allocate a new b_pabd and copy the
6361 * data from buf into it.
6363 if (arc_can_share(hdr, lastbuf)) {
6364 arc_share_buf(hdr, lastbuf);
6366 arc_hdr_alloc_pabd(hdr, B_TRUE);
6367 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6368 buf->b_data, psize);
6370 VERIFY3P(lastbuf->b_data, !=, NULL);
6371 } else if (HDR_SHARED_DATA(hdr)) {
6373 * Uncompressed shared buffers are always at the end
6374 * of the list. Compressed buffers don't have the
6375 * same requirements. This makes it hard to
6376 * simply assert that the lastbuf is shared so
6377 * we rely on the hdr's compression flags to determine
6378 * if we have a compressed, shared buffer.
6380 ASSERT(arc_buf_is_shared(lastbuf) ||
6381 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
6382 ASSERT(!ARC_BUF_SHARED(buf));
6384 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
6385 ASSERT3P(state, !=, arc_l2c_only);
6387 (void) refcount_remove_many(&state->arcs_size,
6388 arc_buf_size(buf), buf);
6390 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6391 ASSERT3P(state, !=, arc_l2c_only);
6392 (void) refcount_remove_many(&state->arcs_esize[type],
6393 arc_buf_size(buf), buf);
6396 hdr->b_l1hdr.b_bufcnt -= 1;
6397 arc_cksum_verify(buf);
6399 arc_buf_unwatch(buf);
6402 mutex_exit(hash_lock);
6405 * Allocate a new hdr. The new hdr will contain a b_pabd
6406 * buffer which will be freed in arc_write().
6408 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
6409 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6410 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6411 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
6412 VERIFY3U(nhdr->b_type, ==, type);
6413 ASSERT(!HDR_SHARED_DATA(nhdr));
6415 nhdr->b_l1hdr.b_buf = buf;
6416 nhdr->b_l1hdr.b_bufcnt = 1;
6417 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6420 mutex_exit(&buf->b_evict_lock);
6421 (void) refcount_add_many(&arc_anon->arcs_size,
6422 arc_buf_size(buf), buf);
6424 mutex_exit(&buf->b_evict_lock);
6425 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6426 /* protected by hash lock, or hdr is on arc_anon */
6427 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6428 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6429 arc_change_state(arc_anon, hdr, hash_lock);
6430 hdr->b_l1hdr.b_arc_access = 0;
6431 mutex_exit(hash_lock);
6433 buf_discard_identity(hdr);
6439 arc_released(arc_buf_t *buf)
6443 mutex_enter(&buf->b_evict_lock);
6444 released = (buf->b_data != NULL &&
6445 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6446 mutex_exit(&buf->b_evict_lock);
6452 arc_referenced(arc_buf_t *buf)
6456 mutex_enter(&buf->b_evict_lock);
6457 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6458 mutex_exit(&buf->b_evict_lock);
6459 return (referenced);
6464 arc_write_ready(zio_t *zio)
6466 arc_write_callback_t *callback = zio->io_private;
6467 arc_buf_t *buf = callback->awcb_buf;
6468 arc_buf_hdr_t *hdr = buf->b_hdr;
6469 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
6471 ASSERT(HDR_HAS_L1HDR(hdr));
6472 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6473 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6476 * If we're reexecuting this zio because the pool suspended, then
6477 * cleanup any state that was previously set the first time the
6478 * callback was invoked.
6480 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6481 arc_cksum_free(hdr);
6483 arc_buf_unwatch(buf);
6485 if (hdr->b_l1hdr.b_pabd != NULL) {
6486 if (arc_buf_is_shared(buf)) {
6487 arc_unshare_buf(hdr, buf);
6489 arc_hdr_free_pabd(hdr);
6493 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6494 ASSERT(!HDR_SHARED_DATA(hdr));
6495 ASSERT(!arc_buf_is_shared(buf));
6497 callback->awcb_ready(zio, buf, callback->awcb_private);
6499 if (HDR_IO_IN_PROGRESS(hdr))
6500 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6502 arc_cksum_compute(buf);
6503 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6505 enum zio_compress compress;
6506 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6507 compress = ZIO_COMPRESS_OFF;
6509 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
6510 compress = BP_GET_COMPRESS(zio->io_bp);
6512 HDR_SET_PSIZE(hdr, psize);
6513 arc_hdr_set_compress(hdr, compress);
6517 * Fill the hdr with data. If the hdr is compressed, the data we want
6518 * is available from the zio, otherwise we can take it from the buf.
6520 * We might be able to share the buf's data with the hdr here. However,
6521 * doing so would cause the ARC to be full of linear ABDs if we write a
6522 * lot of shareable data. As a compromise, we check whether scattered
6523 * ABDs are allowed, and assume that if they are then the user wants
6524 * the ARC to be primarily filled with them regardless of the data being
6525 * written. Therefore, if they're allowed then we allocate one and copy
6526 * the data into it; otherwise, we share the data directly if we can.
6528 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
6529 arc_hdr_alloc_pabd(hdr, B_TRUE);
6532 * Ideally, we would always copy the io_abd into b_pabd, but the
6533 * user may have disabled compressed ARC, thus we must check the
6534 * hdr's compression setting rather than the io_bp's.
6536 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
6537 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=,
6539 ASSERT3U(psize, >, 0);
6541 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6543 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6545 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6549 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6550 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6551 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6553 arc_share_buf(hdr, buf);
6556 arc_hdr_verify(hdr, zio->io_bp);
6560 arc_write_children_ready(zio_t *zio)
6562 arc_write_callback_t *callback = zio->io_private;
6563 arc_buf_t *buf = callback->awcb_buf;
6565 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6569 * The SPA calls this callback for each physical write that happens on behalf
6570 * of a logical write. See the comment in dbuf_write_physdone() for details.
6573 arc_write_physdone(zio_t *zio)
6575 arc_write_callback_t *cb = zio->io_private;
6576 if (cb->awcb_physdone != NULL)
6577 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
6581 arc_write_done(zio_t *zio)
6583 arc_write_callback_t *callback = zio->io_private;
6584 arc_buf_t *buf = callback->awcb_buf;
6585 arc_buf_hdr_t *hdr = buf->b_hdr;
6587 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6589 if (zio->io_error == 0) {
6590 arc_hdr_verify(hdr, zio->io_bp);
6592 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6593 buf_discard_identity(hdr);
6595 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6596 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6599 ASSERT(HDR_EMPTY(hdr));
6603 * If the block to be written was all-zero or compressed enough to be
6604 * embedded in the BP, no write was performed so there will be no
6605 * dva/birth/checksum. The buffer must therefore remain anonymous
6608 if (!HDR_EMPTY(hdr)) {
6609 arc_buf_hdr_t *exists;
6610 kmutex_t *hash_lock;
6612 ASSERT3U(zio->io_error, ==, 0);
6614 arc_cksum_verify(buf);
6616 exists = buf_hash_insert(hdr, &hash_lock);
6617 if (exists != NULL) {
6619 * This can only happen if we overwrite for
6620 * sync-to-convergence, because we remove
6621 * buffers from the hash table when we arc_free().
6623 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6624 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6625 panic("bad overwrite, hdr=%p exists=%p",
6626 (void *)hdr, (void *)exists);
6627 ASSERT(refcount_is_zero(
6628 &exists->b_l1hdr.b_refcnt));
6629 arc_change_state(arc_anon, exists, hash_lock);
6630 mutex_exit(hash_lock);
6631 arc_hdr_destroy(exists);
6632 exists = buf_hash_insert(hdr, &hash_lock);
6633 ASSERT3P(exists, ==, NULL);
6634 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6636 ASSERT(zio->io_prop.zp_nopwrite);
6637 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6638 panic("bad nopwrite, hdr=%p exists=%p",
6639 (void *)hdr, (void *)exists);
6642 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6643 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6644 ASSERT(BP_GET_DEDUP(zio->io_bp));
6645 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6648 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6649 /* if it's not anon, we are doing a scrub */
6650 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6651 arc_access(hdr, hash_lock);
6652 mutex_exit(hash_lock);
6654 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6657 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6658 callback->awcb_done(zio, buf, callback->awcb_private);
6660 abd_put(zio->io_abd);
6661 kmem_free(callback, sizeof (arc_write_callback_t));
6665 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
6666 boolean_t l2arc, const zio_prop_t *zp, arc_write_done_func_t *ready,
6667 arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
6668 arc_write_done_func_t *done, void *private, zio_priority_t priority,
6669 int zio_flags, const zbookmark_phys_t *zb)
6671 arc_buf_hdr_t *hdr = buf->b_hdr;
6672 arc_write_callback_t *callback;
6674 zio_prop_t localprop = *zp;
6676 ASSERT3P(ready, !=, NULL);
6677 ASSERT3P(done, !=, NULL);
6678 ASSERT(!HDR_IO_ERROR(hdr));
6679 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6680 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6681 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6683 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6684 if (ARC_BUF_COMPRESSED(buf)) {
6686 * We're writing a pre-compressed buffer. Make the
6687 * compression algorithm requested by the zio_prop_t match
6688 * the pre-compressed buffer's compression algorithm.
6690 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6692 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6693 zio_flags |= ZIO_FLAG_RAW;
6695 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6696 callback->awcb_ready = ready;
6697 callback->awcb_children_ready = children_ready;
6698 callback->awcb_physdone = physdone;
6699 callback->awcb_done = done;
6700 callback->awcb_private = private;
6701 callback->awcb_buf = buf;
6704 * The hdr's b_pabd is now stale, free it now. A new data block
6705 * will be allocated when the zio pipeline calls arc_write_ready().
6707 if (hdr->b_l1hdr.b_pabd != NULL) {
6709 * If the buf is currently sharing the data block with
6710 * the hdr then we need to break that relationship here.
6711 * The hdr will remain with a NULL data pointer and the
6712 * buf will take sole ownership of the block.
6714 if (arc_buf_is_shared(buf)) {
6715 arc_unshare_buf(hdr, buf);
6717 arc_hdr_free_pabd(hdr);
6719 VERIFY3P(buf->b_data, !=, NULL);
6720 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6722 ASSERT(!arc_buf_is_shared(buf));
6723 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6725 zio = zio_write(pio, spa, txg, bp,
6726 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6727 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6728 (children_ready != NULL) ? arc_write_children_ready : NULL,
6729 arc_write_physdone, arc_write_done, callback,
6730 priority, zio_flags, zb);
6736 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
6739 uint64_t available_memory = ptob(freemem);
6741 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
6742 available_memory = MIN(available_memory, uma_avail());
6745 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
6748 if (txg > spa->spa_lowmem_last_txg) {
6749 spa->spa_lowmem_last_txg = txg;
6750 spa->spa_lowmem_page_load = 0;
6753 * If we are in pageout, we know that memory is already tight,
6754 * the arc is already going to be evicting, so we just want to
6755 * continue to let page writes occur as quickly as possible.
6757 if (curproc == pageproc) {
6758 if (spa->spa_lowmem_page_load >
6759 MAX(ptob(minfree), available_memory) / 4)
6760 return (SET_ERROR(ERESTART));
6761 /* Note: reserve is inflated, so we deflate */
6762 atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8);
6764 } else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) {
6765 /* memory is low, delay before restarting */
6766 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
6767 return (SET_ERROR(EAGAIN));
6769 spa->spa_lowmem_page_load = 0;
6770 #endif /* _KERNEL */
6775 arc_tempreserve_clear(uint64_t reserve)
6777 atomic_add_64(&arc_tempreserve, -reserve);
6778 ASSERT((int64_t)arc_tempreserve >= 0);
6782 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
6787 if (reserve > arc_c/4 && !arc_no_grow) {
6788 arc_c = MIN(arc_c_max, reserve * 4);
6789 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
6791 if (reserve > arc_c)
6792 return (SET_ERROR(ENOMEM));
6795 * Don't count loaned bufs as in flight dirty data to prevent long
6796 * network delays from blocking transactions that are ready to be
6797 * assigned to a txg.
6800 /* assert that it has not wrapped around */
6801 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6803 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
6804 arc_loaned_bytes), 0);
6807 * Writes will, almost always, require additional memory allocations
6808 * in order to compress/encrypt/etc the data. We therefore need to
6809 * make sure that there is sufficient available memory for this.
6811 error = arc_memory_throttle(spa, reserve, txg);
6816 * Throttle writes when the amount of dirty data in the cache
6817 * gets too large. We try to keep the cache less than half full
6818 * of dirty blocks so that our sync times don't grow too large.
6820 * In the case of one pool being built on another pool, we want
6821 * to make sure we don't end up throttling the lower (backing)
6822 * pool when the upper pool is the majority contributor to dirty
6823 * data. To insure we make forward progress during throttling, we
6824 * also check the current pool's net dirty data and only throttle
6825 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
6826 * data in the cache.
6828 * Note: if two requests come in concurrently, we might let them
6829 * both succeed, when one of them should fail. Not a huge deal.
6831 uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
6832 uint64_t spa_dirty_anon = spa_dirty_data(spa);
6834 if (total_dirty > arc_c * zfs_arc_dirty_limit_percent / 100 &&
6835 anon_size > arc_c * zfs_arc_anon_limit_percent / 100 &&
6836 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
6837 uint64_t meta_esize =
6838 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6839 uint64_t data_esize =
6840 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6841 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6842 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6843 arc_tempreserve >> 10, meta_esize >> 10,
6844 data_esize >> 10, reserve >> 10, arc_c >> 10);
6845 return (SET_ERROR(ERESTART));
6847 atomic_add_64(&arc_tempreserve, reserve);
6852 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6853 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6855 size->value.ui64 = refcount_count(&state->arcs_size);
6856 evict_data->value.ui64 =
6857 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6858 evict_metadata->value.ui64 =
6859 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6863 arc_kstat_update(kstat_t *ksp, int rw)
6865 arc_stats_t *as = ksp->ks_data;
6867 if (rw == KSTAT_WRITE) {
6870 arc_kstat_update_state(arc_anon,
6871 &as->arcstat_anon_size,
6872 &as->arcstat_anon_evictable_data,
6873 &as->arcstat_anon_evictable_metadata);
6874 arc_kstat_update_state(arc_mru,
6875 &as->arcstat_mru_size,
6876 &as->arcstat_mru_evictable_data,
6877 &as->arcstat_mru_evictable_metadata);
6878 arc_kstat_update_state(arc_mru_ghost,
6879 &as->arcstat_mru_ghost_size,
6880 &as->arcstat_mru_ghost_evictable_data,
6881 &as->arcstat_mru_ghost_evictable_metadata);
6882 arc_kstat_update_state(arc_mfu,
6883 &as->arcstat_mfu_size,
6884 &as->arcstat_mfu_evictable_data,
6885 &as->arcstat_mfu_evictable_metadata);
6886 arc_kstat_update_state(arc_mfu_ghost,
6887 &as->arcstat_mfu_ghost_size,
6888 &as->arcstat_mfu_ghost_evictable_data,
6889 &as->arcstat_mfu_ghost_evictable_metadata);
6891 ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
6892 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
6893 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
6894 ARCSTAT(arcstat_metadata_size) =
6895 aggsum_value(&astat_metadata_size);
6896 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
6897 ARCSTAT(arcstat_bonus_size) = aggsum_value(&astat_bonus_size);
6898 ARCSTAT(arcstat_dnode_size) = aggsum_value(&astat_dnode_size);
6899 ARCSTAT(arcstat_dbuf_size) = aggsum_value(&astat_dbuf_size);
6900 #if defined(__FreeBSD__) && defined(COMPAT_FREEBSD11)
6901 ARCSTAT(arcstat_other_size) = aggsum_value(&astat_bonus_size) +
6902 aggsum_value(&astat_dnode_size) +
6903 aggsum_value(&astat_dbuf_size);
6905 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
6912 * This function *must* return indices evenly distributed between all
6913 * sublists of the multilist. This is needed due to how the ARC eviction
6914 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6915 * distributed between all sublists and uses this assumption when
6916 * deciding which sublist to evict from and how much to evict from it.
6919 arc_state_multilist_index_func(multilist_t *ml, void *obj)
6921 arc_buf_hdr_t *hdr = obj;
6924 * We rely on b_dva to generate evenly distributed index
6925 * numbers using buf_hash below. So, as an added precaution,
6926 * let's make sure we never add empty buffers to the arc lists.
6928 ASSERT(!HDR_EMPTY(hdr));
6931 * The assumption here, is the hash value for a given
6932 * arc_buf_hdr_t will remain constant throughout it's lifetime
6933 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
6934 * Thus, we don't need to store the header's sublist index
6935 * on insertion, as this index can be recalculated on removal.
6937 * Also, the low order bits of the hash value are thought to be
6938 * distributed evenly. Otherwise, in the case that the multilist
6939 * has a power of two number of sublists, each sublists' usage
6940 * would not be evenly distributed.
6942 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
6943 multilist_get_num_sublists(ml));
6947 static eventhandler_tag arc_event_lowmem = NULL;
6950 arc_lowmem(void *arg __unused, int howto __unused)
6952 int64_t free_memory, to_free;
6954 arc_no_grow = B_TRUE;
6956 arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
6957 free_memory = arc_available_memory();
6958 to_free = (arc_c >> arc_shrink_shift) - MIN(free_memory, 0);
6959 DTRACE_PROBE2(arc__needfree, int64_t, free_memory, int64_t, to_free);
6960 arc_reduce_target_size(to_free);
6962 mutex_enter(&arc_adjust_lock);
6963 arc_adjust_needed = B_TRUE;
6964 zthr_wakeup(arc_adjust_zthr);
6967 * It is unsafe to block here in arbitrary threads, because we can come
6968 * here from ARC itself and may hold ARC locks and thus risk a deadlock
6969 * with ARC reclaim thread.
6971 if (curproc == pageproc)
6972 (void) cv_wait(&arc_adjust_waiters_cv, &arc_adjust_lock);
6973 mutex_exit(&arc_adjust_lock);
6978 arc_state_init(void)
6980 arc_anon = &ARC_anon;
6982 arc_mru_ghost = &ARC_mru_ghost;
6984 arc_mfu_ghost = &ARC_mfu_ghost;
6985 arc_l2c_only = &ARC_l2c_only;
6987 arc_mru->arcs_list[ARC_BUFC_METADATA] =
6988 multilist_create(sizeof (arc_buf_hdr_t),
6989 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6990 arc_state_multilist_index_func);
6991 arc_mru->arcs_list[ARC_BUFC_DATA] =
6992 multilist_create(sizeof (arc_buf_hdr_t),
6993 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6994 arc_state_multilist_index_func);
6995 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
6996 multilist_create(sizeof (arc_buf_hdr_t),
6997 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6998 arc_state_multilist_index_func);
6999 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
7000 multilist_create(sizeof (arc_buf_hdr_t),
7001 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7002 arc_state_multilist_index_func);
7003 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
7004 multilist_create(sizeof (arc_buf_hdr_t),
7005 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7006 arc_state_multilist_index_func);
7007 arc_mfu->arcs_list[ARC_BUFC_DATA] =
7008 multilist_create(sizeof (arc_buf_hdr_t),
7009 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7010 arc_state_multilist_index_func);
7011 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
7012 multilist_create(sizeof (arc_buf_hdr_t),
7013 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7014 arc_state_multilist_index_func);
7015 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
7016 multilist_create(sizeof (arc_buf_hdr_t),
7017 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7018 arc_state_multilist_index_func);
7019 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
7020 multilist_create(sizeof (arc_buf_hdr_t),
7021 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7022 arc_state_multilist_index_func);
7023 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
7024 multilist_create(sizeof (arc_buf_hdr_t),
7025 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7026 arc_state_multilist_index_func);
7028 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7029 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7030 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7031 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7032 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7033 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7034 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7035 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7036 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7037 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7038 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7039 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7041 refcount_create(&arc_anon->arcs_size);
7042 refcount_create(&arc_mru->arcs_size);
7043 refcount_create(&arc_mru_ghost->arcs_size);
7044 refcount_create(&arc_mfu->arcs_size);
7045 refcount_create(&arc_mfu_ghost->arcs_size);
7046 refcount_create(&arc_l2c_only->arcs_size);
7048 aggsum_init(&arc_meta_used, 0);
7049 aggsum_init(&arc_size, 0);
7050 aggsum_init(&astat_data_size, 0);
7051 aggsum_init(&astat_metadata_size, 0);
7052 aggsum_init(&astat_hdr_size, 0);
7053 aggsum_init(&astat_bonus_size, 0);
7054 aggsum_init(&astat_dnode_size, 0);
7055 aggsum_init(&astat_dbuf_size, 0);
7056 aggsum_init(&astat_l2_hdr_size, 0);
7060 arc_state_fini(void)
7062 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7063 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7064 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7065 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7066 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7067 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7068 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7069 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7070 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7071 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7072 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7073 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7075 refcount_destroy(&arc_anon->arcs_size);
7076 refcount_destroy(&arc_mru->arcs_size);
7077 refcount_destroy(&arc_mru_ghost->arcs_size);
7078 refcount_destroy(&arc_mfu->arcs_size);
7079 refcount_destroy(&arc_mfu_ghost->arcs_size);
7080 refcount_destroy(&arc_l2c_only->arcs_size);
7082 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
7083 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
7084 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
7085 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
7086 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
7087 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
7088 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
7089 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7091 aggsum_fini(&arc_meta_used);
7092 aggsum_fini(&arc_size);
7093 aggsum_fini(&astat_data_size);
7094 aggsum_fini(&astat_metadata_size);
7095 aggsum_fini(&astat_hdr_size);
7096 aggsum_fini(&astat_bonus_size);
7097 aggsum_fini(&astat_dnode_size);
7098 aggsum_fini(&astat_dbuf_size);
7099 aggsum_fini(&astat_l2_hdr_size);
7111 int i, prefetch_tunable_set = 0;
7114 * allmem is "all memory that we could possibly use".
7118 uint64_t allmem = ptob(physmem - swapfs_minfree);
7120 uint64_t allmem = (physmem * PAGESIZE) / 2;
7123 uint64_t allmem = kmem_size();
7125 mutex_init(&arc_adjust_lock, NULL, MUTEX_DEFAULT, NULL);
7126 cv_init(&arc_adjust_waiters_cv, NULL, CV_DEFAULT, NULL);
7128 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
7129 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL);
7131 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
7132 arc_c_min = MAX(allmem / 32, arc_abs_min);
7133 /* set max to 5/8 of all memory, or all but 1GB, whichever is more */
7134 if (allmem >= 1 << 30)
7135 arc_c_max = allmem - (1 << 30);
7137 arc_c_max = arc_c_min;
7138 arc_c_max = MAX(allmem * 5 / 8, arc_c_max);
7141 * In userland, there's only the memory pressure that we artificially
7142 * create (see arc_available_memory()). Don't let arc_c get too
7143 * small, because it can cause transactions to be larger than
7144 * arc_c, causing arc_tempreserve_space() to fail.
7147 arc_c_min = arc_c_max / 2;
7152 * Allow the tunables to override our calculations if they are
7155 if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) {
7156 arc_c_max = zfs_arc_max;
7157 arc_c_min = MIN(arc_c_min, arc_c_max);
7159 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
7160 arc_c_min = zfs_arc_min;
7164 arc_p = (arc_c >> 1);
7166 /* limit meta-data to 1/4 of the arc capacity */
7167 arc_meta_limit = arc_c_max / 4;
7171 * Metadata is stored in the kernel's heap. Don't let us
7172 * use more than half the heap for the ARC.
7175 arc_meta_limit = MIN(arc_meta_limit, uma_limit() / 2);
7176 arc_dnode_limit = arc_meta_limit / 10;
7178 arc_meta_limit = MIN(arc_meta_limit,
7179 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
7183 /* Allow the tunable to override if it is reasonable */
7184 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
7185 arc_meta_limit = zfs_arc_meta_limit;
7187 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
7188 arc_c_min = arc_meta_limit / 2;
7190 if (zfs_arc_meta_min > 0) {
7191 arc_meta_min = zfs_arc_meta_min;
7193 arc_meta_min = arc_c_min / 2;
7196 /* Valid range: <arc_meta_min> - <arc_c_max> */
7197 if ((zfs_arc_dnode_limit) && (zfs_arc_dnode_limit != arc_dnode_limit) &&
7198 (zfs_arc_dnode_limit >= zfs_arc_meta_min) &&
7199 (zfs_arc_dnode_limit <= arc_c_max))
7200 arc_dnode_limit = zfs_arc_dnode_limit;
7202 if (zfs_arc_grow_retry > 0)
7203 arc_grow_retry = zfs_arc_grow_retry;
7205 if (zfs_arc_shrink_shift > 0)
7206 arc_shrink_shift = zfs_arc_shrink_shift;
7208 if (zfs_arc_no_grow_shift > 0)
7209 arc_no_grow_shift = zfs_arc_no_grow_shift;
7211 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
7213 if (arc_no_grow_shift >= arc_shrink_shift)
7214 arc_no_grow_shift = arc_shrink_shift - 1;
7216 if (zfs_arc_p_min_shift > 0)
7217 arc_p_min_shift = zfs_arc_p_min_shift;
7219 /* if kmem_flags are set, lets try to use less memory */
7220 if (kmem_debugging())
7222 if (arc_c < arc_c_min)
7225 zfs_arc_min = arc_c_min;
7226 zfs_arc_max = arc_c_max;
7231 * The arc must be "uninitialized", so that hdr_recl() (which is
7232 * registered by buf_init()) will not access arc_reap_zthr before
7235 ASSERT(!arc_initialized);
7238 list_create(&arc_prune_list, sizeof (arc_prune_t),
7239 offsetof(arc_prune_t, p_node));
7240 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
7242 arc_prune_taskq = taskq_create("arc_prune", max_ncpus, minclsyspri,
7243 max_ncpus, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
7245 arc_dnlc_evicts_thread_exit = FALSE;
7247 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
7248 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
7250 if (arc_ksp != NULL) {
7251 arc_ksp->ks_data = &arc_stats;
7252 arc_ksp->ks_update = arc_kstat_update;
7253 kstat_install(arc_ksp);
7256 arc_adjust_zthr = zthr_create_timer(arc_adjust_cb_check,
7257 arc_adjust_cb, NULL, SEC2NSEC(1));
7258 arc_reap_zthr = zthr_create_timer(arc_reap_cb_check,
7259 arc_reap_cb, NULL, SEC2NSEC(1));
7262 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
7263 EVENTHANDLER_PRI_FIRST);
7266 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0,
7267 TS_RUN, minclsyspri);
7269 arc_initialized = B_TRUE;
7273 * Calculate maximum amount of dirty data per pool.
7275 * If it has been set by /etc/system, take that.
7276 * Otherwise, use a percentage of physical memory defined by
7277 * zfs_dirty_data_max_percent (default 10%) with a cap at
7278 * zfs_dirty_data_max_max (default 4GB).
7280 if (zfs_dirty_data_max == 0) {
7281 zfs_dirty_data_max = ptob(physmem) *
7282 zfs_dirty_data_max_percent / 100;
7283 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
7284 zfs_dirty_data_max_max);
7288 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
7289 prefetch_tunable_set = 1;
7292 if (prefetch_tunable_set == 0) {
7293 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
7295 printf(" add \"vfs.zfs.prefetch_disable=0\" "
7296 "to /boot/loader.conf.\n");
7297 zfs_prefetch_disable = 1;
7300 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
7301 prefetch_tunable_set == 0) {
7302 printf("ZFS NOTICE: Prefetch is disabled by default if less "
7303 "than 4GB of RAM is present;\n"
7304 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
7305 "to /boot/loader.conf.\n");
7306 zfs_prefetch_disable = 1;
7309 /* Warn about ZFS memory and address space requirements. */
7310 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
7311 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
7312 "expect unstable behavior.\n");
7314 if (allmem < 512 * (1 << 20)) {
7315 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
7316 "expect unstable behavior.\n");
7317 printf(" Consider tuning vm.kmem_size and "
7318 "vm.kmem_size_max\n");
7319 printf(" in /boot/loader.conf.\n");
7330 if (arc_event_lowmem != NULL)
7331 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
7334 /* Use B_TRUE to ensure *all* buffers are evicted */
7335 arc_flush(NULL, B_TRUE);
7337 mutex_enter(&arc_dnlc_evicts_lock);
7338 arc_dnlc_evicts_thread_exit = TRUE;
7340 * The user evicts thread will set arc_user_evicts_thread_exit
7341 * to FALSE when it is finished exiting; we're waiting for that.
7343 while (arc_dnlc_evicts_thread_exit) {
7344 cv_signal(&arc_dnlc_evicts_cv);
7345 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
7347 mutex_exit(&arc_dnlc_evicts_lock);
7349 arc_initialized = B_FALSE;
7351 if (arc_ksp != NULL) {
7352 kstat_delete(arc_ksp);
7356 taskq_wait(arc_prune_taskq);
7357 taskq_destroy(arc_prune_taskq);
7359 mutex_enter(&arc_prune_mtx);
7360 while ((p = list_head(&arc_prune_list)) != NULL) {
7361 list_remove(&arc_prune_list, p);
7362 refcount_remove(&p->p_refcnt, &arc_prune_list);
7363 refcount_destroy(&p->p_refcnt);
7364 kmem_free(p, sizeof (*p));
7366 mutex_exit(&arc_prune_mtx);
7368 list_destroy(&arc_prune_list);
7369 mutex_destroy(&arc_prune_mtx);
7371 (void) zthr_cancel(arc_adjust_zthr);
7372 zthr_destroy(arc_adjust_zthr);
7374 mutex_destroy(&arc_dnlc_evicts_lock);
7375 cv_destroy(&arc_dnlc_evicts_cv);
7377 (void) zthr_cancel(arc_reap_zthr);
7378 zthr_destroy(arc_reap_zthr);
7380 mutex_destroy(&arc_adjust_lock);
7381 cv_destroy(&arc_adjust_waiters_cv);
7384 * buf_fini() must proceed arc_state_fini() because buf_fin() may
7385 * trigger the release of kmem magazines, which can callback to
7386 * arc_space_return() which accesses aggsums freed in act_state_fini().
7391 ASSERT0(arc_loaned_bytes);
7397 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7398 * It uses dedicated storage devices to hold cached data, which are populated
7399 * using large infrequent writes. The main role of this cache is to boost
7400 * the performance of random read workloads. The intended L2ARC devices
7401 * include short-stroked disks, solid state disks, and other media with
7402 * substantially faster read latency than disk.
7404 * +-----------------------+
7406 * +-----------------------+
7409 * l2arc_feed_thread() arc_read()
7413 * +---------------+ |
7415 * +---------------+ |
7420 * +-------+ +-------+
7422 * | cache | | cache |
7423 * +-------+ +-------+
7424 * +=========+ .-----.
7425 * : L2ARC : |-_____-|
7426 * : devices : | Disks |
7427 * +=========+ `-_____-'
7429 * Read requests are satisfied from the following sources, in order:
7432 * 2) vdev cache of L2ARC devices
7434 * 4) vdev cache of disks
7437 * Some L2ARC device types exhibit extremely slow write performance.
7438 * To accommodate for this there are some significant differences between
7439 * the L2ARC and traditional cache design:
7441 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
7442 * the ARC behave as usual, freeing buffers and placing headers on ghost
7443 * lists. The ARC does not send buffers to the L2ARC during eviction as
7444 * this would add inflated write latencies for all ARC memory pressure.
7446 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7447 * It does this by periodically scanning buffers from the eviction-end of
7448 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7449 * not already there. It scans until a headroom of buffers is satisfied,
7450 * which itself is a buffer for ARC eviction. If a compressible buffer is
7451 * found during scanning and selected for writing to an L2ARC device, we
7452 * temporarily boost scanning headroom during the next scan cycle to make
7453 * sure we adapt to compression effects (which might significantly reduce
7454 * the data volume we write to L2ARC). The thread that does this is
7455 * l2arc_feed_thread(), illustrated below; example sizes are included to
7456 * provide a better sense of ratio than this diagram:
7459 * +---------------------+----------+
7460 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
7461 * +---------------------+----------+ | o L2ARC eligible
7462 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
7463 * +---------------------+----------+ |
7464 * 15.9 Gbytes ^ 32 Mbytes |
7466 * l2arc_feed_thread()
7468 * l2arc write hand <--[oooo]--'
7472 * +==============================+
7473 * L2ARC dev |####|#|###|###| |####| ... |
7474 * +==============================+
7477 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
7478 * evicted, then the L2ARC has cached a buffer much sooner than it probably
7479 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
7480 * safe to say that this is an uncommon case, since buffers at the end of
7481 * the ARC lists have moved there due to inactivity.
7483 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
7484 * then the L2ARC simply misses copying some buffers. This serves as a
7485 * pressure valve to prevent heavy read workloads from both stalling the ARC
7486 * with waits and clogging the L2ARC with writes. This also helps prevent
7487 * the potential for the L2ARC to churn if it attempts to cache content too
7488 * quickly, such as during backups of the entire pool.
7490 * 5. After system boot and before the ARC has filled main memory, there are
7491 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
7492 * lists can remain mostly static. Instead of searching from tail of these
7493 * lists as pictured, the l2arc_feed_thread() will search from the list heads
7494 * for eligible buffers, greatly increasing its chance of finding them.
7496 * The L2ARC device write speed is also boosted during this time so that
7497 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
7498 * there are no L2ARC reads, and no fear of degrading read performance
7499 * through increased writes.
7501 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
7502 * the vdev queue can aggregate them into larger and fewer writes. Each
7503 * device is written to in a rotor fashion, sweeping writes through
7504 * available space then repeating.
7506 * 7. The L2ARC does not store dirty content. It never needs to flush
7507 * write buffers back to disk based storage.
7509 * 8. If an ARC buffer is written (and dirtied) which also exists in the
7510 * L2ARC, the now stale L2ARC buffer is immediately dropped.
7512 * The performance of the L2ARC can be tweaked by a number of tunables, which
7513 * may be necessary for different workloads:
7515 * l2arc_write_max max write bytes per interval
7516 * l2arc_write_boost extra write bytes during device warmup
7517 * l2arc_noprefetch skip caching prefetched buffers
7518 * l2arc_headroom number of max device writes to precache
7519 * l2arc_headroom_boost when we find compressed buffers during ARC
7520 * scanning, we multiply headroom by this
7521 * percentage factor for the next scan cycle,
7522 * since more compressed buffers are likely to
7524 * l2arc_feed_secs seconds between L2ARC writing
7526 * Tunables may be removed or added as future performance improvements are
7527 * integrated, and also may become zpool properties.
7529 * There are three key functions that control how the L2ARC warms up:
7531 * l2arc_write_eligible() check if a buffer is eligible to cache
7532 * l2arc_write_size() calculate how much to write
7533 * l2arc_write_interval() calculate sleep delay between writes
7535 * These three functions determine what to write, how much, and how quickly
7540 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
7543 * A buffer is *not* eligible for the L2ARC if it:
7544 * 1. belongs to a different spa.
7545 * 2. is already cached on the L2ARC.
7546 * 3. has an I/O in progress (it may be an incomplete read).
7547 * 4. is flagged not eligible (zfs property).
7549 if (hdr->b_spa != spa_guid) {
7550 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
7553 if (HDR_HAS_L2HDR(hdr)) {
7554 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
7557 if (HDR_IO_IN_PROGRESS(hdr)) {
7558 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
7561 if (!HDR_L2CACHE(hdr)) {
7562 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
7570 l2arc_write_size(void)
7575 * Make sure our globals have meaningful values in case the user
7578 size = l2arc_write_max;
7580 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
7581 "be greater than zero, resetting it to the default (%d)",
7583 size = l2arc_write_max = L2ARC_WRITE_SIZE;
7586 if (arc_warm == B_FALSE)
7587 size += l2arc_write_boost;
7594 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
7596 clock_t interval, next, now;
7599 * If the ARC lists are busy, increase our write rate; if the
7600 * lists are stale, idle back. This is achieved by checking
7601 * how much we previously wrote - if it was more than half of
7602 * what we wanted, schedule the next write much sooner.
7604 if (l2arc_feed_again && wrote > (wanted / 2))
7605 interval = (hz * l2arc_feed_min_ms) / 1000;
7607 interval = hz * l2arc_feed_secs;
7609 now = ddi_get_lbolt();
7610 next = MAX(now, MIN(now + interval, began + interval));
7616 * Cycle through L2ARC devices. This is how L2ARC load balances.
7617 * If a device is returned, this also returns holding the spa config lock.
7619 static l2arc_dev_t *
7620 l2arc_dev_get_next(void)
7622 l2arc_dev_t *first, *next = NULL;
7625 * Lock out the removal of spas (spa_namespace_lock), then removal
7626 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
7627 * both locks will be dropped and a spa config lock held instead.
7629 mutex_enter(&spa_namespace_lock);
7630 mutex_enter(&l2arc_dev_mtx);
7632 /* if there are no vdevs, there is nothing to do */
7633 if (l2arc_ndev == 0)
7637 next = l2arc_dev_last;
7639 /* loop around the list looking for a non-faulted vdev */
7641 next = list_head(l2arc_dev_list);
7643 next = list_next(l2arc_dev_list, next);
7645 next = list_head(l2arc_dev_list);
7648 /* if we have come back to the start, bail out */
7651 else if (next == first)
7654 } while (vdev_is_dead(next->l2ad_vdev));
7656 /* if we were unable to find any usable vdevs, return NULL */
7657 if (vdev_is_dead(next->l2ad_vdev))
7660 l2arc_dev_last = next;
7663 mutex_exit(&l2arc_dev_mtx);
7666 * Grab the config lock to prevent the 'next' device from being
7667 * removed while we are writing to it.
7670 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
7671 mutex_exit(&spa_namespace_lock);
7677 * Free buffers that were tagged for destruction.
7680 l2arc_do_free_on_write()
7683 l2arc_data_free_t *df, *df_prev;
7685 mutex_enter(&l2arc_free_on_write_mtx);
7686 buflist = l2arc_free_on_write;
7688 for (df = list_tail(buflist); df; df = df_prev) {
7689 df_prev = list_prev(buflist, df);
7690 ASSERT3P(df->l2df_abd, !=, NULL);
7691 abd_free(df->l2df_abd);
7692 list_remove(buflist, df);
7693 kmem_free(df, sizeof (l2arc_data_free_t));
7696 mutex_exit(&l2arc_free_on_write_mtx);
7700 * A write to a cache device has completed. Update all headers to allow
7701 * reads from these buffers to begin.
7704 l2arc_write_done(zio_t *zio)
7706 l2arc_write_callback_t *cb;
7709 arc_buf_hdr_t *head, *hdr, *hdr_prev;
7710 kmutex_t *hash_lock;
7711 int64_t bytes_dropped = 0;
7713 cb = zio->io_private;
7714 ASSERT3P(cb, !=, NULL);
7715 dev = cb->l2wcb_dev;
7716 ASSERT3P(dev, !=, NULL);
7717 head = cb->l2wcb_head;
7718 ASSERT3P(head, !=, NULL);
7719 buflist = &dev->l2ad_buflist;
7720 ASSERT3P(buflist, !=, NULL);
7721 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
7722 l2arc_write_callback_t *, cb);
7724 if (zio->io_error != 0)
7725 ARCSTAT_BUMP(arcstat_l2_writes_error);
7728 * All writes completed, or an error was hit.
7731 mutex_enter(&dev->l2ad_mtx);
7732 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
7733 hdr_prev = list_prev(buflist, hdr);
7735 hash_lock = HDR_LOCK(hdr);
7738 * We cannot use mutex_enter or else we can deadlock
7739 * with l2arc_write_buffers (due to swapping the order
7740 * the hash lock and l2ad_mtx are taken).
7742 if (!mutex_tryenter(hash_lock)) {
7744 * Missed the hash lock. We must retry so we
7745 * don't leave the ARC_FLAG_L2_WRITING bit set.
7747 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
7750 * We don't want to rescan the headers we've
7751 * already marked as having been written out, so
7752 * we reinsert the head node so we can pick up
7753 * where we left off.
7755 list_remove(buflist, head);
7756 list_insert_after(buflist, hdr, head);
7758 mutex_exit(&dev->l2ad_mtx);
7761 * We wait for the hash lock to become available
7762 * to try and prevent busy waiting, and increase
7763 * the chance we'll be able to acquire the lock
7764 * the next time around.
7766 mutex_enter(hash_lock);
7767 mutex_exit(hash_lock);
7772 * We could not have been moved into the arc_l2c_only
7773 * state while in-flight due to our ARC_FLAG_L2_WRITING
7774 * bit being set. Let's just ensure that's being enforced.
7776 ASSERT(HDR_HAS_L1HDR(hdr));
7778 if (zio->io_error != 0) {
7780 * Error - drop L2ARC entry.
7782 list_remove(buflist, hdr);
7784 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
7786 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr));
7787 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
7789 bytes_dropped += arc_hdr_size(hdr);
7790 (void) refcount_remove_many(&dev->l2ad_alloc,
7791 arc_hdr_size(hdr), hdr);
7795 * Allow ARC to begin reads and ghost list evictions to
7798 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
7800 mutex_exit(hash_lock);
7803 atomic_inc_64(&l2arc_writes_done);
7804 list_remove(buflist, head);
7805 ASSERT(!HDR_HAS_L1HDR(head));
7806 kmem_cache_free(hdr_l2only_cache, head);
7807 mutex_exit(&dev->l2ad_mtx);
7809 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
7811 l2arc_do_free_on_write();
7813 kmem_free(cb, sizeof (l2arc_write_callback_t));
7817 * A read to a cache device completed. Validate buffer contents before
7818 * handing over to the regular ARC routines.
7821 l2arc_read_done(zio_t *zio)
7823 l2arc_read_callback_t *cb;
7825 kmutex_t *hash_lock;
7826 boolean_t valid_cksum;
7828 ASSERT3P(zio->io_vd, !=, NULL);
7829 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
7831 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
7833 cb = zio->io_private;
7834 ASSERT3P(cb, !=, NULL);
7835 hdr = cb->l2rcb_hdr;
7836 ASSERT3P(hdr, !=, NULL);
7838 hash_lock = HDR_LOCK(hdr);
7839 mutex_enter(hash_lock);
7840 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
7843 * If the data was read into a temporary buffer,
7844 * move it and free the buffer.
7846 if (cb->l2rcb_abd != NULL) {
7847 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
7848 if (zio->io_error == 0) {
7849 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd,
7854 * The following must be done regardless of whether
7855 * there was an error:
7856 * - free the temporary buffer
7857 * - point zio to the real ARC buffer
7858 * - set zio size accordingly
7859 * These are required because zio is either re-used for
7860 * an I/O of the block in the case of the error
7861 * or the zio is passed to arc_read_done() and it
7864 abd_free(cb->l2rcb_abd);
7865 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
7866 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
7869 ASSERT3P(zio->io_abd, !=, NULL);
7872 * Check this survived the L2ARC journey.
7874 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd);
7875 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
7876 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
7878 valid_cksum = arc_cksum_is_equal(hdr, zio);
7879 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
7880 mutex_exit(hash_lock);
7881 zio->io_private = hdr;
7884 mutex_exit(hash_lock);
7886 * Buffer didn't survive caching. Increment stats and
7887 * reissue to the original storage device.
7889 if (zio->io_error != 0) {
7890 ARCSTAT_BUMP(arcstat_l2_io_error);
7892 zio->io_error = SET_ERROR(EIO);
7895 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
7898 * If there's no waiter, issue an async i/o to the primary
7899 * storage now. If there *is* a waiter, the caller must
7900 * issue the i/o in a context where it's OK to block.
7902 if (zio->io_waiter == NULL) {
7903 zio_t *pio = zio_unique_parent(zio);
7905 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
7907 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
7908 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done,
7909 hdr, zio->io_priority, cb->l2rcb_flags,
7914 kmem_free(cb, sizeof (l2arc_read_callback_t));
7918 * This is the list priority from which the L2ARC will search for pages to
7919 * cache. This is used within loops (0..3) to cycle through lists in the
7920 * desired order. This order can have a significant effect on cache
7923 * Currently the metadata lists are hit first, MFU then MRU, followed by
7924 * the data lists. This function returns a locked list, and also returns
7927 static multilist_sublist_t *
7928 l2arc_sublist_lock(int list_num)
7930 multilist_t *ml = NULL;
7933 ASSERT(list_num >= 0 && list_num <= 3);
7937 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
7940 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
7943 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
7946 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
7951 * Return a randomly-selected sublist. This is acceptable
7952 * because the caller feeds only a little bit of data for each
7953 * call (8MB). Subsequent calls will result in different
7954 * sublists being selected.
7956 idx = multilist_get_random_index(ml);
7957 return (multilist_sublist_lock(ml, idx));
7961 * Evict buffers from the device write hand to the distance specified in
7962 * bytes. This distance may span populated buffers, it may span nothing.
7963 * This is clearing a region on the L2ARC device ready for writing.
7964 * If the 'all' boolean is set, every buffer is evicted.
7967 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
7970 arc_buf_hdr_t *hdr, *hdr_prev;
7971 kmutex_t *hash_lock;
7974 buflist = &dev->l2ad_buflist;
7976 if (!all && dev->l2ad_first) {
7978 * This is the first sweep through the device. There is
7984 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
7986 * When nearing the end of the device, evict to the end
7987 * before the device write hand jumps to the start.
7989 taddr = dev->l2ad_end;
7991 taddr = dev->l2ad_hand + distance;
7993 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
7994 uint64_t, taddr, boolean_t, all);
7997 mutex_enter(&dev->l2ad_mtx);
7998 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
7999 hdr_prev = list_prev(buflist, hdr);
8001 hash_lock = HDR_LOCK(hdr);
8004 * We cannot use mutex_enter or else we can deadlock
8005 * with l2arc_write_buffers (due to swapping the order
8006 * the hash lock and l2ad_mtx are taken).
8008 if (!mutex_tryenter(hash_lock)) {
8010 * Missed the hash lock. Retry.
8012 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
8013 mutex_exit(&dev->l2ad_mtx);
8014 mutex_enter(hash_lock);
8015 mutex_exit(hash_lock);
8020 * A header can't be on this list if it doesn't have L2 header.
8022 ASSERT(HDR_HAS_L2HDR(hdr));
8024 /* Ensure this header has finished being written. */
8025 ASSERT(!HDR_L2_WRITING(hdr));
8026 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
8028 if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
8029 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
8031 * We've evicted to the target address,
8032 * or the end of the device.
8034 mutex_exit(hash_lock);
8038 if (!HDR_HAS_L1HDR(hdr)) {
8039 ASSERT(!HDR_L2_READING(hdr));
8041 * This doesn't exist in the ARC. Destroy.
8042 * arc_hdr_destroy() will call list_remove()
8043 * and decrement arcstat_l2_lsize.
8045 arc_change_state(arc_anon, hdr, hash_lock);
8046 arc_hdr_destroy(hdr);
8048 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
8049 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
8051 * Invalidate issued or about to be issued
8052 * reads, since we may be about to write
8053 * over this location.
8055 if (HDR_L2_READING(hdr)) {
8056 ARCSTAT_BUMP(arcstat_l2_evict_reading);
8057 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
8060 arc_hdr_l2hdr_destroy(hdr);
8062 mutex_exit(hash_lock);
8064 mutex_exit(&dev->l2ad_mtx);
8068 * Find and write ARC buffers to the L2ARC device.
8070 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
8071 * for reading until they have completed writing.
8072 * The headroom_boost is an in-out parameter used to maintain headroom boost
8073 * state between calls to this function.
8075 * Returns the number of bytes actually written (which may be smaller than
8076 * the delta by which the device hand has changed due to alignment).
8079 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
8081 arc_buf_hdr_t *hdr, *hdr_prev, *head;
8082 uint64_t write_asize, write_psize, write_lsize, headroom;
8084 l2arc_write_callback_t *cb;
8086 uint64_t guid = spa_load_guid(spa);
8089 ASSERT3P(dev->l2ad_vdev, !=, NULL);
8092 write_lsize = write_asize = write_psize = 0;
8094 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
8095 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
8097 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
8099 * Copy buffers for L2ARC writing.
8101 for (try = 0; try <= 3; try++) {
8102 multilist_sublist_t *mls = l2arc_sublist_lock(try);
8103 uint64_t passed_sz = 0;
8105 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
8108 * L2ARC fast warmup.
8110 * Until the ARC is warm and starts to evict, read from the
8111 * head of the ARC lists rather than the tail.
8113 if (arc_warm == B_FALSE)
8114 hdr = multilist_sublist_head(mls);
8116 hdr = multilist_sublist_tail(mls);
8118 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
8120 headroom = target_sz * l2arc_headroom;
8121 if (zfs_compressed_arc_enabled)
8122 headroom = (headroom * l2arc_headroom_boost) / 100;
8124 for (; hdr; hdr = hdr_prev) {
8125 kmutex_t *hash_lock;
8127 if (arc_warm == B_FALSE)
8128 hdr_prev = multilist_sublist_next(mls, hdr);
8130 hdr_prev = multilist_sublist_prev(mls, hdr);
8131 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned,
8132 HDR_GET_LSIZE(hdr));
8134 hash_lock = HDR_LOCK(hdr);
8135 if (!mutex_tryenter(hash_lock)) {
8136 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
8138 * Skip this buffer rather than waiting.
8143 passed_sz += HDR_GET_LSIZE(hdr);
8144 if (passed_sz > headroom) {
8148 mutex_exit(hash_lock);
8149 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
8153 if (!l2arc_write_eligible(guid, hdr)) {
8154 mutex_exit(hash_lock);
8159 * We rely on the L1 portion of the header below, so
8160 * it's invalid for this header to have been evicted out
8161 * of the ghost cache, prior to being written out. The
8162 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8164 ASSERT(HDR_HAS_L1HDR(hdr));
8166 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8167 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8168 ASSERT3U(arc_hdr_size(hdr), >, 0);
8169 uint64_t psize = arc_hdr_size(hdr);
8170 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
8173 if ((write_asize + asize) > target_sz) {
8175 mutex_exit(hash_lock);
8176 ARCSTAT_BUMP(arcstat_l2_write_full);
8182 * Insert a dummy header on the buflist so
8183 * l2arc_write_done() can find where the
8184 * write buffers begin without searching.
8186 mutex_enter(&dev->l2ad_mtx);
8187 list_insert_head(&dev->l2ad_buflist, head);
8188 mutex_exit(&dev->l2ad_mtx);
8191 sizeof (l2arc_write_callback_t), KM_SLEEP);
8192 cb->l2wcb_dev = dev;
8193 cb->l2wcb_head = head;
8194 pio = zio_root(spa, l2arc_write_done, cb,
8196 ARCSTAT_BUMP(arcstat_l2_write_pios);
8199 hdr->b_l2hdr.b_dev = dev;
8200 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
8201 arc_hdr_set_flags(hdr,
8202 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
8204 mutex_enter(&dev->l2ad_mtx);
8205 list_insert_head(&dev->l2ad_buflist, hdr);
8206 mutex_exit(&dev->l2ad_mtx);
8208 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr);
8211 * Normally the L2ARC can use the hdr's data, but if
8212 * we're sharing data between the hdr and one of its
8213 * bufs, L2ARC needs its own copy of the data so that
8214 * the ZIO below can't race with the buf consumer.
8215 * Another case where we need to create a copy of the
8216 * data is when the buffer size is not device-aligned
8217 * and we need to pad the block to make it such.
8218 * That also keeps the clock hand suitably aligned.
8220 * To ensure that the copy will be available for the
8221 * lifetime of the ZIO and be cleaned up afterwards, we
8222 * add it to the l2arc_free_on_write queue.
8225 if (!HDR_SHARED_DATA(hdr) && psize == asize) {
8226 to_write = hdr->b_l1hdr.b_pabd;
8228 to_write = abd_alloc_for_io(asize,
8229 HDR_ISTYPE_METADATA(hdr));
8230 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
8231 if (asize != psize) {
8232 abd_zero_off(to_write, psize,
8235 l2arc_free_abd_on_write(to_write, asize,
8238 wzio = zio_write_phys(pio, dev->l2ad_vdev,
8239 hdr->b_l2hdr.b_daddr, asize, to_write,
8240 ZIO_CHECKSUM_OFF, NULL, hdr,
8241 ZIO_PRIORITY_ASYNC_WRITE,
8242 ZIO_FLAG_CANFAIL, B_FALSE);
8244 write_lsize += HDR_GET_LSIZE(hdr);
8245 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
8248 write_psize += psize;
8249 write_asize += asize;
8250 dev->l2ad_hand += asize;
8252 mutex_exit(hash_lock);
8254 (void) zio_nowait(wzio);
8257 multilist_sublist_unlock(mls);
8263 /* No buffers selected for writing? */
8265 ASSERT0(write_lsize);
8266 ASSERT(!HDR_HAS_L1HDR(head));
8267 kmem_cache_free(hdr_l2only_cache, head);
8271 ASSERT3U(write_psize, <=, target_sz);
8272 ARCSTAT_BUMP(arcstat_l2_writes_sent);
8273 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
8274 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
8275 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
8276 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
8279 * Bump device hand to the device start if it is approaching the end.
8280 * l2arc_evict() will already have evicted ahead for this case.
8282 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
8283 dev->l2ad_hand = dev->l2ad_start;
8284 dev->l2ad_first = B_FALSE;
8287 dev->l2ad_writing = B_TRUE;
8288 (void) zio_wait(pio);
8289 dev->l2ad_writing = B_FALSE;
8291 return (write_asize);
8295 * This thread feeds the L2ARC at regular intervals. This is the beating
8296 * heart of the L2ARC.
8300 l2arc_feed_thread(void *unused __unused)
8305 uint64_t size, wrote;
8306 clock_t begin, next = ddi_get_lbolt();
8308 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
8310 mutex_enter(&l2arc_feed_thr_lock);
8312 while (l2arc_thread_exit == 0) {
8313 CALLB_CPR_SAFE_BEGIN(&cpr);
8314 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
8315 next - ddi_get_lbolt());
8316 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
8317 next = ddi_get_lbolt() + hz;
8320 * Quick check for L2ARC devices.
8322 mutex_enter(&l2arc_dev_mtx);
8323 if (l2arc_ndev == 0) {
8324 mutex_exit(&l2arc_dev_mtx);
8327 mutex_exit(&l2arc_dev_mtx);
8328 begin = ddi_get_lbolt();
8331 * This selects the next l2arc device to write to, and in
8332 * doing so the next spa to feed from: dev->l2ad_spa. This
8333 * will return NULL if there are now no l2arc devices or if
8334 * they are all faulted.
8336 * If a device is returned, its spa's config lock is also
8337 * held to prevent device removal. l2arc_dev_get_next()
8338 * will grab and release l2arc_dev_mtx.
8340 if ((dev = l2arc_dev_get_next()) == NULL)
8343 spa = dev->l2ad_spa;
8344 ASSERT3P(spa, !=, NULL);
8347 * If the pool is read-only then force the feed thread to
8348 * sleep a little longer.
8350 if (!spa_writeable(spa)) {
8351 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
8352 spa_config_exit(spa, SCL_L2ARC, dev);
8357 * Avoid contributing to memory pressure.
8359 if (arc_reclaim_needed()) {
8360 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
8361 spa_config_exit(spa, SCL_L2ARC, dev);
8365 ARCSTAT_BUMP(arcstat_l2_feeds);
8367 size = l2arc_write_size();
8370 * Evict L2ARC buffers that will be overwritten.
8372 l2arc_evict(dev, size, B_FALSE);
8375 * Write ARC buffers.
8377 wrote = l2arc_write_buffers(spa, dev, size);
8380 * Calculate interval between writes.
8382 next = l2arc_write_interval(begin, size, wrote);
8383 spa_config_exit(spa, SCL_L2ARC, dev);
8386 l2arc_thread_exit = 0;
8387 cv_broadcast(&l2arc_feed_thr_cv);
8388 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
8393 l2arc_vdev_present(vdev_t *vd)
8397 mutex_enter(&l2arc_dev_mtx);
8398 for (dev = list_head(l2arc_dev_list); dev != NULL;
8399 dev = list_next(l2arc_dev_list, dev)) {
8400 if (dev->l2ad_vdev == vd)
8403 mutex_exit(&l2arc_dev_mtx);
8405 return (dev != NULL);
8409 * Add a vdev for use by the L2ARC. By this point the spa has already
8410 * validated the vdev and opened it.
8413 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
8415 l2arc_dev_t *adddev;
8417 ASSERT(!l2arc_vdev_present(vd));
8419 vdev_ashift_optimize(vd);
8422 * Create a new l2arc device entry.
8424 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
8425 adddev->l2ad_spa = spa;
8426 adddev->l2ad_vdev = vd;
8427 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
8428 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
8429 adddev->l2ad_hand = adddev->l2ad_start;
8430 adddev->l2ad_first = B_TRUE;
8431 adddev->l2ad_writing = B_FALSE;
8433 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
8435 * This is a list of all ARC buffers that are still valid on the
8438 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
8439 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
8441 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
8442 refcount_create(&adddev->l2ad_alloc);
8445 * Add device to global list
8447 mutex_enter(&l2arc_dev_mtx);
8448 list_insert_head(l2arc_dev_list, adddev);
8449 atomic_inc_64(&l2arc_ndev);
8450 mutex_exit(&l2arc_dev_mtx);
8454 * Remove a vdev from the L2ARC.
8457 l2arc_remove_vdev(vdev_t *vd)
8459 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
8462 * Find the device by vdev
8464 mutex_enter(&l2arc_dev_mtx);
8465 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
8466 nextdev = list_next(l2arc_dev_list, dev);
8467 if (vd == dev->l2ad_vdev) {
8472 ASSERT3P(remdev, !=, NULL);
8475 * Remove device from global list
8477 list_remove(l2arc_dev_list, remdev);
8478 l2arc_dev_last = NULL; /* may have been invalidated */
8479 atomic_dec_64(&l2arc_ndev);
8480 mutex_exit(&l2arc_dev_mtx);
8483 * Clear all buflists and ARC references. L2ARC device flush.
8485 l2arc_evict(remdev, 0, B_TRUE);
8486 list_destroy(&remdev->l2ad_buflist);
8487 mutex_destroy(&remdev->l2ad_mtx);
8488 refcount_destroy(&remdev->l2ad_alloc);
8489 kmem_free(remdev, sizeof (l2arc_dev_t));
8495 l2arc_thread_exit = 0;
8497 l2arc_writes_sent = 0;
8498 l2arc_writes_done = 0;
8500 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
8501 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
8502 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
8503 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
8505 l2arc_dev_list = &L2ARC_dev_list;
8506 l2arc_free_on_write = &L2ARC_free_on_write;
8507 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
8508 offsetof(l2arc_dev_t, l2ad_node));
8509 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
8510 offsetof(l2arc_data_free_t, l2df_list_node));
8517 * This is called from dmu_fini(), which is called from spa_fini();
8518 * Because of this, we can assume that all l2arc devices have
8519 * already been removed when the pools themselves were removed.
8522 l2arc_do_free_on_write();
8524 mutex_destroy(&l2arc_feed_thr_lock);
8525 cv_destroy(&l2arc_feed_thr_cv);
8526 mutex_destroy(&l2arc_dev_mtx);
8527 mutex_destroy(&l2arc_free_on_write_mtx);
8529 list_destroy(l2arc_dev_list);
8530 list_destroy(l2arc_free_on_write);
8536 if (!(spa_mode_global & FWRITE))
8539 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
8540 TS_RUN, minclsyspri);
8546 if (!(spa_mode_global & FWRITE))
8549 mutex_enter(&l2arc_feed_thr_lock);
8550 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
8551 l2arc_thread_exit = 1;
8552 while (l2arc_thread_exit != 0)
8553 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
8554 mutex_exit(&l2arc_feed_thr_lock);