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) 2012, Joyent, Inc. All rights reserved.
24 * Copyright (c) 2011, 2016 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2015 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 * Arc buffers may have an associated eviction callback function.
103 * This function will be invoked prior to removing the buffer (e.g.
104 * in arc_do_user_evicts()). Note however that the data associated
105 * with the buffer may be evicted prior to the callback. The callback
106 * must be made with *no locks held* (to prevent deadlock). Additionally,
107 * the users of callbacks must ensure that their private data is
108 * protected from simultaneous callbacks from arc_clear_callback()
109 * and arc_do_user_evicts().
111 * Note that the majority of the performance stats are manipulated
112 * with atomic operations.
114 * The L2ARC uses the l2ad_mtx on each vdev for the following:
116 * - L2ARC buflist creation
117 * - L2ARC buflist eviction
118 * - L2ARC write completion, which walks L2ARC buflists
119 * - ARC header destruction, as it removes from L2ARC buflists
120 * - ARC header release, as it removes from L2ARC buflists
126 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
127 * This structure can point either to a block that is still in the cache or to
128 * one that is only accessible in an L2 ARC device, or it can provide
129 * information about a block that was recently evicted. If a block is
130 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
131 * information to retrieve it from the L2ARC device. This information is
132 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
133 * that is in this state cannot access the data directly.
135 * Blocks that are actively being referenced or have not been evicted
136 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
137 * the arc_buf_hdr_t that will point to the data block in memory. A block can
138 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
139 * caches data in two ways -- in a list of arc buffers (arc_buf_t) and
140 * also in the arc_buf_hdr_t's private physical data block pointer (b_pdata).
141 * Each arc buffer (arc_buf_t) is being actively accessed by a specific ARC
142 * consumer, and always contains uncompressed data. The ARC will provide
143 * references to this data and will keep it cached until it is no longer in
144 * use. Typically, the arc will try to cache only the L1ARC's physical data
145 * block and will aggressively evict any arc_buf_t that is no longer referenced.
146 * The amount of memory consumed by the arc_buf_t's can be seen via the
147 * "overhead_size" kstat.
161 * | b_buf +------------>+---------+ arc_buf_t
162 * | | |b_next +---->+---------+
163 * | b_pdata +-+ |---------| |b_next +-->NULL
164 * +-----------+ | | | +---------+
166 * | +---------+ | |b_data +-+
167 * +->+------+ | +---------+ |
168 * (potentially) | | | |
171 * +->+------+ +------+
172 * uncompressed | | | |
176 * The L1ARC's data pointer, however, may or may not be uncompressed. The
177 * ARC has the ability to store the physical data (b_pdata) associated with
178 * the DVA of the arc_buf_hdr_t. Since the b_pdata is a copy of the on-disk
179 * physical block, it will match its on-disk compression characteristics.
180 * If the block on-disk is compressed, then the physical data block
181 * in the cache will also be compressed and vice-versa. This behavior
182 * can be disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
183 * compressed ARC functionality is disabled, the b_pdata will point to an
184 * uncompressed version of the on-disk data.
186 * When a consumer reads a block, the ARC must first look to see if the
187 * arc_buf_hdr_t is cached. If the hdr is cached and already has an arc_buf_t,
188 * then an additional arc_buf_t is allocated and the uncompressed data is
189 * bcopied from the existing arc_buf_t. If the hdr is cached but does not
190 * have an arc_buf_t, then the ARC allocates a new arc_buf_t and decompresses
191 * the b_pdata contents into the arc_buf_t's b_data. If the arc_buf_hdr_t's
192 * b_pdata is not compressed, then the block is shared with the newly
193 * allocated arc_buf_t. This block sharing only occurs with one arc_buf_t
194 * in the arc buffer chain. Sharing the block reduces the memory overhead
195 * required when the hdr is caching uncompressed blocks or the compressed
196 * arc functionality has been disabled via 'zfs_compressed_arc_enabled'.
198 * The diagram below shows an example of an uncompressed ARC hdr that is
199 * sharing its data with an arc_buf_t:
211 * | | arc_buf_t (shared)
212 * | b_buf +------------>+---------+ arc_buf_t
213 * | | |b_next +---->+---------+
214 * | b_pdata +-+ |---------| |b_next +-->NULL
215 * +-----------+ | | | +---------+
217 * | +---------+ | |b_data +-+
218 * +->+------+ | +---------+ |
220 * uncompressed | | | |
223 * | uncompressed | | |
226 * +---------------------------------+
228 * Writing to the arc requires that the ARC first discard the b_pdata
229 * since the physical block is about to be rewritten. The new data contents
230 * will be contained in the arc_buf_t (uncompressed). As the I/O pipeline
231 * performs the write, it may compress the data before writing it to disk.
232 * The ARC will be called with the transformed data and will bcopy the
233 * transformed on-disk block into a newly allocated b_pdata.
235 * When the L2ARC is in use, it will also take advantage of the b_pdata. The
236 * L2ARC will always write the contents of b_pdata to the L2ARC. This means
237 * that when compressed arc is enabled that the L2ARC blocks are identical
238 * to the on-disk block in the main data pool. This provides a significant
239 * advantage since the ARC can leverage the bp's checksum when reading from the
240 * L2ARC to determine if the contents are valid. However, if the compressed
241 * arc is disabled, then the L2ARC's block must be transformed to look
242 * like the physical block in the main data pool before comparing the
243 * checksum and determining its validity.
248 #include <sys/spa_impl.h>
249 #include <sys/zio_compress.h>
250 #include <sys/zio_checksum.h>
251 #include <sys/zfs_context.h>
253 #include <sys/refcount.h>
254 #include <sys/vdev.h>
255 #include <sys/vdev_impl.h>
256 #include <sys/dsl_pool.h>
257 #include <sys/multilist.h>
259 #include <sys/dnlc.h>
260 #include <sys/racct.h>
262 #include <sys/callb.h>
263 #include <sys/kstat.h>
264 #include <sys/trim_map.h>
265 #include <zfs_fletcher.h>
268 #include <machine/vmparam.h>
272 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
273 boolean_t arc_watch = B_FALSE;
278 static kmutex_t arc_reclaim_lock;
279 static kcondvar_t arc_reclaim_thread_cv;
280 static boolean_t arc_reclaim_thread_exit;
281 static kcondvar_t arc_reclaim_waiters_cv;
283 static kmutex_t arc_dnlc_evicts_lock;
284 static kcondvar_t arc_dnlc_evicts_cv;
285 static boolean_t arc_dnlc_evicts_thread_exit;
287 uint_t arc_reduce_dnlc_percent = 3;
290 * The number of headers to evict in arc_evict_state_impl() before
291 * dropping the sublist lock and evicting from another sublist. A lower
292 * value means we're more likely to evict the "correct" header (i.e. the
293 * oldest header in the arc state), but comes with higher overhead
294 * (i.e. more invocations of arc_evict_state_impl()).
296 int zfs_arc_evict_batch_limit = 10;
299 * The number of sublists used for each of the arc state lists. If this
300 * is not set to a suitable value by the user, it will be configured to
301 * the number of CPUs on the system in arc_init().
303 int zfs_arc_num_sublists_per_state = 0;
305 /* number of seconds before growing cache again */
306 static int arc_grow_retry = 60;
308 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */
309 int zfs_arc_overflow_shift = 8;
311 /* shift of arc_c for calculating both min and max arc_p */
312 static int arc_p_min_shift = 4;
314 /* log2(fraction of arc to reclaim) */
315 static int arc_shrink_shift = 7;
318 * log2(fraction of ARC which must be free to allow growing).
319 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
320 * when reading a new block into the ARC, we will evict an equal-sized block
323 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
324 * we will still not allow it to grow.
326 int arc_no_grow_shift = 5;
330 * minimum lifespan of a prefetch block in clock ticks
331 * (initialized in arc_init())
333 static int arc_min_prefetch_lifespan;
336 * If this percent of memory is free, don't throttle.
338 int arc_lotsfree_percent = 10;
341 extern boolean_t zfs_prefetch_disable;
344 * The arc has filled available memory and has now warmed up.
346 static boolean_t arc_warm;
349 * These tunables are for performance analysis.
351 uint64_t zfs_arc_max;
352 uint64_t zfs_arc_min;
353 uint64_t zfs_arc_meta_limit = 0;
354 uint64_t zfs_arc_meta_min = 0;
355 int zfs_arc_grow_retry = 0;
356 int zfs_arc_shrink_shift = 0;
357 int zfs_arc_p_min_shift = 0;
358 uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
359 u_int zfs_arc_free_target = 0;
361 /* Absolute min for arc min / max is 16MB. */
362 static uint64_t arc_abs_min = 16 << 20;
364 boolean_t zfs_compressed_arc_enabled = B_TRUE;
366 static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
367 static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
368 static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS);
369 static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS);
371 #if defined(__FreeBSD__) && defined(_KERNEL)
373 arc_free_target_init(void *unused __unused)
376 zfs_arc_free_target = vm_pageout_wakeup_thresh;
378 SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
379 arc_free_target_init, NULL);
381 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
382 TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
383 TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
384 SYSCTL_DECL(_vfs_zfs);
385 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN,
386 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size");
387 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN,
388 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size");
389 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
390 &zfs_arc_average_blocksize, 0,
391 "ARC average blocksize");
392 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
393 &arc_shrink_shift, 0,
394 "log2(fraction of arc to reclaim)");
395 SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN,
396 &zfs_compressed_arc_enabled, 0, "Enable compressed ARC");
399 * We don't have a tunable for arc_free_target due to the dependency on
400 * pagedaemon initialisation.
402 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
403 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
404 sysctl_vfs_zfs_arc_free_target, "IU",
405 "Desired number of free pages below which ARC triggers reclaim");
408 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
413 val = zfs_arc_free_target;
414 err = sysctl_handle_int(oidp, &val, 0, req);
415 if (err != 0 || req->newptr == NULL)
420 if (val > vm_cnt.v_page_count)
423 zfs_arc_free_target = val;
429 * Must be declared here, before the definition of corresponding kstat
430 * macro which uses the same names will confuse the compiler.
432 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
433 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
434 sysctl_vfs_zfs_arc_meta_limit, "QU",
435 "ARC metadata limit");
439 * Note that buffers can be in one of 6 states:
440 * ARC_anon - anonymous (discussed below)
441 * ARC_mru - recently used, currently cached
442 * ARC_mru_ghost - recentely used, no longer in cache
443 * ARC_mfu - frequently used, currently cached
444 * ARC_mfu_ghost - frequently used, no longer in cache
445 * ARC_l2c_only - exists in L2ARC but not other states
446 * When there are no active references to the buffer, they are
447 * are linked onto a list in one of these arc states. These are
448 * the only buffers that can be evicted or deleted. Within each
449 * state there are multiple lists, one for meta-data and one for
450 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
451 * etc.) is tracked separately so that it can be managed more
452 * explicitly: favored over data, limited explicitly.
454 * Anonymous buffers are buffers that are not associated with
455 * a DVA. These are buffers that hold dirty block copies
456 * before they are written to stable storage. By definition,
457 * they are "ref'd" and are considered part of arc_mru
458 * that cannot be freed. Generally, they will aquire a DVA
459 * as they are written and migrate onto the arc_mru list.
461 * The ARC_l2c_only state is for buffers that are in the second
462 * level ARC but no longer in any of the ARC_m* lists. The second
463 * level ARC itself may also contain buffers that are in any of
464 * the ARC_m* states - meaning that a buffer can exist in two
465 * places. The reason for the ARC_l2c_only state is to keep the
466 * buffer header in the hash table, so that reads that hit the
467 * second level ARC benefit from these fast lookups.
470 typedef struct arc_state {
472 * list of evictable buffers
474 multilist_t arcs_list[ARC_BUFC_NUMTYPES];
476 * total amount of evictable data in this state
478 refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
480 * total amount of data in this state; this includes: evictable,
481 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
483 refcount_t arcs_size;
487 static arc_state_t ARC_anon;
488 static arc_state_t ARC_mru;
489 static arc_state_t ARC_mru_ghost;
490 static arc_state_t ARC_mfu;
491 static arc_state_t ARC_mfu_ghost;
492 static arc_state_t ARC_l2c_only;
494 typedef struct arc_stats {
495 kstat_named_t arcstat_hits;
496 kstat_named_t arcstat_misses;
497 kstat_named_t arcstat_demand_data_hits;
498 kstat_named_t arcstat_demand_data_misses;
499 kstat_named_t arcstat_demand_metadata_hits;
500 kstat_named_t arcstat_demand_metadata_misses;
501 kstat_named_t arcstat_prefetch_data_hits;
502 kstat_named_t arcstat_prefetch_data_misses;
503 kstat_named_t arcstat_prefetch_metadata_hits;
504 kstat_named_t arcstat_prefetch_metadata_misses;
505 kstat_named_t arcstat_mru_hits;
506 kstat_named_t arcstat_mru_ghost_hits;
507 kstat_named_t arcstat_mfu_hits;
508 kstat_named_t arcstat_mfu_ghost_hits;
509 kstat_named_t arcstat_allocated;
510 kstat_named_t arcstat_deleted;
512 * Number of buffers that could not be evicted because the hash lock
513 * was held by another thread. The lock may not necessarily be held
514 * by something using the same buffer, since hash locks are shared
515 * by multiple buffers.
517 kstat_named_t arcstat_mutex_miss;
519 * Number of buffers skipped because they have I/O in progress, are
520 * indrect prefetch buffers that have not lived long enough, or are
521 * not from the spa we're trying to evict from.
523 kstat_named_t arcstat_evict_skip;
525 * Number of times arc_evict_state() was unable to evict enough
526 * buffers to reach it's target amount.
528 kstat_named_t arcstat_evict_not_enough;
529 kstat_named_t arcstat_evict_l2_cached;
530 kstat_named_t arcstat_evict_l2_eligible;
531 kstat_named_t arcstat_evict_l2_ineligible;
532 kstat_named_t arcstat_evict_l2_skip;
533 kstat_named_t arcstat_hash_elements;
534 kstat_named_t arcstat_hash_elements_max;
535 kstat_named_t arcstat_hash_collisions;
536 kstat_named_t arcstat_hash_chains;
537 kstat_named_t arcstat_hash_chain_max;
538 kstat_named_t arcstat_p;
539 kstat_named_t arcstat_c;
540 kstat_named_t arcstat_c_min;
541 kstat_named_t arcstat_c_max;
542 kstat_named_t arcstat_size;
544 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pdata.
545 * Note that the compressed bytes may match the uncompressed bytes
546 * if the block is either not compressed or compressed arc is disabled.
548 kstat_named_t arcstat_compressed_size;
550 * Uncompressed size of the data stored in b_pdata. If compressed
551 * arc is disabled then this value will be identical to the stat
554 kstat_named_t arcstat_uncompressed_size;
556 * Number of bytes stored in all the arc_buf_t's. This is classified
557 * as "overhead" since this data is typically short-lived and will
558 * be evicted from the arc when it becomes unreferenced unless the
559 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
560 * values have been set (see comment in dbuf.c for more information).
562 kstat_named_t arcstat_overhead_size;
564 * Number of bytes consumed by internal ARC structures necessary
565 * for tracking purposes; these structures are not actually
566 * backed by ARC buffers. This includes arc_buf_hdr_t structures
567 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
568 * caches), and arc_buf_t structures (allocated via arc_buf_t
571 kstat_named_t arcstat_hdr_size;
573 * Number of bytes consumed by ARC buffers of type equal to
574 * ARC_BUFC_DATA. This is generally consumed by buffers backing
575 * on disk user data (e.g. plain file contents).
577 kstat_named_t arcstat_data_size;
579 * Number of bytes consumed by ARC buffers of type equal to
580 * ARC_BUFC_METADATA. This is generally consumed by buffers
581 * backing on disk data that is used for internal ZFS
582 * structures (e.g. ZAP, dnode, indirect blocks, etc).
584 kstat_named_t arcstat_metadata_size;
586 * Number of bytes consumed by various buffers and structures
587 * not actually backed with ARC buffers. This includes bonus
588 * buffers (allocated directly via zio_buf_* functions),
589 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
590 * cache), and dnode_t structures (allocated via dnode_t cache).
592 kstat_named_t arcstat_other_size;
594 * Total number of bytes consumed by ARC buffers residing in the
595 * arc_anon state. This includes *all* buffers in the arc_anon
596 * state; e.g. data, metadata, evictable, and unevictable buffers
597 * are all included in this value.
599 kstat_named_t arcstat_anon_size;
601 * Number of bytes consumed by ARC buffers that meet the
602 * following criteria: backing buffers of type ARC_BUFC_DATA,
603 * residing in the arc_anon state, and are eligible for eviction
604 * (e.g. have no outstanding holds on the buffer).
606 kstat_named_t arcstat_anon_evictable_data;
608 * Number of bytes consumed by ARC buffers that meet the
609 * following criteria: backing buffers of type ARC_BUFC_METADATA,
610 * residing in the arc_anon state, and are eligible for eviction
611 * (e.g. have no outstanding holds on the buffer).
613 kstat_named_t arcstat_anon_evictable_metadata;
615 * Total number of bytes consumed by ARC buffers residing in the
616 * arc_mru state. This includes *all* buffers in the arc_mru
617 * state; e.g. data, metadata, evictable, and unevictable buffers
618 * are all included in this value.
620 kstat_named_t arcstat_mru_size;
622 * Number of bytes consumed by ARC buffers that meet the
623 * following criteria: backing buffers of type ARC_BUFC_DATA,
624 * residing in the arc_mru state, and are eligible for eviction
625 * (e.g. have no outstanding holds on the buffer).
627 kstat_named_t arcstat_mru_evictable_data;
629 * Number of bytes consumed by ARC buffers that meet the
630 * following criteria: backing buffers of type ARC_BUFC_METADATA,
631 * residing in the arc_mru state, and are eligible for eviction
632 * (e.g. have no outstanding holds on the buffer).
634 kstat_named_t arcstat_mru_evictable_metadata;
636 * Total number of bytes that *would have been* consumed by ARC
637 * buffers in the arc_mru_ghost state. The key thing to note
638 * here, is the fact that this size doesn't actually indicate
639 * RAM consumption. The ghost lists only consist of headers and
640 * don't actually have ARC buffers linked off of these headers.
641 * Thus, *if* the headers had associated ARC buffers, these
642 * buffers *would have* consumed this number of bytes.
644 kstat_named_t arcstat_mru_ghost_size;
646 * Number of bytes that *would have been* consumed by ARC
647 * buffers that are eligible for eviction, of type
648 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
650 kstat_named_t arcstat_mru_ghost_evictable_data;
652 * Number of bytes that *would have been* consumed by ARC
653 * buffers that are eligible for eviction, of type
654 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
656 kstat_named_t arcstat_mru_ghost_evictable_metadata;
658 * Total number of bytes consumed by ARC buffers residing in the
659 * arc_mfu state. This includes *all* buffers in the arc_mfu
660 * state; e.g. data, metadata, evictable, and unevictable buffers
661 * are all included in this value.
663 kstat_named_t arcstat_mfu_size;
665 * Number of bytes consumed by ARC buffers that are eligible for
666 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
669 kstat_named_t arcstat_mfu_evictable_data;
671 * Number of bytes consumed by ARC buffers that are eligible for
672 * eviction, of type ARC_BUFC_METADATA, and reside in the
675 kstat_named_t arcstat_mfu_evictable_metadata;
677 * Total number of bytes that *would have been* consumed by ARC
678 * buffers in the arc_mfu_ghost state. See the comment above
679 * arcstat_mru_ghost_size for more details.
681 kstat_named_t arcstat_mfu_ghost_size;
683 * Number of bytes that *would have been* consumed by ARC
684 * buffers that are eligible for eviction, of type
685 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
687 kstat_named_t arcstat_mfu_ghost_evictable_data;
689 * Number of bytes that *would have been* consumed by ARC
690 * buffers that are eligible for eviction, of type
691 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
693 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
694 kstat_named_t arcstat_l2_hits;
695 kstat_named_t arcstat_l2_misses;
696 kstat_named_t arcstat_l2_feeds;
697 kstat_named_t arcstat_l2_rw_clash;
698 kstat_named_t arcstat_l2_read_bytes;
699 kstat_named_t arcstat_l2_write_bytes;
700 kstat_named_t arcstat_l2_writes_sent;
701 kstat_named_t arcstat_l2_writes_done;
702 kstat_named_t arcstat_l2_writes_error;
703 kstat_named_t arcstat_l2_writes_lock_retry;
704 kstat_named_t arcstat_l2_evict_lock_retry;
705 kstat_named_t arcstat_l2_evict_reading;
706 kstat_named_t arcstat_l2_evict_l1cached;
707 kstat_named_t arcstat_l2_free_on_write;
708 kstat_named_t arcstat_l2_abort_lowmem;
709 kstat_named_t arcstat_l2_cksum_bad;
710 kstat_named_t arcstat_l2_io_error;
711 kstat_named_t arcstat_l2_size;
712 kstat_named_t arcstat_l2_asize;
713 kstat_named_t arcstat_l2_hdr_size;
714 kstat_named_t arcstat_l2_padding_needed;
715 kstat_named_t arcstat_l2_write_trylock_fail;
716 kstat_named_t arcstat_l2_write_passed_headroom;
717 kstat_named_t arcstat_l2_write_spa_mismatch;
718 kstat_named_t arcstat_l2_write_in_l2;
719 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
720 kstat_named_t arcstat_l2_write_not_cacheable;
721 kstat_named_t arcstat_l2_write_full;
722 kstat_named_t arcstat_l2_write_buffer_iter;
723 kstat_named_t arcstat_l2_write_pios;
724 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
725 kstat_named_t arcstat_l2_write_buffer_list_iter;
726 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
727 kstat_named_t arcstat_memory_throttle_count;
728 kstat_named_t arcstat_meta_used;
729 kstat_named_t arcstat_meta_limit;
730 kstat_named_t arcstat_meta_max;
731 kstat_named_t arcstat_meta_min;
732 kstat_named_t arcstat_sync_wait_for_async;
733 kstat_named_t arcstat_demand_hit_predictive_prefetch;
736 static arc_stats_t arc_stats = {
737 { "hits", KSTAT_DATA_UINT64 },
738 { "misses", KSTAT_DATA_UINT64 },
739 { "demand_data_hits", KSTAT_DATA_UINT64 },
740 { "demand_data_misses", KSTAT_DATA_UINT64 },
741 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
742 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
743 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
744 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
745 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
746 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
747 { "mru_hits", KSTAT_DATA_UINT64 },
748 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
749 { "mfu_hits", KSTAT_DATA_UINT64 },
750 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
751 { "allocated", KSTAT_DATA_UINT64 },
752 { "deleted", KSTAT_DATA_UINT64 },
753 { "mutex_miss", KSTAT_DATA_UINT64 },
754 { "evict_skip", KSTAT_DATA_UINT64 },
755 { "evict_not_enough", KSTAT_DATA_UINT64 },
756 { "evict_l2_cached", KSTAT_DATA_UINT64 },
757 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
758 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
759 { "evict_l2_skip", KSTAT_DATA_UINT64 },
760 { "hash_elements", KSTAT_DATA_UINT64 },
761 { "hash_elements_max", KSTAT_DATA_UINT64 },
762 { "hash_collisions", KSTAT_DATA_UINT64 },
763 { "hash_chains", KSTAT_DATA_UINT64 },
764 { "hash_chain_max", KSTAT_DATA_UINT64 },
765 { "p", KSTAT_DATA_UINT64 },
766 { "c", KSTAT_DATA_UINT64 },
767 { "c_min", KSTAT_DATA_UINT64 },
768 { "c_max", KSTAT_DATA_UINT64 },
769 { "size", KSTAT_DATA_UINT64 },
770 { "compressed_size", KSTAT_DATA_UINT64 },
771 { "uncompressed_size", KSTAT_DATA_UINT64 },
772 { "overhead_size", KSTAT_DATA_UINT64 },
773 { "hdr_size", KSTAT_DATA_UINT64 },
774 { "data_size", KSTAT_DATA_UINT64 },
775 { "metadata_size", KSTAT_DATA_UINT64 },
776 { "other_size", KSTAT_DATA_UINT64 },
777 { "anon_size", KSTAT_DATA_UINT64 },
778 { "anon_evictable_data", KSTAT_DATA_UINT64 },
779 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
780 { "mru_size", KSTAT_DATA_UINT64 },
781 { "mru_evictable_data", KSTAT_DATA_UINT64 },
782 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
783 { "mru_ghost_size", KSTAT_DATA_UINT64 },
784 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
785 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
786 { "mfu_size", KSTAT_DATA_UINT64 },
787 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
788 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
789 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
790 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
791 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
792 { "l2_hits", KSTAT_DATA_UINT64 },
793 { "l2_misses", KSTAT_DATA_UINT64 },
794 { "l2_feeds", KSTAT_DATA_UINT64 },
795 { "l2_rw_clash", KSTAT_DATA_UINT64 },
796 { "l2_read_bytes", KSTAT_DATA_UINT64 },
797 { "l2_write_bytes", KSTAT_DATA_UINT64 },
798 { "l2_writes_sent", KSTAT_DATA_UINT64 },
799 { "l2_writes_done", KSTAT_DATA_UINT64 },
800 { "l2_writes_error", KSTAT_DATA_UINT64 },
801 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
802 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
803 { "l2_evict_reading", KSTAT_DATA_UINT64 },
804 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
805 { "l2_free_on_write", KSTAT_DATA_UINT64 },
806 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
807 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
808 { "l2_io_error", KSTAT_DATA_UINT64 },
809 { "l2_size", KSTAT_DATA_UINT64 },
810 { "l2_asize", KSTAT_DATA_UINT64 },
811 { "l2_hdr_size", KSTAT_DATA_UINT64 },
812 { "l2_padding_needed", KSTAT_DATA_UINT64 },
813 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
814 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
815 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
816 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
817 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
818 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
819 { "l2_write_full", KSTAT_DATA_UINT64 },
820 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
821 { "l2_write_pios", KSTAT_DATA_UINT64 },
822 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
823 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
824 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
825 { "memory_throttle_count", KSTAT_DATA_UINT64 },
826 { "arc_meta_used", KSTAT_DATA_UINT64 },
827 { "arc_meta_limit", KSTAT_DATA_UINT64 },
828 { "arc_meta_max", KSTAT_DATA_UINT64 },
829 { "arc_meta_min", KSTAT_DATA_UINT64 },
830 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
831 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
834 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
836 #define ARCSTAT_INCR(stat, val) \
837 atomic_add_64(&arc_stats.stat.value.ui64, (val))
839 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
840 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
842 #define ARCSTAT_MAX(stat, val) { \
844 while ((val) > (m = arc_stats.stat.value.ui64) && \
845 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
849 #define ARCSTAT_MAXSTAT(stat) \
850 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
853 * We define a macro to allow ARC hits/misses to be easily broken down by
854 * two separate conditions, giving a total of four different subtypes for
855 * each of hits and misses (so eight statistics total).
857 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
860 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
862 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
866 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
868 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
873 static arc_state_t *arc_anon;
874 static arc_state_t *arc_mru;
875 static arc_state_t *arc_mru_ghost;
876 static arc_state_t *arc_mfu;
877 static arc_state_t *arc_mfu_ghost;
878 static arc_state_t *arc_l2c_only;
881 * There are several ARC variables that are critical to export as kstats --
882 * but we don't want to have to grovel around in the kstat whenever we wish to
883 * manipulate them. For these variables, we therefore define them to be in
884 * terms of the statistic variable. This assures that we are not introducing
885 * the possibility of inconsistency by having shadow copies of the variables,
886 * while still allowing the code to be readable.
888 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
889 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
890 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
891 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
892 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
893 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
894 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
895 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
896 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
898 /* compressed size of entire arc */
899 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
900 /* uncompressed size of entire arc */
901 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
902 /* number of bytes in the arc from arc_buf_t's */
903 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
905 static int arc_no_grow; /* Don't try to grow cache size */
906 static uint64_t arc_tempreserve;
907 static uint64_t arc_loaned_bytes;
909 typedef struct arc_callback arc_callback_t;
911 struct arc_callback {
913 arc_done_func_t *acb_done;
915 zio_t *acb_zio_dummy;
916 arc_callback_t *acb_next;
919 typedef struct arc_write_callback arc_write_callback_t;
921 struct arc_write_callback {
923 arc_done_func_t *awcb_ready;
924 arc_done_func_t *awcb_children_ready;
925 arc_done_func_t *awcb_physdone;
926 arc_done_func_t *awcb_done;
931 * ARC buffers are separated into multiple structs as a memory saving measure:
932 * - Common fields struct, always defined, and embedded within it:
933 * - L2-only fields, always allocated but undefined when not in L2ARC
934 * - L1-only fields, only allocated when in L1ARC
936 * Buffer in L1 Buffer only in L2
937 * +------------------------+ +------------------------+
938 * | arc_buf_hdr_t | | arc_buf_hdr_t |
942 * +------------------------+ +------------------------+
943 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
944 * | (undefined if L1-only) | | |
945 * +------------------------+ +------------------------+
946 * | l1arc_buf_hdr_t |
951 * +------------------------+
953 * Because it's possible for the L2ARC to become extremely large, we can wind
954 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
955 * is minimized by only allocating the fields necessary for an L1-cached buffer
956 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
957 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
958 * words in pointers. arc_hdr_realloc() is used to switch a header between
959 * these two allocation states.
961 typedef struct l1arc_buf_hdr {
962 kmutex_t b_freeze_lock;
963 zio_cksum_t *b_freeze_cksum;
966 * used for debugging wtih kmem_flags - by allocating and freeing
967 * b_thawed when the buffer is thawed, we get a record of the stack
968 * trace that thawed it.
975 /* for waiting on writes to complete */
979 /* protected by arc state mutex */
980 arc_state_t *b_state;
981 multilist_node_t b_arc_node;
983 /* updated atomically */
984 clock_t b_arc_access;
986 /* self protecting */
989 arc_callback_t *b_acb;
993 typedef struct l2arc_dev l2arc_dev_t;
995 typedef struct l2arc_buf_hdr {
996 /* protected by arc_buf_hdr mutex */
997 l2arc_dev_t *b_dev; /* L2ARC device */
998 uint64_t b_daddr; /* disk address, offset byte */
1000 list_node_t b_l2node;
1003 struct arc_buf_hdr {
1004 /* protected by hash lock */
1008 arc_buf_contents_t b_type;
1009 arc_buf_hdr_t *b_hash_next;
1010 arc_flags_t b_flags;
1013 * This field stores the size of the data buffer after
1014 * compression, and is set in the arc's zio completion handlers.
1015 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
1017 * While the block pointers can store up to 32MB in their psize
1018 * field, we can only store up to 32MB minus 512B. This is due
1019 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
1020 * a field of zeros represents 512B in the bp). We can't use a
1021 * bias of 1 since we need to reserve a psize of zero, here, to
1022 * represent holes and embedded blocks.
1024 * This isn't a problem in practice, since the maximum size of a
1025 * buffer is limited to 16MB, so we never need to store 32MB in
1026 * this field. Even in the upstream illumos code base, the
1027 * maximum size of a buffer is limited to 16MB.
1032 * This field stores the size of the data buffer before
1033 * compression, and cannot change once set. It is in units
1034 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
1036 uint16_t b_lsize; /* immutable */
1037 uint64_t b_spa; /* immutable */
1039 /* L2ARC fields. Undefined when not in L2ARC. */
1040 l2arc_buf_hdr_t b_l2hdr;
1041 /* L1ARC fields. Undefined when in l2arc_only state */
1042 l1arc_buf_hdr_t b_l1hdr;
1045 #if defined(__FreeBSD__) && defined(_KERNEL)
1047 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
1052 val = arc_meta_limit;
1053 err = sysctl_handle_64(oidp, &val, 0, req);
1054 if (err != 0 || req->newptr == NULL)
1057 if (val <= 0 || val > arc_c_max)
1060 arc_meta_limit = val;
1065 sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
1071 err = sysctl_handle_64(oidp, &val, 0, req);
1072 if (err != 0 || req->newptr == NULL)
1075 if (zfs_arc_max == 0) {
1076 /* Loader tunable so blindly set */
1081 if (val < arc_abs_min || val > kmem_size())
1083 if (val < arc_c_min)
1085 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
1091 arc_p = (arc_c >> 1);
1093 if (zfs_arc_meta_limit == 0) {
1094 /* limit meta-data to 1/4 of the arc capacity */
1095 arc_meta_limit = arc_c_max / 4;
1098 /* if kmem_flags are set, lets try to use less memory */
1099 if (kmem_debugging())
1102 zfs_arc_max = arc_c;
1108 sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
1114 err = sysctl_handle_64(oidp, &val, 0, req);
1115 if (err != 0 || req->newptr == NULL)
1118 if (zfs_arc_min == 0) {
1119 /* Loader tunable so blindly set */
1124 if (val < arc_abs_min || val > arc_c_max)
1129 if (zfs_arc_meta_min == 0)
1130 arc_meta_min = arc_c_min / 2;
1132 if (arc_c < arc_c_min)
1135 zfs_arc_min = arc_c_min;
1141 #define GHOST_STATE(state) \
1142 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
1143 (state) == arc_l2c_only)
1145 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1146 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1147 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1148 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1149 #define HDR_COMPRESSION_ENABLED(hdr) \
1150 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1152 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1153 #define HDR_L2_READING(hdr) \
1154 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1155 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1156 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1157 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1158 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1159 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1161 #define HDR_ISTYPE_METADATA(hdr) \
1162 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1163 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1165 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1166 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1168 /* For storing compression mode in b_flags */
1169 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1171 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1172 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1173 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1174 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1176 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1182 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1183 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1186 * Hash table routines
1189 #define HT_LOCK_PAD CACHE_LINE_SIZE
1194 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1198 #define BUF_LOCKS 256
1199 typedef struct buf_hash_table {
1201 arc_buf_hdr_t **ht_table;
1202 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1205 static buf_hash_table_t buf_hash_table;
1207 #define BUF_HASH_INDEX(spa, dva, birth) \
1208 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1209 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1210 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1211 #define HDR_LOCK(hdr) \
1212 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1214 uint64_t zfs_crc64_table[256];
1220 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1221 #define L2ARC_HEADROOM 2 /* num of writes */
1223 * If we discover during ARC scan any buffers to be compressed, we boost
1224 * our headroom for the next scanning cycle by this percentage multiple.
1226 #define L2ARC_HEADROOM_BOOST 200
1227 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1228 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1230 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1231 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1233 /* L2ARC Performance Tunables */
1234 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1235 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1236 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1237 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1238 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1239 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1240 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1241 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1242 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1244 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1245 &l2arc_write_max, 0, "max write size");
1246 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1247 &l2arc_write_boost, 0, "extra write during warmup");
1248 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1249 &l2arc_headroom, 0, "number of dev writes");
1250 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1251 &l2arc_feed_secs, 0, "interval seconds");
1252 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1253 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1255 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1256 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1257 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1258 &l2arc_feed_again, 0, "turbo warmup");
1259 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1260 &l2arc_norw, 0, "no reads during writes");
1262 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1263 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1264 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD,
1265 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1266 "size of anonymous state");
1267 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD,
1268 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1269 "size of anonymous state");
1271 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1272 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1273 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD,
1274 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1275 "size of metadata in mru state");
1276 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD,
1277 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1278 "size of data in mru state");
1280 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1281 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1282 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD,
1283 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1284 "size of metadata in mru ghost state");
1285 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD,
1286 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1287 "size of data in mru ghost state");
1289 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1290 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1291 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD,
1292 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1293 "size of metadata in mfu state");
1294 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD,
1295 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1296 "size of data in mfu state");
1298 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1299 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1300 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD,
1301 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1302 "size of metadata in mfu ghost state");
1303 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD,
1304 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1305 "size of data in mfu ghost state");
1307 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1308 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1314 vdev_t *l2ad_vdev; /* vdev */
1315 spa_t *l2ad_spa; /* spa */
1316 uint64_t l2ad_hand; /* next write location */
1317 uint64_t l2ad_start; /* first addr on device */
1318 uint64_t l2ad_end; /* last addr on device */
1319 boolean_t l2ad_first; /* first sweep through */
1320 boolean_t l2ad_writing; /* currently writing */
1321 kmutex_t l2ad_mtx; /* lock for buffer list */
1322 list_t l2ad_buflist; /* buffer list */
1323 list_node_t l2ad_node; /* device list node */
1324 refcount_t l2ad_alloc; /* allocated bytes */
1327 static list_t L2ARC_dev_list; /* device list */
1328 static list_t *l2arc_dev_list; /* device list pointer */
1329 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1330 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1331 static list_t L2ARC_free_on_write; /* free after write buf list */
1332 static list_t *l2arc_free_on_write; /* free after write list ptr */
1333 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1334 static uint64_t l2arc_ndev; /* number of devices */
1336 typedef struct l2arc_read_callback {
1337 arc_buf_hdr_t *l2rcb_hdr; /* read buffer */
1338 blkptr_t l2rcb_bp; /* original blkptr */
1339 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1340 int l2rcb_flags; /* original flags */
1341 void *l2rcb_data; /* temporary buffer */
1342 } l2arc_read_callback_t;
1344 typedef struct l2arc_write_callback {
1345 l2arc_dev_t *l2wcb_dev; /* device info */
1346 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1347 } l2arc_write_callback_t;
1349 typedef struct l2arc_data_free {
1350 /* protected by l2arc_free_on_write_mtx */
1353 arc_buf_contents_t l2df_type;
1354 list_node_t l2df_list_node;
1355 } l2arc_data_free_t;
1357 static kmutex_t l2arc_feed_thr_lock;
1358 static kcondvar_t l2arc_feed_thr_cv;
1359 static uint8_t l2arc_thread_exit;
1361 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1362 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1363 static void arc_hdr_free_pdata(arc_buf_hdr_t *hdr);
1364 static void arc_hdr_alloc_pdata(arc_buf_hdr_t *);
1365 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1366 static boolean_t arc_is_overflowing();
1367 static void arc_buf_watch(arc_buf_t *);
1369 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1370 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1371 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1372 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1374 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1375 static void l2arc_read_done(zio_t *);
1378 l2arc_trim(const arc_buf_hdr_t *hdr)
1380 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1382 ASSERT(HDR_HAS_L2HDR(hdr));
1383 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1385 if (HDR_GET_PSIZE(hdr) != 0) {
1386 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1387 HDR_GET_PSIZE(hdr), 0);
1392 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1394 uint8_t *vdva = (uint8_t *)dva;
1395 uint64_t crc = -1ULL;
1398 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
1400 for (i = 0; i < sizeof (dva_t); i++)
1401 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
1403 crc ^= (spa>>8) ^ birth;
1408 #define HDR_EMPTY(hdr) \
1409 ((hdr)->b_dva.dva_word[0] == 0 && \
1410 (hdr)->b_dva.dva_word[1] == 0)
1412 #define HDR_EQUAL(spa, dva, birth, hdr) \
1413 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1414 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1415 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1418 buf_discard_identity(arc_buf_hdr_t *hdr)
1420 hdr->b_dva.dva_word[0] = 0;
1421 hdr->b_dva.dva_word[1] = 0;
1425 static arc_buf_hdr_t *
1426 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1428 const dva_t *dva = BP_IDENTITY(bp);
1429 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1430 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1431 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1434 mutex_enter(hash_lock);
1435 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1436 hdr = hdr->b_hash_next) {
1437 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1442 mutex_exit(hash_lock);
1448 * Insert an entry into the hash table. If there is already an element
1449 * equal to elem in the hash table, then the already existing element
1450 * will be returned and the new element will not be inserted.
1451 * Otherwise returns NULL.
1452 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1454 static arc_buf_hdr_t *
1455 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1457 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1458 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1459 arc_buf_hdr_t *fhdr;
1462 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1463 ASSERT(hdr->b_birth != 0);
1464 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1466 if (lockp != NULL) {
1468 mutex_enter(hash_lock);
1470 ASSERT(MUTEX_HELD(hash_lock));
1473 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1474 fhdr = fhdr->b_hash_next, i++) {
1475 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1479 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1480 buf_hash_table.ht_table[idx] = hdr;
1481 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1483 /* collect some hash table performance data */
1485 ARCSTAT_BUMP(arcstat_hash_collisions);
1487 ARCSTAT_BUMP(arcstat_hash_chains);
1489 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1492 ARCSTAT_BUMP(arcstat_hash_elements);
1493 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1499 buf_hash_remove(arc_buf_hdr_t *hdr)
1501 arc_buf_hdr_t *fhdr, **hdrp;
1502 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1504 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1505 ASSERT(HDR_IN_HASH_TABLE(hdr));
1507 hdrp = &buf_hash_table.ht_table[idx];
1508 while ((fhdr = *hdrp) != hdr) {
1509 ASSERT3P(fhdr, !=, NULL);
1510 hdrp = &fhdr->b_hash_next;
1512 *hdrp = hdr->b_hash_next;
1513 hdr->b_hash_next = NULL;
1514 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1516 /* collect some hash table performance data */
1517 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1519 if (buf_hash_table.ht_table[idx] &&
1520 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1521 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1525 * Global data structures and functions for the buf kmem cache.
1527 static kmem_cache_t *hdr_full_cache;
1528 static kmem_cache_t *hdr_l2only_cache;
1529 static kmem_cache_t *buf_cache;
1536 kmem_free(buf_hash_table.ht_table,
1537 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1538 for (i = 0; i < BUF_LOCKS; i++)
1539 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1540 kmem_cache_destroy(hdr_full_cache);
1541 kmem_cache_destroy(hdr_l2only_cache);
1542 kmem_cache_destroy(buf_cache);
1546 * Constructor callback - called when the cache is empty
1547 * and a new buf is requested.
1551 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1553 arc_buf_hdr_t *hdr = vbuf;
1555 bzero(hdr, HDR_FULL_SIZE);
1556 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1557 refcount_create(&hdr->b_l1hdr.b_refcnt);
1558 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1559 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1560 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1567 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1569 arc_buf_hdr_t *hdr = vbuf;
1571 bzero(hdr, HDR_L2ONLY_SIZE);
1572 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1579 buf_cons(void *vbuf, void *unused, int kmflag)
1581 arc_buf_t *buf = vbuf;
1583 bzero(buf, sizeof (arc_buf_t));
1584 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1585 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1591 * Destructor callback - called when a cached buf is
1592 * no longer required.
1596 hdr_full_dest(void *vbuf, void *unused)
1598 arc_buf_hdr_t *hdr = vbuf;
1600 ASSERT(HDR_EMPTY(hdr));
1601 cv_destroy(&hdr->b_l1hdr.b_cv);
1602 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1603 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1604 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1605 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1610 hdr_l2only_dest(void *vbuf, void *unused)
1612 arc_buf_hdr_t *hdr = vbuf;
1614 ASSERT(HDR_EMPTY(hdr));
1615 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1620 buf_dest(void *vbuf, void *unused)
1622 arc_buf_t *buf = vbuf;
1624 mutex_destroy(&buf->b_evict_lock);
1625 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1629 * Reclaim callback -- invoked when memory is low.
1633 hdr_recl(void *unused)
1635 dprintf("hdr_recl called\n");
1637 * umem calls the reclaim func when we destroy the buf cache,
1638 * which is after we do arc_fini().
1641 cv_signal(&arc_reclaim_thread_cv);
1648 uint64_t hsize = 1ULL << 12;
1652 * The hash table is big enough to fill all of physical memory
1653 * with an average block size of zfs_arc_average_blocksize (default 8K).
1654 * By default, the table will take up
1655 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1657 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1660 buf_hash_table.ht_mask = hsize - 1;
1661 buf_hash_table.ht_table =
1662 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1663 if (buf_hash_table.ht_table == NULL) {
1664 ASSERT(hsize > (1ULL << 8));
1669 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1670 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1671 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1672 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1674 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1675 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1677 for (i = 0; i < 256; i++)
1678 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1679 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1681 for (i = 0; i < BUF_LOCKS; i++) {
1682 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1683 NULL, MUTEX_DEFAULT, NULL);
1687 #define ARC_MINTIME (hz>>4) /* 62 ms */
1689 static inline boolean_t
1690 arc_buf_is_shared(arc_buf_t *buf)
1692 boolean_t shared = (buf->b_data != NULL &&
1693 buf->b_data == buf->b_hdr->b_l1hdr.b_pdata);
1694 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1699 arc_cksum_free(arc_buf_hdr_t *hdr)
1701 ASSERT(HDR_HAS_L1HDR(hdr));
1702 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1703 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1704 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1705 hdr->b_l1hdr.b_freeze_cksum = NULL;
1707 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1711 arc_cksum_verify(arc_buf_t *buf)
1713 arc_buf_hdr_t *hdr = buf->b_hdr;
1716 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1719 ASSERT(HDR_HAS_L1HDR(hdr));
1721 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1722 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1723 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1726 fletcher_2_native(buf->b_data, HDR_GET_LSIZE(hdr), NULL, &zc);
1727 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1728 panic("buffer modified while frozen!");
1729 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1733 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1735 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1736 boolean_t valid_cksum;
1738 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1739 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1742 * We rely on the blkptr's checksum to determine if the block
1743 * is valid or not. When compressed arc is enabled, the l2arc
1744 * writes the block to the l2arc just as it appears in the pool.
1745 * This allows us to use the blkptr's checksum to validate the
1746 * data that we just read off of the l2arc without having to store
1747 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1748 * arc is disabled, then the data written to the l2arc is always
1749 * uncompressed and won't match the block as it exists in the main
1750 * pool. When this is the case, we must first compress it if it is
1751 * compressed on the main pool before we can validate the checksum.
1753 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1754 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1755 uint64_t lsize = HDR_GET_LSIZE(hdr);
1758 void *cbuf = zio_buf_alloc(HDR_GET_PSIZE(hdr));
1759 csize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1760 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1761 if (csize < HDR_GET_PSIZE(hdr)) {
1763 * Compressed blocks are always a multiple of the
1764 * smallest ashift in the pool. Ideally, we would
1765 * like to round up the csize to the next
1766 * spa_min_ashift but that value may have changed
1767 * since the block was last written. Instead,
1768 * we rely on the fact that the hdr's psize
1769 * was set to the psize of the block when it was
1770 * last written. We set the csize to that value
1771 * and zero out any part that should not contain
1774 bzero((char *)cbuf + csize, HDR_GET_PSIZE(hdr) - csize);
1775 csize = HDR_GET_PSIZE(hdr);
1777 zio_push_transform(zio, cbuf, csize, HDR_GET_PSIZE(hdr), NULL);
1781 * Block pointers always store the checksum for the logical data.
1782 * If the block pointer has the gang bit set, then the checksum
1783 * it represents is for the reconstituted data and not for an
1784 * individual gang member. The zio pipeline, however, must be able to
1785 * determine the checksum of each of the gang constituents so it
1786 * treats the checksum comparison differently than what we need
1787 * for l2arc blocks. This prevents us from using the
1788 * zio_checksum_error() interface directly. Instead we must call the
1789 * zio_checksum_error_impl() so that we can ensure the checksum is
1790 * generated using the correct checksum algorithm and accounts for the
1791 * logical I/O size and not just a gang fragment.
1793 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1794 BP_GET_CHECKSUM(zio->io_bp), zio->io_data, zio->io_size,
1795 zio->io_offset, NULL) == 0);
1796 zio_pop_transforms(zio);
1797 return (valid_cksum);
1801 arc_cksum_compute(arc_buf_t *buf)
1803 arc_buf_hdr_t *hdr = buf->b_hdr;
1805 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1808 ASSERT(HDR_HAS_L1HDR(hdr));
1809 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1810 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1811 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1814 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1816 fletcher_2_native(buf->b_data, HDR_GET_LSIZE(hdr), NULL,
1817 hdr->b_l1hdr.b_freeze_cksum);
1818 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1826 typedef struct procctl {
1834 arc_buf_unwatch(arc_buf_t *buf)
1841 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1842 ctl.prwatch.pr_size = 0;
1843 ctl.prwatch.pr_wflags = 0;
1844 result = write(arc_procfd, &ctl, sizeof (ctl));
1845 ASSERT3U(result, ==, sizeof (ctl));
1852 arc_buf_watch(arc_buf_t *buf)
1859 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1860 ctl.prwatch.pr_size = HDR_GET_LSIZE(buf->b_hdr);
1861 ctl.prwatch.pr_wflags = WA_WRITE;
1862 result = write(arc_procfd, &ctl, sizeof (ctl));
1863 ASSERT3U(result, ==, sizeof (ctl));
1867 #endif /* illumos */
1869 static arc_buf_contents_t
1870 arc_buf_type(arc_buf_hdr_t *hdr)
1872 arc_buf_contents_t type;
1873 if (HDR_ISTYPE_METADATA(hdr)) {
1874 type = ARC_BUFC_METADATA;
1876 type = ARC_BUFC_DATA;
1878 VERIFY3U(hdr->b_type, ==, type);
1883 arc_bufc_to_flags(arc_buf_contents_t type)
1887 /* metadata field is 0 if buffer contains normal data */
1889 case ARC_BUFC_METADATA:
1890 return (ARC_FLAG_BUFC_METADATA);
1894 panic("undefined ARC buffer type!");
1895 return ((uint32_t)-1);
1899 arc_buf_thaw(arc_buf_t *buf)
1901 arc_buf_hdr_t *hdr = buf->b_hdr;
1903 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1904 if (hdr->b_l1hdr.b_state != arc_anon)
1905 panic("modifying non-anon buffer!");
1906 if (HDR_IO_IN_PROGRESS(hdr))
1907 panic("modifying buffer while i/o in progress!");
1908 arc_cksum_verify(buf);
1911 ASSERT(HDR_HAS_L1HDR(hdr));
1912 arc_cksum_free(hdr);
1914 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1916 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1917 if (hdr->b_l1hdr.b_thawed != NULL)
1918 kmem_free(hdr->b_l1hdr.b_thawed, 1);
1919 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1923 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1926 arc_buf_unwatch(buf);
1931 arc_buf_freeze(arc_buf_t *buf)
1933 arc_buf_hdr_t *hdr = buf->b_hdr;
1934 kmutex_t *hash_lock;
1936 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1939 hash_lock = HDR_LOCK(hdr);
1940 mutex_enter(hash_lock);
1942 ASSERT(HDR_HAS_L1HDR(hdr));
1943 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
1944 hdr->b_l1hdr.b_state == arc_anon);
1945 arc_cksum_compute(buf);
1946 mutex_exit(hash_lock);
1951 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1952 * the following functions should be used to ensure that the flags are
1953 * updated in a thread-safe way. When manipulating the flags either
1954 * the hash_lock must be held or the hdr must be undiscoverable. This
1955 * ensures that we're not racing with any other threads when updating
1959 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1961 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1962 hdr->b_flags |= flags;
1966 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1968 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1969 hdr->b_flags &= ~flags;
1973 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1974 * done in a special way since we have to clear and set bits
1975 * at the same time. Consumers that wish to set the compression bits
1976 * must use this function to ensure that the flags are updated in
1977 * thread-safe manner.
1980 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1982 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1985 * Holes and embedded blocks will always have a psize = 0 so
1986 * we ignore the compression of the blkptr and set the
1987 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
1988 * Holes and embedded blocks remain anonymous so we don't
1989 * want to uncompress them. Mark them as uncompressed.
1991 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1992 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1993 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
1994 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1995 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1997 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1998 HDR_SET_COMPRESS(hdr, cmp);
1999 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
2000 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
2005 arc_decompress(arc_buf_t *buf)
2007 arc_buf_hdr_t *hdr = buf->b_hdr;
2008 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2011 if (arc_buf_is_shared(buf)) {
2012 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2013 } else if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) {
2015 * The arc_buf_hdr_t is either not compressed or is
2016 * associated with an embedded block or a hole in which
2017 * case they remain anonymous.
2019 IMPLY(HDR_COMPRESSION_ENABLED(hdr), HDR_GET_PSIZE(hdr) == 0 ||
2020 HDR_GET_PSIZE(hdr) == HDR_GET_LSIZE(hdr));
2021 ASSERT(!HDR_SHARED_DATA(hdr));
2022 bcopy(hdr->b_l1hdr.b_pdata, buf->b_data, HDR_GET_LSIZE(hdr));
2024 ASSERT(!HDR_SHARED_DATA(hdr));
2025 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2026 error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2027 hdr->b_l1hdr.b_pdata, buf->b_data, HDR_GET_PSIZE(hdr),
2028 HDR_GET_LSIZE(hdr));
2030 zfs_dbgmsg("hdr %p, compress %d, psize %d, lsize %d",
2031 hdr, HDR_GET_COMPRESS(hdr), HDR_GET_PSIZE(hdr),
2032 HDR_GET_LSIZE(hdr));
2033 return (SET_ERROR(EIO));
2036 if (bswap != DMU_BSWAP_NUMFUNCS) {
2037 ASSERT(!HDR_SHARED_DATA(hdr));
2038 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2039 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2041 arc_cksum_compute(buf);
2046 * Return the size of the block, b_pdata, that is stored in the arc_buf_hdr_t.
2049 arc_hdr_size(arc_buf_hdr_t *hdr)
2053 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2054 HDR_GET_PSIZE(hdr) > 0) {
2055 size = HDR_GET_PSIZE(hdr);
2057 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2058 size = HDR_GET_LSIZE(hdr);
2064 * Increment the amount of evictable space in the arc_state_t's refcount.
2065 * We account for the space used by the hdr and the arc buf individually
2066 * so that we can add and remove them from the refcount individually.
2069 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2071 arc_buf_contents_t type = arc_buf_type(hdr);
2072 uint64_t lsize = HDR_GET_LSIZE(hdr);
2074 ASSERT(HDR_HAS_L1HDR(hdr));
2076 if (GHOST_STATE(state)) {
2077 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2078 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2079 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2080 (void) refcount_add_many(&state->arcs_esize[type], lsize, hdr);
2084 ASSERT(!GHOST_STATE(state));
2085 if (hdr->b_l1hdr.b_pdata != NULL) {
2086 (void) refcount_add_many(&state->arcs_esize[type],
2087 arc_hdr_size(hdr), hdr);
2089 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2090 buf = buf->b_next) {
2091 if (arc_buf_is_shared(buf)) {
2092 ASSERT(ARC_BUF_LAST(buf));
2095 (void) refcount_add_many(&state->arcs_esize[type], lsize, buf);
2100 * Decrement the amount of evictable space in the arc_state_t's refcount.
2101 * We account for the space used by the hdr and the arc buf individually
2102 * so that we can add and remove them from the refcount individually.
2105 arc_evitable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2107 arc_buf_contents_t type = arc_buf_type(hdr);
2108 uint64_t lsize = HDR_GET_LSIZE(hdr);
2110 ASSERT(HDR_HAS_L1HDR(hdr));
2112 if (GHOST_STATE(state)) {
2113 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2114 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2115 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2116 (void) refcount_remove_many(&state->arcs_esize[type],
2121 ASSERT(!GHOST_STATE(state));
2122 if (hdr->b_l1hdr.b_pdata != NULL) {
2123 (void) refcount_remove_many(&state->arcs_esize[type],
2124 arc_hdr_size(hdr), hdr);
2126 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2127 buf = buf->b_next) {
2128 if (arc_buf_is_shared(buf)) {
2129 ASSERT(ARC_BUF_LAST(buf));
2132 (void) refcount_remove_many(&state->arcs_esize[type],
2138 * Add a reference to this hdr indicating that someone is actively
2139 * referencing that memory. When the refcount transitions from 0 to 1,
2140 * we remove it from the respective arc_state_t list to indicate that
2141 * it is not evictable.
2144 add_reference(arc_buf_hdr_t *hdr, void *tag)
2146 ASSERT(HDR_HAS_L1HDR(hdr));
2147 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2148 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2149 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2150 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2153 arc_state_t *state = hdr->b_l1hdr.b_state;
2155 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2156 (state != arc_anon)) {
2157 /* We don't use the L2-only state list. */
2158 if (state != arc_l2c_only) {
2159 multilist_remove(&state->arcs_list[arc_buf_type(hdr)],
2161 arc_evitable_space_decrement(hdr, state);
2163 /* remove the prefetch flag if we get a reference */
2164 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2169 * Remove a reference from this hdr. When the reference transitions from
2170 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2171 * list making it eligible for eviction.
2174 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2177 arc_state_t *state = hdr->b_l1hdr.b_state;
2179 ASSERT(HDR_HAS_L1HDR(hdr));
2180 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2181 ASSERT(!GHOST_STATE(state));
2184 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2185 * check to prevent usage of the arc_l2c_only list.
2187 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2188 (state != arc_anon)) {
2189 multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr);
2190 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2191 arc_evictable_space_increment(hdr, state);
2197 * Move the supplied buffer to the indicated state. The hash lock
2198 * for the buffer must be held by the caller.
2201 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2202 kmutex_t *hash_lock)
2204 arc_state_t *old_state;
2207 boolean_t update_old, update_new;
2208 arc_buf_contents_t buftype = arc_buf_type(hdr);
2211 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2212 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2213 * L1 hdr doesn't always exist when we change state to arc_anon before
2214 * destroying a header, in which case reallocating to add the L1 hdr is
2217 if (HDR_HAS_L1HDR(hdr)) {
2218 old_state = hdr->b_l1hdr.b_state;
2219 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2220 bufcnt = hdr->b_l1hdr.b_bufcnt;
2221 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pdata != NULL);
2223 old_state = arc_l2c_only;
2226 update_old = B_FALSE;
2228 update_new = update_old;
2230 ASSERT(MUTEX_HELD(hash_lock));
2231 ASSERT3P(new_state, !=, old_state);
2232 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2233 ASSERT(old_state != arc_anon || bufcnt <= 1);
2236 * If this buffer is evictable, transfer it from the
2237 * old state list to the new state list.
2240 if (old_state != arc_anon && old_state != arc_l2c_only) {
2241 ASSERT(HDR_HAS_L1HDR(hdr));
2242 multilist_remove(&old_state->arcs_list[buftype], hdr);
2244 if (GHOST_STATE(old_state)) {
2246 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2247 update_old = B_TRUE;
2249 arc_evitable_space_decrement(hdr, old_state);
2251 if (new_state != arc_anon && new_state != arc_l2c_only) {
2254 * An L1 header always exists here, since if we're
2255 * moving to some L1-cached state (i.e. not l2c_only or
2256 * anonymous), we realloc the header to add an L1hdr
2259 ASSERT(HDR_HAS_L1HDR(hdr));
2260 multilist_insert(&new_state->arcs_list[buftype], hdr);
2262 if (GHOST_STATE(new_state)) {
2264 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2265 update_new = B_TRUE;
2267 arc_evictable_space_increment(hdr, new_state);
2271 ASSERT(!HDR_EMPTY(hdr));
2272 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2273 buf_hash_remove(hdr);
2275 /* adjust state sizes (ignore arc_l2c_only) */
2277 if (update_new && new_state != arc_l2c_only) {
2278 ASSERT(HDR_HAS_L1HDR(hdr));
2279 if (GHOST_STATE(new_state)) {
2283 * When moving a header to a ghost state, we first
2284 * remove all arc buffers. Thus, we'll have a
2285 * bufcnt of zero, and no arc buffer to use for
2286 * the reference. As a result, we use the arc
2287 * header pointer for the reference.
2289 (void) refcount_add_many(&new_state->arcs_size,
2290 HDR_GET_LSIZE(hdr), hdr);
2291 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2293 uint32_t buffers = 0;
2296 * Each individual buffer holds a unique reference,
2297 * thus we must remove each of these references one
2300 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2301 buf = buf->b_next) {
2302 ASSERT3U(bufcnt, !=, 0);
2306 * When the arc_buf_t is sharing the data
2307 * block with the hdr, the owner of the
2308 * reference belongs to the hdr. Only
2309 * add to the refcount if the arc_buf_t is
2312 if (arc_buf_is_shared(buf)) {
2313 ASSERT(ARC_BUF_LAST(buf));
2317 (void) refcount_add_many(&new_state->arcs_size,
2318 HDR_GET_LSIZE(hdr), buf);
2320 ASSERT3U(bufcnt, ==, buffers);
2322 if (hdr->b_l1hdr.b_pdata != NULL) {
2323 (void) refcount_add_many(&new_state->arcs_size,
2324 arc_hdr_size(hdr), hdr);
2326 ASSERT(GHOST_STATE(old_state));
2331 if (update_old && old_state != arc_l2c_only) {
2332 ASSERT(HDR_HAS_L1HDR(hdr));
2333 if (GHOST_STATE(old_state)) {
2337 * When moving a header off of a ghost state,
2338 * the header will not contain any arc buffers.
2339 * We use the arc header pointer for the reference
2340 * which is exactly what we did when we put the
2341 * header on the ghost state.
2344 (void) refcount_remove_many(&old_state->arcs_size,
2345 HDR_GET_LSIZE(hdr), hdr);
2346 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2348 uint32_t buffers = 0;
2351 * Each individual buffer holds a unique reference,
2352 * thus we must remove each of these references one
2355 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2356 buf = buf->b_next) {
2357 ASSERT3P(bufcnt, !=, 0);
2361 * When the arc_buf_t is sharing the data
2362 * block with the hdr, the owner of the
2363 * reference belongs to the hdr. Only
2364 * add to the refcount if the arc_buf_t is
2367 if (arc_buf_is_shared(buf)) {
2368 ASSERT(ARC_BUF_LAST(buf));
2372 (void) refcount_remove_many(
2373 &old_state->arcs_size, HDR_GET_LSIZE(hdr),
2376 ASSERT3U(bufcnt, ==, buffers);
2377 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
2378 (void) refcount_remove_many(
2379 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2383 if (HDR_HAS_L1HDR(hdr))
2384 hdr->b_l1hdr.b_state = new_state;
2387 * L2 headers should never be on the L2 state list since they don't
2388 * have L1 headers allocated.
2390 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2391 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2395 arc_space_consume(uint64_t space, arc_space_type_t type)
2397 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2400 case ARC_SPACE_DATA:
2401 ARCSTAT_INCR(arcstat_data_size, space);
2403 case ARC_SPACE_META:
2404 ARCSTAT_INCR(arcstat_metadata_size, space);
2406 case ARC_SPACE_OTHER:
2407 ARCSTAT_INCR(arcstat_other_size, space);
2409 case ARC_SPACE_HDRS:
2410 ARCSTAT_INCR(arcstat_hdr_size, space);
2412 case ARC_SPACE_L2HDRS:
2413 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
2417 if (type != ARC_SPACE_DATA)
2418 ARCSTAT_INCR(arcstat_meta_used, space);
2420 atomic_add_64(&arc_size, space);
2424 arc_space_return(uint64_t space, arc_space_type_t type)
2426 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2429 case ARC_SPACE_DATA:
2430 ARCSTAT_INCR(arcstat_data_size, -space);
2432 case ARC_SPACE_META:
2433 ARCSTAT_INCR(arcstat_metadata_size, -space);
2435 case ARC_SPACE_OTHER:
2436 ARCSTAT_INCR(arcstat_other_size, -space);
2438 case ARC_SPACE_HDRS:
2439 ARCSTAT_INCR(arcstat_hdr_size, -space);
2441 case ARC_SPACE_L2HDRS:
2442 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
2446 if (type != ARC_SPACE_DATA) {
2447 ASSERT(arc_meta_used >= space);
2448 if (arc_meta_max < arc_meta_used)
2449 arc_meta_max = arc_meta_used;
2450 ARCSTAT_INCR(arcstat_meta_used, -space);
2453 ASSERT(arc_size >= space);
2454 atomic_add_64(&arc_size, -space);
2458 * Allocate an initial buffer for this hdr, subsequent buffers will
2459 * use arc_buf_clone().
2462 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag)
2466 ASSERT(HDR_HAS_L1HDR(hdr));
2467 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2468 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2469 hdr->b_type == ARC_BUFC_METADATA);
2471 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2472 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2473 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2475 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2480 add_reference(hdr, tag);
2483 * We're about to change the hdr's b_flags. We must either
2484 * hold the hash_lock or be undiscoverable.
2486 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2489 * If the hdr's data can be shared (no byteswapping, hdr is
2490 * uncompressed, hdr's data is not currently being written to the
2491 * L2ARC write) then we share the data buffer and set the appropriate
2492 * bit in the hdr's b_flags to indicate the hdr is sharing it's
2493 * b_pdata with the arc_buf_t. Otherwise, we allocate a new buffer to
2494 * store the buf's data.
2496 if (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2497 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF && !HDR_L2_WRITING(hdr)) {
2498 buf->b_data = hdr->b_l1hdr.b_pdata;
2499 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2501 buf->b_data = arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2502 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2503 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2505 VERIFY3P(buf->b_data, !=, NULL);
2507 hdr->b_l1hdr.b_buf = buf;
2508 hdr->b_l1hdr.b_bufcnt += 1;
2514 * Used when allocating additional buffers.
2517 arc_buf_clone(arc_buf_t *from)
2520 arc_buf_hdr_t *hdr = from->b_hdr;
2521 uint64_t size = HDR_GET_LSIZE(hdr);
2523 ASSERT(HDR_HAS_L1HDR(hdr));
2524 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2526 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2529 buf->b_next = hdr->b_l1hdr.b_buf;
2530 hdr->b_l1hdr.b_buf = buf;
2531 buf->b_data = arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2532 bcopy(from->b_data, buf->b_data, size);
2533 hdr->b_l1hdr.b_bufcnt += 1;
2535 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2539 static char *arc_onloan_tag = "onloan";
2542 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2543 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2544 * buffers must be returned to the arc before they can be used by the DMU or
2548 arc_loan_buf(spa_t *spa, int size)
2552 buf = arc_alloc_buf(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
2554 atomic_add_64(&arc_loaned_bytes, size);
2559 * Return a loaned arc buffer to the arc.
2562 arc_return_buf(arc_buf_t *buf, void *tag)
2564 arc_buf_hdr_t *hdr = buf->b_hdr;
2566 ASSERT3P(buf->b_data, !=, NULL);
2567 ASSERT(HDR_HAS_L1HDR(hdr));
2568 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2569 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2571 atomic_add_64(&arc_loaned_bytes, -HDR_GET_LSIZE(hdr));
2574 /* Detach an arc_buf from a dbuf (tag) */
2576 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2578 arc_buf_hdr_t *hdr = buf->b_hdr;
2580 ASSERT3P(buf->b_data, !=, NULL);
2581 ASSERT(HDR_HAS_L1HDR(hdr));
2582 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2583 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2585 atomic_add_64(&arc_loaned_bytes, HDR_GET_LSIZE(hdr));
2589 l2arc_free_data_on_write(void *data, size_t size, arc_buf_contents_t type)
2591 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2593 df->l2df_data = data;
2594 df->l2df_size = size;
2595 df->l2df_type = type;
2596 mutex_enter(&l2arc_free_on_write_mtx);
2597 list_insert_head(l2arc_free_on_write, df);
2598 mutex_exit(&l2arc_free_on_write_mtx);
2602 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
2604 arc_state_t *state = hdr->b_l1hdr.b_state;
2605 arc_buf_contents_t type = arc_buf_type(hdr);
2606 uint64_t size = arc_hdr_size(hdr);
2608 /* protected by hash lock, if in the hash table */
2609 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2610 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2611 ASSERT(state != arc_anon && state != arc_l2c_only);
2613 (void) refcount_remove_many(&state->arcs_esize[type],
2616 (void) refcount_remove_many(&state->arcs_size, size, hdr);
2618 l2arc_free_data_on_write(hdr->b_l1hdr.b_pdata, size, type);
2622 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2623 * data buffer, we transfer the refcount ownership to the hdr and update
2624 * the appropriate kstats.
2627 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2629 arc_state_t *state = hdr->b_l1hdr.b_state;
2631 ASSERT(!HDR_SHARED_DATA(hdr));
2632 ASSERT(!arc_buf_is_shared(buf));
2633 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2634 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2637 * Start sharing the data buffer. We transfer the
2638 * refcount ownership to the hdr since it always owns
2639 * the refcount whenever an arc_buf_t is shared.
2641 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
2642 hdr->b_l1hdr.b_pdata = buf->b_data;
2643 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2646 * Since we've transferred ownership to the hdr we need
2647 * to increment its compressed and uncompressed kstats and
2648 * decrement the overhead size.
2650 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2651 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2652 ARCSTAT_INCR(arcstat_overhead_size, -HDR_GET_LSIZE(hdr));
2656 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2658 arc_state_t *state = hdr->b_l1hdr.b_state;
2660 ASSERT(HDR_SHARED_DATA(hdr));
2661 ASSERT(arc_buf_is_shared(buf));
2662 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
2663 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2666 * We are no longer sharing this buffer so we need
2667 * to transfer its ownership to the rightful owner.
2669 refcount_transfer_ownership(&state->arcs_size, hdr, buf);
2670 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2671 hdr->b_l1hdr.b_pdata = NULL;
2674 * Since the buffer is no longer shared between
2675 * the arc buf and the hdr, count it as overhead.
2677 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2678 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2679 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2683 * Free up buf->b_data and if 'remove' is set, then pull the
2684 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
2687 arc_buf_destroy_impl(arc_buf_t *buf, boolean_t remove)
2690 arc_buf_hdr_t *hdr = buf->b_hdr;
2691 uint64_t size = HDR_GET_LSIZE(hdr);
2692 boolean_t destroyed_buf_is_shared = arc_buf_is_shared(buf);
2695 * Free up the data associated with the buf but only
2696 * if we're not sharing this with the hdr. If we are sharing
2697 * it with the hdr, then hdr will have performed the allocation
2698 * so allow it to do the free.
2700 if (buf->b_data != NULL) {
2702 * We're about to change the hdr's b_flags. We must either
2703 * hold the hash_lock or be undiscoverable.
2705 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2707 arc_cksum_verify(buf);
2709 arc_buf_unwatch(buf);
2712 if (destroyed_buf_is_shared) {
2713 ASSERT(ARC_BUF_LAST(buf));
2714 ASSERT(HDR_SHARED_DATA(hdr));
2715 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2717 arc_free_data_buf(hdr, buf->b_data, size, buf);
2718 ARCSTAT_INCR(arcstat_overhead_size, -size);
2722 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
2723 hdr->b_l1hdr.b_bufcnt -= 1;
2726 /* only remove the buf if requested */
2730 /* remove the buf from the hdr list */
2731 arc_buf_t *lastbuf = NULL;
2732 bufp = &hdr->b_l1hdr.b_buf;
2733 while (*bufp != NULL) {
2735 *bufp = buf->b_next;
2738 * If we've removed a buffer in the middle of
2739 * the list then update the lastbuf and update
2742 if (*bufp != NULL) {
2744 bufp = &(*bufp)->b_next;
2748 ASSERT3P(lastbuf, !=, buf);
2751 * If the current arc_buf_t is sharing its data
2752 * buffer with the hdr, then reassign the hdr's
2753 * b_pdata to share it with the new buffer at the end
2754 * of the list. The shared buffer is always the last one
2755 * on the hdr's buffer list.
2757 if (destroyed_buf_is_shared && lastbuf != NULL) {
2758 ASSERT(ARC_BUF_LAST(buf));
2759 ASSERT(ARC_BUF_LAST(lastbuf));
2760 VERIFY(!arc_buf_is_shared(lastbuf));
2762 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
2763 arc_hdr_free_pdata(hdr);
2766 * We must setup a new shared block between the
2767 * last buffer and the hdr. The data would have
2768 * been allocated by the arc buf so we need to transfer
2769 * ownership to the hdr since it's now being shared.
2771 arc_share_buf(hdr, lastbuf);
2772 } else if (HDR_SHARED_DATA(hdr)) {
2773 ASSERT(arc_buf_is_shared(lastbuf));
2776 if (hdr->b_l1hdr.b_bufcnt == 0)
2777 arc_cksum_free(hdr);
2779 /* clean up the buf */
2781 kmem_cache_free(buf_cache, buf);
2785 arc_hdr_alloc_pdata(arc_buf_hdr_t *hdr)
2787 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2788 ASSERT(HDR_HAS_L1HDR(hdr));
2789 ASSERT(!HDR_SHARED_DATA(hdr));
2791 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2792 hdr->b_l1hdr.b_pdata = arc_get_data_buf(hdr, arc_hdr_size(hdr), hdr);
2793 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
2794 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
2796 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2797 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2801 arc_hdr_free_pdata(arc_buf_hdr_t *hdr)
2803 ASSERT(HDR_HAS_L1HDR(hdr));
2804 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
2807 * If the hdr is currently being written to the l2arc then
2808 * we defer freeing the data by adding it to the l2arc_free_on_write
2809 * list. The l2arc will free the data once it's finished
2810 * writing it to the l2arc device.
2812 if (HDR_L2_WRITING(hdr)) {
2813 arc_hdr_free_on_write(hdr);
2814 ARCSTAT_BUMP(arcstat_l2_free_on_write);
2816 arc_free_data_buf(hdr, hdr->b_l1hdr.b_pdata,
2817 arc_hdr_size(hdr), hdr);
2819 hdr->b_l1hdr.b_pdata = NULL;
2820 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
2822 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2823 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2826 static arc_buf_hdr_t *
2827 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
2828 enum zio_compress compress, arc_buf_contents_t type)
2832 ASSERT3U(lsize, >, 0);
2833 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
2835 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
2836 ASSERT(HDR_EMPTY(hdr));
2837 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
2838 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
2839 HDR_SET_PSIZE(hdr, psize);
2840 HDR_SET_LSIZE(hdr, lsize);
2844 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
2845 arc_hdr_set_compress(hdr, compress);
2847 hdr->b_l1hdr.b_state = arc_anon;
2848 hdr->b_l1hdr.b_arc_access = 0;
2849 hdr->b_l1hdr.b_bufcnt = 0;
2850 hdr->b_l1hdr.b_buf = NULL;
2853 * Allocate the hdr's buffer. This will contain either
2854 * the compressed or uncompressed data depending on the block
2855 * it references and compressed arc enablement.
2857 arc_hdr_alloc_pdata(hdr);
2858 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2864 * Transition between the two allocation states for the arc_buf_hdr struct.
2865 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
2866 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
2867 * version is used when a cache buffer is only in the L2ARC in order to reduce
2870 static arc_buf_hdr_t *
2871 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
2873 ASSERT(HDR_HAS_L2HDR(hdr));
2875 arc_buf_hdr_t *nhdr;
2876 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2878 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
2879 (old == hdr_l2only_cache && new == hdr_full_cache));
2881 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
2883 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
2884 buf_hash_remove(hdr);
2886 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
2888 if (new == hdr_full_cache) {
2889 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
2891 * arc_access and arc_change_state need to be aware that a
2892 * header has just come out of L2ARC, so we set its state to
2893 * l2c_only even though it's about to change.
2895 nhdr->b_l1hdr.b_state = arc_l2c_only;
2897 /* Verify previous threads set to NULL before freeing */
2898 ASSERT3P(nhdr->b_l1hdr.b_pdata, ==, NULL);
2900 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2901 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2902 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
2905 * If we've reached here, We must have been called from
2906 * arc_evict_hdr(), as such we should have already been
2907 * removed from any ghost list we were previously on
2908 * (which protects us from racing with arc_evict_state),
2909 * thus no locking is needed during this check.
2911 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
2914 * A buffer must not be moved into the arc_l2c_only
2915 * state if it's not finished being written out to the
2916 * l2arc device. Otherwise, the b_l1hdr.b_pdata field
2917 * might try to be accessed, even though it was removed.
2919 VERIFY(!HDR_L2_WRITING(hdr));
2920 VERIFY3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2923 if (hdr->b_l1hdr.b_thawed != NULL) {
2924 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2925 hdr->b_l1hdr.b_thawed = NULL;
2929 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
2932 * The header has been reallocated so we need to re-insert it into any
2935 (void) buf_hash_insert(nhdr, NULL);
2937 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
2939 mutex_enter(&dev->l2ad_mtx);
2942 * We must place the realloc'ed header back into the list at
2943 * the same spot. Otherwise, if it's placed earlier in the list,
2944 * l2arc_write_buffers() could find it during the function's
2945 * write phase, and try to write it out to the l2arc.
2947 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
2948 list_remove(&dev->l2ad_buflist, hdr);
2950 mutex_exit(&dev->l2ad_mtx);
2953 * Since we're using the pointer address as the tag when
2954 * incrementing and decrementing the l2ad_alloc refcount, we
2955 * must remove the old pointer (that we're about to destroy) and
2956 * add the new pointer to the refcount. Otherwise we'd remove
2957 * the wrong pointer address when calling arc_hdr_destroy() later.
2960 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
2961 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
2963 buf_discard_identity(hdr);
2964 kmem_cache_free(old, hdr);
2970 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
2971 * The buf is returned thawed since we expect the consumer to modify it.
2974 arc_alloc_buf(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type)
2976 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
2977 ZIO_COMPRESS_OFF, type);
2978 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
2979 arc_buf_t *buf = arc_buf_alloc_impl(hdr, tag);
2985 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
2987 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
2988 l2arc_dev_t *dev = l2hdr->b_dev;
2989 uint64_t asize = arc_hdr_size(hdr);
2991 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
2992 ASSERT(HDR_HAS_L2HDR(hdr));
2994 list_remove(&dev->l2ad_buflist, hdr);
2996 ARCSTAT_INCR(arcstat_l2_asize, -asize);
2997 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr));
2999 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
3001 (void) refcount_remove_many(&dev->l2ad_alloc, asize, hdr);
3002 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3006 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3008 if (HDR_HAS_L1HDR(hdr)) {
3009 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3010 hdr->b_l1hdr.b_bufcnt > 0);
3011 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3012 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3014 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3015 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3017 if (!HDR_EMPTY(hdr))
3018 buf_discard_identity(hdr);
3020 if (HDR_HAS_L2HDR(hdr)) {
3021 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3022 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3025 mutex_enter(&dev->l2ad_mtx);
3028 * Even though we checked this conditional above, we
3029 * need to check this again now that we have the
3030 * l2ad_mtx. This is because we could be racing with
3031 * another thread calling l2arc_evict() which might have
3032 * destroyed this header's L2 portion as we were waiting
3033 * to acquire the l2ad_mtx. If that happens, we don't
3034 * want to re-destroy the header's L2 portion.
3036 if (HDR_HAS_L2HDR(hdr)) {
3038 arc_hdr_l2hdr_destroy(hdr);
3042 mutex_exit(&dev->l2ad_mtx);
3045 if (HDR_HAS_L1HDR(hdr)) {
3046 arc_cksum_free(hdr);
3048 while (hdr->b_l1hdr.b_buf != NULL)
3049 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf, B_TRUE);
3052 if (hdr->b_l1hdr.b_thawed != NULL) {
3053 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3054 hdr->b_l1hdr.b_thawed = NULL;
3058 if (hdr->b_l1hdr.b_pdata != NULL) {
3059 arc_hdr_free_pdata(hdr);
3063 ASSERT3P(hdr->b_hash_next, ==, NULL);
3064 if (HDR_HAS_L1HDR(hdr)) {
3065 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3066 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3067 kmem_cache_free(hdr_full_cache, hdr);
3069 kmem_cache_free(hdr_l2only_cache, hdr);
3074 arc_buf_destroy(arc_buf_t *buf, void* tag)
3076 arc_buf_hdr_t *hdr = buf->b_hdr;
3077 kmutex_t *hash_lock = HDR_LOCK(hdr);
3079 if (hdr->b_l1hdr.b_state == arc_anon) {
3080 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3081 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3082 VERIFY0(remove_reference(hdr, NULL, tag));
3083 arc_hdr_destroy(hdr);
3087 mutex_enter(hash_lock);
3088 ASSERT3P(hdr, ==, buf->b_hdr);
3089 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3090 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3091 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3092 ASSERT3P(buf->b_data, !=, NULL);
3094 (void) remove_reference(hdr, hash_lock, tag);
3095 arc_buf_destroy_impl(buf, B_TRUE);
3096 mutex_exit(hash_lock);
3100 arc_buf_size(arc_buf_t *buf)
3102 return (HDR_GET_LSIZE(buf->b_hdr));
3106 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3107 * state of the header is dependent on it's state prior to entering this
3108 * function. The following transitions are possible:
3110 * - arc_mru -> arc_mru_ghost
3111 * - arc_mfu -> arc_mfu_ghost
3112 * - arc_mru_ghost -> arc_l2c_only
3113 * - arc_mru_ghost -> deleted
3114 * - arc_mfu_ghost -> arc_l2c_only
3115 * - arc_mfu_ghost -> deleted
3118 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3120 arc_state_t *evicted_state, *state;
3121 int64_t bytes_evicted = 0;
3123 ASSERT(MUTEX_HELD(hash_lock));
3124 ASSERT(HDR_HAS_L1HDR(hdr));
3126 state = hdr->b_l1hdr.b_state;
3127 if (GHOST_STATE(state)) {
3128 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3129 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3132 * l2arc_write_buffers() relies on a header's L1 portion
3133 * (i.e. its b_pdata field) during its write phase.
3134 * Thus, we cannot push a header onto the arc_l2c_only
3135 * state (removing it's L1 piece) until the header is
3136 * done being written to the l2arc.
3138 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3139 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3140 return (bytes_evicted);
3143 ARCSTAT_BUMP(arcstat_deleted);
3144 bytes_evicted += HDR_GET_LSIZE(hdr);
3146 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3148 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
3149 if (HDR_HAS_L2HDR(hdr)) {
3150 ASSERT(hdr->b_l1hdr.b_pdata == NULL);
3152 * This buffer is cached on the 2nd Level ARC;
3153 * don't destroy the header.
3155 arc_change_state(arc_l2c_only, hdr, hash_lock);
3157 * dropping from L1+L2 cached to L2-only,
3158 * realloc to remove the L1 header.
3160 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3163 ASSERT(hdr->b_l1hdr.b_pdata == NULL);
3164 arc_change_state(arc_anon, hdr, hash_lock);
3165 arc_hdr_destroy(hdr);
3167 return (bytes_evicted);
3170 ASSERT(state == arc_mru || state == arc_mfu);
3171 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3173 /* prefetch buffers have a minimum lifespan */
3174 if (HDR_IO_IN_PROGRESS(hdr) ||
3175 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3176 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3177 arc_min_prefetch_lifespan)) {
3178 ARCSTAT_BUMP(arcstat_evict_skip);
3179 return (bytes_evicted);
3182 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3183 while (hdr->b_l1hdr.b_buf) {
3184 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3185 if (!mutex_tryenter(&buf->b_evict_lock)) {
3186 ARCSTAT_BUMP(arcstat_mutex_miss);
3189 if (buf->b_data != NULL)
3190 bytes_evicted += HDR_GET_LSIZE(hdr);
3191 mutex_exit(&buf->b_evict_lock);
3192 arc_buf_destroy_impl(buf, B_TRUE);
3195 if (HDR_HAS_L2HDR(hdr)) {
3196 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3198 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3199 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3200 HDR_GET_LSIZE(hdr));
3202 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3203 HDR_GET_LSIZE(hdr));
3207 if (hdr->b_l1hdr.b_bufcnt == 0) {
3208 arc_cksum_free(hdr);
3210 bytes_evicted += arc_hdr_size(hdr);
3213 * If this hdr is being evicted and has a compressed
3214 * buffer then we discard it here before we change states.
3215 * This ensures that the accounting is updated correctly
3216 * in arc_free_data_buf().
3218 arc_hdr_free_pdata(hdr);
3220 arc_change_state(evicted_state, hdr, hash_lock);
3221 ASSERT(HDR_IN_HASH_TABLE(hdr));
3222 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3223 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3226 return (bytes_evicted);
3230 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3231 uint64_t spa, int64_t bytes)
3233 multilist_sublist_t *mls;
3234 uint64_t bytes_evicted = 0;
3236 kmutex_t *hash_lock;
3237 int evict_count = 0;
3239 ASSERT3P(marker, !=, NULL);
3240 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3242 mls = multilist_sublist_lock(ml, idx);
3244 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3245 hdr = multilist_sublist_prev(mls, marker)) {
3246 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3247 (evict_count >= zfs_arc_evict_batch_limit))
3251 * To keep our iteration location, move the marker
3252 * forward. Since we're not holding hdr's hash lock, we
3253 * must be very careful and not remove 'hdr' from the
3254 * sublist. Otherwise, other consumers might mistake the
3255 * 'hdr' as not being on a sublist when they call the
3256 * multilist_link_active() function (they all rely on
3257 * the hash lock protecting concurrent insertions and
3258 * removals). multilist_sublist_move_forward() was
3259 * specifically implemented to ensure this is the case
3260 * (only 'marker' will be removed and re-inserted).
3262 multilist_sublist_move_forward(mls, marker);
3265 * The only case where the b_spa field should ever be
3266 * zero, is the marker headers inserted by
3267 * arc_evict_state(). It's possible for multiple threads
3268 * to be calling arc_evict_state() concurrently (e.g.
3269 * dsl_pool_close() and zio_inject_fault()), so we must
3270 * skip any markers we see from these other threads.
3272 if (hdr->b_spa == 0)
3275 /* we're only interested in evicting buffers of a certain spa */
3276 if (spa != 0 && hdr->b_spa != spa) {
3277 ARCSTAT_BUMP(arcstat_evict_skip);
3281 hash_lock = HDR_LOCK(hdr);
3284 * We aren't calling this function from any code path
3285 * that would already be holding a hash lock, so we're
3286 * asserting on this assumption to be defensive in case
3287 * this ever changes. Without this check, it would be
3288 * possible to incorrectly increment arcstat_mutex_miss
3289 * below (e.g. if the code changed such that we called
3290 * this function with a hash lock held).
3292 ASSERT(!MUTEX_HELD(hash_lock));
3294 if (mutex_tryenter(hash_lock)) {
3295 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3296 mutex_exit(hash_lock);
3298 bytes_evicted += evicted;
3301 * If evicted is zero, arc_evict_hdr() must have
3302 * decided to skip this header, don't increment
3303 * evict_count in this case.
3309 * If arc_size isn't overflowing, signal any
3310 * threads that might happen to be waiting.
3312 * For each header evicted, we wake up a single
3313 * thread. If we used cv_broadcast, we could
3314 * wake up "too many" threads causing arc_size
3315 * to significantly overflow arc_c; since
3316 * arc_get_data_buf() doesn't check for overflow
3317 * when it's woken up (it doesn't because it's
3318 * possible for the ARC to be overflowing while
3319 * full of un-evictable buffers, and the
3320 * function should proceed in this case).
3322 * If threads are left sleeping, due to not
3323 * using cv_broadcast, they will be woken up
3324 * just before arc_reclaim_thread() sleeps.
3326 mutex_enter(&arc_reclaim_lock);
3327 if (!arc_is_overflowing())
3328 cv_signal(&arc_reclaim_waiters_cv);
3329 mutex_exit(&arc_reclaim_lock);
3331 ARCSTAT_BUMP(arcstat_mutex_miss);
3335 multilist_sublist_unlock(mls);
3337 return (bytes_evicted);
3341 * Evict buffers from the given arc state, until we've removed the
3342 * specified number of bytes. Move the removed buffers to the
3343 * appropriate evict state.
3345 * This function makes a "best effort". It skips over any buffers
3346 * it can't get a hash_lock on, and so, may not catch all candidates.
3347 * It may also return without evicting as much space as requested.
3349 * If bytes is specified using the special value ARC_EVICT_ALL, this
3350 * will evict all available (i.e. unlocked and evictable) buffers from
3351 * the given arc state; which is used by arc_flush().
3354 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3355 arc_buf_contents_t type)
3357 uint64_t total_evicted = 0;
3358 multilist_t *ml = &state->arcs_list[type];
3360 arc_buf_hdr_t **markers;
3362 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3364 num_sublists = multilist_get_num_sublists(ml);
3367 * If we've tried to evict from each sublist, made some
3368 * progress, but still have not hit the target number of bytes
3369 * to evict, we want to keep trying. The markers allow us to
3370 * pick up where we left off for each individual sublist, rather
3371 * than starting from the tail each time.
3373 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3374 for (int i = 0; i < num_sublists; i++) {
3375 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3378 * A b_spa of 0 is used to indicate that this header is
3379 * a marker. This fact is used in arc_adjust_type() and
3380 * arc_evict_state_impl().
3382 markers[i]->b_spa = 0;
3384 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3385 multilist_sublist_insert_tail(mls, markers[i]);
3386 multilist_sublist_unlock(mls);
3390 * While we haven't hit our target number of bytes to evict, or
3391 * we're evicting all available buffers.
3393 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3395 * Start eviction using a randomly selected sublist,
3396 * this is to try and evenly balance eviction across all
3397 * sublists. Always starting at the same sublist
3398 * (e.g. index 0) would cause evictions to favor certain
3399 * sublists over others.
3401 int sublist_idx = multilist_get_random_index(ml);
3402 uint64_t scan_evicted = 0;
3404 for (int i = 0; i < num_sublists; i++) {
3405 uint64_t bytes_remaining;
3406 uint64_t bytes_evicted;
3408 if (bytes == ARC_EVICT_ALL)
3409 bytes_remaining = ARC_EVICT_ALL;
3410 else if (total_evicted < bytes)
3411 bytes_remaining = bytes - total_evicted;
3415 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3416 markers[sublist_idx], spa, bytes_remaining);
3418 scan_evicted += bytes_evicted;
3419 total_evicted += bytes_evicted;
3421 /* we've reached the end, wrap to the beginning */
3422 if (++sublist_idx >= num_sublists)
3427 * If we didn't evict anything during this scan, we have
3428 * no reason to believe we'll evict more during another
3429 * scan, so break the loop.
3431 if (scan_evicted == 0) {
3432 /* This isn't possible, let's make that obvious */
3433 ASSERT3S(bytes, !=, 0);
3436 * When bytes is ARC_EVICT_ALL, the only way to
3437 * break the loop is when scan_evicted is zero.
3438 * In that case, we actually have evicted enough,
3439 * so we don't want to increment the kstat.
3441 if (bytes != ARC_EVICT_ALL) {
3442 ASSERT3S(total_evicted, <, bytes);
3443 ARCSTAT_BUMP(arcstat_evict_not_enough);
3450 for (int i = 0; i < num_sublists; i++) {
3451 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3452 multilist_sublist_remove(mls, markers[i]);
3453 multilist_sublist_unlock(mls);
3455 kmem_cache_free(hdr_full_cache, markers[i]);
3457 kmem_free(markers, sizeof (*markers) * num_sublists);
3459 return (total_evicted);
3463 * Flush all "evictable" data of the given type from the arc state
3464 * specified. This will not evict any "active" buffers (i.e. referenced).
3466 * When 'retry' is set to B_FALSE, the function will make a single pass
3467 * over the state and evict any buffers that it can. Since it doesn't
3468 * continually retry the eviction, it might end up leaving some buffers
3469 * in the ARC due to lock misses.
3471 * When 'retry' is set to B_TRUE, the function will continually retry the
3472 * eviction until *all* evictable buffers have been removed from the
3473 * state. As a result, if concurrent insertions into the state are
3474 * allowed (e.g. if the ARC isn't shutting down), this function might
3475 * wind up in an infinite loop, continually trying to evict buffers.
3478 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
3481 uint64_t evicted = 0;
3483 while (refcount_count(&state->arcs_esize[type]) != 0) {
3484 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
3494 * Evict the specified number of bytes from the state specified,
3495 * restricting eviction to the spa and type given. This function
3496 * prevents us from trying to evict more from a state's list than
3497 * is "evictable", and to skip evicting altogether when passed a
3498 * negative value for "bytes". In contrast, arc_evict_state() will
3499 * evict everything it can, when passed a negative value for "bytes".
3502 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3503 arc_buf_contents_t type)
3507 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
3508 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
3509 return (arc_evict_state(state, spa, delta, type));
3516 * Evict metadata buffers from the cache, such that arc_meta_used is
3517 * capped by the arc_meta_limit tunable.
3520 arc_adjust_meta(void)
3522 uint64_t total_evicted = 0;
3526 * If we're over the meta limit, we want to evict enough
3527 * metadata to get back under the meta limit. We don't want to
3528 * evict so much that we drop the MRU below arc_p, though. If
3529 * we're over the meta limit more than we're over arc_p, we
3530 * evict some from the MRU here, and some from the MFU below.
3532 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3533 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3534 refcount_count(&arc_mru->arcs_size) - arc_p));
3536 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3539 * Similar to the above, we want to evict enough bytes to get us
3540 * below the meta limit, but not so much as to drop us below the
3541 * space alloted to the MFU (which is defined as arc_c - arc_p).
3543 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3544 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
3546 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3548 return (total_evicted);
3552 * Return the type of the oldest buffer in the given arc state
3554 * This function will select a random sublist of type ARC_BUFC_DATA and
3555 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3556 * is compared, and the type which contains the "older" buffer will be
3559 static arc_buf_contents_t
3560 arc_adjust_type(arc_state_t *state)
3562 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
3563 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
3564 int data_idx = multilist_get_random_index(data_ml);
3565 int meta_idx = multilist_get_random_index(meta_ml);
3566 multilist_sublist_t *data_mls;
3567 multilist_sublist_t *meta_mls;
3568 arc_buf_contents_t type;
3569 arc_buf_hdr_t *data_hdr;
3570 arc_buf_hdr_t *meta_hdr;
3573 * We keep the sublist lock until we're finished, to prevent
3574 * the headers from being destroyed via arc_evict_state().
3576 data_mls = multilist_sublist_lock(data_ml, data_idx);
3577 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
3580 * These two loops are to ensure we skip any markers that
3581 * might be at the tail of the lists due to arc_evict_state().
3584 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
3585 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
3586 if (data_hdr->b_spa != 0)
3590 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
3591 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
3592 if (meta_hdr->b_spa != 0)
3596 if (data_hdr == NULL && meta_hdr == NULL) {
3597 type = ARC_BUFC_DATA;
3598 } else if (data_hdr == NULL) {
3599 ASSERT3P(meta_hdr, !=, NULL);
3600 type = ARC_BUFC_METADATA;
3601 } else if (meta_hdr == NULL) {
3602 ASSERT3P(data_hdr, !=, NULL);
3603 type = ARC_BUFC_DATA;
3605 ASSERT3P(data_hdr, !=, NULL);
3606 ASSERT3P(meta_hdr, !=, NULL);
3608 /* The headers can't be on the sublist without an L1 header */
3609 ASSERT(HDR_HAS_L1HDR(data_hdr));
3610 ASSERT(HDR_HAS_L1HDR(meta_hdr));
3612 if (data_hdr->b_l1hdr.b_arc_access <
3613 meta_hdr->b_l1hdr.b_arc_access) {
3614 type = ARC_BUFC_DATA;
3616 type = ARC_BUFC_METADATA;
3620 multilist_sublist_unlock(meta_mls);
3621 multilist_sublist_unlock(data_mls);
3627 * Evict buffers from the cache, such that arc_size is capped by arc_c.
3632 uint64_t total_evicted = 0;
3637 * If we're over arc_meta_limit, we want to correct that before
3638 * potentially evicting data buffers below.
3640 total_evicted += arc_adjust_meta();
3645 * If we're over the target cache size, we want to evict enough
3646 * from the list to get back to our target size. We don't want
3647 * to evict too much from the MRU, such that it drops below
3648 * arc_p. So, if we're over our target cache size more than
3649 * the MRU is over arc_p, we'll evict enough to get back to
3650 * arc_p here, and then evict more from the MFU below.
3652 target = MIN((int64_t)(arc_size - arc_c),
3653 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3654 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
3657 * If we're below arc_meta_min, always prefer to evict data.
3658 * Otherwise, try to satisfy the requested number of bytes to
3659 * evict from the type which contains older buffers; in an
3660 * effort to keep newer buffers in the cache regardless of their
3661 * type. If we cannot satisfy the number of bytes from this
3662 * type, spill over into the next type.
3664 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
3665 arc_meta_used > arc_meta_min) {
3666 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3667 total_evicted += bytes;
3670 * If we couldn't evict our target number of bytes from
3671 * metadata, we try to get the rest from data.
3676 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3678 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3679 total_evicted += bytes;
3682 * If we couldn't evict our target number of bytes from
3683 * data, we try to get the rest from metadata.
3688 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3694 * Now that we've tried to evict enough from the MRU to get its
3695 * size back to arc_p, if we're still above the target cache
3696 * size, we evict the rest from the MFU.
3698 target = arc_size - arc_c;
3700 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
3701 arc_meta_used > arc_meta_min) {
3702 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3703 total_evicted += bytes;
3706 * If we couldn't evict our target number of bytes from
3707 * metadata, we try to get the rest from data.
3712 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3714 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3715 total_evicted += bytes;
3718 * If we couldn't evict our target number of bytes from
3719 * data, we try to get the rest from data.
3724 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3728 * Adjust ghost lists
3730 * In addition to the above, the ARC also defines target values
3731 * for the ghost lists. The sum of the mru list and mru ghost
3732 * list should never exceed the target size of the cache, and
3733 * the sum of the mru list, mfu list, mru ghost list, and mfu
3734 * ghost list should never exceed twice the target size of the
3735 * cache. The following logic enforces these limits on the ghost
3736 * caches, and evicts from them as needed.
3738 target = refcount_count(&arc_mru->arcs_size) +
3739 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3741 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3742 total_evicted += bytes;
3747 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3750 * We assume the sum of the mru list and mfu list is less than
3751 * or equal to arc_c (we enforced this above), which means we
3752 * can use the simpler of the two equations below:
3754 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3755 * mru ghost + mfu ghost <= arc_c
3757 target = refcount_count(&arc_mru_ghost->arcs_size) +
3758 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3760 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3761 total_evicted += bytes;
3766 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3768 return (total_evicted);
3772 arc_flush(spa_t *spa, boolean_t retry)
3777 * If retry is B_TRUE, a spa must not be specified since we have
3778 * no good way to determine if all of a spa's buffers have been
3779 * evicted from an arc state.
3781 ASSERT(!retry || spa == 0);
3784 guid = spa_load_guid(spa);
3786 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3787 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3789 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3790 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3792 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3793 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3795 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3796 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3800 arc_shrink(int64_t to_free)
3802 if (arc_c > arc_c_min) {
3803 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
3804 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
3805 if (arc_c > arc_c_min + to_free)
3806 atomic_add_64(&arc_c, -to_free);
3810 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3811 if (arc_c > arc_size)
3812 arc_c = MAX(arc_size, arc_c_min);
3814 arc_p = (arc_c >> 1);
3816 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
3819 ASSERT(arc_c >= arc_c_min);
3820 ASSERT((int64_t)arc_p >= 0);
3823 if (arc_size > arc_c) {
3824 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size,
3826 (void) arc_adjust();
3830 static long needfree = 0;
3832 typedef enum free_memory_reason_t {
3837 FMR_PAGES_PP_MAXIMUM,
3841 } free_memory_reason_t;
3843 int64_t last_free_memory;
3844 free_memory_reason_t last_free_reason;
3847 * Additional reserve of pages for pp_reserve.
3849 int64_t arc_pages_pp_reserve = 64;
3852 * Additional reserve of pages for swapfs.
3854 int64_t arc_swapfs_reserve = 64;
3857 * Return the amount of memory that can be consumed before reclaim will be
3858 * needed. Positive if there is sufficient free memory, negative indicates
3859 * the amount of memory that needs to be freed up.
3862 arc_available_memory(void)
3864 int64_t lowest = INT64_MAX;
3866 free_memory_reason_t r = FMR_UNKNOWN;
3870 n = PAGESIZE * (-needfree);
3878 * Cooperate with pagedaemon when it's time for it to scan
3879 * and reclaim some pages.
3881 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
3889 * check that we're out of range of the pageout scanner. It starts to
3890 * schedule paging if freemem is less than lotsfree and needfree.
3891 * lotsfree is the high-water mark for pageout, and needfree is the
3892 * number of needed free pages. We add extra pages here to make sure
3893 * the scanner doesn't start up while we're freeing memory.
3895 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
3902 * check to make sure that swapfs has enough space so that anon
3903 * reservations can still succeed. anon_resvmem() checks that the
3904 * availrmem is greater than swapfs_minfree, and the number of reserved
3905 * swap pages. We also add a bit of extra here just to prevent
3906 * circumstances from getting really dire.
3908 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
3909 desfree - arc_swapfs_reserve);
3912 r = FMR_SWAPFS_MINFREE;
3917 * Check that we have enough availrmem that memory locking (e.g., via
3918 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3919 * stores the number of pages that cannot be locked; when availrmem
3920 * drops below pages_pp_maximum, page locking mechanisms such as
3921 * page_pp_lock() will fail.)
3923 n = PAGESIZE * (availrmem - pages_pp_maximum -
3924 arc_pages_pp_reserve);
3927 r = FMR_PAGES_PP_MAXIMUM;
3930 #endif /* illumos */
3931 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
3933 * If we're on an i386 platform, it's possible that we'll exhaust the
3934 * kernel heap space before we ever run out of available physical
3935 * memory. Most checks of the size of the heap_area compare against
3936 * tune.t_minarmem, which is the minimum available real memory that we
3937 * can have in the system. However, this is generally fixed at 25 pages
3938 * which is so low that it's useless. In this comparison, we seek to
3939 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3940 * heap is allocated. (Or, in the calculation, if less than 1/4th is
3943 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
3944 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
3949 #define zio_arena NULL
3951 #define zio_arena heap_arena
3955 * If zio data pages are being allocated out of a separate heap segment,
3956 * then enforce that the size of available vmem for this arena remains
3957 * above about 1/16th free.
3959 * Note: The 1/16th arena free requirement was put in place
3960 * to aggressively evict memory from the arc in order to avoid
3961 * memory fragmentation issues.
3963 if (zio_arena != NULL) {
3964 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
3965 (vmem_size(zio_arena, VMEM_ALLOC) >> 4);
3973 * Above limits know nothing about real level of KVA fragmentation.
3974 * Start aggressive reclamation if too little sequential KVA left.
3977 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ?
3978 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) :
3987 /* Every 100 calls, free a small amount */
3988 if (spa_get_random(100) == 0)
3990 #endif /* _KERNEL */
3992 last_free_memory = lowest;
3993 last_free_reason = r;
3994 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
4000 * Determine if the system is under memory pressure and is asking
4001 * to reclaim memory. A return value of B_TRUE indicates that the system
4002 * is under memory pressure and that the arc should adjust accordingly.
4005 arc_reclaim_needed(void)
4007 return (arc_available_memory() < 0);
4010 extern kmem_cache_t *zio_buf_cache[];
4011 extern kmem_cache_t *zio_data_buf_cache[];
4012 extern kmem_cache_t *range_seg_cache;
4014 static __noinline void
4015 arc_kmem_reap_now(void)
4018 kmem_cache_t *prev_cache = NULL;
4019 kmem_cache_t *prev_data_cache = NULL;
4021 DTRACE_PROBE(arc__kmem_reap_start);
4023 if (arc_meta_used >= arc_meta_limit) {
4025 * We are exceeding our meta-data cache limit.
4026 * Purge some DNLC entries to release holds on meta-data.
4028 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4032 * Reclaim unused memory from all kmem caches.
4038 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4039 if (zio_buf_cache[i] != prev_cache) {
4040 prev_cache = zio_buf_cache[i];
4041 kmem_cache_reap_now(zio_buf_cache[i]);
4043 if (zio_data_buf_cache[i] != prev_data_cache) {
4044 prev_data_cache = zio_data_buf_cache[i];
4045 kmem_cache_reap_now(zio_data_buf_cache[i]);
4048 kmem_cache_reap_now(buf_cache);
4049 kmem_cache_reap_now(hdr_full_cache);
4050 kmem_cache_reap_now(hdr_l2only_cache);
4051 kmem_cache_reap_now(range_seg_cache);
4054 if (zio_arena != NULL) {
4056 * Ask the vmem arena to reclaim unused memory from its
4059 vmem_qcache_reap(zio_arena);
4062 DTRACE_PROBE(arc__kmem_reap_end);
4066 * Threads can block in arc_get_data_buf() waiting for this thread to evict
4067 * enough data and signal them to proceed. When this happens, the threads in
4068 * arc_get_data_buf() are sleeping while holding the hash lock for their
4069 * particular arc header. Thus, we must be careful to never sleep on a
4070 * hash lock in this thread. This is to prevent the following deadlock:
4072 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
4073 * waiting for the reclaim thread to signal it.
4075 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4076 * fails, and goes to sleep forever.
4078 * This possible deadlock is avoided by always acquiring a hash lock
4079 * using mutex_tryenter() from arc_reclaim_thread().
4082 arc_reclaim_thread(void *dummy __unused)
4084 hrtime_t growtime = 0;
4087 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4089 mutex_enter(&arc_reclaim_lock);
4090 while (!arc_reclaim_thread_exit) {
4091 int64_t free_memory = arc_available_memory();
4092 uint64_t evicted = 0;
4095 * This is necessary in order for the mdb ::arc dcmd to
4096 * show up to date information. Since the ::arc command
4097 * does not call the kstat's update function, without
4098 * this call, the command may show stale stats for the
4099 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4100 * with this change, the data might be up to 1 second
4101 * out of date; but that should suffice. The arc_state_t
4102 * structures can be queried directly if more accurate
4103 * information is needed.
4105 if (arc_ksp != NULL)
4106 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4108 mutex_exit(&arc_reclaim_lock);
4110 if (free_memory < 0) {
4112 arc_no_grow = B_TRUE;
4116 * Wait at least zfs_grow_retry (default 60) seconds
4117 * before considering growing.
4119 growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4121 arc_kmem_reap_now();
4124 * If we are still low on memory, shrink the ARC
4125 * so that we have arc_shrink_min free space.
4127 free_memory = arc_available_memory();
4130 (arc_c >> arc_shrink_shift) - free_memory;
4133 to_free = MAX(to_free, ptob(needfree));
4135 arc_shrink(to_free);
4137 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4138 arc_no_grow = B_TRUE;
4139 } else if (gethrtime() >= growtime) {
4140 arc_no_grow = B_FALSE;
4143 evicted = arc_adjust();
4145 mutex_enter(&arc_reclaim_lock);
4148 * If evicted is zero, we couldn't evict anything via
4149 * arc_adjust(). This could be due to hash lock
4150 * collisions, but more likely due to the majority of
4151 * arc buffers being unevictable. Therefore, even if
4152 * arc_size is above arc_c, another pass is unlikely to
4153 * be helpful and could potentially cause us to enter an
4156 if (arc_size <= arc_c || evicted == 0) {
4161 * We're either no longer overflowing, or we
4162 * can't evict anything more, so we should wake
4163 * up any threads before we go to sleep.
4165 cv_broadcast(&arc_reclaim_waiters_cv);
4168 * Block until signaled, or after one second (we
4169 * might need to perform arc_kmem_reap_now()
4170 * even if we aren't being signalled)
4172 CALLB_CPR_SAFE_BEGIN(&cpr);
4173 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4174 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4175 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4179 arc_reclaim_thread_exit = B_FALSE;
4180 cv_broadcast(&arc_reclaim_thread_cv);
4181 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
4185 static u_int arc_dnlc_evicts_arg;
4186 extern struct vfsops zfs_vfsops;
4189 arc_dnlc_evicts_thread(void *dummy __unused)
4194 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG);
4196 mutex_enter(&arc_dnlc_evicts_lock);
4197 while (!arc_dnlc_evicts_thread_exit) {
4198 CALLB_CPR_SAFE_BEGIN(&cpr);
4199 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
4200 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock);
4201 if (arc_dnlc_evicts_arg != 0) {
4202 percent = arc_dnlc_evicts_arg;
4203 mutex_exit(&arc_dnlc_evicts_lock);
4205 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops);
4207 mutex_enter(&arc_dnlc_evicts_lock);
4209 * Clear our token only after vnlru_free()
4210 * pass is done, to avoid false queueing of
4213 arc_dnlc_evicts_arg = 0;
4216 arc_dnlc_evicts_thread_exit = FALSE;
4217 cv_broadcast(&arc_dnlc_evicts_cv);
4218 CALLB_CPR_EXIT(&cpr);
4223 dnlc_reduce_cache(void *arg)
4227 percent = (u_int)(uintptr_t)arg;
4228 mutex_enter(&arc_dnlc_evicts_lock);
4229 if (arc_dnlc_evicts_arg == 0) {
4230 arc_dnlc_evicts_arg = percent;
4231 cv_broadcast(&arc_dnlc_evicts_cv);
4233 mutex_exit(&arc_dnlc_evicts_lock);
4237 * Adapt arc info given the number of bytes we are trying to add and
4238 * the state that we are comming from. This function is only called
4239 * when we are adding new content to the cache.
4242 arc_adapt(int bytes, arc_state_t *state)
4245 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4246 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
4247 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
4249 if (state == arc_l2c_only)
4254 * Adapt the target size of the MRU list:
4255 * - if we just hit in the MRU ghost list, then increase
4256 * the target size of the MRU list.
4257 * - if we just hit in the MFU ghost list, then increase
4258 * the target size of the MFU list by decreasing the
4259 * target size of the MRU list.
4261 if (state == arc_mru_ghost) {
4262 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4263 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4265 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4266 } else if (state == arc_mfu_ghost) {
4269 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4270 mult = MIN(mult, 10);
4272 delta = MIN(bytes * mult, arc_p);
4273 arc_p = MAX(arc_p_min, arc_p - delta);
4275 ASSERT((int64_t)arc_p >= 0);
4277 if (arc_reclaim_needed()) {
4278 cv_signal(&arc_reclaim_thread_cv);
4285 if (arc_c >= arc_c_max)
4289 * If we're within (2 * maxblocksize) bytes of the target
4290 * cache size, increment the target cache size
4292 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
4293 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
4294 atomic_add_64(&arc_c, (int64_t)bytes);
4295 if (arc_c > arc_c_max)
4297 else if (state == arc_anon)
4298 atomic_add_64(&arc_p, (int64_t)bytes);
4302 ASSERT((int64_t)arc_p >= 0);
4306 * Check if arc_size has grown past our upper threshold, determined by
4307 * zfs_arc_overflow_shift.
4310 arc_is_overflowing(void)
4312 /* Always allow at least one block of overflow */
4313 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4314 arc_c >> zfs_arc_overflow_shift);
4316 return (arc_size >= arc_c + overflow);
4320 * Allocate a block and return it to the caller. If we are hitting the
4321 * hard limit for the cache size, we must sleep, waiting for the eviction
4322 * thread to catch up. If we're past the target size but below the hard
4323 * limit, we'll only signal the reclaim thread and continue on.
4326 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4329 arc_state_t *state = hdr->b_l1hdr.b_state;
4330 arc_buf_contents_t type = arc_buf_type(hdr);
4332 arc_adapt(size, state);
4335 * If arc_size is currently overflowing, and has grown past our
4336 * upper limit, we must be adding data faster than the evict
4337 * thread can evict. Thus, to ensure we don't compound the
4338 * problem by adding more data and forcing arc_size to grow even
4339 * further past it's target size, we halt and wait for the
4340 * eviction thread to catch up.
4342 * It's also possible that the reclaim thread is unable to evict
4343 * enough buffers to get arc_size below the overflow limit (e.g.
4344 * due to buffers being un-evictable, or hash lock collisions).
4345 * In this case, we want to proceed regardless if we're
4346 * overflowing; thus we don't use a while loop here.
4348 if (arc_is_overflowing()) {
4349 mutex_enter(&arc_reclaim_lock);
4352 * Now that we've acquired the lock, we may no longer be
4353 * over the overflow limit, lets check.
4355 * We're ignoring the case of spurious wake ups. If that
4356 * were to happen, it'd let this thread consume an ARC
4357 * buffer before it should have (i.e. before we're under
4358 * the overflow limit and were signalled by the reclaim
4359 * thread). As long as that is a rare occurrence, it
4360 * shouldn't cause any harm.
4362 if (arc_is_overflowing()) {
4363 cv_signal(&arc_reclaim_thread_cv);
4364 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
4367 mutex_exit(&arc_reclaim_lock);
4370 VERIFY3U(hdr->b_type, ==, type);
4371 if (type == ARC_BUFC_METADATA) {
4372 datap = zio_buf_alloc(size);
4373 arc_space_consume(size, ARC_SPACE_META);
4375 ASSERT(type == ARC_BUFC_DATA);
4376 datap = zio_data_buf_alloc(size);
4377 arc_space_consume(size, ARC_SPACE_DATA);
4381 * Update the state size. Note that ghost states have a
4382 * "ghost size" and so don't need to be updated.
4384 if (!GHOST_STATE(state)) {
4386 (void) refcount_add_many(&state->arcs_size, size, tag);
4389 * If this is reached via arc_read, the link is
4390 * protected by the hash lock. If reached via
4391 * arc_buf_alloc, the header should not be accessed by
4392 * any other thread. And, if reached via arc_read_done,
4393 * the hash lock will protect it if it's found in the
4394 * hash table; otherwise no other thread should be
4395 * trying to [add|remove]_reference it.
4397 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4398 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4399 (void) refcount_add_many(&state->arcs_esize[type],
4404 * If we are growing the cache, and we are adding anonymous
4405 * data, and we have outgrown arc_p, update arc_p
4407 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
4408 (refcount_count(&arc_anon->arcs_size) +
4409 refcount_count(&arc_mru->arcs_size) > arc_p))
4410 arc_p = MIN(arc_c, arc_p + size);
4412 ARCSTAT_BUMP(arcstat_allocated);
4417 * Free the arc data buffer.
4420 arc_free_data_buf(arc_buf_hdr_t *hdr, void *data, uint64_t size, void *tag)
4422 arc_state_t *state = hdr->b_l1hdr.b_state;
4423 arc_buf_contents_t type = arc_buf_type(hdr);
4425 /* protected by hash lock, if in the hash table */
4426 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4427 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4428 ASSERT(state != arc_anon && state != arc_l2c_only);
4430 (void) refcount_remove_many(&state->arcs_esize[type],
4433 (void) refcount_remove_many(&state->arcs_size, size, tag);
4435 VERIFY3U(hdr->b_type, ==, type);
4436 if (type == ARC_BUFC_METADATA) {
4437 zio_buf_free(data, size);
4438 arc_space_return(size, ARC_SPACE_META);
4440 ASSERT(type == ARC_BUFC_DATA);
4441 zio_data_buf_free(data, size);
4442 arc_space_return(size, ARC_SPACE_DATA);
4447 * This routine is called whenever a buffer is accessed.
4448 * NOTE: the hash lock is dropped in this function.
4451 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4455 ASSERT(MUTEX_HELD(hash_lock));
4456 ASSERT(HDR_HAS_L1HDR(hdr));
4458 if (hdr->b_l1hdr.b_state == arc_anon) {
4460 * This buffer is not in the cache, and does not
4461 * appear in our "ghost" list. Add the new buffer
4465 ASSERT0(hdr->b_l1hdr.b_arc_access);
4466 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4467 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4468 arc_change_state(arc_mru, hdr, hash_lock);
4470 } else if (hdr->b_l1hdr.b_state == arc_mru) {
4471 now = ddi_get_lbolt();
4474 * If this buffer is here because of a prefetch, then either:
4475 * - clear the flag if this is a "referencing" read
4476 * (any subsequent access will bump this into the MFU state).
4478 * - move the buffer to the head of the list if this is
4479 * another prefetch (to make it less likely to be evicted).
4481 if (HDR_PREFETCH(hdr)) {
4482 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4483 /* link protected by hash lock */
4484 ASSERT(multilist_link_active(
4485 &hdr->b_l1hdr.b_arc_node));
4487 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4488 ARCSTAT_BUMP(arcstat_mru_hits);
4490 hdr->b_l1hdr.b_arc_access = now;
4495 * This buffer has been "accessed" only once so far,
4496 * but it is still in the cache. Move it to the MFU
4499 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
4501 * More than 125ms have passed since we
4502 * instantiated this buffer. Move it to the
4503 * most frequently used state.
4505 hdr->b_l1hdr.b_arc_access = now;
4506 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4507 arc_change_state(arc_mfu, hdr, hash_lock);
4509 ARCSTAT_BUMP(arcstat_mru_hits);
4510 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
4511 arc_state_t *new_state;
4513 * This buffer has been "accessed" recently, but
4514 * was evicted from the cache. Move it to the
4518 if (HDR_PREFETCH(hdr)) {
4519 new_state = arc_mru;
4520 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
4521 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4522 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4524 new_state = arc_mfu;
4525 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4528 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4529 arc_change_state(new_state, hdr, hash_lock);
4531 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
4532 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
4534 * This buffer has been accessed more than once and is
4535 * still in the cache. Keep it in the MFU state.
4537 * NOTE: an add_reference() that occurred when we did
4538 * the arc_read() will have kicked this off the list.
4539 * If it was a prefetch, we will explicitly move it to
4540 * the head of the list now.
4542 if ((HDR_PREFETCH(hdr)) != 0) {
4543 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4544 /* link protected by hash_lock */
4545 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4547 ARCSTAT_BUMP(arcstat_mfu_hits);
4548 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4549 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
4550 arc_state_t *new_state = arc_mfu;
4552 * This buffer has been accessed more than once but has
4553 * been evicted from the cache. Move it back to the
4557 if (HDR_PREFETCH(hdr)) {
4559 * This is a prefetch access...
4560 * move this block back to the MRU state.
4562 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
4563 new_state = arc_mru;
4566 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4567 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4568 arc_change_state(new_state, hdr, hash_lock);
4570 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
4571 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
4573 * This buffer is on the 2nd Level ARC.
4576 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4577 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4578 arc_change_state(arc_mfu, hdr, hash_lock);
4580 ASSERT(!"invalid arc state");
4584 /* a generic arc_done_func_t which you can use */
4587 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
4589 if (zio == NULL || zio->io_error == 0)
4590 bcopy(buf->b_data, arg, HDR_GET_LSIZE(buf->b_hdr));
4591 arc_buf_destroy(buf, arg);
4594 /* a generic arc_done_func_t */
4596 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
4598 arc_buf_t **bufp = arg;
4599 if (zio && zio->io_error) {
4600 arc_buf_destroy(buf, arg);
4604 ASSERT(buf->b_data);
4609 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
4611 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
4612 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
4613 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
4615 if (HDR_COMPRESSION_ENABLED(hdr)) {
4616 ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
4617 BP_GET_COMPRESS(bp));
4619 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
4620 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
4625 arc_read_done(zio_t *zio)
4627 arc_buf_hdr_t *hdr = zio->io_private;
4628 arc_buf_t *abuf = NULL; /* buffer we're assigning to callback */
4629 kmutex_t *hash_lock = NULL;
4630 arc_callback_t *callback_list, *acb;
4631 int freeable = B_FALSE;
4634 * The hdr was inserted into hash-table and removed from lists
4635 * prior to starting I/O. We should find this header, since
4636 * it's in the hash table, and it should be legit since it's
4637 * not possible to evict it during the I/O. The only possible
4638 * reason for it not to be found is if we were freed during the
4641 if (HDR_IN_HASH_TABLE(hdr)) {
4642 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
4643 ASSERT3U(hdr->b_dva.dva_word[0], ==,
4644 BP_IDENTITY(zio->io_bp)->dva_word[0]);
4645 ASSERT3U(hdr->b_dva.dva_word[1], ==,
4646 BP_IDENTITY(zio->io_bp)->dva_word[1]);
4648 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
4651 ASSERT((found == hdr &&
4652 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
4653 (found == hdr && HDR_L2_READING(hdr)));
4654 ASSERT3P(hash_lock, !=, NULL);
4657 if (zio->io_error == 0) {
4658 /* byteswap if necessary */
4659 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
4660 if (BP_GET_LEVEL(zio->io_bp) > 0) {
4661 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
4663 hdr->b_l1hdr.b_byteswap =
4664 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
4667 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
4671 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
4672 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
4673 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
4675 callback_list = hdr->b_l1hdr.b_acb;
4676 ASSERT3P(callback_list, !=, NULL);
4678 if (hash_lock && zio->io_error == 0 &&
4679 hdr->b_l1hdr.b_state == arc_anon) {
4681 * Only call arc_access on anonymous buffers. This is because
4682 * if we've issued an I/O for an evicted buffer, we've already
4683 * called arc_access (to prevent any simultaneous readers from
4684 * getting confused).
4686 arc_access(hdr, hash_lock);
4689 /* create copies of the data buffer for the callers */
4690 for (acb = callback_list; acb; acb = acb->acb_next) {
4691 if (acb->acb_done != NULL) {
4693 * If we're here, then this must be a demand read
4694 * since prefetch requests don't have callbacks.
4695 * If a read request has a callback (i.e. acb_done is
4696 * not NULL), then we decompress the data for the
4697 * first request and clone the rest. This avoids
4698 * having to waste cpu resources decompressing data
4699 * that nobody is explicitly waiting to read.
4702 acb->acb_buf = arc_buf_alloc_impl(hdr,
4704 if (zio->io_error == 0) {
4706 arc_decompress(acb->acb_buf);
4708 abuf = acb->acb_buf;
4710 add_reference(hdr, acb->acb_private);
4711 acb->acb_buf = arc_buf_clone(abuf);
4715 hdr->b_l1hdr.b_acb = NULL;
4716 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
4719 * This buffer didn't have a callback so it must
4722 ASSERT(HDR_PREFETCH(hdr));
4723 ASSERT0(hdr->b_l1hdr.b_bufcnt);
4724 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
4727 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
4728 callback_list != NULL);
4730 if (zio->io_error == 0) {
4731 arc_hdr_verify(hdr, zio->io_bp);
4733 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
4734 if (hdr->b_l1hdr.b_state != arc_anon)
4735 arc_change_state(arc_anon, hdr, hash_lock);
4736 if (HDR_IN_HASH_TABLE(hdr))
4737 buf_hash_remove(hdr);
4738 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4742 * Broadcast before we drop the hash_lock to avoid the possibility
4743 * that the hdr (and hence the cv) might be freed before we get to
4744 * the cv_broadcast().
4746 cv_broadcast(&hdr->b_l1hdr.b_cv);
4748 if (hash_lock != NULL) {
4749 mutex_exit(hash_lock);
4752 * This block was freed while we waited for the read to
4753 * complete. It has been removed from the hash table and
4754 * moved to the anonymous state (so that it won't show up
4757 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
4758 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4761 /* execute each callback and free its structure */
4762 while ((acb = callback_list) != NULL) {
4764 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
4766 if (acb->acb_zio_dummy != NULL) {
4767 acb->acb_zio_dummy->io_error = zio->io_error;
4768 zio_nowait(acb->acb_zio_dummy);
4771 callback_list = acb->acb_next;
4772 kmem_free(acb, sizeof (arc_callback_t));
4776 arc_hdr_destroy(hdr);
4780 * "Read" the block at the specified DVA (in bp) via the
4781 * cache. If the block is found in the cache, invoke the provided
4782 * callback immediately and return. Note that the `zio' parameter
4783 * in the callback will be NULL in this case, since no IO was
4784 * required. If the block is not in the cache pass the read request
4785 * on to the spa with a substitute callback function, so that the
4786 * requested block will be added to the cache.
4788 * If a read request arrives for a block that has a read in-progress,
4789 * either wait for the in-progress read to complete (and return the
4790 * results); or, if this is a read with a "done" func, add a record
4791 * to the read to invoke the "done" func when the read completes,
4792 * and return; or just return.
4794 * arc_read_done() will invoke all the requested "done" functions
4795 * for readers of this block.
4798 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
4799 void *private, zio_priority_t priority, int zio_flags,
4800 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
4802 arc_buf_hdr_t *hdr = NULL;
4803 kmutex_t *hash_lock = NULL;
4805 uint64_t guid = spa_load_guid(spa);
4807 ASSERT(!BP_IS_EMBEDDED(bp) ||
4808 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
4811 if (!BP_IS_EMBEDDED(bp)) {
4813 * Embedded BP's have no DVA and require no I/O to "read".
4814 * Create an anonymous arc buf to back it.
4816 hdr = buf_hash_find(guid, bp, &hash_lock);
4819 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pdata != NULL) {
4820 arc_buf_t *buf = NULL;
4821 *arc_flags |= ARC_FLAG_CACHED;
4823 if (HDR_IO_IN_PROGRESS(hdr)) {
4825 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
4826 priority == ZIO_PRIORITY_SYNC_READ) {
4828 * This sync read must wait for an
4829 * in-progress async read (e.g. a predictive
4830 * prefetch). Async reads are queued
4831 * separately at the vdev_queue layer, so
4832 * this is a form of priority inversion.
4833 * Ideally, we would "inherit" the demand
4834 * i/o's priority by moving the i/o from
4835 * the async queue to the synchronous queue,
4836 * but there is currently no mechanism to do
4837 * so. Track this so that we can evaluate
4838 * the magnitude of this potential performance
4841 * Note that if the prefetch i/o is already
4842 * active (has been issued to the device),
4843 * the prefetch improved performance, because
4844 * we issued it sooner than we would have
4845 * without the prefetch.
4847 DTRACE_PROBE1(arc__sync__wait__for__async,
4848 arc_buf_hdr_t *, hdr);
4849 ARCSTAT_BUMP(arcstat_sync_wait_for_async);
4851 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4852 arc_hdr_clear_flags(hdr,
4853 ARC_FLAG_PREDICTIVE_PREFETCH);
4856 if (*arc_flags & ARC_FLAG_WAIT) {
4857 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
4858 mutex_exit(hash_lock);
4861 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4864 arc_callback_t *acb = NULL;
4866 acb = kmem_zalloc(sizeof (arc_callback_t),
4868 acb->acb_done = done;
4869 acb->acb_private = private;
4871 acb->acb_zio_dummy = zio_null(pio,
4872 spa, NULL, NULL, NULL, zio_flags);
4874 ASSERT3P(acb->acb_done, !=, NULL);
4875 acb->acb_next = hdr->b_l1hdr.b_acb;
4876 hdr->b_l1hdr.b_acb = acb;
4877 mutex_exit(hash_lock);
4880 mutex_exit(hash_lock);
4884 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4885 hdr->b_l1hdr.b_state == arc_mfu);
4888 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4890 * This is a demand read which does not have to
4891 * wait for i/o because we did a predictive
4892 * prefetch i/o for it, which has completed.
4895 arc__demand__hit__predictive__prefetch,
4896 arc_buf_hdr_t *, hdr);
4898 arcstat_demand_hit_predictive_prefetch);
4899 arc_hdr_clear_flags(hdr,
4900 ARC_FLAG_PREDICTIVE_PREFETCH);
4902 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
4905 * If this block is already in use, create a new
4906 * copy of the data so that we will be guaranteed
4907 * that arc_release() will always succeed.
4909 buf = hdr->b_l1hdr.b_buf;
4911 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
4912 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
4913 buf = arc_buf_alloc_impl(hdr, private);
4914 VERIFY0(arc_decompress(buf));
4916 add_reference(hdr, private);
4917 buf = arc_buf_clone(buf);
4919 ASSERT3P(buf->b_data, !=, NULL);
4921 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
4922 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4923 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
4925 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4926 arc_access(hdr, hash_lock);
4927 if (*arc_flags & ARC_FLAG_L2CACHE)
4928 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
4929 mutex_exit(hash_lock);
4930 ARCSTAT_BUMP(arcstat_hits);
4931 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4932 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4933 data, metadata, hits);
4936 done(NULL, buf, private);
4938 uint64_t lsize = BP_GET_LSIZE(bp);
4939 uint64_t psize = BP_GET_PSIZE(bp);
4940 arc_callback_t *acb;
4943 boolean_t devw = B_FALSE;
4947 /* this block is not in the cache */
4948 arc_buf_hdr_t *exists = NULL;
4949 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
4950 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
4951 BP_GET_COMPRESS(bp), type);
4953 if (!BP_IS_EMBEDDED(bp)) {
4954 hdr->b_dva = *BP_IDENTITY(bp);
4955 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
4956 exists = buf_hash_insert(hdr, &hash_lock);
4958 if (exists != NULL) {
4959 /* somebody beat us to the hash insert */
4960 mutex_exit(hash_lock);
4961 buf_discard_identity(hdr);
4962 arc_hdr_destroy(hdr);
4963 goto top; /* restart the IO request */
4967 * This block is in the ghost cache. If it was L2-only
4968 * (and thus didn't have an L1 hdr), we realloc the
4969 * header to add an L1 hdr.
4971 if (!HDR_HAS_L1HDR(hdr)) {
4972 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
4975 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
4976 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
4977 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4978 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4979 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
4982 * This is a delicate dance that we play here.
4983 * This hdr is in the ghost list so we access it
4984 * to move it out of the ghost list before we
4985 * initiate the read. If it's a prefetch then
4986 * it won't have a callback so we'll remove the
4987 * reference that arc_buf_alloc_impl() created. We
4988 * do this after we've called arc_access() to
4989 * avoid hitting an assert in remove_reference().
4991 arc_access(hdr, hash_lock);
4992 arc_hdr_alloc_pdata(hdr);
4994 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
4995 size = arc_hdr_size(hdr);
4998 * If compression is enabled on the hdr, then will do
4999 * RAW I/O and will store the compressed data in the hdr's
5000 * data block. Otherwise, the hdr's data block will contain
5001 * the uncompressed data.
5003 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5004 zio_flags |= ZIO_FLAG_RAW;
5007 if (*arc_flags & ARC_FLAG_PREFETCH)
5008 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5009 if (*arc_flags & ARC_FLAG_L2CACHE)
5010 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5011 if (BP_GET_LEVEL(bp) > 0)
5012 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5013 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5014 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5015 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5017 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5018 acb->acb_done = done;
5019 acb->acb_private = private;
5021 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5022 hdr->b_l1hdr.b_acb = acb;
5023 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5025 if (HDR_HAS_L2HDR(hdr) &&
5026 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5027 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5028 addr = hdr->b_l2hdr.b_daddr;
5030 * Lock out device removal.
5032 if (vdev_is_dead(vd) ||
5033 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5037 if (priority == ZIO_PRIORITY_ASYNC_READ)
5038 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5040 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5042 if (hash_lock != NULL)
5043 mutex_exit(hash_lock);
5046 * At this point, we have a level 1 cache miss. Try again in
5047 * L2ARC if possible.
5049 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5051 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5052 uint64_t, lsize, zbookmark_phys_t *, zb);
5053 ARCSTAT_BUMP(arcstat_misses);
5054 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5055 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5056 data, metadata, misses);
5061 racct_add_force(curproc, RACCT_READBPS, size);
5062 racct_add_force(curproc, RACCT_READIOPS, 1);
5063 PROC_UNLOCK(curproc);
5066 curthread->td_ru.ru_inblock++;
5069 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5071 * Read from the L2ARC if the following are true:
5072 * 1. The L2ARC vdev was previously cached.
5073 * 2. This buffer still has L2ARC metadata.
5074 * 3. This buffer isn't currently writing to the L2ARC.
5075 * 4. The L2ARC entry wasn't evicted, which may
5076 * also have invalidated the vdev.
5077 * 5. This isn't prefetch and l2arc_noprefetch is set.
5079 if (HDR_HAS_L2HDR(hdr) &&
5080 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5081 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5082 l2arc_read_callback_t *cb;
5085 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5086 ARCSTAT_BUMP(arcstat_l2_hits);
5088 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5090 cb->l2rcb_hdr = hdr;
5093 cb->l2rcb_flags = zio_flags;
5094 uint64_t asize = vdev_psize_to_asize(vd, size);
5095 if (asize != size) {
5096 b_data = zio_data_buf_alloc(asize);
5097 cb->l2rcb_data = b_data;
5099 b_data = hdr->b_l1hdr.b_pdata;
5102 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5103 addr + asize < vd->vdev_psize -
5104 VDEV_LABEL_END_SIZE);
5107 * l2arc read. The SCL_L2ARC lock will be
5108 * released by l2arc_read_done().
5109 * Issue a null zio if the underlying buffer
5110 * was squashed to zero size by compression.
5112 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
5113 ZIO_COMPRESS_EMPTY);
5114 rzio = zio_read_phys(pio, vd, addr,
5117 l2arc_read_done, cb, priority,
5118 zio_flags | ZIO_FLAG_DONT_CACHE |
5120 ZIO_FLAG_DONT_PROPAGATE |
5121 ZIO_FLAG_DONT_RETRY, B_FALSE);
5122 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5124 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
5126 if (*arc_flags & ARC_FLAG_NOWAIT) {
5131 ASSERT(*arc_flags & ARC_FLAG_WAIT);
5132 if (zio_wait(rzio) == 0)
5135 /* l2arc read error; goto zio_read() */
5137 DTRACE_PROBE1(l2arc__miss,
5138 arc_buf_hdr_t *, hdr);
5139 ARCSTAT_BUMP(arcstat_l2_misses);
5140 if (HDR_L2_WRITING(hdr))
5141 ARCSTAT_BUMP(arcstat_l2_rw_clash);
5142 spa_config_exit(spa, SCL_L2ARC, vd);
5146 spa_config_exit(spa, SCL_L2ARC, vd);
5147 if (l2arc_ndev != 0) {
5148 DTRACE_PROBE1(l2arc__miss,
5149 arc_buf_hdr_t *, hdr);
5150 ARCSTAT_BUMP(arcstat_l2_misses);
5154 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pdata, size,
5155 arc_read_done, hdr, priority, zio_flags, zb);
5157 if (*arc_flags & ARC_FLAG_WAIT)
5158 return (zio_wait(rzio));
5160 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5167 * Notify the arc that a block was freed, and thus will never be used again.
5170 arc_freed(spa_t *spa, const blkptr_t *bp)
5173 kmutex_t *hash_lock;
5174 uint64_t guid = spa_load_guid(spa);
5176 ASSERT(!BP_IS_EMBEDDED(bp));
5178 hdr = buf_hash_find(guid, bp, &hash_lock);
5183 * We might be trying to free a block that is still doing I/O
5184 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5185 * dmu_sync-ed block). If this block is being prefetched, then it
5186 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5187 * until the I/O completes. A block may also have a reference if it is
5188 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5189 * have written the new block to its final resting place on disk but
5190 * without the dedup flag set. This would have left the hdr in the MRU
5191 * state and discoverable. When the txg finally syncs it detects that
5192 * the block was overridden in open context and issues an override I/O.
5193 * Since this is a dedup block, the override I/O will determine if the
5194 * block is already in the DDT. If so, then it will replace the io_bp
5195 * with the bp from the DDT and allow the I/O to finish. When the I/O
5196 * reaches the done callback, dbuf_write_override_done, it will
5197 * check to see if the io_bp and io_bp_override are identical.
5198 * If they are not, then it indicates that the bp was replaced with
5199 * the bp in the DDT and the override bp is freed. This allows
5200 * us to arrive here with a reference on a block that is being
5201 * freed. So if we have an I/O in progress, or a reference to
5202 * this hdr, then we don't destroy the hdr.
5204 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
5205 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
5206 arc_change_state(arc_anon, hdr, hash_lock);
5207 arc_hdr_destroy(hdr);
5208 mutex_exit(hash_lock);
5210 mutex_exit(hash_lock);
5216 * Release this buffer from the cache, making it an anonymous buffer. This
5217 * must be done after a read and prior to modifying the buffer contents.
5218 * If the buffer has more than one reference, we must make
5219 * a new hdr for the buffer.
5222 arc_release(arc_buf_t *buf, void *tag)
5224 arc_buf_hdr_t *hdr = buf->b_hdr;
5227 * It would be nice to assert that if it's DMU metadata (level >
5228 * 0 || it's the dnode file), then it must be syncing context.
5229 * But we don't know that information at this level.
5232 mutex_enter(&buf->b_evict_lock);
5234 ASSERT(HDR_HAS_L1HDR(hdr));
5237 * We don't grab the hash lock prior to this check, because if
5238 * the buffer's header is in the arc_anon state, it won't be
5239 * linked into the hash table.
5241 if (hdr->b_l1hdr.b_state == arc_anon) {
5242 mutex_exit(&buf->b_evict_lock);
5243 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5244 ASSERT(!HDR_IN_HASH_TABLE(hdr));
5245 ASSERT(!HDR_HAS_L2HDR(hdr));
5246 ASSERT(HDR_EMPTY(hdr));
5247 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5248 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
5249 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
5251 hdr->b_l1hdr.b_arc_access = 0;
5254 * If the buf is being overridden then it may already
5255 * have a hdr that is not empty.
5257 buf_discard_identity(hdr);
5263 kmutex_t *hash_lock = HDR_LOCK(hdr);
5264 mutex_enter(hash_lock);
5267 * This assignment is only valid as long as the hash_lock is
5268 * held, we must be careful not to reference state or the
5269 * b_state field after dropping the lock.
5271 arc_state_t *state = hdr->b_l1hdr.b_state;
5272 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5273 ASSERT3P(state, !=, arc_anon);
5275 /* this buffer is not on any list */
5276 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0);
5278 if (HDR_HAS_L2HDR(hdr)) {
5279 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5282 * We have to recheck this conditional again now that
5283 * we're holding the l2ad_mtx to prevent a race with
5284 * another thread which might be concurrently calling
5285 * l2arc_evict(). In that case, l2arc_evict() might have
5286 * destroyed the header's L2 portion as we were waiting
5287 * to acquire the l2ad_mtx.
5289 if (HDR_HAS_L2HDR(hdr)) {
5291 arc_hdr_l2hdr_destroy(hdr);
5294 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5298 * Do we have more than one buf?
5300 if (hdr->b_l1hdr.b_bufcnt > 1) {
5301 arc_buf_hdr_t *nhdr;
5303 uint64_t spa = hdr->b_spa;
5304 uint64_t psize = HDR_GET_PSIZE(hdr);
5305 uint64_t lsize = HDR_GET_LSIZE(hdr);
5306 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
5307 arc_buf_contents_t type = arc_buf_type(hdr);
5308 VERIFY3U(hdr->b_type, ==, type);
5310 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
5311 (void) remove_reference(hdr, hash_lock, tag);
5313 if (arc_buf_is_shared(buf)) {
5314 ASSERT(HDR_SHARED_DATA(hdr));
5315 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5316 ASSERT(ARC_BUF_LAST(buf));
5320 * Pull the data off of this hdr and attach it to
5321 * a new anonymous hdr. Also find the last buffer
5322 * in the hdr's buffer list.
5324 arc_buf_t *lastbuf = NULL;
5325 bufp = &hdr->b_l1hdr.b_buf;
5326 while (*bufp != NULL) {
5328 *bufp = buf->b_next;
5332 * If we've removed a buffer in the middle of
5333 * the list then update the lastbuf and update
5336 if (*bufp != NULL) {
5338 bufp = &(*bufp)->b_next;
5342 ASSERT3P(lastbuf, !=, buf);
5343 ASSERT3P(lastbuf, !=, NULL);
5346 * If the current arc_buf_t and the hdr are sharing their data
5347 * buffer, then we must stop sharing that block, transfer
5348 * ownership and setup sharing with a new arc_buf_t at the end
5349 * of the hdr's b_buf list.
5351 if (arc_buf_is_shared(buf)) {
5352 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5353 ASSERT(ARC_BUF_LAST(lastbuf));
5354 VERIFY(!arc_buf_is_shared(lastbuf));
5357 * First, sever the block sharing relationship between
5358 * buf and the arc_buf_hdr_t. Then, setup a new
5359 * block sharing relationship with the last buffer
5360 * on the arc_buf_t list.
5362 arc_unshare_buf(hdr, buf);
5363 arc_share_buf(hdr, lastbuf);
5364 VERIFY3P(lastbuf->b_data, !=, NULL);
5365 } else if (HDR_SHARED_DATA(hdr)) {
5366 ASSERT(arc_buf_is_shared(lastbuf));
5368 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
5369 ASSERT3P(state, !=, arc_l2c_only);
5371 (void) refcount_remove_many(&state->arcs_size,
5372 HDR_GET_LSIZE(hdr), buf);
5374 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5375 ASSERT3P(state, !=, arc_l2c_only);
5376 (void) refcount_remove_many(&state->arcs_esize[type],
5377 HDR_GET_LSIZE(hdr), buf);
5380 hdr->b_l1hdr.b_bufcnt -= 1;
5381 arc_cksum_verify(buf);
5383 arc_buf_unwatch(buf);
5386 mutex_exit(hash_lock);
5389 * Allocate a new hdr. The new hdr will contain a b_pdata
5390 * buffer which will be freed in arc_write().
5392 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
5393 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
5394 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
5395 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
5396 VERIFY3U(nhdr->b_type, ==, type);
5397 ASSERT(!HDR_SHARED_DATA(nhdr));
5399 nhdr->b_l1hdr.b_buf = buf;
5400 nhdr->b_l1hdr.b_bufcnt = 1;
5401 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
5404 mutex_exit(&buf->b_evict_lock);
5405 (void) refcount_add_many(&arc_anon->arcs_size,
5406 HDR_GET_LSIZE(nhdr), buf);
5408 mutex_exit(&buf->b_evict_lock);
5409 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
5410 /* protected by hash lock, or hdr is on arc_anon */
5411 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5412 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5413 arc_change_state(arc_anon, hdr, hash_lock);
5414 hdr->b_l1hdr.b_arc_access = 0;
5415 mutex_exit(hash_lock);
5417 buf_discard_identity(hdr);
5423 arc_released(arc_buf_t *buf)
5427 mutex_enter(&buf->b_evict_lock);
5428 released = (buf->b_data != NULL &&
5429 buf->b_hdr->b_l1hdr.b_state == arc_anon);
5430 mutex_exit(&buf->b_evict_lock);
5436 arc_referenced(arc_buf_t *buf)
5440 mutex_enter(&buf->b_evict_lock);
5441 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
5442 mutex_exit(&buf->b_evict_lock);
5443 return (referenced);
5448 arc_write_ready(zio_t *zio)
5450 arc_write_callback_t *callback = zio->io_private;
5451 arc_buf_t *buf = callback->awcb_buf;
5452 arc_buf_hdr_t *hdr = buf->b_hdr;
5453 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
5455 ASSERT(HDR_HAS_L1HDR(hdr));
5456 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
5457 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
5460 * If we're reexecuting this zio because the pool suspended, then
5461 * cleanup any state that was previously set the first time the
5462 * callback as invoked.
5464 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
5465 arc_cksum_free(hdr);
5467 arc_buf_unwatch(buf);
5469 if (hdr->b_l1hdr.b_pdata != NULL) {
5470 if (arc_buf_is_shared(buf)) {
5471 ASSERT(HDR_SHARED_DATA(hdr));
5473 arc_unshare_buf(hdr, buf);
5475 arc_hdr_free_pdata(hdr);
5479 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
5480 ASSERT(!HDR_SHARED_DATA(hdr));
5481 ASSERT(!arc_buf_is_shared(buf));
5483 callback->awcb_ready(zio, buf, callback->awcb_private);
5485 if (HDR_IO_IN_PROGRESS(hdr))
5486 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
5488 arc_cksum_compute(buf);
5489 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5491 enum zio_compress compress;
5492 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5493 compress = ZIO_COMPRESS_OFF;
5495 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
5496 compress = BP_GET_COMPRESS(zio->io_bp);
5498 HDR_SET_PSIZE(hdr, psize);
5499 arc_hdr_set_compress(hdr, compress);
5502 * If the hdr is compressed, then copy the compressed
5503 * zio contents into arc_buf_hdr_t. Otherwise, copy the original
5504 * data buf into the hdr. Ideally, we would like to always copy the
5505 * io_data into b_pdata but the user may have disabled compressed
5506 * arc thus the on-disk block may or may not match what we maintain
5507 * in the hdr's b_pdata field.
5509 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5510 ASSERT(BP_GET_COMPRESS(zio->io_bp) != ZIO_COMPRESS_OFF);
5511 ASSERT3U(psize, >, 0);
5512 arc_hdr_alloc_pdata(hdr);
5513 bcopy(zio->io_data, hdr->b_l1hdr.b_pdata, psize);
5515 ASSERT3P(buf->b_data, ==, zio->io_orig_data);
5516 ASSERT3U(zio->io_orig_size, ==, HDR_GET_LSIZE(hdr));
5517 ASSERT3U(hdr->b_l1hdr.b_byteswap, ==, DMU_BSWAP_NUMFUNCS);
5518 ASSERT(!HDR_SHARED_DATA(hdr));
5519 ASSERT(!arc_buf_is_shared(buf));
5520 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5521 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
5524 * This hdr is not compressed so we're able to share
5525 * the arc_buf_t data buffer with the hdr.
5527 arc_share_buf(hdr, buf);
5528 VERIFY0(bcmp(zio->io_orig_data, hdr->b_l1hdr.b_pdata,
5529 HDR_GET_LSIZE(hdr)));
5531 arc_hdr_verify(hdr, zio->io_bp);
5535 arc_write_children_ready(zio_t *zio)
5537 arc_write_callback_t *callback = zio->io_private;
5538 arc_buf_t *buf = callback->awcb_buf;
5540 callback->awcb_children_ready(zio, buf, callback->awcb_private);
5544 * The SPA calls this callback for each physical write that happens on behalf
5545 * of a logical write. See the comment in dbuf_write_physdone() for details.
5548 arc_write_physdone(zio_t *zio)
5550 arc_write_callback_t *cb = zio->io_private;
5551 if (cb->awcb_physdone != NULL)
5552 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
5556 arc_write_done(zio_t *zio)
5558 arc_write_callback_t *callback = zio->io_private;
5559 arc_buf_t *buf = callback->awcb_buf;
5560 arc_buf_hdr_t *hdr = buf->b_hdr;
5562 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5564 if (zio->io_error == 0) {
5565 arc_hdr_verify(hdr, zio->io_bp);
5567 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5568 buf_discard_identity(hdr);
5570 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
5571 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
5574 ASSERT(HDR_EMPTY(hdr));
5578 * If the block to be written was all-zero or compressed enough to be
5579 * embedded in the BP, no write was performed so there will be no
5580 * dva/birth/checksum. The buffer must therefore remain anonymous
5583 if (!HDR_EMPTY(hdr)) {
5584 arc_buf_hdr_t *exists;
5585 kmutex_t *hash_lock;
5587 ASSERT(zio->io_error == 0);
5589 arc_cksum_verify(buf);
5591 exists = buf_hash_insert(hdr, &hash_lock);
5592 if (exists != NULL) {
5594 * This can only happen if we overwrite for
5595 * sync-to-convergence, because we remove
5596 * buffers from the hash table when we arc_free().
5598 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
5599 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5600 panic("bad overwrite, hdr=%p exists=%p",
5601 (void *)hdr, (void *)exists);
5602 ASSERT(refcount_is_zero(
5603 &exists->b_l1hdr.b_refcnt));
5604 arc_change_state(arc_anon, exists, hash_lock);
5605 mutex_exit(hash_lock);
5606 arc_hdr_destroy(exists);
5607 exists = buf_hash_insert(hdr, &hash_lock);
5608 ASSERT3P(exists, ==, NULL);
5609 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
5611 ASSERT(zio->io_prop.zp_nopwrite);
5612 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5613 panic("bad nopwrite, hdr=%p exists=%p",
5614 (void *)hdr, (void *)exists);
5617 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
5618 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
5619 ASSERT(BP_GET_DEDUP(zio->io_bp));
5620 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
5623 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5624 /* if it's not anon, we are doing a scrub */
5625 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
5626 arc_access(hdr, hash_lock);
5627 mutex_exit(hash_lock);
5629 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5632 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5633 callback->awcb_done(zio, buf, callback->awcb_private);
5635 kmem_free(callback, sizeof (arc_write_callback_t));
5639 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
5640 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready,
5641 arc_done_func_t *children_ready, arc_done_func_t *physdone,
5642 arc_done_func_t *done, void *private, zio_priority_t priority,
5643 int zio_flags, const zbookmark_phys_t *zb)
5645 arc_buf_hdr_t *hdr = buf->b_hdr;
5646 arc_write_callback_t *callback;
5649 ASSERT3P(ready, !=, NULL);
5650 ASSERT3P(done, !=, NULL);
5651 ASSERT(!HDR_IO_ERROR(hdr));
5652 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5653 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5654 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
5656 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5657 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
5658 callback->awcb_ready = ready;
5659 callback->awcb_children_ready = children_ready;
5660 callback->awcb_physdone = physdone;
5661 callback->awcb_done = done;
5662 callback->awcb_private = private;
5663 callback->awcb_buf = buf;
5666 * The hdr's b_pdata is now stale, free it now. A new data block
5667 * will be allocated when the zio pipeline calls arc_write_ready().
5669 if (hdr->b_l1hdr.b_pdata != NULL) {
5671 * If the buf is currently sharing the data block with
5672 * the hdr then we need to break that relationship here.
5673 * The hdr will remain with a NULL data pointer and the
5674 * buf will take sole ownership of the block.
5676 if (arc_buf_is_shared(buf)) {
5677 ASSERT(ARC_BUF_LAST(buf));
5678 arc_unshare_buf(hdr, buf);
5680 arc_hdr_free_pdata(hdr);
5682 VERIFY3P(buf->b_data, !=, NULL);
5683 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
5685 ASSERT(!arc_buf_is_shared(buf));
5686 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
5688 zio = zio_write(pio, spa, txg, bp, buf->b_data, HDR_GET_LSIZE(hdr), zp,
5690 (children_ready != NULL) ? arc_write_children_ready : NULL,
5691 arc_write_physdone, arc_write_done, callback,
5692 priority, zio_flags, zb);
5698 arc_memory_throttle(uint64_t reserve, uint64_t txg)
5701 uint64_t available_memory = ptob(freemem);
5702 static uint64_t page_load = 0;
5703 static uint64_t last_txg = 0;
5705 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
5707 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE)));
5710 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
5713 if (txg > last_txg) {
5718 * If we are in pageout, we know that memory is already tight,
5719 * the arc is already going to be evicting, so we just want to
5720 * continue to let page writes occur as quickly as possible.
5722 if (curproc == pageproc) {
5723 if (page_load > MAX(ptob(minfree), available_memory) / 4)
5724 return (SET_ERROR(ERESTART));
5725 /* Note: reserve is inflated, so we deflate */
5726 page_load += reserve / 8;
5728 } else if (page_load > 0 && arc_reclaim_needed()) {
5729 /* memory is low, delay before restarting */
5730 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
5731 return (SET_ERROR(EAGAIN));
5739 arc_tempreserve_clear(uint64_t reserve)
5741 atomic_add_64(&arc_tempreserve, -reserve);
5742 ASSERT((int64_t)arc_tempreserve >= 0);
5746 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
5751 if (reserve > arc_c/4 && !arc_no_grow) {
5752 arc_c = MIN(arc_c_max, reserve * 4);
5753 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
5755 if (reserve > arc_c)
5756 return (SET_ERROR(ENOMEM));
5759 * Don't count loaned bufs as in flight dirty data to prevent long
5760 * network delays from blocking transactions that are ready to be
5761 * assigned to a txg.
5763 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
5764 arc_loaned_bytes), 0);
5767 * Writes will, almost always, require additional memory allocations
5768 * in order to compress/encrypt/etc the data. We therefore need to
5769 * make sure that there is sufficient available memory for this.
5771 error = arc_memory_throttle(reserve, txg);
5776 * Throttle writes when the amount of dirty data in the cache
5777 * gets too large. We try to keep the cache less than half full
5778 * of dirty blocks so that our sync times don't grow too large.
5779 * Note: if two requests come in concurrently, we might let them
5780 * both succeed, when one of them should fail. Not a huge deal.
5783 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
5784 anon_size > arc_c / 4) {
5785 uint64_t meta_esize =
5786 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
5787 uint64_t data_esize =
5788 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
5789 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
5790 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
5791 arc_tempreserve >> 10, meta_esize >> 10,
5792 data_esize >> 10, reserve >> 10, arc_c >> 10);
5793 return (SET_ERROR(ERESTART));
5795 atomic_add_64(&arc_tempreserve, reserve);
5800 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
5801 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
5803 size->value.ui64 = refcount_count(&state->arcs_size);
5804 evict_data->value.ui64 =
5805 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
5806 evict_metadata->value.ui64 =
5807 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
5811 arc_kstat_update(kstat_t *ksp, int rw)
5813 arc_stats_t *as = ksp->ks_data;
5815 if (rw == KSTAT_WRITE) {
5818 arc_kstat_update_state(arc_anon,
5819 &as->arcstat_anon_size,
5820 &as->arcstat_anon_evictable_data,
5821 &as->arcstat_anon_evictable_metadata);
5822 arc_kstat_update_state(arc_mru,
5823 &as->arcstat_mru_size,
5824 &as->arcstat_mru_evictable_data,
5825 &as->arcstat_mru_evictable_metadata);
5826 arc_kstat_update_state(arc_mru_ghost,
5827 &as->arcstat_mru_ghost_size,
5828 &as->arcstat_mru_ghost_evictable_data,
5829 &as->arcstat_mru_ghost_evictable_metadata);
5830 arc_kstat_update_state(arc_mfu,
5831 &as->arcstat_mfu_size,
5832 &as->arcstat_mfu_evictable_data,
5833 &as->arcstat_mfu_evictable_metadata);
5834 arc_kstat_update_state(arc_mfu_ghost,
5835 &as->arcstat_mfu_ghost_size,
5836 &as->arcstat_mfu_ghost_evictable_data,
5837 &as->arcstat_mfu_ghost_evictable_metadata);
5844 * This function *must* return indices evenly distributed between all
5845 * sublists of the multilist. This is needed due to how the ARC eviction
5846 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
5847 * distributed between all sublists and uses this assumption when
5848 * deciding which sublist to evict from and how much to evict from it.
5851 arc_state_multilist_index_func(multilist_t *ml, void *obj)
5853 arc_buf_hdr_t *hdr = obj;
5856 * We rely on b_dva to generate evenly distributed index
5857 * numbers using buf_hash below. So, as an added precaution,
5858 * let's make sure we never add empty buffers to the arc lists.
5860 ASSERT(!HDR_EMPTY(hdr));
5863 * The assumption here, is the hash value for a given
5864 * arc_buf_hdr_t will remain constant throughout it's lifetime
5865 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
5866 * Thus, we don't need to store the header's sublist index
5867 * on insertion, as this index can be recalculated on removal.
5869 * Also, the low order bits of the hash value are thought to be
5870 * distributed evenly. Otherwise, in the case that the multilist
5871 * has a power of two number of sublists, each sublists' usage
5872 * would not be evenly distributed.
5874 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
5875 multilist_get_num_sublists(ml));
5879 static eventhandler_tag arc_event_lowmem = NULL;
5882 arc_lowmem(void *arg __unused, int howto __unused)
5885 mutex_enter(&arc_reclaim_lock);
5886 /* XXX: Memory deficit should be passed as argument. */
5887 needfree = btoc(arc_c >> arc_shrink_shift);
5888 DTRACE_PROBE(arc__needfree);
5889 cv_signal(&arc_reclaim_thread_cv);
5892 * It is unsafe to block here in arbitrary threads, because we can come
5893 * here from ARC itself and may hold ARC locks and thus risk a deadlock
5894 * with ARC reclaim thread.
5896 if (curproc == pageproc)
5897 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
5898 mutex_exit(&arc_reclaim_lock);
5903 arc_state_init(void)
5905 arc_anon = &ARC_anon;
5907 arc_mru_ghost = &ARC_mru_ghost;
5909 arc_mfu_ghost = &ARC_mfu_ghost;
5910 arc_l2c_only = &ARC_l2c_only;
5912 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
5913 sizeof (arc_buf_hdr_t),
5914 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5915 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5916 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
5917 sizeof (arc_buf_hdr_t),
5918 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5919 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5920 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
5921 sizeof (arc_buf_hdr_t),
5922 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5923 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5924 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
5925 sizeof (arc_buf_hdr_t),
5926 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5927 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5928 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
5929 sizeof (arc_buf_hdr_t),
5930 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5931 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5932 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
5933 sizeof (arc_buf_hdr_t),
5934 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5935 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5936 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
5937 sizeof (arc_buf_hdr_t),
5938 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5939 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5940 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
5941 sizeof (arc_buf_hdr_t),
5942 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5943 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5944 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
5945 sizeof (arc_buf_hdr_t),
5946 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5947 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5948 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
5949 sizeof (arc_buf_hdr_t),
5950 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5951 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5953 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
5954 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
5955 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
5956 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
5957 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
5958 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
5959 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
5960 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
5961 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
5962 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
5963 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
5964 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
5966 refcount_create(&arc_anon->arcs_size);
5967 refcount_create(&arc_mru->arcs_size);
5968 refcount_create(&arc_mru_ghost->arcs_size);
5969 refcount_create(&arc_mfu->arcs_size);
5970 refcount_create(&arc_mfu_ghost->arcs_size);
5971 refcount_create(&arc_l2c_only->arcs_size);
5975 arc_state_fini(void)
5977 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
5978 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
5979 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
5980 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
5981 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
5982 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
5983 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
5984 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
5985 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
5986 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
5987 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
5988 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
5990 refcount_destroy(&arc_anon->arcs_size);
5991 refcount_destroy(&arc_mru->arcs_size);
5992 refcount_destroy(&arc_mru_ghost->arcs_size);
5993 refcount_destroy(&arc_mfu->arcs_size);
5994 refcount_destroy(&arc_mfu_ghost->arcs_size);
5995 refcount_destroy(&arc_l2c_only->arcs_size);
5997 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
5998 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
5999 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
6000 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
6001 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
6002 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
6003 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
6004 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
6016 int i, prefetch_tunable_set = 0;
6018 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
6019 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
6020 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
6022 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
6023 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL);
6025 /* Convert seconds to clock ticks */
6026 arc_min_prefetch_lifespan = 1 * hz;
6028 /* Start out with 1/8 of all memory */
6029 arc_c = kmem_size() / 8;
6034 * On architectures where the physical memory can be larger
6035 * than the addressable space (intel in 32-bit mode), we may
6036 * need to limit the cache to 1/8 of VM size.
6038 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
6040 #endif /* illumos */
6041 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
6042 arc_c_min = MAX(arc_c / 4, arc_abs_min);
6043 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
6044 if (arc_c * 8 >= 1 << 30)
6045 arc_c_max = (arc_c * 8) - (1 << 30);
6047 arc_c_max = arc_c_min;
6048 arc_c_max = MAX(arc_c * 5, arc_c_max);
6051 * In userland, there's only the memory pressure that we artificially
6052 * create (see arc_available_memory()). Don't let arc_c get too
6053 * small, because it can cause transactions to be larger than
6054 * arc_c, causing arc_tempreserve_space() to fail.
6057 arc_c_min = arc_c_max / 2;
6062 * Allow the tunables to override our calculations if they are
6065 if (zfs_arc_max > arc_abs_min && zfs_arc_max < kmem_size())
6066 arc_c_max = zfs_arc_max;
6067 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
6068 arc_c_min = zfs_arc_min;
6072 arc_p = (arc_c >> 1);
6075 /* limit meta-data to 1/4 of the arc capacity */
6076 arc_meta_limit = arc_c_max / 4;
6078 /* Allow the tunable to override if it is reasonable */
6079 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
6080 arc_meta_limit = zfs_arc_meta_limit;
6082 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
6083 arc_c_min = arc_meta_limit / 2;
6085 if (zfs_arc_meta_min > 0) {
6086 arc_meta_min = zfs_arc_meta_min;
6088 arc_meta_min = arc_c_min / 2;
6091 if (zfs_arc_grow_retry > 0)
6092 arc_grow_retry = zfs_arc_grow_retry;
6094 if (zfs_arc_shrink_shift > 0)
6095 arc_shrink_shift = zfs_arc_shrink_shift;
6098 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6100 if (arc_no_grow_shift >= arc_shrink_shift)
6101 arc_no_grow_shift = arc_shrink_shift - 1;
6103 if (zfs_arc_p_min_shift > 0)
6104 arc_p_min_shift = zfs_arc_p_min_shift;
6106 if (zfs_arc_num_sublists_per_state < 1)
6107 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1);
6109 /* if kmem_flags are set, lets try to use less memory */
6110 if (kmem_debugging())
6112 if (arc_c < arc_c_min)
6115 zfs_arc_min = arc_c_min;
6116 zfs_arc_max = arc_c_max;
6121 arc_reclaim_thread_exit = B_FALSE;
6122 arc_dnlc_evicts_thread_exit = FALSE;
6124 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
6125 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
6127 if (arc_ksp != NULL) {
6128 arc_ksp->ks_data = &arc_stats;
6129 arc_ksp->ks_update = arc_kstat_update;
6130 kstat_install(arc_ksp);
6133 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
6134 TS_RUN, minclsyspri);
6137 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
6138 EVENTHANDLER_PRI_FIRST);
6141 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0,
6142 TS_RUN, minclsyspri);
6148 * Calculate maximum amount of dirty data per pool.
6150 * If it has been set by /etc/system, take that.
6151 * Otherwise, use a percentage of physical memory defined by
6152 * zfs_dirty_data_max_percent (default 10%) with a cap at
6153 * zfs_dirty_data_max_max (default 4GB).
6155 if (zfs_dirty_data_max == 0) {
6156 zfs_dirty_data_max = ptob(physmem) *
6157 zfs_dirty_data_max_percent / 100;
6158 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
6159 zfs_dirty_data_max_max);
6163 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
6164 prefetch_tunable_set = 1;
6167 if (prefetch_tunable_set == 0) {
6168 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
6170 printf(" add \"vfs.zfs.prefetch_disable=0\" "
6171 "to /boot/loader.conf.\n");
6172 zfs_prefetch_disable = 1;
6175 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
6176 prefetch_tunable_set == 0) {
6177 printf("ZFS NOTICE: Prefetch is disabled by default if less "
6178 "than 4GB of RAM is present;\n"
6179 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
6180 "to /boot/loader.conf.\n");
6181 zfs_prefetch_disable = 1;
6184 /* Warn about ZFS memory and address space requirements. */
6185 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
6186 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
6187 "expect unstable behavior.\n");
6189 if (kmem_size() < 512 * (1 << 20)) {
6190 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
6191 "expect unstable behavior.\n");
6192 printf(" Consider tuning vm.kmem_size and "
6193 "vm.kmem_size_max\n");
6194 printf(" in /boot/loader.conf.\n");
6202 mutex_enter(&arc_reclaim_lock);
6203 arc_reclaim_thread_exit = B_TRUE;
6205 * The reclaim thread will set arc_reclaim_thread_exit back to
6206 * B_FALSE when it is finished exiting; we're waiting for that.
6208 while (arc_reclaim_thread_exit) {
6209 cv_signal(&arc_reclaim_thread_cv);
6210 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
6212 mutex_exit(&arc_reclaim_lock);
6214 /* Use B_TRUE to ensure *all* buffers are evicted */
6215 arc_flush(NULL, B_TRUE);
6217 mutex_enter(&arc_dnlc_evicts_lock);
6218 arc_dnlc_evicts_thread_exit = TRUE;
6220 * The user evicts thread will set arc_user_evicts_thread_exit
6221 * to FALSE when it is finished exiting; we're waiting for that.
6223 while (arc_dnlc_evicts_thread_exit) {
6224 cv_signal(&arc_dnlc_evicts_cv);
6225 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
6227 mutex_exit(&arc_dnlc_evicts_lock);
6231 if (arc_ksp != NULL) {
6232 kstat_delete(arc_ksp);
6236 mutex_destroy(&arc_reclaim_lock);
6237 cv_destroy(&arc_reclaim_thread_cv);
6238 cv_destroy(&arc_reclaim_waiters_cv);
6240 mutex_destroy(&arc_dnlc_evicts_lock);
6241 cv_destroy(&arc_dnlc_evicts_cv);
6246 ASSERT0(arc_loaned_bytes);
6249 if (arc_event_lowmem != NULL)
6250 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
6257 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6258 * It uses dedicated storage devices to hold cached data, which are populated
6259 * using large infrequent writes. The main role of this cache is to boost
6260 * the performance of random read workloads. The intended L2ARC devices
6261 * include short-stroked disks, solid state disks, and other media with
6262 * substantially faster read latency than disk.
6264 * +-----------------------+
6266 * +-----------------------+
6269 * l2arc_feed_thread() arc_read()
6273 * +---------------+ |
6275 * +---------------+ |
6280 * +-------+ +-------+
6282 * | cache | | cache |
6283 * +-------+ +-------+
6284 * +=========+ .-----.
6285 * : L2ARC : |-_____-|
6286 * : devices : | Disks |
6287 * +=========+ `-_____-'
6289 * Read requests are satisfied from the following sources, in order:
6292 * 2) vdev cache of L2ARC devices
6294 * 4) vdev cache of disks
6297 * Some L2ARC device types exhibit extremely slow write performance.
6298 * To accommodate for this there are some significant differences between
6299 * the L2ARC and traditional cache design:
6301 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6302 * the ARC behave as usual, freeing buffers and placing headers on ghost
6303 * lists. The ARC does not send buffers to the L2ARC during eviction as
6304 * this would add inflated write latencies for all ARC memory pressure.
6306 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6307 * It does this by periodically scanning buffers from the eviction-end of
6308 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6309 * not already there. It scans until a headroom of buffers is satisfied,
6310 * which itself is a buffer for ARC eviction. If a compressible buffer is
6311 * found during scanning and selected for writing to an L2ARC device, we
6312 * temporarily boost scanning headroom during the next scan cycle to make
6313 * sure we adapt to compression effects (which might significantly reduce
6314 * the data volume we write to L2ARC). The thread that does this is
6315 * l2arc_feed_thread(), illustrated below; example sizes are included to
6316 * provide a better sense of ratio than this diagram:
6319 * +---------------------+----------+
6320 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6321 * +---------------------+----------+ | o L2ARC eligible
6322 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6323 * +---------------------+----------+ |
6324 * 15.9 Gbytes ^ 32 Mbytes |
6326 * l2arc_feed_thread()
6328 * l2arc write hand <--[oooo]--'
6332 * +==============================+
6333 * L2ARC dev |####|#|###|###| |####| ... |
6334 * +==============================+
6337 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6338 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6339 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6340 * safe to say that this is an uncommon case, since buffers at the end of
6341 * the ARC lists have moved there due to inactivity.
6343 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6344 * then the L2ARC simply misses copying some buffers. This serves as a
6345 * pressure valve to prevent heavy read workloads from both stalling the ARC
6346 * with waits and clogging the L2ARC with writes. This also helps prevent
6347 * the potential for the L2ARC to churn if it attempts to cache content too
6348 * quickly, such as during backups of the entire pool.
6350 * 5. After system boot and before the ARC has filled main memory, there are
6351 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6352 * lists can remain mostly static. Instead of searching from tail of these
6353 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6354 * for eligible buffers, greatly increasing its chance of finding them.
6356 * The L2ARC device write speed is also boosted during this time so that
6357 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6358 * there are no L2ARC reads, and no fear of degrading read performance
6359 * through increased writes.
6361 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6362 * the vdev queue can aggregate them into larger and fewer writes. Each
6363 * device is written to in a rotor fashion, sweeping writes through
6364 * available space then repeating.
6366 * 7. The L2ARC does not store dirty content. It never needs to flush
6367 * write buffers back to disk based storage.
6369 * 8. If an ARC buffer is written (and dirtied) which also exists in the
6370 * L2ARC, the now stale L2ARC buffer is immediately dropped.
6372 * The performance of the L2ARC can be tweaked by a number of tunables, which
6373 * may be necessary for different workloads:
6375 * l2arc_write_max max write bytes per interval
6376 * l2arc_write_boost extra write bytes during device warmup
6377 * l2arc_noprefetch skip caching prefetched buffers
6378 * l2arc_headroom number of max device writes to precache
6379 * l2arc_headroom_boost when we find compressed buffers during ARC
6380 * scanning, we multiply headroom by this
6381 * percentage factor for the next scan cycle,
6382 * since more compressed buffers are likely to
6384 * l2arc_feed_secs seconds between L2ARC writing
6386 * Tunables may be removed or added as future performance improvements are
6387 * integrated, and also may become zpool properties.
6389 * There are three key functions that control how the L2ARC warms up:
6391 * l2arc_write_eligible() check if a buffer is eligible to cache
6392 * l2arc_write_size() calculate how much to write
6393 * l2arc_write_interval() calculate sleep delay between writes
6395 * These three functions determine what to write, how much, and how quickly
6400 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
6403 * A buffer is *not* eligible for the L2ARC if it:
6404 * 1. belongs to a different spa.
6405 * 2. is already cached on the L2ARC.
6406 * 3. has an I/O in progress (it may be an incomplete read).
6407 * 4. is flagged not eligible (zfs property).
6409 if (hdr->b_spa != spa_guid) {
6410 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
6413 if (HDR_HAS_L2HDR(hdr)) {
6414 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
6417 if (HDR_IO_IN_PROGRESS(hdr)) {
6418 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
6421 if (!HDR_L2CACHE(hdr)) {
6422 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
6430 l2arc_write_size(void)
6435 * Make sure our globals have meaningful values in case the user
6438 size = l2arc_write_max;
6440 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
6441 "be greater than zero, resetting it to the default (%d)",
6443 size = l2arc_write_max = L2ARC_WRITE_SIZE;
6446 if (arc_warm == B_FALSE)
6447 size += l2arc_write_boost;
6454 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
6456 clock_t interval, next, now;
6459 * If the ARC lists are busy, increase our write rate; if the
6460 * lists are stale, idle back. This is achieved by checking
6461 * how much we previously wrote - if it was more than half of
6462 * what we wanted, schedule the next write much sooner.
6464 if (l2arc_feed_again && wrote > (wanted / 2))
6465 interval = (hz * l2arc_feed_min_ms) / 1000;
6467 interval = hz * l2arc_feed_secs;
6469 now = ddi_get_lbolt();
6470 next = MAX(now, MIN(now + interval, began + interval));
6476 * Cycle through L2ARC devices. This is how L2ARC load balances.
6477 * If a device is returned, this also returns holding the spa config lock.
6479 static l2arc_dev_t *
6480 l2arc_dev_get_next(void)
6482 l2arc_dev_t *first, *next = NULL;
6485 * Lock out the removal of spas (spa_namespace_lock), then removal
6486 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
6487 * both locks will be dropped and a spa config lock held instead.
6489 mutex_enter(&spa_namespace_lock);
6490 mutex_enter(&l2arc_dev_mtx);
6492 /* if there are no vdevs, there is nothing to do */
6493 if (l2arc_ndev == 0)
6497 next = l2arc_dev_last;
6499 /* loop around the list looking for a non-faulted vdev */
6501 next = list_head(l2arc_dev_list);
6503 next = list_next(l2arc_dev_list, next);
6505 next = list_head(l2arc_dev_list);
6508 /* if we have come back to the start, bail out */
6511 else if (next == first)
6514 } while (vdev_is_dead(next->l2ad_vdev));
6516 /* if we were unable to find any usable vdevs, return NULL */
6517 if (vdev_is_dead(next->l2ad_vdev))
6520 l2arc_dev_last = next;
6523 mutex_exit(&l2arc_dev_mtx);
6526 * Grab the config lock to prevent the 'next' device from being
6527 * removed while we are writing to it.
6530 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
6531 mutex_exit(&spa_namespace_lock);
6537 * Free buffers that were tagged for destruction.
6540 l2arc_do_free_on_write()
6543 l2arc_data_free_t *df, *df_prev;
6545 mutex_enter(&l2arc_free_on_write_mtx);
6546 buflist = l2arc_free_on_write;
6548 for (df = list_tail(buflist); df; df = df_prev) {
6549 df_prev = list_prev(buflist, df);
6550 ASSERT3P(df->l2df_data, !=, NULL);
6551 if (df->l2df_type == ARC_BUFC_METADATA) {
6552 zio_buf_free(df->l2df_data, df->l2df_size);
6554 ASSERT(df->l2df_type == ARC_BUFC_DATA);
6555 zio_data_buf_free(df->l2df_data, df->l2df_size);
6557 list_remove(buflist, df);
6558 kmem_free(df, sizeof (l2arc_data_free_t));
6561 mutex_exit(&l2arc_free_on_write_mtx);
6565 * A write to a cache device has completed. Update all headers to allow
6566 * reads from these buffers to begin.
6569 l2arc_write_done(zio_t *zio)
6571 l2arc_write_callback_t *cb;
6574 arc_buf_hdr_t *head, *hdr, *hdr_prev;
6575 kmutex_t *hash_lock;
6576 int64_t bytes_dropped = 0;
6578 cb = zio->io_private;
6579 ASSERT3P(cb, !=, NULL);
6580 dev = cb->l2wcb_dev;
6581 ASSERT3P(dev, !=, NULL);
6582 head = cb->l2wcb_head;
6583 ASSERT3P(head, !=, NULL);
6584 buflist = &dev->l2ad_buflist;
6585 ASSERT3P(buflist, !=, NULL);
6586 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
6587 l2arc_write_callback_t *, cb);
6589 if (zio->io_error != 0)
6590 ARCSTAT_BUMP(arcstat_l2_writes_error);
6593 * All writes completed, or an error was hit.
6596 mutex_enter(&dev->l2ad_mtx);
6597 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
6598 hdr_prev = list_prev(buflist, hdr);
6600 hash_lock = HDR_LOCK(hdr);
6603 * We cannot use mutex_enter or else we can deadlock
6604 * with l2arc_write_buffers (due to swapping the order
6605 * the hash lock and l2ad_mtx are taken).
6607 if (!mutex_tryenter(hash_lock)) {
6609 * Missed the hash lock. We must retry so we
6610 * don't leave the ARC_FLAG_L2_WRITING bit set.
6612 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
6615 * We don't want to rescan the headers we've
6616 * already marked as having been written out, so
6617 * we reinsert the head node so we can pick up
6618 * where we left off.
6620 list_remove(buflist, head);
6621 list_insert_after(buflist, hdr, head);
6623 mutex_exit(&dev->l2ad_mtx);
6626 * We wait for the hash lock to become available
6627 * to try and prevent busy waiting, and increase
6628 * the chance we'll be able to acquire the lock
6629 * the next time around.
6631 mutex_enter(hash_lock);
6632 mutex_exit(hash_lock);
6637 * We could not have been moved into the arc_l2c_only
6638 * state while in-flight due to our ARC_FLAG_L2_WRITING
6639 * bit being set. Let's just ensure that's being enforced.
6641 ASSERT(HDR_HAS_L1HDR(hdr));
6643 if (zio->io_error != 0) {
6645 * Error - drop L2ARC entry.
6647 list_remove(buflist, hdr);
6649 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
6651 ARCSTAT_INCR(arcstat_l2_asize, -arc_hdr_size(hdr));
6652 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr));
6654 bytes_dropped += arc_hdr_size(hdr);
6655 (void) refcount_remove_many(&dev->l2ad_alloc,
6656 arc_hdr_size(hdr), hdr);
6660 * Allow ARC to begin reads and ghost list evictions to
6663 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
6665 mutex_exit(hash_lock);
6668 atomic_inc_64(&l2arc_writes_done);
6669 list_remove(buflist, head);
6670 ASSERT(!HDR_HAS_L1HDR(head));
6671 kmem_cache_free(hdr_l2only_cache, head);
6672 mutex_exit(&dev->l2ad_mtx);
6674 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
6676 l2arc_do_free_on_write();
6678 kmem_free(cb, sizeof (l2arc_write_callback_t));
6682 * A read to a cache device completed. Validate buffer contents before
6683 * handing over to the regular ARC routines.
6686 l2arc_read_done(zio_t *zio)
6688 l2arc_read_callback_t *cb;
6690 kmutex_t *hash_lock;
6691 boolean_t valid_cksum;
6693 ASSERT3P(zio->io_vd, !=, NULL);
6694 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
6696 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
6698 cb = zio->io_private;
6699 ASSERT3P(cb, !=, NULL);
6700 hdr = cb->l2rcb_hdr;
6701 ASSERT3P(hdr, !=, NULL);
6703 hash_lock = HDR_LOCK(hdr);
6704 mutex_enter(hash_lock);
6705 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6708 * If the data was read into a temporary buffer,
6709 * move it and free the buffer.
6711 if (cb->l2rcb_data != NULL) {
6712 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
6713 if (zio->io_error == 0) {
6714 bcopy(cb->l2rcb_data, hdr->b_l1hdr.b_pdata,
6719 * The following must be done regardless of whether
6720 * there was an error:
6721 * - free the temporary buffer
6722 * - point zio to the real ARC buffer
6723 * - set zio size accordingly
6724 * These are required because zio is either re-used for
6725 * an I/O of the block in the case of the error
6726 * or the zio is passed to arc_read_done() and it
6729 zio_data_buf_free(cb->l2rcb_data, zio->io_size);
6730 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
6731 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_pdata;
6734 ASSERT3P(zio->io_data, !=, NULL);
6737 * Check this survived the L2ARC journey.
6739 ASSERT3P(zio->io_data, ==, hdr->b_l1hdr.b_pdata);
6740 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
6741 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
6743 valid_cksum = arc_cksum_is_equal(hdr, zio);
6744 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
6745 mutex_exit(hash_lock);
6746 zio->io_private = hdr;
6749 mutex_exit(hash_lock);
6751 * Buffer didn't survive caching. Increment stats and
6752 * reissue to the original storage device.
6754 if (zio->io_error != 0) {
6755 ARCSTAT_BUMP(arcstat_l2_io_error);
6757 zio->io_error = SET_ERROR(EIO);
6760 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
6763 * If there's no waiter, issue an async i/o to the primary
6764 * storage now. If there *is* a waiter, the caller must
6765 * issue the i/o in a context where it's OK to block.
6767 if (zio->io_waiter == NULL) {
6768 zio_t *pio = zio_unique_parent(zio);
6770 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
6772 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
6773 hdr->b_l1hdr.b_pdata, zio->io_size, arc_read_done,
6774 hdr, zio->io_priority, cb->l2rcb_flags,
6779 kmem_free(cb, sizeof (l2arc_read_callback_t));
6783 * This is the list priority from which the L2ARC will search for pages to
6784 * cache. This is used within loops (0..3) to cycle through lists in the
6785 * desired order. This order can have a significant effect on cache
6788 * Currently the metadata lists are hit first, MFU then MRU, followed by
6789 * the data lists. This function returns a locked list, and also returns
6792 static multilist_sublist_t *
6793 l2arc_sublist_lock(int list_num)
6795 multilist_t *ml = NULL;
6798 ASSERT(list_num >= 0 && list_num <= 3);
6802 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
6805 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
6808 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
6811 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
6816 * Return a randomly-selected sublist. This is acceptable
6817 * because the caller feeds only a little bit of data for each
6818 * call (8MB). Subsequent calls will result in different
6819 * sublists being selected.
6821 idx = multilist_get_random_index(ml);
6822 return (multilist_sublist_lock(ml, idx));
6826 * Evict buffers from the device write hand to the distance specified in
6827 * bytes. This distance may span populated buffers, it may span nothing.
6828 * This is clearing a region on the L2ARC device ready for writing.
6829 * If the 'all' boolean is set, every buffer is evicted.
6832 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
6835 arc_buf_hdr_t *hdr, *hdr_prev;
6836 kmutex_t *hash_lock;
6839 buflist = &dev->l2ad_buflist;
6841 if (!all && dev->l2ad_first) {
6843 * This is the first sweep through the device. There is
6849 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
6851 * When nearing the end of the device, evict to the end
6852 * before the device write hand jumps to the start.
6854 taddr = dev->l2ad_end;
6856 taddr = dev->l2ad_hand + distance;
6858 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
6859 uint64_t, taddr, boolean_t, all);
6862 mutex_enter(&dev->l2ad_mtx);
6863 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
6864 hdr_prev = list_prev(buflist, hdr);
6866 hash_lock = HDR_LOCK(hdr);
6869 * We cannot use mutex_enter or else we can deadlock
6870 * with l2arc_write_buffers (due to swapping the order
6871 * the hash lock and l2ad_mtx are taken).
6873 if (!mutex_tryenter(hash_lock)) {
6875 * Missed the hash lock. Retry.
6877 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
6878 mutex_exit(&dev->l2ad_mtx);
6879 mutex_enter(hash_lock);
6880 mutex_exit(hash_lock);
6884 if (HDR_L2_WRITE_HEAD(hdr)) {
6886 * We hit a write head node. Leave it for
6887 * l2arc_write_done().
6889 list_remove(buflist, hdr);
6890 mutex_exit(hash_lock);
6894 if (!all && HDR_HAS_L2HDR(hdr) &&
6895 (hdr->b_l2hdr.b_daddr > taddr ||
6896 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
6898 * We've evicted to the target address,
6899 * or the end of the device.
6901 mutex_exit(hash_lock);
6905 ASSERT(HDR_HAS_L2HDR(hdr));
6906 if (!HDR_HAS_L1HDR(hdr)) {
6907 ASSERT(!HDR_L2_READING(hdr));
6909 * This doesn't exist in the ARC. Destroy.
6910 * arc_hdr_destroy() will call list_remove()
6911 * and decrement arcstat_l2_size.
6913 arc_change_state(arc_anon, hdr, hash_lock);
6914 arc_hdr_destroy(hdr);
6916 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
6917 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
6919 * Invalidate issued or about to be issued
6920 * reads, since we may be about to write
6921 * over this location.
6923 if (HDR_L2_READING(hdr)) {
6924 ARCSTAT_BUMP(arcstat_l2_evict_reading);
6925 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
6928 /* Ensure this header has finished being written */
6929 ASSERT(!HDR_L2_WRITING(hdr));
6931 arc_hdr_l2hdr_destroy(hdr);
6933 mutex_exit(hash_lock);
6935 mutex_exit(&dev->l2ad_mtx);
6939 * Find and write ARC buffers to the L2ARC device.
6941 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
6942 * for reading until they have completed writing.
6943 * The headroom_boost is an in-out parameter used to maintain headroom boost
6944 * state between calls to this function.
6946 * Returns the number of bytes actually written (which may be smaller than
6947 * the delta by which the device hand has changed due to alignment).
6950 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
6952 arc_buf_hdr_t *hdr, *hdr_prev, *head;
6953 uint64_t write_asize, write_psize, write_sz, headroom;
6955 l2arc_write_callback_t *cb;
6957 uint64_t guid = spa_load_guid(spa);
6960 ASSERT3P(dev->l2ad_vdev, !=, NULL);
6963 write_sz = write_asize = write_psize = 0;
6965 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
6966 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
6968 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
6970 * Copy buffers for L2ARC writing.
6972 for (try = 0; try <= 3; try++) {
6973 multilist_sublist_t *mls = l2arc_sublist_lock(try);
6974 uint64_t passed_sz = 0;
6976 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
6979 * L2ARC fast warmup.
6981 * Until the ARC is warm and starts to evict, read from the
6982 * head of the ARC lists rather than the tail.
6984 if (arc_warm == B_FALSE)
6985 hdr = multilist_sublist_head(mls);
6987 hdr = multilist_sublist_tail(mls);
6989 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
6991 headroom = target_sz * l2arc_headroom;
6992 if (zfs_compressed_arc_enabled)
6993 headroom = (headroom * l2arc_headroom_boost) / 100;
6995 for (; hdr; hdr = hdr_prev) {
6996 kmutex_t *hash_lock;
6998 if (arc_warm == B_FALSE)
6999 hdr_prev = multilist_sublist_next(mls, hdr);
7001 hdr_prev = multilist_sublist_prev(mls, hdr);
7002 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned,
7003 HDR_GET_LSIZE(hdr));
7005 hash_lock = HDR_LOCK(hdr);
7006 if (!mutex_tryenter(hash_lock)) {
7007 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
7009 * Skip this buffer rather than waiting.
7014 passed_sz += HDR_GET_LSIZE(hdr);
7015 if (passed_sz > headroom) {
7019 mutex_exit(hash_lock);
7020 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
7024 if (!l2arc_write_eligible(guid, hdr)) {
7025 mutex_exit(hash_lock);
7029 if ((write_asize + HDR_GET_LSIZE(hdr)) > target_sz) {
7031 mutex_exit(hash_lock);
7032 ARCSTAT_BUMP(arcstat_l2_write_full);
7038 * Insert a dummy header on the buflist so
7039 * l2arc_write_done() can find where the
7040 * write buffers begin without searching.
7042 mutex_enter(&dev->l2ad_mtx);
7043 list_insert_head(&dev->l2ad_buflist, head);
7044 mutex_exit(&dev->l2ad_mtx);
7047 sizeof (l2arc_write_callback_t), KM_SLEEP);
7048 cb->l2wcb_dev = dev;
7049 cb->l2wcb_head = head;
7050 pio = zio_root(spa, l2arc_write_done, cb,
7052 ARCSTAT_BUMP(arcstat_l2_write_pios);
7055 hdr->b_l2hdr.b_dev = dev;
7056 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
7057 arc_hdr_set_flags(hdr,
7058 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
7060 mutex_enter(&dev->l2ad_mtx);
7061 list_insert_head(&dev->l2ad_buflist, hdr);
7062 mutex_exit(&dev->l2ad_mtx);
7065 * We rely on the L1 portion of the header below, so
7066 * it's invalid for this header to have been evicted out
7067 * of the ghost cache, prior to being written out. The
7068 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
7070 ASSERT(HDR_HAS_L1HDR(hdr));
7072 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
7073 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
7074 ASSERT3U(arc_hdr_size(hdr), >, 0);
7075 uint64_t size = arc_hdr_size(hdr);
7076 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
7079 (void) refcount_add_many(&dev->l2ad_alloc, size, hdr);
7082 * Normally the L2ARC can use the hdr's data, but if
7083 * we're sharing data between the hdr and one of its
7084 * bufs, L2ARC needs its own copy of the data so that
7085 * the ZIO below can't race with the buf consumer. To
7086 * ensure that this copy will be available for the
7087 * lifetime of the ZIO and be cleaned up afterwards, we
7088 * add it to the l2arc_free_on_write queue.
7091 if (!HDR_SHARED_DATA(hdr) && size == asize) {
7092 to_write = hdr->b_l1hdr.b_pdata;
7094 arc_buf_contents_t type = arc_buf_type(hdr);
7095 if (type == ARC_BUFC_METADATA) {
7096 to_write = zio_buf_alloc(asize);
7098 ASSERT3U(type, ==, ARC_BUFC_DATA);
7099 to_write = zio_data_buf_alloc(asize);
7102 bcopy(hdr->b_l1hdr.b_pdata, to_write, size);
7104 bzero(to_write + size, asize - size);
7105 l2arc_free_data_on_write(to_write, asize, type);
7107 wzio = zio_write_phys(pio, dev->l2ad_vdev,
7108 hdr->b_l2hdr.b_daddr, asize, to_write,
7109 ZIO_CHECKSUM_OFF, NULL, hdr,
7110 ZIO_PRIORITY_ASYNC_WRITE,
7111 ZIO_FLAG_CANFAIL, B_FALSE);
7113 write_sz += HDR_GET_LSIZE(hdr);
7114 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
7117 write_asize += size;
7118 write_psize += asize;
7119 dev->l2ad_hand += asize;
7121 mutex_exit(hash_lock);
7123 (void) zio_nowait(wzio);
7126 multilist_sublist_unlock(mls);
7132 /* No buffers selected for writing? */
7135 ASSERT(!HDR_HAS_L1HDR(head));
7136 kmem_cache_free(hdr_l2only_cache, head);
7140 ASSERT3U(write_asize, <=, target_sz);
7141 ARCSTAT_BUMP(arcstat_l2_writes_sent);
7142 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
7143 ARCSTAT_INCR(arcstat_l2_size, write_sz);
7144 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
7145 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0);
7148 * Bump device hand to the device start if it is approaching the end.
7149 * l2arc_evict() will already have evicted ahead for this case.
7151 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
7152 dev->l2ad_hand = dev->l2ad_start;
7153 dev->l2ad_first = B_FALSE;
7156 dev->l2ad_writing = B_TRUE;
7157 (void) zio_wait(pio);
7158 dev->l2ad_writing = B_FALSE;
7160 return (write_asize);
7164 * This thread feeds the L2ARC at regular intervals. This is the beating
7165 * heart of the L2ARC.
7168 l2arc_feed_thread(void *dummy __unused)
7173 uint64_t size, wrote;
7174 clock_t begin, next = ddi_get_lbolt();
7176 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
7178 mutex_enter(&l2arc_feed_thr_lock);
7180 while (l2arc_thread_exit == 0) {
7181 CALLB_CPR_SAFE_BEGIN(&cpr);
7182 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
7183 next - ddi_get_lbolt());
7184 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
7185 next = ddi_get_lbolt() + hz;
7188 * Quick check for L2ARC devices.
7190 mutex_enter(&l2arc_dev_mtx);
7191 if (l2arc_ndev == 0) {
7192 mutex_exit(&l2arc_dev_mtx);
7195 mutex_exit(&l2arc_dev_mtx);
7196 begin = ddi_get_lbolt();
7199 * This selects the next l2arc device to write to, and in
7200 * doing so the next spa to feed from: dev->l2ad_spa. This
7201 * will return NULL if there are now no l2arc devices or if
7202 * they are all faulted.
7204 * If a device is returned, its spa's config lock is also
7205 * held to prevent device removal. l2arc_dev_get_next()
7206 * will grab and release l2arc_dev_mtx.
7208 if ((dev = l2arc_dev_get_next()) == NULL)
7211 spa = dev->l2ad_spa;
7212 ASSERT3P(spa, !=, NULL);
7215 * If the pool is read-only then force the feed thread to
7216 * sleep a little longer.
7218 if (!spa_writeable(spa)) {
7219 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
7220 spa_config_exit(spa, SCL_L2ARC, dev);
7225 * Avoid contributing to memory pressure.
7227 if (arc_reclaim_needed()) {
7228 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
7229 spa_config_exit(spa, SCL_L2ARC, dev);
7233 ARCSTAT_BUMP(arcstat_l2_feeds);
7235 size = l2arc_write_size();
7238 * Evict L2ARC buffers that will be overwritten.
7240 l2arc_evict(dev, size, B_FALSE);
7243 * Write ARC buffers.
7245 wrote = l2arc_write_buffers(spa, dev, size);
7248 * Calculate interval between writes.
7250 next = l2arc_write_interval(begin, size, wrote);
7251 spa_config_exit(spa, SCL_L2ARC, dev);
7254 l2arc_thread_exit = 0;
7255 cv_broadcast(&l2arc_feed_thr_cv);
7256 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
7261 l2arc_vdev_present(vdev_t *vd)
7265 mutex_enter(&l2arc_dev_mtx);
7266 for (dev = list_head(l2arc_dev_list); dev != NULL;
7267 dev = list_next(l2arc_dev_list, dev)) {
7268 if (dev->l2ad_vdev == vd)
7271 mutex_exit(&l2arc_dev_mtx);
7273 return (dev != NULL);
7277 * Add a vdev for use by the L2ARC. By this point the spa has already
7278 * validated the vdev and opened it.
7281 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
7283 l2arc_dev_t *adddev;
7285 ASSERT(!l2arc_vdev_present(vd));
7287 vdev_ashift_optimize(vd);
7290 * Create a new l2arc device entry.
7292 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
7293 adddev->l2ad_spa = spa;
7294 adddev->l2ad_vdev = vd;
7295 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
7296 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
7297 adddev->l2ad_hand = adddev->l2ad_start;
7298 adddev->l2ad_first = B_TRUE;
7299 adddev->l2ad_writing = B_FALSE;
7301 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
7303 * This is a list of all ARC buffers that are still valid on the
7306 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
7307 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
7309 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
7310 refcount_create(&adddev->l2ad_alloc);
7313 * Add device to global list
7315 mutex_enter(&l2arc_dev_mtx);
7316 list_insert_head(l2arc_dev_list, adddev);
7317 atomic_inc_64(&l2arc_ndev);
7318 mutex_exit(&l2arc_dev_mtx);
7322 * Remove a vdev from the L2ARC.
7325 l2arc_remove_vdev(vdev_t *vd)
7327 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
7330 * Find the device by vdev
7332 mutex_enter(&l2arc_dev_mtx);
7333 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
7334 nextdev = list_next(l2arc_dev_list, dev);
7335 if (vd == dev->l2ad_vdev) {
7340 ASSERT3P(remdev, !=, NULL);
7343 * Remove device from global list
7345 list_remove(l2arc_dev_list, remdev);
7346 l2arc_dev_last = NULL; /* may have been invalidated */
7347 atomic_dec_64(&l2arc_ndev);
7348 mutex_exit(&l2arc_dev_mtx);
7351 * Clear all buflists and ARC references. L2ARC device flush.
7353 l2arc_evict(remdev, 0, B_TRUE);
7354 list_destroy(&remdev->l2ad_buflist);
7355 mutex_destroy(&remdev->l2ad_mtx);
7356 refcount_destroy(&remdev->l2ad_alloc);
7357 kmem_free(remdev, sizeof (l2arc_dev_t));
7363 l2arc_thread_exit = 0;
7365 l2arc_writes_sent = 0;
7366 l2arc_writes_done = 0;
7368 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
7369 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
7370 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
7371 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
7373 l2arc_dev_list = &L2ARC_dev_list;
7374 l2arc_free_on_write = &L2ARC_free_on_write;
7375 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
7376 offsetof(l2arc_dev_t, l2ad_node));
7377 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
7378 offsetof(l2arc_data_free_t, l2df_list_node));
7385 * This is called from dmu_fini(), which is called from spa_fini();
7386 * Because of this, we can assume that all l2arc devices have
7387 * already been removed when the pools themselves were removed.
7390 l2arc_do_free_on_write();
7392 mutex_destroy(&l2arc_feed_thr_lock);
7393 cv_destroy(&l2arc_feed_thr_cv);
7394 mutex_destroy(&l2arc_dev_mtx);
7395 mutex_destroy(&l2arc_free_on_write_mtx);
7397 list_destroy(l2arc_dev_list);
7398 list_destroy(l2arc_free_on_write);
7404 if (!(spa_mode_global & FWRITE))
7407 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
7408 TS_RUN, minclsyspri);
7414 if (!(spa_mode_global & FWRITE))
7417 mutex_enter(&l2arc_feed_thr_lock);
7418 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
7419 l2arc_thread_exit = 1;
7420 while (l2arc_thread_exit != 0)
7421 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
7422 mutex_exit(&l2arc_feed_thr_lock);