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>
261 #include <sys/callb.h>
262 #include <sys/kstat.h>
263 #include <sys/trim_map.h>
264 #include <zfs_fletcher.h>
267 #include <machine/vmparam.h>
271 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
272 boolean_t arc_watch = B_FALSE;
277 static kmutex_t arc_reclaim_lock;
278 static kcondvar_t arc_reclaim_thread_cv;
279 static boolean_t arc_reclaim_thread_exit;
280 static kcondvar_t arc_reclaim_waiters_cv;
282 uint_t arc_reduce_dnlc_percent = 3;
285 * The number of headers to evict in arc_evict_state_impl() before
286 * dropping the sublist lock and evicting from another sublist. A lower
287 * value means we're more likely to evict the "correct" header (i.e. the
288 * oldest header in the arc state), but comes with higher overhead
289 * (i.e. more invocations of arc_evict_state_impl()).
291 int zfs_arc_evict_batch_limit = 10;
294 * The number of sublists used for each of the arc state lists. If this
295 * is not set to a suitable value by the user, it will be configured to
296 * the number of CPUs on the system in arc_init().
298 int zfs_arc_num_sublists_per_state = 0;
300 /* number of seconds before growing cache again */
301 static int arc_grow_retry = 60;
303 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */
304 int zfs_arc_overflow_shift = 8;
306 /* shift of arc_c for calculating both min and max arc_p */
307 static int arc_p_min_shift = 4;
309 /* log2(fraction of arc to reclaim) */
310 static int arc_shrink_shift = 7;
313 * log2(fraction of ARC which must be free to allow growing).
314 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
315 * when reading a new block into the ARC, we will evict an equal-sized block
318 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
319 * we will still not allow it to grow.
321 int arc_no_grow_shift = 5;
325 * minimum lifespan of a prefetch block in clock ticks
326 * (initialized in arc_init())
328 static int arc_min_prefetch_lifespan;
331 * If this percent of memory is free, don't throttle.
333 int arc_lotsfree_percent = 10;
336 extern boolean_t zfs_prefetch_disable;
339 * The arc has filled available memory and has now warmed up.
341 static boolean_t arc_warm;
344 * These tunables are for performance analysis.
346 uint64_t zfs_arc_max;
347 uint64_t zfs_arc_min;
348 uint64_t zfs_arc_meta_limit = 0;
349 uint64_t zfs_arc_meta_min = 0;
350 int zfs_arc_grow_retry = 0;
351 int zfs_arc_shrink_shift = 0;
352 int zfs_arc_p_min_shift = 0;
353 uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
354 u_int zfs_arc_free_target = 0;
356 /* Absolute min for arc min / max is 16MB. */
357 static uint64_t arc_abs_min = 16 << 20;
359 boolean_t zfs_compressed_arc_enabled = B_TRUE;
361 static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
362 static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
363 static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS);
364 static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS);
366 #if defined(__FreeBSD__) && defined(_KERNEL)
368 arc_free_target_init(void *unused __unused)
371 zfs_arc_free_target = vm_pageout_wakeup_thresh;
373 SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
374 arc_free_target_init, NULL);
376 TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max);
377 TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min);
378 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
379 TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
380 TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize);
381 TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
382 SYSCTL_DECL(_vfs_zfs);
383 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN,
384 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size");
385 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN,
386 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size");
387 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
388 &zfs_arc_average_blocksize, 0,
389 "ARC average blocksize");
390 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
391 &arc_shrink_shift, 0,
392 "log2(fraction of arc to reclaim)");
393 SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN,
394 &zfs_compressed_arc_enabled, 0, "Enable compressed ARC");
397 * We don't have a tunable for arc_free_target due to the dependency on
398 * pagedaemon initialisation.
400 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
401 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
402 sysctl_vfs_zfs_arc_free_target, "IU",
403 "Desired number of free pages below which ARC triggers reclaim");
406 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
411 val = zfs_arc_free_target;
412 err = sysctl_handle_int(oidp, &val, 0, req);
413 if (err != 0 || req->newptr == NULL)
418 if (val > cnt.v_page_count)
421 zfs_arc_free_target = val;
427 * Must be declared here, before the definition of corresponding kstat
428 * macro which uses the same names will confuse the compiler.
430 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
431 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
432 sysctl_vfs_zfs_arc_meta_limit, "QU",
433 "ARC metadata limit");
437 * Note that buffers can be in one of 6 states:
438 * ARC_anon - anonymous (discussed below)
439 * ARC_mru - recently used, currently cached
440 * ARC_mru_ghost - recentely used, no longer in cache
441 * ARC_mfu - frequently used, currently cached
442 * ARC_mfu_ghost - frequently used, no longer in cache
443 * ARC_l2c_only - exists in L2ARC but not other states
444 * When there are no active references to the buffer, they are
445 * are linked onto a list in one of these arc states. These are
446 * the only buffers that can be evicted or deleted. Within each
447 * state there are multiple lists, one for meta-data and one for
448 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
449 * etc.) is tracked separately so that it can be managed more
450 * explicitly: favored over data, limited explicitly.
452 * Anonymous buffers are buffers that are not associated with
453 * a DVA. These are buffers that hold dirty block copies
454 * before they are written to stable storage. By definition,
455 * they are "ref'd" and are considered part of arc_mru
456 * that cannot be freed. Generally, they will aquire a DVA
457 * as they are written and migrate onto the arc_mru list.
459 * The ARC_l2c_only state is for buffers that are in the second
460 * level ARC but no longer in any of the ARC_m* lists. The second
461 * level ARC itself may also contain buffers that are in any of
462 * the ARC_m* states - meaning that a buffer can exist in two
463 * places. The reason for the ARC_l2c_only state is to keep the
464 * buffer header in the hash table, so that reads that hit the
465 * second level ARC benefit from these fast lookups.
468 typedef struct arc_state {
470 * list of evictable buffers
472 multilist_t arcs_list[ARC_BUFC_NUMTYPES];
474 * total amount of evictable data in this state
476 refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
478 * total amount of data in this state; this includes: evictable,
479 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
481 refcount_t arcs_size;
485 static arc_state_t ARC_anon;
486 static arc_state_t ARC_mru;
487 static arc_state_t ARC_mru_ghost;
488 static arc_state_t ARC_mfu;
489 static arc_state_t ARC_mfu_ghost;
490 static arc_state_t ARC_l2c_only;
492 typedef struct arc_stats {
493 kstat_named_t arcstat_hits;
494 kstat_named_t arcstat_misses;
495 kstat_named_t arcstat_demand_data_hits;
496 kstat_named_t arcstat_demand_data_misses;
497 kstat_named_t arcstat_demand_metadata_hits;
498 kstat_named_t arcstat_demand_metadata_misses;
499 kstat_named_t arcstat_prefetch_data_hits;
500 kstat_named_t arcstat_prefetch_data_misses;
501 kstat_named_t arcstat_prefetch_metadata_hits;
502 kstat_named_t arcstat_prefetch_metadata_misses;
503 kstat_named_t arcstat_mru_hits;
504 kstat_named_t arcstat_mru_ghost_hits;
505 kstat_named_t arcstat_mfu_hits;
506 kstat_named_t arcstat_mfu_ghost_hits;
507 kstat_named_t arcstat_allocated;
508 kstat_named_t arcstat_deleted;
510 * Number of buffers that could not be evicted because the hash lock
511 * was held by another thread. The lock may not necessarily be held
512 * by something using the same buffer, since hash locks are shared
513 * by multiple buffers.
515 kstat_named_t arcstat_mutex_miss;
517 * Number of buffers skipped because they have I/O in progress, are
518 * indrect prefetch buffers that have not lived long enough, or are
519 * not from the spa we're trying to evict from.
521 kstat_named_t arcstat_evict_skip;
523 * Number of times arc_evict_state() was unable to evict enough
524 * buffers to reach it's target amount.
526 kstat_named_t arcstat_evict_not_enough;
527 kstat_named_t arcstat_evict_l2_cached;
528 kstat_named_t arcstat_evict_l2_eligible;
529 kstat_named_t arcstat_evict_l2_ineligible;
530 kstat_named_t arcstat_evict_l2_skip;
531 kstat_named_t arcstat_hash_elements;
532 kstat_named_t arcstat_hash_elements_max;
533 kstat_named_t arcstat_hash_collisions;
534 kstat_named_t arcstat_hash_chains;
535 kstat_named_t arcstat_hash_chain_max;
536 kstat_named_t arcstat_p;
537 kstat_named_t arcstat_c;
538 kstat_named_t arcstat_c_min;
539 kstat_named_t arcstat_c_max;
540 kstat_named_t arcstat_size;
542 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pdata.
543 * Note that the compressed bytes may match the uncompressed bytes
544 * if the block is either not compressed or compressed arc is disabled.
546 kstat_named_t arcstat_compressed_size;
548 * Uncompressed size of the data stored in b_pdata. If compressed
549 * arc is disabled then this value will be identical to the stat
552 kstat_named_t arcstat_uncompressed_size;
554 * Number of bytes stored in all the arc_buf_t's. This is classified
555 * as "overhead" since this data is typically short-lived and will
556 * be evicted from the arc when it becomes unreferenced unless the
557 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
558 * values have been set (see comment in dbuf.c for more information).
560 kstat_named_t arcstat_overhead_size;
562 * Number of bytes consumed by internal ARC structures necessary
563 * for tracking purposes; these structures are not actually
564 * backed by ARC buffers. This includes arc_buf_hdr_t structures
565 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
566 * caches), and arc_buf_t structures (allocated via arc_buf_t
569 kstat_named_t arcstat_hdr_size;
571 * Number of bytes consumed by ARC buffers of type equal to
572 * ARC_BUFC_DATA. This is generally consumed by buffers backing
573 * on disk user data (e.g. plain file contents).
575 kstat_named_t arcstat_data_size;
577 * Number of bytes consumed by ARC buffers of type equal to
578 * ARC_BUFC_METADATA. This is generally consumed by buffers
579 * backing on disk data that is used for internal ZFS
580 * structures (e.g. ZAP, dnode, indirect blocks, etc).
582 kstat_named_t arcstat_metadata_size;
584 * Number of bytes consumed by various buffers and structures
585 * not actually backed with ARC buffers. This includes bonus
586 * buffers (allocated directly via zio_buf_* functions),
587 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
588 * cache), and dnode_t structures (allocated via dnode_t cache).
590 kstat_named_t arcstat_other_size;
592 * Total number of bytes consumed by ARC buffers residing in the
593 * arc_anon state. This includes *all* buffers in the arc_anon
594 * state; e.g. data, metadata, evictable, and unevictable buffers
595 * are all included in this value.
597 kstat_named_t arcstat_anon_size;
599 * Number of bytes consumed by ARC buffers that meet the
600 * following criteria: backing buffers of type ARC_BUFC_DATA,
601 * residing in the arc_anon state, and are eligible for eviction
602 * (e.g. have no outstanding holds on the buffer).
604 kstat_named_t arcstat_anon_evictable_data;
606 * Number of bytes consumed by ARC buffers that meet the
607 * following criteria: backing buffers of type ARC_BUFC_METADATA,
608 * residing in the arc_anon state, and are eligible for eviction
609 * (e.g. have no outstanding holds on the buffer).
611 kstat_named_t arcstat_anon_evictable_metadata;
613 * Total number of bytes consumed by ARC buffers residing in the
614 * arc_mru state. This includes *all* buffers in the arc_mru
615 * state; e.g. data, metadata, evictable, and unevictable buffers
616 * are all included in this value.
618 kstat_named_t arcstat_mru_size;
620 * Number of bytes consumed by ARC buffers that meet the
621 * following criteria: backing buffers of type ARC_BUFC_DATA,
622 * residing in the arc_mru state, and are eligible for eviction
623 * (e.g. have no outstanding holds on the buffer).
625 kstat_named_t arcstat_mru_evictable_data;
627 * Number of bytes consumed by ARC buffers that meet the
628 * following criteria: backing buffers of type ARC_BUFC_METADATA,
629 * residing in the arc_mru state, and are eligible for eviction
630 * (e.g. have no outstanding holds on the buffer).
632 kstat_named_t arcstat_mru_evictable_metadata;
634 * Total number of bytes that *would have been* consumed by ARC
635 * buffers in the arc_mru_ghost state. The key thing to note
636 * here, is the fact that this size doesn't actually indicate
637 * RAM consumption. The ghost lists only consist of headers and
638 * don't actually have ARC buffers linked off of these headers.
639 * Thus, *if* the headers had associated ARC buffers, these
640 * buffers *would have* consumed this number of bytes.
642 kstat_named_t arcstat_mru_ghost_size;
644 * Number of bytes that *would have been* consumed by ARC
645 * buffers that are eligible for eviction, of type
646 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
648 kstat_named_t arcstat_mru_ghost_evictable_data;
650 * Number of bytes that *would have been* consumed by ARC
651 * buffers that are eligible for eviction, of type
652 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
654 kstat_named_t arcstat_mru_ghost_evictable_metadata;
656 * Total number of bytes consumed by ARC buffers residing in the
657 * arc_mfu state. This includes *all* buffers in the arc_mfu
658 * state; e.g. data, metadata, evictable, and unevictable buffers
659 * are all included in this value.
661 kstat_named_t arcstat_mfu_size;
663 * Number of bytes consumed by ARC buffers that are eligible for
664 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
667 kstat_named_t arcstat_mfu_evictable_data;
669 * Number of bytes consumed by ARC buffers that are eligible for
670 * eviction, of type ARC_BUFC_METADATA, and reside in the
673 kstat_named_t arcstat_mfu_evictable_metadata;
675 * Total number of bytes that *would have been* consumed by ARC
676 * buffers in the arc_mfu_ghost state. See the comment above
677 * arcstat_mru_ghost_size for more details.
679 kstat_named_t arcstat_mfu_ghost_size;
681 * Number of bytes that *would have been* consumed by ARC
682 * buffers that are eligible for eviction, of type
683 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
685 kstat_named_t arcstat_mfu_ghost_evictable_data;
687 * Number of bytes that *would have been* consumed by ARC
688 * buffers that are eligible for eviction, of type
689 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
691 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
692 kstat_named_t arcstat_l2_hits;
693 kstat_named_t arcstat_l2_misses;
694 kstat_named_t arcstat_l2_feeds;
695 kstat_named_t arcstat_l2_rw_clash;
696 kstat_named_t arcstat_l2_read_bytes;
697 kstat_named_t arcstat_l2_write_bytes;
698 kstat_named_t arcstat_l2_writes_sent;
699 kstat_named_t arcstat_l2_writes_done;
700 kstat_named_t arcstat_l2_writes_error;
701 kstat_named_t arcstat_l2_writes_lock_retry;
702 kstat_named_t arcstat_l2_evict_lock_retry;
703 kstat_named_t arcstat_l2_evict_reading;
704 kstat_named_t arcstat_l2_evict_l1cached;
705 kstat_named_t arcstat_l2_free_on_write;
706 kstat_named_t arcstat_l2_abort_lowmem;
707 kstat_named_t arcstat_l2_cksum_bad;
708 kstat_named_t arcstat_l2_io_error;
709 kstat_named_t arcstat_l2_size;
710 kstat_named_t arcstat_l2_asize;
711 kstat_named_t arcstat_l2_hdr_size;
712 kstat_named_t arcstat_l2_write_trylock_fail;
713 kstat_named_t arcstat_l2_write_passed_headroom;
714 kstat_named_t arcstat_l2_write_spa_mismatch;
715 kstat_named_t arcstat_l2_write_in_l2;
716 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
717 kstat_named_t arcstat_l2_write_not_cacheable;
718 kstat_named_t arcstat_l2_write_full;
719 kstat_named_t arcstat_l2_write_buffer_iter;
720 kstat_named_t arcstat_l2_write_pios;
721 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
722 kstat_named_t arcstat_l2_write_buffer_list_iter;
723 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
724 kstat_named_t arcstat_memory_throttle_count;
725 kstat_named_t arcstat_meta_used;
726 kstat_named_t arcstat_meta_limit;
727 kstat_named_t arcstat_meta_max;
728 kstat_named_t arcstat_meta_min;
729 kstat_named_t arcstat_sync_wait_for_async;
730 kstat_named_t arcstat_demand_hit_predictive_prefetch;
733 static arc_stats_t arc_stats = {
734 { "hits", KSTAT_DATA_UINT64 },
735 { "misses", KSTAT_DATA_UINT64 },
736 { "demand_data_hits", KSTAT_DATA_UINT64 },
737 { "demand_data_misses", KSTAT_DATA_UINT64 },
738 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
739 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
740 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
741 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
742 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
743 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
744 { "mru_hits", KSTAT_DATA_UINT64 },
745 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
746 { "mfu_hits", KSTAT_DATA_UINT64 },
747 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
748 { "allocated", KSTAT_DATA_UINT64 },
749 { "deleted", KSTAT_DATA_UINT64 },
750 { "mutex_miss", KSTAT_DATA_UINT64 },
751 { "evict_skip", KSTAT_DATA_UINT64 },
752 { "evict_not_enough", KSTAT_DATA_UINT64 },
753 { "evict_l2_cached", KSTAT_DATA_UINT64 },
754 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
755 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
756 { "evict_l2_skip", KSTAT_DATA_UINT64 },
757 { "hash_elements", KSTAT_DATA_UINT64 },
758 { "hash_elements_max", KSTAT_DATA_UINT64 },
759 { "hash_collisions", KSTAT_DATA_UINT64 },
760 { "hash_chains", KSTAT_DATA_UINT64 },
761 { "hash_chain_max", KSTAT_DATA_UINT64 },
762 { "p", KSTAT_DATA_UINT64 },
763 { "c", KSTAT_DATA_UINT64 },
764 { "c_min", KSTAT_DATA_UINT64 },
765 { "c_max", KSTAT_DATA_UINT64 },
766 { "size", KSTAT_DATA_UINT64 },
767 { "compressed_size", KSTAT_DATA_UINT64 },
768 { "uncompressed_size", KSTAT_DATA_UINT64 },
769 { "overhead_size", KSTAT_DATA_UINT64 },
770 { "hdr_size", KSTAT_DATA_UINT64 },
771 { "data_size", KSTAT_DATA_UINT64 },
772 { "metadata_size", KSTAT_DATA_UINT64 },
773 { "other_size", KSTAT_DATA_UINT64 },
774 { "anon_size", KSTAT_DATA_UINT64 },
775 { "anon_evictable_data", KSTAT_DATA_UINT64 },
776 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
777 { "mru_size", KSTAT_DATA_UINT64 },
778 { "mru_evictable_data", KSTAT_DATA_UINT64 },
779 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
780 { "mru_ghost_size", KSTAT_DATA_UINT64 },
781 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
782 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
783 { "mfu_size", KSTAT_DATA_UINT64 },
784 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
785 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
786 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
787 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
788 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
789 { "l2_hits", KSTAT_DATA_UINT64 },
790 { "l2_misses", KSTAT_DATA_UINT64 },
791 { "l2_feeds", KSTAT_DATA_UINT64 },
792 { "l2_rw_clash", KSTAT_DATA_UINT64 },
793 { "l2_read_bytes", KSTAT_DATA_UINT64 },
794 { "l2_write_bytes", KSTAT_DATA_UINT64 },
795 { "l2_writes_sent", KSTAT_DATA_UINT64 },
796 { "l2_writes_done", KSTAT_DATA_UINT64 },
797 { "l2_writes_error", KSTAT_DATA_UINT64 },
798 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
799 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
800 { "l2_evict_reading", KSTAT_DATA_UINT64 },
801 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
802 { "l2_free_on_write", KSTAT_DATA_UINT64 },
803 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
804 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
805 { "l2_io_error", KSTAT_DATA_UINT64 },
806 { "l2_size", KSTAT_DATA_UINT64 },
807 { "l2_asize", KSTAT_DATA_UINT64 },
808 { "l2_hdr_size", KSTAT_DATA_UINT64 },
809 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
810 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
811 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
812 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
813 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
814 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
815 { "l2_write_full", KSTAT_DATA_UINT64 },
816 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
817 { "l2_write_pios", KSTAT_DATA_UINT64 },
818 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
819 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
820 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
821 { "memory_throttle_count", KSTAT_DATA_UINT64 },
822 { "arc_meta_used", KSTAT_DATA_UINT64 },
823 { "arc_meta_limit", KSTAT_DATA_UINT64 },
824 { "arc_meta_max", KSTAT_DATA_UINT64 },
825 { "arc_meta_min", KSTAT_DATA_UINT64 },
826 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
827 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
830 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
832 #define ARCSTAT_INCR(stat, val) \
833 atomic_add_64(&arc_stats.stat.value.ui64, (val))
835 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
836 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
838 #define ARCSTAT_MAX(stat, val) { \
840 while ((val) > (m = arc_stats.stat.value.ui64) && \
841 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
845 #define ARCSTAT_MAXSTAT(stat) \
846 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
849 * We define a macro to allow ARC hits/misses to be easily broken down by
850 * two separate conditions, giving a total of four different subtypes for
851 * each of hits and misses (so eight statistics total).
853 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
856 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
858 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
862 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
864 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
869 static arc_state_t *arc_anon;
870 static arc_state_t *arc_mru;
871 static arc_state_t *arc_mru_ghost;
872 static arc_state_t *arc_mfu;
873 static arc_state_t *arc_mfu_ghost;
874 static arc_state_t *arc_l2c_only;
877 * There are several ARC variables that are critical to export as kstats --
878 * but we don't want to have to grovel around in the kstat whenever we wish to
879 * manipulate them. For these variables, we therefore define them to be in
880 * terms of the statistic variable. This assures that we are not introducing
881 * the possibility of inconsistency by having shadow copies of the variables,
882 * while still allowing the code to be readable.
884 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
885 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
886 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
887 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
888 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
889 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
890 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
891 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
892 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
894 /* compressed size of entire arc */
895 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
896 /* uncompressed size of entire arc */
897 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
898 /* number of bytes in the arc from arc_buf_t's */
899 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
901 static int arc_no_grow; /* Don't try to grow cache size */
902 static uint64_t arc_tempreserve;
903 static uint64_t arc_loaned_bytes;
905 typedef struct arc_callback arc_callback_t;
907 struct arc_callback {
909 arc_done_func_t *acb_done;
911 zio_t *acb_zio_dummy;
912 arc_callback_t *acb_next;
915 typedef struct arc_write_callback arc_write_callback_t;
917 struct arc_write_callback {
919 arc_done_func_t *awcb_ready;
920 arc_done_func_t *awcb_children_ready;
921 arc_done_func_t *awcb_physdone;
922 arc_done_func_t *awcb_done;
927 * ARC buffers are separated into multiple structs as a memory saving measure:
928 * - Common fields struct, always defined, and embedded within it:
929 * - L2-only fields, always allocated but undefined when not in L2ARC
930 * - L1-only fields, only allocated when in L1ARC
932 * Buffer in L1 Buffer only in L2
933 * +------------------------+ +------------------------+
934 * | arc_buf_hdr_t | | arc_buf_hdr_t |
938 * +------------------------+ +------------------------+
939 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
940 * | (undefined if L1-only) | | |
941 * +------------------------+ +------------------------+
942 * | l1arc_buf_hdr_t |
947 * +------------------------+
949 * Because it's possible for the L2ARC to become extremely large, we can wind
950 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
951 * is minimized by only allocating the fields necessary for an L1-cached buffer
952 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
953 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
954 * words in pointers. arc_hdr_realloc() is used to switch a header between
955 * these two allocation states.
957 typedef struct l1arc_buf_hdr {
958 kmutex_t b_freeze_lock;
959 zio_cksum_t *b_freeze_cksum;
962 * used for debugging wtih kmem_flags - by allocating and freeing
963 * b_thawed when the buffer is thawed, we get a record of the stack
964 * trace that thawed it.
971 /* for waiting on writes to complete */
975 /* protected by arc state mutex */
976 arc_state_t *b_state;
977 multilist_node_t b_arc_node;
979 /* updated atomically */
980 clock_t b_arc_access;
982 /* self protecting */
985 arc_callback_t *b_acb;
989 typedef struct l2arc_dev l2arc_dev_t;
991 typedef struct l2arc_buf_hdr {
992 /* protected by arc_buf_hdr mutex */
993 l2arc_dev_t *b_dev; /* L2ARC device */
994 uint64_t b_daddr; /* disk address, offset byte */
996 list_node_t b_l2node;
1000 /* protected by hash lock */
1004 arc_buf_contents_t b_type;
1005 arc_buf_hdr_t *b_hash_next;
1006 arc_flags_t b_flags;
1009 * This field stores the size of the data buffer after
1010 * compression, and is set in the arc's zio completion handlers.
1011 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
1013 * While the block pointers can store up to 32MB in their psize
1014 * field, we can only store up to 32MB minus 512B. This is due
1015 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
1016 * a field of zeros represents 512B in the bp). We can't use a
1017 * bias of 1 since we need to reserve a psize of zero, here, to
1018 * represent holes and embedded blocks.
1020 * This isn't a problem in practice, since the maximum size of a
1021 * buffer is limited to 16MB, so we never need to store 32MB in
1022 * this field. Even in the upstream illumos code base, the
1023 * maximum size of a buffer is limited to 16MB.
1028 * This field stores the size of the data buffer before
1029 * compression, and cannot change once set. It is in units
1030 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
1032 uint16_t b_lsize; /* immutable */
1033 uint64_t b_spa; /* immutable */
1035 /* L2ARC fields. Undefined when not in L2ARC. */
1036 l2arc_buf_hdr_t b_l2hdr;
1037 /* L1ARC fields. Undefined when in l2arc_only state */
1038 l1arc_buf_hdr_t b_l1hdr;
1041 #if defined(__FreeBSD__) && defined(_KERNEL)
1043 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
1048 val = arc_meta_limit;
1049 err = sysctl_handle_64(oidp, &val, 0, req);
1050 if (err != 0 || req->newptr == NULL)
1053 if (val <= 0 || val > arc_c_max)
1056 arc_meta_limit = val;
1061 sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
1067 err = sysctl_handle_64(oidp, &val, 0, req);
1068 if (err != 0 || req->newptr == NULL)
1071 if (zfs_arc_max == 0) {
1072 /* Loader tunable so blindly set */
1077 if (val < arc_abs_min || val > kmem_size())
1079 if (val < arc_c_min)
1081 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
1087 arc_p = (arc_c >> 1);
1089 if (zfs_arc_meta_limit == 0) {
1090 /* limit meta-data to 1/4 of the arc capacity */
1091 arc_meta_limit = arc_c_max / 4;
1094 /* if kmem_flags are set, lets try to use less memory */
1095 if (kmem_debugging())
1098 zfs_arc_max = arc_c;
1104 sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
1110 err = sysctl_handle_64(oidp, &val, 0, req);
1111 if (err != 0 || req->newptr == NULL)
1114 if (zfs_arc_min == 0) {
1115 /* Loader tunable so blindly set */
1120 if (val < arc_abs_min || val > arc_c_max)
1125 if (zfs_arc_meta_min == 0)
1126 arc_meta_min = arc_c_min / 2;
1128 if (arc_c < arc_c_min)
1131 zfs_arc_min = arc_c_min;
1137 #define GHOST_STATE(state) \
1138 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
1139 (state) == arc_l2c_only)
1141 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1142 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1143 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1144 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1145 #define HDR_COMPRESSION_ENABLED(hdr) \
1146 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1148 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1149 #define HDR_L2_READING(hdr) \
1150 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1151 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1152 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1153 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1154 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1155 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1157 #define HDR_ISTYPE_METADATA(hdr) \
1158 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1159 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1161 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1162 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1164 /* For storing compression mode in b_flags */
1165 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1167 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1168 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1169 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1170 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1172 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1178 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1179 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1182 * Hash table routines
1185 #define HT_LOCK_PAD CACHE_LINE_SIZE
1190 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1194 #define BUF_LOCKS 256
1195 typedef struct buf_hash_table {
1197 arc_buf_hdr_t **ht_table;
1198 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1201 static buf_hash_table_t buf_hash_table;
1203 #define BUF_HASH_INDEX(spa, dva, birth) \
1204 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1205 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1206 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1207 #define HDR_LOCK(hdr) \
1208 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1210 uint64_t zfs_crc64_table[256];
1216 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1217 #define L2ARC_HEADROOM 2 /* num of writes */
1219 * If we discover during ARC scan any buffers to be compressed, we boost
1220 * our headroom for the next scanning cycle by this percentage multiple.
1222 #define L2ARC_HEADROOM_BOOST 200
1223 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1224 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1226 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1227 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1229 /* L2ARC Performance Tunables */
1230 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1231 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1232 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1233 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1234 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1235 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1236 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1237 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1238 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1240 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1241 &l2arc_write_max, 0, "max write size");
1242 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1243 &l2arc_write_boost, 0, "extra write during warmup");
1244 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1245 &l2arc_headroom, 0, "number of dev writes");
1246 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1247 &l2arc_feed_secs, 0, "interval seconds");
1248 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1249 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1251 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1252 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1253 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1254 &l2arc_feed_again, 0, "turbo warmup");
1255 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1256 &l2arc_norw, 0, "no reads during writes");
1258 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1259 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1260 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD,
1261 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1262 "size of anonymous state");
1263 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD,
1264 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1265 "size of anonymous state");
1267 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1268 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1269 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD,
1270 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1271 "size of metadata in mru state");
1272 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD,
1273 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1274 "size of data in mru state");
1276 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1277 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1278 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD,
1279 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1280 "size of metadata in mru ghost state");
1281 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD,
1282 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1283 "size of data in mru ghost state");
1285 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1286 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1287 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD,
1288 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1289 "size of metadata in mfu state");
1290 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD,
1291 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1292 "size of data in mfu state");
1294 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1295 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1296 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD,
1297 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1298 "size of metadata in mfu ghost state");
1299 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD,
1300 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1301 "size of data in mfu ghost state");
1303 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1304 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1310 vdev_t *l2ad_vdev; /* vdev */
1311 spa_t *l2ad_spa; /* spa */
1312 uint64_t l2ad_hand; /* next write location */
1313 uint64_t l2ad_start; /* first addr on device */
1314 uint64_t l2ad_end; /* last addr on device */
1315 boolean_t l2ad_first; /* first sweep through */
1316 boolean_t l2ad_writing; /* currently writing */
1317 kmutex_t l2ad_mtx; /* lock for buffer list */
1318 list_t l2ad_buflist; /* buffer list */
1319 list_node_t l2ad_node; /* device list node */
1320 refcount_t l2ad_alloc; /* allocated bytes */
1323 static list_t L2ARC_dev_list; /* device list */
1324 static list_t *l2arc_dev_list; /* device list pointer */
1325 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1326 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1327 static list_t L2ARC_free_on_write; /* free after write buf list */
1328 static list_t *l2arc_free_on_write; /* free after write list ptr */
1329 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1330 static uint64_t l2arc_ndev; /* number of devices */
1332 typedef struct l2arc_read_callback {
1333 arc_buf_hdr_t *l2rcb_hdr; /* read buffer */
1334 blkptr_t l2rcb_bp; /* original blkptr */
1335 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1336 int l2rcb_flags; /* original flags */
1337 void *l2rcb_data; /* temporary buffer */
1338 } l2arc_read_callback_t;
1340 typedef struct l2arc_write_callback {
1341 l2arc_dev_t *l2wcb_dev; /* device info */
1342 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1343 } l2arc_write_callback_t;
1345 typedef struct l2arc_data_free {
1346 /* protected by l2arc_free_on_write_mtx */
1349 arc_buf_contents_t l2df_type;
1350 list_node_t l2df_list_node;
1351 } l2arc_data_free_t;
1353 static kmutex_t l2arc_feed_thr_lock;
1354 static kcondvar_t l2arc_feed_thr_cv;
1355 static uint8_t l2arc_thread_exit;
1357 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1358 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1359 static void arc_hdr_free_pdata(arc_buf_hdr_t *hdr);
1360 static void arc_hdr_alloc_pdata(arc_buf_hdr_t *);
1361 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1362 static boolean_t arc_is_overflowing();
1363 static void arc_buf_watch(arc_buf_t *);
1365 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1366 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1367 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1368 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1370 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1371 static void l2arc_read_done(zio_t *);
1374 l2arc_trim(const arc_buf_hdr_t *hdr)
1376 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1378 ASSERT(HDR_HAS_L2HDR(hdr));
1379 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1381 if (HDR_GET_PSIZE(hdr) != 0) {
1382 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1383 HDR_GET_PSIZE(hdr), 0);
1388 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1390 uint8_t *vdva = (uint8_t *)dva;
1391 uint64_t crc = -1ULL;
1394 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
1396 for (i = 0; i < sizeof (dva_t); i++)
1397 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
1399 crc ^= (spa>>8) ^ birth;
1404 #define HDR_EMPTY(hdr) \
1405 ((hdr)->b_dva.dva_word[0] == 0 && \
1406 (hdr)->b_dva.dva_word[1] == 0)
1408 #define HDR_EQUAL(spa, dva, birth, hdr) \
1409 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1410 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1411 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1414 buf_discard_identity(arc_buf_hdr_t *hdr)
1416 hdr->b_dva.dva_word[0] = 0;
1417 hdr->b_dva.dva_word[1] = 0;
1421 static arc_buf_hdr_t *
1422 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1424 const dva_t *dva = BP_IDENTITY(bp);
1425 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1426 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1427 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1430 mutex_enter(hash_lock);
1431 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1432 hdr = hdr->b_hash_next) {
1433 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1438 mutex_exit(hash_lock);
1444 * Insert an entry into the hash table. If there is already an element
1445 * equal to elem in the hash table, then the already existing element
1446 * will be returned and the new element will not be inserted.
1447 * Otherwise returns NULL.
1448 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1450 static arc_buf_hdr_t *
1451 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1453 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1454 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1455 arc_buf_hdr_t *fhdr;
1458 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1459 ASSERT(hdr->b_birth != 0);
1460 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1462 if (lockp != NULL) {
1464 mutex_enter(hash_lock);
1466 ASSERT(MUTEX_HELD(hash_lock));
1469 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1470 fhdr = fhdr->b_hash_next, i++) {
1471 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1475 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1476 buf_hash_table.ht_table[idx] = hdr;
1477 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1479 /* collect some hash table performance data */
1481 ARCSTAT_BUMP(arcstat_hash_collisions);
1483 ARCSTAT_BUMP(arcstat_hash_chains);
1485 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1488 ARCSTAT_BUMP(arcstat_hash_elements);
1489 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1495 buf_hash_remove(arc_buf_hdr_t *hdr)
1497 arc_buf_hdr_t *fhdr, **hdrp;
1498 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1500 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1501 ASSERT(HDR_IN_HASH_TABLE(hdr));
1503 hdrp = &buf_hash_table.ht_table[idx];
1504 while ((fhdr = *hdrp) != hdr) {
1505 ASSERT3P(fhdr, !=, NULL);
1506 hdrp = &fhdr->b_hash_next;
1508 *hdrp = hdr->b_hash_next;
1509 hdr->b_hash_next = NULL;
1510 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1512 /* collect some hash table performance data */
1513 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1515 if (buf_hash_table.ht_table[idx] &&
1516 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1517 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1521 * Global data structures and functions for the buf kmem cache.
1523 static kmem_cache_t *hdr_full_cache;
1524 static kmem_cache_t *hdr_l2only_cache;
1525 static kmem_cache_t *buf_cache;
1532 kmem_free(buf_hash_table.ht_table,
1533 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1534 for (i = 0; i < BUF_LOCKS; i++)
1535 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1536 kmem_cache_destroy(hdr_full_cache);
1537 kmem_cache_destroy(hdr_l2only_cache);
1538 kmem_cache_destroy(buf_cache);
1542 * Constructor callback - called when the cache is empty
1543 * and a new buf is requested.
1547 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1549 arc_buf_hdr_t *hdr = vbuf;
1551 bzero(hdr, HDR_FULL_SIZE);
1552 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1553 refcount_create(&hdr->b_l1hdr.b_refcnt);
1554 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1555 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1556 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1563 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1565 arc_buf_hdr_t *hdr = vbuf;
1567 bzero(hdr, HDR_L2ONLY_SIZE);
1568 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1575 buf_cons(void *vbuf, void *unused, int kmflag)
1577 arc_buf_t *buf = vbuf;
1579 bzero(buf, sizeof (arc_buf_t));
1580 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1581 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1587 * Destructor callback - called when a cached buf is
1588 * no longer required.
1592 hdr_full_dest(void *vbuf, void *unused)
1594 arc_buf_hdr_t *hdr = vbuf;
1596 ASSERT(HDR_EMPTY(hdr));
1597 cv_destroy(&hdr->b_l1hdr.b_cv);
1598 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1599 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1600 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1601 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1606 hdr_l2only_dest(void *vbuf, void *unused)
1608 arc_buf_hdr_t *hdr = vbuf;
1610 ASSERT(HDR_EMPTY(hdr));
1611 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1616 buf_dest(void *vbuf, void *unused)
1618 arc_buf_t *buf = vbuf;
1620 mutex_destroy(&buf->b_evict_lock);
1621 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1625 * Reclaim callback -- invoked when memory is low.
1629 hdr_recl(void *unused)
1631 dprintf("hdr_recl called\n");
1633 * umem calls the reclaim func when we destroy the buf cache,
1634 * which is after we do arc_fini().
1637 cv_signal(&arc_reclaim_thread_cv);
1644 uint64_t hsize = 1ULL << 12;
1648 * The hash table is big enough to fill all of physical memory
1649 * with an average block size of zfs_arc_average_blocksize (default 8K).
1650 * By default, the table will take up
1651 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1653 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1656 buf_hash_table.ht_mask = hsize - 1;
1657 buf_hash_table.ht_table =
1658 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1659 if (buf_hash_table.ht_table == NULL) {
1660 ASSERT(hsize > (1ULL << 8));
1665 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1666 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1667 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1668 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1670 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1671 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1673 for (i = 0; i < 256; i++)
1674 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1675 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1677 for (i = 0; i < BUF_LOCKS; i++) {
1678 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1679 NULL, MUTEX_DEFAULT, NULL);
1683 #define ARC_MINTIME (hz>>4) /* 62 ms */
1685 static inline boolean_t
1686 arc_buf_is_shared(arc_buf_t *buf)
1688 boolean_t shared = (buf->b_data != NULL &&
1689 buf->b_data == buf->b_hdr->b_l1hdr.b_pdata);
1690 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1695 arc_cksum_free(arc_buf_hdr_t *hdr)
1697 ASSERT(HDR_HAS_L1HDR(hdr));
1698 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1699 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1700 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1701 hdr->b_l1hdr.b_freeze_cksum = NULL;
1703 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1707 arc_cksum_verify(arc_buf_t *buf)
1709 arc_buf_hdr_t *hdr = buf->b_hdr;
1712 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1715 ASSERT(HDR_HAS_L1HDR(hdr));
1717 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1718 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1719 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1722 fletcher_2_native(buf->b_data, HDR_GET_LSIZE(hdr), NULL, &zc);
1723 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1724 panic("buffer modified while frozen!");
1725 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1729 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1731 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1732 boolean_t valid_cksum;
1734 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1735 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1738 * We rely on the blkptr's checksum to determine if the block
1739 * is valid or not. When compressed arc is enabled, the l2arc
1740 * writes the block to the l2arc just as it appears in the pool.
1741 * This allows us to use the blkptr's checksum to validate the
1742 * data that we just read off of the l2arc without having to store
1743 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1744 * arc is disabled, then the data written to the l2arc is always
1745 * uncompressed and won't match the block as it exists in the main
1746 * pool. When this is the case, we must first compress it if it is
1747 * compressed on the main pool before we can validate the checksum.
1749 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1750 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1751 uint64_t lsize = HDR_GET_LSIZE(hdr);
1754 void *cbuf = zio_buf_alloc(HDR_GET_PSIZE(hdr));
1755 csize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1756 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1757 if (csize < HDR_GET_PSIZE(hdr)) {
1759 * Compressed blocks are always a multiple of the
1760 * smallest ashift in the pool. Ideally, we would
1761 * like to round up the csize to the next
1762 * spa_min_ashift but that value may have changed
1763 * since the block was last written. Instead,
1764 * we rely on the fact that the hdr's psize
1765 * was set to the psize of the block when it was
1766 * last written. We set the csize to that value
1767 * and zero out any part that should not contain
1770 bzero((char *)cbuf + csize, HDR_GET_PSIZE(hdr) - csize);
1771 csize = HDR_GET_PSIZE(hdr);
1773 zio_push_transform(zio, cbuf, csize, HDR_GET_PSIZE(hdr), NULL);
1777 * Block pointers always store the checksum for the logical data.
1778 * If the block pointer has the gang bit set, then the checksum
1779 * it represents is for the reconstituted data and not for an
1780 * individual gang member. The zio pipeline, however, must be able to
1781 * determine the checksum of each of the gang constituents so it
1782 * treats the checksum comparison differently than what we need
1783 * for l2arc blocks. This prevents us from using the
1784 * zio_checksum_error() interface directly. Instead we must call the
1785 * zio_checksum_error_impl() so that we can ensure the checksum is
1786 * generated using the correct checksum algorithm and accounts for the
1787 * logical I/O size and not just a gang fragment.
1789 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1790 BP_GET_CHECKSUM(zio->io_bp), zio->io_data, zio->io_size,
1791 zio->io_offset, NULL) == 0);
1792 zio_pop_transforms(zio);
1793 return (valid_cksum);
1797 arc_cksum_compute(arc_buf_t *buf)
1799 arc_buf_hdr_t *hdr = buf->b_hdr;
1801 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1804 ASSERT(HDR_HAS_L1HDR(hdr));
1805 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1806 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1807 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1810 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1812 fletcher_2_native(buf->b_data, HDR_GET_LSIZE(hdr), NULL,
1813 hdr->b_l1hdr.b_freeze_cksum);
1814 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1822 typedef struct procctl {
1830 arc_buf_unwatch(arc_buf_t *buf)
1837 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1838 ctl.prwatch.pr_size = 0;
1839 ctl.prwatch.pr_wflags = 0;
1840 result = write(arc_procfd, &ctl, sizeof (ctl));
1841 ASSERT3U(result, ==, sizeof (ctl));
1848 arc_buf_watch(arc_buf_t *buf)
1855 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1856 ctl.prwatch.pr_size = HDR_GET_LSIZE(buf->b_hdr);
1857 ctl.prwatch.pr_wflags = WA_WRITE;
1858 result = write(arc_procfd, &ctl, sizeof (ctl));
1859 ASSERT3U(result, ==, sizeof (ctl));
1863 #endif /* illumos */
1865 static arc_buf_contents_t
1866 arc_buf_type(arc_buf_hdr_t *hdr)
1868 arc_buf_contents_t type;
1869 if (HDR_ISTYPE_METADATA(hdr)) {
1870 type = ARC_BUFC_METADATA;
1872 type = ARC_BUFC_DATA;
1874 VERIFY3U(hdr->b_type, ==, type);
1879 arc_bufc_to_flags(arc_buf_contents_t type)
1883 /* metadata field is 0 if buffer contains normal data */
1885 case ARC_BUFC_METADATA:
1886 return (ARC_FLAG_BUFC_METADATA);
1890 panic("undefined ARC buffer type!");
1891 return ((uint32_t)-1);
1895 arc_buf_thaw(arc_buf_t *buf)
1897 arc_buf_hdr_t *hdr = buf->b_hdr;
1899 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1900 if (hdr->b_l1hdr.b_state != arc_anon)
1901 panic("modifying non-anon buffer!");
1902 if (HDR_IO_IN_PROGRESS(hdr))
1903 panic("modifying buffer while i/o in progress!");
1904 arc_cksum_verify(buf);
1907 ASSERT(HDR_HAS_L1HDR(hdr));
1908 arc_cksum_free(hdr);
1910 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1912 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1913 if (hdr->b_l1hdr.b_thawed != NULL)
1914 kmem_free(hdr->b_l1hdr.b_thawed, 1);
1915 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1919 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1922 arc_buf_unwatch(buf);
1927 arc_buf_freeze(arc_buf_t *buf)
1929 arc_buf_hdr_t *hdr = buf->b_hdr;
1930 kmutex_t *hash_lock;
1932 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1935 hash_lock = HDR_LOCK(hdr);
1936 mutex_enter(hash_lock);
1938 ASSERT(HDR_HAS_L1HDR(hdr));
1939 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
1940 hdr->b_l1hdr.b_state == arc_anon);
1941 arc_cksum_compute(buf);
1942 mutex_exit(hash_lock);
1947 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1948 * the following functions should be used to ensure that the flags are
1949 * updated in a thread-safe way. When manipulating the flags either
1950 * the hash_lock must be held or the hdr must be undiscoverable. This
1951 * ensures that we're not racing with any other threads when updating
1955 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1957 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1958 hdr->b_flags |= flags;
1962 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1964 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1965 hdr->b_flags &= ~flags;
1969 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1970 * done in a special way since we have to clear and set bits
1971 * at the same time. Consumers that wish to set the compression bits
1972 * must use this function to ensure that the flags are updated in
1973 * thread-safe manner.
1976 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1978 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1981 * Holes and embedded blocks will always have a psize = 0 so
1982 * we ignore the compression of the blkptr and set the
1983 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
1984 * Holes and embedded blocks remain anonymous so we don't
1985 * want to uncompress them. Mark them as uncompressed.
1987 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1988 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1989 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
1990 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1991 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1993 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1994 HDR_SET_COMPRESS(hdr, cmp);
1995 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1996 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
2001 arc_decompress(arc_buf_t *buf)
2003 arc_buf_hdr_t *hdr = buf->b_hdr;
2004 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2007 if (arc_buf_is_shared(buf)) {
2008 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2009 } else if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) {
2011 * The arc_buf_hdr_t is either not compressed or is
2012 * associated with an embedded block or a hole in which
2013 * case they remain anonymous.
2015 IMPLY(HDR_COMPRESSION_ENABLED(hdr), HDR_GET_PSIZE(hdr) == 0 ||
2016 HDR_GET_PSIZE(hdr) == HDR_GET_LSIZE(hdr));
2017 ASSERT(!HDR_SHARED_DATA(hdr));
2018 bcopy(hdr->b_l1hdr.b_pdata, buf->b_data, HDR_GET_LSIZE(hdr));
2020 ASSERT(!HDR_SHARED_DATA(hdr));
2021 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2022 error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2023 hdr->b_l1hdr.b_pdata, buf->b_data, HDR_GET_PSIZE(hdr),
2024 HDR_GET_LSIZE(hdr));
2026 zfs_dbgmsg("hdr %p, compress %d, psize %d, lsize %d",
2027 hdr, HDR_GET_COMPRESS(hdr), HDR_GET_PSIZE(hdr),
2028 HDR_GET_LSIZE(hdr));
2029 return (SET_ERROR(EIO));
2032 if (bswap != DMU_BSWAP_NUMFUNCS) {
2033 ASSERT(!HDR_SHARED_DATA(hdr));
2034 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2035 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2037 arc_cksum_compute(buf);
2042 * Return the size of the block, b_pdata, that is stored in the arc_buf_hdr_t.
2045 arc_hdr_size(arc_buf_hdr_t *hdr)
2049 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2050 HDR_GET_PSIZE(hdr) > 0) {
2051 size = HDR_GET_PSIZE(hdr);
2053 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2054 size = HDR_GET_LSIZE(hdr);
2060 * Increment the amount of evictable space in the arc_state_t's refcount.
2061 * We account for the space used by the hdr and the arc buf individually
2062 * so that we can add and remove them from the refcount individually.
2065 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2067 arc_buf_contents_t type = arc_buf_type(hdr);
2068 uint64_t lsize = HDR_GET_LSIZE(hdr);
2070 ASSERT(HDR_HAS_L1HDR(hdr));
2072 if (GHOST_STATE(state)) {
2073 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2074 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2075 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2076 (void) refcount_add_many(&state->arcs_esize[type], lsize, hdr);
2080 ASSERT(!GHOST_STATE(state));
2081 if (hdr->b_l1hdr.b_pdata != NULL) {
2082 (void) refcount_add_many(&state->arcs_esize[type],
2083 arc_hdr_size(hdr), hdr);
2085 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2086 buf = buf->b_next) {
2087 if (arc_buf_is_shared(buf)) {
2088 ASSERT(ARC_BUF_LAST(buf));
2091 (void) refcount_add_many(&state->arcs_esize[type], lsize, buf);
2096 * Decrement the amount of evictable space in the arc_state_t's refcount.
2097 * We account for the space used by the hdr and the arc buf individually
2098 * so that we can add and remove them from the refcount individually.
2101 arc_evitable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2103 arc_buf_contents_t type = arc_buf_type(hdr);
2104 uint64_t lsize = HDR_GET_LSIZE(hdr);
2106 ASSERT(HDR_HAS_L1HDR(hdr));
2108 if (GHOST_STATE(state)) {
2109 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2110 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2111 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2112 (void) refcount_remove_many(&state->arcs_esize[type],
2117 ASSERT(!GHOST_STATE(state));
2118 if (hdr->b_l1hdr.b_pdata != NULL) {
2119 (void) refcount_remove_many(&state->arcs_esize[type],
2120 arc_hdr_size(hdr), hdr);
2122 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2123 buf = buf->b_next) {
2124 if (arc_buf_is_shared(buf)) {
2125 ASSERT(ARC_BUF_LAST(buf));
2128 (void) refcount_remove_many(&state->arcs_esize[type],
2134 * Add a reference to this hdr indicating that someone is actively
2135 * referencing that memory. When the refcount transitions from 0 to 1,
2136 * we remove it from the respective arc_state_t list to indicate that
2137 * it is not evictable.
2140 add_reference(arc_buf_hdr_t *hdr, void *tag)
2142 ASSERT(HDR_HAS_L1HDR(hdr));
2143 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2144 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2145 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2146 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2149 arc_state_t *state = hdr->b_l1hdr.b_state;
2151 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2152 (state != arc_anon)) {
2153 /* We don't use the L2-only state list. */
2154 if (state != arc_l2c_only) {
2155 multilist_remove(&state->arcs_list[arc_buf_type(hdr)],
2157 arc_evitable_space_decrement(hdr, state);
2159 /* remove the prefetch flag if we get a reference */
2160 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2165 * Remove a reference from this hdr. When the reference transitions from
2166 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2167 * list making it eligible for eviction.
2170 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2173 arc_state_t *state = hdr->b_l1hdr.b_state;
2175 ASSERT(HDR_HAS_L1HDR(hdr));
2176 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2177 ASSERT(!GHOST_STATE(state));
2180 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2181 * check to prevent usage of the arc_l2c_only list.
2183 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2184 (state != arc_anon)) {
2185 multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr);
2186 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2187 arc_evictable_space_increment(hdr, state);
2193 * Move the supplied buffer to the indicated state. The hash lock
2194 * for the buffer must be held by the caller.
2197 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2198 kmutex_t *hash_lock)
2200 arc_state_t *old_state;
2203 boolean_t update_old, update_new;
2204 arc_buf_contents_t buftype = arc_buf_type(hdr);
2207 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2208 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2209 * L1 hdr doesn't always exist when we change state to arc_anon before
2210 * destroying a header, in which case reallocating to add the L1 hdr is
2213 if (HDR_HAS_L1HDR(hdr)) {
2214 old_state = hdr->b_l1hdr.b_state;
2215 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2216 bufcnt = hdr->b_l1hdr.b_bufcnt;
2217 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pdata != NULL);
2219 old_state = arc_l2c_only;
2222 update_old = B_FALSE;
2224 update_new = update_old;
2226 ASSERT(MUTEX_HELD(hash_lock));
2227 ASSERT3P(new_state, !=, old_state);
2228 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2229 ASSERT(old_state != arc_anon || bufcnt <= 1);
2232 * If this buffer is evictable, transfer it from the
2233 * old state list to the new state list.
2236 if (old_state != arc_anon && old_state != arc_l2c_only) {
2237 ASSERT(HDR_HAS_L1HDR(hdr));
2238 multilist_remove(&old_state->arcs_list[buftype], hdr);
2240 if (GHOST_STATE(old_state)) {
2242 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2243 update_old = B_TRUE;
2245 arc_evitable_space_decrement(hdr, old_state);
2247 if (new_state != arc_anon && new_state != arc_l2c_only) {
2250 * An L1 header always exists here, since if we're
2251 * moving to some L1-cached state (i.e. not l2c_only or
2252 * anonymous), we realloc the header to add an L1hdr
2255 ASSERT(HDR_HAS_L1HDR(hdr));
2256 multilist_insert(&new_state->arcs_list[buftype], hdr);
2258 if (GHOST_STATE(new_state)) {
2260 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2261 update_new = B_TRUE;
2263 arc_evictable_space_increment(hdr, new_state);
2267 ASSERT(!HDR_EMPTY(hdr));
2268 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2269 buf_hash_remove(hdr);
2271 /* adjust state sizes (ignore arc_l2c_only) */
2273 if (update_new && new_state != arc_l2c_only) {
2274 ASSERT(HDR_HAS_L1HDR(hdr));
2275 if (GHOST_STATE(new_state)) {
2279 * When moving a header to a ghost state, we first
2280 * remove all arc buffers. Thus, we'll have a
2281 * bufcnt of zero, and no arc buffer to use for
2282 * the reference. As a result, we use the arc
2283 * header pointer for the reference.
2285 (void) refcount_add_many(&new_state->arcs_size,
2286 HDR_GET_LSIZE(hdr), hdr);
2287 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2289 uint32_t buffers = 0;
2292 * Each individual buffer holds a unique reference,
2293 * thus we must remove each of these references one
2296 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2297 buf = buf->b_next) {
2298 ASSERT3U(bufcnt, !=, 0);
2302 * When the arc_buf_t is sharing the data
2303 * block with the hdr, the owner of the
2304 * reference belongs to the hdr. Only
2305 * add to the refcount if the arc_buf_t is
2308 if (arc_buf_is_shared(buf)) {
2309 ASSERT(ARC_BUF_LAST(buf));
2313 (void) refcount_add_many(&new_state->arcs_size,
2314 HDR_GET_LSIZE(hdr), buf);
2316 ASSERT3U(bufcnt, ==, buffers);
2318 if (hdr->b_l1hdr.b_pdata != NULL) {
2319 (void) refcount_add_many(&new_state->arcs_size,
2320 arc_hdr_size(hdr), hdr);
2322 ASSERT(GHOST_STATE(old_state));
2327 if (update_old && old_state != arc_l2c_only) {
2328 ASSERT(HDR_HAS_L1HDR(hdr));
2329 if (GHOST_STATE(old_state)) {
2333 * When moving a header off of a ghost state,
2334 * the header will not contain any arc buffers.
2335 * We use the arc header pointer for the reference
2336 * which is exactly what we did when we put the
2337 * header on the ghost state.
2340 (void) refcount_remove_many(&old_state->arcs_size,
2341 HDR_GET_LSIZE(hdr), hdr);
2342 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2344 uint32_t buffers = 0;
2347 * Each individual buffer holds a unique reference,
2348 * thus we must remove each of these references one
2351 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2352 buf = buf->b_next) {
2353 ASSERT3P(bufcnt, !=, 0);
2357 * When the arc_buf_t is sharing the data
2358 * block with the hdr, the owner of the
2359 * reference belongs to the hdr. Only
2360 * add to the refcount if the arc_buf_t is
2363 if (arc_buf_is_shared(buf)) {
2364 ASSERT(ARC_BUF_LAST(buf));
2368 (void) refcount_remove_many(
2369 &old_state->arcs_size, HDR_GET_LSIZE(hdr),
2372 ASSERT3U(bufcnt, ==, buffers);
2373 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
2374 (void) refcount_remove_many(
2375 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2379 if (HDR_HAS_L1HDR(hdr))
2380 hdr->b_l1hdr.b_state = new_state;
2383 * L2 headers should never be on the L2 state list since they don't
2384 * have L1 headers allocated.
2386 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2387 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2391 arc_space_consume(uint64_t space, arc_space_type_t type)
2393 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2396 case ARC_SPACE_DATA:
2397 ARCSTAT_INCR(arcstat_data_size, space);
2399 case ARC_SPACE_META:
2400 ARCSTAT_INCR(arcstat_metadata_size, space);
2402 case ARC_SPACE_OTHER:
2403 ARCSTAT_INCR(arcstat_other_size, space);
2405 case ARC_SPACE_HDRS:
2406 ARCSTAT_INCR(arcstat_hdr_size, space);
2408 case ARC_SPACE_L2HDRS:
2409 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
2413 if (type != ARC_SPACE_DATA)
2414 ARCSTAT_INCR(arcstat_meta_used, space);
2416 atomic_add_64(&arc_size, space);
2420 arc_space_return(uint64_t space, arc_space_type_t type)
2422 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2425 case ARC_SPACE_DATA:
2426 ARCSTAT_INCR(arcstat_data_size, -space);
2428 case ARC_SPACE_META:
2429 ARCSTAT_INCR(arcstat_metadata_size, -space);
2431 case ARC_SPACE_OTHER:
2432 ARCSTAT_INCR(arcstat_other_size, -space);
2434 case ARC_SPACE_HDRS:
2435 ARCSTAT_INCR(arcstat_hdr_size, -space);
2437 case ARC_SPACE_L2HDRS:
2438 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
2442 if (type != ARC_SPACE_DATA) {
2443 ASSERT(arc_meta_used >= space);
2444 if (arc_meta_max < arc_meta_used)
2445 arc_meta_max = arc_meta_used;
2446 ARCSTAT_INCR(arcstat_meta_used, -space);
2449 ASSERT(arc_size >= space);
2450 atomic_add_64(&arc_size, -space);
2454 * Allocate an initial buffer for this hdr, subsequent buffers will
2455 * use arc_buf_clone().
2458 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag)
2462 ASSERT(HDR_HAS_L1HDR(hdr));
2463 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2464 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2465 hdr->b_type == ARC_BUFC_METADATA);
2467 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2468 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2469 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2471 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2476 add_reference(hdr, tag);
2479 * We're about to change the hdr's b_flags. We must either
2480 * hold the hash_lock or be undiscoverable.
2482 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2485 * If the hdr's data can be shared (no byteswapping, hdr is
2486 * uncompressed, hdr's data is not currently being written to the
2487 * L2ARC write) then we share the data buffer and set the appropriate
2488 * bit in the hdr's b_flags to indicate the hdr is sharing it's
2489 * b_pdata with the arc_buf_t. Otherwise, we allocate a new buffer to
2490 * store the buf's data.
2492 if (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2493 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF && !HDR_L2_WRITING(hdr)) {
2494 buf->b_data = hdr->b_l1hdr.b_pdata;
2495 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2497 buf->b_data = arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2498 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2499 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2501 VERIFY3P(buf->b_data, !=, NULL);
2503 hdr->b_l1hdr.b_buf = buf;
2504 hdr->b_l1hdr.b_bufcnt += 1;
2510 * Used when allocating additional buffers.
2513 arc_buf_clone(arc_buf_t *from)
2516 arc_buf_hdr_t *hdr = from->b_hdr;
2517 uint64_t size = HDR_GET_LSIZE(hdr);
2519 ASSERT(HDR_HAS_L1HDR(hdr));
2520 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2522 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2525 buf->b_next = hdr->b_l1hdr.b_buf;
2526 hdr->b_l1hdr.b_buf = buf;
2527 buf->b_data = arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2528 bcopy(from->b_data, buf->b_data, size);
2529 hdr->b_l1hdr.b_bufcnt += 1;
2531 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2535 static char *arc_onloan_tag = "onloan";
2538 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2539 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2540 * buffers must be returned to the arc before they can be used by the DMU or
2544 arc_loan_buf(spa_t *spa, int size)
2548 buf = arc_alloc_buf(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
2550 atomic_add_64(&arc_loaned_bytes, size);
2555 * Return a loaned arc buffer to the arc.
2558 arc_return_buf(arc_buf_t *buf, void *tag)
2560 arc_buf_hdr_t *hdr = buf->b_hdr;
2562 ASSERT3P(buf->b_data, !=, NULL);
2563 ASSERT(HDR_HAS_L1HDR(hdr));
2564 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2565 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2567 atomic_add_64(&arc_loaned_bytes, -HDR_GET_LSIZE(hdr));
2570 /* Detach an arc_buf from a dbuf (tag) */
2572 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2574 arc_buf_hdr_t *hdr = buf->b_hdr;
2576 ASSERT3P(buf->b_data, !=, NULL);
2577 ASSERT(HDR_HAS_L1HDR(hdr));
2578 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2579 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2581 atomic_add_64(&arc_loaned_bytes, HDR_GET_LSIZE(hdr));
2585 l2arc_free_data_on_write(void *data, size_t size, arc_buf_contents_t type)
2587 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2589 df->l2df_data = data;
2590 df->l2df_size = size;
2591 df->l2df_type = type;
2592 mutex_enter(&l2arc_free_on_write_mtx);
2593 list_insert_head(l2arc_free_on_write, df);
2594 mutex_exit(&l2arc_free_on_write_mtx);
2598 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
2600 arc_state_t *state = hdr->b_l1hdr.b_state;
2601 arc_buf_contents_t type = arc_buf_type(hdr);
2602 uint64_t size = arc_hdr_size(hdr);
2604 /* protected by hash lock, if in the hash table */
2605 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2606 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2607 ASSERT(state != arc_anon && state != arc_l2c_only);
2609 (void) refcount_remove_many(&state->arcs_esize[type],
2612 (void) refcount_remove_many(&state->arcs_size, size, hdr);
2613 if (type == ARC_BUFC_METADATA) {
2614 arc_space_return(size, ARC_SPACE_META);
2616 ASSERT(type == ARC_BUFC_DATA);
2617 arc_space_return(size, ARC_SPACE_DATA);
2620 l2arc_free_data_on_write(hdr->b_l1hdr.b_pdata, size, type);
2624 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2625 * data buffer, we transfer the refcount ownership to the hdr and update
2626 * the appropriate kstats.
2629 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2631 arc_state_t *state = hdr->b_l1hdr.b_state;
2633 ASSERT(!HDR_SHARED_DATA(hdr));
2634 ASSERT(!arc_buf_is_shared(buf));
2635 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2636 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2639 * Start sharing the data buffer. We transfer the
2640 * refcount ownership to the hdr since it always owns
2641 * the refcount whenever an arc_buf_t is shared.
2643 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
2644 hdr->b_l1hdr.b_pdata = buf->b_data;
2645 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2648 * Since we've transferred ownership to the hdr we need
2649 * to increment its compressed and uncompressed kstats and
2650 * decrement the overhead size.
2652 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2653 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2654 ARCSTAT_INCR(arcstat_overhead_size, -HDR_GET_LSIZE(hdr));
2658 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2660 arc_state_t *state = hdr->b_l1hdr.b_state;
2662 ASSERT(HDR_SHARED_DATA(hdr));
2663 ASSERT(arc_buf_is_shared(buf));
2664 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
2665 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2668 * We are no longer sharing this buffer so we need
2669 * to transfer its ownership to the rightful owner.
2671 refcount_transfer_ownership(&state->arcs_size, hdr, buf);
2672 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2673 hdr->b_l1hdr.b_pdata = NULL;
2676 * Since the buffer is no longer shared between
2677 * the arc buf and the hdr, count it as overhead.
2679 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2680 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2681 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2685 * Free up buf->b_data and if 'remove' is set, then pull the
2686 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
2689 arc_buf_destroy_impl(arc_buf_t *buf, boolean_t remove)
2692 arc_buf_hdr_t *hdr = buf->b_hdr;
2693 uint64_t size = HDR_GET_LSIZE(hdr);
2694 boolean_t destroyed_buf_is_shared = arc_buf_is_shared(buf);
2697 * Free up the data associated with the buf but only
2698 * if we're not sharing this with the hdr. If we are sharing
2699 * it with the hdr, then hdr will have performed the allocation
2700 * so allow it to do the free.
2702 if (buf->b_data != NULL) {
2704 * We're about to change the hdr's b_flags. We must either
2705 * hold the hash_lock or be undiscoverable.
2707 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2709 arc_cksum_verify(buf);
2711 arc_buf_unwatch(buf);
2714 if (destroyed_buf_is_shared) {
2715 ASSERT(ARC_BUF_LAST(buf));
2716 ASSERT(HDR_SHARED_DATA(hdr));
2717 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2719 arc_free_data_buf(hdr, buf->b_data, size, buf);
2720 ARCSTAT_INCR(arcstat_overhead_size, -size);
2724 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
2725 hdr->b_l1hdr.b_bufcnt -= 1;
2728 /* only remove the buf if requested */
2732 /* remove the buf from the hdr list */
2733 arc_buf_t *lastbuf = NULL;
2734 bufp = &hdr->b_l1hdr.b_buf;
2735 while (*bufp != NULL) {
2737 *bufp = buf->b_next;
2740 * If we've removed a buffer in the middle of
2741 * the list then update the lastbuf and update
2744 if (*bufp != NULL) {
2746 bufp = &(*bufp)->b_next;
2750 ASSERT3P(lastbuf, !=, buf);
2753 * If the current arc_buf_t is sharing its data
2754 * buffer with the hdr, then reassign the hdr's
2755 * b_pdata to share it with the new buffer at the end
2756 * of the list. The shared buffer is always the last one
2757 * on the hdr's buffer list.
2759 if (destroyed_buf_is_shared && lastbuf != NULL) {
2760 ASSERT(ARC_BUF_LAST(buf));
2761 ASSERT(ARC_BUF_LAST(lastbuf));
2762 VERIFY(!arc_buf_is_shared(lastbuf));
2764 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
2765 arc_hdr_free_pdata(hdr);
2768 * We must setup a new shared block between the
2769 * last buffer and the hdr. The data would have
2770 * been allocated by the arc buf so we need to transfer
2771 * ownership to the hdr since it's now being shared.
2773 arc_share_buf(hdr, lastbuf);
2774 } else if (HDR_SHARED_DATA(hdr)) {
2775 ASSERT(arc_buf_is_shared(lastbuf));
2778 if (hdr->b_l1hdr.b_bufcnt == 0)
2779 arc_cksum_free(hdr);
2781 /* clean up the buf */
2783 kmem_cache_free(buf_cache, buf);
2787 arc_hdr_alloc_pdata(arc_buf_hdr_t *hdr)
2789 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2790 ASSERT(HDR_HAS_L1HDR(hdr));
2791 ASSERT(!HDR_SHARED_DATA(hdr));
2793 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2794 hdr->b_l1hdr.b_pdata = arc_get_data_buf(hdr, arc_hdr_size(hdr), hdr);
2795 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
2796 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
2798 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2799 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2803 arc_hdr_free_pdata(arc_buf_hdr_t *hdr)
2805 ASSERT(HDR_HAS_L1HDR(hdr));
2806 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
2809 * If the hdr is currently being written to the l2arc then
2810 * we defer freeing the data by adding it to the l2arc_free_on_write
2811 * list. The l2arc will free the data once it's finished
2812 * writing it to the l2arc device.
2814 if (HDR_L2_WRITING(hdr)) {
2815 arc_hdr_free_on_write(hdr);
2816 ARCSTAT_BUMP(arcstat_l2_free_on_write);
2818 arc_free_data_buf(hdr, hdr->b_l1hdr.b_pdata,
2819 arc_hdr_size(hdr), hdr);
2821 hdr->b_l1hdr.b_pdata = NULL;
2822 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
2824 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2825 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2828 static arc_buf_hdr_t *
2829 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
2830 enum zio_compress compress, arc_buf_contents_t type)
2834 ASSERT3U(lsize, >, 0);
2835 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
2837 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
2838 ASSERT(HDR_EMPTY(hdr));
2839 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
2840 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
2841 HDR_SET_PSIZE(hdr, psize);
2842 HDR_SET_LSIZE(hdr, lsize);
2846 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
2847 arc_hdr_set_compress(hdr, compress);
2849 hdr->b_l1hdr.b_state = arc_anon;
2850 hdr->b_l1hdr.b_arc_access = 0;
2851 hdr->b_l1hdr.b_bufcnt = 0;
2852 hdr->b_l1hdr.b_buf = NULL;
2855 * Allocate the hdr's buffer. This will contain either
2856 * the compressed or uncompressed data depending on the block
2857 * it references and compressed arc enablement.
2859 arc_hdr_alloc_pdata(hdr);
2860 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2866 * Transition between the two allocation states for the arc_buf_hdr struct.
2867 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
2868 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
2869 * version is used when a cache buffer is only in the L2ARC in order to reduce
2872 static arc_buf_hdr_t *
2873 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
2875 ASSERT(HDR_HAS_L2HDR(hdr));
2877 arc_buf_hdr_t *nhdr;
2878 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2880 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
2881 (old == hdr_l2only_cache && new == hdr_full_cache));
2883 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
2885 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
2886 buf_hash_remove(hdr);
2888 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
2890 if (new == hdr_full_cache) {
2891 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
2893 * arc_access and arc_change_state need to be aware that a
2894 * header has just come out of L2ARC, so we set its state to
2895 * l2c_only even though it's about to change.
2897 nhdr->b_l1hdr.b_state = arc_l2c_only;
2899 /* Verify previous threads set to NULL before freeing */
2900 ASSERT3P(nhdr->b_l1hdr.b_pdata, ==, NULL);
2902 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2903 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2904 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
2907 * If we've reached here, We must have been called from
2908 * arc_evict_hdr(), as such we should have already been
2909 * removed from any ghost list we were previously on
2910 * (which protects us from racing with arc_evict_state),
2911 * thus no locking is needed during this check.
2913 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
2916 * A buffer must not be moved into the arc_l2c_only
2917 * state if it's not finished being written out to the
2918 * l2arc device. Otherwise, the b_l1hdr.b_pdata field
2919 * might try to be accessed, even though it was removed.
2921 VERIFY(!HDR_L2_WRITING(hdr));
2922 VERIFY3P(hdr->b_l1hdr.b_pdata, ==, NULL);
2925 if (hdr->b_l1hdr.b_thawed != NULL) {
2926 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2927 hdr->b_l1hdr.b_thawed = NULL;
2931 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
2934 * The header has been reallocated so we need to re-insert it into any
2937 (void) buf_hash_insert(nhdr, NULL);
2939 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
2941 mutex_enter(&dev->l2ad_mtx);
2944 * We must place the realloc'ed header back into the list at
2945 * the same spot. Otherwise, if it's placed earlier in the list,
2946 * l2arc_write_buffers() could find it during the function's
2947 * write phase, and try to write it out to the l2arc.
2949 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
2950 list_remove(&dev->l2ad_buflist, hdr);
2952 mutex_exit(&dev->l2ad_mtx);
2955 * Since we're using the pointer address as the tag when
2956 * incrementing and decrementing the l2ad_alloc refcount, we
2957 * must remove the old pointer (that we're about to destroy) and
2958 * add the new pointer to the refcount. Otherwise we'd remove
2959 * the wrong pointer address when calling arc_hdr_destroy() later.
2962 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
2963 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
2965 buf_discard_identity(hdr);
2966 kmem_cache_free(old, hdr);
2972 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
2973 * The buf is returned thawed since we expect the consumer to modify it.
2976 arc_alloc_buf(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type)
2978 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
2979 ZIO_COMPRESS_OFF, type);
2980 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
2981 arc_buf_t *buf = arc_buf_alloc_impl(hdr, tag);
2987 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
2989 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
2990 l2arc_dev_t *dev = l2hdr->b_dev;
2991 uint64_t asize = arc_hdr_size(hdr);
2993 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
2994 ASSERT(HDR_HAS_L2HDR(hdr));
2996 list_remove(&dev->l2ad_buflist, hdr);
2998 ARCSTAT_INCR(arcstat_l2_asize, -asize);
2999 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr));
3001 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
3003 (void) refcount_remove_many(&dev->l2ad_alloc, asize, hdr);
3004 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3008 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3010 if (HDR_HAS_L1HDR(hdr)) {
3011 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3012 hdr->b_l1hdr.b_bufcnt > 0);
3013 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3014 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3016 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3017 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3019 if (!HDR_EMPTY(hdr))
3020 buf_discard_identity(hdr);
3022 if (HDR_HAS_L2HDR(hdr)) {
3023 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3024 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3027 mutex_enter(&dev->l2ad_mtx);
3030 * Even though we checked this conditional above, we
3031 * need to check this again now that we have the
3032 * l2ad_mtx. This is because we could be racing with
3033 * another thread calling l2arc_evict() which might have
3034 * destroyed this header's L2 portion as we were waiting
3035 * to acquire the l2ad_mtx. If that happens, we don't
3036 * want to re-destroy the header's L2 portion.
3038 if (HDR_HAS_L2HDR(hdr)) {
3040 arc_hdr_l2hdr_destroy(hdr);
3044 mutex_exit(&dev->l2ad_mtx);
3047 if (HDR_HAS_L1HDR(hdr)) {
3048 arc_cksum_free(hdr);
3050 while (hdr->b_l1hdr.b_buf != NULL)
3051 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf, B_TRUE);
3054 if (hdr->b_l1hdr.b_thawed != NULL) {
3055 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3056 hdr->b_l1hdr.b_thawed = NULL;
3060 if (hdr->b_l1hdr.b_pdata != NULL) {
3061 arc_hdr_free_pdata(hdr);
3065 ASSERT3P(hdr->b_hash_next, ==, NULL);
3066 if (HDR_HAS_L1HDR(hdr)) {
3067 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3068 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3069 kmem_cache_free(hdr_full_cache, hdr);
3071 kmem_cache_free(hdr_l2only_cache, hdr);
3076 arc_buf_destroy(arc_buf_t *buf, void* tag)
3078 arc_buf_hdr_t *hdr = buf->b_hdr;
3079 kmutex_t *hash_lock = HDR_LOCK(hdr);
3081 if (hdr->b_l1hdr.b_state == arc_anon) {
3082 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3083 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3084 VERIFY0(remove_reference(hdr, NULL, tag));
3085 arc_hdr_destroy(hdr);
3089 mutex_enter(hash_lock);
3090 ASSERT3P(hdr, ==, buf->b_hdr);
3091 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3092 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3093 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3094 ASSERT3P(buf->b_data, !=, NULL);
3096 (void) remove_reference(hdr, hash_lock, tag);
3097 arc_buf_destroy_impl(buf, B_TRUE);
3098 mutex_exit(hash_lock);
3102 arc_buf_size(arc_buf_t *buf)
3104 return (HDR_GET_LSIZE(buf->b_hdr));
3108 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3109 * state of the header is dependent on it's state prior to entering this
3110 * function. The following transitions are possible:
3112 * - arc_mru -> arc_mru_ghost
3113 * - arc_mfu -> arc_mfu_ghost
3114 * - arc_mru_ghost -> arc_l2c_only
3115 * - arc_mru_ghost -> deleted
3116 * - arc_mfu_ghost -> arc_l2c_only
3117 * - arc_mfu_ghost -> deleted
3120 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3122 arc_state_t *evicted_state, *state;
3123 int64_t bytes_evicted = 0;
3125 ASSERT(MUTEX_HELD(hash_lock));
3126 ASSERT(HDR_HAS_L1HDR(hdr));
3128 state = hdr->b_l1hdr.b_state;
3129 if (GHOST_STATE(state)) {
3130 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3131 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3134 * l2arc_write_buffers() relies on a header's L1 portion
3135 * (i.e. its b_pdata field) during its write phase.
3136 * Thus, we cannot push a header onto the arc_l2c_only
3137 * state (removing it's L1 piece) until the header is
3138 * done being written to the l2arc.
3140 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3141 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3142 return (bytes_evicted);
3145 ARCSTAT_BUMP(arcstat_deleted);
3146 bytes_evicted += HDR_GET_LSIZE(hdr);
3148 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3150 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
3151 if (HDR_HAS_L2HDR(hdr)) {
3152 ASSERT(hdr->b_l1hdr.b_pdata == NULL);
3154 * This buffer is cached on the 2nd Level ARC;
3155 * don't destroy the header.
3157 arc_change_state(arc_l2c_only, hdr, hash_lock);
3159 * dropping from L1+L2 cached to L2-only,
3160 * realloc to remove the L1 header.
3162 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3165 ASSERT(hdr->b_l1hdr.b_pdata == NULL);
3166 arc_change_state(arc_anon, hdr, hash_lock);
3167 arc_hdr_destroy(hdr);
3169 return (bytes_evicted);
3172 ASSERT(state == arc_mru || state == arc_mfu);
3173 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3175 /* prefetch buffers have a minimum lifespan */
3176 if (HDR_IO_IN_PROGRESS(hdr) ||
3177 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3178 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3179 arc_min_prefetch_lifespan)) {
3180 ARCSTAT_BUMP(arcstat_evict_skip);
3181 return (bytes_evicted);
3184 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3185 while (hdr->b_l1hdr.b_buf) {
3186 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3187 if (!mutex_tryenter(&buf->b_evict_lock)) {
3188 ARCSTAT_BUMP(arcstat_mutex_miss);
3191 if (buf->b_data != NULL)
3192 bytes_evicted += HDR_GET_LSIZE(hdr);
3193 mutex_exit(&buf->b_evict_lock);
3194 arc_buf_destroy_impl(buf, B_TRUE);
3197 if (HDR_HAS_L2HDR(hdr)) {
3198 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3200 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3201 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3202 HDR_GET_LSIZE(hdr));
3204 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3205 HDR_GET_LSIZE(hdr));
3209 if (hdr->b_l1hdr.b_bufcnt == 0) {
3210 arc_cksum_free(hdr);
3212 bytes_evicted += arc_hdr_size(hdr);
3215 * If this hdr is being evicted and has a compressed
3216 * buffer then we discard it here before we change states.
3217 * This ensures that the accounting is updated correctly
3218 * in arc_free_data_buf().
3220 arc_hdr_free_pdata(hdr);
3222 arc_change_state(evicted_state, hdr, hash_lock);
3223 ASSERT(HDR_IN_HASH_TABLE(hdr));
3224 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3225 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3228 return (bytes_evicted);
3232 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3233 uint64_t spa, int64_t bytes)
3235 multilist_sublist_t *mls;
3236 uint64_t bytes_evicted = 0;
3238 kmutex_t *hash_lock;
3239 int evict_count = 0;
3241 ASSERT3P(marker, !=, NULL);
3242 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3244 mls = multilist_sublist_lock(ml, idx);
3246 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3247 hdr = multilist_sublist_prev(mls, marker)) {
3248 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3249 (evict_count >= zfs_arc_evict_batch_limit))
3253 * To keep our iteration location, move the marker
3254 * forward. Since we're not holding hdr's hash lock, we
3255 * must be very careful and not remove 'hdr' from the
3256 * sublist. Otherwise, other consumers might mistake the
3257 * 'hdr' as not being on a sublist when they call the
3258 * multilist_link_active() function (they all rely on
3259 * the hash lock protecting concurrent insertions and
3260 * removals). multilist_sublist_move_forward() was
3261 * specifically implemented to ensure this is the case
3262 * (only 'marker' will be removed and re-inserted).
3264 multilist_sublist_move_forward(mls, marker);
3267 * The only case where the b_spa field should ever be
3268 * zero, is the marker headers inserted by
3269 * arc_evict_state(). It's possible for multiple threads
3270 * to be calling arc_evict_state() concurrently (e.g.
3271 * dsl_pool_close() and zio_inject_fault()), so we must
3272 * skip any markers we see from these other threads.
3274 if (hdr->b_spa == 0)
3277 /* we're only interested in evicting buffers of a certain spa */
3278 if (spa != 0 && hdr->b_spa != spa) {
3279 ARCSTAT_BUMP(arcstat_evict_skip);
3283 hash_lock = HDR_LOCK(hdr);
3286 * We aren't calling this function from any code path
3287 * that would already be holding a hash lock, so we're
3288 * asserting on this assumption to be defensive in case
3289 * this ever changes. Without this check, it would be
3290 * possible to incorrectly increment arcstat_mutex_miss
3291 * below (e.g. if the code changed such that we called
3292 * this function with a hash lock held).
3294 ASSERT(!MUTEX_HELD(hash_lock));
3296 if (mutex_tryenter(hash_lock)) {
3297 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3298 mutex_exit(hash_lock);
3300 bytes_evicted += evicted;
3303 * If evicted is zero, arc_evict_hdr() must have
3304 * decided to skip this header, don't increment
3305 * evict_count in this case.
3311 * If arc_size isn't overflowing, signal any
3312 * threads that might happen to be waiting.
3314 * For each header evicted, we wake up a single
3315 * thread. If we used cv_broadcast, we could
3316 * wake up "too many" threads causing arc_size
3317 * to significantly overflow arc_c; since
3318 * arc_get_data_buf() doesn't check for overflow
3319 * when it's woken up (it doesn't because it's
3320 * possible for the ARC to be overflowing while
3321 * full of un-evictable buffers, and the
3322 * function should proceed in this case).
3324 * If threads are left sleeping, due to not
3325 * using cv_broadcast, they will be woken up
3326 * just before arc_reclaim_thread() sleeps.
3328 mutex_enter(&arc_reclaim_lock);
3329 if (!arc_is_overflowing())
3330 cv_signal(&arc_reclaim_waiters_cv);
3331 mutex_exit(&arc_reclaim_lock);
3333 ARCSTAT_BUMP(arcstat_mutex_miss);
3337 multilist_sublist_unlock(mls);
3339 return (bytes_evicted);
3343 * Evict buffers from the given arc state, until we've removed the
3344 * specified number of bytes. Move the removed buffers to the
3345 * appropriate evict state.
3347 * This function makes a "best effort". It skips over any buffers
3348 * it can't get a hash_lock on, and so, may not catch all candidates.
3349 * It may also return without evicting as much space as requested.
3351 * If bytes is specified using the special value ARC_EVICT_ALL, this
3352 * will evict all available (i.e. unlocked and evictable) buffers from
3353 * the given arc state; which is used by arc_flush().
3356 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3357 arc_buf_contents_t type)
3359 uint64_t total_evicted = 0;
3360 multilist_t *ml = &state->arcs_list[type];
3362 arc_buf_hdr_t **markers;
3364 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3366 num_sublists = multilist_get_num_sublists(ml);
3369 * If we've tried to evict from each sublist, made some
3370 * progress, but still have not hit the target number of bytes
3371 * to evict, we want to keep trying. The markers allow us to
3372 * pick up where we left off for each individual sublist, rather
3373 * than starting from the tail each time.
3375 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3376 for (int i = 0; i < num_sublists; i++) {
3377 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3380 * A b_spa of 0 is used to indicate that this header is
3381 * a marker. This fact is used in arc_adjust_type() and
3382 * arc_evict_state_impl().
3384 markers[i]->b_spa = 0;
3386 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3387 multilist_sublist_insert_tail(mls, markers[i]);
3388 multilist_sublist_unlock(mls);
3392 * While we haven't hit our target number of bytes to evict, or
3393 * we're evicting all available buffers.
3395 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3397 * Start eviction using a randomly selected sublist,
3398 * this is to try and evenly balance eviction across all
3399 * sublists. Always starting at the same sublist
3400 * (e.g. index 0) would cause evictions to favor certain
3401 * sublists over others.
3403 int sublist_idx = multilist_get_random_index(ml);
3404 uint64_t scan_evicted = 0;
3406 for (int i = 0; i < num_sublists; i++) {
3407 uint64_t bytes_remaining;
3408 uint64_t bytes_evicted;
3410 if (bytes == ARC_EVICT_ALL)
3411 bytes_remaining = ARC_EVICT_ALL;
3412 else if (total_evicted < bytes)
3413 bytes_remaining = bytes - total_evicted;
3417 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3418 markers[sublist_idx], spa, bytes_remaining);
3420 scan_evicted += bytes_evicted;
3421 total_evicted += bytes_evicted;
3423 /* we've reached the end, wrap to the beginning */
3424 if (++sublist_idx >= num_sublists)
3429 * If we didn't evict anything during this scan, we have
3430 * no reason to believe we'll evict more during another
3431 * scan, so break the loop.
3433 if (scan_evicted == 0) {
3434 /* This isn't possible, let's make that obvious */
3435 ASSERT3S(bytes, !=, 0);
3438 * When bytes is ARC_EVICT_ALL, the only way to
3439 * break the loop is when scan_evicted is zero.
3440 * In that case, we actually have evicted enough,
3441 * so we don't want to increment the kstat.
3443 if (bytes != ARC_EVICT_ALL) {
3444 ASSERT3S(total_evicted, <, bytes);
3445 ARCSTAT_BUMP(arcstat_evict_not_enough);
3452 for (int i = 0; i < num_sublists; i++) {
3453 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3454 multilist_sublist_remove(mls, markers[i]);
3455 multilist_sublist_unlock(mls);
3457 kmem_cache_free(hdr_full_cache, markers[i]);
3459 kmem_free(markers, sizeof (*markers) * num_sublists);
3461 return (total_evicted);
3465 * Flush all "evictable" data of the given type from the arc state
3466 * specified. This will not evict any "active" buffers (i.e. referenced).
3468 * When 'retry' is set to B_FALSE, the function will make a single pass
3469 * over the state and evict any buffers that it can. Since it doesn't
3470 * continually retry the eviction, it might end up leaving some buffers
3471 * in the ARC due to lock misses.
3473 * When 'retry' is set to B_TRUE, the function will continually retry the
3474 * eviction until *all* evictable buffers have been removed from the
3475 * state. As a result, if concurrent insertions into the state are
3476 * allowed (e.g. if the ARC isn't shutting down), this function might
3477 * wind up in an infinite loop, continually trying to evict buffers.
3480 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
3483 uint64_t evicted = 0;
3485 while (refcount_count(&state->arcs_esize[type]) != 0) {
3486 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
3496 * Evict the specified number of bytes from the state specified,
3497 * restricting eviction to the spa and type given. This function
3498 * prevents us from trying to evict more from a state's list than
3499 * is "evictable", and to skip evicting altogether when passed a
3500 * negative value for "bytes". In contrast, arc_evict_state() will
3501 * evict everything it can, when passed a negative value for "bytes".
3504 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3505 arc_buf_contents_t type)
3509 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
3510 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
3511 return (arc_evict_state(state, spa, delta, type));
3518 * Evict metadata buffers from the cache, such that arc_meta_used is
3519 * capped by the arc_meta_limit tunable.
3522 arc_adjust_meta(void)
3524 uint64_t total_evicted = 0;
3528 * If we're over the meta limit, we want to evict enough
3529 * metadata to get back under the meta limit. We don't want to
3530 * evict so much that we drop the MRU below arc_p, though. If
3531 * we're over the meta limit more than we're over arc_p, we
3532 * evict some from the MRU here, and some from the MFU below.
3534 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3535 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3536 refcount_count(&arc_mru->arcs_size) - arc_p));
3538 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3541 * Similar to the above, we want to evict enough bytes to get us
3542 * below the meta limit, but not so much as to drop us below the
3543 * space alloted to the MFU (which is defined as arc_c - arc_p).
3545 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3546 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
3548 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3550 return (total_evicted);
3554 * Return the type of the oldest buffer in the given arc state
3556 * This function will select a random sublist of type ARC_BUFC_DATA and
3557 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3558 * is compared, and the type which contains the "older" buffer will be
3561 static arc_buf_contents_t
3562 arc_adjust_type(arc_state_t *state)
3564 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
3565 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
3566 int data_idx = multilist_get_random_index(data_ml);
3567 int meta_idx = multilist_get_random_index(meta_ml);
3568 multilist_sublist_t *data_mls;
3569 multilist_sublist_t *meta_mls;
3570 arc_buf_contents_t type;
3571 arc_buf_hdr_t *data_hdr;
3572 arc_buf_hdr_t *meta_hdr;
3575 * We keep the sublist lock until we're finished, to prevent
3576 * the headers from being destroyed via arc_evict_state().
3578 data_mls = multilist_sublist_lock(data_ml, data_idx);
3579 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
3582 * These two loops are to ensure we skip any markers that
3583 * might be at the tail of the lists due to arc_evict_state().
3586 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
3587 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
3588 if (data_hdr->b_spa != 0)
3592 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
3593 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
3594 if (meta_hdr->b_spa != 0)
3598 if (data_hdr == NULL && meta_hdr == NULL) {
3599 type = ARC_BUFC_DATA;
3600 } else if (data_hdr == NULL) {
3601 ASSERT3P(meta_hdr, !=, NULL);
3602 type = ARC_BUFC_METADATA;
3603 } else if (meta_hdr == NULL) {
3604 ASSERT3P(data_hdr, !=, NULL);
3605 type = ARC_BUFC_DATA;
3607 ASSERT3P(data_hdr, !=, NULL);
3608 ASSERT3P(meta_hdr, !=, NULL);
3610 /* The headers can't be on the sublist without an L1 header */
3611 ASSERT(HDR_HAS_L1HDR(data_hdr));
3612 ASSERT(HDR_HAS_L1HDR(meta_hdr));
3614 if (data_hdr->b_l1hdr.b_arc_access <
3615 meta_hdr->b_l1hdr.b_arc_access) {
3616 type = ARC_BUFC_DATA;
3618 type = ARC_BUFC_METADATA;
3622 multilist_sublist_unlock(meta_mls);
3623 multilist_sublist_unlock(data_mls);
3629 * Evict buffers from the cache, such that arc_size is capped by arc_c.
3634 uint64_t total_evicted = 0;
3639 * If we're over arc_meta_limit, we want to correct that before
3640 * potentially evicting data buffers below.
3642 total_evicted += arc_adjust_meta();
3647 * If we're over the target cache size, we want to evict enough
3648 * from the list to get back to our target size. We don't want
3649 * to evict too much from the MRU, such that it drops below
3650 * arc_p. So, if we're over our target cache size more than
3651 * the MRU is over arc_p, we'll evict enough to get back to
3652 * arc_p here, and then evict more from the MFU below.
3654 target = MIN((int64_t)(arc_size - arc_c),
3655 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3656 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
3659 * If we're below arc_meta_min, always prefer to evict data.
3660 * Otherwise, try to satisfy the requested number of bytes to
3661 * evict from the type which contains older buffers; in an
3662 * effort to keep newer buffers in the cache regardless of their
3663 * type. If we cannot satisfy the number of bytes from this
3664 * type, spill over into the next type.
3666 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
3667 arc_meta_used > arc_meta_min) {
3668 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3669 total_evicted += bytes;
3672 * If we couldn't evict our target number of bytes from
3673 * metadata, we try to get the rest from data.
3678 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3680 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3681 total_evicted += bytes;
3684 * If we couldn't evict our target number of bytes from
3685 * data, we try to get the rest from metadata.
3690 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3696 * Now that we've tried to evict enough from the MRU to get its
3697 * size back to arc_p, if we're still above the target cache
3698 * size, we evict the rest from the MFU.
3700 target = arc_size - arc_c;
3702 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
3703 arc_meta_used > arc_meta_min) {
3704 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3705 total_evicted += bytes;
3708 * If we couldn't evict our target number of bytes from
3709 * metadata, we try to get the rest from data.
3714 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3716 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3717 total_evicted += bytes;
3720 * If we couldn't evict our target number of bytes from
3721 * data, we try to get the rest from data.
3726 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3730 * Adjust ghost lists
3732 * In addition to the above, the ARC also defines target values
3733 * for the ghost lists. The sum of the mru list and mru ghost
3734 * list should never exceed the target size of the cache, and
3735 * the sum of the mru list, mfu list, mru ghost list, and mfu
3736 * ghost list should never exceed twice the target size of the
3737 * cache. The following logic enforces these limits on the ghost
3738 * caches, and evicts from them as needed.
3740 target = refcount_count(&arc_mru->arcs_size) +
3741 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3743 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3744 total_evicted += bytes;
3749 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3752 * We assume the sum of the mru list and mfu list is less than
3753 * or equal to arc_c (we enforced this above), which means we
3754 * can use the simpler of the two equations below:
3756 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3757 * mru ghost + mfu ghost <= arc_c
3759 target = refcount_count(&arc_mru_ghost->arcs_size) +
3760 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3762 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3763 total_evicted += bytes;
3768 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3770 return (total_evicted);
3774 arc_flush(spa_t *spa, boolean_t retry)
3779 * If retry is B_TRUE, a spa must not be specified since we have
3780 * no good way to determine if all of a spa's buffers have been
3781 * evicted from an arc state.
3783 ASSERT(!retry || spa == 0);
3786 guid = spa_load_guid(spa);
3788 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3789 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3791 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3792 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3794 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3795 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3797 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3798 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3802 arc_shrink(int64_t to_free)
3804 if (arc_c > arc_c_min) {
3805 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
3806 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
3807 if (arc_c > arc_c_min + to_free)
3808 atomic_add_64(&arc_c, -to_free);
3812 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3813 if (arc_c > arc_size)
3814 arc_c = MAX(arc_size, arc_c_min);
3816 arc_p = (arc_c >> 1);
3818 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
3821 ASSERT(arc_c >= arc_c_min);
3822 ASSERT((int64_t)arc_p >= 0);
3825 if (arc_size > arc_c) {
3826 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size,
3828 (void) arc_adjust();
3832 static long needfree = 0;
3834 typedef enum free_memory_reason_t {
3839 FMR_PAGES_PP_MAXIMUM,
3843 } free_memory_reason_t;
3845 int64_t last_free_memory;
3846 free_memory_reason_t last_free_reason;
3849 * Additional reserve of pages for pp_reserve.
3851 int64_t arc_pages_pp_reserve = 64;
3854 * Additional reserve of pages for swapfs.
3856 int64_t arc_swapfs_reserve = 64;
3859 * Return the amount of memory that can be consumed before reclaim will be
3860 * needed. Positive if there is sufficient free memory, negative indicates
3861 * the amount of memory that needs to be freed up.
3864 arc_available_memory(void)
3866 int64_t lowest = INT64_MAX;
3868 free_memory_reason_t r = FMR_UNKNOWN;
3872 n = PAGESIZE * (-needfree);
3880 * Cooperate with pagedaemon when it's time for it to scan
3881 * and reclaim some pages.
3883 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
3891 * check that we're out of range of the pageout scanner. It starts to
3892 * schedule paging if freemem is less than lotsfree and needfree.
3893 * lotsfree is the high-water mark for pageout, and needfree is the
3894 * number of needed free pages. We add extra pages here to make sure
3895 * the scanner doesn't start up while we're freeing memory.
3897 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
3904 * check to make sure that swapfs has enough space so that anon
3905 * reservations can still succeed. anon_resvmem() checks that the
3906 * availrmem is greater than swapfs_minfree, and the number of reserved
3907 * swap pages. We also add a bit of extra here just to prevent
3908 * circumstances from getting really dire.
3910 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
3911 desfree - arc_swapfs_reserve);
3914 r = FMR_SWAPFS_MINFREE;
3919 * Check that we have enough availrmem that memory locking (e.g., via
3920 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3921 * stores the number of pages that cannot be locked; when availrmem
3922 * drops below pages_pp_maximum, page locking mechanisms such as
3923 * page_pp_lock() will fail.)
3925 n = PAGESIZE * (availrmem - pages_pp_maximum -
3926 arc_pages_pp_reserve);
3929 r = FMR_PAGES_PP_MAXIMUM;
3932 #endif /* illumos */
3933 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
3935 * If we're on an i386 platform, it's possible that we'll exhaust the
3936 * kernel heap space before we ever run out of available physical
3937 * memory. Most checks of the size of the heap_area compare against
3938 * tune.t_minarmem, which is the minimum available real memory that we
3939 * can have in the system. However, this is generally fixed at 25 pages
3940 * which is so low that it's useless. In this comparison, we seek to
3941 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3942 * heap is allocated. (Or, in the calculation, if less than 1/4th is
3945 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
3946 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
3951 #define zio_arena NULL
3953 #define zio_arena heap_arena
3957 * If zio data pages are being allocated out of a separate heap segment,
3958 * then enforce that the size of available vmem for this arena remains
3959 * above about 1/16th free.
3961 * Note: The 1/16th arena free requirement was put in place
3962 * to aggressively evict memory from the arc in order to avoid
3963 * memory fragmentation issues.
3965 if (zio_arena != NULL) {
3966 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
3967 (vmem_size(zio_arena, VMEM_ALLOC) >> 4);
3975 * Above limits know nothing about real level of KVA fragmentation.
3976 * Start aggressive reclamation if too little sequential KVA left.
3979 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ?
3980 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) :
3989 /* Every 100 calls, free a small amount */
3990 if (spa_get_random(100) == 0)
3992 #endif /* _KERNEL */
3994 last_free_memory = lowest;
3995 last_free_reason = r;
3996 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
4002 * Determine if the system is under memory pressure and is asking
4003 * to reclaim memory. A return value of B_TRUE indicates that the system
4004 * is under memory pressure and that the arc should adjust accordingly.
4007 arc_reclaim_needed(void)
4009 return (arc_available_memory() < 0);
4012 extern kmem_cache_t *zio_buf_cache[];
4013 extern kmem_cache_t *zio_data_buf_cache[];
4014 extern kmem_cache_t *range_seg_cache;
4016 static __noinline void
4017 arc_kmem_reap_now(void)
4020 kmem_cache_t *prev_cache = NULL;
4021 kmem_cache_t *prev_data_cache = NULL;
4023 DTRACE_PROBE(arc__kmem_reap_start);
4025 if (arc_meta_used >= arc_meta_limit) {
4027 * We are exceeding our meta-data cache limit.
4028 * Purge some DNLC entries to release holds on meta-data.
4030 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4034 * Reclaim unused memory from all kmem caches.
4040 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4041 if (zio_buf_cache[i] != prev_cache) {
4042 prev_cache = zio_buf_cache[i];
4043 kmem_cache_reap_now(zio_buf_cache[i]);
4045 if (zio_data_buf_cache[i] != prev_data_cache) {
4046 prev_data_cache = zio_data_buf_cache[i];
4047 kmem_cache_reap_now(zio_data_buf_cache[i]);
4050 kmem_cache_reap_now(buf_cache);
4051 kmem_cache_reap_now(hdr_full_cache);
4052 kmem_cache_reap_now(hdr_l2only_cache);
4053 kmem_cache_reap_now(range_seg_cache);
4056 if (zio_arena != NULL) {
4058 * Ask the vmem arena to reclaim unused memory from its
4061 vmem_qcache_reap(zio_arena);
4064 DTRACE_PROBE(arc__kmem_reap_end);
4068 * Threads can block in arc_get_data_buf() waiting for this thread to evict
4069 * enough data and signal them to proceed. When this happens, the threads in
4070 * arc_get_data_buf() are sleeping while holding the hash lock for their
4071 * particular arc header. Thus, we must be careful to never sleep on a
4072 * hash lock in this thread. This is to prevent the following deadlock:
4074 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
4075 * waiting for the reclaim thread to signal it.
4077 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4078 * fails, and goes to sleep forever.
4080 * This possible deadlock is avoided by always acquiring a hash lock
4081 * using mutex_tryenter() from arc_reclaim_thread().
4084 arc_reclaim_thread(void *dummy __unused)
4086 hrtime_t growtime = 0;
4089 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4091 mutex_enter(&arc_reclaim_lock);
4092 while (!arc_reclaim_thread_exit) {
4093 uint64_t evicted = 0;
4096 * This is necessary in order for the mdb ::arc dcmd to
4097 * show up to date information. Since the ::arc command
4098 * does not call the kstat's update function, without
4099 * this call, the command may show stale stats for the
4100 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4101 * with this change, the data might be up to 1 second
4102 * out of date; but that should suffice. The arc_state_t
4103 * structures can be queried directly if more accurate
4104 * information is needed.
4106 if (arc_ksp != NULL)
4107 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4109 mutex_exit(&arc_reclaim_lock);
4112 * We call arc_adjust() before (possibly) calling
4113 * arc_kmem_reap_now(), so that we can wake up
4114 * arc_get_data_buf() sooner.
4116 evicted = arc_adjust();
4118 int64_t free_memory = arc_available_memory();
4119 if (free_memory < 0) {
4121 arc_no_grow = B_TRUE;
4125 * Wait at least zfs_grow_retry (default 60) seconds
4126 * before considering growing.
4128 growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4130 arc_kmem_reap_now();
4133 * If we are still low on memory, shrink the ARC
4134 * so that we have arc_shrink_min free space.
4136 free_memory = arc_available_memory();
4139 (arc_c >> arc_shrink_shift) - free_memory;
4142 to_free = MAX(to_free, ptob(needfree));
4144 arc_shrink(to_free);
4146 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4147 arc_no_grow = B_TRUE;
4148 } else if (gethrtime() >= growtime) {
4149 arc_no_grow = B_FALSE;
4152 mutex_enter(&arc_reclaim_lock);
4155 * If evicted is zero, we couldn't evict anything via
4156 * arc_adjust(). This could be due to hash lock
4157 * collisions, but more likely due to the majority of
4158 * arc buffers being unevictable. Therefore, even if
4159 * arc_size is above arc_c, another pass is unlikely to
4160 * be helpful and could potentially cause us to enter an
4163 if (arc_size <= arc_c || evicted == 0) {
4168 * We're either no longer overflowing, or we
4169 * can't evict anything more, so we should wake
4170 * up any threads before we go to sleep.
4172 cv_broadcast(&arc_reclaim_waiters_cv);
4175 * Block until signaled, or after one second (we
4176 * might need to perform arc_kmem_reap_now()
4177 * even if we aren't being signalled)
4179 CALLB_CPR_SAFE_BEGIN(&cpr);
4180 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4181 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4182 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4186 arc_reclaim_thread_exit = B_FALSE;
4187 cv_broadcast(&arc_reclaim_thread_cv);
4188 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
4193 * Adapt arc info given the number of bytes we are trying to add and
4194 * the state that we are comming from. This function is only called
4195 * when we are adding new content to the cache.
4198 arc_adapt(int bytes, arc_state_t *state)
4201 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4202 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
4203 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
4205 if (state == arc_l2c_only)
4210 * Adapt the target size of the MRU list:
4211 * - if we just hit in the MRU ghost list, then increase
4212 * the target size of the MRU list.
4213 * - if we just hit in the MFU ghost list, then increase
4214 * the target size of the MFU list by decreasing the
4215 * target size of the MRU list.
4217 if (state == arc_mru_ghost) {
4218 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4219 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4221 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4222 } else if (state == arc_mfu_ghost) {
4225 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4226 mult = MIN(mult, 10);
4228 delta = MIN(bytes * mult, arc_p);
4229 arc_p = MAX(arc_p_min, arc_p - delta);
4231 ASSERT((int64_t)arc_p >= 0);
4233 if (arc_reclaim_needed()) {
4234 cv_signal(&arc_reclaim_thread_cv);
4241 if (arc_c >= arc_c_max)
4245 * If we're within (2 * maxblocksize) bytes of the target
4246 * cache size, increment the target cache size
4248 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
4249 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
4250 atomic_add_64(&arc_c, (int64_t)bytes);
4251 if (arc_c > arc_c_max)
4253 else if (state == arc_anon)
4254 atomic_add_64(&arc_p, (int64_t)bytes);
4258 ASSERT((int64_t)arc_p >= 0);
4262 * Check if arc_size has grown past our upper threshold, determined by
4263 * zfs_arc_overflow_shift.
4266 arc_is_overflowing(void)
4268 /* Always allow at least one block of overflow */
4269 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4270 arc_c >> zfs_arc_overflow_shift);
4272 return (arc_size >= arc_c + overflow);
4276 * Allocate a block and return it to the caller. If we are hitting the
4277 * hard limit for the cache size, we must sleep, waiting for the eviction
4278 * thread to catch up. If we're past the target size but below the hard
4279 * limit, we'll only signal the reclaim thread and continue on.
4282 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4285 arc_state_t *state = hdr->b_l1hdr.b_state;
4286 arc_buf_contents_t type = arc_buf_type(hdr);
4288 arc_adapt(size, state);
4291 * If arc_size is currently overflowing, and has grown past our
4292 * upper limit, we must be adding data faster than the evict
4293 * thread can evict. Thus, to ensure we don't compound the
4294 * problem by adding more data and forcing arc_size to grow even
4295 * further past it's target size, we halt and wait for the
4296 * eviction thread to catch up.
4298 * It's also possible that the reclaim thread is unable to evict
4299 * enough buffers to get arc_size below the overflow limit (e.g.
4300 * due to buffers being un-evictable, or hash lock collisions).
4301 * In this case, we want to proceed regardless if we're
4302 * overflowing; thus we don't use a while loop here.
4304 if (arc_is_overflowing()) {
4305 mutex_enter(&arc_reclaim_lock);
4308 * Now that we've acquired the lock, we may no longer be
4309 * over the overflow limit, lets check.
4311 * We're ignoring the case of spurious wake ups. If that
4312 * were to happen, it'd let this thread consume an ARC
4313 * buffer before it should have (i.e. before we're under
4314 * the overflow limit and were signalled by the reclaim
4315 * thread). As long as that is a rare occurrence, it
4316 * shouldn't cause any harm.
4318 if (arc_is_overflowing()) {
4319 cv_signal(&arc_reclaim_thread_cv);
4320 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
4323 mutex_exit(&arc_reclaim_lock);
4326 VERIFY3U(hdr->b_type, ==, type);
4327 if (type == ARC_BUFC_METADATA) {
4328 datap = zio_buf_alloc(size);
4329 arc_space_consume(size, ARC_SPACE_META);
4331 ASSERT(type == ARC_BUFC_DATA);
4332 datap = zio_data_buf_alloc(size);
4333 arc_space_consume(size, ARC_SPACE_DATA);
4337 * Update the state size. Note that ghost states have a
4338 * "ghost size" and so don't need to be updated.
4340 if (!GHOST_STATE(state)) {
4342 (void) refcount_add_many(&state->arcs_size, size, tag);
4345 * If this is reached via arc_read, the link is
4346 * protected by the hash lock. If reached via
4347 * arc_buf_alloc, the header should not be accessed by
4348 * any other thread. And, if reached via arc_read_done,
4349 * the hash lock will protect it if it's found in the
4350 * hash table; otherwise no other thread should be
4351 * trying to [add|remove]_reference it.
4353 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4354 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4355 (void) refcount_add_many(&state->arcs_esize[type],
4360 * If we are growing the cache, and we are adding anonymous
4361 * data, and we have outgrown arc_p, update arc_p
4363 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
4364 (refcount_count(&arc_anon->arcs_size) +
4365 refcount_count(&arc_mru->arcs_size) > arc_p))
4366 arc_p = MIN(arc_c, arc_p + size);
4368 ARCSTAT_BUMP(arcstat_allocated);
4373 * Free the arc data buffer.
4376 arc_free_data_buf(arc_buf_hdr_t *hdr, void *data, uint64_t size, void *tag)
4378 arc_state_t *state = hdr->b_l1hdr.b_state;
4379 arc_buf_contents_t type = arc_buf_type(hdr);
4381 /* protected by hash lock, if in the hash table */
4382 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4383 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4384 ASSERT(state != arc_anon && state != arc_l2c_only);
4386 (void) refcount_remove_many(&state->arcs_esize[type],
4389 (void) refcount_remove_many(&state->arcs_size, size, tag);
4391 VERIFY3U(hdr->b_type, ==, type);
4392 if (type == ARC_BUFC_METADATA) {
4393 zio_buf_free(data, size);
4394 arc_space_return(size, ARC_SPACE_META);
4396 ASSERT(type == ARC_BUFC_DATA);
4397 zio_data_buf_free(data, size);
4398 arc_space_return(size, ARC_SPACE_DATA);
4403 * This routine is called whenever a buffer is accessed.
4404 * NOTE: the hash lock is dropped in this function.
4407 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4411 ASSERT(MUTEX_HELD(hash_lock));
4412 ASSERT(HDR_HAS_L1HDR(hdr));
4414 if (hdr->b_l1hdr.b_state == arc_anon) {
4416 * This buffer is not in the cache, and does not
4417 * appear in our "ghost" list. Add the new buffer
4421 ASSERT0(hdr->b_l1hdr.b_arc_access);
4422 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4423 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4424 arc_change_state(arc_mru, hdr, hash_lock);
4426 } else if (hdr->b_l1hdr.b_state == arc_mru) {
4427 now = ddi_get_lbolt();
4430 * If this buffer is here because of a prefetch, then either:
4431 * - clear the flag if this is a "referencing" read
4432 * (any subsequent access will bump this into the MFU state).
4434 * - move the buffer to the head of the list if this is
4435 * another prefetch (to make it less likely to be evicted).
4437 if (HDR_PREFETCH(hdr)) {
4438 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4439 /* link protected by hash lock */
4440 ASSERT(multilist_link_active(
4441 &hdr->b_l1hdr.b_arc_node));
4443 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4444 ARCSTAT_BUMP(arcstat_mru_hits);
4446 hdr->b_l1hdr.b_arc_access = now;
4451 * This buffer has been "accessed" only once so far,
4452 * but it is still in the cache. Move it to the MFU
4455 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
4457 * More than 125ms have passed since we
4458 * instantiated this buffer. Move it to the
4459 * most frequently used state.
4461 hdr->b_l1hdr.b_arc_access = now;
4462 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4463 arc_change_state(arc_mfu, hdr, hash_lock);
4465 ARCSTAT_BUMP(arcstat_mru_hits);
4466 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
4467 arc_state_t *new_state;
4469 * This buffer has been "accessed" recently, but
4470 * was evicted from the cache. Move it to the
4474 if (HDR_PREFETCH(hdr)) {
4475 new_state = arc_mru;
4476 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
4477 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4478 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4480 new_state = arc_mfu;
4481 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4484 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4485 arc_change_state(new_state, hdr, hash_lock);
4487 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
4488 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
4490 * This buffer has been accessed more than once and is
4491 * still in the cache. Keep it in the MFU state.
4493 * NOTE: an add_reference() that occurred when we did
4494 * the arc_read() will have kicked this off the list.
4495 * If it was a prefetch, we will explicitly move it to
4496 * the head of the list now.
4498 if ((HDR_PREFETCH(hdr)) != 0) {
4499 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4500 /* link protected by hash_lock */
4501 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4503 ARCSTAT_BUMP(arcstat_mfu_hits);
4504 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4505 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
4506 arc_state_t *new_state = arc_mfu;
4508 * This buffer has been accessed more than once but has
4509 * been evicted from the cache. Move it back to the
4513 if (HDR_PREFETCH(hdr)) {
4515 * This is a prefetch access...
4516 * move this block back to the MRU state.
4518 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
4519 new_state = arc_mru;
4522 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4523 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4524 arc_change_state(new_state, hdr, hash_lock);
4526 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
4527 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
4529 * This buffer is on the 2nd Level ARC.
4532 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4533 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4534 arc_change_state(arc_mfu, hdr, hash_lock);
4536 ASSERT(!"invalid arc state");
4540 /* a generic arc_done_func_t which you can use */
4543 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
4545 if (zio == NULL || zio->io_error == 0)
4546 bcopy(buf->b_data, arg, HDR_GET_LSIZE(buf->b_hdr));
4547 arc_buf_destroy(buf, arg);
4550 /* a generic arc_done_func_t */
4552 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
4554 arc_buf_t **bufp = arg;
4555 if (zio && zio->io_error) {
4556 arc_buf_destroy(buf, arg);
4560 ASSERT(buf->b_data);
4565 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
4567 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
4568 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
4569 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
4571 if (HDR_COMPRESSION_ENABLED(hdr)) {
4572 ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
4573 BP_GET_COMPRESS(bp));
4575 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
4576 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
4581 arc_read_done(zio_t *zio)
4583 arc_buf_hdr_t *hdr = zio->io_private;
4584 arc_buf_t *abuf = NULL; /* buffer we're assigning to callback */
4585 kmutex_t *hash_lock = NULL;
4586 arc_callback_t *callback_list, *acb;
4587 int freeable = B_FALSE;
4590 * The hdr was inserted into hash-table and removed from lists
4591 * prior to starting I/O. We should find this header, since
4592 * it's in the hash table, and it should be legit since it's
4593 * not possible to evict it during the I/O. The only possible
4594 * reason for it not to be found is if we were freed during the
4597 if (HDR_IN_HASH_TABLE(hdr)) {
4598 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
4599 ASSERT3U(hdr->b_dva.dva_word[0], ==,
4600 BP_IDENTITY(zio->io_bp)->dva_word[0]);
4601 ASSERT3U(hdr->b_dva.dva_word[1], ==,
4602 BP_IDENTITY(zio->io_bp)->dva_word[1]);
4604 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
4607 ASSERT((found == hdr &&
4608 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
4609 (found == hdr && HDR_L2_READING(hdr)));
4610 ASSERT3P(hash_lock, !=, NULL);
4613 if (zio->io_error == 0) {
4614 /* byteswap if necessary */
4615 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
4616 if (BP_GET_LEVEL(zio->io_bp) > 0) {
4617 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
4619 hdr->b_l1hdr.b_byteswap =
4620 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
4623 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
4627 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
4628 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
4629 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
4631 callback_list = hdr->b_l1hdr.b_acb;
4632 ASSERT3P(callback_list, !=, NULL);
4634 if (hash_lock && zio->io_error == 0 &&
4635 hdr->b_l1hdr.b_state == arc_anon) {
4637 * Only call arc_access on anonymous buffers. This is because
4638 * if we've issued an I/O for an evicted buffer, we've already
4639 * called arc_access (to prevent any simultaneous readers from
4640 * getting confused).
4642 arc_access(hdr, hash_lock);
4645 /* create copies of the data buffer for the callers */
4646 for (acb = callback_list; acb; acb = acb->acb_next) {
4647 if (acb->acb_done != NULL) {
4649 * If we're here, then this must be a demand read
4650 * since prefetch requests don't have callbacks.
4651 * If a read request has a callback (i.e. acb_done is
4652 * not NULL), then we decompress the data for the
4653 * first request and clone the rest. This avoids
4654 * having to waste cpu resources decompressing data
4655 * that nobody is explicitly waiting to read.
4658 acb->acb_buf = arc_buf_alloc_impl(hdr,
4660 if (zio->io_error == 0) {
4662 arc_decompress(acb->acb_buf);
4664 abuf = acb->acb_buf;
4666 add_reference(hdr, acb->acb_private);
4667 acb->acb_buf = arc_buf_clone(abuf);
4671 hdr->b_l1hdr.b_acb = NULL;
4672 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
4675 * This buffer didn't have a callback so it must
4678 ASSERT(HDR_PREFETCH(hdr));
4679 ASSERT0(hdr->b_l1hdr.b_bufcnt);
4680 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
4683 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
4684 callback_list != NULL);
4686 if (zio->io_error == 0) {
4687 arc_hdr_verify(hdr, zio->io_bp);
4689 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
4690 if (hdr->b_l1hdr.b_state != arc_anon)
4691 arc_change_state(arc_anon, hdr, hash_lock);
4692 if (HDR_IN_HASH_TABLE(hdr))
4693 buf_hash_remove(hdr);
4694 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4698 * Broadcast before we drop the hash_lock to avoid the possibility
4699 * that the hdr (and hence the cv) might be freed before we get to
4700 * the cv_broadcast().
4702 cv_broadcast(&hdr->b_l1hdr.b_cv);
4704 if (hash_lock != NULL) {
4705 mutex_exit(hash_lock);
4708 * This block was freed while we waited for the read to
4709 * complete. It has been removed from the hash table and
4710 * moved to the anonymous state (so that it won't show up
4713 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
4714 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4717 /* execute each callback and free its structure */
4718 while ((acb = callback_list) != NULL) {
4720 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
4722 if (acb->acb_zio_dummy != NULL) {
4723 acb->acb_zio_dummy->io_error = zio->io_error;
4724 zio_nowait(acb->acb_zio_dummy);
4727 callback_list = acb->acb_next;
4728 kmem_free(acb, sizeof (arc_callback_t));
4732 arc_hdr_destroy(hdr);
4736 * "Read" the block at the specified DVA (in bp) via the
4737 * cache. If the block is found in the cache, invoke the provided
4738 * callback immediately and return. Note that the `zio' parameter
4739 * in the callback will be NULL in this case, since no IO was
4740 * required. If the block is not in the cache pass the read request
4741 * on to the spa with a substitute callback function, so that the
4742 * requested block will be added to the cache.
4744 * If a read request arrives for a block that has a read in-progress,
4745 * either wait for the in-progress read to complete (and return the
4746 * results); or, if this is a read with a "done" func, add a record
4747 * to the read to invoke the "done" func when the read completes,
4748 * and return; or just return.
4750 * arc_read_done() will invoke all the requested "done" functions
4751 * for readers of this block.
4754 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
4755 void *private, zio_priority_t priority, int zio_flags,
4756 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
4758 arc_buf_hdr_t *hdr = NULL;
4759 kmutex_t *hash_lock = NULL;
4761 uint64_t guid = spa_load_guid(spa);
4763 ASSERT(!BP_IS_EMBEDDED(bp) ||
4764 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
4767 if (!BP_IS_EMBEDDED(bp)) {
4769 * Embedded BP's have no DVA and require no I/O to "read".
4770 * Create an anonymous arc buf to back it.
4772 hdr = buf_hash_find(guid, bp, &hash_lock);
4775 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pdata != NULL) {
4776 arc_buf_t *buf = NULL;
4777 *arc_flags |= ARC_FLAG_CACHED;
4779 if (HDR_IO_IN_PROGRESS(hdr)) {
4781 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
4782 priority == ZIO_PRIORITY_SYNC_READ) {
4784 * This sync read must wait for an
4785 * in-progress async read (e.g. a predictive
4786 * prefetch). Async reads are queued
4787 * separately at the vdev_queue layer, so
4788 * this is a form of priority inversion.
4789 * Ideally, we would "inherit" the demand
4790 * i/o's priority by moving the i/o from
4791 * the async queue to the synchronous queue,
4792 * but there is currently no mechanism to do
4793 * so. Track this so that we can evaluate
4794 * the magnitude of this potential performance
4797 * Note that if the prefetch i/o is already
4798 * active (has been issued to the device),
4799 * the prefetch improved performance, because
4800 * we issued it sooner than we would have
4801 * without the prefetch.
4803 DTRACE_PROBE1(arc__sync__wait__for__async,
4804 arc_buf_hdr_t *, hdr);
4805 ARCSTAT_BUMP(arcstat_sync_wait_for_async);
4807 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4808 arc_hdr_clear_flags(hdr,
4809 ARC_FLAG_PREDICTIVE_PREFETCH);
4812 if (*arc_flags & ARC_FLAG_WAIT) {
4813 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
4814 mutex_exit(hash_lock);
4817 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4820 arc_callback_t *acb = NULL;
4822 acb = kmem_zalloc(sizeof (arc_callback_t),
4824 acb->acb_done = done;
4825 acb->acb_private = private;
4827 acb->acb_zio_dummy = zio_null(pio,
4828 spa, NULL, NULL, NULL, zio_flags);
4830 ASSERT3P(acb->acb_done, !=, NULL);
4831 acb->acb_next = hdr->b_l1hdr.b_acb;
4832 hdr->b_l1hdr.b_acb = acb;
4833 mutex_exit(hash_lock);
4836 mutex_exit(hash_lock);
4840 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4841 hdr->b_l1hdr.b_state == arc_mfu);
4844 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4846 * This is a demand read which does not have to
4847 * wait for i/o because we did a predictive
4848 * prefetch i/o for it, which has completed.
4851 arc__demand__hit__predictive__prefetch,
4852 arc_buf_hdr_t *, hdr);
4854 arcstat_demand_hit_predictive_prefetch);
4855 arc_hdr_clear_flags(hdr,
4856 ARC_FLAG_PREDICTIVE_PREFETCH);
4858 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
4861 * If this block is already in use, create a new
4862 * copy of the data so that we will be guaranteed
4863 * that arc_release() will always succeed.
4865 buf = hdr->b_l1hdr.b_buf;
4867 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
4868 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
4869 buf = arc_buf_alloc_impl(hdr, private);
4870 VERIFY0(arc_decompress(buf));
4872 add_reference(hdr, private);
4873 buf = arc_buf_clone(buf);
4875 ASSERT3P(buf->b_data, !=, NULL);
4877 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
4878 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4879 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
4881 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4882 arc_access(hdr, hash_lock);
4883 if (*arc_flags & ARC_FLAG_L2CACHE)
4884 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
4885 mutex_exit(hash_lock);
4886 ARCSTAT_BUMP(arcstat_hits);
4887 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4888 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4889 data, metadata, hits);
4892 done(NULL, buf, private);
4894 uint64_t lsize = BP_GET_LSIZE(bp);
4895 uint64_t psize = BP_GET_PSIZE(bp);
4896 arc_callback_t *acb;
4899 boolean_t devw = B_FALSE;
4903 /* this block is not in the cache */
4904 arc_buf_hdr_t *exists = NULL;
4905 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
4906 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
4907 BP_GET_COMPRESS(bp), type);
4909 if (!BP_IS_EMBEDDED(bp)) {
4910 hdr->b_dva = *BP_IDENTITY(bp);
4911 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
4912 exists = buf_hash_insert(hdr, &hash_lock);
4914 if (exists != NULL) {
4915 /* somebody beat us to the hash insert */
4916 mutex_exit(hash_lock);
4917 buf_discard_identity(hdr);
4918 arc_hdr_destroy(hdr);
4919 goto top; /* restart the IO request */
4923 * This block is in the ghost cache. If it was L2-only
4924 * (and thus didn't have an L1 hdr), we realloc the
4925 * header to add an L1 hdr.
4927 if (!HDR_HAS_L1HDR(hdr)) {
4928 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
4931 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
4932 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
4933 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4934 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4935 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
4938 * This is a delicate dance that we play here.
4939 * This hdr is in the ghost list so we access it
4940 * to move it out of the ghost list before we
4941 * initiate the read. If it's a prefetch then
4942 * it won't have a callback so we'll remove the
4943 * reference that arc_buf_alloc_impl() created. We
4944 * do this after we've called arc_access() to
4945 * avoid hitting an assert in remove_reference().
4947 arc_access(hdr, hash_lock);
4948 arc_hdr_alloc_pdata(hdr);
4950 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
4951 size = arc_hdr_size(hdr);
4954 * If compression is enabled on the hdr, then will do
4955 * RAW I/O and will store the compressed data in the hdr's
4956 * data block. Otherwise, the hdr's data block will contain
4957 * the uncompressed data.
4959 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
4960 zio_flags |= ZIO_FLAG_RAW;
4963 if (*arc_flags & ARC_FLAG_PREFETCH)
4964 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
4965 if (*arc_flags & ARC_FLAG_L2CACHE)
4966 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
4967 if (BP_GET_LEVEL(bp) > 0)
4968 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
4969 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
4970 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
4971 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
4973 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
4974 acb->acb_done = done;
4975 acb->acb_private = private;
4977 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
4978 hdr->b_l1hdr.b_acb = acb;
4979 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
4981 if (HDR_HAS_L2HDR(hdr) &&
4982 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
4983 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
4984 addr = hdr->b_l2hdr.b_daddr;
4986 * Lock out device removal.
4988 if (vdev_is_dead(vd) ||
4989 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
4993 if (priority == ZIO_PRIORITY_ASYNC_READ)
4994 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
4996 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
4998 if (hash_lock != NULL)
4999 mutex_exit(hash_lock);
5002 * At this point, we have a level 1 cache miss. Try again in
5003 * L2ARC if possible.
5005 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5007 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5008 uint64_t, lsize, zbookmark_phys_t *, zb);
5009 ARCSTAT_BUMP(arcstat_misses);
5010 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5011 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5012 data, metadata, misses);
5014 curthread->td_ru.ru_inblock++;
5017 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5019 * Read from the L2ARC if the following are true:
5020 * 1. The L2ARC vdev was previously cached.
5021 * 2. This buffer still has L2ARC metadata.
5022 * 3. This buffer isn't currently writing to the L2ARC.
5023 * 4. The L2ARC entry wasn't evicted, which may
5024 * also have invalidated the vdev.
5025 * 5. This isn't prefetch and l2arc_noprefetch is set.
5027 if (HDR_HAS_L2HDR(hdr) &&
5028 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5029 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5030 l2arc_read_callback_t *cb;
5033 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5034 ARCSTAT_BUMP(arcstat_l2_hits);
5036 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5038 cb->l2rcb_hdr = hdr;
5041 cb->l2rcb_flags = zio_flags;
5042 uint64_t asize = vdev_psize_to_asize(vd, size);
5043 if (asize != size) {
5044 b_data = zio_data_buf_alloc(asize);
5045 cb->l2rcb_data = b_data;
5047 b_data = hdr->b_l1hdr.b_pdata;
5050 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5051 addr + asize < vd->vdev_psize -
5052 VDEV_LABEL_END_SIZE);
5055 * l2arc read. The SCL_L2ARC lock will be
5056 * released by l2arc_read_done().
5057 * Issue a null zio if the underlying buffer
5058 * was squashed to zero size by compression.
5060 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
5061 ZIO_COMPRESS_EMPTY);
5062 rzio = zio_read_phys(pio, vd, addr,
5065 l2arc_read_done, cb, priority,
5066 zio_flags | ZIO_FLAG_DONT_CACHE |
5068 ZIO_FLAG_DONT_PROPAGATE |
5069 ZIO_FLAG_DONT_RETRY, B_FALSE);
5070 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5072 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
5074 if (*arc_flags & ARC_FLAG_NOWAIT) {
5079 ASSERT(*arc_flags & ARC_FLAG_WAIT);
5080 if (zio_wait(rzio) == 0)
5083 /* l2arc read error; goto zio_read() */
5085 DTRACE_PROBE1(l2arc__miss,
5086 arc_buf_hdr_t *, hdr);
5087 ARCSTAT_BUMP(arcstat_l2_misses);
5088 if (HDR_L2_WRITING(hdr))
5089 ARCSTAT_BUMP(arcstat_l2_rw_clash);
5090 spa_config_exit(spa, SCL_L2ARC, vd);
5094 spa_config_exit(spa, SCL_L2ARC, vd);
5095 if (l2arc_ndev != 0) {
5096 DTRACE_PROBE1(l2arc__miss,
5097 arc_buf_hdr_t *, hdr);
5098 ARCSTAT_BUMP(arcstat_l2_misses);
5102 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pdata, size,
5103 arc_read_done, hdr, priority, zio_flags, zb);
5105 if (*arc_flags & ARC_FLAG_WAIT)
5106 return (zio_wait(rzio));
5108 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5115 * Notify the arc that a block was freed, and thus will never be used again.
5118 arc_freed(spa_t *spa, const blkptr_t *bp)
5121 kmutex_t *hash_lock;
5122 uint64_t guid = spa_load_guid(spa);
5124 ASSERT(!BP_IS_EMBEDDED(bp));
5126 hdr = buf_hash_find(guid, bp, &hash_lock);
5131 * We might be trying to free a block that is still doing I/O
5132 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5133 * dmu_sync-ed block). If this block is being prefetched, then it
5134 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5135 * until the I/O completes. A block may also have a reference if it is
5136 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5137 * have written the new block to its final resting place on disk but
5138 * without the dedup flag set. This would have left the hdr in the MRU
5139 * state and discoverable. When the txg finally syncs it detects that
5140 * the block was overridden in open context and issues an override I/O.
5141 * Since this is a dedup block, the override I/O will determine if the
5142 * block is already in the DDT. If so, then it will replace the io_bp
5143 * with the bp from the DDT and allow the I/O to finish. When the I/O
5144 * reaches the done callback, dbuf_write_override_done, it will
5145 * check to see if the io_bp and io_bp_override are identical.
5146 * If they are not, then it indicates that the bp was replaced with
5147 * the bp in the DDT and the override bp is freed. This allows
5148 * us to arrive here with a reference on a block that is being
5149 * freed. So if we have an I/O in progress, or a reference to
5150 * this hdr, then we don't destroy the hdr.
5152 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
5153 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
5154 arc_change_state(arc_anon, hdr, hash_lock);
5155 arc_hdr_destroy(hdr);
5156 mutex_exit(hash_lock);
5158 mutex_exit(hash_lock);
5164 * Release this buffer from the cache, making it an anonymous buffer. This
5165 * must be done after a read and prior to modifying the buffer contents.
5166 * If the buffer has more than one reference, we must make
5167 * a new hdr for the buffer.
5170 arc_release(arc_buf_t *buf, void *tag)
5172 arc_buf_hdr_t *hdr = buf->b_hdr;
5175 * It would be nice to assert that if it's DMU metadata (level >
5176 * 0 || it's the dnode file), then it must be syncing context.
5177 * But we don't know that information at this level.
5180 mutex_enter(&buf->b_evict_lock);
5182 ASSERT(HDR_HAS_L1HDR(hdr));
5185 * We don't grab the hash lock prior to this check, because if
5186 * the buffer's header is in the arc_anon state, it won't be
5187 * linked into the hash table.
5189 if (hdr->b_l1hdr.b_state == arc_anon) {
5190 mutex_exit(&buf->b_evict_lock);
5191 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5192 ASSERT(!HDR_IN_HASH_TABLE(hdr));
5193 ASSERT(!HDR_HAS_L2HDR(hdr));
5194 ASSERT(HDR_EMPTY(hdr));
5195 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5196 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
5197 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
5199 hdr->b_l1hdr.b_arc_access = 0;
5202 * If the buf is being overridden then it may already
5203 * have a hdr that is not empty.
5205 buf_discard_identity(hdr);
5211 kmutex_t *hash_lock = HDR_LOCK(hdr);
5212 mutex_enter(hash_lock);
5215 * This assignment is only valid as long as the hash_lock is
5216 * held, we must be careful not to reference state or the
5217 * b_state field after dropping the lock.
5219 arc_state_t *state = hdr->b_l1hdr.b_state;
5220 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5221 ASSERT3P(state, !=, arc_anon);
5223 /* this buffer is not on any list */
5224 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0);
5226 if (HDR_HAS_L2HDR(hdr)) {
5227 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5230 * We have to recheck this conditional again now that
5231 * we're holding the l2ad_mtx to prevent a race with
5232 * another thread which might be concurrently calling
5233 * l2arc_evict(). In that case, l2arc_evict() might have
5234 * destroyed the header's L2 portion as we were waiting
5235 * to acquire the l2ad_mtx.
5237 if (HDR_HAS_L2HDR(hdr)) {
5239 arc_hdr_l2hdr_destroy(hdr);
5242 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5246 * Do we have more than one buf?
5248 if (hdr->b_l1hdr.b_bufcnt > 1) {
5249 arc_buf_hdr_t *nhdr;
5251 uint64_t spa = hdr->b_spa;
5252 uint64_t psize = HDR_GET_PSIZE(hdr);
5253 uint64_t lsize = HDR_GET_LSIZE(hdr);
5254 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
5255 arc_buf_contents_t type = arc_buf_type(hdr);
5256 VERIFY3U(hdr->b_type, ==, type);
5258 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
5259 (void) remove_reference(hdr, hash_lock, tag);
5261 if (arc_buf_is_shared(buf)) {
5262 ASSERT(HDR_SHARED_DATA(hdr));
5263 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5264 ASSERT(ARC_BUF_LAST(buf));
5268 * Pull the data off of this hdr and attach it to
5269 * a new anonymous hdr. Also find the last buffer
5270 * in the hdr's buffer list.
5272 arc_buf_t *lastbuf = NULL;
5273 bufp = &hdr->b_l1hdr.b_buf;
5274 while (*bufp != NULL) {
5276 *bufp = buf->b_next;
5280 * If we've removed a buffer in the middle of
5281 * the list then update the lastbuf and update
5284 if (*bufp != NULL) {
5286 bufp = &(*bufp)->b_next;
5290 ASSERT3P(lastbuf, !=, buf);
5291 ASSERT3P(lastbuf, !=, NULL);
5294 * If the current arc_buf_t and the hdr are sharing their data
5295 * buffer, then we must stop sharing that block, transfer
5296 * ownership and setup sharing with a new arc_buf_t at the end
5297 * of the hdr's b_buf list.
5299 if (arc_buf_is_shared(buf)) {
5300 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5301 ASSERT(ARC_BUF_LAST(lastbuf));
5302 VERIFY(!arc_buf_is_shared(lastbuf));
5305 * First, sever the block sharing relationship between
5306 * buf and the arc_buf_hdr_t. Then, setup a new
5307 * block sharing relationship with the last buffer
5308 * on the arc_buf_t list.
5310 arc_unshare_buf(hdr, buf);
5311 arc_share_buf(hdr, lastbuf);
5312 VERIFY3P(lastbuf->b_data, !=, NULL);
5313 } else if (HDR_SHARED_DATA(hdr)) {
5314 ASSERT(arc_buf_is_shared(lastbuf));
5316 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
5317 ASSERT3P(state, !=, arc_l2c_only);
5319 (void) refcount_remove_many(&state->arcs_size,
5320 HDR_GET_LSIZE(hdr), buf);
5322 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5323 ASSERT3P(state, !=, arc_l2c_only);
5324 (void) refcount_remove_many(&state->arcs_esize[type],
5325 HDR_GET_LSIZE(hdr), buf);
5328 hdr->b_l1hdr.b_bufcnt -= 1;
5329 arc_cksum_verify(buf);
5331 arc_buf_unwatch(buf);
5334 mutex_exit(hash_lock);
5337 * Allocate a new hdr. The new hdr will contain a b_pdata
5338 * buffer which will be freed in arc_write().
5340 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
5341 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
5342 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
5343 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
5344 VERIFY3U(nhdr->b_type, ==, type);
5345 ASSERT(!HDR_SHARED_DATA(nhdr));
5347 nhdr->b_l1hdr.b_buf = buf;
5348 nhdr->b_l1hdr.b_bufcnt = 1;
5349 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
5352 mutex_exit(&buf->b_evict_lock);
5353 (void) refcount_add_many(&arc_anon->arcs_size,
5354 HDR_GET_LSIZE(nhdr), buf);
5356 mutex_exit(&buf->b_evict_lock);
5357 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
5358 /* protected by hash lock, or hdr is on arc_anon */
5359 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5360 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5361 arc_change_state(arc_anon, hdr, hash_lock);
5362 hdr->b_l1hdr.b_arc_access = 0;
5363 mutex_exit(hash_lock);
5365 buf_discard_identity(hdr);
5371 arc_released(arc_buf_t *buf)
5375 mutex_enter(&buf->b_evict_lock);
5376 released = (buf->b_data != NULL &&
5377 buf->b_hdr->b_l1hdr.b_state == arc_anon);
5378 mutex_exit(&buf->b_evict_lock);
5384 arc_referenced(arc_buf_t *buf)
5388 mutex_enter(&buf->b_evict_lock);
5389 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
5390 mutex_exit(&buf->b_evict_lock);
5391 return (referenced);
5396 arc_write_ready(zio_t *zio)
5398 arc_write_callback_t *callback = zio->io_private;
5399 arc_buf_t *buf = callback->awcb_buf;
5400 arc_buf_hdr_t *hdr = buf->b_hdr;
5401 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
5403 ASSERT(HDR_HAS_L1HDR(hdr));
5404 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
5405 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
5408 * If we're reexecuting this zio because the pool suspended, then
5409 * cleanup any state that was previously set the first time the
5410 * callback as invoked.
5412 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
5413 arc_cksum_free(hdr);
5415 arc_buf_unwatch(buf);
5417 if (hdr->b_l1hdr.b_pdata != NULL) {
5418 if (arc_buf_is_shared(buf)) {
5419 ASSERT(HDR_SHARED_DATA(hdr));
5421 arc_unshare_buf(hdr, buf);
5423 arc_hdr_free_pdata(hdr);
5427 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
5428 ASSERT(!HDR_SHARED_DATA(hdr));
5429 ASSERT(!arc_buf_is_shared(buf));
5431 callback->awcb_ready(zio, buf, callback->awcb_private);
5433 if (HDR_IO_IN_PROGRESS(hdr))
5434 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
5436 arc_cksum_compute(buf);
5437 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5439 enum zio_compress compress;
5440 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5441 compress = ZIO_COMPRESS_OFF;
5443 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
5444 compress = BP_GET_COMPRESS(zio->io_bp);
5446 HDR_SET_PSIZE(hdr, psize);
5447 arc_hdr_set_compress(hdr, compress);
5450 * If the hdr is compressed, then copy the compressed
5451 * zio contents into arc_buf_hdr_t. Otherwise, copy the original
5452 * data buf into the hdr. Ideally, we would like to always copy the
5453 * io_data into b_pdata but the user may have disabled compressed
5454 * arc thus the on-disk block may or may not match what we maintain
5455 * in the hdr's b_pdata field.
5457 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5458 ASSERT(BP_GET_COMPRESS(zio->io_bp) != ZIO_COMPRESS_OFF);
5459 ASSERT3U(psize, >, 0);
5460 arc_hdr_alloc_pdata(hdr);
5461 bcopy(zio->io_data, hdr->b_l1hdr.b_pdata, psize);
5463 ASSERT3P(buf->b_data, ==, zio->io_orig_data);
5464 ASSERT3U(zio->io_orig_size, ==, HDR_GET_LSIZE(hdr));
5465 ASSERT3U(hdr->b_l1hdr.b_byteswap, ==, DMU_BSWAP_NUMFUNCS);
5466 ASSERT(!HDR_SHARED_DATA(hdr));
5467 ASSERT(!arc_buf_is_shared(buf));
5468 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5469 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
5472 * This hdr is not compressed so we're able to share
5473 * the arc_buf_t data buffer with the hdr.
5475 arc_share_buf(hdr, buf);
5476 VERIFY0(bcmp(zio->io_orig_data, hdr->b_l1hdr.b_pdata,
5477 HDR_GET_LSIZE(hdr)));
5479 arc_hdr_verify(hdr, zio->io_bp);
5483 arc_write_children_ready(zio_t *zio)
5485 arc_write_callback_t *callback = zio->io_private;
5486 arc_buf_t *buf = callback->awcb_buf;
5488 callback->awcb_children_ready(zio, buf, callback->awcb_private);
5492 * The SPA calls this callback for each physical write that happens on behalf
5493 * of a logical write. See the comment in dbuf_write_physdone() for details.
5496 arc_write_physdone(zio_t *zio)
5498 arc_write_callback_t *cb = zio->io_private;
5499 if (cb->awcb_physdone != NULL)
5500 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
5504 arc_write_done(zio_t *zio)
5506 arc_write_callback_t *callback = zio->io_private;
5507 arc_buf_t *buf = callback->awcb_buf;
5508 arc_buf_hdr_t *hdr = buf->b_hdr;
5510 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5512 if (zio->io_error == 0) {
5513 arc_hdr_verify(hdr, zio->io_bp);
5515 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5516 buf_discard_identity(hdr);
5518 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
5519 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
5522 ASSERT(HDR_EMPTY(hdr));
5526 * If the block to be written was all-zero or compressed enough to be
5527 * embedded in the BP, no write was performed so there will be no
5528 * dva/birth/checksum. The buffer must therefore remain anonymous
5531 if (!HDR_EMPTY(hdr)) {
5532 arc_buf_hdr_t *exists;
5533 kmutex_t *hash_lock;
5535 ASSERT(zio->io_error == 0);
5537 arc_cksum_verify(buf);
5539 exists = buf_hash_insert(hdr, &hash_lock);
5540 if (exists != NULL) {
5542 * This can only happen if we overwrite for
5543 * sync-to-convergence, because we remove
5544 * buffers from the hash table when we arc_free().
5546 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
5547 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5548 panic("bad overwrite, hdr=%p exists=%p",
5549 (void *)hdr, (void *)exists);
5550 ASSERT(refcount_is_zero(
5551 &exists->b_l1hdr.b_refcnt));
5552 arc_change_state(arc_anon, exists, hash_lock);
5553 mutex_exit(hash_lock);
5554 arc_hdr_destroy(exists);
5555 exists = buf_hash_insert(hdr, &hash_lock);
5556 ASSERT3P(exists, ==, NULL);
5557 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
5559 ASSERT(zio->io_prop.zp_nopwrite);
5560 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5561 panic("bad nopwrite, hdr=%p exists=%p",
5562 (void *)hdr, (void *)exists);
5565 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
5566 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
5567 ASSERT(BP_GET_DEDUP(zio->io_bp));
5568 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
5571 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5572 /* if it's not anon, we are doing a scrub */
5573 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
5574 arc_access(hdr, hash_lock);
5575 mutex_exit(hash_lock);
5577 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5580 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5581 callback->awcb_done(zio, buf, callback->awcb_private);
5583 kmem_free(callback, sizeof (arc_write_callback_t));
5587 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
5588 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready,
5589 arc_done_func_t *children_ready, arc_done_func_t *physdone,
5590 arc_done_func_t *done, void *private, zio_priority_t priority,
5591 int zio_flags, const zbookmark_phys_t *zb)
5593 arc_buf_hdr_t *hdr = buf->b_hdr;
5594 arc_write_callback_t *callback;
5597 ASSERT3P(ready, !=, NULL);
5598 ASSERT3P(done, !=, NULL);
5599 ASSERT(!HDR_IO_ERROR(hdr));
5600 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5601 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5602 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
5604 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5605 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
5606 callback->awcb_ready = ready;
5607 callback->awcb_children_ready = children_ready;
5608 callback->awcb_physdone = physdone;
5609 callback->awcb_done = done;
5610 callback->awcb_private = private;
5611 callback->awcb_buf = buf;
5614 * The hdr's b_pdata is now stale, free it now. A new data block
5615 * will be allocated when the zio pipeline calls arc_write_ready().
5617 if (hdr->b_l1hdr.b_pdata != NULL) {
5619 * If the buf is currently sharing the data block with
5620 * the hdr then we need to break that relationship here.
5621 * The hdr will remain with a NULL data pointer and the
5622 * buf will take sole ownership of the block.
5624 if (arc_buf_is_shared(buf)) {
5625 ASSERT(ARC_BUF_LAST(buf));
5626 arc_unshare_buf(hdr, buf);
5628 arc_hdr_free_pdata(hdr);
5630 VERIFY3P(buf->b_data, !=, NULL);
5631 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
5633 ASSERT(!arc_buf_is_shared(buf));
5634 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL);
5636 zio = zio_write(pio, spa, txg, bp, buf->b_data, HDR_GET_LSIZE(hdr), zp,
5638 (children_ready != NULL) ? arc_write_children_ready : NULL,
5639 arc_write_physdone, arc_write_done, callback,
5640 priority, zio_flags, zb);
5646 arc_memory_throttle(uint64_t reserve, uint64_t txg)
5649 uint64_t available_memory = ptob(freemem);
5650 static uint64_t page_load = 0;
5651 static uint64_t last_txg = 0;
5653 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
5655 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE)));
5658 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
5661 if (txg > last_txg) {
5666 * If we are in pageout, we know that memory is already tight,
5667 * the arc is already going to be evicting, so we just want to
5668 * continue to let page writes occur as quickly as possible.
5670 if (curproc == pageproc) {
5671 if (page_load > MAX(ptob(minfree), available_memory) / 4)
5672 return (SET_ERROR(ERESTART));
5673 /* Note: reserve is inflated, so we deflate */
5674 page_load += reserve / 8;
5676 } else if (page_load > 0 && arc_reclaim_needed()) {
5677 /* memory is low, delay before restarting */
5678 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
5679 return (SET_ERROR(EAGAIN));
5687 arc_tempreserve_clear(uint64_t reserve)
5689 atomic_add_64(&arc_tempreserve, -reserve);
5690 ASSERT((int64_t)arc_tempreserve >= 0);
5694 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
5699 if (reserve > arc_c/4 && !arc_no_grow) {
5700 arc_c = MIN(arc_c_max, reserve * 4);
5701 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
5703 if (reserve > arc_c)
5704 return (SET_ERROR(ENOMEM));
5707 * Don't count loaned bufs as in flight dirty data to prevent long
5708 * network delays from blocking transactions that are ready to be
5709 * assigned to a txg.
5711 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
5712 arc_loaned_bytes), 0);
5715 * Writes will, almost always, require additional memory allocations
5716 * in order to compress/encrypt/etc the data. We therefore need to
5717 * make sure that there is sufficient available memory for this.
5719 error = arc_memory_throttle(reserve, txg);
5724 * Throttle writes when the amount of dirty data in the cache
5725 * gets too large. We try to keep the cache less than half full
5726 * of dirty blocks so that our sync times don't grow too large.
5727 * Note: if two requests come in concurrently, we might let them
5728 * both succeed, when one of them should fail. Not a huge deal.
5731 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
5732 anon_size > arc_c / 4) {
5733 uint64_t meta_esize =
5734 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
5735 uint64_t data_esize =
5736 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
5737 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
5738 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
5739 arc_tempreserve >> 10, meta_esize >> 10,
5740 data_esize >> 10, reserve >> 10, arc_c >> 10);
5741 return (SET_ERROR(ERESTART));
5743 atomic_add_64(&arc_tempreserve, reserve);
5748 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
5749 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
5751 size->value.ui64 = refcount_count(&state->arcs_size);
5752 evict_data->value.ui64 =
5753 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
5754 evict_metadata->value.ui64 =
5755 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
5759 arc_kstat_update(kstat_t *ksp, int rw)
5761 arc_stats_t *as = ksp->ks_data;
5763 if (rw == KSTAT_WRITE) {
5766 arc_kstat_update_state(arc_anon,
5767 &as->arcstat_anon_size,
5768 &as->arcstat_anon_evictable_data,
5769 &as->arcstat_anon_evictable_metadata);
5770 arc_kstat_update_state(arc_mru,
5771 &as->arcstat_mru_size,
5772 &as->arcstat_mru_evictable_data,
5773 &as->arcstat_mru_evictable_metadata);
5774 arc_kstat_update_state(arc_mru_ghost,
5775 &as->arcstat_mru_ghost_size,
5776 &as->arcstat_mru_ghost_evictable_data,
5777 &as->arcstat_mru_ghost_evictable_metadata);
5778 arc_kstat_update_state(arc_mfu,
5779 &as->arcstat_mfu_size,
5780 &as->arcstat_mfu_evictable_data,
5781 &as->arcstat_mfu_evictable_metadata);
5782 arc_kstat_update_state(arc_mfu_ghost,
5783 &as->arcstat_mfu_ghost_size,
5784 &as->arcstat_mfu_ghost_evictable_data,
5785 &as->arcstat_mfu_ghost_evictable_metadata);
5792 * This function *must* return indices evenly distributed between all
5793 * sublists of the multilist. This is needed due to how the ARC eviction
5794 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
5795 * distributed between all sublists and uses this assumption when
5796 * deciding which sublist to evict from and how much to evict from it.
5799 arc_state_multilist_index_func(multilist_t *ml, void *obj)
5801 arc_buf_hdr_t *hdr = obj;
5804 * We rely on b_dva to generate evenly distributed index
5805 * numbers using buf_hash below. So, as an added precaution,
5806 * let's make sure we never add empty buffers to the arc lists.
5808 ASSERT(!HDR_EMPTY(hdr));
5811 * The assumption here, is the hash value for a given
5812 * arc_buf_hdr_t will remain constant throughout it's lifetime
5813 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
5814 * Thus, we don't need to store the header's sublist index
5815 * on insertion, as this index can be recalculated on removal.
5817 * Also, the low order bits of the hash value are thought to be
5818 * distributed evenly. Otherwise, in the case that the multilist
5819 * has a power of two number of sublists, each sublists' usage
5820 * would not be evenly distributed.
5822 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
5823 multilist_get_num_sublists(ml));
5827 static eventhandler_tag arc_event_lowmem = NULL;
5830 arc_lowmem(void *arg __unused, int howto __unused)
5833 mutex_enter(&arc_reclaim_lock);
5834 /* XXX: Memory deficit should be passed as argument. */
5835 needfree = btoc(arc_c >> arc_shrink_shift);
5836 DTRACE_PROBE(arc__needfree);
5837 cv_signal(&arc_reclaim_thread_cv);
5840 * It is unsafe to block here in arbitrary threads, because we can come
5841 * here from ARC itself and may hold ARC locks and thus risk a deadlock
5842 * with ARC reclaim thread.
5844 if (curproc == pageproc)
5845 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
5846 mutex_exit(&arc_reclaim_lock);
5851 arc_state_init(void)
5853 arc_anon = &ARC_anon;
5855 arc_mru_ghost = &ARC_mru_ghost;
5857 arc_mfu_ghost = &ARC_mfu_ghost;
5858 arc_l2c_only = &ARC_l2c_only;
5860 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
5861 sizeof (arc_buf_hdr_t),
5862 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5863 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5864 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
5865 sizeof (arc_buf_hdr_t),
5866 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5867 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5868 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
5869 sizeof (arc_buf_hdr_t),
5870 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5871 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5872 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
5873 sizeof (arc_buf_hdr_t),
5874 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5875 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5876 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
5877 sizeof (arc_buf_hdr_t),
5878 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5879 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5880 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
5881 sizeof (arc_buf_hdr_t),
5882 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5883 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5884 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
5885 sizeof (arc_buf_hdr_t),
5886 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5887 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5888 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
5889 sizeof (arc_buf_hdr_t),
5890 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5891 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5892 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
5893 sizeof (arc_buf_hdr_t),
5894 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5895 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5896 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
5897 sizeof (arc_buf_hdr_t),
5898 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5899 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5901 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
5902 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
5903 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
5904 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
5905 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
5906 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
5907 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
5908 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
5909 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
5910 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
5911 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
5912 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
5914 refcount_create(&arc_anon->arcs_size);
5915 refcount_create(&arc_mru->arcs_size);
5916 refcount_create(&arc_mru_ghost->arcs_size);
5917 refcount_create(&arc_mfu->arcs_size);
5918 refcount_create(&arc_mfu_ghost->arcs_size);
5919 refcount_create(&arc_l2c_only->arcs_size);
5923 arc_state_fini(void)
5925 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
5926 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
5927 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
5928 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
5929 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
5930 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
5931 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
5932 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
5933 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
5934 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
5935 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
5936 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
5938 refcount_destroy(&arc_anon->arcs_size);
5939 refcount_destroy(&arc_mru->arcs_size);
5940 refcount_destroy(&arc_mru_ghost->arcs_size);
5941 refcount_destroy(&arc_mfu->arcs_size);
5942 refcount_destroy(&arc_mfu_ghost->arcs_size);
5943 refcount_destroy(&arc_l2c_only->arcs_size);
5945 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
5946 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
5947 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
5948 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
5949 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
5950 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
5951 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
5952 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
5964 int i, prefetch_tunable_set = 0;
5966 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
5967 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
5968 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
5970 /* Convert seconds to clock ticks */
5971 arc_min_prefetch_lifespan = 1 * hz;
5973 /* Start out with 1/8 of all memory */
5974 arc_c = kmem_size() / 8;
5979 * On architectures where the physical memory can be larger
5980 * than the addressable space (intel in 32-bit mode), we may
5981 * need to limit the cache to 1/8 of VM size.
5983 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
5985 #endif /* illumos */
5986 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
5987 arc_c_min = MAX(arc_c / 4, arc_abs_min);
5988 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
5989 if (arc_c * 8 >= 1 << 30)
5990 arc_c_max = (arc_c * 8) - (1 << 30);
5992 arc_c_max = arc_c_min;
5993 arc_c_max = MAX(arc_c * 5, arc_c_max);
5996 * In userland, there's only the memory pressure that we artificially
5997 * create (see arc_available_memory()). Don't let arc_c get too
5998 * small, because it can cause transactions to be larger than
5999 * arc_c, causing arc_tempreserve_space() to fail.
6002 arc_c_min = arc_c_max / 2;
6007 * Allow the tunables to override our calculations if they are
6010 if (zfs_arc_max > arc_abs_min && zfs_arc_max < kmem_size()) {
6011 arc_c_max = zfs_arc_max;
6012 arc_c_min = MIN(arc_c_min, arc_c_max);
6014 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
6015 arc_c_min = zfs_arc_min;
6019 arc_p = (arc_c >> 1);
6022 /* limit meta-data to 1/4 of the arc capacity */
6023 arc_meta_limit = arc_c_max / 4;
6025 /* Allow the tunable to override if it is reasonable */
6026 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
6027 arc_meta_limit = zfs_arc_meta_limit;
6029 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
6030 arc_c_min = arc_meta_limit / 2;
6032 if (zfs_arc_meta_min > 0) {
6033 arc_meta_min = zfs_arc_meta_min;
6035 arc_meta_min = arc_c_min / 2;
6038 if (zfs_arc_grow_retry > 0)
6039 arc_grow_retry = zfs_arc_grow_retry;
6041 if (zfs_arc_shrink_shift > 0)
6042 arc_shrink_shift = zfs_arc_shrink_shift;
6045 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6047 if (arc_no_grow_shift >= arc_shrink_shift)
6048 arc_no_grow_shift = arc_shrink_shift - 1;
6050 if (zfs_arc_p_min_shift > 0)
6051 arc_p_min_shift = zfs_arc_p_min_shift;
6053 if (zfs_arc_num_sublists_per_state < 1)
6054 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1);
6056 /* if kmem_flags are set, lets try to use less memory */
6057 if (kmem_debugging())
6059 if (arc_c < arc_c_min)
6062 zfs_arc_min = arc_c_min;
6063 zfs_arc_max = arc_c_max;
6068 arc_reclaim_thread_exit = B_FALSE;
6070 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
6071 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
6073 if (arc_ksp != NULL) {
6074 arc_ksp->ks_data = &arc_stats;
6075 arc_ksp->ks_update = arc_kstat_update;
6076 kstat_install(arc_ksp);
6079 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
6080 TS_RUN, minclsyspri);
6083 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
6084 EVENTHANDLER_PRI_FIRST);
6091 * Calculate maximum amount of dirty data per pool.
6093 * If it has been set by /etc/system, take that.
6094 * Otherwise, use a percentage of physical memory defined by
6095 * zfs_dirty_data_max_percent (default 10%) with a cap at
6096 * zfs_dirty_data_max_max (default 4GB).
6098 if (zfs_dirty_data_max == 0) {
6099 zfs_dirty_data_max = ptob(physmem) *
6100 zfs_dirty_data_max_percent / 100;
6101 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
6102 zfs_dirty_data_max_max);
6106 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
6107 prefetch_tunable_set = 1;
6110 if (prefetch_tunable_set == 0) {
6111 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
6113 printf(" add \"vfs.zfs.prefetch_disable=0\" "
6114 "to /boot/loader.conf.\n");
6115 zfs_prefetch_disable = 1;
6118 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
6119 prefetch_tunable_set == 0) {
6120 printf("ZFS NOTICE: Prefetch is disabled by default if less "
6121 "than 4GB of RAM is present;\n"
6122 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
6123 "to /boot/loader.conf.\n");
6124 zfs_prefetch_disable = 1;
6127 /* Warn about ZFS memory and address space requirements. */
6128 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
6129 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
6130 "expect unstable behavior.\n");
6132 if (kmem_size() < 512 * (1 << 20)) {
6133 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
6134 "expect unstable behavior.\n");
6135 printf(" Consider tuning vm.kmem_size and "
6136 "vm.kmem_size_max\n");
6137 printf(" in /boot/loader.conf.\n");
6145 mutex_enter(&arc_reclaim_lock);
6146 arc_reclaim_thread_exit = B_TRUE;
6148 * The reclaim thread will set arc_reclaim_thread_exit back to
6149 * B_FALSE when it is finished exiting; we're waiting for that.
6151 while (arc_reclaim_thread_exit) {
6152 cv_signal(&arc_reclaim_thread_cv);
6153 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
6155 mutex_exit(&arc_reclaim_lock);
6157 /* Use B_TRUE to ensure *all* buffers are evicted */
6158 arc_flush(NULL, B_TRUE);
6162 if (arc_ksp != NULL) {
6163 kstat_delete(arc_ksp);
6167 mutex_destroy(&arc_reclaim_lock);
6168 cv_destroy(&arc_reclaim_thread_cv);
6169 cv_destroy(&arc_reclaim_waiters_cv);
6174 ASSERT0(arc_loaned_bytes);
6177 if (arc_event_lowmem != NULL)
6178 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
6185 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6186 * It uses dedicated storage devices to hold cached data, which are populated
6187 * using large infrequent writes. The main role of this cache is to boost
6188 * the performance of random read workloads. The intended L2ARC devices
6189 * include short-stroked disks, solid state disks, and other media with
6190 * substantially faster read latency than disk.
6192 * +-----------------------+
6194 * +-----------------------+
6197 * l2arc_feed_thread() arc_read()
6201 * +---------------+ |
6203 * +---------------+ |
6208 * +-------+ +-------+
6210 * | cache | | cache |
6211 * +-------+ +-------+
6212 * +=========+ .-----.
6213 * : L2ARC : |-_____-|
6214 * : devices : | Disks |
6215 * +=========+ `-_____-'
6217 * Read requests are satisfied from the following sources, in order:
6220 * 2) vdev cache of L2ARC devices
6222 * 4) vdev cache of disks
6225 * Some L2ARC device types exhibit extremely slow write performance.
6226 * To accommodate for this there are some significant differences between
6227 * the L2ARC and traditional cache design:
6229 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6230 * the ARC behave as usual, freeing buffers and placing headers on ghost
6231 * lists. The ARC does not send buffers to the L2ARC during eviction as
6232 * this would add inflated write latencies for all ARC memory pressure.
6234 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6235 * It does this by periodically scanning buffers from the eviction-end of
6236 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6237 * not already there. It scans until a headroom of buffers is satisfied,
6238 * which itself is a buffer for ARC eviction. If a compressible buffer is
6239 * found during scanning and selected for writing to an L2ARC device, we
6240 * temporarily boost scanning headroom during the next scan cycle to make
6241 * sure we adapt to compression effects (which might significantly reduce
6242 * the data volume we write to L2ARC). The thread that does this is
6243 * l2arc_feed_thread(), illustrated below; example sizes are included to
6244 * provide a better sense of ratio than this diagram:
6247 * +---------------------+----------+
6248 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6249 * +---------------------+----------+ | o L2ARC eligible
6250 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6251 * +---------------------+----------+ |
6252 * 15.9 Gbytes ^ 32 Mbytes |
6254 * l2arc_feed_thread()
6256 * l2arc write hand <--[oooo]--'
6260 * +==============================+
6261 * L2ARC dev |####|#|###|###| |####| ... |
6262 * +==============================+
6265 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6266 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6267 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6268 * safe to say that this is an uncommon case, since buffers at the end of
6269 * the ARC lists have moved there due to inactivity.
6271 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6272 * then the L2ARC simply misses copying some buffers. This serves as a
6273 * pressure valve to prevent heavy read workloads from both stalling the ARC
6274 * with waits and clogging the L2ARC with writes. This also helps prevent
6275 * the potential for the L2ARC to churn if it attempts to cache content too
6276 * quickly, such as during backups of the entire pool.
6278 * 5. After system boot and before the ARC has filled main memory, there are
6279 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6280 * lists can remain mostly static. Instead of searching from tail of these
6281 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6282 * for eligible buffers, greatly increasing its chance of finding them.
6284 * The L2ARC device write speed is also boosted during this time so that
6285 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6286 * there are no L2ARC reads, and no fear of degrading read performance
6287 * through increased writes.
6289 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6290 * the vdev queue can aggregate them into larger and fewer writes. Each
6291 * device is written to in a rotor fashion, sweeping writes through
6292 * available space then repeating.
6294 * 7. The L2ARC does not store dirty content. It never needs to flush
6295 * write buffers back to disk based storage.
6297 * 8. If an ARC buffer is written (and dirtied) which also exists in the
6298 * L2ARC, the now stale L2ARC buffer is immediately dropped.
6300 * The performance of the L2ARC can be tweaked by a number of tunables, which
6301 * may be necessary for different workloads:
6303 * l2arc_write_max max write bytes per interval
6304 * l2arc_write_boost extra write bytes during device warmup
6305 * l2arc_noprefetch skip caching prefetched buffers
6306 * l2arc_headroom number of max device writes to precache
6307 * l2arc_headroom_boost when we find compressed buffers during ARC
6308 * scanning, we multiply headroom by this
6309 * percentage factor for the next scan cycle,
6310 * since more compressed buffers are likely to
6312 * l2arc_feed_secs seconds between L2ARC writing
6314 * Tunables may be removed or added as future performance improvements are
6315 * integrated, and also may become zpool properties.
6317 * There are three key functions that control how the L2ARC warms up:
6319 * l2arc_write_eligible() check if a buffer is eligible to cache
6320 * l2arc_write_size() calculate how much to write
6321 * l2arc_write_interval() calculate sleep delay between writes
6323 * These three functions determine what to write, how much, and how quickly
6328 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
6331 * A buffer is *not* eligible for the L2ARC if it:
6332 * 1. belongs to a different spa.
6333 * 2. is already cached on the L2ARC.
6334 * 3. has an I/O in progress (it may be an incomplete read).
6335 * 4. is flagged not eligible (zfs property).
6337 if (hdr->b_spa != spa_guid) {
6338 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
6341 if (HDR_HAS_L2HDR(hdr)) {
6342 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
6345 if (HDR_IO_IN_PROGRESS(hdr)) {
6346 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
6349 if (!HDR_L2CACHE(hdr)) {
6350 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
6358 l2arc_write_size(void)
6363 * Make sure our globals have meaningful values in case the user
6366 size = l2arc_write_max;
6368 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
6369 "be greater than zero, resetting it to the default (%d)",
6371 size = l2arc_write_max = L2ARC_WRITE_SIZE;
6374 if (arc_warm == B_FALSE)
6375 size += l2arc_write_boost;
6382 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
6384 clock_t interval, next, now;
6387 * If the ARC lists are busy, increase our write rate; if the
6388 * lists are stale, idle back. This is achieved by checking
6389 * how much we previously wrote - if it was more than half of
6390 * what we wanted, schedule the next write much sooner.
6392 if (l2arc_feed_again && wrote > (wanted / 2))
6393 interval = (hz * l2arc_feed_min_ms) / 1000;
6395 interval = hz * l2arc_feed_secs;
6397 now = ddi_get_lbolt();
6398 next = MAX(now, MIN(now + interval, began + interval));
6404 * Cycle through L2ARC devices. This is how L2ARC load balances.
6405 * If a device is returned, this also returns holding the spa config lock.
6407 static l2arc_dev_t *
6408 l2arc_dev_get_next(void)
6410 l2arc_dev_t *first, *next = NULL;
6413 * Lock out the removal of spas (spa_namespace_lock), then removal
6414 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
6415 * both locks will be dropped and a spa config lock held instead.
6417 mutex_enter(&spa_namespace_lock);
6418 mutex_enter(&l2arc_dev_mtx);
6420 /* if there are no vdevs, there is nothing to do */
6421 if (l2arc_ndev == 0)
6425 next = l2arc_dev_last;
6427 /* loop around the list looking for a non-faulted vdev */
6429 next = list_head(l2arc_dev_list);
6431 next = list_next(l2arc_dev_list, next);
6433 next = list_head(l2arc_dev_list);
6436 /* if we have come back to the start, bail out */
6439 else if (next == first)
6442 } while (vdev_is_dead(next->l2ad_vdev));
6444 /* if we were unable to find any usable vdevs, return NULL */
6445 if (vdev_is_dead(next->l2ad_vdev))
6448 l2arc_dev_last = next;
6451 mutex_exit(&l2arc_dev_mtx);
6454 * Grab the config lock to prevent the 'next' device from being
6455 * removed while we are writing to it.
6458 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
6459 mutex_exit(&spa_namespace_lock);
6465 * Free buffers that were tagged for destruction.
6468 l2arc_do_free_on_write()
6471 l2arc_data_free_t *df, *df_prev;
6473 mutex_enter(&l2arc_free_on_write_mtx);
6474 buflist = l2arc_free_on_write;
6476 for (df = list_tail(buflist); df; df = df_prev) {
6477 df_prev = list_prev(buflist, df);
6478 ASSERT3P(df->l2df_data, !=, NULL);
6479 if (df->l2df_type == ARC_BUFC_METADATA) {
6480 zio_buf_free(df->l2df_data, df->l2df_size);
6482 ASSERT(df->l2df_type == ARC_BUFC_DATA);
6483 zio_data_buf_free(df->l2df_data, df->l2df_size);
6485 list_remove(buflist, df);
6486 kmem_free(df, sizeof (l2arc_data_free_t));
6489 mutex_exit(&l2arc_free_on_write_mtx);
6493 * A write to a cache device has completed. Update all headers to allow
6494 * reads from these buffers to begin.
6497 l2arc_write_done(zio_t *zio)
6499 l2arc_write_callback_t *cb;
6502 arc_buf_hdr_t *head, *hdr, *hdr_prev;
6503 kmutex_t *hash_lock;
6504 int64_t bytes_dropped = 0;
6506 cb = zio->io_private;
6507 ASSERT3P(cb, !=, NULL);
6508 dev = cb->l2wcb_dev;
6509 ASSERT3P(dev, !=, NULL);
6510 head = cb->l2wcb_head;
6511 ASSERT3P(head, !=, NULL);
6512 buflist = &dev->l2ad_buflist;
6513 ASSERT3P(buflist, !=, NULL);
6514 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
6515 l2arc_write_callback_t *, cb);
6517 if (zio->io_error != 0)
6518 ARCSTAT_BUMP(arcstat_l2_writes_error);
6521 * All writes completed, or an error was hit.
6524 mutex_enter(&dev->l2ad_mtx);
6525 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
6526 hdr_prev = list_prev(buflist, hdr);
6528 hash_lock = HDR_LOCK(hdr);
6531 * We cannot use mutex_enter or else we can deadlock
6532 * with l2arc_write_buffers (due to swapping the order
6533 * the hash lock and l2ad_mtx are taken).
6535 if (!mutex_tryenter(hash_lock)) {
6537 * Missed the hash lock. We must retry so we
6538 * don't leave the ARC_FLAG_L2_WRITING bit set.
6540 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
6543 * We don't want to rescan the headers we've
6544 * already marked as having been written out, so
6545 * we reinsert the head node so we can pick up
6546 * where we left off.
6548 list_remove(buflist, head);
6549 list_insert_after(buflist, hdr, head);
6551 mutex_exit(&dev->l2ad_mtx);
6554 * We wait for the hash lock to become available
6555 * to try and prevent busy waiting, and increase
6556 * the chance we'll be able to acquire the lock
6557 * the next time around.
6559 mutex_enter(hash_lock);
6560 mutex_exit(hash_lock);
6565 * We could not have been moved into the arc_l2c_only
6566 * state while in-flight due to our ARC_FLAG_L2_WRITING
6567 * bit being set. Let's just ensure that's being enforced.
6569 ASSERT(HDR_HAS_L1HDR(hdr));
6571 if (zio->io_error != 0) {
6573 * Error - drop L2ARC entry.
6575 list_remove(buflist, hdr);
6577 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
6579 ARCSTAT_INCR(arcstat_l2_asize, -arc_hdr_size(hdr));
6580 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr));
6582 bytes_dropped += arc_hdr_size(hdr);
6583 (void) refcount_remove_many(&dev->l2ad_alloc,
6584 arc_hdr_size(hdr), hdr);
6588 * Allow ARC to begin reads and ghost list evictions to
6591 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
6593 mutex_exit(hash_lock);
6596 atomic_inc_64(&l2arc_writes_done);
6597 list_remove(buflist, head);
6598 ASSERT(!HDR_HAS_L1HDR(head));
6599 kmem_cache_free(hdr_l2only_cache, head);
6600 mutex_exit(&dev->l2ad_mtx);
6602 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
6604 l2arc_do_free_on_write();
6606 kmem_free(cb, sizeof (l2arc_write_callback_t));
6610 * A read to a cache device completed. Validate buffer contents before
6611 * handing over to the regular ARC routines.
6614 l2arc_read_done(zio_t *zio)
6616 l2arc_read_callback_t *cb;
6618 kmutex_t *hash_lock;
6619 boolean_t valid_cksum;
6621 ASSERT3P(zio->io_vd, !=, NULL);
6622 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
6624 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
6626 cb = zio->io_private;
6627 ASSERT3P(cb, !=, NULL);
6628 hdr = cb->l2rcb_hdr;
6629 ASSERT3P(hdr, !=, NULL);
6631 hash_lock = HDR_LOCK(hdr);
6632 mutex_enter(hash_lock);
6633 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6636 * If the data was read into a temporary buffer,
6637 * move it and free the buffer.
6639 if (cb->l2rcb_data != NULL) {
6640 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
6641 if (zio->io_error == 0) {
6642 bcopy(cb->l2rcb_data, hdr->b_l1hdr.b_pdata,
6647 * The following must be done regardless of whether
6648 * there was an error:
6649 * - free the temporary buffer
6650 * - point zio to the real ARC buffer
6651 * - set zio size accordingly
6652 * These are required because zio is either re-used for
6653 * an I/O of the block in the case of the error
6654 * or the zio is passed to arc_read_done() and it
6657 zio_data_buf_free(cb->l2rcb_data, zio->io_size);
6658 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
6659 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_pdata;
6662 ASSERT3P(zio->io_data, !=, NULL);
6665 * Check this survived the L2ARC journey.
6667 ASSERT3P(zio->io_data, ==, hdr->b_l1hdr.b_pdata);
6668 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
6669 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
6671 valid_cksum = arc_cksum_is_equal(hdr, zio);
6672 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
6673 mutex_exit(hash_lock);
6674 zio->io_private = hdr;
6677 mutex_exit(hash_lock);
6679 * Buffer didn't survive caching. Increment stats and
6680 * reissue to the original storage device.
6682 if (zio->io_error != 0) {
6683 ARCSTAT_BUMP(arcstat_l2_io_error);
6685 zio->io_error = SET_ERROR(EIO);
6688 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
6691 * If there's no waiter, issue an async i/o to the primary
6692 * storage now. If there *is* a waiter, the caller must
6693 * issue the i/o in a context where it's OK to block.
6695 if (zio->io_waiter == NULL) {
6696 zio_t *pio = zio_unique_parent(zio);
6698 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
6700 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
6701 hdr->b_l1hdr.b_pdata, zio->io_size, arc_read_done,
6702 hdr, zio->io_priority, cb->l2rcb_flags,
6707 kmem_free(cb, sizeof (l2arc_read_callback_t));
6711 * This is the list priority from which the L2ARC will search for pages to
6712 * cache. This is used within loops (0..3) to cycle through lists in the
6713 * desired order. This order can have a significant effect on cache
6716 * Currently the metadata lists are hit first, MFU then MRU, followed by
6717 * the data lists. This function returns a locked list, and also returns
6720 static multilist_sublist_t *
6721 l2arc_sublist_lock(int list_num)
6723 multilist_t *ml = NULL;
6726 ASSERT(list_num >= 0 && list_num <= 3);
6730 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
6733 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
6736 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
6739 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
6744 * Return a randomly-selected sublist. This is acceptable
6745 * because the caller feeds only a little bit of data for each
6746 * call (8MB). Subsequent calls will result in different
6747 * sublists being selected.
6749 idx = multilist_get_random_index(ml);
6750 return (multilist_sublist_lock(ml, idx));
6754 * Evict buffers from the device write hand to the distance specified in
6755 * bytes. This distance may span populated buffers, it may span nothing.
6756 * This is clearing a region on the L2ARC device ready for writing.
6757 * If the 'all' boolean is set, every buffer is evicted.
6760 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
6763 arc_buf_hdr_t *hdr, *hdr_prev;
6764 kmutex_t *hash_lock;
6767 buflist = &dev->l2ad_buflist;
6769 if (!all && dev->l2ad_first) {
6771 * This is the first sweep through the device. There is
6777 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
6779 * When nearing the end of the device, evict to the end
6780 * before the device write hand jumps to the start.
6782 taddr = dev->l2ad_end;
6784 taddr = dev->l2ad_hand + distance;
6786 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
6787 uint64_t, taddr, boolean_t, all);
6790 mutex_enter(&dev->l2ad_mtx);
6791 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
6792 hdr_prev = list_prev(buflist, hdr);
6794 hash_lock = HDR_LOCK(hdr);
6797 * We cannot use mutex_enter or else we can deadlock
6798 * with l2arc_write_buffers (due to swapping the order
6799 * the hash lock and l2ad_mtx are taken).
6801 if (!mutex_tryenter(hash_lock)) {
6803 * Missed the hash lock. Retry.
6805 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
6806 mutex_exit(&dev->l2ad_mtx);
6807 mutex_enter(hash_lock);
6808 mutex_exit(hash_lock);
6812 if (HDR_L2_WRITE_HEAD(hdr)) {
6814 * We hit a write head node. Leave it for
6815 * l2arc_write_done().
6817 list_remove(buflist, hdr);
6818 mutex_exit(hash_lock);
6822 if (!all && HDR_HAS_L2HDR(hdr) &&
6823 (hdr->b_l2hdr.b_daddr >= taddr ||
6824 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
6826 * We've evicted to the target address,
6827 * or the end of the device.
6829 mutex_exit(hash_lock);
6833 ASSERT(HDR_HAS_L2HDR(hdr));
6834 if (!HDR_HAS_L1HDR(hdr)) {
6835 ASSERT(!HDR_L2_READING(hdr));
6837 * This doesn't exist in the ARC. Destroy.
6838 * arc_hdr_destroy() will call list_remove()
6839 * and decrement arcstat_l2_size.
6841 arc_change_state(arc_anon, hdr, hash_lock);
6842 arc_hdr_destroy(hdr);
6844 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
6845 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
6847 * Invalidate issued or about to be issued
6848 * reads, since we may be about to write
6849 * over this location.
6851 if (HDR_L2_READING(hdr)) {
6852 ARCSTAT_BUMP(arcstat_l2_evict_reading);
6853 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
6856 /* Ensure this header has finished being written */
6857 ASSERT(!HDR_L2_WRITING(hdr));
6859 arc_hdr_l2hdr_destroy(hdr);
6861 mutex_exit(hash_lock);
6863 mutex_exit(&dev->l2ad_mtx);
6867 * Find and write ARC buffers to the L2ARC device.
6869 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
6870 * for reading until they have completed writing.
6871 * The headroom_boost is an in-out parameter used to maintain headroom boost
6872 * state between calls to this function.
6874 * Returns the number of bytes actually written (which may be smaller than
6875 * the delta by which the device hand has changed due to alignment).
6878 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
6880 arc_buf_hdr_t *hdr, *hdr_prev, *head;
6881 uint64_t write_asize, write_psize, write_sz, headroom;
6883 l2arc_write_callback_t *cb;
6885 uint64_t guid = spa_load_guid(spa);
6888 ASSERT3P(dev->l2ad_vdev, !=, NULL);
6891 write_sz = write_asize = write_psize = 0;
6893 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
6894 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
6896 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
6898 * Copy buffers for L2ARC writing.
6900 for (try = 0; try <= 3; try++) {
6901 multilist_sublist_t *mls = l2arc_sublist_lock(try);
6902 uint64_t passed_sz = 0;
6904 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
6907 * L2ARC fast warmup.
6909 * Until the ARC is warm and starts to evict, read from the
6910 * head of the ARC lists rather than the tail.
6912 if (arc_warm == B_FALSE)
6913 hdr = multilist_sublist_head(mls);
6915 hdr = multilist_sublist_tail(mls);
6917 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
6919 headroom = target_sz * l2arc_headroom;
6920 if (zfs_compressed_arc_enabled)
6921 headroom = (headroom * l2arc_headroom_boost) / 100;
6923 for (; hdr; hdr = hdr_prev) {
6924 kmutex_t *hash_lock;
6926 if (arc_warm == B_FALSE)
6927 hdr_prev = multilist_sublist_next(mls, hdr);
6929 hdr_prev = multilist_sublist_prev(mls, hdr);
6930 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned,
6931 HDR_GET_LSIZE(hdr));
6933 hash_lock = HDR_LOCK(hdr);
6934 if (!mutex_tryenter(hash_lock)) {
6935 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
6937 * Skip this buffer rather than waiting.
6942 passed_sz += HDR_GET_LSIZE(hdr);
6943 if (passed_sz > headroom) {
6947 mutex_exit(hash_lock);
6948 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
6952 if (!l2arc_write_eligible(guid, hdr)) {
6953 mutex_exit(hash_lock);
6958 * We rely on the L1 portion of the header below, so
6959 * it's invalid for this header to have been evicted out
6960 * of the ghost cache, prior to being written out. The
6961 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6963 ASSERT(HDR_HAS_L1HDR(hdr));
6965 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
6966 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL);
6967 ASSERT3U(arc_hdr_size(hdr), >, 0);
6968 uint64_t size = arc_hdr_size(hdr);
6969 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
6972 if ((write_psize + asize) > target_sz) {
6974 mutex_exit(hash_lock);
6975 ARCSTAT_BUMP(arcstat_l2_write_full);
6981 * Insert a dummy header on the buflist so
6982 * l2arc_write_done() can find where the
6983 * write buffers begin without searching.
6985 mutex_enter(&dev->l2ad_mtx);
6986 list_insert_head(&dev->l2ad_buflist, head);
6987 mutex_exit(&dev->l2ad_mtx);
6990 sizeof (l2arc_write_callback_t), KM_SLEEP);
6991 cb->l2wcb_dev = dev;
6992 cb->l2wcb_head = head;
6993 pio = zio_root(spa, l2arc_write_done, cb,
6995 ARCSTAT_BUMP(arcstat_l2_write_pios);
6998 hdr->b_l2hdr.b_dev = dev;
6999 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
7000 arc_hdr_set_flags(hdr,
7001 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
7003 mutex_enter(&dev->l2ad_mtx);
7004 list_insert_head(&dev->l2ad_buflist, hdr);
7005 mutex_exit(&dev->l2ad_mtx);
7007 (void) refcount_add_many(&dev->l2ad_alloc, size, hdr);
7010 * Normally the L2ARC can use the hdr's data, but if
7011 * we're sharing data between the hdr and one of its
7012 * bufs, L2ARC needs its own copy of the data so that
7013 * the ZIO below can't race with the buf consumer. To
7014 * ensure that this copy will be available for the
7015 * lifetime of the ZIO and be cleaned up afterwards, we
7016 * add it to the l2arc_free_on_write queue.
7019 if (!HDR_SHARED_DATA(hdr) && size == asize) {
7020 to_write = hdr->b_l1hdr.b_pdata;
7022 arc_buf_contents_t type = arc_buf_type(hdr);
7023 if (type == ARC_BUFC_METADATA) {
7024 to_write = zio_buf_alloc(asize);
7026 ASSERT3U(type, ==, ARC_BUFC_DATA);
7027 to_write = zio_data_buf_alloc(asize);
7030 bcopy(hdr->b_l1hdr.b_pdata, to_write, size);
7032 bzero(to_write + size, asize - size);
7033 l2arc_free_data_on_write(to_write, asize, type);
7035 wzio = zio_write_phys(pio, dev->l2ad_vdev,
7036 hdr->b_l2hdr.b_daddr, asize, to_write,
7037 ZIO_CHECKSUM_OFF, NULL, hdr,
7038 ZIO_PRIORITY_ASYNC_WRITE,
7039 ZIO_FLAG_CANFAIL, B_FALSE);
7041 write_sz += HDR_GET_LSIZE(hdr);
7042 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
7045 write_asize += size;
7046 write_psize += asize;
7047 dev->l2ad_hand += asize;
7049 mutex_exit(hash_lock);
7051 (void) zio_nowait(wzio);
7054 multilist_sublist_unlock(mls);
7060 /* No buffers selected for writing? */
7063 ASSERT(!HDR_HAS_L1HDR(head));
7064 kmem_cache_free(hdr_l2only_cache, head);
7068 ASSERT3U(write_psize, <=, target_sz);
7069 ARCSTAT_BUMP(arcstat_l2_writes_sent);
7070 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
7071 ARCSTAT_INCR(arcstat_l2_size, write_sz);
7072 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
7073 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0);
7076 * Bump device hand to the device start if it is approaching the end.
7077 * l2arc_evict() will already have evicted ahead for this case.
7079 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
7080 dev->l2ad_hand = dev->l2ad_start;
7081 dev->l2ad_first = B_FALSE;
7084 dev->l2ad_writing = B_TRUE;
7085 (void) zio_wait(pio);
7086 dev->l2ad_writing = B_FALSE;
7088 return (write_asize);
7092 * This thread feeds the L2ARC at regular intervals. This is the beating
7093 * heart of the L2ARC.
7096 l2arc_feed_thread(void *dummy __unused)
7101 uint64_t size, wrote;
7102 clock_t begin, next = ddi_get_lbolt();
7104 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
7106 mutex_enter(&l2arc_feed_thr_lock);
7108 while (l2arc_thread_exit == 0) {
7109 CALLB_CPR_SAFE_BEGIN(&cpr);
7110 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
7111 next - ddi_get_lbolt());
7112 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
7113 next = ddi_get_lbolt() + hz;
7116 * Quick check for L2ARC devices.
7118 mutex_enter(&l2arc_dev_mtx);
7119 if (l2arc_ndev == 0) {
7120 mutex_exit(&l2arc_dev_mtx);
7123 mutex_exit(&l2arc_dev_mtx);
7124 begin = ddi_get_lbolt();
7127 * This selects the next l2arc device to write to, and in
7128 * doing so the next spa to feed from: dev->l2ad_spa. This
7129 * will return NULL if there are now no l2arc devices or if
7130 * they are all faulted.
7132 * If a device is returned, its spa's config lock is also
7133 * held to prevent device removal. l2arc_dev_get_next()
7134 * will grab and release l2arc_dev_mtx.
7136 if ((dev = l2arc_dev_get_next()) == NULL)
7139 spa = dev->l2ad_spa;
7140 ASSERT3P(spa, !=, NULL);
7143 * If the pool is read-only then force the feed thread to
7144 * sleep a little longer.
7146 if (!spa_writeable(spa)) {
7147 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
7148 spa_config_exit(spa, SCL_L2ARC, dev);
7153 * Avoid contributing to memory pressure.
7155 if (arc_reclaim_needed()) {
7156 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
7157 spa_config_exit(spa, SCL_L2ARC, dev);
7161 ARCSTAT_BUMP(arcstat_l2_feeds);
7163 size = l2arc_write_size();
7166 * Evict L2ARC buffers that will be overwritten.
7168 l2arc_evict(dev, size, B_FALSE);
7171 * Write ARC buffers.
7173 wrote = l2arc_write_buffers(spa, dev, size);
7176 * Calculate interval between writes.
7178 next = l2arc_write_interval(begin, size, wrote);
7179 spa_config_exit(spa, SCL_L2ARC, dev);
7182 l2arc_thread_exit = 0;
7183 cv_broadcast(&l2arc_feed_thr_cv);
7184 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
7189 l2arc_vdev_present(vdev_t *vd)
7193 mutex_enter(&l2arc_dev_mtx);
7194 for (dev = list_head(l2arc_dev_list); dev != NULL;
7195 dev = list_next(l2arc_dev_list, dev)) {
7196 if (dev->l2ad_vdev == vd)
7199 mutex_exit(&l2arc_dev_mtx);
7201 return (dev != NULL);
7205 * Add a vdev for use by the L2ARC. By this point the spa has already
7206 * validated the vdev and opened it.
7209 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
7211 l2arc_dev_t *adddev;
7213 ASSERT(!l2arc_vdev_present(vd));
7215 vdev_ashift_optimize(vd);
7218 * Create a new l2arc device entry.
7220 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
7221 adddev->l2ad_spa = spa;
7222 adddev->l2ad_vdev = vd;
7223 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
7224 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
7225 adddev->l2ad_hand = adddev->l2ad_start;
7226 adddev->l2ad_first = B_TRUE;
7227 adddev->l2ad_writing = B_FALSE;
7229 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
7231 * This is a list of all ARC buffers that are still valid on the
7234 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
7235 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
7237 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
7238 refcount_create(&adddev->l2ad_alloc);
7241 * Add device to global list
7243 mutex_enter(&l2arc_dev_mtx);
7244 list_insert_head(l2arc_dev_list, adddev);
7245 atomic_inc_64(&l2arc_ndev);
7246 mutex_exit(&l2arc_dev_mtx);
7250 * Remove a vdev from the L2ARC.
7253 l2arc_remove_vdev(vdev_t *vd)
7255 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
7258 * Find the device by vdev
7260 mutex_enter(&l2arc_dev_mtx);
7261 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
7262 nextdev = list_next(l2arc_dev_list, dev);
7263 if (vd == dev->l2ad_vdev) {
7268 ASSERT3P(remdev, !=, NULL);
7271 * Remove device from global list
7273 list_remove(l2arc_dev_list, remdev);
7274 l2arc_dev_last = NULL; /* may have been invalidated */
7275 atomic_dec_64(&l2arc_ndev);
7276 mutex_exit(&l2arc_dev_mtx);
7279 * Clear all buflists and ARC references. L2ARC device flush.
7281 l2arc_evict(remdev, 0, B_TRUE);
7282 list_destroy(&remdev->l2ad_buflist);
7283 mutex_destroy(&remdev->l2ad_mtx);
7284 refcount_destroy(&remdev->l2ad_alloc);
7285 kmem_free(remdev, sizeof (l2arc_dev_t));
7291 l2arc_thread_exit = 0;
7293 l2arc_writes_sent = 0;
7294 l2arc_writes_done = 0;
7296 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
7297 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
7298 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
7299 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
7301 l2arc_dev_list = &L2ARC_dev_list;
7302 l2arc_free_on_write = &L2ARC_free_on_write;
7303 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
7304 offsetof(l2arc_dev_t, l2ad_node));
7305 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
7306 offsetof(l2arc_data_free_t, l2df_list_node));
7313 * This is called from dmu_fini(), which is called from spa_fini();
7314 * Because of this, we can assume that all l2arc devices have
7315 * already been removed when the pools themselves were removed.
7318 l2arc_do_free_on_write();
7320 mutex_destroy(&l2arc_feed_thr_lock);
7321 cv_destroy(&l2arc_feed_thr_cv);
7322 mutex_destroy(&l2arc_dev_mtx);
7323 mutex_destroy(&l2arc_free_on_write_mtx);
7325 list_destroy(l2arc_dev_list);
7326 list_destroy(l2arc_free_on_write);
7332 if (!(spa_mode_global & FWRITE))
7335 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
7336 TS_RUN, minclsyspri);
7342 if (!(spa_mode_global & FWRITE))
7345 mutex_enter(&l2arc_feed_thr_lock);
7346 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
7347 l2arc_thread_exit = 1;
7348 while (l2arc_thread_exit != 0)
7349 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
7350 mutex_exit(&l2arc_feed_thr_lock);