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 * It as also possible to register a callback which is run when the
103 * arc_meta_limit is reached and no buffers can be safely evicted. In
104 * this case the arc user should drop a reference on some arc buffers so
105 * they can be reclaimed and the arc_meta_limit honored. For example,
106 * when using the ZPL each dentry holds a references on a znode. These
107 * dentries must be pruned before the arc buffer holding the znode can
110 * Note that the majority of the performance stats are manipulated
111 * with atomic operations.
113 * The L2ARC uses the l2ad_mtx on each vdev for the following:
115 * - L2ARC buflist creation
116 * - L2ARC buflist eviction
117 * - L2ARC write completion, which walks L2ARC buflists
118 * - ARC header destruction, as it removes from L2ARC buflists
119 * - ARC header release, as it removes from L2ARC buflists
125 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
126 * This structure can point either to a block that is still in the cache or to
127 * one that is only accessible in an L2 ARC device, or it can provide
128 * information about a block that was recently evicted. If a block is
129 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
130 * information to retrieve it from the L2ARC device. This information is
131 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
132 * that is in this state cannot access the data directly.
134 * Blocks that are actively being referenced or have not been evicted
135 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
136 * the arc_buf_hdr_t that will point to the data block in memory. A block can
137 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
138 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
139 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
141 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
142 * ability to store the physical data (b_pabd) associated with the DVA of the
143 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
144 * it will match its on-disk compression characteristics. This behavior can be
145 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
146 * compressed ARC functionality is disabled, the b_pabd will point to an
147 * uncompressed version of the on-disk data.
149 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
150 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
151 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
152 * consumer. The ARC will provide references to this data and will keep it
153 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
154 * data block and will evict any arc_buf_t that is no longer referenced. The
155 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
156 * "overhead_size" kstat.
158 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
159 * compressed form. The typical case is that consumers will want uncompressed
160 * data, and when that happens a new data buffer is allocated where the data is
161 * decompressed for them to use. Currently the only consumer who wants
162 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
163 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
164 * with the arc_buf_hdr_t.
166 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
167 * first one is owned by a compressed send consumer (and therefore references
168 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
169 * used by any other consumer (and has its own uncompressed copy of the data
184 * | b_buf +------------>+-----------+ arc_buf_t
185 * | b_pabd +-+ |b_next +---->+-----------+
186 * +-----------+ | |-----------| |b_next +-->NULL
187 * | |b_comp = T | +-----------+
188 * | |b_data +-+ |b_comp = F |
189 * | +-----------+ | |b_data +-+
190 * +->+------+ | +-----------+ |
192 * data | |<--------------+ | uncompressed
193 * +------+ compressed, | data
194 * shared +-->+------+
199 * When a consumer reads a block, the ARC must first look to see if the
200 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
201 * arc_buf_t and either copies uncompressed data into a new data buffer from an
202 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
203 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
204 * hdr is compressed and the desired compression characteristics of the
205 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
206 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
207 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
208 * be anywhere in the hdr's list.
210 * The diagram below shows an example of an uncompressed ARC hdr that is
211 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
212 * the last element in the buf list):
224 * | | arc_buf_t (shared)
225 * | b_buf +------------>+---------+ arc_buf_t
226 * | | |b_next +---->+---------+
227 * | b_pabd +-+ |---------| |b_next +-->NULL
228 * +-----------+ | | | +---------+
230 * | +---------+ | |b_data +-+
231 * +->+------+ | +---------+ |
233 * uncompressed | | | |
236 * | uncompressed | | |
239 * +---------------------------------+
241 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
242 * since the physical block is about to be rewritten. The new data contents
243 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
244 * it may compress the data before writing it to disk. The ARC will be called
245 * with the transformed data and will bcopy the transformed on-disk block into
246 * a newly allocated b_pabd. Writes are always done into buffers which have
247 * either been loaned (and hence are new and don't have other readers) or
248 * buffers which have been released (and hence have their own hdr, if there
249 * were originally other readers of the buf's original hdr). This ensures that
250 * the ARC only needs to update a single buf and its hdr after a write occurs.
252 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
253 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
254 * that when compressed ARC is enabled that the L2ARC blocks are identical
255 * to the on-disk block in the main data pool. This provides a significant
256 * advantage since the ARC can leverage the bp's checksum when reading from the
257 * L2ARC to determine if the contents are valid. However, if the compressed
258 * ARC is disabled, then the L2ARC's block must be transformed to look
259 * like the physical block in the main data pool before comparing the
260 * checksum and determining its validity.
265 #include <sys/spa_impl.h>
266 #include <sys/zio_compress.h>
267 #include <sys/zio_checksum.h>
268 #include <sys/zfs_context.h>
270 #include <sys/refcount.h>
271 #include <sys/vdev.h>
272 #include <sys/vdev_impl.h>
273 #include <sys/dsl_pool.h>
274 #include <sys/zio_checksum.h>
275 #include <sys/multilist.h>
278 #include <sys/vmsystm.h>
280 #include <sys/fs/swapnode.h>
282 #include <linux/mm_compat.h>
284 #include <sys/callb.h>
285 #include <sys/kstat.h>
286 #include <sys/dmu_tx.h>
287 #include <zfs_fletcher.h>
288 #include <sys/arc_impl.h>
289 #include <sys/trace_arc.h>
292 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
293 boolean_t arc_watch
= B_FALSE
;
296 static kmutex_t arc_reclaim_lock
;
297 static kcondvar_t arc_reclaim_thread_cv
;
298 static boolean_t arc_reclaim_thread_exit
;
299 static kcondvar_t arc_reclaim_waiters_cv
;
302 * The number of headers to evict in arc_evict_state_impl() before
303 * dropping the sublist lock and evicting from another sublist. A lower
304 * value means we're more likely to evict the "correct" header (i.e. the
305 * oldest header in the arc state), but comes with higher overhead
306 * (i.e. more invocations of arc_evict_state_impl()).
308 int zfs_arc_evict_batch_limit
= 10;
311 * The number of sublists used for each of the arc state lists. If this
312 * is not set to a suitable value by the user, it will be configured to
313 * the number of CPUs on the system in arc_init().
315 int zfs_arc_num_sublists_per_state
= 0;
317 /* number of seconds before growing cache again */
318 static int arc_grow_retry
= 5;
320 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
321 int zfs_arc_overflow_shift
= 8;
323 /* shift of arc_c for calculating both min and max arc_p */
324 static int arc_p_min_shift
= 4;
326 /* log2(fraction of arc to reclaim) */
327 static int arc_shrink_shift
= 7;
330 * log2(fraction of ARC which must be free to allow growing).
331 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
332 * when reading a new block into the ARC, we will evict an equal-sized block
335 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
336 * we will still not allow it to grow.
338 int arc_no_grow_shift
= 5;
342 * minimum lifespan of a prefetch block in clock ticks
343 * (initialized in arc_init())
345 static int arc_min_prefetch_lifespan
;
348 * If this percent of memory is free, don't throttle.
350 int arc_lotsfree_percent
= 10;
355 * The arc has filled available memory and has now warmed up.
357 static boolean_t arc_warm
;
360 * log2 fraction of the zio arena to keep free.
362 int arc_zio_arena_free_shift
= 2;
365 * These tunables are for performance analysis.
367 unsigned long zfs_arc_max
= 0;
368 unsigned long zfs_arc_min
= 0;
369 unsigned long zfs_arc_meta_limit
= 0;
370 unsigned long zfs_arc_meta_min
= 0;
371 unsigned long zfs_arc_dnode_limit
= 0;
372 unsigned long zfs_arc_dnode_reduce_percent
= 10;
373 int zfs_arc_grow_retry
= 0;
374 int zfs_arc_shrink_shift
= 0;
375 int zfs_arc_p_min_shift
= 0;
376 int zfs_arc_average_blocksize
= 8 * 1024; /* 8KB */
378 int zfs_compressed_arc_enabled
= B_TRUE
;
381 * ARC will evict meta buffers that exceed arc_meta_limit. This
382 * tunable make arc_meta_limit adjustable for different workloads.
384 unsigned long zfs_arc_meta_limit_percent
= 75;
387 * Percentage that can be consumed by dnodes of ARC meta buffers.
389 unsigned long zfs_arc_dnode_limit_percent
= 10;
392 * These tunables are Linux specific
394 unsigned long zfs_arc_sys_free
= 0;
395 int zfs_arc_min_prefetch_lifespan
= 0;
396 int zfs_arc_p_aggressive_disable
= 1;
397 int zfs_arc_p_dampener_disable
= 1;
398 int zfs_arc_meta_prune
= 10000;
399 int zfs_arc_meta_strategy
= ARC_STRATEGY_META_BALANCED
;
400 int zfs_arc_meta_adjust_restarts
= 4096;
401 int zfs_arc_lotsfree_percent
= 10;
404 static arc_state_t ARC_anon
;
405 static arc_state_t ARC_mru
;
406 static arc_state_t ARC_mru_ghost
;
407 static arc_state_t ARC_mfu
;
408 static arc_state_t ARC_mfu_ghost
;
409 static arc_state_t ARC_l2c_only
;
411 typedef struct arc_stats
{
412 kstat_named_t arcstat_hits
;
413 kstat_named_t arcstat_misses
;
414 kstat_named_t arcstat_demand_data_hits
;
415 kstat_named_t arcstat_demand_data_misses
;
416 kstat_named_t arcstat_demand_metadata_hits
;
417 kstat_named_t arcstat_demand_metadata_misses
;
418 kstat_named_t arcstat_prefetch_data_hits
;
419 kstat_named_t arcstat_prefetch_data_misses
;
420 kstat_named_t arcstat_prefetch_metadata_hits
;
421 kstat_named_t arcstat_prefetch_metadata_misses
;
422 kstat_named_t arcstat_mru_hits
;
423 kstat_named_t arcstat_mru_ghost_hits
;
424 kstat_named_t arcstat_mfu_hits
;
425 kstat_named_t arcstat_mfu_ghost_hits
;
426 kstat_named_t arcstat_deleted
;
428 * Number of buffers that could not be evicted because the hash lock
429 * was held by another thread. The lock may not necessarily be held
430 * by something using the same buffer, since hash locks are shared
431 * by multiple buffers.
433 kstat_named_t arcstat_mutex_miss
;
435 * Number of buffers skipped because they have I/O in progress, are
436 * indrect prefetch buffers that have not lived long enough, or are
437 * not from the spa we're trying to evict from.
439 kstat_named_t arcstat_evict_skip
;
441 * Number of times arc_evict_state() was unable to evict enough
442 * buffers to reach its target amount.
444 kstat_named_t arcstat_evict_not_enough
;
445 kstat_named_t arcstat_evict_l2_cached
;
446 kstat_named_t arcstat_evict_l2_eligible
;
447 kstat_named_t arcstat_evict_l2_ineligible
;
448 kstat_named_t arcstat_evict_l2_skip
;
449 kstat_named_t arcstat_hash_elements
;
450 kstat_named_t arcstat_hash_elements_max
;
451 kstat_named_t arcstat_hash_collisions
;
452 kstat_named_t arcstat_hash_chains
;
453 kstat_named_t arcstat_hash_chain_max
;
454 kstat_named_t arcstat_p
;
455 kstat_named_t arcstat_c
;
456 kstat_named_t arcstat_c_min
;
457 kstat_named_t arcstat_c_max
;
458 kstat_named_t arcstat_size
;
460 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
461 * Note that the compressed bytes may match the uncompressed bytes
462 * if the block is either not compressed or compressed arc is disabled.
464 kstat_named_t arcstat_compressed_size
;
466 * Uncompressed size of the data stored in b_pabd. If compressed
467 * arc is disabled then this value will be identical to the stat
470 kstat_named_t arcstat_uncompressed_size
;
472 * Number of bytes stored in all the arc_buf_t's. This is classified
473 * as "overhead" since this data is typically short-lived and will
474 * be evicted from the arc when it becomes unreferenced unless the
475 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
476 * values have been set (see comment in dbuf.c for more information).
478 kstat_named_t arcstat_overhead_size
;
480 * Number of bytes consumed by internal ARC structures necessary
481 * for tracking purposes; these structures are not actually
482 * backed by ARC buffers. This includes arc_buf_hdr_t structures
483 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
484 * caches), and arc_buf_t structures (allocated via arc_buf_t
487 kstat_named_t arcstat_hdr_size
;
489 * Number of bytes consumed by ARC buffers of type equal to
490 * ARC_BUFC_DATA. This is generally consumed by buffers backing
491 * on disk user data (e.g. plain file contents).
493 kstat_named_t arcstat_data_size
;
495 * Number of bytes consumed by ARC buffers of type equal to
496 * ARC_BUFC_METADATA. This is generally consumed by buffers
497 * backing on disk data that is used for internal ZFS
498 * structures (e.g. ZAP, dnode, indirect blocks, etc).
500 kstat_named_t arcstat_metadata_size
;
502 * Number of bytes consumed by dmu_buf_impl_t objects.
504 kstat_named_t arcstat_dbuf_size
;
506 * Number of bytes consumed by dnode_t objects.
508 kstat_named_t arcstat_dnode_size
;
510 * Number of bytes consumed by bonus buffers.
512 kstat_named_t arcstat_bonus_size
;
514 * Total number of bytes consumed by ARC buffers residing in the
515 * arc_anon state. This includes *all* buffers in the arc_anon
516 * state; e.g. data, metadata, evictable, and unevictable buffers
517 * are all included in this value.
519 kstat_named_t arcstat_anon_size
;
521 * Number of bytes consumed by ARC buffers that meet the
522 * following criteria: backing buffers of type ARC_BUFC_DATA,
523 * residing in the arc_anon state, and are eligible for eviction
524 * (e.g. have no outstanding holds on the buffer).
526 kstat_named_t arcstat_anon_evictable_data
;
528 * Number of bytes consumed by ARC buffers that meet the
529 * following criteria: backing buffers of type ARC_BUFC_METADATA,
530 * residing in the arc_anon state, and are eligible for eviction
531 * (e.g. have no outstanding holds on the buffer).
533 kstat_named_t arcstat_anon_evictable_metadata
;
535 * Total number of bytes consumed by ARC buffers residing in the
536 * arc_mru state. This includes *all* buffers in the arc_mru
537 * state; e.g. data, metadata, evictable, and unevictable buffers
538 * are all included in this value.
540 kstat_named_t arcstat_mru_size
;
542 * Number of bytes consumed by ARC buffers that meet the
543 * following criteria: backing buffers of type ARC_BUFC_DATA,
544 * residing in the arc_mru state, and are eligible for eviction
545 * (e.g. have no outstanding holds on the buffer).
547 kstat_named_t arcstat_mru_evictable_data
;
549 * Number of bytes consumed by ARC buffers that meet the
550 * following criteria: backing buffers of type ARC_BUFC_METADATA,
551 * residing in the arc_mru state, and are eligible for eviction
552 * (e.g. have no outstanding holds on the buffer).
554 kstat_named_t arcstat_mru_evictable_metadata
;
556 * Total number of bytes that *would have been* consumed by ARC
557 * buffers in the arc_mru_ghost state. The key thing to note
558 * here, is the fact that this size doesn't actually indicate
559 * RAM consumption. The ghost lists only consist of headers and
560 * don't actually have ARC buffers linked off of these headers.
561 * Thus, *if* the headers had associated ARC buffers, these
562 * buffers *would have* consumed this number of bytes.
564 kstat_named_t arcstat_mru_ghost_size
;
566 * Number of bytes that *would have been* consumed by ARC
567 * buffers that are eligible for eviction, of type
568 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
570 kstat_named_t arcstat_mru_ghost_evictable_data
;
572 * Number of bytes that *would have been* consumed by ARC
573 * buffers that are eligible for eviction, of type
574 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
576 kstat_named_t arcstat_mru_ghost_evictable_metadata
;
578 * Total number of bytes consumed by ARC buffers residing in the
579 * arc_mfu state. This includes *all* buffers in the arc_mfu
580 * state; e.g. data, metadata, evictable, and unevictable buffers
581 * are all included in this value.
583 kstat_named_t arcstat_mfu_size
;
585 * Number of bytes consumed by ARC buffers that are eligible for
586 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
589 kstat_named_t arcstat_mfu_evictable_data
;
591 * Number of bytes consumed by ARC buffers that are eligible for
592 * eviction, of type ARC_BUFC_METADATA, and reside in the
595 kstat_named_t arcstat_mfu_evictable_metadata
;
597 * Total number of bytes that *would have been* consumed by ARC
598 * buffers in the arc_mfu_ghost state. See the comment above
599 * arcstat_mru_ghost_size for more details.
601 kstat_named_t arcstat_mfu_ghost_size
;
603 * Number of bytes that *would have been* consumed by ARC
604 * buffers that are eligible for eviction, of type
605 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
607 kstat_named_t arcstat_mfu_ghost_evictable_data
;
609 * Number of bytes that *would have been* consumed by ARC
610 * buffers that are eligible for eviction, of type
611 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
613 kstat_named_t arcstat_mfu_ghost_evictable_metadata
;
614 kstat_named_t arcstat_l2_hits
;
615 kstat_named_t arcstat_l2_misses
;
616 kstat_named_t arcstat_l2_feeds
;
617 kstat_named_t arcstat_l2_rw_clash
;
618 kstat_named_t arcstat_l2_read_bytes
;
619 kstat_named_t arcstat_l2_write_bytes
;
620 kstat_named_t arcstat_l2_writes_sent
;
621 kstat_named_t arcstat_l2_writes_done
;
622 kstat_named_t arcstat_l2_writes_error
;
623 kstat_named_t arcstat_l2_writes_lock_retry
;
624 kstat_named_t arcstat_l2_evict_lock_retry
;
625 kstat_named_t arcstat_l2_evict_reading
;
626 kstat_named_t arcstat_l2_evict_l1cached
;
627 kstat_named_t arcstat_l2_free_on_write
;
628 kstat_named_t arcstat_l2_abort_lowmem
;
629 kstat_named_t arcstat_l2_cksum_bad
;
630 kstat_named_t arcstat_l2_io_error
;
631 kstat_named_t arcstat_l2_size
;
632 kstat_named_t arcstat_l2_asize
;
633 kstat_named_t arcstat_l2_hdr_size
;
634 kstat_named_t arcstat_memory_throttle_count
;
635 kstat_named_t arcstat_memory_direct_count
;
636 kstat_named_t arcstat_memory_indirect_count
;
637 kstat_named_t arcstat_no_grow
;
638 kstat_named_t arcstat_tempreserve
;
639 kstat_named_t arcstat_loaned_bytes
;
640 kstat_named_t arcstat_prune
;
641 kstat_named_t arcstat_meta_used
;
642 kstat_named_t arcstat_meta_limit
;
643 kstat_named_t arcstat_dnode_limit
;
644 kstat_named_t arcstat_meta_max
;
645 kstat_named_t arcstat_meta_min
;
646 kstat_named_t arcstat_sync_wait_for_async
;
647 kstat_named_t arcstat_demand_hit_predictive_prefetch
;
648 kstat_named_t arcstat_need_free
;
649 kstat_named_t arcstat_sys_free
;
652 static arc_stats_t arc_stats
= {
653 { "hits", KSTAT_DATA_UINT64
},
654 { "misses", KSTAT_DATA_UINT64
},
655 { "demand_data_hits", KSTAT_DATA_UINT64
},
656 { "demand_data_misses", KSTAT_DATA_UINT64
},
657 { "demand_metadata_hits", KSTAT_DATA_UINT64
},
658 { "demand_metadata_misses", KSTAT_DATA_UINT64
},
659 { "prefetch_data_hits", KSTAT_DATA_UINT64
},
660 { "prefetch_data_misses", KSTAT_DATA_UINT64
},
661 { "prefetch_metadata_hits", KSTAT_DATA_UINT64
},
662 { "prefetch_metadata_misses", KSTAT_DATA_UINT64
},
663 { "mru_hits", KSTAT_DATA_UINT64
},
664 { "mru_ghost_hits", KSTAT_DATA_UINT64
},
665 { "mfu_hits", KSTAT_DATA_UINT64
},
666 { "mfu_ghost_hits", KSTAT_DATA_UINT64
},
667 { "deleted", KSTAT_DATA_UINT64
},
668 { "mutex_miss", KSTAT_DATA_UINT64
},
669 { "evict_skip", KSTAT_DATA_UINT64
},
670 { "evict_not_enough", KSTAT_DATA_UINT64
},
671 { "evict_l2_cached", KSTAT_DATA_UINT64
},
672 { "evict_l2_eligible", KSTAT_DATA_UINT64
},
673 { "evict_l2_ineligible", KSTAT_DATA_UINT64
},
674 { "evict_l2_skip", KSTAT_DATA_UINT64
},
675 { "hash_elements", KSTAT_DATA_UINT64
},
676 { "hash_elements_max", KSTAT_DATA_UINT64
},
677 { "hash_collisions", KSTAT_DATA_UINT64
},
678 { "hash_chains", KSTAT_DATA_UINT64
},
679 { "hash_chain_max", KSTAT_DATA_UINT64
},
680 { "p", KSTAT_DATA_UINT64
},
681 { "c", KSTAT_DATA_UINT64
},
682 { "c_min", KSTAT_DATA_UINT64
},
683 { "c_max", KSTAT_DATA_UINT64
},
684 { "size", KSTAT_DATA_UINT64
},
685 { "compressed_size", KSTAT_DATA_UINT64
},
686 { "uncompressed_size", KSTAT_DATA_UINT64
},
687 { "overhead_size", KSTAT_DATA_UINT64
},
688 { "hdr_size", KSTAT_DATA_UINT64
},
689 { "data_size", KSTAT_DATA_UINT64
},
690 { "metadata_size", KSTAT_DATA_UINT64
},
691 { "dbuf_size", KSTAT_DATA_UINT64
},
692 { "dnode_size", KSTAT_DATA_UINT64
},
693 { "bonus_size", KSTAT_DATA_UINT64
},
694 { "anon_size", KSTAT_DATA_UINT64
},
695 { "anon_evictable_data", KSTAT_DATA_UINT64
},
696 { "anon_evictable_metadata", KSTAT_DATA_UINT64
},
697 { "mru_size", KSTAT_DATA_UINT64
},
698 { "mru_evictable_data", KSTAT_DATA_UINT64
},
699 { "mru_evictable_metadata", KSTAT_DATA_UINT64
},
700 { "mru_ghost_size", KSTAT_DATA_UINT64
},
701 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64
},
702 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
703 { "mfu_size", KSTAT_DATA_UINT64
},
704 { "mfu_evictable_data", KSTAT_DATA_UINT64
},
705 { "mfu_evictable_metadata", KSTAT_DATA_UINT64
},
706 { "mfu_ghost_size", KSTAT_DATA_UINT64
},
707 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64
},
708 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
709 { "l2_hits", KSTAT_DATA_UINT64
},
710 { "l2_misses", KSTAT_DATA_UINT64
},
711 { "l2_feeds", KSTAT_DATA_UINT64
},
712 { "l2_rw_clash", KSTAT_DATA_UINT64
},
713 { "l2_read_bytes", KSTAT_DATA_UINT64
},
714 { "l2_write_bytes", KSTAT_DATA_UINT64
},
715 { "l2_writes_sent", KSTAT_DATA_UINT64
},
716 { "l2_writes_done", KSTAT_DATA_UINT64
},
717 { "l2_writes_error", KSTAT_DATA_UINT64
},
718 { "l2_writes_lock_retry", KSTAT_DATA_UINT64
},
719 { "l2_evict_lock_retry", KSTAT_DATA_UINT64
},
720 { "l2_evict_reading", KSTAT_DATA_UINT64
},
721 { "l2_evict_l1cached", KSTAT_DATA_UINT64
},
722 { "l2_free_on_write", KSTAT_DATA_UINT64
},
723 { "l2_abort_lowmem", KSTAT_DATA_UINT64
},
724 { "l2_cksum_bad", KSTAT_DATA_UINT64
},
725 { "l2_io_error", KSTAT_DATA_UINT64
},
726 { "l2_size", KSTAT_DATA_UINT64
},
727 { "l2_asize", KSTAT_DATA_UINT64
},
728 { "l2_hdr_size", KSTAT_DATA_UINT64
},
729 { "memory_throttle_count", KSTAT_DATA_UINT64
},
730 { "memory_direct_count", KSTAT_DATA_UINT64
},
731 { "memory_indirect_count", KSTAT_DATA_UINT64
},
732 { "arc_no_grow", KSTAT_DATA_UINT64
},
733 { "arc_tempreserve", KSTAT_DATA_UINT64
},
734 { "arc_loaned_bytes", KSTAT_DATA_UINT64
},
735 { "arc_prune", KSTAT_DATA_UINT64
},
736 { "arc_meta_used", KSTAT_DATA_UINT64
},
737 { "arc_meta_limit", KSTAT_DATA_UINT64
},
738 { "arc_dnode_limit", KSTAT_DATA_UINT64
},
739 { "arc_meta_max", KSTAT_DATA_UINT64
},
740 { "arc_meta_min", KSTAT_DATA_UINT64
},
741 { "sync_wait_for_async", KSTAT_DATA_UINT64
},
742 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64
},
743 { "arc_need_free", KSTAT_DATA_UINT64
},
744 { "arc_sys_free", KSTAT_DATA_UINT64
}
747 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
749 #define ARCSTAT_INCR(stat, val) \
750 atomic_add_64(&arc_stats.stat.value.ui64, (val))
752 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
753 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
755 #define ARCSTAT_MAX(stat, val) { \
757 while ((val) > (m = arc_stats.stat.value.ui64) && \
758 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
762 #define ARCSTAT_MAXSTAT(stat) \
763 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
766 * We define a macro to allow ARC hits/misses to be easily broken down by
767 * two separate conditions, giving a total of four different subtypes for
768 * each of hits and misses (so eight statistics total).
770 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
773 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
775 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
779 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
781 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
786 static arc_state_t
*arc_anon
;
787 static arc_state_t
*arc_mru
;
788 static arc_state_t
*arc_mru_ghost
;
789 static arc_state_t
*arc_mfu
;
790 static arc_state_t
*arc_mfu_ghost
;
791 static arc_state_t
*arc_l2c_only
;
794 * There are several ARC variables that are critical to export as kstats --
795 * but we don't want to have to grovel around in the kstat whenever we wish to
796 * manipulate them. For these variables, we therefore define them to be in
797 * terms of the statistic variable. This assures that we are not introducing
798 * the possibility of inconsistency by having shadow copies of the variables,
799 * while still allowing the code to be readable.
801 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
802 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
803 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
804 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
805 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
806 #define arc_no_grow ARCSTAT(arcstat_no_grow) /* do not grow cache size */
807 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
808 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
809 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
810 #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
811 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
812 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
813 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
814 #define arc_dbuf_size ARCSTAT(arcstat_dbuf_size) /* dbuf metadata */
815 #define arc_dnode_size ARCSTAT(arcstat_dnode_size) /* dnode metadata */
816 #define arc_bonus_size ARCSTAT(arcstat_bonus_size) /* bonus buffer metadata */
817 #define arc_need_free ARCSTAT(arcstat_need_free) /* bytes to be freed */
818 #define arc_sys_free ARCSTAT(arcstat_sys_free) /* target system free bytes */
820 /* compressed size of entire arc */
821 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
822 /* uncompressed size of entire arc */
823 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
824 /* number of bytes in the arc from arc_buf_t's */
825 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
827 static list_t arc_prune_list
;
828 static kmutex_t arc_prune_mtx
;
829 static taskq_t
*arc_prune_taskq
;
831 #define GHOST_STATE(state) \
832 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
833 (state) == arc_l2c_only)
835 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
836 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
837 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
838 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
839 #define HDR_COMPRESSION_ENABLED(hdr) \
840 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
842 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
843 #define HDR_L2_READING(hdr) \
844 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
845 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
846 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
847 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
848 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
849 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
851 #define HDR_ISTYPE_METADATA(hdr) \
852 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
853 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
855 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
856 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
858 /* For storing compression mode in b_flags */
859 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
861 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
862 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
863 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
864 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
866 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
867 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
868 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
874 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
875 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
878 * Hash table routines
881 #define HT_LOCK_ALIGN 64
882 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
887 unsigned char pad
[HT_LOCK_PAD
];
891 #define BUF_LOCKS 8192
892 typedef struct buf_hash_table
{
894 arc_buf_hdr_t
**ht_table
;
895 struct ht_lock ht_locks
[BUF_LOCKS
];
898 static buf_hash_table_t buf_hash_table
;
900 #define BUF_HASH_INDEX(spa, dva, birth) \
901 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
902 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
903 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
904 #define HDR_LOCK(hdr) \
905 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
907 uint64_t zfs_crc64_table
[256];
913 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
914 #define L2ARC_HEADROOM 2 /* num of writes */
917 * If we discover during ARC scan any buffers to be compressed, we boost
918 * our headroom for the next scanning cycle by this percentage multiple.
920 #define L2ARC_HEADROOM_BOOST 200
921 #define L2ARC_FEED_SECS 1 /* caching interval secs */
922 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
925 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
926 * and each of the state has two types: data and metadata.
928 #define L2ARC_FEED_TYPES 4
930 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
931 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
933 /* L2ARC Performance Tunables */
934 unsigned long l2arc_write_max
= L2ARC_WRITE_SIZE
; /* def max write size */
935 unsigned long l2arc_write_boost
= L2ARC_WRITE_SIZE
; /* extra warmup write */
936 unsigned long l2arc_headroom
= L2ARC_HEADROOM
; /* # of dev writes */
937 unsigned long l2arc_headroom_boost
= L2ARC_HEADROOM_BOOST
;
938 unsigned long l2arc_feed_secs
= L2ARC_FEED_SECS
; /* interval seconds */
939 unsigned long l2arc_feed_min_ms
= L2ARC_FEED_MIN_MS
; /* min interval msecs */
940 int l2arc_noprefetch
= B_TRUE
; /* don't cache prefetch bufs */
941 int l2arc_feed_again
= B_TRUE
; /* turbo warmup */
942 int l2arc_norw
= B_FALSE
; /* no reads during writes */
947 static list_t L2ARC_dev_list
; /* device list */
948 static list_t
*l2arc_dev_list
; /* device list pointer */
949 static kmutex_t l2arc_dev_mtx
; /* device list mutex */
950 static l2arc_dev_t
*l2arc_dev_last
; /* last device used */
951 static list_t L2ARC_free_on_write
; /* free after write buf list */
952 static list_t
*l2arc_free_on_write
; /* free after write list ptr */
953 static kmutex_t l2arc_free_on_write_mtx
; /* mutex for list */
954 static uint64_t l2arc_ndev
; /* number of devices */
956 typedef struct l2arc_read_callback
{
957 arc_buf_hdr_t
*l2rcb_hdr
; /* read header */
958 blkptr_t l2rcb_bp
; /* original blkptr */
959 zbookmark_phys_t l2rcb_zb
; /* original bookmark */
960 int l2rcb_flags
; /* original flags */
961 } l2arc_read_callback_t
;
963 typedef struct l2arc_data_free
{
964 /* protected by l2arc_free_on_write_mtx */
967 arc_buf_contents_t l2df_type
;
968 list_node_t l2df_list_node
;
971 static kmutex_t l2arc_feed_thr_lock
;
972 static kcondvar_t l2arc_feed_thr_cv
;
973 static uint8_t l2arc_thread_exit
;
975 static abd_t
*arc_get_data_abd(arc_buf_hdr_t
*, uint64_t, void *);
976 static void *arc_get_data_buf(arc_buf_hdr_t
*, uint64_t, void *);
977 static void arc_get_data_impl(arc_buf_hdr_t
*, uint64_t, void *);
978 static void arc_free_data_abd(arc_buf_hdr_t
*, abd_t
*, uint64_t, void *);
979 static void arc_free_data_buf(arc_buf_hdr_t
*, void *, uint64_t, void *);
980 static void arc_free_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
);
981 static void arc_hdr_free_pabd(arc_buf_hdr_t
*);
982 static void arc_hdr_alloc_pabd(arc_buf_hdr_t
*);
983 static void arc_access(arc_buf_hdr_t
*, kmutex_t
*);
984 static boolean_t
arc_is_overflowing(void);
985 static void arc_buf_watch(arc_buf_t
*);
986 static void arc_tuning_update(void);
987 static void arc_prune_async(int64_t);
988 static uint64_t arc_all_memory(void);
990 static arc_buf_contents_t
arc_buf_type(arc_buf_hdr_t
*);
991 static uint32_t arc_bufc_to_flags(arc_buf_contents_t
);
992 static inline void arc_hdr_set_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
);
993 static inline void arc_hdr_clear_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
);
995 static boolean_t
l2arc_write_eligible(uint64_t, arc_buf_hdr_t
*);
996 static void l2arc_read_done(zio_t
*);
999 buf_hash(uint64_t spa
, const dva_t
*dva
, uint64_t birth
)
1001 uint8_t *vdva
= (uint8_t *)dva
;
1002 uint64_t crc
= -1ULL;
1005 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
1007 for (i
= 0; i
< sizeof (dva_t
); i
++)
1008 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ vdva
[i
]) & 0xFF];
1010 crc
^= (spa
>>8) ^ birth
;
1015 #define HDR_EMPTY(hdr) \
1016 ((hdr)->b_dva.dva_word[0] == 0 && \
1017 (hdr)->b_dva.dva_word[1] == 0)
1019 #define HDR_EQUAL(spa, dva, birth, hdr) \
1020 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1021 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1022 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1025 buf_discard_identity(arc_buf_hdr_t
*hdr
)
1027 hdr
->b_dva
.dva_word
[0] = 0;
1028 hdr
->b_dva
.dva_word
[1] = 0;
1032 static arc_buf_hdr_t
*
1033 buf_hash_find(uint64_t spa
, const blkptr_t
*bp
, kmutex_t
**lockp
)
1035 const dva_t
*dva
= BP_IDENTITY(bp
);
1036 uint64_t birth
= BP_PHYSICAL_BIRTH(bp
);
1037 uint64_t idx
= BUF_HASH_INDEX(spa
, dva
, birth
);
1038 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
1041 mutex_enter(hash_lock
);
1042 for (hdr
= buf_hash_table
.ht_table
[idx
]; hdr
!= NULL
;
1043 hdr
= hdr
->b_hash_next
) {
1044 if (HDR_EQUAL(spa
, dva
, birth
, hdr
)) {
1049 mutex_exit(hash_lock
);
1055 * Insert an entry into the hash table. If there is already an element
1056 * equal to elem in the hash table, then the already existing element
1057 * will be returned and the new element will not be inserted.
1058 * Otherwise returns NULL.
1059 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1061 static arc_buf_hdr_t
*
1062 buf_hash_insert(arc_buf_hdr_t
*hdr
, kmutex_t
**lockp
)
1064 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
1065 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
1066 arc_buf_hdr_t
*fhdr
;
1069 ASSERT(!DVA_IS_EMPTY(&hdr
->b_dva
));
1070 ASSERT(hdr
->b_birth
!= 0);
1071 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
1073 if (lockp
!= NULL
) {
1075 mutex_enter(hash_lock
);
1077 ASSERT(MUTEX_HELD(hash_lock
));
1080 for (fhdr
= buf_hash_table
.ht_table
[idx
], i
= 0; fhdr
!= NULL
;
1081 fhdr
= fhdr
->b_hash_next
, i
++) {
1082 if (HDR_EQUAL(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
, fhdr
))
1086 hdr
->b_hash_next
= buf_hash_table
.ht_table
[idx
];
1087 buf_hash_table
.ht_table
[idx
] = hdr
;
1088 arc_hdr_set_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
1090 /* collect some hash table performance data */
1092 ARCSTAT_BUMP(arcstat_hash_collisions
);
1094 ARCSTAT_BUMP(arcstat_hash_chains
);
1096 ARCSTAT_MAX(arcstat_hash_chain_max
, i
);
1099 ARCSTAT_BUMP(arcstat_hash_elements
);
1100 ARCSTAT_MAXSTAT(arcstat_hash_elements
);
1106 buf_hash_remove(arc_buf_hdr_t
*hdr
)
1108 arc_buf_hdr_t
*fhdr
, **hdrp
;
1109 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
1111 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx
)));
1112 ASSERT(HDR_IN_HASH_TABLE(hdr
));
1114 hdrp
= &buf_hash_table
.ht_table
[idx
];
1115 while ((fhdr
= *hdrp
) != hdr
) {
1116 ASSERT3P(fhdr
, !=, NULL
);
1117 hdrp
= &fhdr
->b_hash_next
;
1119 *hdrp
= hdr
->b_hash_next
;
1120 hdr
->b_hash_next
= NULL
;
1121 arc_hdr_clear_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
1123 /* collect some hash table performance data */
1124 ARCSTAT_BUMPDOWN(arcstat_hash_elements
);
1126 if (buf_hash_table
.ht_table
[idx
] &&
1127 buf_hash_table
.ht_table
[idx
]->b_hash_next
== NULL
)
1128 ARCSTAT_BUMPDOWN(arcstat_hash_chains
);
1132 * Global data structures and functions for the buf kmem cache.
1134 static kmem_cache_t
*hdr_full_cache
;
1135 static kmem_cache_t
*hdr_l2only_cache
;
1136 static kmem_cache_t
*buf_cache
;
1143 #if defined(_KERNEL) && defined(HAVE_SPL)
1145 * Large allocations which do not require contiguous pages
1146 * should be using vmem_free() in the linux kernel\
1148 vmem_free(buf_hash_table
.ht_table
,
1149 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
1151 kmem_free(buf_hash_table
.ht_table
,
1152 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
1154 for (i
= 0; i
< BUF_LOCKS
; i
++)
1155 mutex_destroy(&buf_hash_table
.ht_locks
[i
].ht_lock
);
1156 kmem_cache_destroy(hdr_full_cache
);
1157 kmem_cache_destroy(hdr_l2only_cache
);
1158 kmem_cache_destroy(buf_cache
);
1162 * Constructor callback - called when the cache is empty
1163 * and a new buf is requested.
1167 hdr_full_cons(void *vbuf
, void *unused
, int kmflag
)
1169 arc_buf_hdr_t
*hdr
= vbuf
;
1171 bzero(hdr
, HDR_FULL_SIZE
);
1172 cv_init(&hdr
->b_l1hdr
.b_cv
, NULL
, CV_DEFAULT
, NULL
);
1173 refcount_create(&hdr
->b_l1hdr
.b_refcnt
);
1174 mutex_init(&hdr
->b_l1hdr
.b_freeze_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1175 list_link_init(&hdr
->b_l1hdr
.b_arc_node
);
1176 list_link_init(&hdr
->b_l2hdr
.b_l2node
);
1177 multilist_link_init(&hdr
->b_l1hdr
.b_arc_node
);
1178 arc_space_consume(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1185 hdr_l2only_cons(void *vbuf
, void *unused
, int kmflag
)
1187 arc_buf_hdr_t
*hdr
= vbuf
;
1189 bzero(hdr
, HDR_L2ONLY_SIZE
);
1190 arc_space_consume(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1197 buf_cons(void *vbuf
, void *unused
, int kmflag
)
1199 arc_buf_t
*buf
= vbuf
;
1201 bzero(buf
, sizeof (arc_buf_t
));
1202 mutex_init(&buf
->b_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1203 arc_space_consume(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1209 * Destructor callback - called when a cached buf is
1210 * no longer required.
1214 hdr_full_dest(void *vbuf
, void *unused
)
1216 arc_buf_hdr_t
*hdr
= vbuf
;
1218 ASSERT(HDR_EMPTY(hdr
));
1219 cv_destroy(&hdr
->b_l1hdr
.b_cv
);
1220 refcount_destroy(&hdr
->b_l1hdr
.b_refcnt
);
1221 mutex_destroy(&hdr
->b_l1hdr
.b_freeze_lock
);
1222 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
1223 arc_space_return(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1228 hdr_l2only_dest(void *vbuf
, void *unused
)
1230 ASSERTV(arc_buf_hdr_t
*hdr
= vbuf
);
1232 ASSERT(HDR_EMPTY(hdr
));
1233 arc_space_return(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1238 buf_dest(void *vbuf
, void *unused
)
1240 arc_buf_t
*buf
= vbuf
;
1242 mutex_destroy(&buf
->b_evict_lock
);
1243 arc_space_return(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1247 * Reclaim callback -- invoked when memory is low.
1251 hdr_recl(void *unused
)
1253 dprintf("hdr_recl called\n");
1255 * umem calls the reclaim func when we destroy the buf cache,
1256 * which is after we do arc_fini().
1259 cv_signal(&arc_reclaim_thread_cv
);
1265 uint64_t *ct
= NULL
;
1266 uint64_t hsize
= 1ULL << 12;
1270 * The hash table is big enough to fill all of physical memory
1271 * with an average block size of zfs_arc_average_blocksize (default 8K).
1272 * By default, the table will take up
1273 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1275 while (hsize
* zfs_arc_average_blocksize
< arc_all_memory())
1278 buf_hash_table
.ht_mask
= hsize
- 1;
1279 #if defined(_KERNEL) && defined(HAVE_SPL)
1281 * Large allocations which do not require contiguous pages
1282 * should be using vmem_alloc() in the linux kernel
1284 buf_hash_table
.ht_table
=
1285 vmem_zalloc(hsize
* sizeof (void*), KM_SLEEP
);
1287 buf_hash_table
.ht_table
=
1288 kmem_zalloc(hsize
* sizeof (void*), KM_NOSLEEP
);
1290 if (buf_hash_table
.ht_table
== NULL
) {
1291 ASSERT(hsize
> (1ULL << 8));
1296 hdr_full_cache
= kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE
,
1297 0, hdr_full_cons
, hdr_full_dest
, hdr_recl
, NULL
, NULL
, 0);
1298 hdr_l2only_cache
= kmem_cache_create("arc_buf_hdr_t_l2only",
1299 HDR_L2ONLY_SIZE
, 0, hdr_l2only_cons
, hdr_l2only_dest
, hdr_recl
,
1301 buf_cache
= kmem_cache_create("arc_buf_t", sizeof (arc_buf_t
),
1302 0, buf_cons
, buf_dest
, NULL
, NULL
, NULL
, 0);
1304 for (i
= 0; i
< 256; i
++)
1305 for (ct
= zfs_crc64_table
+ i
, *ct
= i
, j
= 8; j
> 0; j
--)
1306 *ct
= (*ct
>> 1) ^ (-(*ct
& 1) & ZFS_CRC64_POLY
);
1308 for (i
= 0; i
< BUF_LOCKS
; i
++) {
1309 mutex_init(&buf_hash_table
.ht_locks
[i
].ht_lock
,
1310 NULL
, MUTEX_DEFAULT
, NULL
);
1314 #define ARC_MINTIME (hz>>4) /* 62 ms */
1317 * This is the size that the buf occupies in memory. If the buf is compressed,
1318 * it will correspond to the compressed size. You should use this method of
1319 * getting the buf size unless you explicitly need the logical size.
1322 arc_buf_size(arc_buf_t
*buf
)
1324 return (ARC_BUF_COMPRESSED(buf
) ?
1325 HDR_GET_PSIZE(buf
->b_hdr
) : HDR_GET_LSIZE(buf
->b_hdr
));
1329 arc_buf_lsize(arc_buf_t
*buf
)
1331 return (HDR_GET_LSIZE(buf
->b_hdr
));
1335 arc_get_compression(arc_buf_t
*buf
)
1337 return (ARC_BUF_COMPRESSED(buf
) ?
1338 HDR_GET_COMPRESS(buf
->b_hdr
) : ZIO_COMPRESS_OFF
);
1341 static inline boolean_t
1342 arc_buf_is_shared(arc_buf_t
*buf
)
1344 boolean_t shared
= (buf
->b_data
!= NULL
&&
1345 buf
->b_hdr
->b_l1hdr
.b_pabd
!= NULL
&&
1346 abd_is_linear(buf
->b_hdr
->b_l1hdr
.b_pabd
) &&
1347 buf
->b_data
== abd_to_buf(buf
->b_hdr
->b_l1hdr
.b_pabd
));
1348 IMPLY(shared
, HDR_SHARED_DATA(buf
->b_hdr
));
1349 IMPLY(shared
, ARC_BUF_SHARED(buf
));
1350 IMPLY(shared
, ARC_BUF_COMPRESSED(buf
) || ARC_BUF_LAST(buf
));
1353 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1354 * already being shared" requirement prevents us from doing that.
1361 arc_cksum_free(arc_buf_hdr_t
*hdr
)
1363 ASSERT(HDR_HAS_L1HDR(hdr
));
1364 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
1365 if (hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
) {
1366 kmem_free(hdr
->b_l1hdr
.b_freeze_cksum
, sizeof (zio_cksum_t
));
1367 hdr
->b_l1hdr
.b_freeze_cksum
= NULL
;
1369 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1373 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1374 * matches the checksum that is stored in the hdr. If there is no checksum,
1375 * or if the buf is compressed, this is a no-op.
1378 arc_cksum_verify(arc_buf_t
*buf
)
1380 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1383 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1386 if (ARC_BUF_COMPRESSED(buf
)) {
1390 ASSERT(HDR_HAS_L1HDR(hdr
));
1392 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
1393 if (hdr
->b_l1hdr
.b_freeze_cksum
== NULL
|| HDR_IO_ERROR(hdr
)) {
1394 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1398 fletcher_2_native(buf
->b_data
, arc_buf_size(buf
), NULL
, &zc
);
1399 if (!ZIO_CHECKSUM_EQUAL(*hdr
->b_l1hdr
.b_freeze_cksum
, zc
))
1400 panic("buffer modified while frozen!");
1401 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1405 arc_cksum_is_equal(arc_buf_hdr_t
*hdr
, zio_t
*zio
)
1407 enum zio_compress compress
= BP_GET_COMPRESS(zio
->io_bp
);
1408 boolean_t valid_cksum
;
1410 ASSERT(!BP_IS_EMBEDDED(zio
->io_bp
));
1411 VERIFY3U(BP_GET_PSIZE(zio
->io_bp
), ==, HDR_GET_PSIZE(hdr
));
1414 * We rely on the blkptr's checksum to determine if the block
1415 * is valid or not. When compressed arc is enabled, the l2arc
1416 * writes the block to the l2arc just as it appears in the pool.
1417 * This allows us to use the blkptr's checksum to validate the
1418 * data that we just read off of the l2arc without having to store
1419 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1420 * arc is disabled, then the data written to the l2arc is always
1421 * uncompressed and won't match the block as it exists in the main
1422 * pool. When this is the case, we must first compress it if it is
1423 * compressed on the main pool before we can validate the checksum.
1425 if (!HDR_COMPRESSION_ENABLED(hdr
) && compress
!= ZIO_COMPRESS_OFF
) {
1429 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
1431 cbuf
= zio_buf_alloc(HDR_GET_PSIZE(hdr
));
1432 lsize
= HDR_GET_LSIZE(hdr
);
1433 csize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
);
1435 ASSERT3U(csize
, <=, HDR_GET_PSIZE(hdr
));
1436 if (csize
< HDR_GET_PSIZE(hdr
)) {
1438 * Compressed blocks are always a multiple of the
1439 * smallest ashift in the pool. Ideally, we would
1440 * like to round up the csize to the next
1441 * spa_min_ashift but that value may have changed
1442 * since the block was last written. Instead,
1443 * we rely on the fact that the hdr's psize
1444 * was set to the psize of the block when it was
1445 * last written. We set the csize to that value
1446 * and zero out any part that should not contain
1449 bzero((char *)cbuf
+ csize
, HDR_GET_PSIZE(hdr
) - csize
);
1450 csize
= HDR_GET_PSIZE(hdr
);
1452 zio_push_transform(zio
, cbuf
, csize
, HDR_GET_PSIZE(hdr
), NULL
);
1456 * Block pointers always store the checksum for the logical data.
1457 * If the block pointer has the gang bit set, then the checksum
1458 * it represents is for the reconstituted data and not for an
1459 * individual gang member. The zio pipeline, however, must be able to
1460 * determine the checksum of each of the gang constituents so it
1461 * treats the checksum comparison differently than what we need
1462 * for l2arc blocks. This prevents us from using the
1463 * zio_checksum_error() interface directly. Instead we must call the
1464 * zio_checksum_error_impl() so that we can ensure the checksum is
1465 * generated using the correct checksum algorithm and accounts for the
1466 * logical I/O size and not just a gang fragment.
1468 valid_cksum
= (zio_checksum_error_impl(zio
->io_spa
, zio
->io_bp
,
1469 BP_GET_CHECKSUM(zio
->io_bp
), zio
->io_abd
, zio
->io_size
,
1470 zio
->io_offset
, NULL
) == 0);
1471 zio_pop_transforms(zio
);
1472 return (valid_cksum
);
1476 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1477 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1478 * isn't modified later on. If buf is compressed or there is already a checksum
1479 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1482 arc_cksum_compute(arc_buf_t
*buf
)
1484 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1486 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1489 ASSERT(HDR_HAS_L1HDR(hdr
));
1491 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1492 if (hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
) {
1493 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1495 } else if (ARC_BUF_COMPRESSED(buf
)) {
1496 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1500 ASSERT(!ARC_BUF_COMPRESSED(buf
));
1501 hdr
->b_l1hdr
.b_freeze_cksum
= kmem_alloc(sizeof (zio_cksum_t
),
1503 fletcher_2_native(buf
->b_data
, arc_buf_size(buf
), NULL
,
1504 hdr
->b_l1hdr
.b_freeze_cksum
);
1505 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1511 arc_buf_sigsegv(int sig
, siginfo_t
*si
, void *unused
)
1513 panic("Got SIGSEGV at address: 0x%lx\n", (long) si
->si_addr
);
1519 arc_buf_unwatch(arc_buf_t
*buf
)
1523 ASSERT0(mprotect(buf
->b_data
, HDR_GET_LSIZE(buf
->b_hdr
),
1524 PROT_READ
| PROT_WRITE
));
1531 arc_buf_watch(arc_buf_t
*buf
)
1535 ASSERT0(mprotect(buf
->b_data
, arc_buf_size(buf
),
1540 static arc_buf_contents_t
1541 arc_buf_type(arc_buf_hdr_t
*hdr
)
1543 arc_buf_contents_t type
;
1544 if (HDR_ISTYPE_METADATA(hdr
)) {
1545 type
= ARC_BUFC_METADATA
;
1547 type
= ARC_BUFC_DATA
;
1549 VERIFY3U(hdr
->b_type
, ==, type
);
1554 arc_is_metadata(arc_buf_t
*buf
)
1556 return (HDR_ISTYPE_METADATA(buf
->b_hdr
) != 0);
1560 arc_bufc_to_flags(arc_buf_contents_t type
)
1564 /* metadata field is 0 if buffer contains normal data */
1566 case ARC_BUFC_METADATA
:
1567 return (ARC_FLAG_BUFC_METADATA
);
1571 panic("undefined ARC buffer type!");
1572 return ((uint32_t)-1);
1576 arc_buf_thaw(arc_buf_t
*buf
)
1578 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1580 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
1581 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
1583 arc_cksum_verify(buf
);
1586 * Compressed buffers do not manipulate the b_freeze_cksum or
1587 * allocate b_thawed.
1589 if (ARC_BUF_COMPRESSED(buf
)) {
1593 ASSERT(HDR_HAS_L1HDR(hdr
));
1594 arc_cksum_free(hdr
);
1595 arc_buf_unwatch(buf
);
1599 arc_buf_freeze(arc_buf_t
*buf
)
1601 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1602 kmutex_t
*hash_lock
;
1604 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1607 if (ARC_BUF_COMPRESSED(buf
)) {
1611 hash_lock
= HDR_LOCK(hdr
);
1612 mutex_enter(hash_lock
);
1614 ASSERT(HDR_HAS_L1HDR(hdr
));
1615 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
||
1616 hdr
->b_l1hdr
.b_state
== arc_anon
);
1617 arc_cksum_compute(buf
);
1618 mutex_exit(hash_lock
);
1622 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1623 * the following functions should be used to ensure that the flags are
1624 * updated in a thread-safe way. When manipulating the flags either
1625 * the hash_lock must be held or the hdr must be undiscoverable. This
1626 * ensures that we're not racing with any other threads when updating
1630 arc_hdr_set_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
)
1632 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1633 hdr
->b_flags
|= flags
;
1637 arc_hdr_clear_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
)
1639 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1640 hdr
->b_flags
&= ~flags
;
1644 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1645 * done in a special way since we have to clear and set bits
1646 * at the same time. Consumers that wish to set the compression bits
1647 * must use this function to ensure that the flags are updated in
1648 * thread-safe manner.
1651 arc_hdr_set_compress(arc_buf_hdr_t
*hdr
, enum zio_compress cmp
)
1653 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1656 * Holes and embedded blocks will always have a psize = 0 so
1657 * we ignore the compression of the blkptr and set the
1658 * want to uncompress them. Mark them as uncompressed.
1660 if (!zfs_compressed_arc_enabled
|| HDR_GET_PSIZE(hdr
) == 0) {
1661 arc_hdr_clear_flags(hdr
, ARC_FLAG_COMPRESSED_ARC
);
1662 HDR_SET_COMPRESS(hdr
, ZIO_COMPRESS_OFF
);
1663 ASSERT(!HDR_COMPRESSION_ENABLED(hdr
));
1664 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
1666 arc_hdr_set_flags(hdr
, ARC_FLAG_COMPRESSED_ARC
);
1667 HDR_SET_COMPRESS(hdr
, cmp
);
1668 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, cmp
);
1669 ASSERT(HDR_COMPRESSION_ENABLED(hdr
));
1674 * Looks for another buf on the same hdr which has the data decompressed, copies
1675 * from it, and returns true. If no such buf exists, returns false.
1678 arc_buf_try_copy_decompressed_data(arc_buf_t
*buf
)
1680 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1682 boolean_t copied
= B_FALSE
;
1684 ASSERT(HDR_HAS_L1HDR(hdr
));
1685 ASSERT3P(buf
->b_data
, !=, NULL
);
1686 ASSERT(!ARC_BUF_COMPRESSED(buf
));
1688 for (from
= hdr
->b_l1hdr
.b_buf
; from
!= NULL
;
1689 from
= from
->b_next
) {
1690 /* can't use our own data buffer */
1695 if (!ARC_BUF_COMPRESSED(from
)) {
1696 bcopy(from
->b_data
, buf
->b_data
, arc_buf_size(buf
));
1703 * There were no decompressed bufs, so there should not be a
1704 * checksum on the hdr either.
1706 EQUIV(!copied
, hdr
->b_l1hdr
.b_freeze_cksum
== NULL
);
1712 * Given a buf that has a data buffer attached to it, this function will
1713 * efficiently fill the buf with data of the specified compression setting from
1714 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1715 * are already sharing a data buf, no copy is performed.
1717 * If the buf is marked as compressed but uncompressed data was requested, this
1718 * will allocate a new data buffer for the buf, remove that flag, and fill the
1719 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1720 * uncompressed data, and (since we haven't added support for it yet) if you
1721 * want compressed data your buf must already be marked as compressed and have
1722 * the correct-sized data buffer.
1725 arc_buf_fill(arc_buf_t
*buf
, boolean_t compressed
)
1727 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1728 boolean_t hdr_compressed
= (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
);
1729 dmu_object_byteswap_t bswap
= hdr
->b_l1hdr
.b_byteswap
;
1731 ASSERT3P(buf
->b_data
, !=, NULL
);
1732 IMPLY(compressed
, hdr_compressed
);
1733 IMPLY(compressed
, ARC_BUF_COMPRESSED(buf
));
1735 if (hdr_compressed
== compressed
) {
1736 if (!arc_buf_is_shared(buf
)) {
1737 abd_copy_to_buf(buf
->b_data
, hdr
->b_l1hdr
.b_pabd
,
1741 ASSERT(hdr_compressed
);
1742 ASSERT(!compressed
);
1743 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, HDR_GET_PSIZE(hdr
));
1746 * If the buf is sharing its data with the hdr, unlink it and
1747 * allocate a new data buffer for the buf.
1749 if (arc_buf_is_shared(buf
)) {
1750 ASSERT(ARC_BUF_COMPRESSED(buf
));
1752 /* We need to give the buf it's own b_data */
1753 buf
->b_flags
&= ~ARC_BUF_FLAG_SHARED
;
1755 arc_get_data_buf(hdr
, HDR_GET_LSIZE(hdr
), buf
);
1756 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
1758 /* Previously overhead was 0; just add new overhead */
1759 ARCSTAT_INCR(arcstat_overhead_size
, HDR_GET_LSIZE(hdr
));
1760 } else if (ARC_BUF_COMPRESSED(buf
)) {
1761 /* We need to reallocate the buf's b_data */
1762 arc_free_data_buf(hdr
, buf
->b_data
, HDR_GET_PSIZE(hdr
),
1765 arc_get_data_buf(hdr
, HDR_GET_LSIZE(hdr
), buf
);
1767 /* We increased the size of b_data; update overhead */
1768 ARCSTAT_INCR(arcstat_overhead_size
,
1769 HDR_GET_LSIZE(hdr
) - HDR_GET_PSIZE(hdr
));
1773 * Regardless of the buf's previous compression settings, it
1774 * should not be compressed at the end of this function.
1776 buf
->b_flags
&= ~ARC_BUF_FLAG_COMPRESSED
;
1779 * Try copying the data from another buf which already has a
1780 * decompressed version. If that's not possible, it's time to
1781 * bite the bullet and decompress the data from the hdr.
1783 if (arc_buf_try_copy_decompressed_data(buf
)) {
1784 /* Skip byteswapping and checksumming (already done) */
1785 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, !=, NULL
);
1788 int error
= zio_decompress_data(HDR_GET_COMPRESS(hdr
),
1789 hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
1790 HDR_GET_PSIZE(hdr
), HDR_GET_LSIZE(hdr
));
1793 * Absent hardware errors or software bugs, this should
1794 * be impossible, but log it anyway so we can debug it.
1798 "hdr %p, compress %d, psize %d, lsize %d",
1799 hdr
, HDR_GET_COMPRESS(hdr
),
1800 HDR_GET_PSIZE(hdr
), HDR_GET_LSIZE(hdr
));
1801 return (SET_ERROR(EIO
));
1806 /* Byteswap the buf's data if necessary */
1807 if (bswap
!= DMU_BSWAP_NUMFUNCS
) {
1808 ASSERT(!HDR_SHARED_DATA(hdr
));
1809 ASSERT3U(bswap
, <, DMU_BSWAP_NUMFUNCS
);
1810 dmu_ot_byteswap
[bswap
].ob_func(buf
->b_data
, HDR_GET_LSIZE(hdr
));
1813 /* Compute the hdr's checksum if necessary */
1814 arc_cksum_compute(buf
);
1820 arc_decompress(arc_buf_t
*buf
)
1822 return (arc_buf_fill(buf
, B_FALSE
));
1826 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1829 arc_hdr_size(arc_buf_hdr_t
*hdr
)
1833 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
&&
1834 HDR_GET_PSIZE(hdr
) > 0) {
1835 size
= HDR_GET_PSIZE(hdr
);
1837 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, 0);
1838 size
= HDR_GET_LSIZE(hdr
);
1844 * Increment the amount of evictable space in the arc_state_t's refcount.
1845 * We account for the space used by the hdr and the arc buf individually
1846 * so that we can add and remove them from the refcount individually.
1849 arc_evictable_space_increment(arc_buf_hdr_t
*hdr
, arc_state_t
*state
)
1851 arc_buf_contents_t type
= arc_buf_type(hdr
);
1854 ASSERT(HDR_HAS_L1HDR(hdr
));
1856 if (GHOST_STATE(state
)) {
1857 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
1858 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1859 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
1860 (void) refcount_add_many(&state
->arcs_esize
[type
],
1861 HDR_GET_LSIZE(hdr
), hdr
);
1865 ASSERT(!GHOST_STATE(state
));
1866 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
1867 (void) refcount_add_many(&state
->arcs_esize
[type
],
1868 arc_hdr_size(hdr
), hdr
);
1870 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
; buf
= buf
->b_next
) {
1871 if (arc_buf_is_shared(buf
))
1873 (void) refcount_add_many(&state
->arcs_esize
[type
],
1874 arc_buf_size(buf
), buf
);
1879 * Decrement the amount of evictable space in the arc_state_t's refcount.
1880 * We account for the space used by the hdr and the arc buf individually
1881 * so that we can add and remove them from the refcount individually.
1884 arc_evictable_space_decrement(arc_buf_hdr_t
*hdr
, arc_state_t
*state
)
1886 arc_buf_contents_t type
= arc_buf_type(hdr
);
1889 ASSERT(HDR_HAS_L1HDR(hdr
));
1891 if (GHOST_STATE(state
)) {
1892 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
1893 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1894 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
1895 (void) refcount_remove_many(&state
->arcs_esize
[type
],
1896 HDR_GET_LSIZE(hdr
), hdr
);
1900 ASSERT(!GHOST_STATE(state
));
1901 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
1902 (void) refcount_remove_many(&state
->arcs_esize
[type
],
1903 arc_hdr_size(hdr
), hdr
);
1905 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
; buf
= buf
->b_next
) {
1906 if (arc_buf_is_shared(buf
))
1908 (void) refcount_remove_many(&state
->arcs_esize
[type
],
1909 arc_buf_size(buf
), buf
);
1914 * Add a reference to this hdr indicating that someone is actively
1915 * referencing that memory. When the refcount transitions from 0 to 1,
1916 * we remove it from the respective arc_state_t list to indicate that
1917 * it is not evictable.
1920 add_reference(arc_buf_hdr_t
*hdr
, void *tag
)
1924 ASSERT(HDR_HAS_L1HDR(hdr
));
1925 if (!MUTEX_HELD(HDR_LOCK(hdr
))) {
1926 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
1927 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
1928 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1931 state
= hdr
->b_l1hdr
.b_state
;
1933 if ((refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
) == 1) &&
1934 (state
!= arc_anon
)) {
1935 /* We don't use the L2-only state list. */
1936 if (state
!= arc_l2c_only
) {
1937 multilist_remove(&state
->arcs_list
[arc_buf_type(hdr
)],
1939 arc_evictable_space_decrement(hdr
, state
);
1941 /* remove the prefetch flag if we get a reference */
1942 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
1947 * Remove a reference from this hdr. When the reference transitions from
1948 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
1949 * list making it eligible for eviction.
1952 remove_reference(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, void *tag
)
1955 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
1957 ASSERT(HDR_HAS_L1HDR(hdr
));
1958 ASSERT(state
== arc_anon
|| MUTEX_HELD(hash_lock
));
1959 ASSERT(!GHOST_STATE(state
));
1962 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1963 * check to prevent usage of the arc_l2c_only list.
1965 if (((cnt
= refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
)) == 0) &&
1966 (state
!= arc_anon
)) {
1967 multilist_insert(&state
->arcs_list
[arc_buf_type(hdr
)], hdr
);
1968 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, >, 0);
1969 arc_evictable_space_increment(hdr
, state
);
1975 * Returns detailed information about a specific arc buffer. When the
1976 * state_index argument is set the function will calculate the arc header
1977 * list position for its arc state. Since this requires a linear traversal
1978 * callers are strongly encourage not to do this. However, it can be helpful
1979 * for targeted analysis so the functionality is provided.
1982 arc_buf_info(arc_buf_t
*ab
, arc_buf_info_t
*abi
, int state_index
)
1984 arc_buf_hdr_t
*hdr
= ab
->b_hdr
;
1985 l1arc_buf_hdr_t
*l1hdr
= NULL
;
1986 l2arc_buf_hdr_t
*l2hdr
= NULL
;
1987 arc_state_t
*state
= NULL
;
1989 memset(abi
, 0, sizeof (arc_buf_info_t
));
1994 abi
->abi_flags
= hdr
->b_flags
;
1996 if (HDR_HAS_L1HDR(hdr
)) {
1997 l1hdr
= &hdr
->b_l1hdr
;
1998 state
= l1hdr
->b_state
;
2000 if (HDR_HAS_L2HDR(hdr
))
2001 l2hdr
= &hdr
->b_l2hdr
;
2004 abi
->abi_bufcnt
= l1hdr
->b_bufcnt
;
2005 abi
->abi_access
= l1hdr
->b_arc_access
;
2006 abi
->abi_mru_hits
= l1hdr
->b_mru_hits
;
2007 abi
->abi_mru_ghost_hits
= l1hdr
->b_mru_ghost_hits
;
2008 abi
->abi_mfu_hits
= l1hdr
->b_mfu_hits
;
2009 abi
->abi_mfu_ghost_hits
= l1hdr
->b_mfu_ghost_hits
;
2010 abi
->abi_holds
= refcount_count(&l1hdr
->b_refcnt
);
2014 abi
->abi_l2arc_dattr
= l2hdr
->b_daddr
;
2015 abi
->abi_l2arc_hits
= l2hdr
->b_hits
;
2018 abi
->abi_state_type
= state
? state
->arcs_state
: ARC_STATE_ANON
;
2019 abi
->abi_state_contents
= arc_buf_type(hdr
);
2020 abi
->abi_size
= arc_hdr_size(hdr
);
2024 * Move the supplied buffer to the indicated state. The hash lock
2025 * for the buffer must be held by the caller.
2028 arc_change_state(arc_state_t
*new_state
, arc_buf_hdr_t
*hdr
,
2029 kmutex_t
*hash_lock
)
2031 arc_state_t
*old_state
;
2034 boolean_t update_old
, update_new
;
2035 arc_buf_contents_t buftype
= arc_buf_type(hdr
);
2038 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2039 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2040 * L1 hdr doesn't always exist when we change state to arc_anon before
2041 * destroying a header, in which case reallocating to add the L1 hdr is
2044 if (HDR_HAS_L1HDR(hdr
)) {
2045 old_state
= hdr
->b_l1hdr
.b_state
;
2046 refcnt
= refcount_count(&hdr
->b_l1hdr
.b_refcnt
);
2047 bufcnt
= hdr
->b_l1hdr
.b_bufcnt
;
2048 update_old
= (bufcnt
> 0 || hdr
->b_l1hdr
.b_pabd
!= NULL
);
2050 old_state
= arc_l2c_only
;
2053 update_old
= B_FALSE
;
2055 update_new
= update_old
;
2057 ASSERT(MUTEX_HELD(hash_lock
));
2058 ASSERT3P(new_state
, !=, old_state
);
2059 ASSERT(!GHOST_STATE(new_state
) || bufcnt
== 0);
2060 ASSERT(old_state
!= arc_anon
|| bufcnt
<= 1);
2063 * If this buffer is evictable, transfer it from the
2064 * old state list to the new state list.
2067 if (old_state
!= arc_anon
&& old_state
!= arc_l2c_only
) {
2068 ASSERT(HDR_HAS_L1HDR(hdr
));
2069 multilist_remove(&old_state
->arcs_list
[buftype
], hdr
);
2071 if (GHOST_STATE(old_state
)) {
2073 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2074 update_old
= B_TRUE
;
2076 arc_evictable_space_decrement(hdr
, old_state
);
2078 if (new_state
!= arc_anon
&& new_state
!= arc_l2c_only
) {
2080 * An L1 header always exists here, since if we're
2081 * moving to some L1-cached state (i.e. not l2c_only or
2082 * anonymous), we realloc the header to add an L1hdr
2085 ASSERT(HDR_HAS_L1HDR(hdr
));
2086 multilist_insert(&new_state
->arcs_list
[buftype
], hdr
);
2088 if (GHOST_STATE(new_state
)) {
2090 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2091 update_new
= B_TRUE
;
2093 arc_evictable_space_increment(hdr
, new_state
);
2097 ASSERT(!HDR_EMPTY(hdr
));
2098 if (new_state
== arc_anon
&& HDR_IN_HASH_TABLE(hdr
))
2099 buf_hash_remove(hdr
);
2101 /* adjust state sizes (ignore arc_l2c_only) */
2103 if (update_new
&& new_state
!= arc_l2c_only
) {
2104 ASSERT(HDR_HAS_L1HDR(hdr
));
2105 if (GHOST_STATE(new_state
)) {
2109 * When moving a header to a ghost state, we first
2110 * remove all arc buffers. Thus, we'll have a
2111 * bufcnt of zero, and no arc buffer to use for
2112 * the reference. As a result, we use the arc
2113 * header pointer for the reference.
2115 (void) refcount_add_many(&new_state
->arcs_size
,
2116 HDR_GET_LSIZE(hdr
), hdr
);
2117 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2120 uint32_t buffers
= 0;
2123 * Each individual buffer holds a unique reference,
2124 * thus we must remove each of these references one
2127 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2128 buf
= buf
->b_next
) {
2129 ASSERT3U(bufcnt
, !=, 0);
2133 * When the arc_buf_t is sharing the data
2134 * block with the hdr, the owner of the
2135 * reference belongs to the hdr. Only
2136 * add to the refcount if the arc_buf_t is
2139 if (arc_buf_is_shared(buf
))
2142 (void) refcount_add_many(&new_state
->arcs_size
,
2143 arc_buf_size(buf
), buf
);
2145 ASSERT3U(bufcnt
, ==, buffers
);
2147 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
2148 (void) refcount_add_many(&new_state
->arcs_size
,
2149 arc_hdr_size(hdr
), hdr
);
2151 ASSERT(GHOST_STATE(old_state
));
2156 if (update_old
&& old_state
!= arc_l2c_only
) {
2157 ASSERT(HDR_HAS_L1HDR(hdr
));
2158 if (GHOST_STATE(old_state
)) {
2160 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2163 * When moving a header off of a ghost state,
2164 * the header will not contain any arc buffers.
2165 * We use the arc header pointer for the reference
2166 * which is exactly what we did when we put the
2167 * header on the ghost state.
2170 (void) refcount_remove_many(&old_state
->arcs_size
,
2171 HDR_GET_LSIZE(hdr
), hdr
);
2174 uint32_t buffers
= 0;
2177 * Each individual buffer holds a unique reference,
2178 * thus we must remove each of these references one
2181 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2182 buf
= buf
->b_next
) {
2183 ASSERT3U(bufcnt
, !=, 0);
2187 * When the arc_buf_t is sharing the data
2188 * block with the hdr, the owner of the
2189 * reference belongs to the hdr. Only
2190 * add to the refcount if the arc_buf_t is
2193 if (arc_buf_is_shared(buf
))
2196 (void) refcount_remove_many(
2197 &old_state
->arcs_size
, arc_buf_size(buf
),
2200 ASSERT3U(bufcnt
, ==, buffers
);
2201 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2202 (void) refcount_remove_many(
2203 &old_state
->arcs_size
, arc_hdr_size(hdr
), hdr
);
2207 if (HDR_HAS_L1HDR(hdr
))
2208 hdr
->b_l1hdr
.b_state
= new_state
;
2211 * L2 headers should never be on the L2 state list since they don't
2212 * have L1 headers allocated.
2214 ASSERT(multilist_is_empty(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]) &&
2215 multilist_is_empty(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]));
2219 arc_space_consume(uint64_t space
, arc_space_type_t type
)
2221 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
2226 case ARC_SPACE_DATA
:
2227 ARCSTAT_INCR(arcstat_data_size
, space
);
2229 case ARC_SPACE_META
:
2230 ARCSTAT_INCR(arcstat_metadata_size
, space
);
2232 case ARC_SPACE_BONUS
:
2233 ARCSTAT_INCR(arcstat_bonus_size
, space
);
2235 case ARC_SPACE_DNODE
:
2236 ARCSTAT_INCR(arcstat_dnode_size
, space
);
2238 case ARC_SPACE_DBUF
:
2239 ARCSTAT_INCR(arcstat_dbuf_size
, space
);
2241 case ARC_SPACE_HDRS
:
2242 ARCSTAT_INCR(arcstat_hdr_size
, space
);
2244 case ARC_SPACE_L2HDRS
:
2245 ARCSTAT_INCR(arcstat_l2_hdr_size
, space
);
2249 if (type
!= ARC_SPACE_DATA
)
2250 ARCSTAT_INCR(arcstat_meta_used
, space
);
2252 atomic_add_64(&arc_size
, space
);
2256 arc_space_return(uint64_t space
, arc_space_type_t type
)
2258 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
2263 case ARC_SPACE_DATA
:
2264 ARCSTAT_INCR(arcstat_data_size
, -space
);
2266 case ARC_SPACE_META
:
2267 ARCSTAT_INCR(arcstat_metadata_size
, -space
);
2269 case ARC_SPACE_BONUS
:
2270 ARCSTAT_INCR(arcstat_bonus_size
, -space
);
2272 case ARC_SPACE_DNODE
:
2273 ARCSTAT_INCR(arcstat_dnode_size
, -space
);
2275 case ARC_SPACE_DBUF
:
2276 ARCSTAT_INCR(arcstat_dbuf_size
, -space
);
2278 case ARC_SPACE_HDRS
:
2279 ARCSTAT_INCR(arcstat_hdr_size
, -space
);
2281 case ARC_SPACE_L2HDRS
:
2282 ARCSTAT_INCR(arcstat_l2_hdr_size
, -space
);
2286 if (type
!= ARC_SPACE_DATA
) {
2287 ASSERT(arc_meta_used
>= space
);
2288 if (arc_meta_max
< arc_meta_used
)
2289 arc_meta_max
= arc_meta_used
;
2290 ARCSTAT_INCR(arcstat_meta_used
, -space
);
2293 ASSERT(arc_size
>= space
);
2294 atomic_add_64(&arc_size
, -space
);
2298 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2299 * with the hdr's b_pabd.
2302 arc_can_share(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2304 boolean_t hdr_compressed
, buf_compressed
;
2306 * The criteria for sharing a hdr's data are:
2307 * 1. the hdr's compression matches the buf's compression
2308 * 2. the hdr doesn't need to be byteswapped
2309 * 3. the hdr isn't already being shared
2310 * 4. the buf is either compressed or it is the last buf in the hdr list
2312 * Criterion #4 maintains the invariant that shared uncompressed
2313 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2314 * might ask, "if a compressed buf is allocated first, won't that be the
2315 * last thing in the list?", but in that case it's impossible to create
2316 * a shared uncompressed buf anyway (because the hdr must be compressed
2317 * to have the compressed buf). You might also think that #3 is
2318 * sufficient to make this guarantee, however it's possible
2319 * (specifically in the rare L2ARC write race mentioned in
2320 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2321 * is sharable, but wasn't at the time of its allocation. Rather than
2322 * allow a new shared uncompressed buf to be created and then shuffle
2323 * the list around to make it the last element, this simply disallows
2324 * sharing if the new buf isn't the first to be added.
2326 ASSERT3P(buf
->b_hdr
, ==, hdr
);
2327 hdr_compressed
= HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
;
2328 buf_compressed
= ARC_BUF_COMPRESSED(buf
) != 0;
2329 return (buf_compressed
== hdr_compressed
&&
2330 hdr
->b_l1hdr
.b_byteswap
== DMU_BSWAP_NUMFUNCS
&&
2331 !HDR_SHARED_DATA(hdr
) &&
2332 (ARC_BUF_LAST(buf
) || ARC_BUF_COMPRESSED(buf
)));
2336 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2337 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2338 * copy was made successfully, or an error code otherwise.
2341 arc_buf_alloc_impl(arc_buf_hdr_t
*hdr
, void *tag
, boolean_t compressed
,
2342 boolean_t fill
, arc_buf_t
**ret
)
2345 boolean_t can_share
;
2347 ASSERT(HDR_HAS_L1HDR(hdr
));
2348 ASSERT3U(HDR_GET_LSIZE(hdr
), >, 0);
2349 VERIFY(hdr
->b_type
== ARC_BUFC_DATA
||
2350 hdr
->b_type
== ARC_BUFC_METADATA
);
2351 ASSERT3P(ret
, !=, NULL
);
2352 ASSERT3P(*ret
, ==, NULL
);
2354 hdr
->b_l1hdr
.b_mru_hits
= 0;
2355 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
2356 hdr
->b_l1hdr
.b_mfu_hits
= 0;
2357 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
2358 hdr
->b_l1hdr
.b_l2_hits
= 0;
2360 buf
= *ret
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
2363 buf
->b_next
= hdr
->b_l1hdr
.b_buf
;
2366 add_reference(hdr
, tag
);
2369 * We're about to change the hdr's b_flags. We must either
2370 * hold the hash_lock or be undiscoverable.
2372 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2375 * Only honor requests for compressed bufs if the hdr is actually
2378 if (compressed
&& HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
)
2379 buf
->b_flags
|= ARC_BUF_FLAG_COMPRESSED
;
2382 * Although the ARC should handle it correctly, levels above the ARC
2383 * should prevent us from having multiple compressed bufs off the same
2384 * hdr. To ensure we notice it if this behavior changes, we assert this
2385 * here the best we can.
2387 IMPLY(ARC_BUF_COMPRESSED(buf
), !HDR_SHARED_DATA(hdr
));
2390 * If the hdr's data can be shared then we share the data buffer and
2391 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2392 * allocate a new buffer to store the buf's data.
2394 * There are two additional restrictions here because we're sharing
2395 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2396 * actively involved in an L2ARC write, because if this buf is used by
2397 * an arc_write() then the hdr's data buffer will be released when the
2398 * write completes, even though the L2ARC write might still be using it.
2399 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2400 * need to be ABD-aware.
2402 can_share
= arc_can_share(hdr
, buf
) && !HDR_L2_WRITING(hdr
) &&
2403 abd_is_linear(hdr
->b_l1hdr
.b_pabd
);
2405 /* Set up b_data and sharing */
2407 buf
->b_data
= abd_to_buf(hdr
->b_l1hdr
.b_pabd
);
2408 buf
->b_flags
|= ARC_BUF_FLAG_SHARED
;
2409 arc_hdr_set_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2412 arc_get_data_buf(hdr
, arc_buf_size(buf
), buf
);
2413 ARCSTAT_INCR(arcstat_overhead_size
, arc_buf_size(buf
));
2415 VERIFY3P(buf
->b_data
, !=, NULL
);
2417 hdr
->b_l1hdr
.b_buf
= buf
;
2418 hdr
->b_l1hdr
.b_bufcnt
+= 1;
2421 * If the user wants the data from the hdr, we need to either copy or
2422 * decompress the data.
2425 return (arc_buf_fill(buf
, ARC_BUF_COMPRESSED(buf
) != 0));
2431 static char *arc_onloan_tag
= "onloan";
2434 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2435 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2436 * buffers must be returned to the arc before they can be used by the DMU or
2440 arc_loan_buf(spa_t
*spa
, boolean_t is_metadata
, int size
)
2442 arc_buf_t
*buf
= arc_alloc_buf(spa
, arc_onloan_tag
,
2443 is_metadata
? ARC_BUFC_METADATA
: ARC_BUFC_DATA
, size
);
2445 atomic_add_64(&arc_loaned_bytes
, size
);
2450 arc_loan_compressed_buf(spa_t
*spa
, uint64_t psize
, uint64_t lsize
,
2451 enum zio_compress compression_type
)
2453 arc_buf_t
*buf
= arc_alloc_compressed_buf(spa
, arc_onloan_tag
,
2454 psize
, lsize
, compression_type
);
2456 atomic_add_64(&arc_loaned_bytes
, psize
);
2462 * Return a loaned arc buffer to the arc.
2465 arc_return_buf(arc_buf_t
*buf
, void *tag
)
2467 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2469 ASSERT3P(buf
->b_data
, !=, NULL
);
2470 ASSERT(HDR_HAS_L1HDR(hdr
));
2471 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
);
2472 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
2474 atomic_add_64(&arc_loaned_bytes
, -arc_buf_size(buf
));
2477 /* Detach an arc_buf from a dbuf (tag) */
2479 arc_loan_inuse_buf(arc_buf_t
*buf
, void *tag
)
2481 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2483 ASSERT3P(buf
->b_data
, !=, NULL
);
2484 ASSERT(HDR_HAS_L1HDR(hdr
));
2485 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
2486 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
);
2488 atomic_add_64(&arc_loaned_bytes
, -arc_buf_size(buf
));
2492 l2arc_free_abd_on_write(abd_t
*abd
, size_t size
, arc_buf_contents_t type
)
2494 l2arc_data_free_t
*df
= kmem_alloc(sizeof (*df
), KM_SLEEP
);
2497 df
->l2df_size
= size
;
2498 df
->l2df_type
= type
;
2499 mutex_enter(&l2arc_free_on_write_mtx
);
2500 list_insert_head(l2arc_free_on_write
, df
);
2501 mutex_exit(&l2arc_free_on_write_mtx
);
2505 arc_hdr_free_on_write(arc_buf_hdr_t
*hdr
)
2507 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
2508 arc_buf_contents_t type
= arc_buf_type(hdr
);
2509 uint64_t size
= arc_hdr_size(hdr
);
2511 /* protected by hash lock, if in the hash table */
2512 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
2513 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2514 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
2516 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2519 (void) refcount_remove_many(&state
->arcs_size
, size
, hdr
);
2521 l2arc_free_abd_on_write(hdr
->b_l1hdr
.b_pabd
, size
, type
);
2525 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2526 * data buffer, we transfer the refcount ownership to the hdr and update
2527 * the appropriate kstats.
2530 arc_share_buf(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2532 ASSERT(arc_can_share(hdr
, buf
));
2533 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2534 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2537 * Start sharing the data buffer. We transfer the
2538 * refcount ownership to the hdr since it always owns
2539 * the refcount whenever an arc_buf_t is shared.
2541 refcount_transfer_ownership(&hdr
->b_l1hdr
.b_state
->arcs_size
, buf
, hdr
);
2542 hdr
->b_l1hdr
.b_pabd
= abd_get_from_buf(buf
->b_data
, arc_buf_size(buf
));
2543 abd_take_ownership_of_buf(hdr
->b_l1hdr
.b_pabd
,
2544 HDR_ISTYPE_METADATA(hdr
));
2545 arc_hdr_set_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2546 buf
->b_flags
|= ARC_BUF_FLAG_SHARED
;
2549 * Since we've transferred ownership to the hdr we need
2550 * to increment its compressed and uncompressed kstats and
2551 * decrement the overhead size.
2553 ARCSTAT_INCR(arcstat_compressed_size
, arc_hdr_size(hdr
));
2554 ARCSTAT_INCR(arcstat_uncompressed_size
, HDR_GET_LSIZE(hdr
));
2555 ARCSTAT_INCR(arcstat_overhead_size
, -arc_buf_size(buf
));
2559 arc_unshare_buf(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2561 ASSERT(arc_buf_is_shared(buf
));
2562 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2563 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2566 * We are no longer sharing this buffer so we need
2567 * to transfer its ownership to the rightful owner.
2569 refcount_transfer_ownership(&hdr
->b_l1hdr
.b_state
->arcs_size
, hdr
, buf
);
2570 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2571 abd_release_ownership_of_buf(hdr
->b_l1hdr
.b_pabd
);
2572 abd_put(hdr
->b_l1hdr
.b_pabd
);
2573 hdr
->b_l1hdr
.b_pabd
= NULL
;
2574 buf
->b_flags
&= ~ARC_BUF_FLAG_SHARED
;
2577 * Since the buffer is no longer shared between
2578 * the arc buf and the hdr, count it as overhead.
2580 ARCSTAT_INCR(arcstat_compressed_size
, -arc_hdr_size(hdr
));
2581 ARCSTAT_INCR(arcstat_uncompressed_size
, -HDR_GET_LSIZE(hdr
));
2582 ARCSTAT_INCR(arcstat_overhead_size
, arc_buf_size(buf
));
2586 * Remove an arc_buf_t from the hdr's buf list and return the last
2587 * arc_buf_t on the list. If no buffers remain on the list then return
2591 arc_buf_remove(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2593 arc_buf_t
**bufp
= &hdr
->b_l1hdr
.b_buf
;
2594 arc_buf_t
*lastbuf
= NULL
;
2596 ASSERT(HDR_HAS_L1HDR(hdr
));
2597 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2600 * Remove the buf from the hdr list and locate the last
2601 * remaining buffer on the list.
2603 while (*bufp
!= NULL
) {
2605 *bufp
= buf
->b_next
;
2608 * If we've removed a buffer in the middle of
2609 * the list then update the lastbuf and update
2612 if (*bufp
!= NULL
) {
2614 bufp
= &(*bufp
)->b_next
;
2618 ASSERT3P(lastbuf
, !=, buf
);
2619 IMPLY(hdr
->b_l1hdr
.b_bufcnt
> 0, lastbuf
!= NULL
);
2620 IMPLY(hdr
->b_l1hdr
.b_bufcnt
> 0, hdr
->b_l1hdr
.b_buf
!= NULL
);
2621 IMPLY(lastbuf
!= NULL
, ARC_BUF_LAST(lastbuf
));
2627 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
2631 arc_buf_destroy_impl(arc_buf_t
*buf
)
2634 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2637 * Free up the data associated with the buf but only if we're not
2638 * sharing this with the hdr. If we are sharing it with the hdr, the
2639 * hdr is responsible for doing the free.
2641 if (buf
->b_data
!= NULL
) {
2643 * We're about to change the hdr's b_flags. We must either
2644 * hold the hash_lock or be undiscoverable.
2646 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2648 arc_cksum_verify(buf
);
2649 arc_buf_unwatch(buf
);
2651 if (arc_buf_is_shared(buf
)) {
2652 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2654 uint64_t size
= arc_buf_size(buf
);
2655 arc_free_data_buf(hdr
, buf
->b_data
, size
, buf
);
2656 ARCSTAT_INCR(arcstat_overhead_size
, -size
);
2660 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
2661 hdr
->b_l1hdr
.b_bufcnt
-= 1;
2664 lastbuf
= arc_buf_remove(hdr
, buf
);
2666 if (ARC_BUF_SHARED(buf
) && !ARC_BUF_COMPRESSED(buf
)) {
2668 * If the current arc_buf_t is sharing its data buffer with the
2669 * hdr, then reassign the hdr's b_pabd to share it with the new
2670 * buffer at the end of the list. The shared buffer is always
2671 * the last one on the hdr's buffer list.
2673 * There is an equivalent case for compressed bufs, but since
2674 * they aren't guaranteed to be the last buf in the list and
2675 * that is an exceedingly rare case, we just allow that space be
2676 * wasted temporarily.
2678 if (lastbuf
!= NULL
) {
2679 /* Only one buf can be shared at once */
2680 VERIFY(!arc_buf_is_shared(lastbuf
));
2681 /* hdr is uncompressed so can't have compressed buf */
2682 VERIFY(!ARC_BUF_COMPRESSED(lastbuf
));
2684 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2685 arc_hdr_free_pabd(hdr
);
2688 * We must setup a new shared block between the
2689 * last buffer and the hdr. The data would have
2690 * been allocated by the arc buf so we need to transfer
2691 * ownership to the hdr since it's now being shared.
2693 arc_share_buf(hdr
, lastbuf
);
2695 } else if (HDR_SHARED_DATA(hdr
)) {
2697 * Uncompressed shared buffers are always at the end
2698 * of the list. Compressed buffers don't have the
2699 * same requirements. This makes it hard to
2700 * simply assert that the lastbuf is shared so
2701 * we rely on the hdr's compression flags to determine
2702 * if we have a compressed, shared buffer.
2704 ASSERT3P(lastbuf
, !=, NULL
);
2705 ASSERT(arc_buf_is_shared(lastbuf
) ||
2706 HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
);
2709 if (hdr
->b_l1hdr
.b_bufcnt
== 0)
2710 arc_cksum_free(hdr
);
2712 /* clean up the buf */
2714 kmem_cache_free(buf_cache
, buf
);
2718 arc_hdr_alloc_pabd(arc_buf_hdr_t
*hdr
)
2720 ASSERT3U(HDR_GET_LSIZE(hdr
), >, 0);
2721 ASSERT(HDR_HAS_L1HDR(hdr
));
2722 ASSERT(!HDR_SHARED_DATA(hdr
));
2724 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2725 hdr
->b_l1hdr
.b_pabd
= arc_get_data_abd(hdr
, arc_hdr_size(hdr
), hdr
);
2726 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
2727 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2729 ARCSTAT_INCR(arcstat_compressed_size
, arc_hdr_size(hdr
));
2730 ARCSTAT_INCR(arcstat_uncompressed_size
, HDR_GET_LSIZE(hdr
));
2734 arc_hdr_free_pabd(arc_buf_hdr_t
*hdr
)
2736 ASSERT(HDR_HAS_L1HDR(hdr
));
2737 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2740 * If the hdr is currently being written to the l2arc then
2741 * we defer freeing the data by adding it to the l2arc_free_on_write
2742 * list. The l2arc will free the data once it's finished
2743 * writing it to the l2arc device.
2745 if (HDR_L2_WRITING(hdr
)) {
2746 arc_hdr_free_on_write(hdr
);
2747 ARCSTAT_BUMP(arcstat_l2_free_on_write
);
2749 arc_free_data_abd(hdr
, hdr
->b_l1hdr
.b_pabd
,
2750 arc_hdr_size(hdr
), hdr
);
2752 hdr
->b_l1hdr
.b_pabd
= NULL
;
2753 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
2755 ARCSTAT_INCR(arcstat_compressed_size
, -arc_hdr_size(hdr
));
2756 ARCSTAT_INCR(arcstat_uncompressed_size
, -HDR_GET_LSIZE(hdr
));
2759 static arc_buf_hdr_t
*
2760 arc_hdr_alloc(uint64_t spa
, int32_t psize
, int32_t lsize
,
2761 enum zio_compress compression_type
, arc_buf_contents_t type
)
2765 VERIFY(type
== ARC_BUFC_DATA
|| type
== ARC_BUFC_METADATA
);
2767 hdr
= kmem_cache_alloc(hdr_full_cache
, KM_PUSHPAGE
);
2768 ASSERT(HDR_EMPTY(hdr
));
2769 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
2770 HDR_SET_PSIZE(hdr
, psize
);
2771 HDR_SET_LSIZE(hdr
, lsize
);
2775 arc_hdr_set_flags(hdr
, arc_bufc_to_flags(type
) | ARC_FLAG_HAS_L1HDR
);
2776 arc_hdr_set_compress(hdr
, compression_type
);
2778 hdr
->b_l1hdr
.b_state
= arc_anon
;
2779 hdr
->b_l1hdr
.b_arc_access
= 0;
2780 hdr
->b_l1hdr
.b_bufcnt
= 0;
2781 hdr
->b_l1hdr
.b_buf
= NULL
;
2784 * Allocate the hdr's buffer. This will contain either
2785 * the compressed or uncompressed data depending on the block
2786 * it references and compressed arc enablement.
2788 arc_hdr_alloc_pabd(hdr
);
2789 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2795 * Transition between the two allocation states for the arc_buf_hdr struct.
2796 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
2797 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
2798 * version is used when a cache buffer is only in the L2ARC in order to reduce
2801 static arc_buf_hdr_t
*
2802 arc_hdr_realloc(arc_buf_hdr_t
*hdr
, kmem_cache_t
*old
, kmem_cache_t
*new)
2804 arc_buf_hdr_t
*nhdr
;
2805 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
2807 ASSERT(HDR_HAS_L2HDR(hdr
));
2808 ASSERT((old
== hdr_full_cache
&& new == hdr_l2only_cache
) ||
2809 (old
== hdr_l2only_cache
&& new == hdr_full_cache
));
2811 nhdr
= kmem_cache_alloc(new, KM_PUSHPAGE
);
2813 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)));
2814 buf_hash_remove(hdr
);
2816 bcopy(hdr
, nhdr
, HDR_L2ONLY_SIZE
);
2818 if (new == hdr_full_cache
) {
2819 arc_hdr_set_flags(nhdr
, ARC_FLAG_HAS_L1HDR
);
2821 * arc_access and arc_change_state need to be aware that a
2822 * header has just come out of L2ARC, so we set its state to
2823 * l2c_only even though it's about to change.
2825 nhdr
->b_l1hdr
.b_state
= arc_l2c_only
;
2827 /* Verify previous threads set to NULL before freeing */
2828 ASSERT3P(nhdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2830 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2831 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
2832 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
2835 * If we've reached here, We must have been called from
2836 * arc_evict_hdr(), as such we should have already been
2837 * removed from any ghost list we were previously on
2838 * (which protects us from racing with arc_evict_state),
2839 * thus no locking is needed during this check.
2841 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
2844 * A buffer must not be moved into the arc_l2c_only
2845 * state if it's not finished being written out to the
2846 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
2847 * might try to be accessed, even though it was removed.
2849 VERIFY(!HDR_L2_WRITING(hdr
));
2850 VERIFY3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2852 arc_hdr_clear_flags(nhdr
, ARC_FLAG_HAS_L1HDR
);
2855 * The header has been reallocated so we need to re-insert it into any
2858 (void) buf_hash_insert(nhdr
, NULL
);
2860 ASSERT(list_link_active(&hdr
->b_l2hdr
.b_l2node
));
2862 mutex_enter(&dev
->l2ad_mtx
);
2865 * We must place the realloc'ed header back into the list at
2866 * the same spot. Otherwise, if it's placed earlier in the list,
2867 * l2arc_write_buffers() could find it during the function's
2868 * write phase, and try to write it out to the l2arc.
2870 list_insert_after(&dev
->l2ad_buflist
, hdr
, nhdr
);
2871 list_remove(&dev
->l2ad_buflist
, hdr
);
2873 mutex_exit(&dev
->l2ad_mtx
);
2876 * Since we're using the pointer address as the tag when
2877 * incrementing and decrementing the l2ad_alloc refcount, we
2878 * must remove the old pointer (that we're about to destroy) and
2879 * add the new pointer to the refcount. Otherwise we'd remove
2880 * the wrong pointer address when calling arc_hdr_destroy() later.
2883 (void) refcount_remove_many(&dev
->l2ad_alloc
, arc_hdr_size(hdr
), hdr
);
2884 (void) refcount_add_many(&dev
->l2ad_alloc
, arc_hdr_size(nhdr
), nhdr
);
2886 buf_discard_identity(hdr
);
2887 kmem_cache_free(old
, hdr
);
2893 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
2894 * The buf is returned thawed since we expect the consumer to modify it.
2897 arc_alloc_buf(spa_t
*spa
, void *tag
, arc_buf_contents_t type
, int32_t size
)
2900 arc_buf_hdr_t
*hdr
= arc_hdr_alloc(spa_load_guid(spa
), size
, size
,
2901 ZIO_COMPRESS_OFF
, type
);
2902 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
2905 VERIFY0(arc_buf_alloc_impl(hdr
, tag
, B_FALSE
, B_FALSE
, &buf
));
2912 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
2913 * for bufs containing metadata.
2916 arc_alloc_compressed_buf(spa_t
*spa
, void *tag
, uint64_t psize
, uint64_t lsize
,
2917 enum zio_compress compression_type
)
2921 ASSERT3U(lsize
, >, 0);
2922 ASSERT3U(lsize
, >=, psize
);
2923 ASSERT(compression_type
> ZIO_COMPRESS_OFF
);
2924 ASSERT(compression_type
< ZIO_COMPRESS_FUNCTIONS
);
2926 hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
,
2927 compression_type
, ARC_BUFC_DATA
);
2928 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
2931 VERIFY0(arc_buf_alloc_impl(hdr
, tag
, B_TRUE
, B_FALSE
, &buf
));
2933 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
2935 if (!arc_buf_is_shared(buf
)) {
2937 * To ensure that the hdr has the correct data in it if we call
2938 * arc_decompress() on this buf before it's been written to
2939 * disk, it's easiest if we just set up sharing between the
2942 ASSERT(!abd_is_linear(hdr
->b_l1hdr
.b_pabd
));
2943 arc_hdr_free_pabd(hdr
);
2944 arc_share_buf(hdr
, buf
);
2951 arc_hdr_l2hdr_destroy(arc_buf_hdr_t
*hdr
)
2953 l2arc_buf_hdr_t
*l2hdr
= &hdr
->b_l2hdr
;
2954 l2arc_dev_t
*dev
= l2hdr
->b_dev
;
2955 uint64_t asize
= arc_hdr_size(hdr
);
2957 ASSERT(MUTEX_HELD(&dev
->l2ad_mtx
));
2958 ASSERT(HDR_HAS_L2HDR(hdr
));
2960 list_remove(&dev
->l2ad_buflist
, hdr
);
2962 ARCSTAT_INCR(arcstat_l2_asize
, -asize
);
2963 ARCSTAT_INCR(arcstat_l2_size
, -HDR_GET_LSIZE(hdr
));
2965 vdev_space_update(dev
->l2ad_vdev
, -asize
, 0, 0);
2967 (void) refcount_remove_many(&dev
->l2ad_alloc
, asize
, hdr
);
2968 arc_hdr_clear_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
2972 arc_hdr_destroy(arc_buf_hdr_t
*hdr
)
2974 if (HDR_HAS_L1HDR(hdr
)) {
2975 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
||
2976 hdr
->b_l1hdr
.b_bufcnt
> 0);
2977 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2978 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
2980 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
2981 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
2983 if (!HDR_EMPTY(hdr
))
2984 buf_discard_identity(hdr
);
2986 if (HDR_HAS_L2HDR(hdr
)) {
2987 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
2988 boolean_t buflist_held
= MUTEX_HELD(&dev
->l2ad_mtx
);
2991 mutex_enter(&dev
->l2ad_mtx
);
2994 * Even though we checked this conditional above, we
2995 * need to check this again now that we have the
2996 * l2ad_mtx. This is because we could be racing with
2997 * another thread calling l2arc_evict() which might have
2998 * destroyed this header's L2 portion as we were waiting
2999 * to acquire the l2ad_mtx. If that happens, we don't
3000 * want to re-destroy the header's L2 portion.
3002 if (HDR_HAS_L2HDR(hdr
))
3003 arc_hdr_l2hdr_destroy(hdr
);
3006 mutex_exit(&dev
->l2ad_mtx
);
3009 if (HDR_HAS_L1HDR(hdr
)) {
3010 arc_cksum_free(hdr
);
3012 while (hdr
->b_l1hdr
.b_buf
!= NULL
)
3013 arc_buf_destroy_impl(hdr
->b_l1hdr
.b_buf
);
3015 if (hdr
->b_l1hdr
.b_pabd
!= NULL
)
3016 arc_hdr_free_pabd(hdr
);
3019 ASSERT3P(hdr
->b_hash_next
, ==, NULL
);
3020 if (HDR_HAS_L1HDR(hdr
)) {
3021 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
3022 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
3023 kmem_cache_free(hdr_full_cache
, hdr
);
3025 kmem_cache_free(hdr_l2only_cache
, hdr
);
3030 arc_buf_destroy(arc_buf_t
*buf
, void* tag
)
3032 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3033 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
3035 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
3036 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
3037 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3038 VERIFY0(remove_reference(hdr
, NULL
, tag
));
3039 arc_hdr_destroy(hdr
);
3043 mutex_enter(hash_lock
);
3044 ASSERT3P(hdr
, ==, buf
->b_hdr
);
3045 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
3046 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
3047 ASSERT3P(hdr
->b_l1hdr
.b_state
, !=, arc_anon
);
3048 ASSERT3P(buf
->b_data
, !=, NULL
);
3050 (void) remove_reference(hdr
, hash_lock
, tag
);
3051 arc_buf_destroy_impl(buf
);
3052 mutex_exit(hash_lock
);
3056 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3057 * state of the header is dependent on its state prior to entering this
3058 * function. The following transitions are possible:
3060 * - arc_mru -> arc_mru_ghost
3061 * - arc_mfu -> arc_mfu_ghost
3062 * - arc_mru_ghost -> arc_l2c_only
3063 * - arc_mru_ghost -> deleted
3064 * - arc_mfu_ghost -> arc_l2c_only
3065 * - arc_mfu_ghost -> deleted
3068 arc_evict_hdr(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
3070 arc_state_t
*evicted_state
, *state
;
3071 int64_t bytes_evicted
= 0;
3073 ASSERT(MUTEX_HELD(hash_lock
));
3074 ASSERT(HDR_HAS_L1HDR(hdr
));
3076 state
= hdr
->b_l1hdr
.b_state
;
3077 if (GHOST_STATE(state
)) {
3078 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3079 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
3082 * l2arc_write_buffers() relies on a header's L1 portion
3083 * (i.e. its b_pabd field) during it's write phase.
3084 * Thus, we cannot push a header onto the arc_l2c_only
3085 * state (removing its L1 piece) until the header is
3086 * done being written to the l2arc.
3088 if (HDR_HAS_L2HDR(hdr
) && HDR_L2_WRITING(hdr
)) {
3089 ARCSTAT_BUMP(arcstat_evict_l2_skip
);
3090 return (bytes_evicted
);
3093 ARCSTAT_BUMP(arcstat_deleted
);
3094 bytes_evicted
+= HDR_GET_LSIZE(hdr
);
3096 DTRACE_PROBE1(arc__delete
, arc_buf_hdr_t
*, hdr
);
3098 if (HDR_HAS_L2HDR(hdr
)) {
3099 ASSERT(hdr
->b_l1hdr
.b_pabd
== NULL
);
3101 * This buffer is cached on the 2nd Level ARC;
3102 * don't destroy the header.
3104 arc_change_state(arc_l2c_only
, hdr
, hash_lock
);
3106 * dropping from L1+L2 cached to L2-only,
3107 * realloc to remove the L1 header.
3109 hdr
= arc_hdr_realloc(hdr
, hdr_full_cache
,
3112 arc_change_state(arc_anon
, hdr
, hash_lock
);
3113 arc_hdr_destroy(hdr
);
3115 return (bytes_evicted
);
3118 ASSERT(state
== arc_mru
|| state
== arc_mfu
);
3119 evicted_state
= (state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
3121 /* prefetch buffers have a minimum lifespan */
3122 if (HDR_IO_IN_PROGRESS(hdr
) ||
3123 ((hdr
->b_flags
& (ARC_FLAG_PREFETCH
| ARC_FLAG_INDIRECT
)) &&
3124 ddi_get_lbolt() - hdr
->b_l1hdr
.b_arc_access
<
3125 arc_min_prefetch_lifespan
)) {
3126 ARCSTAT_BUMP(arcstat_evict_skip
);
3127 return (bytes_evicted
);
3130 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
3131 while (hdr
->b_l1hdr
.b_buf
) {
3132 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
3133 if (!mutex_tryenter(&buf
->b_evict_lock
)) {
3134 ARCSTAT_BUMP(arcstat_mutex_miss
);
3137 if (buf
->b_data
!= NULL
)
3138 bytes_evicted
+= HDR_GET_LSIZE(hdr
);
3139 mutex_exit(&buf
->b_evict_lock
);
3140 arc_buf_destroy_impl(buf
);
3143 if (HDR_HAS_L2HDR(hdr
)) {
3144 ARCSTAT_INCR(arcstat_evict_l2_cached
, HDR_GET_LSIZE(hdr
));
3146 if (l2arc_write_eligible(hdr
->b_spa
, hdr
)) {
3147 ARCSTAT_INCR(arcstat_evict_l2_eligible
,
3148 HDR_GET_LSIZE(hdr
));
3150 ARCSTAT_INCR(arcstat_evict_l2_ineligible
,
3151 HDR_GET_LSIZE(hdr
));
3155 if (hdr
->b_l1hdr
.b_bufcnt
== 0) {
3156 arc_cksum_free(hdr
);
3158 bytes_evicted
+= arc_hdr_size(hdr
);
3161 * If this hdr is being evicted and has a compressed
3162 * buffer then we discard it here before we change states.
3163 * This ensures that the accounting is updated correctly
3164 * in arc_free_data_impl().
3166 arc_hdr_free_pabd(hdr
);
3168 arc_change_state(evicted_state
, hdr
, hash_lock
);
3169 ASSERT(HDR_IN_HASH_TABLE(hdr
));
3170 arc_hdr_set_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
3171 DTRACE_PROBE1(arc__evict
, arc_buf_hdr_t
*, hdr
);
3174 return (bytes_evicted
);
3178 arc_evict_state_impl(multilist_t
*ml
, int idx
, arc_buf_hdr_t
*marker
,
3179 uint64_t spa
, int64_t bytes
)
3181 multilist_sublist_t
*mls
;
3182 uint64_t bytes_evicted
= 0;
3184 kmutex_t
*hash_lock
;
3185 int evict_count
= 0;
3187 ASSERT3P(marker
, !=, NULL
);
3188 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
3190 mls
= multilist_sublist_lock(ml
, idx
);
3192 for (hdr
= multilist_sublist_prev(mls
, marker
); hdr
!= NULL
;
3193 hdr
= multilist_sublist_prev(mls
, marker
)) {
3194 if ((bytes
!= ARC_EVICT_ALL
&& bytes_evicted
>= bytes
) ||
3195 (evict_count
>= zfs_arc_evict_batch_limit
))
3199 * To keep our iteration location, move the marker
3200 * forward. Since we're not holding hdr's hash lock, we
3201 * must be very careful and not remove 'hdr' from the
3202 * sublist. Otherwise, other consumers might mistake the
3203 * 'hdr' as not being on a sublist when they call the
3204 * multilist_link_active() function (they all rely on
3205 * the hash lock protecting concurrent insertions and
3206 * removals). multilist_sublist_move_forward() was
3207 * specifically implemented to ensure this is the case
3208 * (only 'marker' will be removed and re-inserted).
3210 multilist_sublist_move_forward(mls
, marker
);
3213 * The only case where the b_spa field should ever be
3214 * zero, is the marker headers inserted by
3215 * arc_evict_state(). It's possible for multiple threads
3216 * to be calling arc_evict_state() concurrently (e.g.
3217 * dsl_pool_close() and zio_inject_fault()), so we must
3218 * skip any markers we see from these other threads.
3220 if (hdr
->b_spa
== 0)
3223 /* we're only interested in evicting buffers of a certain spa */
3224 if (spa
!= 0 && hdr
->b_spa
!= spa
) {
3225 ARCSTAT_BUMP(arcstat_evict_skip
);
3229 hash_lock
= HDR_LOCK(hdr
);
3232 * We aren't calling this function from any code path
3233 * that would already be holding a hash lock, so we're
3234 * asserting on this assumption to be defensive in case
3235 * this ever changes. Without this check, it would be
3236 * possible to incorrectly increment arcstat_mutex_miss
3237 * below (e.g. if the code changed such that we called
3238 * this function with a hash lock held).
3240 ASSERT(!MUTEX_HELD(hash_lock
));
3242 if (mutex_tryenter(hash_lock
)) {
3243 uint64_t evicted
= arc_evict_hdr(hdr
, hash_lock
);
3244 mutex_exit(hash_lock
);
3246 bytes_evicted
+= evicted
;
3249 * If evicted is zero, arc_evict_hdr() must have
3250 * decided to skip this header, don't increment
3251 * evict_count in this case.
3257 * If arc_size isn't overflowing, signal any
3258 * threads that might happen to be waiting.
3260 * For each header evicted, we wake up a single
3261 * thread. If we used cv_broadcast, we could
3262 * wake up "too many" threads causing arc_size
3263 * to significantly overflow arc_c; since
3264 * arc_get_data_impl() doesn't check for overflow
3265 * when it's woken up (it doesn't because it's
3266 * possible for the ARC to be overflowing while
3267 * full of un-evictable buffers, and the
3268 * function should proceed in this case).
3270 * If threads are left sleeping, due to not
3271 * using cv_broadcast, they will be woken up
3272 * just before arc_reclaim_thread() sleeps.
3274 mutex_enter(&arc_reclaim_lock
);
3275 if (!arc_is_overflowing())
3276 cv_signal(&arc_reclaim_waiters_cv
);
3277 mutex_exit(&arc_reclaim_lock
);
3279 ARCSTAT_BUMP(arcstat_mutex_miss
);
3283 multilist_sublist_unlock(mls
);
3285 return (bytes_evicted
);
3289 * Evict buffers from the given arc state, until we've removed the
3290 * specified number of bytes. Move the removed buffers to the
3291 * appropriate evict state.
3293 * This function makes a "best effort". It skips over any buffers
3294 * it can't get a hash_lock on, and so, may not catch all candidates.
3295 * It may also return without evicting as much space as requested.
3297 * If bytes is specified using the special value ARC_EVICT_ALL, this
3298 * will evict all available (i.e. unlocked and evictable) buffers from
3299 * the given arc state; which is used by arc_flush().
3302 arc_evict_state(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
3303 arc_buf_contents_t type
)
3305 uint64_t total_evicted
= 0;
3306 multilist_t
*ml
= &state
->arcs_list
[type
];
3308 arc_buf_hdr_t
**markers
;
3311 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
3313 num_sublists
= multilist_get_num_sublists(ml
);
3316 * If we've tried to evict from each sublist, made some
3317 * progress, but still have not hit the target number of bytes
3318 * to evict, we want to keep trying. The markers allow us to
3319 * pick up where we left off for each individual sublist, rather
3320 * than starting from the tail each time.
3322 markers
= kmem_zalloc(sizeof (*markers
) * num_sublists
, KM_SLEEP
);
3323 for (i
= 0; i
< num_sublists
; i
++) {
3324 multilist_sublist_t
*mls
;
3326 markers
[i
] = kmem_cache_alloc(hdr_full_cache
, KM_SLEEP
);
3329 * A b_spa of 0 is used to indicate that this header is
3330 * a marker. This fact is used in arc_adjust_type() and
3331 * arc_evict_state_impl().
3333 markers
[i
]->b_spa
= 0;
3335 mls
= multilist_sublist_lock(ml
, i
);
3336 multilist_sublist_insert_tail(mls
, markers
[i
]);
3337 multilist_sublist_unlock(mls
);
3341 * While we haven't hit our target number of bytes to evict, or
3342 * we're evicting all available buffers.
3344 while (total_evicted
< bytes
|| bytes
== ARC_EVICT_ALL
) {
3345 int sublist_idx
= multilist_get_random_index(ml
);
3346 uint64_t scan_evicted
= 0;
3349 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
3350 * Request that 10% of the LRUs be scanned by the superblock
3353 if (type
== ARC_BUFC_DATA
&& arc_dnode_size
> arc_dnode_limit
)
3354 arc_prune_async((arc_dnode_size
- arc_dnode_limit
) /
3355 sizeof (dnode_t
) / zfs_arc_dnode_reduce_percent
);
3358 * Start eviction using a randomly selected sublist,
3359 * this is to try and evenly balance eviction across all
3360 * sublists. Always starting at the same sublist
3361 * (e.g. index 0) would cause evictions to favor certain
3362 * sublists over others.
3364 for (i
= 0; i
< num_sublists
; i
++) {
3365 uint64_t bytes_remaining
;
3366 uint64_t bytes_evicted
;
3368 if (bytes
== ARC_EVICT_ALL
)
3369 bytes_remaining
= ARC_EVICT_ALL
;
3370 else if (total_evicted
< bytes
)
3371 bytes_remaining
= bytes
- total_evicted
;
3375 bytes_evicted
= arc_evict_state_impl(ml
, sublist_idx
,
3376 markers
[sublist_idx
], spa
, bytes_remaining
);
3378 scan_evicted
+= bytes_evicted
;
3379 total_evicted
+= bytes_evicted
;
3381 /* we've reached the end, wrap to the beginning */
3382 if (++sublist_idx
>= num_sublists
)
3387 * If we didn't evict anything during this scan, we have
3388 * no reason to believe we'll evict more during another
3389 * scan, so break the loop.
3391 if (scan_evicted
== 0) {
3392 /* This isn't possible, let's make that obvious */
3393 ASSERT3S(bytes
, !=, 0);
3396 * When bytes is ARC_EVICT_ALL, the only way to
3397 * break the loop is when scan_evicted is zero.
3398 * In that case, we actually have evicted enough,
3399 * so we don't want to increment the kstat.
3401 if (bytes
!= ARC_EVICT_ALL
) {
3402 ASSERT3S(total_evicted
, <, bytes
);
3403 ARCSTAT_BUMP(arcstat_evict_not_enough
);
3410 for (i
= 0; i
< num_sublists
; i
++) {
3411 multilist_sublist_t
*mls
= multilist_sublist_lock(ml
, i
);
3412 multilist_sublist_remove(mls
, markers
[i
]);
3413 multilist_sublist_unlock(mls
);
3415 kmem_cache_free(hdr_full_cache
, markers
[i
]);
3417 kmem_free(markers
, sizeof (*markers
) * num_sublists
);
3419 return (total_evicted
);
3423 * Flush all "evictable" data of the given type from the arc state
3424 * specified. This will not evict any "active" buffers (i.e. referenced).
3426 * When 'retry' is set to B_FALSE, the function will make a single pass
3427 * over the state and evict any buffers that it can. Since it doesn't
3428 * continually retry the eviction, it might end up leaving some buffers
3429 * in the ARC due to lock misses.
3431 * When 'retry' is set to B_TRUE, the function will continually retry the
3432 * eviction until *all* evictable buffers have been removed from the
3433 * state. As a result, if concurrent insertions into the state are
3434 * allowed (e.g. if the ARC isn't shutting down), this function might
3435 * wind up in an infinite loop, continually trying to evict buffers.
3438 arc_flush_state(arc_state_t
*state
, uint64_t spa
, arc_buf_contents_t type
,
3441 uint64_t evicted
= 0;
3443 while (refcount_count(&state
->arcs_esize
[type
]) != 0) {
3444 evicted
+= arc_evict_state(state
, spa
, ARC_EVICT_ALL
, type
);
3454 * Helper function for arc_prune_async() it is responsible for safely
3455 * handling the execution of a registered arc_prune_func_t.
3458 arc_prune_task(void *ptr
)
3460 arc_prune_t
*ap
= (arc_prune_t
*)ptr
;
3461 arc_prune_func_t
*func
= ap
->p_pfunc
;
3464 func(ap
->p_adjust
, ap
->p_private
);
3466 refcount_remove(&ap
->p_refcnt
, func
);
3470 * Notify registered consumers they must drop holds on a portion of the ARC
3471 * buffered they reference. This provides a mechanism to ensure the ARC can
3472 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
3473 * is analogous to dnlc_reduce_cache() but more generic.
3475 * This operation is performed asynchronously so it may be safely called
3476 * in the context of the arc_reclaim_thread(). A reference is taken here
3477 * for each registered arc_prune_t and the arc_prune_task() is responsible
3478 * for releasing it once the registered arc_prune_func_t has completed.
3481 arc_prune_async(int64_t adjust
)
3485 mutex_enter(&arc_prune_mtx
);
3486 for (ap
= list_head(&arc_prune_list
); ap
!= NULL
;
3487 ap
= list_next(&arc_prune_list
, ap
)) {
3489 if (refcount_count(&ap
->p_refcnt
) >= 2)
3492 refcount_add(&ap
->p_refcnt
, ap
->p_pfunc
);
3493 ap
->p_adjust
= adjust
;
3494 if (taskq_dispatch(arc_prune_taskq
, arc_prune_task
,
3495 ap
, TQ_SLEEP
) == TASKQID_INVALID
) {
3496 refcount_remove(&ap
->p_refcnt
, ap
->p_pfunc
);
3499 ARCSTAT_BUMP(arcstat_prune
);
3501 mutex_exit(&arc_prune_mtx
);
3505 * Evict the specified number of bytes from the state specified,
3506 * restricting eviction to the spa and type given. This function
3507 * prevents us from trying to evict more from a state's list than
3508 * is "evictable", and to skip evicting altogether when passed a
3509 * negative value for "bytes". In contrast, arc_evict_state() will
3510 * evict everything it can, when passed a negative value for "bytes".
3513 arc_adjust_impl(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
3514 arc_buf_contents_t type
)
3518 if (bytes
> 0 && refcount_count(&state
->arcs_esize
[type
]) > 0) {
3519 delta
= MIN(refcount_count(&state
->arcs_esize
[type
]), bytes
);
3520 return (arc_evict_state(state
, spa
, delta
, type
));
3527 * The goal of this function is to evict enough meta data buffers from the
3528 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
3529 * more complicated than it appears because it is common for data buffers
3530 * to have holds on meta data buffers. In addition, dnode meta data buffers
3531 * will be held by the dnodes in the block preventing them from being freed.
3532 * This means we can't simply traverse the ARC and expect to always find
3533 * enough unheld meta data buffer to release.
3535 * Therefore, this function has been updated to make alternating passes
3536 * over the ARC releasing data buffers and then newly unheld meta data
3537 * buffers. This ensures forward progress is maintained and arc_meta_used
3538 * will decrease. Normally this is sufficient, but if required the ARC
3539 * will call the registered prune callbacks causing dentry and inodes to
3540 * be dropped from the VFS cache. This will make dnode meta data buffers
3541 * available for reclaim.
3544 arc_adjust_meta_balanced(void)
3546 int64_t delta
, prune
= 0, adjustmnt
;
3547 uint64_t total_evicted
= 0;
3548 arc_buf_contents_t type
= ARC_BUFC_DATA
;
3549 int restarts
= MAX(zfs_arc_meta_adjust_restarts
, 0);
3553 * This slightly differs than the way we evict from the mru in
3554 * arc_adjust because we don't have a "target" value (i.e. no
3555 * "meta" arc_p). As a result, I think we can completely
3556 * cannibalize the metadata in the MRU before we evict the
3557 * metadata from the MFU. I think we probably need to implement a
3558 * "metadata arc_p" value to do this properly.
3560 adjustmnt
= arc_meta_used
- arc_meta_limit
;
3562 if (adjustmnt
> 0 && refcount_count(&arc_mru
->arcs_esize
[type
]) > 0) {
3563 delta
= MIN(refcount_count(&arc_mru
->arcs_esize
[type
]),
3565 total_evicted
+= arc_adjust_impl(arc_mru
, 0, delta
, type
);
3570 * We can't afford to recalculate adjustmnt here. If we do,
3571 * new metadata buffers can sneak into the MRU or ANON lists,
3572 * thus penalize the MFU metadata. Although the fudge factor is
3573 * small, it has been empirically shown to be significant for
3574 * certain workloads (e.g. creating many empty directories). As
3575 * such, we use the original calculation for adjustmnt, and
3576 * simply decrement the amount of data evicted from the MRU.
3579 if (adjustmnt
> 0 && refcount_count(&arc_mfu
->arcs_esize
[type
]) > 0) {
3580 delta
= MIN(refcount_count(&arc_mfu
->arcs_esize
[type
]),
3582 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, delta
, type
);
3585 adjustmnt
= arc_meta_used
- arc_meta_limit
;
3587 if (adjustmnt
> 0 &&
3588 refcount_count(&arc_mru_ghost
->arcs_esize
[type
]) > 0) {
3589 delta
= MIN(adjustmnt
,
3590 refcount_count(&arc_mru_ghost
->arcs_esize
[type
]));
3591 total_evicted
+= arc_adjust_impl(arc_mru_ghost
, 0, delta
, type
);
3595 if (adjustmnt
> 0 &&
3596 refcount_count(&arc_mfu_ghost
->arcs_esize
[type
]) > 0) {
3597 delta
= MIN(adjustmnt
,
3598 refcount_count(&arc_mfu_ghost
->arcs_esize
[type
]));
3599 total_evicted
+= arc_adjust_impl(arc_mfu_ghost
, 0, delta
, type
);
3603 * If after attempting to make the requested adjustment to the ARC
3604 * the meta limit is still being exceeded then request that the
3605 * higher layers drop some cached objects which have holds on ARC
3606 * meta buffers. Requests to the upper layers will be made with
3607 * increasingly large scan sizes until the ARC is below the limit.
3609 if (arc_meta_used
> arc_meta_limit
) {
3610 if (type
== ARC_BUFC_DATA
) {
3611 type
= ARC_BUFC_METADATA
;
3613 type
= ARC_BUFC_DATA
;
3615 if (zfs_arc_meta_prune
) {
3616 prune
+= zfs_arc_meta_prune
;
3617 arc_prune_async(prune
);
3626 return (total_evicted
);
3630 * Evict metadata buffers from the cache, such that arc_meta_used is
3631 * capped by the arc_meta_limit tunable.
3634 arc_adjust_meta_only(void)
3636 uint64_t total_evicted
= 0;
3640 * If we're over the meta limit, we want to evict enough
3641 * metadata to get back under the meta limit. We don't want to
3642 * evict so much that we drop the MRU below arc_p, though. If
3643 * we're over the meta limit more than we're over arc_p, we
3644 * evict some from the MRU here, and some from the MFU below.
3646 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
3647 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
3648 refcount_count(&arc_mru
->arcs_size
) - arc_p
));
3650 total_evicted
+= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3653 * Similar to the above, we want to evict enough bytes to get us
3654 * below the meta limit, but not so much as to drop us below the
3655 * space allotted to the MFU (which is defined as arc_c - arc_p).
3657 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
3658 (int64_t)(refcount_count(&arc_mfu
->arcs_size
) - (arc_c
- arc_p
)));
3660 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3662 return (total_evicted
);
3666 arc_adjust_meta(void)
3668 if (zfs_arc_meta_strategy
== ARC_STRATEGY_META_ONLY
)
3669 return (arc_adjust_meta_only());
3671 return (arc_adjust_meta_balanced());
3675 * Return the type of the oldest buffer in the given arc state
3677 * This function will select a random sublist of type ARC_BUFC_DATA and
3678 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3679 * is compared, and the type which contains the "older" buffer will be
3682 static arc_buf_contents_t
3683 arc_adjust_type(arc_state_t
*state
)
3685 multilist_t
*data_ml
= &state
->arcs_list
[ARC_BUFC_DATA
];
3686 multilist_t
*meta_ml
= &state
->arcs_list
[ARC_BUFC_METADATA
];
3687 int data_idx
= multilist_get_random_index(data_ml
);
3688 int meta_idx
= multilist_get_random_index(meta_ml
);
3689 multilist_sublist_t
*data_mls
;
3690 multilist_sublist_t
*meta_mls
;
3691 arc_buf_contents_t type
;
3692 arc_buf_hdr_t
*data_hdr
;
3693 arc_buf_hdr_t
*meta_hdr
;
3696 * We keep the sublist lock until we're finished, to prevent
3697 * the headers from being destroyed via arc_evict_state().
3699 data_mls
= multilist_sublist_lock(data_ml
, data_idx
);
3700 meta_mls
= multilist_sublist_lock(meta_ml
, meta_idx
);
3703 * These two loops are to ensure we skip any markers that
3704 * might be at the tail of the lists due to arc_evict_state().
3707 for (data_hdr
= multilist_sublist_tail(data_mls
); data_hdr
!= NULL
;
3708 data_hdr
= multilist_sublist_prev(data_mls
, data_hdr
)) {
3709 if (data_hdr
->b_spa
!= 0)
3713 for (meta_hdr
= multilist_sublist_tail(meta_mls
); meta_hdr
!= NULL
;
3714 meta_hdr
= multilist_sublist_prev(meta_mls
, meta_hdr
)) {
3715 if (meta_hdr
->b_spa
!= 0)
3719 if (data_hdr
== NULL
&& meta_hdr
== NULL
) {
3720 type
= ARC_BUFC_DATA
;
3721 } else if (data_hdr
== NULL
) {
3722 ASSERT3P(meta_hdr
, !=, NULL
);
3723 type
= ARC_BUFC_METADATA
;
3724 } else if (meta_hdr
== NULL
) {
3725 ASSERT3P(data_hdr
, !=, NULL
);
3726 type
= ARC_BUFC_DATA
;
3728 ASSERT3P(data_hdr
, !=, NULL
);
3729 ASSERT3P(meta_hdr
, !=, NULL
);
3731 /* The headers can't be on the sublist without an L1 header */
3732 ASSERT(HDR_HAS_L1HDR(data_hdr
));
3733 ASSERT(HDR_HAS_L1HDR(meta_hdr
));
3735 if (data_hdr
->b_l1hdr
.b_arc_access
<
3736 meta_hdr
->b_l1hdr
.b_arc_access
) {
3737 type
= ARC_BUFC_DATA
;
3739 type
= ARC_BUFC_METADATA
;
3743 multilist_sublist_unlock(meta_mls
);
3744 multilist_sublist_unlock(data_mls
);
3750 * Evict buffers from the cache, such that arc_size is capped by arc_c.
3755 uint64_t total_evicted
= 0;
3760 * If we're over arc_meta_limit, we want to correct that before
3761 * potentially evicting data buffers below.
3763 total_evicted
+= arc_adjust_meta();
3768 * If we're over the target cache size, we want to evict enough
3769 * from the list to get back to our target size. We don't want
3770 * to evict too much from the MRU, such that it drops below
3771 * arc_p. So, if we're over our target cache size more than
3772 * the MRU is over arc_p, we'll evict enough to get back to
3773 * arc_p here, and then evict more from the MFU below.
3775 target
= MIN((int64_t)(arc_size
- arc_c
),
3776 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
3777 refcount_count(&arc_mru
->arcs_size
) + arc_meta_used
- arc_p
));
3780 * If we're below arc_meta_min, always prefer to evict data.
3781 * Otherwise, try to satisfy the requested number of bytes to
3782 * evict from the type which contains older buffers; in an
3783 * effort to keep newer buffers in the cache regardless of their
3784 * type. If we cannot satisfy the number of bytes from this
3785 * type, spill over into the next type.
3787 if (arc_adjust_type(arc_mru
) == ARC_BUFC_METADATA
&&
3788 arc_meta_used
> arc_meta_min
) {
3789 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3790 total_evicted
+= bytes
;
3793 * If we couldn't evict our target number of bytes from
3794 * metadata, we try to get the rest from data.
3799 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3801 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3802 total_evicted
+= bytes
;
3805 * If we couldn't evict our target number of bytes from
3806 * data, we try to get the rest from metadata.
3811 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3817 * Now that we've tried to evict enough from the MRU to get its
3818 * size back to arc_p, if we're still above the target cache
3819 * size, we evict the rest from the MFU.
3821 target
= arc_size
- arc_c
;
3823 if (arc_adjust_type(arc_mfu
) == ARC_BUFC_METADATA
&&
3824 arc_meta_used
> arc_meta_min
) {
3825 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3826 total_evicted
+= bytes
;
3829 * If we couldn't evict our target number of bytes from
3830 * metadata, we try to get the rest from data.
3835 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3837 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3838 total_evicted
+= bytes
;
3841 * If we couldn't evict our target number of bytes from
3842 * data, we try to get the rest from data.
3847 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3851 * Adjust ghost lists
3853 * In addition to the above, the ARC also defines target values
3854 * for the ghost lists. The sum of the mru list and mru ghost
3855 * list should never exceed the target size of the cache, and
3856 * the sum of the mru list, mfu list, mru ghost list, and mfu
3857 * ghost list should never exceed twice the target size of the
3858 * cache. The following logic enforces these limits on the ghost
3859 * caches, and evicts from them as needed.
3861 target
= refcount_count(&arc_mru
->arcs_size
) +
3862 refcount_count(&arc_mru_ghost
->arcs_size
) - arc_c
;
3864 bytes
= arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_DATA
);
3865 total_evicted
+= bytes
;
3870 arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_METADATA
);
3873 * We assume the sum of the mru list and mfu list is less than
3874 * or equal to arc_c (we enforced this above), which means we
3875 * can use the simpler of the two equations below:
3877 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3878 * mru ghost + mfu ghost <= arc_c
3880 target
= refcount_count(&arc_mru_ghost
->arcs_size
) +
3881 refcount_count(&arc_mfu_ghost
->arcs_size
) - arc_c
;
3883 bytes
= arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_DATA
);
3884 total_evicted
+= bytes
;
3889 arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_METADATA
);
3891 return (total_evicted
);
3895 arc_flush(spa_t
*spa
, boolean_t retry
)
3900 * If retry is B_TRUE, a spa must not be specified since we have
3901 * no good way to determine if all of a spa's buffers have been
3902 * evicted from an arc state.
3904 ASSERT(!retry
|| spa
== 0);
3907 guid
= spa_load_guid(spa
);
3909 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_DATA
, retry
);
3910 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_METADATA
, retry
);
3912 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_DATA
, retry
);
3913 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_METADATA
, retry
);
3915 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3916 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3918 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3919 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3923 arc_shrink(int64_t to_free
)
3927 if (c
> to_free
&& c
- to_free
> arc_c_min
) {
3928 arc_c
= c
- to_free
;
3929 atomic_add_64(&arc_p
, -(arc_p
>> arc_shrink_shift
));
3930 if (arc_c
> arc_size
)
3931 arc_c
= MAX(arc_size
, arc_c_min
);
3933 arc_p
= (arc_c
>> 1);
3934 ASSERT(arc_c
>= arc_c_min
);
3935 ASSERT((int64_t)arc_p
>= 0);
3940 if (arc_size
> arc_c
)
3941 (void) arc_adjust();
3945 * Return maximum amount of memory that we could possibly use. Reduced
3946 * to half of all memory in user space which is primarily used for testing.
3949 arc_all_memory(void)
3952 return (MIN(ptob(physmem
),
3953 vmem_size(heap_arena
, VMEM_FREE
| VMEM_ALLOC
)));
3955 return (ptob(physmem
) / 2);
3959 typedef enum free_memory_reason_t
{
3964 FMR_PAGES_PP_MAXIMUM
,
3967 } free_memory_reason_t
;
3969 int64_t last_free_memory
;
3970 free_memory_reason_t last_free_reason
;
3974 * Additional reserve of pages for pp_reserve.
3976 int64_t arc_pages_pp_reserve
= 64;
3979 * Additional reserve of pages for swapfs.
3981 int64_t arc_swapfs_reserve
= 64;
3982 #endif /* _KERNEL */
3985 * Return the amount of memory that can be consumed before reclaim will be
3986 * needed. Positive if there is sufficient free memory, negative indicates
3987 * the amount of memory that needs to be freed up.
3990 arc_available_memory(void)
3992 int64_t lowest
= INT64_MAX
;
3993 free_memory_reason_t r
= FMR_UNKNOWN
;
3995 uint64_t available_memory
= ptob(freemem
);
3998 pgcnt_t needfree
= btop(arc_need_free
);
3999 pgcnt_t lotsfree
= btop(arc_sys_free
);
4000 pgcnt_t desfree
= 0;
4005 MIN(available_memory
, vmem_size(heap_arena
, VMEM_FREE
));
4009 n
= PAGESIZE
* (-needfree
);
4017 * check that we're out of range of the pageout scanner. It starts to
4018 * schedule paging if freemem is less than lotsfree and needfree.
4019 * lotsfree is the high-water mark for pageout, and needfree is the
4020 * number of needed free pages. We add extra pages here to make sure
4021 * the scanner doesn't start up while we're freeing memory.
4023 n
= PAGESIZE
* (btop(available_memory
) - lotsfree
- needfree
- desfree
);
4031 * check to make sure that swapfs has enough space so that anon
4032 * reservations can still succeed. anon_resvmem() checks that the
4033 * availrmem is greater than swapfs_minfree, and the number of reserved
4034 * swap pages. We also add a bit of extra here just to prevent
4035 * circumstances from getting really dire.
4037 n
= PAGESIZE
* (availrmem
- swapfs_minfree
- swapfs_reserve
-
4038 desfree
- arc_swapfs_reserve
);
4041 r
= FMR_SWAPFS_MINFREE
;
4046 * Check that we have enough availrmem that memory locking (e.g., via
4047 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4048 * stores the number of pages that cannot be locked; when availrmem
4049 * drops below pages_pp_maximum, page locking mechanisms such as
4050 * page_pp_lock() will fail.)
4052 n
= PAGESIZE
* (availrmem
- pages_pp_maximum
-
4053 arc_pages_pp_reserve
);
4056 r
= FMR_PAGES_PP_MAXIMUM
;
4062 * If we're on an i386 platform, it's possible that we'll exhaust the
4063 * kernel heap space before we ever run out of available physical
4064 * memory. Most checks of the size of the heap_area compare against
4065 * tune.t_minarmem, which is the minimum available real memory that we
4066 * can have in the system. However, this is generally fixed at 25 pages
4067 * which is so low that it's useless. In this comparison, we seek to
4068 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4069 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4072 n
= vmem_size(heap_arena
, VMEM_FREE
) -
4073 (vmem_size(heap_arena
, VMEM_FREE
| VMEM_ALLOC
) >> 2);
4081 * If zio data pages are being allocated out of a separate heap segment,
4082 * then enforce that the size of available vmem for this arena remains
4083 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4085 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4086 * memory (in the zio_arena) free, which can avoid memory
4087 * fragmentation issues.
4089 if (zio_arena
!= NULL
) {
4090 n
= (int64_t)vmem_size(zio_arena
, VMEM_FREE
) -
4091 (vmem_size(zio_arena
, VMEM_ALLOC
) >>
4092 arc_zio_arena_free_shift
);
4099 /* Every 100 calls, free a small amount */
4100 if (spa_get_random(100) == 0)
4102 #endif /* _KERNEL */
4104 last_free_memory
= lowest
;
4105 last_free_reason
= r
;
4111 * Determine if the system is under memory pressure and is asking
4112 * to reclaim memory. A return value of B_TRUE indicates that the system
4113 * is under memory pressure and that the arc should adjust accordingly.
4116 arc_reclaim_needed(void)
4118 return (arc_available_memory() < 0);
4122 arc_kmem_reap_now(void)
4125 kmem_cache_t
*prev_cache
= NULL
;
4126 kmem_cache_t
*prev_data_cache
= NULL
;
4127 extern kmem_cache_t
*zio_buf_cache
[];
4128 extern kmem_cache_t
*zio_data_buf_cache
[];
4129 extern kmem_cache_t
*range_seg_cache
;
4131 if ((arc_meta_used
>= arc_meta_limit
) && zfs_arc_meta_prune
) {
4133 * We are exceeding our meta-data cache limit.
4134 * Prune some entries to release holds on meta-data.
4136 arc_prune_async(zfs_arc_meta_prune
);
4139 for (i
= 0; i
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; i
++) {
4141 /* reach upper limit of cache size on 32-bit */
4142 if (zio_buf_cache
[i
] == NULL
)
4145 if (zio_buf_cache
[i
] != prev_cache
) {
4146 prev_cache
= zio_buf_cache
[i
];
4147 kmem_cache_reap_now(zio_buf_cache
[i
]);
4149 if (zio_data_buf_cache
[i
] != prev_data_cache
) {
4150 prev_data_cache
= zio_data_buf_cache
[i
];
4151 kmem_cache_reap_now(zio_data_buf_cache
[i
]);
4154 kmem_cache_reap_now(buf_cache
);
4155 kmem_cache_reap_now(hdr_full_cache
);
4156 kmem_cache_reap_now(hdr_l2only_cache
);
4157 kmem_cache_reap_now(range_seg_cache
);
4159 if (zio_arena
!= NULL
) {
4161 * Ask the vmem arena to reclaim unused memory from its
4164 vmem_qcache_reap(zio_arena
);
4169 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4170 * enough data and signal them to proceed. When this happens, the threads in
4171 * arc_get_data_impl() are sleeping while holding the hash lock for their
4172 * particular arc header. Thus, we must be careful to never sleep on a
4173 * hash lock in this thread. This is to prevent the following deadlock:
4175 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4176 * waiting for the reclaim thread to signal it.
4178 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4179 * fails, and goes to sleep forever.
4181 * This possible deadlock is avoided by always acquiring a hash lock
4182 * using mutex_tryenter() from arc_reclaim_thread().
4185 arc_reclaim_thread(void)
4187 fstrans_cookie_t cookie
= spl_fstrans_mark();
4188 hrtime_t growtime
= 0;
4191 CALLB_CPR_INIT(&cpr
, &arc_reclaim_lock
, callb_generic_cpr
, FTAG
);
4193 mutex_enter(&arc_reclaim_lock
);
4194 while (!arc_reclaim_thread_exit
) {
4196 int64_t free_memory
= arc_available_memory();
4197 uint64_t evicted
= 0;
4199 arc_tuning_update();
4202 * This is necessary in order for the mdb ::arc dcmd to
4203 * show up to date information. Since the ::arc command
4204 * does not call the kstat's update function, without
4205 * this call, the command may show stale stats for the
4206 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4207 * with this change, the data might be up to 1 second
4208 * out of date; but that should suffice. The arc_state_t
4209 * structures can be queried directly if more accurate
4210 * information is needed.
4213 if (arc_ksp
!= NULL
)
4214 arc_ksp
->ks_update(arc_ksp
, KSTAT_READ
);
4216 mutex_exit(&arc_reclaim_lock
);
4218 if (free_memory
< 0) {
4220 arc_no_grow
= B_TRUE
;
4224 * Wait at least zfs_grow_retry (default 5) seconds
4225 * before considering growing.
4227 growtime
= gethrtime() + SEC2NSEC(arc_grow_retry
);
4229 arc_kmem_reap_now();
4232 * If we are still low on memory, shrink the ARC
4233 * so that we have arc_shrink_min free space.
4235 free_memory
= arc_available_memory();
4237 to_free
= (arc_c
>> arc_shrink_shift
) - free_memory
;
4240 to_free
= MAX(to_free
, arc_need_free
);
4242 arc_shrink(to_free
);
4244 } else if (free_memory
< arc_c
>> arc_no_grow_shift
) {
4245 arc_no_grow
= B_TRUE
;
4246 } else if (gethrtime() >= growtime
) {
4247 arc_no_grow
= B_FALSE
;
4250 evicted
= arc_adjust();
4252 mutex_enter(&arc_reclaim_lock
);
4255 * If evicted is zero, we couldn't evict anything via
4256 * arc_adjust(). This could be due to hash lock
4257 * collisions, but more likely due to the majority of
4258 * arc buffers being unevictable. Therefore, even if
4259 * arc_size is above arc_c, another pass is unlikely to
4260 * be helpful and could potentially cause us to enter an
4263 if (arc_size
<= arc_c
|| evicted
== 0) {
4265 * We're either no longer overflowing, or we
4266 * can't evict anything more, so we should wake
4267 * up any threads before we go to sleep and clear
4268 * arc_need_free since nothing more can be done.
4270 cv_broadcast(&arc_reclaim_waiters_cv
);
4274 * Block until signaled, or after one second (we
4275 * might need to perform arc_kmem_reap_now()
4276 * even if we aren't being signalled)
4278 CALLB_CPR_SAFE_BEGIN(&cpr
);
4279 (void) cv_timedwait_sig_hires(&arc_reclaim_thread_cv
,
4280 &arc_reclaim_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
4281 CALLB_CPR_SAFE_END(&cpr
, &arc_reclaim_lock
);
4285 arc_reclaim_thread_exit
= B_FALSE
;
4286 cv_broadcast(&arc_reclaim_thread_cv
);
4287 CALLB_CPR_EXIT(&cpr
); /* drops arc_reclaim_lock */
4288 spl_fstrans_unmark(cookie
);
4294 * Determine the amount of memory eligible for eviction contained in the
4295 * ARC. All clean data reported by the ghost lists can always be safely
4296 * evicted. Due to arc_c_min, the same does not hold for all clean data
4297 * contained by the regular mru and mfu lists.
4299 * In the case of the regular mru and mfu lists, we need to report as
4300 * much clean data as possible, such that evicting that same reported
4301 * data will not bring arc_size below arc_c_min. Thus, in certain
4302 * circumstances, the total amount of clean data in the mru and mfu
4303 * lists might not actually be evictable.
4305 * The following two distinct cases are accounted for:
4307 * 1. The sum of the amount of dirty data contained by both the mru and
4308 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4309 * is greater than or equal to arc_c_min.
4310 * (i.e. amount of dirty data >= arc_c_min)
4312 * This is the easy case; all clean data contained by the mru and mfu
4313 * lists is evictable. Evicting all clean data can only drop arc_size
4314 * to the amount of dirty data, which is greater than arc_c_min.
4316 * 2. The sum of the amount of dirty data contained by both the mru and
4317 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4318 * is less than arc_c_min.
4319 * (i.e. arc_c_min > amount of dirty data)
4321 * 2.1. arc_size is greater than or equal arc_c_min.
4322 * (i.e. arc_size >= arc_c_min > amount of dirty data)
4324 * In this case, not all clean data from the regular mru and mfu
4325 * lists is actually evictable; we must leave enough clean data
4326 * to keep arc_size above arc_c_min. Thus, the maximum amount of
4327 * evictable data from the two lists combined, is exactly the
4328 * difference between arc_size and arc_c_min.
4330 * 2.2. arc_size is less than arc_c_min
4331 * (i.e. arc_c_min > arc_size > amount of dirty data)
4333 * In this case, none of the data contained in the mru and mfu
4334 * lists is evictable, even if it's clean. Since arc_size is
4335 * already below arc_c_min, evicting any more would only
4336 * increase this negative difference.
4339 arc_evictable_memory(void) {
4340 uint64_t arc_clean
=
4341 refcount_count(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]) +
4342 refcount_count(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]) +
4343 refcount_count(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]) +
4344 refcount_count(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
4345 uint64_t ghost_clean
=
4346 refcount_count(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_DATA
]) +
4347 refcount_count(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_METADATA
]) +
4348 refcount_count(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_DATA
]) +
4349 refcount_count(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
4350 uint64_t arc_dirty
= MAX((int64_t)arc_size
- (int64_t)arc_clean
, 0);
4352 if (arc_dirty
>= arc_c_min
)
4353 return (ghost_clean
+ arc_clean
);
4355 return (ghost_clean
+ MAX((int64_t)arc_size
- (int64_t)arc_c_min
, 0));
4359 * If sc->nr_to_scan is zero, the caller is requesting a query of the
4360 * number of objects which can potentially be freed. If it is nonzero,
4361 * the request is to free that many objects.
4363 * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
4364 * in struct shrinker and also require the shrinker to return the number
4367 * Older kernels require the shrinker to return the number of freeable
4368 * objects following the freeing of nr_to_free.
4370 static spl_shrinker_t
4371 __arc_shrinker_func(struct shrinker
*shrink
, struct shrink_control
*sc
)
4375 /* The arc is considered warm once reclaim has occurred */
4376 if (unlikely(arc_warm
== B_FALSE
))
4379 /* Return the potential number of reclaimable pages */
4380 pages
= btop((int64_t)arc_evictable_memory());
4381 if (sc
->nr_to_scan
== 0)
4384 /* Not allowed to perform filesystem reclaim */
4385 if (!(sc
->gfp_mask
& __GFP_FS
))
4386 return (SHRINK_STOP
);
4388 /* Reclaim in progress */
4389 if (mutex_tryenter(&arc_reclaim_lock
) == 0)
4390 return (SHRINK_STOP
);
4392 mutex_exit(&arc_reclaim_lock
);
4395 * Evict the requested number of pages by shrinking arc_c the
4396 * requested amount. If there is nothing left to evict just
4397 * reap whatever we can from the various arc slabs.
4400 arc_shrink(ptob(sc
->nr_to_scan
));
4401 arc_kmem_reap_now();
4402 #ifdef HAVE_SPLIT_SHRINKER_CALLBACK
4403 pages
= MAX(pages
- btop(arc_evictable_memory()), 0);
4405 pages
= btop(arc_evictable_memory());
4408 arc_kmem_reap_now();
4409 pages
= SHRINK_STOP
;
4413 * We've reaped what we can, wake up threads.
4415 cv_broadcast(&arc_reclaim_waiters_cv
);
4418 * When direct reclaim is observed it usually indicates a rapid
4419 * increase in memory pressure. This occurs because the kswapd
4420 * threads were unable to asynchronously keep enough free memory
4421 * available. In this case set arc_no_grow to briefly pause arc
4422 * growth to avoid compounding the memory pressure.
4424 if (current_is_kswapd()) {
4425 ARCSTAT_BUMP(arcstat_memory_indirect_count
);
4427 arc_no_grow
= B_TRUE
;
4428 arc_need_free
= ptob(sc
->nr_to_scan
);
4429 ARCSTAT_BUMP(arcstat_memory_direct_count
);
4434 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func
);
4436 SPL_SHRINKER_DECLARE(arc_shrinker
, arc_shrinker_func
, DEFAULT_SEEKS
);
4437 #endif /* _KERNEL */
4440 * Adapt arc info given the number of bytes we are trying to add and
4441 * the state that we are comming from. This function is only called
4442 * when we are adding new content to the cache.
4445 arc_adapt(int bytes
, arc_state_t
*state
)
4448 uint64_t arc_p_min
= (arc_c
>> arc_p_min_shift
);
4449 int64_t mrug_size
= refcount_count(&arc_mru_ghost
->arcs_size
);
4450 int64_t mfug_size
= refcount_count(&arc_mfu_ghost
->arcs_size
);
4452 if (state
== arc_l2c_only
)
4457 * Adapt the target size of the MRU list:
4458 * - if we just hit in the MRU ghost list, then increase
4459 * the target size of the MRU list.
4460 * - if we just hit in the MFU ghost list, then increase
4461 * the target size of the MFU list by decreasing the
4462 * target size of the MRU list.
4464 if (state
== arc_mru_ghost
) {
4465 mult
= (mrug_size
>= mfug_size
) ? 1 : (mfug_size
/ mrug_size
);
4466 if (!zfs_arc_p_dampener_disable
)
4467 mult
= MIN(mult
, 10); /* avoid wild arc_p adjustment */
4469 arc_p
= MIN(arc_c
- arc_p_min
, arc_p
+ bytes
* mult
);
4470 } else if (state
== arc_mfu_ghost
) {
4473 mult
= (mfug_size
>= mrug_size
) ? 1 : (mrug_size
/ mfug_size
);
4474 if (!zfs_arc_p_dampener_disable
)
4475 mult
= MIN(mult
, 10);
4477 delta
= MIN(bytes
* mult
, arc_p
);
4478 arc_p
= MAX(arc_p_min
, arc_p
- delta
);
4480 ASSERT((int64_t)arc_p
>= 0);
4482 if (arc_reclaim_needed()) {
4483 cv_signal(&arc_reclaim_thread_cv
);
4490 if (arc_c
>= arc_c_max
)
4494 * If we're within (2 * maxblocksize) bytes of the target
4495 * cache size, increment the target cache size
4497 ASSERT3U(arc_c
, >=, 2ULL << SPA_MAXBLOCKSHIFT
);
4498 if (arc_size
>= arc_c
- (2ULL << SPA_MAXBLOCKSHIFT
)) {
4499 atomic_add_64(&arc_c
, (int64_t)bytes
);
4500 if (arc_c
> arc_c_max
)
4502 else if (state
== arc_anon
)
4503 atomic_add_64(&arc_p
, (int64_t)bytes
);
4507 ASSERT((int64_t)arc_p
>= 0);
4511 * Check if arc_size has grown past our upper threshold, determined by
4512 * zfs_arc_overflow_shift.
4515 arc_is_overflowing(void)
4517 /* Always allow at least one block of overflow */
4518 uint64_t overflow
= MAX(SPA_MAXBLOCKSIZE
,
4519 arc_c
>> zfs_arc_overflow_shift
);
4521 return (arc_size
>= arc_c
+ overflow
);
4525 arc_get_data_abd(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4527 arc_buf_contents_t type
= arc_buf_type(hdr
);
4529 arc_get_data_impl(hdr
, size
, tag
);
4530 if (type
== ARC_BUFC_METADATA
) {
4531 return (abd_alloc(size
, B_TRUE
));
4533 ASSERT(type
== ARC_BUFC_DATA
);
4534 return (abd_alloc(size
, B_FALSE
));
4539 arc_get_data_buf(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4541 arc_buf_contents_t type
= arc_buf_type(hdr
);
4543 arc_get_data_impl(hdr
, size
, tag
);
4544 if (type
== ARC_BUFC_METADATA
) {
4545 return (zio_buf_alloc(size
));
4547 ASSERT(type
== ARC_BUFC_DATA
);
4548 return (zio_data_buf_alloc(size
));
4553 * Allocate a block and return it to the caller. If we are hitting the
4554 * hard limit for the cache size, we must sleep, waiting for the eviction
4555 * thread to catch up. If we're past the target size but below the hard
4556 * limit, we'll only signal the reclaim thread and continue on.
4559 arc_get_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4561 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
4562 arc_buf_contents_t type
= arc_buf_type(hdr
);
4564 arc_adapt(size
, state
);
4567 * If arc_size is currently overflowing, and has grown past our
4568 * upper limit, we must be adding data faster than the evict
4569 * thread can evict. Thus, to ensure we don't compound the
4570 * problem by adding more data and forcing arc_size to grow even
4571 * further past it's target size, we halt and wait for the
4572 * eviction thread to catch up.
4574 * It's also possible that the reclaim thread is unable to evict
4575 * enough buffers to get arc_size below the overflow limit (e.g.
4576 * due to buffers being un-evictable, or hash lock collisions).
4577 * In this case, we want to proceed regardless if we're
4578 * overflowing; thus we don't use a while loop here.
4580 if (arc_is_overflowing()) {
4581 mutex_enter(&arc_reclaim_lock
);
4584 * Now that we've acquired the lock, we may no longer be
4585 * over the overflow limit, lets check.
4587 * We're ignoring the case of spurious wake ups. If that
4588 * were to happen, it'd let this thread consume an ARC
4589 * buffer before it should have (i.e. before we're under
4590 * the overflow limit and were signalled by the reclaim
4591 * thread). As long as that is a rare occurrence, it
4592 * shouldn't cause any harm.
4594 if (arc_is_overflowing()) {
4595 cv_signal(&arc_reclaim_thread_cv
);
4596 cv_wait(&arc_reclaim_waiters_cv
, &arc_reclaim_lock
);
4599 mutex_exit(&arc_reclaim_lock
);
4602 VERIFY3U(hdr
->b_type
, ==, type
);
4603 if (type
== ARC_BUFC_METADATA
) {
4604 arc_space_consume(size
, ARC_SPACE_META
);
4606 arc_space_consume(size
, ARC_SPACE_DATA
);
4610 * Update the state size. Note that ghost states have a
4611 * "ghost size" and so don't need to be updated.
4613 if (!GHOST_STATE(state
)) {
4615 (void) refcount_add_many(&state
->arcs_size
, size
, tag
);
4618 * If this is reached via arc_read, the link is
4619 * protected by the hash lock. If reached via
4620 * arc_buf_alloc, the header should not be accessed by
4621 * any other thread. And, if reached via arc_read_done,
4622 * the hash lock will protect it if it's found in the
4623 * hash table; otherwise no other thread should be
4624 * trying to [add|remove]_reference it.
4626 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
4627 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4628 (void) refcount_add_many(&state
->arcs_esize
[type
],
4633 * If we are growing the cache, and we are adding anonymous
4634 * data, and we have outgrown arc_p, update arc_p
4636 if (arc_size
< arc_c
&& hdr
->b_l1hdr
.b_state
== arc_anon
&&
4637 (refcount_count(&arc_anon
->arcs_size
) +
4638 refcount_count(&arc_mru
->arcs_size
) > arc_p
))
4639 arc_p
= MIN(arc_c
, arc_p
+ size
);
4644 arc_free_data_abd(arc_buf_hdr_t
*hdr
, abd_t
*abd
, uint64_t size
, void *tag
)
4646 arc_free_data_impl(hdr
, size
, tag
);
4651 arc_free_data_buf(arc_buf_hdr_t
*hdr
, void *buf
, uint64_t size
, void *tag
)
4653 arc_buf_contents_t type
= arc_buf_type(hdr
);
4655 arc_free_data_impl(hdr
, size
, tag
);
4656 if (type
== ARC_BUFC_METADATA
) {
4657 zio_buf_free(buf
, size
);
4659 ASSERT(type
== ARC_BUFC_DATA
);
4660 zio_data_buf_free(buf
, size
);
4665 * Free the arc data buffer.
4668 arc_free_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4670 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
4671 arc_buf_contents_t type
= arc_buf_type(hdr
);
4673 /* protected by hash lock, if in the hash table */
4674 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
4675 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4676 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
4678 (void) refcount_remove_many(&state
->arcs_esize
[type
],
4681 (void) refcount_remove_many(&state
->arcs_size
, size
, tag
);
4683 VERIFY3U(hdr
->b_type
, ==, type
);
4684 if (type
== ARC_BUFC_METADATA
) {
4685 arc_space_return(size
, ARC_SPACE_META
);
4687 ASSERT(type
== ARC_BUFC_DATA
);
4688 arc_space_return(size
, ARC_SPACE_DATA
);
4693 * This routine is called whenever a buffer is accessed.
4694 * NOTE: the hash lock is dropped in this function.
4697 arc_access(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
4701 ASSERT(MUTEX_HELD(hash_lock
));
4702 ASSERT(HDR_HAS_L1HDR(hdr
));
4704 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
4706 * This buffer is not in the cache, and does not
4707 * appear in our "ghost" list. Add the new buffer
4711 ASSERT0(hdr
->b_l1hdr
.b_arc_access
);
4712 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4713 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
4714 arc_change_state(arc_mru
, hdr
, hash_lock
);
4716 } else if (hdr
->b_l1hdr
.b_state
== arc_mru
) {
4717 now
= ddi_get_lbolt();
4720 * If this buffer is here because of a prefetch, then either:
4721 * - clear the flag if this is a "referencing" read
4722 * (any subsequent access will bump this into the MFU state).
4724 * - move the buffer to the head of the list if this is
4725 * another prefetch (to make it less likely to be evicted).
4727 if (HDR_PREFETCH(hdr
)) {
4728 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
4729 /* link protected by hash lock */
4730 ASSERT(multilist_link_active(
4731 &hdr
->b_l1hdr
.b_arc_node
));
4733 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
4734 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
4735 ARCSTAT_BUMP(arcstat_mru_hits
);
4737 hdr
->b_l1hdr
.b_arc_access
= now
;
4742 * This buffer has been "accessed" only once so far,
4743 * but it is still in the cache. Move it to the MFU
4746 if (ddi_time_after(now
, hdr
->b_l1hdr
.b_arc_access
+
4749 * More than 125ms have passed since we
4750 * instantiated this buffer. Move it to the
4751 * most frequently used state.
4753 hdr
->b_l1hdr
.b_arc_access
= now
;
4754 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4755 arc_change_state(arc_mfu
, hdr
, hash_lock
);
4757 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
4758 ARCSTAT_BUMP(arcstat_mru_hits
);
4759 } else if (hdr
->b_l1hdr
.b_state
== arc_mru_ghost
) {
4760 arc_state_t
*new_state
;
4762 * This buffer has been "accessed" recently, but
4763 * was evicted from the cache. Move it to the
4767 if (HDR_PREFETCH(hdr
)) {
4768 new_state
= arc_mru
;
4769 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) > 0)
4770 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
4771 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
4773 new_state
= arc_mfu
;
4774 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4777 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4778 arc_change_state(new_state
, hdr
, hash_lock
);
4780 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_ghost_hits
);
4781 ARCSTAT_BUMP(arcstat_mru_ghost_hits
);
4782 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu
) {
4784 * This buffer has been accessed more than once and is
4785 * still in the cache. Keep it in the MFU state.
4787 * NOTE: an add_reference() that occurred when we did
4788 * the arc_read() will have kicked this off the list.
4789 * If it was a prefetch, we will explicitly move it to
4790 * the head of the list now.
4792 if ((HDR_PREFETCH(hdr
)) != 0) {
4793 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4794 /* link protected by hash_lock */
4795 ASSERT(multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
4797 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_hits
);
4798 ARCSTAT_BUMP(arcstat_mfu_hits
);
4799 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4800 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu_ghost
) {
4801 arc_state_t
*new_state
= arc_mfu
;
4803 * This buffer has been accessed more than once but has
4804 * been evicted from the cache. Move it back to the
4808 if (HDR_PREFETCH(hdr
)) {
4810 * This is a prefetch access...
4811 * move this block back to the MRU state.
4813 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
4814 new_state
= arc_mru
;
4817 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4818 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4819 arc_change_state(new_state
, hdr
, hash_lock
);
4821 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_ghost_hits
);
4822 ARCSTAT_BUMP(arcstat_mfu_ghost_hits
);
4823 } else if (hdr
->b_l1hdr
.b_state
== arc_l2c_only
) {
4825 * This buffer is on the 2nd Level ARC.
4828 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4829 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4830 arc_change_state(arc_mfu
, hdr
, hash_lock
);
4832 cmn_err(CE_PANIC
, "invalid arc state 0x%p",
4833 hdr
->b_l1hdr
.b_state
);
4837 /* a generic arc_done_func_t which you can use */
4840 arc_bcopy_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4842 if (zio
== NULL
|| zio
->io_error
== 0)
4843 bcopy(buf
->b_data
, arg
, arc_buf_size(buf
));
4844 arc_buf_destroy(buf
, arg
);
4847 /* a generic arc_done_func_t */
4849 arc_getbuf_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4851 arc_buf_t
**bufp
= arg
;
4852 if (zio
&& zio
->io_error
) {
4853 arc_buf_destroy(buf
, arg
);
4857 ASSERT(buf
->b_data
);
4862 arc_hdr_verify(arc_buf_hdr_t
*hdr
, blkptr_t
*bp
)
4864 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
)) {
4865 ASSERT3U(HDR_GET_PSIZE(hdr
), ==, 0);
4866 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
4868 if (HDR_COMPRESSION_ENABLED(hdr
)) {
4869 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==,
4870 BP_GET_COMPRESS(bp
));
4872 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, BP_GET_LSIZE(bp
));
4873 ASSERT3U(HDR_GET_PSIZE(hdr
), ==, BP_GET_PSIZE(bp
));
4878 arc_read_done(zio_t
*zio
)
4880 arc_buf_hdr_t
*hdr
= zio
->io_private
;
4881 kmutex_t
*hash_lock
= NULL
;
4882 arc_callback_t
*callback_list
;
4883 arc_callback_t
*acb
;
4884 boolean_t freeable
= B_FALSE
;
4885 boolean_t no_zio_error
= (zio
->io_error
== 0);
4886 int callback_cnt
= 0;
4888 * The hdr was inserted into hash-table and removed from lists
4889 * prior to starting I/O. We should find this header, since
4890 * it's in the hash table, and it should be legit since it's
4891 * not possible to evict it during the I/O. The only possible
4892 * reason for it not to be found is if we were freed during the
4895 if (HDR_IN_HASH_TABLE(hdr
)) {
4896 arc_buf_hdr_t
*found
;
4898 ASSERT3U(hdr
->b_birth
, ==, BP_PHYSICAL_BIRTH(zio
->io_bp
));
4899 ASSERT3U(hdr
->b_dva
.dva_word
[0], ==,
4900 BP_IDENTITY(zio
->io_bp
)->dva_word
[0]);
4901 ASSERT3U(hdr
->b_dva
.dva_word
[1], ==,
4902 BP_IDENTITY(zio
->io_bp
)->dva_word
[1]);
4904 found
= buf_hash_find(hdr
->b_spa
, zio
->io_bp
, &hash_lock
);
4906 ASSERT((found
== hdr
&&
4907 DVA_EQUAL(&hdr
->b_dva
, BP_IDENTITY(zio
->io_bp
))) ||
4908 (found
== hdr
&& HDR_L2_READING(hdr
)));
4909 ASSERT3P(hash_lock
, !=, NULL
);
4913 /* byteswap if necessary */
4914 if (BP_SHOULD_BYTESWAP(zio
->io_bp
)) {
4915 if (BP_GET_LEVEL(zio
->io_bp
) > 0) {
4916 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_UINT64
;
4918 hdr
->b_l1hdr
.b_byteswap
=
4919 DMU_OT_BYTESWAP(BP_GET_TYPE(zio
->io_bp
));
4922 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
4926 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2_EVICTED
);
4927 if (l2arc_noprefetch
&& HDR_PREFETCH(hdr
))
4928 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2CACHE
);
4930 callback_list
= hdr
->b_l1hdr
.b_acb
;
4931 ASSERT3P(callback_list
, !=, NULL
);
4933 if (hash_lock
&& no_zio_error
&& hdr
->b_l1hdr
.b_state
== arc_anon
) {
4935 * Only call arc_access on anonymous buffers. This is because
4936 * if we've issued an I/O for an evicted buffer, we've already
4937 * called arc_access (to prevent any simultaneous readers from
4938 * getting confused).
4940 arc_access(hdr
, hash_lock
);
4944 * If a read request has a callback (i.e. acb_done is not NULL), then we
4945 * make a buf containing the data according to the parameters which were
4946 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
4947 * aren't needlessly decompressing the data multiple times.
4949 for (acb
= callback_list
; acb
!= NULL
; acb
= acb
->acb_next
) {
4954 /* This is a demand read since prefetches don't use callbacks */
4958 error
= arc_buf_alloc_impl(hdr
, acb
->acb_private
,
4959 acb
->acb_compressed
, no_zio_error
, &acb
->acb_buf
);
4961 zio
->io_error
= error
;
4964 hdr
->b_l1hdr
.b_acb
= NULL
;
4965 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
4966 if (callback_cnt
== 0) {
4967 ASSERT(HDR_PREFETCH(hdr
));
4968 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
4969 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
4972 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
) ||
4973 callback_list
!= NULL
);
4976 arc_hdr_verify(hdr
, zio
->io_bp
);
4978 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_ERROR
);
4979 if (hdr
->b_l1hdr
.b_state
!= arc_anon
)
4980 arc_change_state(arc_anon
, hdr
, hash_lock
);
4981 if (HDR_IN_HASH_TABLE(hdr
))
4982 buf_hash_remove(hdr
);
4983 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
4987 * Broadcast before we drop the hash_lock to avoid the possibility
4988 * that the hdr (and hence the cv) might be freed before we get to
4989 * the cv_broadcast().
4991 cv_broadcast(&hdr
->b_l1hdr
.b_cv
);
4993 if (hash_lock
!= NULL
) {
4994 mutex_exit(hash_lock
);
4997 * This block was freed while we waited for the read to
4998 * complete. It has been removed from the hash table and
4999 * moved to the anonymous state (so that it won't show up
5002 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
5003 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
5006 /* execute each callback and free its structure */
5007 while ((acb
= callback_list
) != NULL
) {
5009 acb
->acb_done(zio
, acb
->acb_buf
, acb
->acb_private
);
5011 if (acb
->acb_zio_dummy
!= NULL
) {
5012 acb
->acb_zio_dummy
->io_error
= zio
->io_error
;
5013 zio_nowait(acb
->acb_zio_dummy
);
5016 callback_list
= acb
->acb_next
;
5017 kmem_free(acb
, sizeof (arc_callback_t
));
5021 arc_hdr_destroy(hdr
);
5025 * "Read" the block at the specified DVA (in bp) via the
5026 * cache. If the block is found in the cache, invoke the provided
5027 * callback immediately and return. Note that the `zio' parameter
5028 * in the callback will be NULL in this case, since no IO was
5029 * required. If the block is not in the cache pass the read request
5030 * on to the spa with a substitute callback function, so that the
5031 * requested block will be added to the cache.
5033 * If a read request arrives for a block that has a read in-progress,
5034 * either wait for the in-progress read to complete (and return the
5035 * results); or, if this is a read with a "done" func, add a record
5036 * to the read to invoke the "done" func when the read completes,
5037 * and return; or just return.
5039 * arc_read_done() will invoke all the requested "done" functions
5040 * for readers of this block.
5043 arc_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, arc_done_func_t
*done
,
5044 void *private, zio_priority_t priority
, int zio_flags
,
5045 arc_flags_t
*arc_flags
, const zbookmark_phys_t
*zb
)
5047 arc_buf_hdr_t
*hdr
= NULL
;
5048 kmutex_t
*hash_lock
= NULL
;
5050 uint64_t guid
= spa_load_guid(spa
);
5051 boolean_t compressed_read
= (zio_flags
& ZIO_FLAG_RAW
) != 0;
5054 ASSERT(!BP_IS_EMBEDDED(bp
) ||
5055 BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
);
5058 if (!BP_IS_EMBEDDED(bp
)) {
5060 * Embedded BP's have no DVA and require no I/O to "read".
5061 * Create an anonymous arc buf to back it.
5063 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
5066 if (hdr
!= NULL
&& HDR_HAS_L1HDR(hdr
) && hdr
->b_l1hdr
.b_pabd
!= NULL
) {
5067 arc_buf_t
*buf
= NULL
;
5068 *arc_flags
|= ARC_FLAG_CACHED
;
5070 if (HDR_IO_IN_PROGRESS(hdr
)) {
5072 if ((hdr
->b_flags
& ARC_FLAG_PRIO_ASYNC_READ
) &&
5073 priority
== ZIO_PRIORITY_SYNC_READ
) {
5075 * This sync read must wait for an
5076 * in-progress async read (e.g. a predictive
5077 * prefetch). Async reads are queued
5078 * separately at the vdev_queue layer, so
5079 * this is a form of priority inversion.
5080 * Ideally, we would "inherit" the demand
5081 * i/o's priority by moving the i/o from
5082 * the async queue to the synchronous queue,
5083 * but there is currently no mechanism to do
5084 * so. Track this so that we can evaluate
5085 * the magnitude of this potential performance
5088 * Note that if the prefetch i/o is already
5089 * active (has been issued to the device),
5090 * the prefetch improved performance, because
5091 * we issued it sooner than we would have
5092 * without the prefetch.
5094 DTRACE_PROBE1(arc__sync__wait__for__async
,
5095 arc_buf_hdr_t
*, hdr
);
5096 ARCSTAT_BUMP(arcstat_sync_wait_for_async
);
5098 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
5099 arc_hdr_clear_flags(hdr
,
5100 ARC_FLAG_PREDICTIVE_PREFETCH
);
5103 if (*arc_flags
& ARC_FLAG_WAIT
) {
5104 cv_wait(&hdr
->b_l1hdr
.b_cv
, hash_lock
);
5105 mutex_exit(hash_lock
);
5108 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
5111 arc_callback_t
*acb
= NULL
;
5113 acb
= kmem_zalloc(sizeof (arc_callback_t
),
5115 acb
->acb_done
= done
;
5116 acb
->acb_private
= private;
5118 acb
->acb_zio_dummy
= zio_null(pio
,
5119 spa
, NULL
, NULL
, NULL
, zio_flags
);
5121 ASSERT3P(acb
->acb_done
, !=, NULL
);
5122 acb
->acb_next
= hdr
->b_l1hdr
.b_acb
;
5123 hdr
->b_l1hdr
.b_acb
= acb
;
5124 mutex_exit(hash_lock
);
5127 mutex_exit(hash_lock
);
5131 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
5132 hdr
->b_l1hdr
.b_state
== arc_mfu
);
5135 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
5137 * This is a demand read which does not have to
5138 * wait for i/o because we did a predictive
5139 * prefetch i/o for it, which has completed.
5142 arc__demand__hit__predictive__prefetch
,
5143 arc_buf_hdr_t
*, hdr
);
5145 arcstat_demand_hit_predictive_prefetch
);
5146 arc_hdr_clear_flags(hdr
,
5147 ARC_FLAG_PREDICTIVE_PREFETCH
);
5149 ASSERT(!BP_IS_EMBEDDED(bp
) || !BP_IS_HOLE(bp
));
5151 /* Get a buf with the desired data in it. */
5152 VERIFY0(arc_buf_alloc_impl(hdr
, private,
5153 compressed_read
, B_TRUE
, &buf
));
5154 } else if (*arc_flags
& ARC_FLAG_PREFETCH
&&
5155 refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
5156 arc_hdr_set_flags(hdr
, ARC_FLAG_PREFETCH
);
5158 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
5159 arc_access(hdr
, hash_lock
);
5160 if (*arc_flags
& ARC_FLAG_L2CACHE
)
5161 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
5162 mutex_exit(hash_lock
);
5163 ARCSTAT_BUMP(arcstat_hits
);
5164 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
5165 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
5166 data
, metadata
, hits
);
5169 done(NULL
, buf
, private);
5171 uint64_t lsize
= BP_GET_LSIZE(bp
);
5172 uint64_t psize
= BP_GET_PSIZE(bp
);
5173 arc_callback_t
*acb
;
5176 boolean_t devw
= B_FALSE
;
5180 * Gracefully handle a damaged logical block size as a
5183 if (lsize
> spa_maxblocksize(spa
)) {
5184 rc
= SET_ERROR(ECKSUM
);
5189 /* this block is not in the cache */
5190 arc_buf_hdr_t
*exists
= NULL
;
5191 arc_buf_contents_t type
= BP_GET_BUFC_TYPE(bp
);
5192 hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
,
5193 BP_GET_COMPRESS(bp
), type
);
5195 if (!BP_IS_EMBEDDED(bp
)) {
5196 hdr
->b_dva
= *BP_IDENTITY(bp
);
5197 hdr
->b_birth
= BP_PHYSICAL_BIRTH(bp
);
5198 exists
= buf_hash_insert(hdr
, &hash_lock
);
5200 if (exists
!= NULL
) {
5201 /* somebody beat us to the hash insert */
5202 mutex_exit(hash_lock
);
5203 buf_discard_identity(hdr
);
5204 arc_hdr_destroy(hdr
);
5205 goto top
; /* restart the IO request */
5209 * This block is in the ghost cache. If it was L2-only
5210 * (and thus didn't have an L1 hdr), we realloc the
5211 * header to add an L1 hdr.
5213 if (!HDR_HAS_L1HDR(hdr
)) {
5214 hdr
= arc_hdr_realloc(hdr
, hdr_l2only_cache
,
5218 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
5219 ASSERT(GHOST_STATE(hdr
->b_l1hdr
.b_state
));
5220 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5221 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5222 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
5223 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
5226 * This is a delicate dance that we play here.
5227 * This hdr is in the ghost list so we access it
5228 * to move it out of the ghost list before we
5229 * initiate the read. If it's a prefetch then
5230 * it won't have a callback so we'll remove the
5231 * reference that arc_buf_alloc_impl() created. We
5232 * do this after we've called arc_access() to
5233 * avoid hitting an assert in remove_reference().
5235 arc_access(hdr
, hash_lock
);
5236 arc_hdr_alloc_pabd(hdr
);
5238 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
5239 size
= arc_hdr_size(hdr
);
5242 * If compression is enabled on the hdr, then will do
5243 * RAW I/O and will store the compressed data in the hdr's
5244 * data block. Otherwise, the hdr's data block will contain
5245 * the uncompressed data.
5247 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
) {
5248 zio_flags
|= ZIO_FLAG_RAW
;
5251 if (*arc_flags
& ARC_FLAG_PREFETCH
)
5252 arc_hdr_set_flags(hdr
, ARC_FLAG_PREFETCH
);
5253 if (*arc_flags
& ARC_FLAG_L2CACHE
)
5254 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
5255 if (BP_GET_LEVEL(bp
) > 0)
5256 arc_hdr_set_flags(hdr
, ARC_FLAG_INDIRECT
);
5257 if (*arc_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
)
5258 arc_hdr_set_flags(hdr
, ARC_FLAG_PREDICTIVE_PREFETCH
);
5259 ASSERT(!GHOST_STATE(hdr
->b_l1hdr
.b_state
));
5261 acb
= kmem_zalloc(sizeof (arc_callback_t
), KM_SLEEP
);
5262 acb
->acb_done
= done
;
5263 acb
->acb_private
= private;
5264 acb
->acb_compressed
= compressed_read
;
5266 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
5267 hdr
->b_l1hdr
.b_acb
= acb
;
5268 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5270 if (HDR_HAS_L2HDR(hdr
) &&
5271 (vd
= hdr
->b_l2hdr
.b_dev
->l2ad_vdev
) != NULL
) {
5272 devw
= hdr
->b_l2hdr
.b_dev
->l2ad_writing
;
5273 addr
= hdr
->b_l2hdr
.b_daddr
;
5275 * Lock out device removal.
5277 if (vdev_is_dead(vd
) ||
5278 !spa_config_tryenter(spa
, SCL_L2ARC
, vd
, RW_READER
))
5282 if (priority
== ZIO_PRIORITY_ASYNC_READ
)
5283 arc_hdr_set_flags(hdr
, ARC_FLAG_PRIO_ASYNC_READ
);
5285 arc_hdr_clear_flags(hdr
, ARC_FLAG_PRIO_ASYNC_READ
);
5287 if (hash_lock
!= NULL
)
5288 mutex_exit(hash_lock
);
5291 * At this point, we have a level 1 cache miss. Try again in
5292 * L2ARC if possible.
5294 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, lsize
);
5296 DTRACE_PROBE4(arc__miss
, arc_buf_hdr_t
*, hdr
, blkptr_t
*, bp
,
5297 uint64_t, lsize
, zbookmark_phys_t
*, zb
);
5298 ARCSTAT_BUMP(arcstat_misses
);
5299 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
5300 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
5301 data
, metadata
, misses
);
5303 if (vd
!= NULL
&& l2arc_ndev
!= 0 && !(l2arc_norw
&& devw
)) {
5305 * Read from the L2ARC if the following are true:
5306 * 1. The L2ARC vdev was previously cached.
5307 * 2. This buffer still has L2ARC metadata.
5308 * 3. This buffer isn't currently writing to the L2ARC.
5309 * 4. The L2ARC entry wasn't evicted, which may
5310 * also have invalidated the vdev.
5311 * 5. This isn't prefetch and l2arc_noprefetch is set.
5313 if (HDR_HAS_L2HDR(hdr
) &&
5314 !HDR_L2_WRITING(hdr
) && !HDR_L2_EVICTED(hdr
) &&
5315 !(l2arc_noprefetch
&& HDR_PREFETCH(hdr
))) {
5316 l2arc_read_callback_t
*cb
;
5318 DTRACE_PROBE1(l2arc__hit
, arc_buf_hdr_t
*, hdr
);
5319 ARCSTAT_BUMP(arcstat_l2_hits
);
5320 atomic_inc_32(&hdr
->b_l2hdr
.b_hits
);
5322 cb
= kmem_zalloc(sizeof (l2arc_read_callback_t
),
5324 cb
->l2rcb_hdr
= hdr
;
5327 cb
->l2rcb_flags
= zio_flags
;
5329 ASSERT(addr
>= VDEV_LABEL_START_SIZE
&&
5330 addr
+ lsize
< vd
->vdev_psize
-
5331 VDEV_LABEL_END_SIZE
);
5334 * l2arc read. The SCL_L2ARC lock will be
5335 * released by l2arc_read_done().
5336 * Issue a null zio if the underlying buffer
5337 * was squashed to zero size by compression.
5339 ASSERT3U(HDR_GET_COMPRESS(hdr
), !=,
5340 ZIO_COMPRESS_EMPTY
);
5341 rzio
= zio_read_phys(pio
, vd
, addr
,
5342 size
, hdr
->b_l1hdr
.b_pabd
,
5344 l2arc_read_done
, cb
, priority
,
5345 zio_flags
| ZIO_FLAG_DONT_CACHE
|
5347 ZIO_FLAG_DONT_PROPAGATE
|
5348 ZIO_FLAG_DONT_RETRY
, B_FALSE
);
5350 DTRACE_PROBE2(l2arc__read
, vdev_t
*, vd
,
5352 ARCSTAT_INCR(arcstat_l2_read_bytes
, size
);
5354 if (*arc_flags
& ARC_FLAG_NOWAIT
) {
5359 ASSERT(*arc_flags
& ARC_FLAG_WAIT
);
5360 if (zio_wait(rzio
) == 0)
5363 /* l2arc read error; goto zio_read() */
5365 DTRACE_PROBE1(l2arc__miss
,
5366 arc_buf_hdr_t
*, hdr
);
5367 ARCSTAT_BUMP(arcstat_l2_misses
);
5368 if (HDR_L2_WRITING(hdr
))
5369 ARCSTAT_BUMP(arcstat_l2_rw_clash
);
5370 spa_config_exit(spa
, SCL_L2ARC
, vd
);
5374 spa_config_exit(spa
, SCL_L2ARC
, vd
);
5375 if (l2arc_ndev
!= 0) {
5376 DTRACE_PROBE1(l2arc__miss
,
5377 arc_buf_hdr_t
*, hdr
);
5378 ARCSTAT_BUMP(arcstat_l2_misses
);
5382 rzio
= zio_read(pio
, spa
, bp
, hdr
->b_l1hdr
.b_pabd
, size
,
5383 arc_read_done
, hdr
, priority
, zio_flags
, zb
);
5385 if (*arc_flags
& ARC_FLAG_WAIT
) {
5386 rc
= zio_wait(rzio
);
5390 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
5395 spa_read_history_add(spa
, zb
, *arc_flags
);
5400 arc_add_prune_callback(arc_prune_func_t
*func
, void *private)
5404 p
= kmem_alloc(sizeof (*p
), KM_SLEEP
);
5406 p
->p_private
= private;
5407 list_link_init(&p
->p_node
);
5408 refcount_create(&p
->p_refcnt
);
5410 mutex_enter(&arc_prune_mtx
);
5411 refcount_add(&p
->p_refcnt
, &arc_prune_list
);
5412 list_insert_head(&arc_prune_list
, p
);
5413 mutex_exit(&arc_prune_mtx
);
5419 arc_remove_prune_callback(arc_prune_t
*p
)
5421 boolean_t wait
= B_FALSE
;
5422 mutex_enter(&arc_prune_mtx
);
5423 list_remove(&arc_prune_list
, p
);
5424 if (refcount_remove(&p
->p_refcnt
, &arc_prune_list
) > 0)
5426 mutex_exit(&arc_prune_mtx
);
5428 /* wait for arc_prune_task to finish */
5430 taskq_wait_outstanding(arc_prune_taskq
, 0);
5431 ASSERT0(refcount_count(&p
->p_refcnt
));
5432 refcount_destroy(&p
->p_refcnt
);
5433 kmem_free(p
, sizeof (*p
));
5437 * Notify the arc that a block was freed, and thus will never be used again.
5440 arc_freed(spa_t
*spa
, const blkptr_t
*bp
)
5443 kmutex_t
*hash_lock
;
5444 uint64_t guid
= spa_load_guid(spa
);
5446 ASSERT(!BP_IS_EMBEDDED(bp
));
5448 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
5453 * We might be trying to free a block that is still doing I/O
5454 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5455 * dmu_sync-ed block). If this block is being prefetched, then it
5456 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5457 * until the I/O completes. A block may also have a reference if it is
5458 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5459 * have written the new block to its final resting place on disk but
5460 * without the dedup flag set. This would have left the hdr in the MRU
5461 * state and discoverable. When the txg finally syncs it detects that
5462 * the block was overridden in open context and issues an override I/O.
5463 * Since this is a dedup block, the override I/O will determine if the
5464 * block is already in the DDT. If so, then it will replace the io_bp
5465 * with the bp from the DDT and allow the I/O to finish. When the I/O
5466 * reaches the done callback, dbuf_write_override_done, it will
5467 * check to see if the io_bp and io_bp_override are identical.
5468 * If they are not, then it indicates that the bp was replaced with
5469 * the bp in the DDT and the override bp is freed. This allows
5470 * us to arrive here with a reference on a block that is being
5471 * freed. So if we have an I/O in progress, or a reference to
5472 * this hdr, then we don't destroy the hdr.
5474 if (!HDR_HAS_L1HDR(hdr
) || (!HDR_IO_IN_PROGRESS(hdr
) &&
5475 refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
))) {
5476 arc_change_state(arc_anon
, hdr
, hash_lock
);
5477 arc_hdr_destroy(hdr
);
5478 mutex_exit(hash_lock
);
5480 mutex_exit(hash_lock
);
5486 * Release this buffer from the cache, making it an anonymous buffer. This
5487 * must be done after a read and prior to modifying the buffer contents.
5488 * If the buffer has more than one reference, we must make
5489 * a new hdr for the buffer.
5492 arc_release(arc_buf_t
*buf
, void *tag
)
5494 kmutex_t
*hash_lock
;
5496 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5499 * It would be nice to assert that if its DMU metadata (level >
5500 * 0 || it's the dnode file), then it must be syncing context.
5501 * But we don't know that information at this level.
5504 mutex_enter(&buf
->b_evict_lock
);
5506 ASSERT(HDR_HAS_L1HDR(hdr
));
5509 * We don't grab the hash lock prior to this check, because if
5510 * the buffer's header is in the arc_anon state, it won't be
5511 * linked into the hash table.
5513 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
5514 mutex_exit(&buf
->b_evict_lock
);
5515 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5516 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
5517 ASSERT(!HDR_HAS_L2HDR(hdr
));
5518 ASSERT(HDR_EMPTY(hdr
));
5520 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
5521 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), ==, 1);
5522 ASSERT(!list_link_active(&hdr
->b_l1hdr
.b_arc_node
));
5524 hdr
->b_l1hdr
.b_arc_access
= 0;
5527 * If the buf is being overridden then it may already
5528 * have a hdr that is not empty.
5530 buf_discard_identity(hdr
);
5536 hash_lock
= HDR_LOCK(hdr
);
5537 mutex_enter(hash_lock
);
5540 * This assignment is only valid as long as the hash_lock is
5541 * held, we must be careful not to reference state or the
5542 * b_state field after dropping the lock.
5544 state
= hdr
->b_l1hdr
.b_state
;
5545 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
5546 ASSERT3P(state
, !=, arc_anon
);
5548 /* this buffer is not on any list */
5549 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), >, 0);
5551 if (HDR_HAS_L2HDR(hdr
)) {
5552 mutex_enter(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
5555 * We have to recheck this conditional again now that
5556 * we're holding the l2ad_mtx to prevent a race with
5557 * another thread which might be concurrently calling
5558 * l2arc_evict(). In that case, l2arc_evict() might have
5559 * destroyed the header's L2 portion as we were waiting
5560 * to acquire the l2ad_mtx.
5562 if (HDR_HAS_L2HDR(hdr
))
5563 arc_hdr_l2hdr_destroy(hdr
);
5565 mutex_exit(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
5569 * Do we have more than one buf?
5571 if (hdr
->b_l1hdr
.b_bufcnt
> 1) {
5572 arc_buf_hdr_t
*nhdr
;
5573 uint64_t spa
= hdr
->b_spa
;
5574 uint64_t psize
= HDR_GET_PSIZE(hdr
);
5575 uint64_t lsize
= HDR_GET_LSIZE(hdr
);
5576 enum zio_compress compress
= HDR_GET_COMPRESS(hdr
);
5577 arc_buf_contents_t type
= arc_buf_type(hdr
);
5578 arc_buf_t
*lastbuf
= NULL
;
5579 VERIFY3U(hdr
->b_type
, ==, type
);
5581 ASSERT(hdr
->b_l1hdr
.b_buf
!= buf
|| buf
->b_next
!= NULL
);
5582 (void) remove_reference(hdr
, hash_lock
, tag
);
5584 if (arc_buf_is_shared(buf
) && !ARC_BUF_COMPRESSED(buf
)) {
5585 ASSERT3P(hdr
->b_l1hdr
.b_buf
, !=, buf
);
5586 ASSERT(ARC_BUF_LAST(buf
));
5590 * Pull the data off of this hdr and attach it to
5591 * a new anonymous hdr. Also find the last buffer
5592 * in the hdr's buffer list.
5594 lastbuf
= arc_buf_remove(hdr
, buf
);
5595 ASSERT3P(lastbuf
, !=, NULL
);
5598 * If the current arc_buf_t and the hdr are sharing their data
5599 * buffer, then we must stop sharing that block.
5601 if (arc_buf_is_shared(buf
)) {
5602 ASSERT3P(hdr
->b_l1hdr
.b_buf
, !=, buf
);
5603 VERIFY(!arc_buf_is_shared(lastbuf
));
5606 * First, sever the block sharing relationship between
5607 * buf and the arc_buf_hdr_t. Then, setup a new
5608 * block sharing relationship with the last buffer
5609 * on the arc_buf_t list.
5611 arc_unshare_buf(hdr
, buf
);
5614 * Now we need to recreate the hdr's b_pabd. Since we
5615 * have lastbuf handy, we try to share with it, but if
5616 * we can't then we allocate a new b_pabd and copy the
5617 * data from buf into it.
5619 if (arc_can_share(hdr
, lastbuf
)) {
5620 arc_share_buf(hdr
, lastbuf
);
5622 arc_hdr_alloc_pabd(hdr
);
5623 abd_copy_from_buf(hdr
->b_l1hdr
.b_pabd
,
5624 buf
->b_data
, psize
);
5626 VERIFY3P(lastbuf
->b_data
, !=, NULL
);
5627 } else if (HDR_SHARED_DATA(hdr
)) {
5629 * Uncompressed shared buffers are always at the end
5630 * of the list. Compressed buffers don't have the
5631 * same requirements. This makes it hard to
5632 * simply assert that the lastbuf is shared so
5633 * we rely on the hdr's compression flags to determine
5634 * if we have a compressed, shared buffer.
5636 ASSERT(arc_buf_is_shared(lastbuf
) ||
5637 HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
);
5638 ASSERT(!ARC_BUF_SHARED(buf
));
5640 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
5641 ASSERT3P(state
, !=, arc_l2c_only
);
5643 (void) refcount_remove_many(&state
->arcs_size
,
5644 arc_buf_size(buf
), buf
);
5646 if (refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
)) {
5647 ASSERT3P(state
, !=, arc_l2c_only
);
5648 (void) refcount_remove_many(&state
->arcs_esize
[type
],
5649 arc_buf_size(buf
), buf
);
5652 hdr
->b_l1hdr
.b_bufcnt
-= 1;
5653 arc_cksum_verify(buf
);
5654 arc_buf_unwatch(buf
);
5656 mutex_exit(hash_lock
);
5659 * Allocate a new hdr. The new hdr will contain a b_pabd
5660 * buffer which will be freed in arc_write().
5662 nhdr
= arc_hdr_alloc(spa
, psize
, lsize
, compress
, type
);
5663 ASSERT3P(nhdr
->b_l1hdr
.b_buf
, ==, NULL
);
5664 ASSERT0(nhdr
->b_l1hdr
.b_bufcnt
);
5665 ASSERT0(refcount_count(&nhdr
->b_l1hdr
.b_refcnt
));
5666 VERIFY3U(nhdr
->b_type
, ==, type
);
5667 ASSERT(!HDR_SHARED_DATA(nhdr
));
5669 nhdr
->b_l1hdr
.b_buf
= buf
;
5670 nhdr
->b_l1hdr
.b_bufcnt
= 1;
5671 nhdr
->b_l1hdr
.b_mru_hits
= 0;
5672 nhdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
5673 nhdr
->b_l1hdr
.b_mfu_hits
= 0;
5674 nhdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
5675 nhdr
->b_l1hdr
.b_l2_hits
= 0;
5676 (void) refcount_add(&nhdr
->b_l1hdr
.b_refcnt
, tag
);
5679 mutex_exit(&buf
->b_evict_lock
);
5680 (void) refcount_add_many(&arc_anon
->arcs_size
,
5681 HDR_GET_LSIZE(nhdr
), buf
);
5683 mutex_exit(&buf
->b_evict_lock
);
5684 ASSERT(refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 1);
5685 /* protected by hash lock, or hdr is on arc_anon */
5686 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
5687 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5688 hdr
->b_l1hdr
.b_mru_hits
= 0;
5689 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
5690 hdr
->b_l1hdr
.b_mfu_hits
= 0;
5691 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
5692 hdr
->b_l1hdr
.b_l2_hits
= 0;
5693 arc_change_state(arc_anon
, hdr
, hash_lock
);
5694 hdr
->b_l1hdr
.b_arc_access
= 0;
5695 mutex_exit(hash_lock
);
5697 buf_discard_identity(hdr
);
5703 arc_released(arc_buf_t
*buf
)
5707 mutex_enter(&buf
->b_evict_lock
);
5708 released
= (buf
->b_data
!= NULL
&&
5709 buf
->b_hdr
->b_l1hdr
.b_state
== arc_anon
);
5710 mutex_exit(&buf
->b_evict_lock
);
5716 arc_referenced(arc_buf_t
*buf
)
5720 mutex_enter(&buf
->b_evict_lock
);
5721 referenced
= (refcount_count(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
5722 mutex_exit(&buf
->b_evict_lock
);
5723 return (referenced
);
5728 arc_write_ready(zio_t
*zio
)
5730 arc_write_callback_t
*callback
= zio
->io_private
;
5731 arc_buf_t
*buf
= callback
->awcb_buf
;
5732 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5733 uint64_t psize
= BP_IS_HOLE(zio
->io_bp
) ? 0 : BP_GET_PSIZE(zio
->io_bp
);
5734 enum zio_compress compress
;
5735 fstrans_cookie_t cookie
= spl_fstrans_mark();
5737 ASSERT(HDR_HAS_L1HDR(hdr
));
5738 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
5739 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
5742 * If we're reexecuting this zio because the pool suspended, then
5743 * cleanup any state that was previously set the first time the
5744 * callback was invoked.
5746 if (zio
->io_flags
& ZIO_FLAG_REEXECUTED
) {
5747 arc_cksum_free(hdr
);
5748 arc_buf_unwatch(buf
);
5749 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
5750 if (arc_buf_is_shared(buf
)) {
5751 arc_unshare_buf(hdr
, buf
);
5753 arc_hdr_free_pabd(hdr
);
5757 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
5758 ASSERT(!HDR_SHARED_DATA(hdr
));
5759 ASSERT(!arc_buf_is_shared(buf
));
5761 callback
->awcb_ready(zio
, buf
, callback
->awcb_private
);
5763 if (HDR_IO_IN_PROGRESS(hdr
))
5764 ASSERT(zio
->io_flags
& ZIO_FLAG_REEXECUTED
);
5766 arc_cksum_compute(buf
);
5767 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5769 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
5770 compress
= ZIO_COMPRESS_OFF
;
5772 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, BP_GET_LSIZE(zio
->io_bp
));
5773 compress
= BP_GET_COMPRESS(zio
->io_bp
);
5775 HDR_SET_PSIZE(hdr
, psize
);
5776 arc_hdr_set_compress(hdr
, compress
);
5779 * Fill the hdr with data. If the hdr is compressed, the data we want
5780 * is available from the zio, otherwise we can take it from the buf.
5782 * We might be able to share the buf's data with the hdr here. However,
5783 * doing so would cause the ARC to be full of linear ABDs if we write a
5784 * lot of shareable data. As a compromise, we check whether scattered
5785 * ABDs are allowed, and assume that if they are then the user wants
5786 * the ARC to be primarily filled with them regardless of the data being
5787 * written. Therefore, if they're allowed then we allocate one and copy
5788 * the data into it; otherwise, we share the data directly if we can.
5790 if (zfs_abd_scatter_enabled
|| !arc_can_share(hdr
, buf
)) {
5791 arc_hdr_alloc_pabd(hdr
);
5794 * Ideally, we would always copy the io_abd into b_pabd, but the
5795 * user may have disabled compressed ARC, thus we must check the
5796 * hdr's compression setting rather than the io_bp's.
5798 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
) {
5799 ASSERT3U(BP_GET_COMPRESS(zio
->io_bp
), !=,
5801 ASSERT3U(psize
, >, 0);
5803 abd_copy(hdr
->b_l1hdr
.b_pabd
, zio
->io_abd
, psize
);
5805 ASSERT3U(zio
->io_orig_size
, ==, arc_hdr_size(hdr
));
5807 abd_copy_from_buf(hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
5811 ASSERT3P(buf
->b_data
, ==, abd_to_buf(zio
->io_orig_abd
));
5812 ASSERT3U(zio
->io_orig_size
, ==, arc_buf_size(buf
));
5813 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
5815 arc_share_buf(hdr
, buf
);
5818 arc_hdr_verify(hdr
, zio
->io_bp
);
5819 spl_fstrans_unmark(cookie
);
5823 arc_write_children_ready(zio_t
*zio
)
5825 arc_write_callback_t
*callback
= zio
->io_private
;
5826 arc_buf_t
*buf
= callback
->awcb_buf
;
5828 callback
->awcb_children_ready(zio
, buf
, callback
->awcb_private
);
5832 * The SPA calls this callback for each physical write that happens on behalf
5833 * of a logical write. See the comment in dbuf_write_physdone() for details.
5836 arc_write_physdone(zio_t
*zio
)
5838 arc_write_callback_t
*cb
= zio
->io_private
;
5839 if (cb
->awcb_physdone
!= NULL
)
5840 cb
->awcb_physdone(zio
, cb
->awcb_buf
, cb
->awcb_private
);
5844 arc_write_done(zio_t
*zio
)
5846 arc_write_callback_t
*callback
= zio
->io_private
;
5847 arc_buf_t
*buf
= callback
->awcb_buf
;
5848 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5850 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
5852 if (zio
->io_error
== 0) {
5853 arc_hdr_verify(hdr
, zio
->io_bp
);
5855 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
5856 buf_discard_identity(hdr
);
5858 hdr
->b_dva
= *BP_IDENTITY(zio
->io_bp
);
5859 hdr
->b_birth
= BP_PHYSICAL_BIRTH(zio
->io_bp
);
5862 ASSERT(HDR_EMPTY(hdr
));
5866 * If the block to be written was all-zero or compressed enough to be
5867 * embedded in the BP, no write was performed so there will be no
5868 * dva/birth/checksum. The buffer must therefore remain anonymous
5871 if (!HDR_EMPTY(hdr
)) {
5872 arc_buf_hdr_t
*exists
;
5873 kmutex_t
*hash_lock
;
5875 ASSERT3U(zio
->io_error
, ==, 0);
5877 arc_cksum_verify(buf
);
5879 exists
= buf_hash_insert(hdr
, &hash_lock
);
5880 if (exists
!= NULL
) {
5882 * This can only happen if we overwrite for
5883 * sync-to-convergence, because we remove
5884 * buffers from the hash table when we arc_free().
5886 if (zio
->io_flags
& ZIO_FLAG_IO_REWRITE
) {
5887 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
5888 panic("bad overwrite, hdr=%p exists=%p",
5889 (void *)hdr
, (void *)exists
);
5890 ASSERT(refcount_is_zero(
5891 &exists
->b_l1hdr
.b_refcnt
));
5892 arc_change_state(arc_anon
, exists
, hash_lock
);
5893 mutex_exit(hash_lock
);
5894 arc_hdr_destroy(exists
);
5895 exists
= buf_hash_insert(hdr
, &hash_lock
);
5896 ASSERT3P(exists
, ==, NULL
);
5897 } else if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
5899 ASSERT(zio
->io_prop
.zp_nopwrite
);
5900 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
5901 panic("bad nopwrite, hdr=%p exists=%p",
5902 (void *)hdr
, (void *)exists
);
5905 ASSERT(hdr
->b_l1hdr
.b_bufcnt
== 1);
5906 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
5907 ASSERT(BP_GET_DEDUP(zio
->io_bp
));
5908 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
5911 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5912 /* if it's not anon, we are doing a scrub */
5913 if (exists
== NULL
&& hdr
->b_l1hdr
.b_state
== arc_anon
)
5914 arc_access(hdr
, hash_lock
);
5915 mutex_exit(hash_lock
);
5917 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5920 ASSERT(!refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5921 callback
->awcb_done(zio
, buf
, callback
->awcb_private
);
5923 abd_put(zio
->io_abd
);
5924 kmem_free(callback
, sizeof (arc_write_callback_t
));
5928 arc_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
5929 blkptr_t
*bp
, arc_buf_t
*buf
, boolean_t l2arc
,
5930 const zio_prop_t
*zp
, arc_done_func_t
*ready
,
5931 arc_done_func_t
*children_ready
, arc_done_func_t
*physdone
,
5932 arc_done_func_t
*done
, void *private, zio_priority_t priority
,
5933 int zio_flags
, const zbookmark_phys_t
*zb
)
5935 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5936 arc_write_callback_t
*callback
;
5939 ASSERT3P(ready
, !=, NULL
);
5940 ASSERT3P(done
, !=, NULL
);
5941 ASSERT(!HDR_IO_ERROR(hdr
));
5942 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5943 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
5944 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, >, 0);
5946 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
5947 if (ARC_BUF_COMPRESSED(buf
)) {
5948 ASSERT3U(zp
->zp_compress
, !=, ZIO_COMPRESS_OFF
);
5949 zio_flags
|= ZIO_FLAG_RAW
;
5951 callback
= kmem_zalloc(sizeof (arc_write_callback_t
), KM_SLEEP
);
5952 callback
->awcb_ready
= ready
;
5953 callback
->awcb_children_ready
= children_ready
;
5954 callback
->awcb_physdone
= physdone
;
5955 callback
->awcb_done
= done
;
5956 callback
->awcb_private
= private;
5957 callback
->awcb_buf
= buf
;
5960 * The hdr's b_pabd is now stale, free it now. A new data block
5961 * will be allocated when the zio pipeline calls arc_write_ready().
5963 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
5965 * If the buf is currently sharing the data block with
5966 * the hdr then we need to break that relationship here.
5967 * The hdr will remain with a NULL data pointer and the
5968 * buf will take sole ownership of the block.
5970 if (arc_buf_is_shared(buf
)) {
5971 arc_unshare_buf(hdr
, buf
);
5973 arc_hdr_free_pabd(hdr
);
5975 VERIFY3P(buf
->b_data
, !=, NULL
);
5976 arc_hdr_set_compress(hdr
, ZIO_COMPRESS_OFF
);
5978 ASSERT(!arc_buf_is_shared(buf
));
5979 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
5981 zio
= zio_write(pio
, spa
, txg
, bp
,
5982 abd_get_from_buf(buf
->b_data
, HDR_GET_LSIZE(hdr
)),
5983 HDR_GET_LSIZE(hdr
), arc_buf_size(buf
), zp
,
5985 (children_ready
!= NULL
) ? arc_write_children_ready
: NULL
,
5986 arc_write_physdone
, arc_write_done
, callback
,
5987 priority
, zio_flags
, zb
);
5993 arc_memory_throttle(uint64_t reserve
, uint64_t txg
)
5996 uint64_t available_memory
= ptob(freemem
);
5997 static uint64_t page_load
= 0;
5998 static uint64_t last_txg
= 0;
6000 pgcnt_t minfree
= btop(arc_sys_free
/ 4);
6005 MIN(available_memory
, vmem_size(heap_arena
, VMEM_FREE
));
6008 if (available_memory
> arc_all_memory() * arc_lotsfree_percent
/ 100)
6011 if (txg
> last_txg
) {
6016 * If we are in pageout, we know that memory is already tight,
6017 * the arc is already going to be evicting, so we just want to
6018 * continue to let page writes occur as quickly as possible.
6020 if (current_is_kswapd()) {
6021 if (page_load
> MAX(ptob(minfree
), available_memory
) / 4) {
6022 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
6023 return (SET_ERROR(ERESTART
));
6025 /* Note: reserve is inflated, so we deflate */
6026 page_load
+= reserve
/ 8;
6028 } else if (page_load
> 0 && arc_reclaim_needed()) {
6029 /* memory is low, delay before restarting */
6030 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
6031 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
6032 return (SET_ERROR(EAGAIN
));
6040 arc_tempreserve_clear(uint64_t reserve
)
6042 atomic_add_64(&arc_tempreserve
, -reserve
);
6043 ASSERT((int64_t)arc_tempreserve
>= 0);
6047 arc_tempreserve_space(uint64_t reserve
, uint64_t txg
)
6053 reserve
> arc_c
/4 &&
6054 reserve
* 4 > (2ULL << SPA_MAXBLOCKSHIFT
))
6055 arc_c
= MIN(arc_c_max
, reserve
* 4);
6058 * Throttle when the calculated memory footprint for the TXG
6059 * exceeds the target ARC size.
6061 if (reserve
> arc_c
) {
6062 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve
);
6063 return (SET_ERROR(ERESTART
));
6067 * Don't count loaned bufs as in flight dirty data to prevent long
6068 * network delays from blocking transactions that are ready to be
6069 * assigned to a txg.
6071 anon_size
= MAX((int64_t)(refcount_count(&arc_anon
->arcs_size
) -
6072 arc_loaned_bytes
), 0);
6075 * Writes will, almost always, require additional memory allocations
6076 * in order to compress/encrypt/etc the data. We therefore need to
6077 * make sure that there is sufficient available memory for this.
6079 error
= arc_memory_throttle(reserve
, txg
);
6084 * Throttle writes when the amount of dirty data in the cache
6085 * gets too large. We try to keep the cache less than half full
6086 * of dirty blocks so that our sync times don't grow too large.
6087 * Note: if two requests come in concurrently, we might let them
6088 * both succeed, when one of them should fail. Not a huge deal.
6091 if (reserve
+ arc_tempreserve
+ anon_size
> arc_c
/ 2 &&
6092 anon_size
> arc_c
/ 4) {
6093 uint64_t meta_esize
=
6094 refcount_count(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
6095 uint64_t data_esize
=
6096 refcount_count(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
6097 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6098 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6099 arc_tempreserve
>> 10, meta_esize
>> 10,
6100 data_esize
>> 10, reserve
>> 10, arc_c
>> 10);
6101 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle
);
6102 return (SET_ERROR(ERESTART
));
6104 atomic_add_64(&arc_tempreserve
, reserve
);
6109 arc_kstat_update_state(arc_state_t
*state
, kstat_named_t
*size
,
6110 kstat_named_t
*evict_data
, kstat_named_t
*evict_metadata
)
6112 size
->value
.ui64
= refcount_count(&state
->arcs_size
);
6113 evict_data
->value
.ui64
=
6114 refcount_count(&state
->arcs_esize
[ARC_BUFC_DATA
]);
6115 evict_metadata
->value
.ui64
=
6116 refcount_count(&state
->arcs_esize
[ARC_BUFC_METADATA
]);
6120 arc_kstat_update(kstat_t
*ksp
, int rw
)
6122 arc_stats_t
*as
= ksp
->ks_data
;
6124 if (rw
== KSTAT_WRITE
) {
6127 arc_kstat_update_state(arc_anon
,
6128 &as
->arcstat_anon_size
,
6129 &as
->arcstat_anon_evictable_data
,
6130 &as
->arcstat_anon_evictable_metadata
);
6131 arc_kstat_update_state(arc_mru
,
6132 &as
->arcstat_mru_size
,
6133 &as
->arcstat_mru_evictable_data
,
6134 &as
->arcstat_mru_evictable_metadata
);
6135 arc_kstat_update_state(arc_mru_ghost
,
6136 &as
->arcstat_mru_ghost_size
,
6137 &as
->arcstat_mru_ghost_evictable_data
,
6138 &as
->arcstat_mru_ghost_evictable_metadata
);
6139 arc_kstat_update_state(arc_mfu
,
6140 &as
->arcstat_mfu_size
,
6141 &as
->arcstat_mfu_evictable_data
,
6142 &as
->arcstat_mfu_evictable_metadata
);
6143 arc_kstat_update_state(arc_mfu_ghost
,
6144 &as
->arcstat_mfu_ghost_size
,
6145 &as
->arcstat_mfu_ghost_evictable_data
,
6146 &as
->arcstat_mfu_ghost_evictable_metadata
);
6153 * This function *must* return indices evenly distributed between all
6154 * sublists of the multilist. This is needed due to how the ARC eviction
6155 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6156 * distributed between all sublists and uses this assumption when
6157 * deciding which sublist to evict from and how much to evict from it.
6160 arc_state_multilist_index_func(multilist_t
*ml
, void *obj
)
6162 arc_buf_hdr_t
*hdr
= obj
;
6165 * We rely on b_dva to generate evenly distributed index
6166 * numbers using buf_hash below. So, as an added precaution,
6167 * let's make sure we never add empty buffers to the arc lists.
6169 ASSERT(!HDR_EMPTY(hdr
));
6172 * The assumption here, is the hash value for a given
6173 * arc_buf_hdr_t will remain constant throughout its lifetime
6174 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
6175 * Thus, we don't need to store the header's sublist index
6176 * on insertion, as this index can be recalculated on removal.
6178 * Also, the low order bits of the hash value are thought to be
6179 * distributed evenly. Otherwise, in the case that the multilist
6180 * has a power of two number of sublists, each sublists' usage
6181 * would not be evenly distributed.
6183 return (buf_hash(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
) %
6184 multilist_get_num_sublists(ml
));
6188 * Called during module initialization and periodically thereafter to
6189 * apply reasonable changes to the exposed performance tunings. Non-zero
6190 * zfs_* values which differ from the currently set values will be applied.
6193 arc_tuning_update(void)
6195 uint64_t percent
, allmem
= arc_all_memory();
6197 /* Valid range: 64M - <all physical memory> */
6198 if ((zfs_arc_max
) && (zfs_arc_max
!= arc_c_max
) &&
6199 (zfs_arc_max
> 64 << 20) && (zfs_arc_max
< allmem
) &&
6200 (zfs_arc_max
> arc_c_min
)) {
6201 arc_c_max
= zfs_arc_max
;
6203 arc_p
= (arc_c
>> 1);
6204 /* Valid range of arc_meta_limit: arc_meta_min - arc_c_max */
6205 percent
= MIN(zfs_arc_meta_limit_percent
, 100);
6206 arc_meta_limit
= MAX(arc_meta_min
, (percent
* arc_c_max
) / 100);
6207 percent
= MIN(zfs_arc_dnode_limit_percent
, 100);
6208 arc_dnode_limit
= (percent
* arc_meta_limit
) / 100;
6211 /* Valid range: 32M - <arc_c_max> */
6212 if ((zfs_arc_min
) && (zfs_arc_min
!= arc_c_min
) &&
6213 (zfs_arc_min
>= 2ULL << SPA_MAXBLOCKSHIFT
) &&
6214 (zfs_arc_min
<= arc_c_max
)) {
6215 arc_c_min
= zfs_arc_min
;
6216 arc_c
= MAX(arc_c
, arc_c_min
);
6219 /* Valid range: 16M - <arc_c_max> */
6220 if ((zfs_arc_meta_min
) && (zfs_arc_meta_min
!= arc_meta_min
) &&
6221 (zfs_arc_meta_min
>= 1ULL << SPA_MAXBLOCKSHIFT
) &&
6222 (zfs_arc_meta_min
<= arc_c_max
)) {
6223 arc_meta_min
= zfs_arc_meta_min
;
6224 arc_meta_limit
= MAX(arc_meta_limit
, arc_meta_min
);
6225 arc_dnode_limit
= arc_meta_limit
/ 10;
6228 /* Valid range: <arc_meta_min> - <arc_c_max> */
6229 if ((zfs_arc_meta_limit
) && (zfs_arc_meta_limit
!= arc_meta_limit
) &&
6230 (zfs_arc_meta_limit
>= zfs_arc_meta_min
) &&
6231 (zfs_arc_meta_limit
<= arc_c_max
))
6232 arc_meta_limit
= zfs_arc_meta_limit
;
6234 /* Valid range: <arc_meta_min> - <arc_c_max> */
6235 if ((zfs_arc_dnode_limit
) && (zfs_arc_dnode_limit
!= arc_dnode_limit
) &&
6236 (zfs_arc_dnode_limit
>= zfs_arc_meta_min
) &&
6237 (zfs_arc_dnode_limit
<= arc_c_max
))
6238 arc_dnode_limit
= zfs_arc_dnode_limit
;
6240 /* Valid range: 1 - N */
6241 if (zfs_arc_grow_retry
)
6242 arc_grow_retry
= zfs_arc_grow_retry
;
6244 /* Valid range: 1 - N */
6245 if (zfs_arc_shrink_shift
) {
6246 arc_shrink_shift
= zfs_arc_shrink_shift
;
6247 arc_no_grow_shift
= MIN(arc_no_grow_shift
, arc_shrink_shift
-1);
6250 /* Valid range: 1 - N */
6251 if (zfs_arc_p_min_shift
)
6252 arc_p_min_shift
= zfs_arc_p_min_shift
;
6254 /* Valid range: 1 - N ticks */
6255 if (zfs_arc_min_prefetch_lifespan
)
6256 arc_min_prefetch_lifespan
= zfs_arc_min_prefetch_lifespan
;
6258 /* Valid range: 0 - 100 */
6259 if ((zfs_arc_lotsfree_percent
>= 0) &&
6260 (zfs_arc_lotsfree_percent
<= 100))
6261 arc_lotsfree_percent
= zfs_arc_lotsfree_percent
;
6263 /* Valid range: 0 - <all physical memory> */
6264 if ((zfs_arc_sys_free
) && (zfs_arc_sys_free
!= arc_sys_free
))
6265 arc_sys_free
= MIN(MAX(zfs_arc_sys_free
, 0), allmem
);
6270 arc_state_init(void)
6272 arc_anon
= &ARC_anon
;
6274 arc_mru_ghost
= &ARC_mru_ghost
;
6276 arc_mfu_ghost
= &ARC_mfu_ghost
;
6277 arc_l2c_only
= &ARC_l2c_only
;
6279 multilist_create(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
],
6280 sizeof (arc_buf_hdr_t
),
6281 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6282 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
6283 multilist_create(&arc_mru
->arcs_list
[ARC_BUFC_DATA
],
6284 sizeof (arc_buf_hdr_t
),
6285 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6286 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
6287 multilist_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
],
6288 sizeof (arc_buf_hdr_t
),
6289 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6290 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
6291 multilist_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
],
6292 sizeof (arc_buf_hdr_t
),
6293 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6294 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
6295 multilist_create(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
],
6296 sizeof (arc_buf_hdr_t
),
6297 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6298 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
6299 multilist_create(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
],
6300 sizeof (arc_buf_hdr_t
),
6301 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6302 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
6303 multilist_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
],
6304 sizeof (arc_buf_hdr_t
),
6305 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6306 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
6307 multilist_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
],
6308 sizeof (arc_buf_hdr_t
),
6309 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6310 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
6311 multilist_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
],
6312 sizeof (arc_buf_hdr_t
),
6313 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6314 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
6315 multilist_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
],
6316 sizeof (arc_buf_hdr_t
),
6317 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6318 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
6320 refcount_create(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
6321 refcount_create(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
6322 refcount_create(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]);
6323 refcount_create(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]);
6324 refcount_create(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6325 refcount_create(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6326 refcount_create(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
6327 refcount_create(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]);
6328 refcount_create(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6329 refcount_create(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6330 refcount_create(&arc_l2c_only
->arcs_esize
[ARC_BUFC_METADATA
]);
6331 refcount_create(&arc_l2c_only
->arcs_esize
[ARC_BUFC_DATA
]);
6333 refcount_create(&arc_anon
->arcs_size
);
6334 refcount_create(&arc_mru
->arcs_size
);
6335 refcount_create(&arc_mru_ghost
->arcs_size
);
6336 refcount_create(&arc_mfu
->arcs_size
);
6337 refcount_create(&arc_mfu_ghost
->arcs_size
);
6338 refcount_create(&arc_l2c_only
->arcs_size
);
6340 arc_anon
->arcs_state
= ARC_STATE_ANON
;
6341 arc_mru
->arcs_state
= ARC_STATE_MRU
;
6342 arc_mru_ghost
->arcs_state
= ARC_STATE_MRU_GHOST
;
6343 arc_mfu
->arcs_state
= ARC_STATE_MFU
;
6344 arc_mfu_ghost
->arcs_state
= ARC_STATE_MFU_GHOST
;
6345 arc_l2c_only
->arcs_state
= ARC_STATE_L2C_ONLY
;
6349 arc_state_fini(void)
6351 refcount_destroy(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
6352 refcount_destroy(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
6353 refcount_destroy(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]);
6354 refcount_destroy(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]);
6355 refcount_destroy(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6356 refcount_destroy(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6357 refcount_destroy(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
6358 refcount_destroy(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]);
6359 refcount_destroy(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6360 refcount_destroy(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6361 refcount_destroy(&arc_l2c_only
->arcs_esize
[ARC_BUFC_METADATA
]);
6362 refcount_destroy(&arc_l2c_only
->arcs_esize
[ARC_BUFC_DATA
]);
6364 refcount_destroy(&arc_anon
->arcs_size
);
6365 refcount_destroy(&arc_mru
->arcs_size
);
6366 refcount_destroy(&arc_mru_ghost
->arcs_size
);
6367 refcount_destroy(&arc_mfu
->arcs_size
);
6368 refcount_destroy(&arc_mfu_ghost
->arcs_size
);
6369 refcount_destroy(&arc_l2c_only
->arcs_size
);
6371 multilist_destroy(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
]);
6372 multilist_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
6373 multilist_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
]);
6374 multilist_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
6375 multilist_destroy(&arc_mru
->arcs_list
[ARC_BUFC_DATA
]);
6376 multilist_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
]);
6377 multilist_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
]);
6378 multilist_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
]);
6379 multilist_destroy(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]);
6380 multilist_destroy(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]);
6392 uint64_t percent
, allmem
= arc_all_memory();
6394 mutex_init(&arc_reclaim_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
6395 cv_init(&arc_reclaim_thread_cv
, NULL
, CV_DEFAULT
, NULL
);
6396 cv_init(&arc_reclaim_waiters_cv
, NULL
, CV_DEFAULT
, NULL
);
6398 /* Convert seconds to clock ticks */
6399 arc_min_prefetch_lifespan
= 1 * hz
;
6403 * Register a shrinker to support synchronous (direct) memory
6404 * reclaim from the arc. This is done to prevent kswapd from
6405 * swapping out pages when it is preferable to shrink the arc.
6407 spl_register_shrinker(&arc_shrinker
);
6409 /* Set to 1/64 of all memory or a minimum of 512K */
6410 arc_sys_free
= MAX(allmem
/ 64, (512 * 1024));
6414 /* Set max to 1/2 of all memory */
6415 arc_c_max
= allmem
/ 2;
6418 * In userland, there's only the memory pressure that we artificially
6419 * create (see arc_available_memory()). Don't let arc_c get too
6420 * small, because it can cause transactions to be larger than
6421 * arc_c, causing arc_tempreserve_space() to fail.
6424 arc_c_min
= MAX(arc_c_max
/ 2, 2ULL << SPA_MAXBLOCKSHIFT
);
6426 arc_c_min
= 2ULL << SPA_MAXBLOCKSHIFT
;
6430 arc_p
= (arc_c
>> 1);
6433 /* Set min to 1/2 of arc_c_min */
6434 arc_meta_min
= 1ULL << SPA_MAXBLOCKSHIFT
;
6435 /* Initialize maximum observed usage to zero */
6438 * Set arc_meta_limit to a percent of arc_c_max with a floor of
6439 * arc_meta_min, and a ceiling of arc_c_max.
6441 percent
= MIN(zfs_arc_meta_limit_percent
, 100);
6442 arc_meta_limit
= MAX(arc_meta_min
, (percent
* arc_c_max
) / 100);
6443 percent
= MIN(zfs_arc_dnode_limit_percent
, 100);
6444 arc_dnode_limit
= (percent
* arc_meta_limit
) / 100;
6446 /* Apply user specified tunings */
6447 arc_tuning_update();
6449 if (zfs_arc_num_sublists_per_state
< 1)
6450 zfs_arc_num_sublists_per_state
= MAX(boot_ncpus
, 1);
6452 /* if kmem_flags are set, lets try to use less memory */
6453 if (kmem_debugging())
6455 if (arc_c
< arc_c_min
)
6461 list_create(&arc_prune_list
, sizeof (arc_prune_t
),
6462 offsetof(arc_prune_t
, p_node
));
6463 mutex_init(&arc_prune_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
6465 arc_prune_taskq
= taskq_create("arc_prune", max_ncpus
, defclsyspri
,
6466 max_ncpus
, INT_MAX
, TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
);
6468 arc_reclaim_thread_exit
= B_FALSE
;
6470 arc_ksp
= kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED
,
6471 sizeof (arc_stats
) / sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
);
6473 if (arc_ksp
!= NULL
) {
6474 arc_ksp
->ks_data
= &arc_stats
;
6475 arc_ksp
->ks_update
= arc_kstat_update
;
6476 kstat_install(arc_ksp
);
6479 (void) thread_create(NULL
, 0, arc_reclaim_thread
, NULL
, 0, &p0
,
6480 TS_RUN
, defclsyspri
);
6486 * Calculate maximum amount of dirty data per pool.
6488 * If it has been set by a module parameter, take that.
6489 * Otherwise, use a percentage of physical memory defined by
6490 * zfs_dirty_data_max_percent (default 10%) with a cap at
6491 * zfs_dirty_data_max_max (default 25% of physical memory).
6493 if (zfs_dirty_data_max_max
== 0)
6494 zfs_dirty_data_max_max
= allmem
*
6495 zfs_dirty_data_max_max_percent
/ 100;
6497 if (zfs_dirty_data_max
== 0) {
6498 zfs_dirty_data_max
= allmem
*
6499 zfs_dirty_data_max_percent
/ 100;
6500 zfs_dirty_data_max
= MIN(zfs_dirty_data_max
,
6501 zfs_dirty_data_max_max
);
6511 spl_unregister_shrinker(&arc_shrinker
);
6512 #endif /* _KERNEL */
6514 mutex_enter(&arc_reclaim_lock
);
6515 arc_reclaim_thread_exit
= B_TRUE
;
6517 * The reclaim thread will set arc_reclaim_thread_exit back to
6518 * B_FALSE when it is finished exiting; we're waiting for that.
6520 while (arc_reclaim_thread_exit
) {
6521 cv_signal(&arc_reclaim_thread_cv
);
6522 cv_wait(&arc_reclaim_thread_cv
, &arc_reclaim_lock
);
6524 mutex_exit(&arc_reclaim_lock
);
6526 /* Use B_TRUE to ensure *all* buffers are evicted */
6527 arc_flush(NULL
, B_TRUE
);
6531 if (arc_ksp
!= NULL
) {
6532 kstat_delete(arc_ksp
);
6536 taskq_wait(arc_prune_taskq
);
6537 taskq_destroy(arc_prune_taskq
);
6539 mutex_enter(&arc_prune_mtx
);
6540 while ((p
= list_head(&arc_prune_list
)) != NULL
) {
6541 list_remove(&arc_prune_list
, p
);
6542 refcount_remove(&p
->p_refcnt
, &arc_prune_list
);
6543 refcount_destroy(&p
->p_refcnt
);
6544 kmem_free(p
, sizeof (*p
));
6546 mutex_exit(&arc_prune_mtx
);
6548 list_destroy(&arc_prune_list
);
6549 mutex_destroy(&arc_prune_mtx
);
6550 mutex_destroy(&arc_reclaim_lock
);
6551 cv_destroy(&arc_reclaim_thread_cv
);
6552 cv_destroy(&arc_reclaim_waiters_cv
);
6557 ASSERT0(arc_loaned_bytes
);
6563 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6564 * It uses dedicated storage devices to hold cached data, which are populated
6565 * using large infrequent writes. The main role of this cache is to boost
6566 * the performance of random read workloads. The intended L2ARC devices
6567 * include short-stroked disks, solid state disks, and other media with
6568 * substantially faster read latency than disk.
6570 * +-----------------------+
6572 * +-----------------------+
6575 * l2arc_feed_thread() arc_read()
6579 * +---------------+ |
6581 * +---------------+ |
6586 * +-------+ +-------+
6588 * | cache | | cache |
6589 * +-------+ +-------+
6590 * +=========+ .-----.
6591 * : L2ARC : |-_____-|
6592 * : devices : | Disks |
6593 * +=========+ `-_____-'
6595 * Read requests are satisfied from the following sources, in order:
6598 * 2) vdev cache of L2ARC devices
6600 * 4) vdev cache of disks
6603 * Some L2ARC device types exhibit extremely slow write performance.
6604 * To accommodate for this there are some significant differences between
6605 * the L2ARC and traditional cache design:
6607 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6608 * the ARC behave as usual, freeing buffers and placing headers on ghost
6609 * lists. The ARC does not send buffers to the L2ARC during eviction as
6610 * this would add inflated write latencies for all ARC memory pressure.
6612 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6613 * It does this by periodically scanning buffers from the eviction-end of
6614 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6615 * not already there. It scans until a headroom of buffers is satisfied,
6616 * which itself is a buffer for ARC eviction. If a compressible buffer is
6617 * found during scanning and selected for writing to an L2ARC device, we
6618 * temporarily boost scanning headroom during the next scan cycle to make
6619 * sure we adapt to compression effects (which might significantly reduce
6620 * the data volume we write to L2ARC). The thread that does this is
6621 * l2arc_feed_thread(), illustrated below; example sizes are included to
6622 * provide a better sense of ratio than this diagram:
6625 * +---------------------+----------+
6626 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6627 * +---------------------+----------+ | o L2ARC eligible
6628 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6629 * +---------------------+----------+ |
6630 * 15.9 Gbytes ^ 32 Mbytes |
6632 * l2arc_feed_thread()
6634 * l2arc write hand <--[oooo]--'
6638 * +==============================+
6639 * L2ARC dev |####|#|###|###| |####| ... |
6640 * +==============================+
6643 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6644 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6645 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6646 * safe to say that this is an uncommon case, since buffers at the end of
6647 * the ARC lists have moved there due to inactivity.
6649 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6650 * then the L2ARC simply misses copying some buffers. This serves as a
6651 * pressure valve to prevent heavy read workloads from both stalling the ARC
6652 * with waits and clogging the L2ARC with writes. This also helps prevent
6653 * the potential for the L2ARC to churn if it attempts to cache content too
6654 * quickly, such as during backups of the entire pool.
6656 * 5. After system boot and before the ARC has filled main memory, there are
6657 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6658 * lists can remain mostly static. Instead of searching from tail of these
6659 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6660 * for eligible buffers, greatly increasing its chance of finding them.
6662 * The L2ARC device write speed is also boosted during this time so that
6663 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6664 * there are no L2ARC reads, and no fear of degrading read performance
6665 * through increased writes.
6667 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6668 * the vdev queue can aggregate them into larger and fewer writes. Each
6669 * device is written to in a rotor fashion, sweeping writes through
6670 * available space then repeating.
6672 * 7. The L2ARC does not store dirty content. It never needs to flush
6673 * write buffers back to disk based storage.
6675 * 8. If an ARC buffer is written (and dirtied) which also exists in the
6676 * L2ARC, the now stale L2ARC buffer is immediately dropped.
6678 * The performance of the L2ARC can be tweaked by a number of tunables, which
6679 * may be necessary for different workloads:
6681 * l2arc_write_max max write bytes per interval
6682 * l2arc_write_boost extra write bytes during device warmup
6683 * l2arc_noprefetch skip caching prefetched buffers
6684 * l2arc_headroom number of max device writes to precache
6685 * l2arc_headroom_boost when we find compressed buffers during ARC
6686 * scanning, we multiply headroom by this
6687 * percentage factor for the next scan cycle,
6688 * since more compressed buffers are likely to
6690 * l2arc_feed_secs seconds between L2ARC writing
6692 * Tunables may be removed or added as future performance improvements are
6693 * integrated, and also may become zpool properties.
6695 * There are three key functions that control how the L2ARC warms up:
6697 * l2arc_write_eligible() check if a buffer is eligible to cache
6698 * l2arc_write_size() calculate how much to write
6699 * l2arc_write_interval() calculate sleep delay between writes
6701 * These three functions determine what to write, how much, and how quickly
6706 l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*hdr
)
6709 * A buffer is *not* eligible for the L2ARC if it:
6710 * 1. belongs to a different spa.
6711 * 2. is already cached on the L2ARC.
6712 * 3. has an I/O in progress (it may be an incomplete read).
6713 * 4. is flagged not eligible (zfs property).
6715 if (hdr
->b_spa
!= spa_guid
|| HDR_HAS_L2HDR(hdr
) ||
6716 HDR_IO_IN_PROGRESS(hdr
) || !HDR_L2CACHE(hdr
))
6723 l2arc_write_size(void)
6728 * Make sure our globals have meaningful values in case the user
6731 size
= l2arc_write_max
;
6733 cmn_err(CE_NOTE
, "Bad value for l2arc_write_max, value must "
6734 "be greater than zero, resetting it to the default (%d)",
6736 size
= l2arc_write_max
= L2ARC_WRITE_SIZE
;
6739 if (arc_warm
== B_FALSE
)
6740 size
+= l2arc_write_boost
;
6747 l2arc_write_interval(clock_t began
, uint64_t wanted
, uint64_t wrote
)
6749 clock_t interval
, next
, now
;
6752 * If the ARC lists are busy, increase our write rate; if the
6753 * lists are stale, idle back. This is achieved by checking
6754 * how much we previously wrote - if it was more than half of
6755 * what we wanted, schedule the next write much sooner.
6757 if (l2arc_feed_again
&& wrote
> (wanted
/ 2))
6758 interval
= (hz
* l2arc_feed_min_ms
) / 1000;
6760 interval
= hz
* l2arc_feed_secs
;
6762 now
= ddi_get_lbolt();
6763 next
= MAX(now
, MIN(now
+ interval
, began
+ interval
));
6769 * Cycle through L2ARC devices. This is how L2ARC load balances.
6770 * If a device is returned, this also returns holding the spa config lock.
6772 static l2arc_dev_t
*
6773 l2arc_dev_get_next(void)
6775 l2arc_dev_t
*first
, *next
= NULL
;
6778 * Lock out the removal of spas (spa_namespace_lock), then removal
6779 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
6780 * both locks will be dropped and a spa config lock held instead.
6782 mutex_enter(&spa_namespace_lock
);
6783 mutex_enter(&l2arc_dev_mtx
);
6785 /* if there are no vdevs, there is nothing to do */
6786 if (l2arc_ndev
== 0)
6790 next
= l2arc_dev_last
;
6792 /* loop around the list looking for a non-faulted vdev */
6794 next
= list_head(l2arc_dev_list
);
6796 next
= list_next(l2arc_dev_list
, next
);
6798 next
= list_head(l2arc_dev_list
);
6801 /* if we have come back to the start, bail out */
6804 else if (next
== first
)
6807 } while (vdev_is_dead(next
->l2ad_vdev
));
6809 /* if we were unable to find any usable vdevs, return NULL */
6810 if (vdev_is_dead(next
->l2ad_vdev
))
6813 l2arc_dev_last
= next
;
6816 mutex_exit(&l2arc_dev_mtx
);
6819 * Grab the config lock to prevent the 'next' device from being
6820 * removed while we are writing to it.
6823 spa_config_enter(next
->l2ad_spa
, SCL_L2ARC
, next
, RW_READER
);
6824 mutex_exit(&spa_namespace_lock
);
6830 * Free buffers that were tagged for destruction.
6833 l2arc_do_free_on_write(void)
6836 l2arc_data_free_t
*df
, *df_prev
;
6838 mutex_enter(&l2arc_free_on_write_mtx
);
6839 buflist
= l2arc_free_on_write
;
6841 for (df
= list_tail(buflist
); df
; df
= df_prev
) {
6842 df_prev
= list_prev(buflist
, df
);
6843 ASSERT3P(df
->l2df_abd
, !=, NULL
);
6844 abd_free(df
->l2df_abd
);
6845 list_remove(buflist
, df
);
6846 kmem_free(df
, sizeof (l2arc_data_free_t
));
6849 mutex_exit(&l2arc_free_on_write_mtx
);
6853 * A write to a cache device has completed. Update all headers to allow
6854 * reads from these buffers to begin.
6857 l2arc_write_done(zio_t
*zio
)
6859 l2arc_write_callback_t
*cb
;
6862 arc_buf_hdr_t
*head
, *hdr
, *hdr_prev
;
6863 kmutex_t
*hash_lock
;
6864 int64_t bytes_dropped
= 0;
6866 cb
= zio
->io_private
;
6867 ASSERT3P(cb
, !=, NULL
);
6868 dev
= cb
->l2wcb_dev
;
6869 ASSERT3P(dev
, !=, NULL
);
6870 head
= cb
->l2wcb_head
;
6871 ASSERT3P(head
, !=, NULL
);
6872 buflist
= &dev
->l2ad_buflist
;
6873 ASSERT3P(buflist
, !=, NULL
);
6874 DTRACE_PROBE2(l2arc__iodone
, zio_t
*, zio
,
6875 l2arc_write_callback_t
*, cb
);
6877 if (zio
->io_error
!= 0)
6878 ARCSTAT_BUMP(arcstat_l2_writes_error
);
6881 * All writes completed, or an error was hit.
6884 mutex_enter(&dev
->l2ad_mtx
);
6885 for (hdr
= list_prev(buflist
, head
); hdr
; hdr
= hdr_prev
) {
6886 hdr_prev
= list_prev(buflist
, hdr
);
6888 hash_lock
= HDR_LOCK(hdr
);
6891 * We cannot use mutex_enter or else we can deadlock
6892 * with l2arc_write_buffers (due to swapping the order
6893 * the hash lock and l2ad_mtx are taken).
6895 if (!mutex_tryenter(hash_lock
)) {
6897 * Missed the hash lock. We must retry so we
6898 * don't leave the ARC_FLAG_L2_WRITING bit set.
6900 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry
);
6903 * We don't want to rescan the headers we've
6904 * already marked as having been written out, so
6905 * we reinsert the head node so we can pick up
6906 * where we left off.
6908 list_remove(buflist
, head
);
6909 list_insert_after(buflist
, hdr
, head
);
6911 mutex_exit(&dev
->l2ad_mtx
);
6914 * We wait for the hash lock to become available
6915 * to try and prevent busy waiting, and increase
6916 * the chance we'll be able to acquire the lock
6917 * the next time around.
6919 mutex_enter(hash_lock
);
6920 mutex_exit(hash_lock
);
6925 * We could not have been moved into the arc_l2c_only
6926 * state while in-flight due to our ARC_FLAG_L2_WRITING
6927 * bit being set. Let's just ensure that's being enforced.
6929 ASSERT(HDR_HAS_L1HDR(hdr
));
6932 * Skipped - drop L2ARC entry and mark the header as no
6933 * longer L2 eligibile.
6935 if (zio
->io_error
!= 0) {
6937 * Error - drop L2ARC entry.
6939 list_remove(buflist
, hdr
);
6940 arc_hdr_clear_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
6942 ARCSTAT_INCR(arcstat_l2_asize
, -arc_hdr_size(hdr
));
6943 ARCSTAT_INCR(arcstat_l2_size
, -HDR_GET_LSIZE(hdr
));
6945 bytes_dropped
+= arc_hdr_size(hdr
);
6946 (void) refcount_remove_many(&dev
->l2ad_alloc
,
6947 arc_hdr_size(hdr
), hdr
);
6951 * Allow ARC to begin reads and ghost list evictions to
6954 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2_WRITING
);
6956 mutex_exit(hash_lock
);
6959 atomic_inc_64(&l2arc_writes_done
);
6960 list_remove(buflist
, head
);
6961 ASSERT(!HDR_HAS_L1HDR(head
));
6962 kmem_cache_free(hdr_l2only_cache
, head
);
6963 mutex_exit(&dev
->l2ad_mtx
);
6965 vdev_space_update(dev
->l2ad_vdev
, -bytes_dropped
, 0, 0);
6967 l2arc_do_free_on_write();
6969 kmem_free(cb
, sizeof (l2arc_write_callback_t
));
6973 * A read to a cache device completed. Validate buffer contents before
6974 * handing over to the regular ARC routines.
6977 l2arc_read_done(zio_t
*zio
)
6979 l2arc_read_callback_t
*cb
;
6981 kmutex_t
*hash_lock
;
6982 boolean_t valid_cksum
;
6984 ASSERT3P(zio
->io_vd
, !=, NULL
);
6985 ASSERT(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
);
6987 spa_config_exit(zio
->io_spa
, SCL_L2ARC
, zio
->io_vd
);
6989 cb
= zio
->io_private
;
6990 ASSERT3P(cb
, !=, NULL
);
6991 hdr
= cb
->l2rcb_hdr
;
6992 ASSERT3P(hdr
, !=, NULL
);
6994 hash_lock
= HDR_LOCK(hdr
);
6995 mutex_enter(hash_lock
);
6996 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
6998 ASSERT3P(zio
->io_abd
, !=, NULL
);
7001 * Check this survived the L2ARC journey.
7003 ASSERT3P(zio
->io_abd
, ==, hdr
->b_l1hdr
.b_pabd
);
7004 zio
->io_bp_copy
= cb
->l2rcb_bp
; /* XXX fix in L2ARC 2.0 */
7005 zio
->io_bp
= &zio
->io_bp_copy
; /* XXX fix in L2ARC 2.0 */
7007 valid_cksum
= arc_cksum_is_equal(hdr
, zio
);
7008 if (valid_cksum
&& zio
->io_error
== 0 && !HDR_L2_EVICTED(hdr
)) {
7009 mutex_exit(hash_lock
);
7010 zio
->io_private
= hdr
;
7013 mutex_exit(hash_lock
);
7015 * Buffer didn't survive caching. Increment stats and
7016 * reissue to the original storage device.
7018 if (zio
->io_error
!= 0) {
7019 ARCSTAT_BUMP(arcstat_l2_io_error
);
7021 zio
->io_error
= SET_ERROR(EIO
);
7024 ARCSTAT_BUMP(arcstat_l2_cksum_bad
);
7027 * If there's no waiter, issue an async i/o to the primary
7028 * storage now. If there *is* a waiter, the caller must
7029 * issue the i/o in a context where it's OK to block.
7031 if (zio
->io_waiter
== NULL
) {
7032 zio_t
*pio
= zio_unique_parent(zio
);
7034 ASSERT(!pio
|| pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
7036 zio_nowait(zio_read(pio
, zio
->io_spa
, zio
->io_bp
,
7037 hdr
->b_l1hdr
.b_pabd
, zio
->io_size
, arc_read_done
,
7038 hdr
, zio
->io_priority
, cb
->l2rcb_flags
,
7043 kmem_free(cb
, sizeof (l2arc_read_callback_t
));
7047 * This is the list priority from which the L2ARC will search for pages to
7048 * cache. This is used within loops (0..3) to cycle through lists in the
7049 * desired order. This order can have a significant effect on cache
7052 * Currently the metadata lists are hit first, MFU then MRU, followed by
7053 * the data lists. This function returns a locked list, and also returns
7056 static multilist_sublist_t
*
7057 l2arc_sublist_lock(int list_num
)
7059 multilist_t
*ml
= NULL
;
7062 ASSERT(list_num
>= 0 && list_num
< L2ARC_FEED_TYPES
);
7066 ml
= &arc_mfu
->arcs_list
[ARC_BUFC_METADATA
];
7069 ml
= &arc_mru
->arcs_list
[ARC_BUFC_METADATA
];
7072 ml
= &arc_mfu
->arcs_list
[ARC_BUFC_DATA
];
7075 ml
= &arc_mru
->arcs_list
[ARC_BUFC_DATA
];
7082 * Return a randomly-selected sublist. This is acceptable
7083 * because the caller feeds only a little bit of data for each
7084 * call (8MB). Subsequent calls will result in different
7085 * sublists being selected.
7087 idx
= multilist_get_random_index(ml
);
7088 return (multilist_sublist_lock(ml
, idx
));
7092 * Evict buffers from the device write hand to the distance specified in
7093 * bytes. This distance may span populated buffers, it may span nothing.
7094 * This is clearing a region on the L2ARC device ready for writing.
7095 * If the 'all' boolean is set, every buffer is evicted.
7098 l2arc_evict(l2arc_dev_t
*dev
, uint64_t distance
, boolean_t all
)
7101 arc_buf_hdr_t
*hdr
, *hdr_prev
;
7102 kmutex_t
*hash_lock
;
7105 buflist
= &dev
->l2ad_buflist
;
7107 if (!all
&& dev
->l2ad_first
) {
7109 * This is the first sweep through the device. There is
7115 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- (2 * distance
))) {
7117 * When nearing the end of the device, evict to the end
7118 * before the device write hand jumps to the start.
7120 taddr
= dev
->l2ad_end
;
7122 taddr
= dev
->l2ad_hand
+ distance
;
7124 DTRACE_PROBE4(l2arc__evict
, l2arc_dev_t
*, dev
, list_t
*, buflist
,
7125 uint64_t, taddr
, boolean_t
, all
);
7128 mutex_enter(&dev
->l2ad_mtx
);
7129 for (hdr
= list_tail(buflist
); hdr
; hdr
= hdr_prev
) {
7130 hdr_prev
= list_prev(buflist
, hdr
);
7132 hash_lock
= HDR_LOCK(hdr
);
7135 * We cannot use mutex_enter or else we can deadlock
7136 * with l2arc_write_buffers (due to swapping the order
7137 * the hash lock and l2ad_mtx are taken).
7139 if (!mutex_tryenter(hash_lock
)) {
7141 * Missed the hash lock. Retry.
7143 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry
);
7144 mutex_exit(&dev
->l2ad_mtx
);
7145 mutex_enter(hash_lock
);
7146 mutex_exit(hash_lock
);
7150 if (HDR_L2_WRITE_HEAD(hdr
)) {
7152 * We hit a write head node. Leave it for
7153 * l2arc_write_done().
7155 list_remove(buflist
, hdr
);
7156 mutex_exit(hash_lock
);
7160 if (!all
&& HDR_HAS_L2HDR(hdr
) &&
7161 (hdr
->b_l2hdr
.b_daddr
> taddr
||
7162 hdr
->b_l2hdr
.b_daddr
< dev
->l2ad_hand
)) {
7164 * We've evicted to the target address,
7165 * or the end of the device.
7167 mutex_exit(hash_lock
);
7171 ASSERT(HDR_HAS_L2HDR(hdr
));
7172 if (!HDR_HAS_L1HDR(hdr
)) {
7173 ASSERT(!HDR_L2_READING(hdr
));
7175 * This doesn't exist in the ARC. Destroy.
7176 * arc_hdr_destroy() will call list_remove()
7177 * and decrement arcstat_l2_size.
7179 arc_change_state(arc_anon
, hdr
, hash_lock
);
7180 arc_hdr_destroy(hdr
);
7182 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_l2c_only
);
7183 ARCSTAT_BUMP(arcstat_l2_evict_l1cached
);
7185 * Invalidate issued or about to be issued
7186 * reads, since we may be about to write
7187 * over this location.
7189 if (HDR_L2_READING(hdr
)) {
7190 ARCSTAT_BUMP(arcstat_l2_evict_reading
);
7191 arc_hdr_set_flags(hdr
, ARC_FLAG_L2_EVICTED
);
7194 /* Ensure this header has finished being written */
7195 ASSERT(!HDR_L2_WRITING(hdr
));
7197 arc_hdr_l2hdr_destroy(hdr
);
7199 mutex_exit(hash_lock
);
7201 mutex_exit(&dev
->l2ad_mtx
);
7205 * Find and write ARC buffers to the L2ARC device.
7207 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7208 * for reading until they have completed writing.
7209 * The headroom_boost is an in-out parameter used to maintain headroom boost
7210 * state between calls to this function.
7212 * Returns the number of bytes actually written (which may be smaller than
7213 * the delta by which the device hand has changed due to alignment).
7216 l2arc_write_buffers(spa_t
*spa
, l2arc_dev_t
*dev
, uint64_t target_sz
)
7218 arc_buf_hdr_t
*hdr
, *hdr_prev
, *head
;
7219 uint64_t write_asize
, write_psize
, write_sz
, headroom
;
7221 l2arc_write_callback_t
*cb
;
7223 uint64_t guid
= spa_load_guid(spa
);
7226 ASSERT3P(dev
->l2ad_vdev
, !=, NULL
);
7229 write_sz
= write_asize
= write_psize
= 0;
7231 head
= kmem_cache_alloc(hdr_l2only_cache
, KM_PUSHPAGE
);
7232 arc_hdr_set_flags(head
, ARC_FLAG_L2_WRITE_HEAD
| ARC_FLAG_HAS_L2HDR
);
7235 * Copy buffers for L2ARC writing.
7237 for (try = 0; try < L2ARC_FEED_TYPES
; try++) {
7238 multilist_sublist_t
*mls
= l2arc_sublist_lock(try);
7239 uint64_t passed_sz
= 0;
7241 VERIFY3P(mls
, !=, NULL
);
7244 * L2ARC fast warmup.
7246 * Until the ARC is warm and starts to evict, read from the
7247 * head of the ARC lists rather than the tail.
7249 if (arc_warm
== B_FALSE
)
7250 hdr
= multilist_sublist_head(mls
);
7252 hdr
= multilist_sublist_tail(mls
);
7254 headroom
= target_sz
* l2arc_headroom
;
7255 if (zfs_compressed_arc_enabled
)
7256 headroom
= (headroom
* l2arc_headroom_boost
) / 100;
7258 for (; hdr
; hdr
= hdr_prev
) {
7259 kmutex_t
*hash_lock
;
7260 uint64_t asize
, size
;
7263 if (arc_warm
== B_FALSE
)
7264 hdr_prev
= multilist_sublist_next(mls
, hdr
);
7266 hdr_prev
= multilist_sublist_prev(mls
, hdr
);
7268 hash_lock
= HDR_LOCK(hdr
);
7269 if (!mutex_tryenter(hash_lock
)) {
7271 * Skip this buffer rather than waiting.
7276 passed_sz
+= HDR_GET_LSIZE(hdr
);
7277 if (passed_sz
> headroom
) {
7281 mutex_exit(hash_lock
);
7285 if (!l2arc_write_eligible(guid
, hdr
)) {
7286 mutex_exit(hash_lock
);
7290 if ((write_asize
+ HDR_GET_LSIZE(hdr
)) > target_sz
) {
7292 mutex_exit(hash_lock
);
7298 * Insert a dummy header on the buflist so
7299 * l2arc_write_done() can find where the
7300 * write buffers begin without searching.
7302 mutex_enter(&dev
->l2ad_mtx
);
7303 list_insert_head(&dev
->l2ad_buflist
, head
);
7304 mutex_exit(&dev
->l2ad_mtx
);
7307 sizeof (l2arc_write_callback_t
), KM_SLEEP
);
7308 cb
->l2wcb_dev
= dev
;
7309 cb
->l2wcb_head
= head
;
7310 pio
= zio_root(spa
, l2arc_write_done
, cb
,
7314 hdr
->b_l2hdr
.b_dev
= dev
;
7315 hdr
->b_l2hdr
.b_hits
= 0;
7317 hdr
->b_l2hdr
.b_daddr
= dev
->l2ad_hand
;
7318 arc_hdr_set_flags(hdr
,
7319 ARC_FLAG_L2_WRITING
| ARC_FLAG_HAS_L2HDR
);
7321 mutex_enter(&dev
->l2ad_mtx
);
7322 list_insert_head(&dev
->l2ad_buflist
, hdr
);
7323 mutex_exit(&dev
->l2ad_mtx
);
7326 * We rely on the L1 portion of the header below, so
7327 * it's invalid for this header to have been evicted out
7328 * of the ghost cache, prior to being written out. The
7329 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
7331 ASSERT(HDR_HAS_L1HDR(hdr
));
7333 ASSERT3U(HDR_GET_PSIZE(hdr
), >, 0);
7334 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
7335 ASSERT3U(arc_hdr_size(hdr
), >, 0);
7336 size
= arc_hdr_size(hdr
);
7338 (void) refcount_add_many(&dev
->l2ad_alloc
, size
, hdr
);
7341 * Normally the L2ARC can use the hdr's data, but if
7342 * we're sharing data between the hdr and one of its
7343 * bufs, L2ARC needs its own copy of the data so that
7344 * the ZIO below can't race with the buf consumer. To
7345 * ensure that this copy will be available for the
7346 * lifetime of the ZIO and be cleaned up afterwards, we
7347 * add it to the l2arc_free_on_write queue.
7349 if (!HDR_SHARED_DATA(hdr
)) {
7350 to_write
= hdr
->b_l1hdr
.b_pabd
;
7352 to_write
= abd_alloc_for_io(size
,
7353 HDR_ISTYPE_METADATA(hdr
));
7354 abd_copy(to_write
, hdr
->b_l1hdr
.b_pabd
, size
);
7355 l2arc_free_abd_on_write(to_write
, size
,
7358 wzio
= zio_write_phys(pio
, dev
->l2ad_vdev
,
7359 hdr
->b_l2hdr
.b_daddr
, size
, to_write
,
7360 ZIO_CHECKSUM_OFF
, NULL
, hdr
,
7361 ZIO_PRIORITY_ASYNC_WRITE
,
7362 ZIO_FLAG_CANFAIL
, B_FALSE
);
7364 write_sz
+= HDR_GET_LSIZE(hdr
);
7365 DTRACE_PROBE2(l2arc__write
, vdev_t
*, dev
->l2ad_vdev
,
7368 write_asize
+= size
;
7370 * Keep the clock hand suitably device-aligned.
7372 asize
= vdev_psize_to_asize(dev
->l2ad_vdev
, size
);
7373 write_psize
+= asize
;
7374 dev
->l2ad_hand
+= asize
;
7376 mutex_exit(hash_lock
);
7378 (void) zio_nowait(wzio
);
7381 multilist_sublist_unlock(mls
);
7387 /* No buffers selected for writing? */
7390 ASSERT(!HDR_HAS_L1HDR(head
));
7391 kmem_cache_free(hdr_l2only_cache
, head
);
7395 ASSERT3U(write_asize
, <=, target_sz
);
7396 ARCSTAT_BUMP(arcstat_l2_writes_sent
);
7397 ARCSTAT_INCR(arcstat_l2_write_bytes
, write_asize
);
7398 ARCSTAT_INCR(arcstat_l2_size
, write_sz
);
7399 ARCSTAT_INCR(arcstat_l2_asize
, write_asize
);
7400 vdev_space_update(dev
->l2ad_vdev
, write_asize
, 0, 0);
7403 * Bump device hand to the device start if it is approaching the end.
7404 * l2arc_evict() will already have evicted ahead for this case.
7406 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- target_sz
)) {
7407 dev
->l2ad_hand
= dev
->l2ad_start
;
7408 dev
->l2ad_first
= B_FALSE
;
7411 dev
->l2ad_writing
= B_TRUE
;
7412 (void) zio_wait(pio
);
7413 dev
->l2ad_writing
= B_FALSE
;
7415 return (write_asize
);
7419 * This thread feeds the L2ARC at regular intervals. This is the beating
7420 * heart of the L2ARC.
7423 l2arc_feed_thread(void)
7428 uint64_t size
, wrote
;
7429 clock_t begin
, next
= ddi_get_lbolt();
7430 fstrans_cookie_t cookie
;
7432 CALLB_CPR_INIT(&cpr
, &l2arc_feed_thr_lock
, callb_generic_cpr
, FTAG
);
7434 mutex_enter(&l2arc_feed_thr_lock
);
7436 cookie
= spl_fstrans_mark();
7437 while (l2arc_thread_exit
== 0) {
7438 CALLB_CPR_SAFE_BEGIN(&cpr
);
7439 (void) cv_timedwait_sig(&l2arc_feed_thr_cv
,
7440 &l2arc_feed_thr_lock
, next
);
7441 CALLB_CPR_SAFE_END(&cpr
, &l2arc_feed_thr_lock
);
7442 next
= ddi_get_lbolt() + hz
;
7445 * Quick check for L2ARC devices.
7447 mutex_enter(&l2arc_dev_mtx
);
7448 if (l2arc_ndev
== 0) {
7449 mutex_exit(&l2arc_dev_mtx
);
7452 mutex_exit(&l2arc_dev_mtx
);
7453 begin
= ddi_get_lbolt();
7456 * This selects the next l2arc device to write to, and in
7457 * doing so the next spa to feed from: dev->l2ad_spa. This
7458 * will return NULL if there are now no l2arc devices or if
7459 * they are all faulted.
7461 * If a device is returned, its spa's config lock is also
7462 * held to prevent device removal. l2arc_dev_get_next()
7463 * will grab and release l2arc_dev_mtx.
7465 if ((dev
= l2arc_dev_get_next()) == NULL
)
7468 spa
= dev
->l2ad_spa
;
7469 ASSERT3P(spa
, !=, NULL
);
7472 * If the pool is read-only then force the feed thread to
7473 * sleep a little longer.
7475 if (!spa_writeable(spa
)) {
7476 next
= ddi_get_lbolt() + 5 * l2arc_feed_secs
* hz
;
7477 spa_config_exit(spa
, SCL_L2ARC
, dev
);
7482 * Avoid contributing to memory pressure.
7484 if (arc_reclaim_needed()) {
7485 ARCSTAT_BUMP(arcstat_l2_abort_lowmem
);
7486 spa_config_exit(spa
, SCL_L2ARC
, dev
);
7490 ARCSTAT_BUMP(arcstat_l2_feeds
);
7492 size
= l2arc_write_size();
7495 * Evict L2ARC buffers that will be overwritten.
7497 l2arc_evict(dev
, size
, B_FALSE
);
7500 * Write ARC buffers.
7502 wrote
= l2arc_write_buffers(spa
, dev
, size
);
7505 * Calculate interval between writes.
7507 next
= l2arc_write_interval(begin
, size
, wrote
);
7508 spa_config_exit(spa
, SCL_L2ARC
, dev
);
7510 spl_fstrans_unmark(cookie
);
7512 l2arc_thread_exit
= 0;
7513 cv_broadcast(&l2arc_feed_thr_cv
);
7514 CALLB_CPR_EXIT(&cpr
); /* drops l2arc_feed_thr_lock */
7519 l2arc_vdev_present(vdev_t
*vd
)
7523 mutex_enter(&l2arc_dev_mtx
);
7524 for (dev
= list_head(l2arc_dev_list
); dev
!= NULL
;
7525 dev
= list_next(l2arc_dev_list
, dev
)) {
7526 if (dev
->l2ad_vdev
== vd
)
7529 mutex_exit(&l2arc_dev_mtx
);
7531 return (dev
!= NULL
);
7535 * Add a vdev for use by the L2ARC. By this point the spa has already
7536 * validated the vdev and opened it.
7539 l2arc_add_vdev(spa_t
*spa
, vdev_t
*vd
)
7541 l2arc_dev_t
*adddev
;
7543 ASSERT(!l2arc_vdev_present(vd
));
7546 * Create a new l2arc device entry.
7548 adddev
= kmem_zalloc(sizeof (l2arc_dev_t
), KM_SLEEP
);
7549 adddev
->l2ad_spa
= spa
;
7550 adddev
->l2ad_vdev
= vd
;
7551 adddev
->l2ad_start
= VDEV_LABEL_START_SIZE
;
7552 adddev
->l2ad_end
= VDEV_LABEL_START_SIZE
+ vdev_get_min_asize(vd
);
7553 adddev
->l2ad_hand
= adddev
->l2ad_start
;
7554 adddev
->l2ad_first
= B_TRUE
;
7555 adddev
->l2ad_writing
= B_FALSE
;
7556 list_link_init(&adddev
->l2ad_node
);
7558 mutex_init(&adddev
->l2ad_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7560 * This is a list of all ARC buffers that are still valid on the
7563 list_create(&adddev
->l2ad_buflist
, sizeof (arc_buf_hdr_t
),
7564 offsetof(arc_buf_hdr_t
, b_l2hdr
.b_l2node
));
7566 vdev_space_update(vd
, 0, 0, adddev
->l2ad_end
- adddev
->l2ad_hand
);
7567 refcount_create(&adddev
->l2ad_alloc
);
7570 * Add device to global list
7572 mutex_enter(&l2arc_dev_mtx
);
7573 list_insert_head(l2arc_dev_list
, adddev
);
7574 atomic_inc_64(&l2arc_ndev
);
7575 mutex_exit(&l2arc_dev_mtx
);
7579 * Remove a vdev from the L2ARC.
7582 l2arc_remove_vdev(vdev_t
*vd
)
7584 l2arc_dev_t
*dev
, *nextdev
, *remdev
= NULL
;
7587 * Find the device by vdev
7589 mutex_enter(&l2arc_dev_mtx
);
7590 for (dev
= list_head(l2arc_dev_list
); dev
; dev
= nextdev
) {
7591 nextdev
= list_next(l2arc_dev_list
, dev
);
7592 if (vd
== dev
->l2ad_vdev
) {
7597 ASSERT3P(remdev
, !=, NULL
);
7600 * Remove device from global list
7602 list_remove(l2arc_dev_list
, remdev
);
7603 l2arc_dev_last
= NULL
; /* may have been invalidated */
7604 atomic_dec_64(&l2arc_ndev
);
7605 mutex_exit(&l2arc_dev_mtx
);
7608 * Clear all buflists and ARC references. L2ARC device flush.
7610 l2arc_evict(remdev
, 0, B_TRUE
);
7611 list_destroy(&remdev
->l2ad_buflist
);
7612 mutex_destroy(&remdev
->l2ad_mtx
);
7613 refcount_destroy(&remdev
->l2ad_alloc
);
7614 kmem_free(remdev
, sizeof (l2arc_dev_t
));
7620 l2arc_thread_exit
= 0;
7622 l2arc_writes_sent
= 0;
7623 l2arc_writes_done
= 0;
7625 mutex_init(&l2arc_feed_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
7626 cv_init(&l2arc_feed_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
7627 mutex_init(&l2arc_dev_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7628 mutex_init(&l2arc_free_on_write_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7630 l2arc_dev_list
= &L2ARC_dev_list
;
7631 l2arc_free_on_write
= &L2ARC_free_on_write
;
7632 list_create(l2arc_dev_list
, sizeof (l2arc_dev_t
),
7633 offsetof(l2arc_dev_t
, l2ad_node
));
7634 list_create(l2arc_free_on_write
, sizeof (l2arc_data_free_t
),
7635 offsetof(l2arc_data_free_t
, l2df_list_node
));
7642 * This is called from dmu_fini(), which is called from spa_fini();
7643 * Because of this, we can assume that all l2arc devices have
7644 * already been removed when the pools themselves were removed.
7647 l2arc_do_free_on_write();
7649 mutex_destroy(&l2arc_feed_thr_lock
);
7650 cv_destroy(&l2arc_feed_thr_cv
);
7651 mutex_destroy(&l2arc_dev_mtx
);
7652 mutex_destroy(&l2arc_free_on_write_mtx
);
7654 list_destroy(l2arc_dev_list
);
7655 list_destroy(l2arc_free_on_write
);
7661 if (!(spa_mode_global
& FWRITE
))
7664 (void) thread_create(NULL
, 0, l2arc_feed_thread
, NULL
, 0, &p0
,
7665 TS_RUN
, defclsyspri
);
7671 if (!(spa_mode_global
& FWRITE
))
7674 mutex_enter(&l2arc_feed_thr_lock
);
7675 cv_signal(&l2arc_feed_thr_cv
); /* kick thread out of startup */
7676 l2arc_thread_exit
= 1;
7677 while (l2arc_thread_exit
!= 0)
7678 cv_wait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
);
7679 mutex_exit(&l2arc_feed_thr_lock
);
7682 #if defined(_KERNEL) && defined(HAVE_SPL)
7683 EXPORT_SYMBOL(arc_buf_size
);
7684 EXPORT_SYMBOL(arc_write
);
7685 EXPORT_SYMBOL(arc_read
);
7686 EXPORT_SYMBOL(arc_buf_info
);
7687 EXPORT_SYMBOL(arc_getbuf_func
);
7688 EXPORT_SYMBOL(arc_add_prune_callback
);
7689 EXPORT_SYMBOL(arc_remove_prune_callback
);
7691 module_param(zfs_arc_min
, ulong
, 0644);
7692 MODULE_PARM_DESC(zfs_arc_min
, "Min arc size");
7694 module_param(zfs_arc_max
, ulong
, 0644);
7695 MODULE_PARM_DESC(zfs_arc_max
, "Max arc size");
7697 module_param(zfs_arc_meta_limit
, ulong
, 0644);
7698 MODULE_PARM_DESC(zfs_arc_meta_limit
, "Meta limit for arc size");
7700 module_param(zfs_arc_meta_limit_percent
, ulong
, 0644);
7701 MODULE_PARM_DESC(zfs_arc_meta_limit_percent
,
7702 "Percent of arc size for arc meta limit");
7704 module_param(zfs_arc_meta_min
, ulong
, 0644);
7705 MODULE_PARM_DESC(zfs_arc_meta_min
, "Min arc metadata");
7707 module_param(zfs_arc_meta_prune
, int, 0644);
7708 MODULE_PARM_DESC(zfs_arc_meta_prune
, "Meta objects to scan for prune");
7710 module_param(zfs_arc_meta_adjust_restarts
, int, 0644);
7711 MODULE_PARM_DESC(zfs_arc_meta_adjust_restarts
,
7712 "Limit number of restarts in arc_adjust_meta");
7714 module_param(zfs_arc_meta_strategy
, int, 0644);
7715 MODULE_PARM_DESC(zfs_arc_meta_strategy
, "Meta reclaim strategy");
7717 module_param(zfs_arc_grow_retry
, int, 0644);
7718 MODULE_PARM_DESC(zfs_arc_grow_retry
, "Seconds before growing arc size");
7720 module_param(zfs_arc_p_aggressive_disable
, int, 0644);
7721 MODULE_PARM_DESC(zfs_arc_p_aggressive_disable
, "disable aggressive arc_p grow");
7723 module_param(zfs_arc_p_dampener_disable
, int, 0644);
7724 MODULE_PARM_DESC(zfs_arc_p_dampener_disable
, "disable arc_p adapt dampener");
7726 module_param(zfs_arc_shrink_shift
, int, 0644);
7727 MODULE_PARM_DESC(zfs_arc_shrink_shift
, "log2(fraction of arc to reclaim)");
7729 module_param(zfs_arc_p_min_shift
, int, 0644);
7730 MODULE_PARM_DESC(zfs_arc_p_min_shift
, "arc_c shift to calc min/max arc_p");
7732 module_param(zfs_arc_average_blocksize
, int, 0444);
7733 MODULE_PARM_DESC(zfs_arc_average_blocksize
, "Target average block size");
7735 module_param(zfs_compressed_arc_enabled
, int, 0644);
7736 MODULE_PARM_DESC(zfs_arc_average_blocksize
, "Disable compressed arc buffers");
7738 module_param(zfs_arc_min_prefetch_lifespan
, int, 0644);
7739 MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan
, "Min life of prefetch block");
7741 module_param(zfs_arc_num_sublists_per_state
, int, 0644);
7742 MODULE_PARM_DESC(zfs_arc_num_sublists_per_state
,
7743 "Number of sublists used in each of the ARC state lists");
7745 module_param(l2arc_write_max
, ulong
, 0644);
7746 MODULE_PARM_DESC(l2arc_write_max
, "Max write bytes per interval");
7748 module_param(l2arc_write_boost
, ulong
, 0644);
7749 MODULE_PARM_DESC(l2arc_write_boost
, "Extra write bytes during device warmup");
7751 module_param(l2arc_headroom
, ulong
, 0644);
7752 MODULE_PARM_DESC(l2arc_headroom
, "Number of max device writes to precache");
7754 module_param(l2arc_headroom_boost
, ulong
, 0644);
7755 MODULE_PARM_DESC(l2arc_headroom_boost
, "Compressed l2arc_headroom multiplier");
7757 module_param(l2arc_feed_secs
, ulong
, 0644);
7758 MODULE_PARM_DESC(l2arc_feed_secs
, "Seconds between L2ARC writing");
7760 module_param(l2arc_feed_min_ms
, ulong
, 0644);
7761 MODULE_PARM_DESC(l2arc_feed_min_ms
, "Min feed interval in milliseconds");
7763 module_param(l2arc_noprefetch
, int, 0644);
7764 MODULE_PARM_DESC(l2arc_noprefetch
, "Skip caching prefetched buffers");
7766 module_param(l2arc_feed_again
, int, 0644);
7767 MODULE_PARM_DESC(l2arc_feed_again
, "Turbo L2ARC warmup");
7769 module_param(l2arc_norw
, int, 0644);
7770 MODULE_PARM_DESC(l2arc_norw
, "No reads during writes");
7772 module_param(zfs_arc_lotsfree_percent
, int, 0644);
7773 MODULE_PARM_DESC(zfs_arc_lotsfree_percent
,
7774 "System free memory I/O throttle in bytes");
7776 module_param(zfs_arc_sys_free
, ulong
, 0644);
7777 MODULE_PARM_DESC(zfs_arc_sys_free
, "System free memory target size in bytes");
7779 module_param(zfs_arc_dnode_limit
, ulong
, 0644);
7780 MODULE_PARM_DESC(zfs_arc_dnode_limit
, "Minimum bytes of dnodes in arc");
7782 module_param(zfs_arc_dnode_limit_percent
, ulong
, 0644);
7783 MODULE_PARM_DESC(zfs_arc_dnode_limit_percent
,
7784 "Percent of ARC meta buffers for dnodes");
7786 module_param(zfs_arc_dnode_reduce_percent
, ulong
, 0644);
7787 MODULE_PARM_DESC(zfs_arc_dnode_reduce_percent
,
7788 "Percentage of excess dnodes to try to unpin");